Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

default · ~99.6k rows (showing the first 120k)

Split (2)

train · ~72.4k rows (showing the first 93.1k)

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1505.02968	1	1505	2015-05-12T11:43:05	Certain actions of finitely generated groups on higher dimensional noncommutative tori	[ "math.OA" ]	We study certain actions of finitely generated abelian groups on higher dimensional noncommutative tori. Given a dimension $d$ and a finitely generated abelian group $G$, we apply a certain function to detect whether there is a simple noncommutative $d$-torus which admits a specific action by $G$. We also describe the condition of $G$ under which the associated crossed product is an AF algebra.	math.OA	math	CERTAIN ACTIONS OF FINITELY GENERATED GROUPS ON HIGHER DIMENSIONAL NONCOMMUTATIVE TORI ZHUOFENG HE Abstract. We study certain actions of finitely generated abelian groups on higher dimensional noncommutative tori. Given a dimension d and a finitely generated abelian group G, we apply a certain function to de- tect whether there is a simple noncommutative d-torus which admits a specific action by G. We also describe the condition of G under which the associated crossed product is an AF algebra. . A O h t a m [ 1 v 8 6 9 2 0 . 5 0 5 1 : v i X r a Contents 1. Introduction 2. Preliminaries 3. Realization of cyclic groups acting on noncomutative d-tori 4. Case d − W (n) = 1 and actions on odd-dimensional noncommutative tori 5. Finitely generated abelian group actions Acknowledgment References 1 4 8 14 17 18 19 1. Introduction Certain finite group actions on the rotation algebras have already been systematically studied in [1] by S. Echterhoff, W. Luck, N. C. Phillips and S. Walters. It is known that up to conjugacy, a finite subgroup F of SL2(Z) satisfies F = Zk where k = 2, 3, 4 or 6. According to their paper, by realizing certain generators for Zk, the action of Zk on a rotation algebra can be given for each such k. We note that this action is the natural generalization of the action of SL2(Z) on a torus T2 = R2/Z2. If we require the rotation algebra to be a simple one, i.e. for the irrational rotation algebra, denoted by Aθ where θ is an irrational number, the crosed product Aθ ⋊ Zk is an AF- algebra for k = 2, 3, 4 and 6. Moreover for each k, the K0-group K0(Aθ ⋊Zk) ∼= Aθ′ ⋊ Zk′ if and only if has been calculated, which shows that Aθ ⋊ Zk k = k′ and θ = θ′ mod Z, see [1, Theorem 0.1]. Also they show that higher dimensional noncommutative tori admit flip actions of Z2, and for simple ones the crossed products are AF algebras. 1 Certain actions of finitely generated groups on higher dimensional noncommutative tori 2 As stated in their paper, the proof of the above theorem breaks into the following three steps. (1) Computation of the K-theory of the crossed product. (2) Proof that the crossed product satisfies the Universal Coefficient Theorem. (3) Proof that the action has the tracial Rohlin property. Their argument and proof in step (2) and (3) are general enough so that they are also effective for higher dimensional cases. Hence it is accessible if one obtains a proper realization of some finite group, a proper generalization of actions above to higher dimensional cases, and a way to compute the K- theory of the crossed product. In [2], the authors manage to obtain results in this way. They give a proper definition of such actions which generalize the action above, which they call it by "canonical actions". Throughout this paper we are interested in actions of this type. A detailed discussion of such actions is in the Preliminaries. By realizing a cyclic group with the com- panion matrix of a cyclotomic polynomial, the authors study such actions of cyclic group on simple higher dimensional tori for certain dimensions. To finish step (1), i.e. to compute the K-groups of the associated crossed prod- uct, the authors first apply [1, Lemma 2.1] and [5, Theorem 4.1] to write the crossed product in terms of another twisted group C ∗-algebra, then to com- pute its K-theory, according to [1, Theorem 0.3], it is the same to compute the K-theory of untwisted group C ∗-algebra, and finally by [4, Theorem 0.1] they manage to obtain desired the K-groups. The authors then show that for certain dimension d and order n there is a simple noncommutative d-torus on which Zn acts and the K1-group of the crossed product is not trivial, see [2, Theorem 3.6]. We continue their work and study such actions of finitely generated abelian groups on simple higher dimensional noncommutative tori. By combining the tensor structure of higher dimensional noncommutatvie tori and the re- alization method in [2] for cyclic groups and actions, we manage to realize all of the possible actions under cyclic groups. Then for a given dimension d by the structure of finite cyclic group of GLd(Zd), we apply a certain func- tion to tell when we can realize the action of the cyclic group on a simple noncommutative d-torus. Moreover we describe the condition under which the crossed product is an AF algebra. The function W (n) is defined later. Hereby we remark that through this paper we consider noncommutative d-tori, where we usually suppose d > 1, since there is no 1-dimensional non- commutative tori. However some of our results are actually compatible with the case when we allow d = 1. Theorem 1.1 (Theorem 3.1). Given a dimension d and an order n > 1, there is an action of Zn on a simple noncommutative d-tori AΘ if either d − W (n) > 1 or d − W (n) = 0. The crossed product AΘ ⋊ Zn is an AT algebra, and it is an AF algebra if and only if d − W (n) = 0, and n either admits a form of 2m where m = 3j5ipel l , or admits a form of 2k3j5i where Certain actions of finitely generated groups on higher dimensional noncommutative tori 3 k 6= 1, j ≤ 2, i ≤ 1 and el are all nonnegative integers, and pl > 5 is a prime number. Then we generalize the function W for finite abelian groups, and realize actions with a similar argument. We obtain the following. Theorem 1.2 (Theorem 3.6). Given a dimension d and an abelian finite subgroup G ≤ GLd(Z) with d − W (G) > 1 or d − W (G) = 0, there is an action of G on a simple noncommutative d-torus AΘ. The crossed product AΘ ⋊G is an AT algebra, and it is an AF algebra if and only if d−W (G) = 0 and, if we write G as G ∼=Qt j=1 Znj such that W (G) = W (Znl), t Xl=1 each nl either admits a form of 2m where m = 3j5iper r , or admits a form of 2k3j5i where k 6= 1, j ≤ 2, i ≤ 1 and er are all nonnegative integers, and pr > 5 is a prime number. Especially we have Corollary 3.7 which states that the only possibility of a finite abelian group F acting on a irrational rotation algebra is when F = Zk where k = 2, 3, 4 and 6. In Section 3 we prove above theorems and discuss related results. Then we find out that the function W contains more information. We have the following. Theorem 1.3 (Theorem 4.3). Given a dimension d and an order n, if d − W (n) = 1, then there is no simple noncommutative torus on which Zn acts in the above way. In Section 4 we give the proof of this theorem. With enough knowledge of the function W when considering cyclic group actions, we move to the topic of finitely generated abelian group acting on simple higher dimensional noncommutative tori. The action of torsion free part of the finitely generated group is defined to be adding dimensions concerning the crossed product. Similarly we study the condition by which the crossed product is an AF algebra. Theorem 1.4 (Theorem 5.1). For a given dimension d and a finitely gen- erated abelian group G ∼= (cid:16)Qs i=1 Zp ei i (cid:17) × Zr, if d − W (n) = 1 and r > 0, or d − W (n) 6= 1 then there is a noncommutative d-torus AΘ admitting an action of G, denoted by α : G → Aut(AΘ). We require AΘ is pseudo-simple in the first case and simple in the second case. Then the crossed product AΘ ⋊α G is a simple AT algebra, and it is an AF algebra if and only if r = 0, d − W (G) = 0 and G satisfies the last condition in Theorem 1.2. A detailed discussion is in Section 5. Certain actions of finitely generated groups on higher dimensional noncommutative tori 4 2. Preliminaries For a second countable locally compact Hausdorff topological group G with its modular function ∆G : G → (0, +∞) and a Borel 2-cocycle ω ∈ full) twisted group C ∗- Z(G, T), we define the associated reduced (resp. r (G, ω) (resp. C ∗(G, ω)) in the following way. Regard L1(G) as algebras, C ∗ a vector space, and equip it with twisted convolution given by f ∗ω g(t) =ZG and the involution given by f (s)g(s−1t)ω(s, s−1t)ds f ∗(t) = ∆G(t−1)ω(t, t−1)f (t−1). Then L1(G, ω) := (L1(G), ∗ω , ∗) becomes a ∗-algebra which is named the twisted convolution algebra. As similar to the definition of group C ∗-algebras, to consider nondegenerate representations of L1(G, ω), we turn to a twisted analogue of unitary representations of G. More precisely, an ω-representation of G on a Hilbert space H is a Borel map V : G → U (H) such that V (t)V (s) = ω(t, s)V (ts) where U (H) stands for the unitary group of H endowed with the strong operator topology. For example we may define Lω : G → U (L2(G)), the regular ω-representation of G, by Lω(s)ξ(t) = ω(s, s−1t)ξ(s−1t) for all ξ ∈ L2(G). For arbitrary ω-representation V : G → U (H), define π : L1(G, ω) → B(H) via π(f ) =ZG f (s)V (s)ds for f ∈ L1(G, ω). Then π is a contractive ∗-homomorphism. As in the definition of a group C ∗-algebra, with abuse of notations we may also denote π by V : L1(G, ω) → B(H). Hence we can define the reduced twisted group C ∗-algebra as C ∗ r (G, ω) = Lω(L1(G, ω)). Moreover every nondegenerate representation of L1(G, ω) is induced by an ω-representation of G. We can see this by taking a norm one approximate identity in L1(G, ω). Thus we obtain the universal representation πu of L1(G, ω), and define the full twisted group C ∗-algebra as C ∗(G, ω) = πu(L1(G, ω)). We remark that C ∗(G, ω) has the universal property and the ∗-homomorphism Lω : C ∗(G, ω) → C ∗ C ∗-algebras coincide if G is amenable. r (G, ω) is surjective. The reduced and full twisted group We realize higher dimensional noncommutative tori with twisted group C ∗-algebras. For a given dimension d, take G = Zd. Throughout this paper we denote by Td(R) the set of all d × d skew symmetric real matrices. Then Certain actions of finitely generated groups on higher dimensional noncommutative tori 5 for Θ = (θij) ∈ Td(R) define the induced 2-cocycle ωΘ : Zd × Zd → T by formula ωΘ(x, y) = exp(πihΘx, yi) for x, y ∈ Zd. Then AΘ := C ∗(Zd, ωΘ) is called a noncommutative d-torus. Since Zd is discrete and amenable, if we take the standard basis of it, say ei for i = 1, . . . , d, and denote the regular ωΘ-representation by lΘ : Zd → U (ℓ2(Zd)), then we have AΘ := C ∗(Zd, ωΘ) = C ∗{lΘ(ei) i = 1, . . . , d} where each lΘ(ei) is a unitary and they satisfy the commuting relations lΘ(ei)lΘ(ej) = ωΘ(ei, ej)2lΘ(ej )lΘ(ei) = exp(2πiθji)lΘ(ej )lΘ(ei). This is clearly a generalization of the rotation algebra to the d-dimensional case. We refer to the rotation algebra which is the case when d = 2. One may take the zero d × d matrix as Θ and then get AΘ = C(Td), however this is not so much in our interests since it is not simple. Thus we have the following. Definition 2.1. The matrix Θ ∈ Td(R) is said to be nondegerate if when- ever x ∈ Zd satisfies exp(2πihΘx, yi) = 1 for all y ∈ Zd, we have x = 0. Theorem 2.2 ([6, Theorem 1.9]). Let Θ ∈ Td(R). Then the noncommuta- tive d-torus AΘ is simple if and only if so is Θ. Proposition 2.3. For Θi ∈ Tdi(R), i = 1, . . . , m, denote Θ := Lm Td(R) where d := Pm i=1 Θi ∈ i=1 di. Then Θ is nondegenerate if and only if each Θi is nondegenerate. Proof. By induction we only show the the case when m = 2. If Θ1 and Θ2 are both nondegenerate, for x = (x1, x2) ∈ Zd satisfying exp(2πihΘx, yi) = 1 for all y ∈ Zd, by taking y1 ∈ Zd1 × {0} and y2 ∈ {0} × Zd2 we obtain exp(2πihΘixi, yii) = 1 for all yi ∈ Zdi, i = 1, 2. Then by nondegeneracy of Θ1 and Θ2, we obtain x = (x1, x2) = 0, in other words Θ is nondegenerate. Conversely if Θ is nondegenerate, then for x1 ∈ Zd1 satisfying exp(2πihΘ1x1, y1i) = 1 for all y1 ∈ Zd1, then we have for x = (x1, 0) ∈ Zd, exp(2πihΘx, yi) = exp(2πihΘ1x1, y1i) = 1 for all y ∈ Zd. Thus x1 = 0, and similarly x2 = 0. Hence Θ1 and Θ2 are both nondegenerate. (cid:3) Then we discuss actions on noncommutatitive tori. We denote by AΘ a noncommutative d-torus as above, which is not necessarily simple. Through the regular representation lΘ : ℓ1(Zd, ωΘ) → B(ℓ2(Zd)), we regard AΘ as a C ∗-subalgebra in B(ℓ2(Zd)). For a matrix A ∈ GLd(Z), we have a unitary uA in B(ℓ2(Zd)) given by uAξ(x) = ξ(A−1x) Certain actions of finitely generated groups on higher dimensional noncommutative tori 6 for ξ ∈ ℓ2(Zd) and x ∈ Zd. Thus we have Ad uA ∈ Aut(B(ℓ2(Zd))). Then we have an action denoted by Ad u• : GLd(Z) → Aut(B(ℓ2(Zd))). Then for AΘ = C ∗{lΘ(ei) i = 1, . . . , d} = span{lΘ(x) x ∈ Zd}, we have the fomula Ad(uA)(lΘ(x))ξ(y) = l(A−1)tΘA−1(Ax)ξ(y) for ξ ∈ ℓ2(Zd) and y ∈ Zd. Thus for A ∈ GLd(Z) such that Θ = (A−1)tΘA−1 we have Ad(uA) ∈ Aut(AΘ) by restriction. Then denote GΘ := {A ∈ GLd(Z) Θ = AtΘA}, we obtain an action Ad u• : GΘ → Aut(AΘ). This is an action on AΘ which generalizes the action of SL2(Z) on the rotation algebra Aθ discussed in [1]. Particularly, when the dimension d = 2, we have GΘ = SL2(Z). We are especially interested in such actions of some subgroup G of GΘ on AΘ. Then we roughly state classification results for our cases, and refer to [1], [2],[6] and [7] for details. A simple AΘ is known to be with a unique tracial state and has tracial rank zero. For an action of a finite group on it, which has the tracial Rokhlin property, say α : G → Aut(AΘ), the resulting crossed product AΘ ⋊α G becomes also simple, with a unique tracial state and with tracial rank zero. Particularly, for a finite subgroup G ≤ GΘ, the action described in last paragraph of G on simple AΘ is known to have the tracial Rokhlin property. Moreover the resulting crossed product AΘ ⋊ G satisfies the Universal Coefficient Theorem. Thus the crossed products of such actions becomes classifiable. We also need the following theorem. Theorem 2.4 ([6, Preposition 3.7]). Let A be a simple infinite dimensional separable unital nuclear C ∗-algebra with tracial rank zero and which satisfies the Universal Coefficient Theorem. Then A is a simple AH algebra with real rank zero and no dimension growth. If K∗(A) is torsion free, A is an AT algebra. If, in addition, K1(A) = 0, then A is an AF algebra. Then we introduce results of J. A. Jeong and J. H. Lee in [2]. We use their realization method as a building block to obtain more realizations and actions. For a give dimension d and an order n with d = φ(n), where φ is the Euler's totient function. The nth cyclotomic polynomial Φn(x) is defined by Φn(x) = Y0<k≤n gcd(k,n)=1 (x − exp(2πi k n )). Φn(x) is known to be a monic polynomial of degree d = φ(n), and with integer coefficients. Thus set Certain actions of finitely generated groups on higher dimensional noncommutative tori 7 0 0 1 0 0 1 ... ... 0 0 0 0 0 0 0 . . . 0 0 · · · · · · · · · ... . . . · · · 0 −a0 0 −a1 0 −a2 ... ... 0 −ad−2 1 −ad−1   ,   Cn = which is the companion matrix of Φn(x). Then we have Cn is in GLd(Z) and of order n. Theorem 2.5 ([2, Theorem 4.2]). Let n ≥ 3 and d := φ(n). Then there exist simple d-dimensional tori on which the group Zn = hCni acts canonically. We remark in this theorem the "canonical action" is the action we dis- cussed above and what we are interested in. To compute the K-theory for associated crossed product of such action, we need the following theorem. Theorem 2.6 ([4, Theorem 0.1]). Let n, d ∈ N. Consider the extension of groups 1 → Zd → Zd ⋊α Zn → Zn → 1 such that conjugation action α of Zn on Zd is free outside of the origin 0 ∈ Zd. Then K0(C ∗(Zd ⋊α Zn)) ∼= Zs0 for some s0 ∈ Z and K1(C ∗(Zd ⋊α Zn)) ∼= Zs1, where s1 =Pl≥0 rkZ((V2l+1 Zd)Zn ). If n is even, s1 = 0. If n > 2 is prime and d = n − 1, then s1 = 2n−1−(n−1)2 2n . The authors in [2] apply this method and obtain that for certain dimension d and order n, the resulting crossed product is not an AF algebra. The theorem is stated as the following. Theorem 2.7. [2, Theorem 3.6] Let n ≥ 7 be an odd integer and d := φ(n). Consider the extension of groups 1 → Zd → Zd ⋊α Zn → Zn → 1 with Zn = hCni. If 2d ≥ n + 5, then K1(C ∗(Zd ⋊α Zn)) 6= 0. We denote by N the smallest subcategory of the category of separable C ∗-algebras, which contains separable Type I algebras and is closed under taking ideals, quotients, extensions, inductive limits, stable isomorphisms and crossed products by Z and R, then the following theorem holds. Theorem 2.8 ([8, Kunneth theorem]). Let A and B be C ∗-algebras with A in N. Then there is a natural short exact sequence 0 → K∗(A) ⊗ K∗(B) α−→ K∗(A ⊗ B) β −→ Tor(K∗(A), K∗(B)) → 0. The sequence is Z2-graded with deg α = 1, deg β = 1, where Kp ⊗ Kq and Tor(Kp ⊗ Kq) are given degree p + q for p, q ∈ Z2. Certain actions of finitely generated groups on higher dimensional noncommutative tori 8 3. Realization of cyclic groups acting on noncomutative d-tori Given skew symmetric real matrices Θ1 and Θ2, suppose that Θ = Θ1 ⊕ Θ2, then naturally we have AΘ ∼= AΘ1 ⊗ AΘ2. Then for any finite group actions described in the Preliminaries of G1 ≤ GΘ1 on AΘ1, and G2 ≤ GΘ2 on AΘ2, we could naturally get a finite group action of G1 × G2 on AΘ, one can easily verify that G = G1 × G2 ≤ GΘ and the resulting action is the restriction of Ad u• : GLd(Z) → Aut(B(ℓ2(Zd))), where we denote by d the degree of Θ. Moreover we have AΘ ⋊ G ∼= (AΘ1 ⋊ G1) ⊗ (AΘ2 ⋊ G2) by an isomorphism defined by (a ⊗ b, (g, h)) 7→ (a, g) ⊗ (b, h). For a given dimension d, consider GLd(Z). By [3, Theroem 2.7], GLd(Z) if and only W (n) ≤ d, where p1 < i=1 pei i has an element of order n = Qt · · · < pt are all prime numbers and W (n) :=(Pt Pt i=1(pi − 1)pei−1 i=1(pi − 1)pei−1 i i − 1, , pe1 1 = 2; otherwise, thus we are able to obtain finitely many candidates for n, which is the possible order of a realizable cyclic group in GLd(Z). For those n's with φ(n) = d, where φ is the Euler's totient function, by [2, Theorem 4.2] we may find a noncommutative d-torus AΘ which is simple, namely figure out the form of the nondegenerate Θ, and realize an action of Zn on AΘ. Moreover we could compute K∗(AΘ ⋊ Zn) and then know whether AΘ ⋊ Zn is AF or not. However this way does not work for those n's with φ(n) 6= d, paticularly when d is odd. We manage a way of realizing cyclic groups for more candidates n to find actions in the following. If pe1 1 i=1 pei By Kuzmanovich's proof in [3], suppose n = Qt such that W (n) ≤ d. could construct a di × di matrix Ai of order pei i ) = (pi − 1)pei−1 We note that φ(pei W (n) × W (n) matrix with W (n) ≤ d. i with p1 < · · · < pt 6= 2, by applying the companion matrix we i where di = (pi − 1)pei−1 . i=1 Ai, which is a . Then set B := Lt Thus by [2, Theorem 4.2], for each i, we could find nondegenerate Θi acts on AΘi. If d − W (n) > 1, set A := in Tdi(R) such that hAii = Zp B ⊕ Id−W (n), which is a d × d matrix of order n. It is possible to find a nondegenerate Θ0. Then set ei i i i Θ :=(cid:16) t Mi=1 Θi(cid:17) ⊕ Θ0, Certain actions of finitely generated groups on higher dimensional noncommutative tori 9 which is nondegenerate, thus we have an action of hAi = Zn on AΘ. Moreover by our former discussion one may check that AΘ ⋊ Zn ∼=(cid:16) (AΘi ⋊ Zp ei i )(cid:17) ⊗ AΘ0. t Oi=1 By the method above we may realize more finite cyclic group actions on higher dimensional noncommutative tori. To compute their K-groups, we apply the Kunneth theorem in [8], stated as Theorem 2.8. To verify the conditions for our problem, we have the following confirma- tion. Firstly by [2, Proposition 3.4] each AΘi ⋊ Zp is an AT algebra, and so is AΘ0. Since C(T) ⊗ F is separable and of Type I, where F stands for a finite dimensional C ∗-algebra, then every AΘi ⋊ Zp and AΘ0 are in N. Secondly by [2, Remark 3.2] and [4, Theorem 0.1] we know that for each i, ei i ei i K∗(AΘi ⋊ Zp ei i ) ∼= K∗(C ∗(Zdi ⋊ Zp ei i , ωΘi) ∼= K∗(C ∗(Zdi ⋊ Zp )), ei i which is torsion free. Then applying Kunneth theorem, we obtain K∗(AΘ ⋊ Zn) =(cid:16) K∗(AΘi ⋊ Zp ei i )(cid:17) ⊗ K∗(AΘ0). t Oi=1 By [2, Propersition 3.1] AΘ ⋊ Zn satisfies the Universal Coefficient Theorem. As we have already stated before, AΘ ⋊Zn is then classifiable. We obtain the K∗(AΘ ⋊ Zn) is torsion free, thus under this realization the crossed product AΘ ⋊ Zn is an AT algebra. However, We note that d − W (n) > 1 and K0(AΘ0) ∼= K1(AΘ0) ∼= Z2d−W (n)−1. Thus in this case K1(AΘ ⋊ Zn) 6= 0, and AΘ ⋊ Zn is not an AF-algebra. If d − W (n) = 0, the situation is essentially the same to our discussion i=1 Ai, which is a d × d matrix of order n. Then set before. Instead we set A := B = Lt Θ :=(cid:16) t Mi=1 Θi(cid:17), which is nondegenerate, thus we have an action of hAi = Zn on AΘ. Then we obtain AΘ ⋊ Zn ∼= (AΘi ⋊ Zp ) ei i and again by the Kunneth theorem, t Oi=1 t Oi=1 K∗(AΘ ⋊ Zn) = K∗(AΘi ⋊ Zp ). ei i Thus K1(AΘ ⋊ Zn) = 0 if and only if K1(AΘi ⋊ Zp ) = 0 for each i. ei i Certain actions of finitely generated groups on higher dimensional noncommutative tori 10 Then our attention is driven to each AΘi ⋊ Zp ei i . Since in the case when ) 6= 0 if (pi − 2) ≥ 5. Thus the only candidates of K1(AΘi ⋊ Zp pi is an odd prime, by [2, Theorem 3.6], we have K1(AΘi ⋊ Zp pei−1 ) = 0 are i pei i = 3, 32, 5 or 2k where k > 1. To compute K1-groups, we apply [4, Theorem 0.1], stated as Theorem 2.6. ei i ei i Hence the corresponding K1-groups vanish when pei k > 1. In the following we only check the case when pei above enables us to only compute s1 where i = 3, 5 and 2k where i = 32. The theorem s1 = rkZ((^1 Z6)Z9) + rkZ((^3 Z6)Z9 ) + rkZ((^5 We obtain rkZ((V1 Z6)Z9) = 0 for the action is free outside the origin 0 ∈ Z6. According to our realization we have Z9 is the cyclic group generalized by A in GL6(Z), where Z6)Z9). A =   0 0 0 0 0 −a0 1 0 0 0 0 −a1 0 1 0 0 0 −a2 0 0 1 0 0 −a3 0 0 0 1 0 −a4 0 0 0 0 1 −a5   Φ9(x) = Y0<k<9 gcd(k,9)=1 (x − ζ k) = aixi, ζ = exp(2πi 1 9 ). 5 Xi=0 and Write A = Q−1diag(ζ, ζ 2, ζ 4, ζ 5, ζ 7, ζ 8)Q, where Q ∈ GL6(C). Notice that x a fixed point of A acting on V3 Z if and only if Qx is a fixed point of diag(ζ, ζ 2, ζ 4, ζ 5, ζ 7, ζ 8) acting onV3 QZ. Then for 1 ≤ i < j < k ≤ 6, if diag(ζ, ζ 2, ζ 4, ζ 5, ζ 7, ζ 8)ei ∧ ej ∧ ek = ei ∧ ej ∧ ek, it is the same to stating that there is a way of pick three number in 1, 2, 4, 5, 7, 8 such that the sum of them can be divided by 9, which is impossible. Thus we obtain Hence when pei which Z9 acts in the way we are interested in, and i = 32, we obtain a simple noncommutative 6-torus AΘ on in other words AΘ ⋊ Z9 is an AF algebra. K1(AΘ ⋊ Z9) = 0, Similarly we also obtain Z6)Z9 ) = 0. Z6)Z9 ) = 0. rkZ((^3 rkZ((^5 Certain actions of finitely generated groups on higher dimensional noncommutative tori 11 We leave the case when d − W (n) = 1 to next section. To complete the proof, we discuss the case when given the dimension d and the order n such i=1 pei that W (n) ≤ d, and with pe1 i . By proof in [3] notice that W ( n 2 , by applying the companion matrix we could construct a di × di matrix Ai of order pei i where di = (pi − 1)pei−1 i=2 Ai, which is a W (n) × W (n) matrix with W (n) ≤ d, and of order n 2 . 1 = 2 in the prime factorization of n =Qt for i = 2, . . . , t. Then set B := Lt 2 ) = W (n) ≤ d. Then for n i Again we could find nondegenerate Θi in Tdi(R) such that hAii acts on AΘi 2 ) > 1, 2 . It is possible to in the way we described above for i = 2, . . . , t. If d − W (n) = d − W ( n set A := B ⊕ Id−W (n), which is a d × d matrix of order n find a nondegenerate Θ0 since d − W (n) > 1. Then set Θ :=(cid:16) t Mi=1 Θi(cid:17) ⊕ Θ0, which is nondegenerate. Since (−A)tΘ(−A) = AtΘA = Θ, we have an action of h−Ai = Zn on AΘ. However, in this case we can only write Zn as Zn ∼=(cid:16) Y2≤i≤t i6=j Zp ei i (cid:17) × Z 2p ej j where 2 ≤ j ≤ t. Thus instead of A, we consider A(j) as a generator of Zn where A(j) is defined by A(j) = 0 A1 0 A2 ... ... 0 0 ... ... 0 0 0 0   0 · · · 0 · · · ... . . . · · · −Aj ... ... 0 · · · · · · 0 0 0 0 0 ... ... 0 0 ... . . . 0 At 0 0 0 0 ... 0 ... 0 Id−W (n)   where 1 ≤ j ≤ t. Then we consider actions of Zn = hA(j)i on AΘ. Similarly we have AΘ ⋊ Zn ∼=(cid:16) O2≤i≤t i6=j (AΘi ⋊ Zp ei i )(cid:17) ⊗ (AΘj ⋊ Z ) ⊗ AΘ0, 2p ej j where 2 ≤ j ≤ t. By the Kunneth therorem K∗(AΘ ⋊ Zn) is the tensor of K∗ of each factor. Again since the part of K∗(AΘ0), K1(AΘ ⋊ Zn) 6= 0, which is equivalently to say that AΘ ⋊ Zn is not an AF algebra. Certain actions of finitely generated groups on higher dimensional noncommutative tori 12 It is similar to above when d−W (n) = d−W ( n 2 ) = 0. However we remark that K∗(AΘ ⋊ Zn) ∼=(cid:16) O2≤i≤t i6=j K∗(AΘi ⋊ Zp ei i )(cid:17) ⊗ K∗(AΘj ⋊ Z ) 2p ej j where 2 ≤ j ≤ t. Thus by [4, Theorem 0.1], we have K1(AΘj ⋊ Z2p since 2pej is even. Hence in this case, AΘ ⋊Zn is an AF-algebra exactly when j 2 admits a form of 3j5ipel n l where j, i and el are all nonnegetive integers, j ≤ 2, i ≤ 1 and pl > 5 is a prime number. As a summarize we have the following theorem. ) = 0 ei j Theorem 3.1. Given a dimension d and an order n > 1, there is an action of Zn on a simple noncommutative d-tori AΘ if either d − W (n) > 1 or d − W (n) = 0. The crossed product AΘ ⋊ Zn is an AT algebra, and it is an AF algebra if and only if d − W (n) = 0, and n either admits a form of 2m where m = 3j5ipel l , or admits a form of 2k3j5i where k 6= 1, j ≤ 2, i ≤ 1 and el are all nonnegative integers, and pl > 5 is a prime number. Corollary 3.2. Given a dimension d which is even and order n, there is an action of Zn on simple noncommutative d-tori if and only if W (n) ≤ d. Proof. Since W (n) is always even when n ≥ 3. (cid:3) Note that we have the following example which is already mentioned in [1]. Example 3.3. Consider an action of Zn on a simple 2-dimensional non- commutative torus , say AΘ. Suppose that the action is of the above type. By Corollary 3.2, it is equivalent to say that d − W (n) = 0 or d − W (n) = 2 where d = 2. Then by knowledge of W (n), we obtain n = 2, 3, 4 or 6. Moreover by Theorem 3.1, each crossed product AΘ ⋊ Zn is an AF algebra. Then generally for a finite abelian group G, we may generalize our method of construct such actions of G on noncommutative tori. By the structure theorem for finitely generated abelian groups we write G as s G ∼= Yi=1 we cannot simply define W (G) := Ps where p1 ≤ · · · ≤ ps are primes and ei ≤ ei+1 whenever pi = pi+1. However i ). Since if so, suppose the case when G = Z2 × Z2, in which we have W (G) = 0, but we can only realize it in GL4(Z) as hA, Bi where i=1 W (pei Zp ei i A =  0 −1 0 0 0 −1 0 0 0 1 0 0 1 0 0 0   and B =  0 1 0 0 1 0 0 0 −1 0 0 0 0 0 0 −1 .   Certain actions of finitely generated groups on higher dimensional noncommutative tori 13 1 = 2, and there is no other pei i such that pei i as a normal subgroup, where Z2p i=1 Zp This is because if for a finite abelian group G ∼=Qs , if Z2 is a normal group of it, i.e. pe1 6= 2, then we cannot write Z2 into Z2p is a normal subgroup of G. Thus we cannot realize Z2 as the signature part of the matrix which need not a cost of a dimension, and since the determinant of realized matrix is 1, the degree should be even. Thus instead we have to cost two of dimensions to set a diagonal part as example above. Hence in this case we should modify our test funtion to taking value 2 of Z2. ei i ei i ei i Definition 3.4. For a cyclic group Zn, define the function W as follows. W (Zn) :=(W (n), n 6= 2; n = 2. 2, Definition 3.5. For a finitely generated abelian group G, by the structure theorem for finitely generated abelian groups, we have G ∼=(cid:16) s Yi=1 Zp ei i (cid:17) × Zr, where p1 ≤ · · · ≤ ps are primes and ei ≤ ei+1 whenever pi = pi+1. Let Gtor denote the torsion part Gtor. Define the function W of G as t W (G) := minn W (Znl) Gtor ∼= Xl=1 l Yj=1 Znjo. Theorem 3.6. For a given dimension d and an abelian finite subgroup G ≤ GLd(Z) with d − W (G) > 1 or d − W (G) = 0, there is an action of G on a simple noncommutative d-torus AΘ. The crossed product AΘ ⋊ G is an AT algebra, and it is an AF algebra if and only if d − W (G) = 0 and, if we write G as G ∼=Qt j=1 Znj such that t W (G) = W (Znl), Xl=1 Xl=1 each nl either admits a form of 2m where m = 3j5iper r , or admits a form of 2k3j5i where k 6= 1, j ≤ 2, i ≤ 1 and er are all nonnegative integers, and pr > 5 is a prime number. Proof. By definition we can write G as G ∼=Qt t W (G) = W (Znl). j=1 Znl such that Thus we have d − W (Znl) > 1 or d − W (Znl) = 0 for all l. Then for each j by Theorem 3.1 we obtain a simple noncommutative torus AΘl on which Znl acts in the above way. We emphasis that if nl = 2, we obtain an irrational rotation algebra. Then by the tensor product argument similar in section Certain actions of finitely generated groups on higher dimensional noncommutative tori 14 3 we obtain a simple noncommutative d-torus AΘ on which G acts in the above way. The rest of the proof is similar to the one in section 3 and from Theorem (cid:3) 3.1. Corollary 3.7. The only actions of finite abelian group on irrational rota- tion algebras of the above type are by Zk where k = 2, 3, 4 and 6. Proof. Note that in the current case d = 2. suppose G is a finite abelian group acting on the irrational rotation algebra Aθ in the above way. It is a conflict if W (G) > 2 then by our statement above. Thus W (G) = 2. Then by the definition of the function W we have G = Zk where k = 2, 3, 4 and 6. (cid:3) 4. Case d − W (n) = 1 and actions on odd-dimensional noncommutative tori d − W (n) = 1 for a given dimension d and order n = Qt As a remaining problem of the last section, we need to study the case i with primes p1 < · · · < pt. We start our discussion again from Kuzmanovich's result which shows that for any A ∈ GLd(Z) of order n, there is Q ∈ GLd(Q) ⊆ GLd(R) such that i=1 pei 0 0 ... 0 · · · 0 · · · ... . . . · · · At 0 · · · 1 0    A1 0 0 A2 ... ... 0 0 0 0 where Ai is the companion matrix for pei i . Λ := QAQ−1 =  In general for a subgroup G ≤ GLd(Z) and an element Q ∈ GLd(R) with QGQ−1 ≤ GLd(Z), it is routine to check that G ≤ GΘ1 if and only if QGQ−1 ≤ GΘ2 where Θ2 := (Q−1)tΘ1Q−1. Consider a category G defined as following: an object is a pair (G, ω) where G is a discete group and ω is a 2-cocycle in Z(G, T). For objects (G1, ω1) and (G2, ω2), HomG((G1, ω1), (G2, ω2)) := {ϕ : G1 → G2 ω2 ◦ ϕ = ω1}. If we denote by C the category of C ∗-algebras and define C ∗(·, ·)(G, ω) := C ∗(G, ω), C ∗(·, ·)ϕ : ℓ1(G2, ω2) → ℓ1(G1, ω1), f 7→ f ◦ ϕ. By the universal property of a full twisted C ∗-algebra, one may check that C ∗(·, ·) is a functor from G to C. Thus for arbitrary Θ1 ∈ Td(R) and Q ∈ GLd(R), we have Θ2 := QtΘ1Q ∈ Td(R). By taking different generators we can obtain an isomorphism Q−1 : Zd → Zd, Certain actions of finitely generated groups on higher dimensional noncommutative tori 15 moreover ωΘ2 ◦ Q−1(x, y) = exp(πihΘ2Q−1x, Q−1yi) = exp(πihΘ1x, yi) = ωΘ1(x, y). Thus we obtain AΘ1 = C ∗(Zd, ωΘ1) ∼= C ∗(Zd, ωΘ2) = AΘ2. Since Td,A(R) = Td,Q−1ΛQ(R) = QtTd,Λ(R)Q, and we have the following preposition in general. Proposition 4.1. For Λ ∈ GLd(Z) where · · · 0 · · · 0 ... . . . · · · At where each Ai is the companion matrix for pei Td,Λ(R), then A1 0 0 A2 ... ... 0 0 Λ =    i of degree di, and for Θ ∈ Θ =  Θ11 −Θt ... −Θt Θ12 12 Θ22 ... 1t −Θt 2t · · · Θ1t · · · Θ2t ... . . . · · · Θtt   where for i ≤ j, Θij is a di × dj matrix , and Θij = (θkl) such that θkl = θ(k+1)(l+1) for all k, l. Proof. Write Θ = (Θij) then by ΛtΘΛ = Θ we obtain At iΘijAj = Θij. Then by a similar proof to the one in [2, Lemma 4.3], we maintain the form of Θij. (cid:3) Proposition 4.2. Θ ∈ Td,Λ(R) has the following form, for some Θ′ ∈ Td−1(R). Θ =(cid:18) Θ′ 0 0 (cid:19) 0 Proof. We write Θdi ΛtΘΛ = Θ, we have := (θdk) where Pi−1 l=1 dl < k ≤ Pi ΘdiAi = Θdi. l=1 dl. Then for Thus it suffices to show when i = 1. Since A1 is the companion matrix for pe1 1 , we have (θd,1, θd,2, . . . , θd,d1−1, θd,d1) = (θd,2, θd,3, . . . , θd,d1 , − d1−1 Xl=0 alθd,l+1) Certain actions of finitely generated groups on higher dimensional noncommutative tori 16 where d1−1 alxl = Φp e1 1 (x) = Xl=0 Hence we have θ := θd,1 = · · · = θd,d1 and xip ei −1 i . pi−1 Xk=0 thus θ = 0 and then Θdi = 0. −(cid:16) d1−1 Xl=0 al(cid:17)θ = θ, (cid:3) Thus for any d-dimensional noncommutative torus AΘ on which Zn = hAi acts, we have Θ = QtΘ′′Q where Then we have AΘ ∼= AΘ′′ ∼= AΘ′ ⊗ C(T), which is clearly not simple. Θ′′ :=(cid:18) Θ′ 0 0 (cid:19) . 0 Hence we have the following theorem. Theorem 4.3. For a given dimension d and order n, if d − W (n) = 1, then there is no simple noncommutative tori on which Zn acts in the above way. Corollary 4.4. For a given dimension d and order n, there is a action of Zn on a simple noncommutative d-torus AΘ if and only if either d − W (n) > 1 or d − W (n) = 0. Corollary 4.5. For a given dimension d > 3 which is odd and an order n, then there is an action of Zn on simple d-dimensional noncommutative tori if and only if there is an action of Zn of the above type on d − 3-dimensional noncommutative tori. Proof. Notice that d − W (n) is an odd integer then by Corollary 4.4 and Corollary 3.2 we draw the conclusion. (cid:3) We mention that Theorem 4.3 is a generalization of [2, Theorem 5.2] by the following example. Example 4.6. Consider an action of Zn on a simple 3-dimensional noncom- mutative torus. Then by Theorem 4.3 and by the fact that W (n) is always even, it is equivalent to say that d − W (n) = 3. Thus we obtain n = 2, in other words, the only action by a nontrivial finite cyclic group on a simple 3-dimensional torus of the above type is the flip action by Z2, as stated in [2, Theorem 5.2]. We rewrite a fact mentioned in [6, Lemma 1.5] as following. In general for a d-dimensional noncommutative torus AΘ where Θ ∈ Td(R), write Θ =(cid:18) Θ′ −(θid)t (θid) 0 (cid:19) and the standard generator of AΘ as ui for i = 1, . . . , d. Then AΘ′ = C ∗{ui i = 1, . . . , d − 1}, Certain actions of finitely generated groups on higher dimensional noncommutative tori 17 moreover we can define an action α : Z → Aut(AΘ′) by α(1)(ui) = exp(2πiθid)ui which is homotopic to the identity and AΘ ∼= AΘ′ ⋊α Z. Return to our current case for the given dimension d and the order n with d − W (n) = 1. We can construct a noncommutative d-torus AΘ where Θ ∈ Td(R) and Θ =(cid:18) Θ′ 0 0 (cid:19) , 0 and AΘ admits an action of Zn. Although AΘ ∼= AΘ′ ⊗ C(T) is never simple, or equivalently Θ is never nondegenerate, by our construction Θ′ is nondegenerate, and the simple noncommutative d − 1-tori AΘ′ admits an action of Zn which is denoted by α′. Hence write ui for i = 1, . . . , d the standard generators of AΘ, we could define an action by the following, α : Z × Zn → Aut(AΘ) α(1, I)(ui) = ui, α(1, I)(ui) = exp(2πiθ)ui, α(0, A)(ui) = α′(ui), α(0, A)(ui) = ui, i = 1, . . . , d − 1; i = d; i = 1, . . . , d − 1; i = d   where θ is an irrational number. It is actually the product action of two commuting actions. Then we have AΘ ⋊α (Z × Zn) ∼= (AΘ ⋊α Z) ⋊ α Zn ∼= (AΘ′ ⊗ Aθ) ⋊α′ Zn. Since Θ′′ = Θ′ ⊕(cid:18) 0 −θ 0 (cid:19) θ is simple, and action α′ of Zn on AΘ′′ ∼= AΘ′ ⊗ Aθ is of the above type, the the crossed product AΘ ⋊α (Z × Zn) is covered by a discussion in Sec- tion 3. We comment that action α restricted to Z is adding dimension to noncommutative torus AΘ and making it a simple one on which there is an action of Zn of above type, in our context this should be the action of Z on a noncommutative torus in which we are interested. 5. Finitely generated abelian group actions Motived by the last part of the former section, we have the following discussion. For a given dimension d and a finitely generated abelian group G, by the structure theorem we have G ∼=(cid:16) s Yi=1 Zp ei i (cid:17) × Zr, Certain actions of finitely generated groups on higher dimensional noncommutative tori 18 if d − W (G) = 0 or d − W (G) > 0, there is an action of Qs on a simple noncommutative d-torus AΘ by Theorem 3.6, say α′. If r > 0 we may define an action α : G → Aut(AΘ) with a nondegenerate Θ0 ∈ Tr(R) such that i=1 Zp ei i s AΘ ⋊α G ∼= (AΘ ⊗ AΘ0) ⋊α′ Zp ei i . Yi=1 Note that there is also a similar action when r = 0 stated in Theorem 3.6. If d−W (n) = 1, there is a noncommutative d-torus AΘ of form AΘ′⊗C(T), or equivalently, with Θ =(cid:18) Θ′ 0 0 (cid:19) 0 where Θ′ ∈ Td−1(R) is nondegenerate. Both AΘ and AΘ′ admit an action of Zn. Denote the latter one by α′. If additionally r > 0, we can obtain an action α : G → Aut(AΘ). In such case if r > 1 we obtain α with an irrational number θ and a nondegenerate Θ0 ∈ Tr−1(R) such that AΘ ⋊α G ∼= (AΘ′ ⊗ Aθ ⊗ AΘ0) ⋊α′ s Yi=1 Zp ei i . Similarly if r = 1 we obtain α with only an irrational number θ such that AΘ ⋊α G ∼= (AΘ′ ⊗ Aθ) ⋊α′ s Yi=1 Zp ei i . Theorem 5.1. For a given dimension d and a finitely generated abelian ei i=1 Zp group G ∼= (cid:16)Qs i (cid:17) × Zr, If d − W (n) = 1 and r > 0, or d − W (n) 6= 1 then there is a noncommutative d-torus AΘ admitting an action of G, denoted by α : G → Aut(AΘ). We require AΘ is pseudo-simple in the first case and simple in the second case. Then the crossed product AΘ ⋊α G is a simple AT algebra, and it is an AF algebra if and only if r = 0, d−W (G) = 0 and G satisfies the last condition in Theorem 3.6. Proof. The existence is by our discussion in this section. Then by Theorem 3.6 we draw the conclusion. (cid:3) Acknowledgment The author would like to express his deep gratitude to his supervisor, Professor Yasuyuki Kawahigashi, for helpful suggestions, warm encourage- ment and support during two years of his master course. The author would like to thank Dr. Chen Jiang, Dr. Yuki Arano, Dr. Yosuke Kubota, Dr. Shuhei Masumoto, Dr. Takuya Takeishi and Dr. Lu Xu for many useful discussions. The author is grateful to the University of Tokyo for Special Scholarship for International Students (Todai Fellowship). Certain actions of finitely generated groups on higher dimensional noncommutative tori 19 References [1] S. Echterhorff, W. Luck, N. C. Phillips, S. Walters, The structure of crossed products of irrational rotation algebras by finite subgroups of SL2(Z), J. Reine Angew. Math. 639 (2010), 173 -- 221. [2] J. A. Jeong, J. H. Lee, Finite groups acting on higher dimensonal noncommutative tori, arXiv:1402.1826v2. [3] J. Kuzmanovich, A. Pavlichenkov, Finite groups of matrices whose entires are interg- ers, Amer. Math. Monthly 109 (2002), 173 -- 186. [4] M. Langer, W. Luck, Topological K-theroy of the group C ∗-algebra of a semi-direct product Zn ⋊ Z/m for a free conjugation action, J. Topol. Anal. 4 (2012), 121 -- 172. [5] J. A. Packer, I. Raeburn, Twisted crossed products of C ∗-algebras, Math. Proc. Cam- bridge Philos. Soc. 106 (1989), 293 -- 311. [6] N. C. Phillips, Every higher noncommutative Torus is an AT-algebra, arXiv:math/0609783. [7] N. C. Phillips, The tracial Rokhlin property for actions of finite groups on C ∗-algebras, Amer. J. Math. 133 (2011), 581 -- 636. [8] C. Schochet, Topological methods for C ∗-algebras II: Geometric resolutions and the Kunneth formula, Pacific J. Math. 98 (1982), 443 -- 458. Graduate School of Mathematical Sciences, the University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo 153-8914, Japan. E-mail address: [email protected]
1711.08786	2	1711	2018-04-11T13:16:38	On a class of determinant preserving maps for finite von Neumann algebras	[ "math.OA", "math.FA" ]	Let $\mathscr{R}$ be a finite von Neumann algebra with a faithful tracial state $\tau $ and let $\Delta$ denote the associated Fuglede-Kadison determinant. In this paper, we characterize all unital bijective maps $\phi$ on the set of invertible positive elements in $\mathscr{R}$ which satisfy $$\Delta(\phi(A)+\phi(B)) = \Delta(A+B).$$ We show that any such map originates from a $\tau$-preserving Jordan $$-automorphism of $\mathscr{R}$ (either $$-automorphism or $*$-anti-automorphism in the more restrictive case of finite factors). In establishing the aforementioned result, we make crucial use of the solutions to the equation $\Delta(A + B) = \Delta(A) + \Delta(B)$ in the set of invertible positive operators in $\mathscr{R}$. To this end, we give a new proof of the inequality $$\Delta(A+B) \ge \Delta(A) + \Delta(B),$$ using a generalized version of the Hadamard determinant inequality and conclude that equality holds for invertible $B$ if and only if $A$ is a nonnegative scalar multiple of $B$.	math.OA	math	ON A CLASS OF DETERMINANT PRESERVING MAPS FOR FINITE VON NEUMANN ALGEBRAS MARCELL GA ´AL AND SOUMYASHANT NAYAK Abstract. Let R be a finite von Neumann algebra with a faithful tracial state τ and let ∆ denote the associated Fuglede-Kadison determinant. In this paper, we characterize all unital bijective maps φ on the set of invertible positive elements in R which satisfy ∆(φ(A) + φ(B)) = ∆(A + B). We show that any such map originates from a τ -preserving Jordan ∗-automorphism of R (either ∗-automorphism or ∗-anti-automorphism in the more restrictive case of finite factors). In establishing the aforementioned result, we make crucial use of the solutions to the equation ∆(A + B) = ∆(A) + ∆(B) in the set of invertible positive operators in R. To this end, we give a new proof of the inequality using a generalized version of the Hadamard determinant inequality and conclude that equality holds for invertible B if and only if A is a nonnegative scalar multiple of B. ∆(A + B) ≥ ∆(A) + ∆(B), 8 1 0 2 r p A 1 1 ] . A O h t a m [ 2 v 6 8 7 8 0 . 1 1 7 1 : v i X r a 1. Introduction In 1897 Frobenius [9] proved that if φ is a linear map on the matrix algebra Mn(C) of n × n complex matrices preserving the determinant, then there are matrices M, N ∈ Mn(C) such that det(MN) = 1 and φ can be written in one of the following forms: a) φ(A) = MAN, A ∈ Mn(C); b) φ(A) = MAtN, A ∈ Mn(C) where (·)t denotes transposition of a matrix. In the past decades this result of Frobenius has inspired many researchers to deal with different sorts of preserver problems involving various notions of determinant [2, 5, 7, 13, 17, 20]. Among others, in [13] Huang et al. completely described all maps on the positive definite cone Pn of Mn(C) which satisfy the sole property det(φ(A) + φ(B)) = det φ(I) · det(A + B) for all A, B ∈ Pn. Note that when φ is a unital linear map on Pn, the above property simply means that φ is det preserving. In this paper, we consider an identical operator algebraic counterpart of this problem in the setting of finite von Neumann algebras. Our approach to the solution is based on a generalization of the Minkowski determinant inequality to the setting of von Neumann algebras. Note that the usual Minkowski determi- nant inequality for matrices A, B ∈ Pn asserts that npdet(A + B) ≥ npdet(A) + npdet(B), with equality if and only if A, B are positive scalar multiples of each other. In [1, Corol- lary 4.3.3 (i)], Arveson gives a variational proof of a version of the Minkowski determinant 1 2 MARCELL GA ´AL AND SOUMYASHANT NAYAK inequality in finite von Neumann algebras involving the Fuglede-Kadison determinant. Re- cently, this result has been subsumed by the study of the anti-norm property of a wide class of functionals by Bourin and Hiai [4, Corollary 7.6]. The equality conditions are harder to isolate from these proofs because of limiting arguments and are not explicitly docu- mented. As that will play an important role in our results, we first need to establish when ∆(A+ B) = ∆(A) + ∆(B) holds for positive operators A, B in a finite von Neumann algebra. To this end, we give a new proof of the inequality using a generalized version of the Hadamard determinant inequality [11, 8] in a bootstrapping argument. Consequently, we are able to solve the aforementioned preserver problem concerning Fuglede-Kadison determinants on finite von Neumann algebras. The paper is organized as follows. In § 2 we fix the notation and briefly review some facts from the theory of determinants on von Neumann algebras. The precise formulations of our corresponding results and their proofs are collected in § 3 and § 4. 2. Preliminaries Throughout this paper, R denotes a finite von Neumann algebra acting on the complex (separable) Hilbert space H and containing the identity operator I. Let τ be a faithful tracial state on R, by which we mean a linear functional τ : R → C such that for all A, B ∈ R, we have (i) τ (AB) = τ (BA), (ii) τ (A∗A) ≥ 0 with equality if and only if A = 0, (iii) τ (I) = 1. The set of invertible operators in R is denoted by GL1(R). We denote the cone of positive operators in R by R+ and use GL1(R)+ to denote the set of invertible operators in R+. For A ∈ GL1(R), the Fuglede-Kadison determinant ∆ associated with τ is defined as ∆(A) = exp(τ (log √A∗A)). The dependence of ∆ on τ is suppressed in the notation and it is to be assumed that a choice of a faithful tracial state has already been made. Although this concept of determinant was developed in [10] in the context of type II1 factors, it naturally extends to finite von Neumann algebras as above. Example 2.1. The simplest examples of finite von Neumann algebras are given by Mn(C), the full matrix algebra of n × n complex matrices. For A ∈ Mn(C), the Fuglede-Kadison determinant ∆(A) is given by npdet(A) where det is the usual matrix determinant. Example 2.2. On M2(C) the unique faithful tracial state is given by tr2 : M2(C) → C, tr2(A) = a11 + a22 2 where aij ∈ C (1 ≤ i, j ≤ 2) denotes the (i, j)th entry of the matrix A in M2(C). Denote by D2(C) the ∗-subalgebra of diagonal matrices in M2(C). The von Neumann algebra M2(R) ∼= R ⊗ M2(C) (acting on H ⊕ H ) is also finite and the faithful tracial state on M2(R) is given by τ2 = τ ⊗ tr2, that is, for an operator A in M2(R), we have τ (A11) + τ (A22) . τ2(A) = 2 DETERMINANT PRESERVING MAPS 3 We denote by ∆2 the Fuglede-Kadison determinant on M2(R) corresponding to τ ⊗ tr2. For operators A1, A2 in R, we define diag(A1, A2) := (cid:20)A1 0 A2(cid:21) ∈ D2(R) ∼= R ⊗ D2(C) ⊂ M2(R). 0 It is straightforward to see that for invertible operators A1, A2 in R, the operator diag(A1, A2) in M2(R) is invertible and ∆2(diag(A1, A2)) = p∆(A1) · ∆(A2). Example 2.3. Let X be a compact (Hausdorff) topological space with a probability Radon measure ν. The space of essentially bounded complex-valued functions on (X, ν), denoted by L∞(X, ν), which acts via left multiplication on L2(X, ν), forms an abelian von Neumann algebra. The involution operation is given by f ∗(x) := f (x). A faithful tracial state on L∞(X, ν) is obtained by τν(f ) = ZG f (x) dν(x), for f ∈ L∞(X, ν) and the corresponding Fuglede-Kadison determinant determinant is given by ∆ν(f ) = exp(cid:18)ZG log(f (x))dν(x)(cid:19) . Group von Neumann algebras provide another important class of examples of finite von Neumann algebras, see e.g. [19, §3.2]. One of the most remarkable properties of ∆ is that it is a group homomorphism of GL1(R) into the multiplicative group of positive real numbers. However, there may be several exten- sions of ∆ from GL1(R) to the whole of R. From the proof of [10, Lemma 6], note that for a projection E 6= I in R, we have ∆′(E) = 0 for any extension ∆′ of ∆. In this paper, we con- sider only the analytic extension which is defined as follows. For A ∈ R, let σ(A) ⊂ [0,∞) denote the spectrum of √A∗A and let µ be the probability measure supported on σ(A) and induced by the tracial state τ . Then we define ∆(A) := exp(cid:18)Zσ(A) log λ dµ(λ)(cid:19) and denote this extension also by ∆. with understanding that ∆(A) = 0 whenever Rσ(A) log λ dµ(λ) = −∞. We abuse notation Below we summarize some properties of ∆ which we shall need in § 3. (p1) ∆(U) = 1 for a unitary U in R; (p2) ∆(AB) = ∆(A) · ∆(B) for A, B ∈ R; (p3) ∆ is norm continuous on GL1(R); (p4) ∆(λA) = λ∆(A) for λ ∈ C, A ∈ R; (p5) limε→0+ ∆(A + εI) = ∆(A) for a positive operator A in R. 3. The Minkowski determinant inequality In this section, we aim to establish the following version of the Minkowski determinant inequality. 4 MARCELL GA ´AL AND SOUMYASHANT NAYAK Theorem 3.1 (generalized Minkowski determinant inequality). For positive operators A, B in R, we have (3.1) ∆(A + B) ≥ ∆(A) + ∆(B). Moreover, if B is invertible, then equality holds in (3.1) if and only if A is a nonnegative scalar multiple of B. We work towards the proof of Theorem 3.1 using several lemmas. Before turning to their proof, let us explain the main ideas implemented in them. A proof of the Minkowski determinant inequality (see [16, p. 115]) for matrices is based on the 'traditional' Hadamard determinant inequality which states that for a positive definite matrix A in Mn(C), the determinant of A is less than or equal to the product of its diagonal entries and equality holds if and only if A is a diagonal matrix. For a given A ∈ Mn(C), considering the positive semidefinite matrix √A∗A, one may derive from this inequality the geometrically intuitive fact that the volume of an n-parallelepiped with prescribed lengths of edges is maximized when the edges are mutually orthogonal. In this paper, we make use of an 'abstract' Hadamard-type determinant inequality in our proof to reflect the geometric origins of inequality (3.1). Recall that if S is a von Neumann subalgebra of R, then by a conditional expectation we mean a unital (identity preserving) positive linear map Φ : R → S which satisfies Φ(SAT ) = SΦ(A)T for all A ∈ R and S, T ∈ S . Concerning τ -preserving conditional expectations, in [18, Theorem 4.1] the second author has proved the following generalization of the Hadamard determinant inequality: Theorem 3.2. For a τ -preserving conditional expectation Φ on R and an invertible positive operator A in R, we have that ∆(Φ(A−1)−1) ≤ ∆(A) ≤ ∆(Φ(A)) and equality holds in either of the above two inequalities (and hence in both inequalities) if and only if Φ(A) = A. We consider the map Φ2 : M2(R) → M2(R) defined by Φ2(A) := diag(A11, A22) ∈ D2(R) ⊂ M2(R) where Aij ∈ R(1 ≤ i, j ≤ 2) is the (i, j)th entry of A. Note that Φ2 is a (τ ⊗ tr2)-preserving normal conditional expectation from M2(R) onto the von Neumann subalgebra D2(R). In Lemma 3.3, we will use Theorem 3.2 in the context of the von Neumann algebra M2(R) and the (τ ⊗ tr2)-preserving normal conditional expectation Φ2. More precisely, for positive operators A1, A2 in R we first prove in Lemma 3.3 that (3.2) ∆(A1) · ∆(A2) ≤ ∆(tA1 + (1 − t)A2) · ∆(tA2 + (1 − t)A1), t ∈ [0, 1]. Choosing t = 1/2, we arrive at ∆(A1) · ∆(A2) ≤ ∆(cid:18)A1 + A2 2 (cid:19)2 DETERMINANT PRESERVING MAPS 5 which readily implies that 2p∆(A1) · ∆(A2) ≤ ∆(A1 +A2), which is weaker than the desired inequality. We then use a "tensor power trick" to proceed with a bootstrapping argument to prove the required inequality. Now we are in a position to prove our first lemma. Lemma 3.3. For positive operators A1, A2 in R and t ∈ [0, 1], the following inequality holds: (3.3) ∆(A1) · ∆(A2) ≤ ∆(tA1 + (1 − t)A2) · ∆(tA2 + (1 − t)A1). Further if A1, A2 are invertible, then equality holds if and only if either t ∈ {0, 1} or A1 = A2. Proof. Consider the unitary operator U in M2(R) given by U := (cid:20) √tI √1 − tI √1 − tI −√tI (cid:21) . Note that U ∗diag(A1, A2)U = (cid:20) tA1 + (1 − t)A2 pt(1 − t)(A1 − A2) tA2 + (1 − t)A1 (cid:21) . pt(1 − t)(A1 − A2) Clearly, Φ2(U ∗diag(A1, A2)U) = diag(tA1 + (1 − t)A2, tA2 + (1 − t)A1). Using Theorem 3.2 and property (p5) concerning ∆, we get that p∆(A1) ·p∆(A2) = ∆2(diag(A1, A2)) = ∆2(U ∗diag(A1, A2)U) ≤ ∆2(Φ2(U ∗diag(A1, A2)U)) = p∆(tA1 + (1 − t)A2) ·p∆(tA2 + (1 − t)A1). If A1, A2 are invertible, then U ∗diag(A1, A2)U is also invertible and equality holds if and (cid:3) only if pt(1 − t)(A1 − A2) = 0, that is, either t ∈ {0, 1} or A1 = A2. Lemma 3.4. Let n be a positive integer. For positive operators A1, A2, . . . , An in R, the following inequality holds: (3.4) ∆(cid:16) A1 + · · · + An n (cid:17) ≥ (∆(A1) · · · ∆(An))1/n. Further if A1, . . . , An are invertible, then equality holds if and only if A1 = A2 = . . . = An. Proof. First for n = 2k with k ∈ N we prove by induction the inequality (3.4) along with the equality condition and then employ a standard argument to establish it for all n ∈ N. Choosing t = 1/2 in Lemma 3.3, we get for A1, A2 ∈ R+ that (cid:17) ≥ (∆(A1) · ∆(A2))1/2. (3.5) ∆(cid:16)A1 + A2 2 Further if A1, A2 are invertible, equality holds if and only if A1 = A2. This proves the case when n = 2. Now assume that inequality (3.4) holds for n = 2k−1 along with the equality condition. For A1, A2, . . . , A2k ∈ R+, we define B1 := A1 + · · · + A2k−1 2k−1 , B2 := A2k−1+1 + · · · + A2k 2k−1 . 6 MARCELL GA ´AL AND SOUMYASHANT NAYAK From (3.5), we infer that (3.6) Furthermore, the induction hypothesis furnishes ∆(cid:16)B1 + B2 2 (cid:17) ≥ (∆(B1)∆(B2))1/2. (3.7) ∆(B1) ≥ (∆(A1) · · · ∆(A2k−1))1/2k−1 , ∆(B2) ≥ (∆(A2k−1+1) · · · ∆(A2k))1/2k−1 . Combining (3.6) and (3.7), we have (3.8) ∆(cid:16)A1 + · · · + A2k 2k (cid:17) ≥ (∆(A1) · · · ∆(A2k ))1/2k . If A1, . . . , A2k are invertible, then so are B1, B2, and equality holds if and only if B1 = B2 A1 = · · · = A2k−1 and A2k−1+1 = · · · = A2k or, in other words, if and only if A1 = · · · = A2k . Thus by induction, for n a power of 2, we have established inequality (3.4) along with the equality condition. Next we consider an arbitrary positive integer m. Let k be a positive integer such that 2k−1 ≤ m < 2k. For positive operators A1, . . . , Am, we define a positive operator B by B := (A1 + · · · + Am)/m. It follows that ∆(B) = ∆(cid:16) A1 + · · · + Am + (2k − m)B 2k (cid:17) ≥ (∆(A1) · · · ∆(Am))1/2k (∆(B))1−m/2k and using property (p5), we conclude that ≥ (∆(A1) · · · ∆(Am))1/2k ∆(B)m/2k =⇒ ∆(B) ≥ (∆(A1) · · · ∆(Am))1/m. If A1, . . . , Am are invertible, then so is B and equality holds if and only if A1 = A2 = . . . = Am = B. (cid:3) Theorem 3.5. For a positive operator A in R and t ≥ 0, the following inequality holds: (3.9) ∆(tI + A) ≥ t + ∆(A) with equality if and only if either t = 0 or A is a nonnegative scalar multiple of I. Proof. Let A be an invertible positive operator such that ∆(A) = 1. For p, q ∈ N, an application of Lemma 3.4 gives us that ∆(cid:18) pI + qA p + q (cid:19) ≥ p+qp∆(I)p∆(A)q = 1. Thus ∆((p/q)I + A) ≥ p/q + 1 with equality if and only if A = I. Approximating with strictly positive rational numbers, we have by property (p5) for ∆ that (3.10) Note that for A ∈ GL1(R)+ the operator B := (1/∆(A))A is an invertible positive operator satisfying ∆(B) = 1. As ∆(tI + B) ≥ t + 1, for t ≥ 0 substituting s = t∆(A), we get the desired inequality ∆(tI + A) ≥ t + 1, for t ≥ 0. (3.11) ∆(sI + A) ≥ s + ∆(A), for s ≥ 0. Next we derive conditions for the case of equality in (3.11). Note that for a particular value of s under consideration, if s/∆(A) is rational, then equality holds in (3.11) if and only DETERMINANT PRESERVING MAPS 7 if B = I ⇔ A = ∆(A)I. For s > 0, if ∆(sI + A) = s + ∆(A), using (3.11) repeatedly along with the multiplicativity of ∆, we get that s2 + ∆(A)(2s + ∆(A)) = (s + ∆(A))2 = ∆(sI + A)2 = ∆(s2I + A(2sI + A)) Thus we conclude that ∆(2sI + A) = 2s + ∆(A). For r ∈]0, s[, using (3.11) we deduce that ≥ s2 + ∆(A)∆(2sI + A) ≥ s2 + ∆(A)(2s + ∆(A)). 2s + ∆(A) = ∆(2sI + A) = ∆((s + r)I + (s − r)I + A) ≥ (s + r) + ∆((s − r)I + A) ≥ (s + r) + (s − r) + ∆(A) = 2s + ∆(A). As (s + r) + ∆((s − r)I + A) = 2s + ∆(A), we have that ∆((s − r)I + A) = s − r + ∆(A) for all r ∈]0, s[. We may choose r such that (s − r)/∆(A) is rational and thus conclude that A is a scalar multiple of the identity. Hence equality holds in (3.11) if and only if either s = 0 or A is a scalar multiple of the identity. Next we consider the case when A is not necessarily invertible. For some t > 0 and any s ∈]0, t], define As := sI + A. As documented in property (p5) for ∆, note that limε→0+ ∆(Aε) = ∆(A). We have ∆(At) = ∆(cid:18) t I + At/2(cid:19) ≥ Taking the limit k → ∞, we conclude that ∆(tI + A) = ∆(At) ≥ 2 t 2 + ∆(At/2) ≥ k Xi=1 t 2i + ∆(At/2k ). ∞ Xi=1 t 2i + ∆(A) = t + ∆(A). If A is a scalar multiple of the identity, equality trivially holds. If ∆(tI + A) = t + ∆(A), we must have ∆(At) = t/2 + ∆(At/2) and thus At/2 is a scalar multiple of I implying that A is a scalar multiple of I. (cid:3) Proof of Theorem 3.1. If B is invertible, then by (3.9) we infer that 2 ) ≥ 1 + ∆(B− 1 2 AB− 1 2 ) ∆(I + B− 1 2 AB− 1 2 AB− 1 with equality if and only if B− 1 2 = λI with some λ ≥ 0. Using the multiplicative property of the determinant ∆, we conclude that ∆(A + B) ≥ ∆(A) + ∆(B) with equality if and only if A = λB with some λ ≥ 0. If B is not invertible, we consider the invertible positive operator Bε := B + εI (ε > 0). Then we have ∆(A + Bε) ≥ ∆(A) + ∆(Bε) and taking the limit ε → 0+, we see that ∆(A + B) ≥ ∆(A) + ∆(B), as required. Remark 3.6. If A or B is invertible, then the condition of equality has a straightforward form demanding a scaling relationship between the operators unless one of them is 0. But if neither A nor B is invertible, then the conditions under which equality holds are less easy to characterize. Assume that E, F are orthogonal projections in R such that E + F < I. We then have ∆(E + F ) = 0 = ∆(E) + ∆(F ) and thus we cannot expect the operators in (cid:3) 8 MARCELL GA ´AL AND SOUMYASHANT NAYAK question to have any specific 'correlation', a term which we do not define but whose spirit is captured in the above example. 4. A class of determinant preserving maps In this section, we present our result concerning certain determinant preserving bijective maps on GL1(R)+. Before we do so let us recall the concept of Jordan ∗-isomorphisms. Definition 4.1. A linear map J : R → R is called a morphism; (i) Jordan homomorphism if J(A2) = J(A)2, for all A ∈ R; (ii) Jordan ∗-homomorphism if J(A∗) = J(A)∗ for all A ∈ R and J is a Jordan homo- (ii) Jordan ∗-isomorphism if J is a bijective Jordan ∗-homomorphism. Whenever we use the terms ∗-homomorphism, ∗-isomorphisms and ∗-automorphism, it is implicitly understood to refer to the C ∗-algebraic structure of R. A celebrated result of Kadison [14, Corollary 5] states that a (linear) order automorphism of a C ∗-algebra is necessarily implemented by a Jordan ∗-automorphism. Roughly speaking, this means that in a C ∗-algebra the order and the Jordan structures are intimately connected and in fact, determine each other. The mentioned result of Kadison was crucially used in [3, Lemma 8] where the structure of additive bijective maps was determined on the cone GL1(R)+. We apply this in the proof of the main result of the section which is as follows. Theorem 4.2. Let φ : GL1(R)+ → GL1(R)+ be a bijection. Then ∆(φ(A) + φ(B)) = ∆(φ(I)) · ∆(A + B) holds for all A, B in GL1(R)+ if and only if there is a τ -preserving Jordan ∗-isomorphism J : R → R and a positive invertible element T ∈ GL1(R)+ such that φ(A) = T J(A)T, A ∈ GL1(R)+. The proof uses the main ideas in [13], however, we must adapt them to the much more general setting of finite von Neumann algebras. Before turning to the proof of Theorem 4.2, we paraphrase an auxiliary lemma from [10] which makes use of the Riesz-Dunford holomorphic functional calculus for Banach algebras [6] to derive a pertinent corollary. Lemma 4.3 ([10, Lemma 2]). Let B be a complex Banach algebra with a norm-continuous tracial linear functional T. Let f (λ) be a holomorphic function on a domain Λ ⊂ C bounded by a curve Γ and let γ : [0, 1] → R be a differentiable family of operators in R, such that the spectrum of each operator γ(t) lies in Λ. Then f (γ(t)) is differentiable with respect to t and T((f ◦ γ)′(t))) = T(f ′(γ(t)) · γ′(t)). Corollary 4.4. For invertible positive operators A, B in R, the function g : [0, 1] → R defined by g(t) = ∆(tA + (1− t)B) is differentiable at 0+ and g′(0+) = ∆(B)(τ (B−1A)− 1). DETERMINANT PRESERVING MAPS 9 Proof. Let γ : [0, 1] → GL1(R)+ be the line segment γ(t) = tA + (1 − t)B. Clearly, γ is a continuously differentiable curve with γ′(t) = A − B for all t ∈ [0, 1]. If ε > 0 is such that εI ≤ A and εI ≤ B, we have that εI ≤ γ(t) for any t ∈ [0, 1]. Thus we may choose a domain Λ ⊆ C not containing 0 that is bounded by a curve Γ which surrounds the spectra of γ(t) for all t ∈ [0, 1] and does not wind around 0. On the domain Λ, f = log is a holomorphic function. Define G(t) := τ (log(γ(t)))). Using Lemma 4.3, we get that G′(0+) = τ (B−1(A − B)) = τ (B−1A − I) = τ (B−1A) − 1. As g(t) = exp G(t), we conclude that g′(0+) = exp(G(0))·G′(0+) = ∆(B)(τ (B−1A)−1). (cid:3) Proof of Theorem 4.2. Let us begin with the necessity part. Observe that the transformation ψ : GL1(R)+ → GL1(R)+ defined by ψ(A) = φ(I)−1/2φ(A)φ(I)−1/2 is unital. From the multiplicativity of ∆, we see that ∆(ψ(A) + ψ(B)) = ∆(A + B), (4.1) Plugging A = B into equation (4.1), we deduce that ∆(ψ(A)) = ∆(A) for every A ∈ GL1(R)+ and thus for a positive real number λ > 0 we obtain for A, B ∈ GL1(R)+. ∆(ψ(A) + ψ(λA)) = ∆(A + λA) = ∆(A) + ∆(λA) = ∆(ψ(A)) + ∆(ψ(λA)). An application of Theorem 3.1 entails that ψ(λA) = µψ(A) for some µ > 0. As noted earlier, ∆(ψ(A)) = ∆(A) which implies λ = µ meaning that ψ is positive homogeneous. For A, B ∈ GL1(R)+ and a number t ∈]0, 1[, we get that ∆(tψ(A) + (1 − t)ψ(B)) = ∆ (ψ(tA) + ψ((1 − t)B)) = ∆(tA + (1 − t)B). For s ∈ {0, 1}, as ∆(ψ(A)) = ∆(A), ∆(ψ(B)) = ∆(B), it follows that ∆(sψ(A) + (1 − s)ψ(B)) = ∆(sA+(1−s)B). In summary, we have for t ∈ [0, 1] that ∆(tψ(A)+(1−t)ψ(B)) = ∆(tA + (1 − t)B). Taking the derivative of both sides with respect to t at 0+, and using Corollary 4.4, we obtain that (4.2) τ (ψ(B)−1ψ(A)) = τ (B−1A), for all A, B ∈ GL1(R)+. The right hand side of (4.2) is additive in the variable A. As B runs through the whole of GL1(R)+, substituting X = ψ(B)−1, it follows from (4.2) that for all A, C, X ∈ GL1(R)+, we must have τ (Xψ(A + C)) = τ (Xψ(A)) + τ (Xψ(C)), or, equivalently, τ (X[ψ(A + C) − (ψ(A) + ψ(C))]) = 0. Since a self-adjoint operator X in R may be written as the difference of two invertible positive operators X + (kXk + ε)I, (kXk + ε)I for ε > 0, we further have that τ (X[ψ(A + C) − (ψ(A) + ψ(C))]) = 0 for all A, C ∈ GL1(R)+ and all self-adjoint operators X in R. Choosing X = ψ(A + C) − (ψ(A) + ψ(C)) and using the faithfulness of the tracial state τ , we conclude that ψ(A + C) − (ψ(A) + ψ(C)) = 0 for all A, C ∈ GL1(R)+. Thus, ψ is an additive bijection. The structure of such maps is described in [3]. According to [3, Lemma 8] there is a Jordan ∗-isomorphism J : R → R such that ψ(A) = J(A) for all A ∈ GL1(R)+. 10 MARCELL GA ´AL AND SOUMYASHANT NAYAK the necessity part. The desired τ -preserving property also follows from (4.2). Setting T := pφ(I) completes We next prove the sufficiency. It is well-known that for a Jordan ∗-homomorphism J on R and a continuous function f defined on the spectrum of a self-adjoint operator A in R, we have J(f (A)) = f (J(A)). As J is assumed to be τ -preserving, we observe that ∆(J(A)) = ∆(A), by taking f = log. With these considerations in mind and by the multiplicativity of ∆ we finally conclude that, if there is a τ -preserving Jordan ∗-homomorphism J : R → R and an operator T in GL1(R)+ such that Φ(A) = T J(A)T for all A ∈ GL1(R)+, then we must have ∆(φ(A + B)) = ∆(T )∆(J(A + B))∆(T ) = ∆(T 2)∆(A + B) = ∆(φ(I)) · ∆(A + B). (cid:3) In the particular case of finite factors (that is, von Neumann algebras with trivial center CI), the structure of Jordan ∗-automorphisms is quite straightforward from [12, Theorem I] as finite factors are simple rings (see [15, Corollary 6.8.4]). This helps us elucidate the solution to the preserver problem considered in this paper in a simple manner. Note that ∗-automorphisms and ∗-anti-automorphisms of a finite factor are trace preserving. Corollary 4.5. Let R be a finite factor and φ : GL1(R)+ → GL1(R)+ be a bijective map. Then we have ∆(φ(A) + φ(B)) = ∆(φ(I)) · ∆(A + B), for all A, B ∈ GL1(R)+ if and only if there is a ∗-automorphism (or ∗-anti-automorphism) θ of R and a positive invertible element T ∈ GL1(R)+ such that φ(A) = T θ(A)T, A ∈ GL1(R)+. 5. Acknowledgement The first author was supported by the National Research, Development and Innovation Office NKFIH Reg. No. K-115383. References [1] W. Arveson, Analyticity in operator algebras, Amer. J. Math. 89 (1967), 578-642. [2] B. Aupetit, Spectrum-preserving linear mappings between Banach algebras or Jordan-Banach algebras, J. London Math. Soc. 62 (2000), 917 -- 924. [3] R. Beneduci and L. Moln´ar, On the standard K-loop structure of positive invertible elements in a C ∗- algebra, J. Math. Anal. Appl. 420 (2014), 551 -- 562. [4] J. Bourin and F. Hiai, Anti-norms on finite von Neumann algebras, Publ. Res. Inst. Math. Sci. 51 (2015), 207 -- 235. [5] G. Dolinar and P. Semrl, Determinant preserving maps on matrix algebras, Linear Algebra Appl. 348 (2002), 189 -- 192. [6] N. Dunford, Spectral theory, Bull. Amer. Math. Soc. vol. 49 (1943), 637 -- 651. [7] M. L. Eaton, On linear transformations which preserve the determinant, Illinois J. Math. 13 (1969) 722 -- 727. [8] E. Fischer, Uber den Hadamardschen Determinentsatz, Arch. Math. u. Phys. (3), 13 (1907), 32 -- 40. DETERMINANT PRESERVING MAPS 11 [9] G. Frobenius, Uber die Darstellung der endlichen Gruppen durch lineare Substitutionen, Si. (1897) 994 -- 1015. [10] B. Fuglede and R. V. Kadison, Determinant theory in finite factors, Annals of Math. 55 (1952), 520 -- 530. [11] J. Hadamard, R´esolution dune question relative aux determinants, Bull. des sciences math. (17)(1893), 240 -- 246. [12] I. N. Herstein, Jordan homomorphisms, Transactions of the American Mathematical Society, Vol. 81, No. 2 (Mar., 1956), 331 -- 341. [13] H. Huang, C. N. Liu, P. Szokol, M. C. Tsai and J. Zhang, Trace and determinant preserving maps of matrices, Linear Algebra Appl. 507 (2016), 373 -- 388. [14] R. V. Kadison, A Generalized Schwarz Inequality and Algebraic Invariants for Operator Algebras, Annals of Mathematics, Second Series, Vol. 56, No. 3 (Nov., 1952), 494 -- 503. [15] R. Kadison, J. Ringrose, Fundamentals of the Theory of Operator Algebras, Vol. I-II, Academic Press, San Diego, 1986 (Reprinted by AMS in 1997). [16] M. Marcus, H. Minc, A Survey of Matrix Theory and Matrix Inequalities, Dover, 1992. [17] G. Nagy, Determinant preserving maps: an infinite dimensional version of a theorem of Frobenius, Linear Multilinear Algebra 65 (2017), 351 -- 360. [18] S. Nayak, The Hadamard determinant inequality - extensions to operators on a Hilbert space, Journal of Functional Analysis, [DOI:10.1016/j.jfa.2017.10.009]. [19] A. Sinclair and R. Smith, Finite von Neumann Algebras and Masas, LMS Lecture Note Series (Book 351), Cambridge University Press (2008). [20] V. Tan and F. Wang, On determinant preserver problems, Linear Algebra Appl. 369 (2003), 311 -- 317. Bolyai Institute, Functional Analysis Research Group, University of Szeged, H-6720 Szeged, Aradi v´ertan´uk tere 1, Hungary E-mail address: [email protected] Smilow Center for Translational Research, University of Pennsylvania, Philadelphia, PA 19104 E-mail address: [email protected] URL: https://nsoum.github.io/
1603.00209	1	1603	2016-03-01T10:05:13	Group C*-algebras without the completely bounded approximation property	[ "math.OA" ]	It is proved that: (1) The Fourier algebra A(G) of a simple Lie group G of real rank at least 2 with finite center does not have a multiplier bounded approximate unit. (2) The reduced C*-algebra of any lattice in a non-compact simple Lie group of real rank at least 2 with finite center does not have the completely bounded approximation property. Hence, the results obtained by J. de Canniere and the author for SO(n,1), n at least 2, and by M. Cowling for SU(n,1) do not generalize to simple Lie groups of real rank at least 2.	math.OA	math	GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY UFFE HAAGERUP ABSTRACT. It is proved that: (1) The Fourier algebra A(G) of a simple Lie group G of real rank at least 2 with finite center does not have a multiplier bounded approximate unit. (2) The reduced C ∗-algebra C ∗ r (Γ) of any lattice Γ in a non-compact simple Lie group of real rank at least 2 with finite center does not have the completely bounded ap- proximation property. Hence, the results obtained by J. de Canniere and the author [dCH85] for SOe(n, 1), n ≥ 2, and by M. Cowling [Cow83] for SU(n, 1) do not generalize to simple Lie groups of real rank at least 2. PREAMBLE BY ALAIN VALETTE In spring 2015, I contacted Uffe Haagerup about his manuscript "Group C∗-algebras without the completely bounded approximation property", written in 1986, and never published. I mentioned that Journal of Lie Theory might be a good place to publish it. Uffe Haagerup liked the idea and said he was willing to update the paper after the summer. By a sad twist of fate, Uffe Haagerup tragically passed away in July 2015. After his untimely death, Maria Ramirez Solano volunteered to type the manuscript, and Søren Knudby accepted to write an introduc- tion and update the bibliography. We heartily thank both of them for their help in making the manuscript available to the community. We also thank Søren Haagerup for giving us permission to publish his father's paper. INTRODUCTION The Fourier algebra A(G) of a locally compact group, introduced by Eymard [Eym64], consists of the matrix coefficients of the regular representation. The Fourier algebra is the predual of the group von Neumann algebra M(G) generated by the regular representation. The multipliers M A(G) of the Fourier algebra consists of those continuous functions ϕ on G such that ϕψ ∈ A(G) for every ψ ∈ A(G). One identifies ϕ with the corresponding operator mϕ on A(G), and the multiplier norm kϕkMA is the operator norm of mϕ. If the transposed operator m∗ ϕ on M(G) is completely bounded, we say that ϕ is a completely bounded multiplier. The space of completely bounded multipliers is denoted M0A(G), and the completely bounded multiplier norm kϕkM0A is the completely bounded operator norm of m∗ ϕ. We refer to the papers [Cow89, dCH85] for details. Date: March 2, 2016. 1 GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 2 Leptin [Lep68] showed that the Fourier algebra A(G) has an approximate unit bounded in norm if and only if G is amenable. In [dCH85], de Cannière and the author showed that the Fourier algebra of the non-amenable group SOe(n, 1), n ≥ 2, admits an approximate unit bounded in the completely bounded multiplier norm. In [Cow83], Cowling obtained similar results for SU(n, 1). In the first half of the paper, we show these results do not generalize to simple Lie groups of real rank at least 2: Theorem 1. The Fourier algebra A(G) of a simple Lie group G of real rank at least 2 with finite center does not have an approximate unit bounded in multiplier norm. The second half of the paper is concerned with applications to operator algebras. It is shown that the Fourier algebra A(Γ) of a lattice Γ in a second countable locally compact group G has an approximate unit bounded in the completely bounded multiplier norm if and only if the Fourier algebra A(G) of G has such an approximate unit (Theorem 2.3). It is also shown that, for a discrete group Γ, the Fourier algebra A(Γ) has an approximate unit bounded in the completely bounded multiplier norm if and only if the reduced group r (Γ) has the completely bounded approximation property, if and only if the C∗-algebra C∗ group von Neumann algebra M(Γ) has the (σ-weak) completely bounded approximation property (Theorem 2.6). As a corollary, C∗ r (Γ) does not have the completely bounded approximation property when Γ is a lattice in a simple Lie group of real rank at least 2 with finite center (Corollary 2.7). The paper ends with an appendix containing a characterization, due to Bozejko and Fendler, of completely bounded multipliers. A preliminary version of this paper was completed in 1986 and has circulated among ex- perts in the field. We now mention some of the developments related to this paper up until its publication. In 1989, Cowling and the author [CH89] showed that the Fourier algebra of a simple Lie group with finite center and real rank 1 admits an approximate unit bounded in the com- pletely bounded multiplier norm. This generalized the results of [Cow83, dCH85]. The condition of finite center was subsequently removed by Hansen [Han90]. In 1996, Dorofaeff [Dor93, Dor96] removed the finite center condition from Theorem 1, thus giving a complete characterization of simple Lie groups whose Fourier algebras ad- mits multiplier bounded approximate units. In 2005, Cowling, Dorofaeff, Seeger, and Wright [CDSW05] extended the previous results to cover many non-simple Lie groups including all real algebraic linear groups. In 2012, Ozawa gave a short proof of Theorem 1 (see [Oza12] and [Knu15]). In 1994, a weaker approximation property (called the Approximation Property or simply the AP) than the one considered in Theorem 1 was introduced by Kraus and the author in [HK94]. In 2011–2013, it was shown by Lafforgue, de la Salle, de Laat and the author [HdL13, LDlS11] that simple Lie groups G of real rank at least 2 with finite center do not even have the AP, thus improving Theorem 1. The finite center condition was subsequently removed in [HdL16]. 1. PROOF OF THEOREM 1 Reducing the problem to SL(3, R) and Sp(2, R). In [Wan69], Wang proved that any simple Lie group G with finite center and real rank at least 2 has Kazhdan's property (T ), GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 3 by using the fact that all these groups contain a closed subgroup G′ with finite center and locally isomorphic to either SL(3, R) or SO(2, 3) (cf. [BT65, Theorem 7.2]). If G′ fails to have a multiplier bounded approximate unit for its Fourier algebra, so does G (cf. [dCH85, Proposition 1.12]). It is elementary to check, that if F is a finite normal subgroup of a locally compact group H, then A(H) has a multiplier bounded approximate unit if and only if A(H/F ) has a multiplier bounded approximate unit. Thus if G′ and G′′ are locally isomorphic simple Lie groups with finite center, then A(G′) has a multiplier bounded approximate unit if and only if A(G′′) has a multiplier bounded approximate unit. Hence, to prove Theorem 1 it suffices to consider the two groups SL(3, R) and SO(2, 3) of real rank 2. But SO(2, 3) is locally isomorphic to Sp(2, R) (cf. [Hel78, p. 519]). So we can as well work with SL(3, R) and Sp(2, R). Case SL(3, R). Consider the closed subgroup G0 of SL(3, R) consisting of the 3 × 3- matrices of the form A ∈ SL(2, R), b1, b2 ∈ R. Note that G0 is isomorphic to the semidirect product SL(2, R)×α R2, where α is the usual action of SL(2, R) on R2. We will show that already G0 fails to have a multiplier bounded approximate unit. Put  A 0 0 0 0 1  , b1 b2 0 0 1  K0 = SO(2)  , p1 + x2/4 z 1 0 1 x z + 1 2 xy y 0 0 1  γ(x, y, z) = −x, − (x, y, z) ∈ R3. , yp1 + x2/4! . and let N ⊆ G0 be the nilpotent group of upper triangular matrices with 1's in the diagonal (N is the Heisenberg group). It is convenient to use the following coordinates for N : Lemma A. Let γ be the diffeomorphism of N given by If ϕ : N → C is the restriction of a K0-biinvariant functioneϕ on G0 to N then ϕ = ϕ ◦ γ. Proof. Define u, v ∈ SO(2) by 1 u = By direct computation one gets: 2 1 x −1 2 (cid:19) p1 + x2/4(cid:18) x (cid:18) v 1 x z + 1 2 xy y 0 0 1 1 0 0 1 (cid:19) (cid:18) u 0 0 where (x′, y′, z′) = γ(x, y, z). This proves the lemma. and v = −u. 0 1 (cid:19) = 1 x′ 1 0 0 0 z′ + 1 2 x′y′ y′ 1  (cid:3) GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 4 Lemma B ("Failure of Fubini's Theorem"). Let ϕ ∈ C∞ integrals J(ϕ) =ZRZR I(ϕ) =ZRZR y2 − z2 dz dy, ϕ(y, z) ϕ(y, z) y2 − z2 dy dz c (R2). Then the two double exist, when the inner integrals are taken in the principal value sense (exclude symmetric intervals around the zeroes of y2 − z2, and let the length of the intervals go to zero). Moreover I(ϕ) − J(ϕ) = π2ϕ(0, 0). Proof (sketch). I, J are the distributions on R2, whose Fourier transforms are given by the L∞-functions bI(t, u) = π2χ{u2>t2}, bJ(t, u) = −π2χ{u2<t2}, where χ denotes the characteristic function, and D → bD is the natural extension of the Fourier transform on R2, which on the L1-functions is given by ei(ty+uz)f (y, z) dy dz. bf (t, u) =ZZR2 Hence bI − bJ = π2, which is the Fourier transform of π2 times the Dirac measure at (0, 0). (cid:3) Lemma C. The map is a well-defined distribution K on R2, independent of the order of integration, and bK is the L∞-function ϕ ∈ C∞ c (R2) ϕ(y, z) 1 + y2 − z2 dy dz, ϕ 7→ZZR2 bK(t, u) =(cid:26) J0(√u2 − t2) u2 > t2 u2 < t2 0 where J0 is the zero-order Bessel function. Proof (sketch). To compute bK, observe that K is SO(1, 1)-invariant. Hence bK can only depend on u2−t2, so it is sufficient to compute bK(t, 0) and bK(0, u). But bK(0, u) = J0(u) follows from the formula J0(u) =Z 1 −1 1 √1 − x2 eiux dx. Lemma D. Let D be the distribution on N given by D(ϕ) =Z ∞ −∞Z ∞ −∞ ϕ(x, y, z) (1 + x2/4)y2 − z2 dz dy dx, −∞Z ∞ D(ϕ) ≤ 2π3kϕkA(N ) ∀ϕ ∈ C∞ c (N ). Then (cid:3) ϕ ∈ C∞ c (N ) GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 5 Proof (sketch). Note that one cannot permute the order of integrations dz, dy (cf. Lemma B). However, it is not hard to check that dy and dx can be permuted. We have to prove that D corresponds to an operator T ∈ M(N ), the von Neumann algebra associated to the left regular representation of N , such that kTk ≤ 2π3 (cf. [Eym64]). The Heisenberg group N is of type I, and the infinite dimensional irreducible representations of N are given by (ρa)a∈R\{0} acting on L2(R) by: (ρa(x, 0, 0)f )(t) = eiaxtf (t) (ρa(0, y, 0)f )(t) = f (t − y) (ρa(0, 0, z)f )(t) = eiazf (t) where f ∈ L2(R). The remaining irreducible representations are all one-dimensional, and they form a null set for the Plancherel measure on bN . For f, g ∈ C∞ hρa(x, y, x)f ), gi (1 + x2/4)y2 − z2 dz dy dx Z ∞ −∞Z ∞ −∞Z ∞ c (R) the integral can be computed by permuting the order of integration dy, dx and applying Lemma C. After some reduction, one finds that the integral is equal to −∞ Since J0(x) ≤ 1 for all x ∈ R, st < 0, st > 0. (1.1) where where where g(s)k(s, t)f (t) dt, ZZR2 k(s, t) =( 2π2 s−t J0(a√−4st) 0 k(s, t) ≤ 2π2K(s, t) K(s, t) =(cid:26) 1 s−t st < 0, st > 0. 0 π 2 (U H − HU ), H(f )(s) = 1 πZ ∞ −∞ f (t) s − t dt Moreover, K is the kernel of a bounded operator on L2(R) of norm ≤ π because it is the kernel of the operator is the Hilbert transform (unitary on L2(R)), and U is the unitary operator given by mul- tiplication with sign(s), s ∈ R. Therefore (1.1) implies that k(s, t) is the kernel of an integral operator on L2(R) with norm ≤ 2π3. Hence D corresponds to an operator in M(N ) =R ⊕ Lemma E. If ϕ ∈ C∞ R\{0} ρa(N )′′ da of norm ≤ 2π3. c (N ) and ϕ ◦ γ = ϕ then (cid:3) where γ is the diffeomorphism of N defined in Lemma A. Z ∞ −∞ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ(x, 0, 0) p1 + x2/4 dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 4πkϕkA(N ), GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 6 Proof. Clearly, D(ϕ ◦ γ) = Z ∞ = Z ∞ −∞Z ∞ −∞Z ∞ −∞Z ∞ −∞Z ∞ −∞ −∞ Hence, using Lemma B , yp1 + x2/4(cid:19) dz dy dx ϕ(cid:18)−x,− z√1+x2/4 (1 + x2/4)y2 − z2 z2 − (1 + x2/4)y2 dy dz dx ϕ(x, y, z) 2D(ϕ) = D(ϕ) + D(ϕ ◦ γ) −∞ = Z ∞ −∞Z ∞ −∞Z ∞ −Z ∞ −∞Z ∞ −∞Z ∞ = π2Z ∞ p1 + x2/4 ϕ(x, 0, 0) −∞ −∞ ϕ(x, y, z) (1 + x2/4)y2 − z2 dz dy dx (1 + x2/4)y2 − z2 dy dz dx ϕ(x, y, z) dx. (cid:3) The lemma follows now from Lemma D. Proof of Theorem 1 for SL(3, R). It is sufficient to show that the subgroup G0 does not have a multiplier bounded approximate unit for A(G0): Assume that there existseϕn ∈ A(G0), n ∈ N, such that supn keϕnkMA(G0) < ∞ and for all ϕ ∈ A(G0). Since C∞ c (G0) is dense in A(G0), we can choose the eϕn-functions in C∞ averaging by K0-elements from left and right, we can also obtain that that eϕn is K0- biinvariant. Note that eϕn → 1 uniformly on compact subsets of G0. Put c (G0), and by lim n→∞kϕeϕn − ϕkA(G0) = 0 Then ϕn = eϕnN . n we get ϕn(x, 0, 0)2 kϕnkA(N ) = kϕnkMA(N ) ≤ keϕnkMA(G0). The first equality holds because N is amenable. Thus supn kϕnkA(N ) < ∞, and by Lemma A, ϕn = ϕn ◦ γ for all n ∈ N. By using Lemma E to ϕ2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 4πkϕnk2 Z ∞ 4πZ ∞ p1 + x2/4 p1 + x2/4 This gives a contradiction. Thus by Fatou's Lemma, n→∞ kϕnk2 lim inf A(N ) ≥ = ∞. A(N ). dx −∞ −∞ (cid:3) 1 GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 7 Case Sp(2, R). The group Sp(2, R) is the set of g ∈ GL(4, R) that leaves invariant the exterior form x1x3 − x3x1 + x2x4 − x4x2, (cf. [Hel78, p. 445]). For our purpose, it is convenient to permute the third and fourth coordinate in R4, so we will instead consider the group G ∼= Sp(2, R) of invertible 4 × 4- matrices that leave invariant the form: x1x4 − x4x1 + x2x3 − x3x2. G is a 10 dimensional connected Lie group with Lie algebra a11 + a44 = a22 + a33 = a12 + a34 = a21 + a43 = 0, a13 = a24, a31 = a42 (cid:27) . Let g0 ⊆ g be the Lie algebra g =(cid:26)(aij)i,j=1,...,4(cid:12)(cid:12)(cid:12)(cid:12) g0 =  0 −c2 0 −b2 b3 0 0 0 c1 b1 b2 0 c3 c1 c2 0  (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b1, b2, b3, c1, c2, c3 ∈ R . Then g0 = g1⊕sg2 (semidirect sum), where g1 ∼= sl(2, R) is the Lie algebra corresponding to c1 = c2 = c3 = 0, and g2 is the Lie algebra corresponding to b1 = b2 = b3 = 0. (Note that g2 is isomorphic to the Lie algebra of the three dimensional Heisenberg group). Hence exp(g0) generates a closed subgroup G0 of G, namely, the semidirect product of and c1 0 1 0 where the action α : G1 → Aut(G2) is given by 1 −c2 1 0 0 0 0 0 G1 =  G2 = exp(g2) = α(a) c1 c2 c3 1 0 0 0 0 0 a 0 0 0 0 0 1    = a K0 = 1 0 0 0 0 0 0 0 SO(2) 0 0 0 0 0 1 , (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) , c3 c1 c2 1 a ∈ SL(2, R) c1, c2, c3 ∈ R (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)   . 1   c1 c2 c3 0 0 Consider next the compact subgroup K0 of G0 given by GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 8 .   (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) z 1 −y 1 0 0 0 0 0 (x, y, z, w) ∈ R4. w z + 1 2 xy y 1 z − 1 2 xy x 1 0 c3 − 1 6 b1c2 2 c1 + 1 2 b1c2 c2 1 Lemma F. Let γ′ be the diffeomorphism of N given by b1, c1, c2, c3 ∈ R The group N is isomorphic to the group Γ4 considered by Dixmier in [Dix58]. In the rest of this section, we will use the following coordinates for N : and the nilpotent subgroup of G0 given by 0 −c2 c1 c3 0 0 b1 c1 0 0 0 c2 0 0 0 0 c1 − 1 2 b1c2 b1 1 0 1 −c2 1 0 0 0 0 0 N = exp =   n(x, y, z, w) = γ′(x, y, z, w) = −x,− b1, c1, c2, c3 ∈ R (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)   , , yp1 + x2/4, w! . If ϕ : N → C is the restriction of a K0-biinvariant functioneϕ on G0 to N , then ϕ = ϕ◦ γ′.  = n(γ′(x, y, z, w)).  D′(ϕ) =Z ∞ p1 + x2/4 Proof. Let u, v ∈ SO(2) be as in Lemma A. Then 1 0 0 0 0 0 0 0 0 0 0 1  n(x, y, z, w) −∞Z ∞ −∞Z ∞ D′(ϕ) ≤ 2π3kϕkA(N ) −∞ (1 + x2/4)y2 − z2 dz dy dx, for all ϕ ∈ C∞ Lemma G. Let D′ be the distribution on N given by 1 0 0 0 0 0 u v 0 0 0 0 0 1 ϕ ∈ C∞ c (N ). ϕ(x, y, z, 0) (cid:3) Then Proof (sketch). N is the semidirect product of the abelian subgroup c (N ). isomorphic to R3, and the one-parameter group 1 0 0 0 B =  C =  0 z w 1 x z 0 0 0 1 1 0 1 −y 1 0 0 0 0 0 0 0 0 0 1 y 1 0  (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) , (x, z, w) ∈ R3 y ∈ R (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)  , GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 9 i.e. N ∼= R3 ×θ R where the action θ is given by Let (s, u, v) be the dual coordinates to (x, z, w). The transpose of θy is θy(x, z, w) = (x, z − yx, w − 2yz + y2x). bθy(s, u, v) = (s − yu + y2v, u − 2yv, v). [Mac52], [Dix58]): The orbits forbθy are (v 6= 0), (v = 0, u 6= 0), (v = 0, u = 0). (1) parabolas (2) straight lines (3) single points The irreducible representation of N can now be obtained by the Mackey machine (cf. Since {(s, u, v) v = 0} is a null set in R3, the first type of orbits gives sufficiently many irreducible representations to disintegrate the regular representation of N . Let ρa,b be the irreducible representation coming from the parabolic orbit with vertex (b, 0, a), a ∈ R\{0}, b ∈ R. Then, ρa,b can be realized on L2(R) as follows: (ρa,b(x, 0, 0, 0))(t) = ei(at2+b)xf (t) (ρa,b(0, y, 0, 0))(t) = f (t − y) (ρa,b(0, 0, z, 0))(t) = ei2atzf (t) (ρa,b(0, 0, 0, w))(t) = eiawf (t), where f ∈ L2(R). A computation similar to the one in Lemma D gives now, where −∞ −∞Z ∞ −∞Z ∞ Z ∞ k(s, t) =( 2π2 hρa,b(x, y, z, 0)f, gi (1 + x2/4)y2 − z2 dz dy dx =ZZR2 s−t J0(p−(as2 + b)(at2 + b)) 0 If ab ≥ 0 then k(s, t) = 0 almost everywhere in R2. If ab < 0 we put c =p−b/a. Then where K(s, t) = 1 operator on L2(R) of norm ≤ π, namely, the operator s−t χ{(s2−c2)(t2−c2)<0}. But K(s, t) is the kernel of a bounded integral k(s, t) ≤ 2π2K(s, t), g(s)k(s, t)f (t) ds dt, (as2 + b)(at2 + b) < 0 (as2 + b)(at2 + b) > 0. π 2 (U2HU1 − U1HU2), where H is the Hilbert transform, and U1 and U2 are the unitary multiplication operators on L2(R) given by the functions sign(t + c) and sign(t − c), respectively. Hence k(s, t) is the kernel of a bounded integral operator of norm ≤ 2π3. Since ρa,b(N )′′ da db, M(N ) =ZZ ⊕ a6=0,b∈R it follows that D′ corresponds to an element in M(N ) ∼= A(N )∗ of norm ≤ 2π3. Proof of Theorem 1 for Sp(2, R). Exactly as in the proof of Lemma E, we get that if ϕ ∈ c (N ) and ϕ ◦ γ′ = ϕ, then C∞ Z ∞ ϕ(x, 0, 0, 0) −∞ (cid:3) p1 + x2/4 dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 4πkϕkA(N ), (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 10 and the proof for Sp(2, R) can be completed by the same arguments as we used for SL(3, R). (cid:3) Remark. It follows from the above proofs that R2 ×s SL(2, R) and HG ×s SL(2, R) (the semidirect product of the Heisenberg group HG by SL(2, R) described above) both fail to have a multiplier bounded approximate unit for their Fourier algebras, although the Fourier algebras of R2, HG and SL(2, R) all have multiplier bounded approximate units. (R2 and HG are amenable, for SL(2, R) see [dCH85]). 2. RESULTS ABOUT LATTICES IN LIE GROUPS A lattice Γ in a locally compact group G is a closed discrete subgroup, for which G/Γ has a bounded G-invariant measure. A locally compact group that admits a lattice is necessarily unimodular (cf. Definition 1.8 and Remark 1.9 in [Rag72]). In the following, Γ denotes a lattice in a second countable locally compact group G. In this case, Γ is countable and the quotient map ρ : G → G/Γ has a a Borel cross section. Let Ω be the range of a Borel cross section. Then G = [γ∈Γ Ωγ (disjoint union). Let µ be the Haar measure on G. Since Γ is countable, µ(Ω) > 0. Moreover, the quotient map ρ is a bijection of Ω onto G/Γ, that carries µΩ onto a G-invariant measure on G/Γ. Thus by the assumption that Γ is a lattice, µ(Ω) < ∞. In the following we will assume that µ is normalized such that µ(Ω) = 1. Let µΓ be the counting measure on Γ. For every bounded function ϕ on Γ, the function is a well-defined bounded continuous function on G because χΩ ∈ L1(G), G is unimodu- lar, and because χΩ ∗ ϕµΓ is the bounded function given by bϕ = χΩ ∗ ϕµΓ ∗eχΩ Lemma 2.1. (χΩ ∗ ϕµΓ)(ωγ) = ϕ(γ), γ ∈ Γ, ω ∈ Ω. kbϕkA(G) ≤ kϕkA(Γ) kbϕkM0A(G) ≤ kϕkM0A(Γ). (1) If ϕ ∈ A(Γ) then bϕ ∈ A(G) and (2) If ϕ ∈ M0A(Γ) thenbϕ ∈ M0A(G) and (1) If ϕ ∈ A(Γ), then there exists f, g ∈ ℓ2(Γ) such that ϕ = f ∗eg, kfk2kgk2 = kϕkA(Γ). kϕkA(Γ), and f1 ∗eg1 = bϕ. This proves (1). Put f1 = χΩ ∗ f µΓ, g1 = χΩ ∗ gµΓ. Then f1, g1 ∈ L2(G), kf1k2kg1k2 = kfk2kgk2 = (2) Every z ∈ G has a unique decomposition Proof. z = ωγ, ω ∈ Ω, γ ∈ Γ. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 11 The γ-part in the decomposition will be denoted by γ(z). For x ∈ G, we let τx : Ω → Ω be the map given by τx(ω) = ω′, where ω′ is the Ω-component of xw in the decomposition xω = ω′γ′, ω′ ∈ Ω, γ′ ∈ Γ. Each τx is a Borel isomorphism of Ω because τx corresponds to left translation by x in G/Γ with the Borel isomorphism Ω → G/Γ given by the quotient map. Since the latter Borel isomorphism carries µΩ to an invariant measure on G/Γ, it follows that µΩ is τx- invariant. We rewrite the functionbϕ = χΩ ∗ ϕµΓ ∗eχΩ in a suitable way. Observe first that (χΩ ∗ ϕµΓ)(ωγ) = ϕ(γ), ω ∈ Ω, γ ∈ Γ, or equivalently, Therefore, (χΩ ∗ ϕµΓ)(x) = ϕ(γ(x)), x ∈ G. bϕ(x) =ZG =ZxΩ =ZΩ ϕ(γ(y))χΩ(x−1y) dµ(y) ϕ(γ(y)) dy ϕ(γ(xω)) dµ(ω). For x, y ∈ G and ω ∈ Ω: Thus, xω = τx(ω)γ(xω) yω = τy(ω)γ(yω). yx−1τx(ω) = yω(xω)−1τx(ω) = τy(ω)γ(yω)γ(xω)−1. Since τx(ω) ∈ Ω and γ(yω)γ(xω)−1 ∈ Γ, it follows that γ(yx−1τx(ω)) = γ(yω)γ(xω)−1. Hence ZΩ ϕ(γ(yω)γ(xω)−1) dµ(ω) =ZΩ However, since dµ(ω) is invariant under τx, the latter integral is equal to ϕ(γ(yx−1τx(ω))) dµ(ω). (2.1) Hence, ∀x, y ∈ G: ZΩ ϕ(γ(yx−1ω)) dµ(ω) = bϕ(yx−1). bϕ(yx−1) =ZΩ ϕ(γ(yω)γ(xω)−1) dµ(ω). We can now apply Bozejko-Fendler's result that M0A(G) coincides with the space of Herz-Shur multipliers on G with same norm [BF84]: Since Γ is discrete, it implies, that there exist a Hilbert space H and bounded maps ξ, η from Γ to H such that and ϕ(γ2γ−1 1 ) = hξ(γ1), η(γ2)i, γ1, γ2 ∈ Γ, kξk∞kηk∞ = kϕkM0A(Γ). GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 12 This follows from Gilbert's characterization of Herz-Schur multipliers [Gil]. Define now sup bξ(x)(ω) = ξ(γ(xω)) bη(x)(ω) = η(γ(xω)). bξ,bη : Γ → L2(Ω, H, dµ) by Thenbξ,bη are bounded Borel functions from G to L2(Ω, H, dµ) and x∈Gkbξ(x)k2 ≤ kξk∞, x∈Gkbη(x)k2 ≤ kηk∞, bϕ(yx−1) =ZΩhξ(γ(xω)), η(γ(yω))i dµ(ω) = hbξ(x),bη(y)i. kbϕkM0A(G) ≤ kbξk∞kbηk∞ ≤ kξk∞kηk∞ = kϕkM0A(Γ), Since bϕ is continuous, it implies that bϕ ∈ M0A(G) and and by (2.1), sup (cf. proof of [Cow83, Proposition 1.1] or [Gil]). This proves (2). Lemma 2.2. Let G be a locally compact group and let k ≥ 1. Then the following condi- tions are equivalent: (cid:3) σ(L∞, L1)-topology. (1) There exists a net (ϕα) in A(G) such that supα kϕαkM0A(G) ≤ k and ϕα → 1 in (2) There exists a net (ϕα) in A(G) such that supα kϕαkM0A(G) ≤ k and ϕα → 1 (3) There exists an approximate unit (ϕα) for A(G) such that supα kϕαkM0A(G) ≤ k. uniformly on compact sets. (By an approximate unit in A(G), we just mean a net (ϕα) such that ∀ψ ∈ A(G). lim α kϕαψ − ψkA(G) = 0 The net (ϕα) will in general be unbounded in A(G)-norm.) Proof. (3) =⇒ (2) =⇒ (1) is clear. (1) =⇒ (2): Assume that (ϕα) satisfies (1) and put ϕ′ α = h ∗ ϕα, where h ∈ Cc(G)+,R h dµ = 1. Then α(x) =ZG Let K ⊆ G be compact. Then the functions h(x · ), x ∈ K form a compact subset of L1(G). Since ϕα → 1 in σ(L∞, L1), and since supα kϕαk∞ ≤ k, the convergence is uniform on compact subsets of L1(G). Hence, h(xy)ϕα(y−1)dy. ϕ′ ϕ′ α(x) = h qϕα, h(x · )i α is contained converges to h1, h(x · )i = (h ∗ 1)(x) = 1 uniformly on K. Moreover, ϕ′ in the σ(L∞, L1)-closed convex hull of left translates of ϕα, and since the unit ball in M0A(G) is σ(L∞, L1)-closed (cf. [dCH85]), we have kϕ′ αkM0A(G) ≤ kϕαkM0A(G) ≤ k. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 13 (2) =⇒ (3): Assume that (ϕα) satisfies (2). Choose h ∈ Cc(G)+,RG h dµ = 1, and put αkM0A(G) ≤ k for all α. Let ψ ∈ A(G)∩ Cc(G), and α = h∗ ϕα. As above, we have kϕ′ ϕ′ put K = supp(h), L = supp(ψ). Moreover, set gα = ϕαχK−1L (where χE is the characteristic function of a set E). Then for x ∈ L, h(y)ϕα(y−1x) dy = (h ∗ gα)(x) (2.2) (h ∗ ϕα)(x) =ZK Similarly, (2.3) By the assumption on ϕα, gα → χK−1L uniformly on K −1L. Moreover, gα vanishes outside the compact set K −1L. Hence, (h ∗ 1)(x) = (h ∗ χK−1L)(x), x ∈ L. ega → eχK−1L, and since h ∈ L2(G), this implies that h ∗ gα → h ∗ χK−1L Hence also in L2(G)-norm, in A(G)-norm. which by (2.2) and (2.3) is equivalent to (h ∗ gα)ψ → (h ∗ χK−1L)ψ in A(G)-norm, ϕ′ αψ = (h ∗ ϕα)ψ → (h ∗ 1)ψ = ψ in A(G)-norm. Since A(G) ∩ Cc(G) is dense in A(G) and since α kϕ′ αkMA(G) ≤ sup α kϕ′ sup αkM0A(G) < ∞, it follows that kϕ′ αψ − ψkA(G) → 0, for all ψ ∈ A(G). (cid:3) Remark. The proof of (2) =⇒ (3) in Lemma 2.2 is due to Michael Cowling (private communication). It substitutes a previous proof of ours, that was valid only for Lie groups. Theorem 2.3. Let Γ be a lattice in a second countable locally compact group G, and let k ∈ [1,∞[. then the following conditions are equivalent. (1) A(G) has an approximate unit (ϕα), such that kϕαkM0A(G) ≤ k, for all α. (2) A(Γ) has an approximate unit (ψα), such that kψαkM0A(Γ) ≤ k for all α. Proof. (1) =⇒ (2) follows from [dCH85], and is valid for any closed subgroup Γ ⊆ G. (2) =⇒ (1): Assume that (ψα) ⊆ A(Γ) satisfies (2), and put bψα = χΩ ∗ ψαµΓ∗eχΩ, where Ω is chosen as in Lemma 2.1. Then bψα ∈ A(G) by Lemma 2.1. Since Γ is discrete, the net (ψα) is uniformly bounded, and since ψα → 1 pointwise in Γ, also It is easy to check that the map ϕ 7→ bϕ = χΩ ∗ ϕµΓ ∗ eχΩ from ℓ∞(Γ) to L∞(G) is the transpose of a bounded map from L1(G) to ℓ1(Γ). Hence, bψα → 1G in σ(L∞, L1)- topology. Moreover, kbψαkM0A(G) ≤ kψαkMA(Γ) by Lemma 2.1. Hence (1) =⇒ (2) follows from (1) ⇐⇒ (3) in Lemma 2.2. σ(ℓ∞(Γ), ℓ1(Γ)). ψα → 1Γ, (cid:3) GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 14 Corollary 2.4. Every lattice Γ in a simple Lie group of real rank at least 2 fails to have a complete multiplier bounded approximate unit. Proof. Follows from Theorem 2.3 and Theorem 1. (cid:3) Let Γ be a discrete group, and let M(Γ) be the von Neumann algebra associated with the left regular representation of Γ on ℓ2(Γ), and let δx(y) =(cid:26) 1, x = y 0, x 6= y be the standard basis in ℓ2(Γ). Put Then tr is a normal faithful trace on M(Γ). tr(a) = haδe, δei, a ∈ M(Γ). Lemma 2.5. Let T be a bounded map from C∗ to itself. Let r (Γ) to itself or a bounded map from M(Γ) ϕT (x) = tr(λ(x)∗T λ(x)), x ∈ Γ. (1) If T is completely bounded, then ϕT ∈ M0A(G), and kϕTkM0A(Γ) ≤ kTkCB. (2) If T is of finite rank, then ϕT ∈ ℓ2(Γ). Proof. (1): Since λ ⊗ λ is unitarily equivalent to λ ⊗ 1ℓ2(Γ) where 1ℓ2(Γ) is the trivial [Dix77, 13.11.3] and [Fel62]), there exists a normal representation of Γ on ℓ2(Γ) (cf. *-isomorphism π of M = M(Γ) onto a von Neumann subalgebra N of M ⊗ M such that π(λ(x)) = λ(x) ⊗ λ(x), x ∈ G. Let ε be the normal conditional expectation of M⊗ M onto N that leaves tr⊗ tr invariant. Then, since ε is orthogonal with respect to the inner product given by tr ⊗ tr, one gets easily that ε(λ(x) ⊗ λ(y)) =(cid:26) λ(x) ⊗ λ(x), x = y, x 6= y. 0, Put ρ = π−1ε. Since one has and norm C∗ r (Γ) = span{λ(g) g ∈ Γ} r (Γ) ⊗ C∗ π(C∗ r (Γ)) ⊆ C∗ r (Γ) into itself, r (Γ)) ⊆ C∗ r (Γ) ⊗ C∗ ρ(C∗ r (Γ), r (Γ). Hence, if T is completely bounded of C∗ S = ρ ◦ (T ⊗ idC ∗ is a well-defined completely bounded map on C∗ r (Γ)) ◦ π r (G) and kSkCB ≤ kTkCB. Now, cx,yλ(y), where cx,y ∈ C and the sum is convergent in the k k2-norm associated with tr. Hence (T ⊗ idC ∗ cx,yλ(y) ⊗ λ(x). T λ(x) =Xy∈G r (Γ))π(λ(x)) =Xy∈G GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 15 By the orthogonality property of ε we have Hence But ε(T ⊗ idC ∗ r (Γ))π(λ(x)) = cx,xλ(x) ⊗ λ(x). Sλ(x) = cx,xλ(x), x ∈ G. cx,x = hT λ(x), λ(x)itr = ϕT (x). ϕT , and hence ϕT ∈ M0A(Γ) This shows that (with the notation of [dCH85]) S = m∗ and kϕTkM0A(G) = kSkCB ≤ kTkCB, because π, π−1 and ε are completely positive. If instead T is a completely bounded map on M(Γ), then we let, S = ρ ◦ (T ⊗ idM(Γ)) ◦ π, and the same proof applies (cf. [dCH85, Section 1]). (2): It is sufficient to consider the rank one case: A rank one map on C∗ r (Γ) is of the form T (a) = f (a)b Thus where f ∈ C∗ r (Γ)∗, b ∈ C∗ r (Γ). ϕT (x) = tr(λ(x)∗b)f (λ(x)). But x 7→ tr(λ(x)∗b) is in ℓ2(Γ), because (λ(x))x∈Γ is an orthonormal family in L2(C∗ and x 7→ f (λ(x)) is bounded. This proves (2) in the C∗ by the same arguments. Theorem 2.6. Let Γ be a discrete group and let k ≥ 1. Then the following three conditions are equivalent. r (Γ)-case. The M(Γ)-case follows (cid:3) r (Γ), tr), (1) A(Γ) has an approximate unit (ϕα) , such that kϕαkM0A(Γ) ≤ k for all α. (2) There exists a net (Tα) of finite rank operators on C∗ r (Γ) such that kTαkCB ≤ k for all α, and such that kTαx − xk → 0 for all x ∈ C∗ r (Γ). (3) There exists a net (Tα) of σ-weakly continuous finite rank maps on M(Γ), such that kTαkCB ≤ k for all α and hϕ, Tαx − xi → 0, for all x ∈ M(Γ), ϕ ∈ M(Γ)∗. Proof. (1) =⇒ (2) and (1) =⇒ (3) follows from [dCH85]. (2) =⇒ (1): Assume Tα are finite rank operators of C∗ k and r (Γ) into itself, such that supα kTαkCB ≤ Let ϕα = ϕTα be as defined in the preceding lemma. Then ϕα ∈ ℓ2(Γ) ⊆ A(Γ) and kTαa − ak → 0 for all a ∈ C∗ r (Γ). sup α kϕαkM0A(Γ) < ∞, and since ϕα(x) = tr(λ(x)∗Tα(λ(x))) → 1 for all x ∈ Γ: kϕαψ − ψkA(Γ) → 0, (2.4) for all ψ ∈ A(Γ) with finite support, i.e. for all ψ ∈ A(Γ) ∩ Cc(Γ) which form a dense subset of A(Γ). Using that kϕαkMA(Γ) ≤ kϕαkM0A(Γ) ≤ k for all α, one gets that (2.4) holds for all ψ ∈ A(Γ). GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 16 (3) =⇒ (1): Let (Tα) be a net satisfying the conditions in (3). Since the functional, is in M(Γ)∗ (x fixed) we have that a 7→ tr(λ(x)∗a), a ∈ M(Γ) tr(λ(x)∗Tα(λ(x))) → 1, ∀x ∈ G. The proof can now be completed exactly as in (2) =⇒ (1) by use of Lemma 2.5. Corollary 2.7. Let Γ be a lattice in a simple Lie group of real rank at least 2 with finite center. Then (cid:3) r (Γ) does not have the completely bounded approximation property. (1) C∗ (2) M(Γ) does not have the (σ-weak) completely bounded approximation property. 3. APPENDIX: ON COMPLETELY BOUNDED MULTIPLIERS AND HERZ-SHUR-MULTIPLIERS. In Section 2, we used Bozejko and Fendler's results [BF84] that M0A(G) coincides iso- metrically with the space B2(G) introduced by Herz in [Her74]. Their result relies heavily on a characterization of B2(G) found by Gilbert ([Gil, Theorem 4.7]). However, Gilbert's paper, has never been published, so for the convenience of the reader, we give below a self-contained proof of the result needed in Section 2. Let a∗ b denote the Schur product (a∗ b)ij = aij bij of complex n× n-matrices in Mn(C), the operator norm of a when Mn(C) acts and let kak be the C∗-norm on Mn(C), i.e. on the Euclidean space ℓ2({1, . . . , n}). We let k kS denote the Schur multiplier norm on Mn(C), i.e. kakS = sup{ka ∗ bk b ∈ Mn(C), kbk ≤ 1}. Let ≤ be the usual ordering in Mn(C) as a C∗-algebra, i.e. a ≥ 0 iff a = a∗ and all eigenvalues of a are nonnegative. It is well known that Hence, if a ≥ 0 then the operator Ma : b 7→ a ∗ b is positive on the C∗-algebra Mn(C), and therefore, a ≥ 0, b ≥ 0 =⇒ a ∗ b ≥ 0. kakS = kMa(1)k =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  a1 0 ... 0 an  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = max{a1, . . . , an} for every a ∈ Mn(C)+ (see for instance [RD66, Corollary 1] or [KR97, 10.5.10]). The following lemma is a special case of [Gil, Theorem 4.7]. The proof given here is inspired by [Pau84]. Lemma 3.1 ([Gil]). Let a ∈ Mn(C). The following three conditions are equivalent: (1) kakS ≤ 1. (2) ∃b, c ∈ Mn(C)+ such that(cid:18) b a∗ c (cid:19) ≥ 0 and bii ≤ 1 cii ≤ 1, i = 1, . . . , n. (3) There exist a Hilbert space H and 2n vectors ξ1, . . . , ξn, η1, . . . , ηn in the closed a unit ball of H such that aij = hξi, ηji, i, j = 1, . . . , n. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 17 Proof. (1) =⇒ (2): Consider the real subspace E of M2n(C)s.a. consisting of the matrices of the form x (cid:18) y x∗ z (cid:19) (yi + zi) + 2Re nXi,j=1 where y, z are self-adjoint diagonal matrices y = diag(y1, . . . , yn), z = diag(z1, . . . , zn). Let ϕ : E → C be the linear form given by z (cid:19) = ϕ(cid:18) y x∗ We will prove that ϕ ≥ 0, i.e. ϕ(w) is non-negative on positive hermitian matrices in E. Note that 12n ∈ E. Since nXi=1 xij aij . (3.1) x ϕ(w) ≥ 0 ⇐⇒ ϕ(w + ε12n) ≥ 0, ∀ε > 0, it is sufficient to check that ϕ(w) ≥ 0 for all strictly positive w =(cid:18) y x∗ z (cid:19) x i.e. those w ∈ E+ for which yi > 0, zi > 0, i = 1, . . . , n. However, w ≥ 0 implies that (cid:18) 1n y−1/2x∗z−1/2 z−1/2xy−1/2 1n (cid:19) ≥ 0 which is equivalent to kz−1/2xy−1/2k ≤ 1 (C∗-norm). Since kakS ≤ 1, the matrix, e = (aij z−1/2 i xij y−1/2 j )i,j=1,...,n has also C∗-norm ≤ 1. Hence which implies that eij y1/2 j nXi,j=1 =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) xij aij(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) nXi,j=1 ϕ(cid:18) y x∗ z (cid:19) ≥ nXj=1 x yj + nXi=1 zi!1/2 z1/2 i (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ nXj=1 zi − 2 nXj=1 yj!1/2 nXi=1 zi!1/2 yj!1/2 nXi=1 ≥ 0. Hence ϕ is a positive functional on E ⊆ M2n(C). Let eϕ be a Hahn-Banach extension of ϕ to M2n(C)s.a.. Then eϕ ≥ 0, because keϕk = kϕk = ϕ(12n), and because Hence, there exists a positive hermitian matrix 0 ≤ w ≤ 12n =⇒ eϕ(12n − w) ≤ keϕk =⇒ eϕ(w) ≥ 0. such that (dij )i,j=1,...,2n 2nXi,j=1 eϕ(w) = dij wij , w ∈ M2n(C)s.a.. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 18 x then by (3.1) x ∈ Mn(C). Therefore dii = 1, i = 1, . . . , 2n. Moreover, if w is of the form For diagonal matrices w = diag(y1, . . . , yn, z1, . . . , zn), eϕ(w) =Pn w =(cid:18) 0 x∗ 0 (cid:19) , di,n+j(x∗)ij = 2Re nXi,j=1 nXi,j=1 for all (xij ) ∈ Mn(C). Hence aij = dn+i,j, i.e. d =(cid:18) b a∗ c (cid:19) aij xij = 2Re nXi,j=1 dn+i,jxij + a where bii = cii = 1, i = 1, . . . , n. This proves (2). (2) =⇒ (3): Let i=1(yi + zi) by (3.1). nXi,j=1 dn+i,j xij a d =(cid:18) b a∗ c (cid:19) dij =Xk fikfjk d =(cid:18) b a∗ c (cid:19) . a be as in (2). Let H = ℓ2({1, . . . , 2n}), and let η1, . . . , ηn, ξ1, . . . , ξn be the row vectors of the operator f = d1/2. Since we get aij = hξi, ηji, and kξik2 = dn+i,n+i ≤ 1, kηik2 = dii ≤ 1, for i = 1 . . . , n. (3) =⇒ (2): Let ξi, ηj be as in (3), put bij = hηi, ηji, cij = hξi, ξji i, j = 1 . . . , n, and let If we let ηn+i = ξi, i = 1, . . . , n. Then dij = hηi, ηji, i, j = 1, . . . , 2n. 2 ≥ 0. 2nXi,j=1 This implies that d is positive hermitian, because, for λ1, . . . , λ2n ∈ C, (2) =⇒ (1): If dij λiλj =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) λiηi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2nXi=1 d =(cid:18) b a∗ c (cid:19) ≥ 0 and bii ≤ 1, cii ≤ 1, then the map e 7→ d ∗ e, e ∈ M2n(C) is positive and kd ∗ ek ≤ max1≤i≤2n diikek. Taking e of the form e = (cid:18) 0 x 0 (cid:19), x ∈ Mn(C), we see that ka ∗ xk ≤ kxk, x ∈ Mn(C). This proves (2) =⇒ (1). Theorem 3.2 ([BF84]). Let ϕ be a continuous function on a locally compact group G, and let k ≥ 0. Then the following conditions are equivalent. (cid:3) a 0 (1) ϕ ∈ M0A(G) and kϕkM0A(G) ≤ k. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 19 (2) For every finite set x1, . . . , xn in G, kϕ(x−1 j xi)i,j=1,...,nkS ≤ k (3) There exist a Hilbert space H and two bounded maps ξ, η : G → H such that ϕ(y−1x) = hξ(x), η(y)i, ∀x, y ∈ G and (sup x∈G kξ(x)k)(sup y∈Gkη(y)k) ≤ k. (4) There exist a Hilbert space H and two bounded maps ξ, η : G → H as in (3) with the additional property that the coordinate functions ξi and ηi (with respect to any orthonormal basis (ei)i∈I of H) are continuous and the families {ξi}i∈I and {ηi}i∈I are locally countable. i,j=1 be the matrix units It is elementary to check that f ∗ spani,j{fij} is a subalgebra of M(G)⊗Mn(C) ∗-isomorphic to Mn(C). Since ∗-isomorphisms preserve norms, i=1 fii = 1, so N = fij = λ(x−1 Proof. (1) =⇒ (2) (cf. [BF84]): Let x1, . . . , xn ∈ G and let (eij)n in Mn(C). Put i xj ) ⊗ eij ∈ M(G) ⊗ Mn(C). ij = fji, fijfkl = δjkfil andPn =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αij fij(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi,j=1 αij ∈ C. Let Mϕ : M(G) → M(G) be as in [dCH85, Section 1]. Then by (1), αij eij(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi,j=1 , where in is the identity on Mn(C). But, kMϕ ⊗ ink ≤ k, Therefore and hence for aij ∈ C, Mϕλ(x) = ϕ(x)λ(x), x ∈ G. (Mϕ ⊗ in)fij = ϕ(x−1 i xj )fij, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi,j=1 ϕ(x−1 i xj)αij fij(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi,j=1 αijfij(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . Since fij in this inequality can be exchanged by the standard matrix units eij of Mn(C), it follows that the matrix (ϕ(x−1 i xj ))i,j=1,...,n as well as its transpose has Schur multiplier norm ≤ k. (2) =⇒ (3): Let (F ,⊆) be the family of all finite subsets of G ordered by inclusion. Assuming (2), we can for each F ∈ F find a Hilbert space HF and bounded maps with and ξF , ηF : F → HF . x∈F kξF (x)k ≤ k1/2, sup x∈F kηF (x)k ≤ k1/2, sup ϕ(y−1x) = hξ(x), η(y)i, x, y ∈ F. (3) follows now easily by a standard ultraproduct argument: Let namely U be a cofinal ultrafilter on (F ,⊆) (i.e. an ultrafilter that contains all sets of the form {F ′ ∈ F F ′ ⊇ GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 20 F}, where F ∈ F), and let HU be the ultraproduct of (HF )F ∈F corresponding to U. For x ∈ G, we let ξ(x), and η(x) be the elements in HU with representing sequences ξ(x) = (ξF (x))F ∈F η(x) = (ηF (x))F ∈F where we set ξF (x) = ηF (x) = 0 if x 6∈ F . Since for fixed x, y ∈ G, x, y ∈ F eventually, we have Moreover, hξ(x), η(y)i = lim U hξF (x), ηF (y)i = ϕ(y−1x), x, y ∈ G. x∈Gkξ(x)k ≤ k1/2, sup x∈Gkη(x)k ≤ k1/2, sup (3) =⇒ (4): Let ξ, η : G → H satisfy the conditions in (3). Let ξ′(x) = P ξ(x), where P is the orthogonal projection on the closed linear span of {η(x) x ∈ G}, and put η′(y) = Qη(y), where Q is the orthogonal projection on the closed linear span of {ξ′(x) x ∈ G}. Then ϕ(y−1x) = hξ′(x), η(y)i = hξ′(x), η′(y)i, x, y ∈ H, and both {ξ′(x)}x∈G and {η′(x)}x∈G are total sets in the Hilbert space H ′ = Q(H). There is therefore no loss of generality to assume, that {ξ(x)}x∈G and {η(x)}x∈G are total in H. Since ϕ is continuous, the map (x, y) → hξ(x), η(y)i is separately continuous, and by the uniform boundedness, and totality of {ξ(x)}x∈G and {η(x)}x∈G, it follows that ξ, η : G → H are continuous from G to H with σ(H, H ∗)- topology. Let {ξi(x)}i∈I and {ηi(x)}i∈I be the coordinates of ξ(x) and η(x) with respect to a fixed basis (ei)i∈I in H. Then ξi and ηi are continuous complex valued functions. Moreover, for any open relative compact set U in G, Xi∈IZU ξi(x)2 dx ≤ZU kξ(x)k2 dx < ∞. By the continuity of ξi and the fact, that no non-empty open set has Haar measure zero, it follows that all except countably many of the ξi's vanish on the set U , i.e. {ξi}i∈I is locally countable, and similarly, {ηi}i∈I is locally countable. (4) =⇒ (1): For f ∈ L∞(G), we let m(f ) denote the multiplication operator g 7→ f g on L2(G). For x ∈ G where γx(f )(y) = f (x−1y). Put now, λ(x)m(f )λ(x)−1 = m(γx(f )), ai = m(qξi), bi = m(qηi), a∗ 1/2 b∗ i bi are convergent and where qg(x) = g(x−1). ThenPi∈I a∗ i ai andPi∈I b∗ i bi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I i ai(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Φ(s) =Xi∈I (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I ≤ sup x∈Gkη(x)k ≤ k. Thus, we can define a bounded σ-weakly continuous map Φ on B(L2(G)) by x∈Gkξ(x)k · sup b∗ i sai. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 21 Now Φ(λ(x)) = Xi∈I = Xi∈I = m Xi∈I m(qηi)∗λ(x)m(qξi) m(qηi)∗m(γx(qξi))λ(x) qηiγx(qξi)! λ(x). Here we have used that {ξi} and {ηi} are locally countable, so pointwise convergence of the sum implies σ(L∞, L1)-convergence. But for y ∈ G, qηi(y)(γx(qξi)(y)) =Xi∈I Xi∈I ηi(y−1)ξi(y−1x) = hξ(y−1x), η(y−1)i = ϕ(y(y−1x)) = ϕ(x). Hence Φ(λ(x)) = ϕ(x)λ(x). This implies that Fores ∈ B(L2(G)) ⊗ Mn(C), Hence, Φ(M(G)) ⊆ M(G). (Φ ⊗ in)(es) =Xi∈I (bi ⊗ 1n)∗es(ai ⊗ 1n). 1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I i bi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kΦ ⊗ ink ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I i ai(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) a∗ b∗ 1/2 ϕ ∈ M0A(G) and kϕkM0A(G) ≤ k. ≤ k. Thus kΦkCB ≤ k, so by [dCH85, Proposition 1.2] Corollary 3.3 ([BF84] and [Her74]). Let Gd be the group G with discrete topology. Then and the M0A-norms on the two spaces coincide. M0A(G) = M0A(Gd) ∩ C(G), (cid:3) Proof. Immediate from (1) ⇐⇒ (2). Remark 3.4. It is not hard to see that the functions ξ, η : G → H in (3) of Theorem 3.2 can actually be chosen to be continuous: From the proof of (3) =⇒ (4) it follows that ξ, η can be chosen such that they are contin- uous from G to H with respect to the σ(H, H ∗)-topology on H, and such that for every open relative compact set U ⊆ G, ξ(U ) and η(U ) are contained in a separable subspace of H. This implies that ξ, η ∈ L∞(G, H). Let now f ∈ Cc(G) with kfk2 = 1, and define (cid:3) by bξ(x),bη(x) ∈ L2(G, H) bξ(x)(z) = f (z)ξ(zx) bη(x)(z) = f (z)η(zx) for all x ∈ G z ∈ G, z ∈ G. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 22 Then it is clear that Moreover, x∈Gkη(x)k. sup x∈Gkbξ(x)k2 ≤ sup sup x∈Gkξ(x)k, x∈Gkbη(x)k2 ≤ sup hbξ(x),bη(y)i = Z f (z)2hξ(zx), η(zy)idz = Z f (z)2ϕ(y−1x)dz = ϕ(y−1x). Finally, using the fact that the group of right translations (Rx)x∈G defined by (Rxh)(z) = ∆1/2 G (x)h(zx), h ∈ L2(G, H) acts norm-continuously on L2(G, H), it is not not hard to check that x 7→ bξ(x) and y 7→ bη(y) are norm continuous from G to L2(G, H). REFERENCES [BF84] [BT65] Marek Bozejko and Gero Fendler. Herz-Schur multipliers and completely bounded multipliers of the Fourier algebra of a locally compact group. Boll. Un. Mat. Ital. A (6), 3(2):297–302, 1984. Armand Borel and Jacques Tits. Groupes réductifs. Inst. Hautes Études Sci. Publ. Math., (27):55– 150, 1965. [CDSW05] Michael Cowling, Brian Dorofaeff, Andreas Seeger, and James Wright. A family of singular oscilla- [CH89] tory integral operators and failure of weak amenability. Duke Math. J., 127(3):429–486, 2005. Michael Cowling and Uffe Haagerup. Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one. Invent. Math., 96(3):507–549, 1989. [Cow83] Michael Cowling. Harmonic analysis on some nilpotent Lie groups (with application to the repre- sentation theory of some semisimple Lie groups). In Topics in modern harmonic analysis, Vol. I, II (Turin/Milan, 1982), pages 81–123. Ist. Naz. Alta Mat. Francesco Severi, Rome, 1983. [Cow89] M. Cowling. Rigidity for lattices in semisimple Lie groups: von Neumann algebras and ergodic [dCH85] [Dix58] [Dix77] [Dor93] [Dor96] [Eym64] [Fel62] actions. Rend. Sem. Mat. Univ. Politec. Torino, 47(1):1–37 (1991), 1989. Jean de Cannière and Uffe Haagerup. Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups. Amer. J. Math., 107(2):455–500, 1985. Jacques Dixmier. Sur les représentations unitaires des groupes de Lie nilpotents. III. Canad. J. Math., 10:321–348, 1958. Jacques Dixmier. C ∗-algebras. North-Holland Publishing Co., Amsterdam, 1977. Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. Brian Dorofaeff. The Fourier algebra of SL(2, R) ⋊ Rn, n ≥ 2, has no multiplier bounded approx- imate unit. Math. Ann., 297(4):707–724, 1993. Brian Dorofaeff. Weak amenability and semidirect products in simple Lie groups. Math. Ann., 306(4):737–742, 1996. Pierre Eymard. L'algèbre de Fourier d'un groupe localement compact. Bull. Soc. Math. France, 92:181–236, 1964. J. M. G. Fell. Weak containment and induced representations of groups. Canad. J. Math., 14:237– 268, 1962. J.E. Gilbert. Convolution operators of Banach space tensor products III. Unpublished. [Gil] [Han90] Mogens Lemvig Hansen. Weak amenability of the universal covering group of SU(1, n). Math. [HdL13] [HdL16] [Hel78] Ann., 288(3):445–472, 1990. Uffe Haagerup and Tim de Laat. Simple Lie groups without the Approximation Property. Duke Math. J., 162(5):925–964, 2013. Uffe Haagerup and Tim de Laat. Simple Lie groups without the Approximation Property II. Trans. Amer. Math. Soc., 368(6):3777–3809, 2016. Sigurdur Helgason. Differential geometry, Lie groups, and symmetric spaces, volume 80 of Pure and Applied Mathematics. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York, 1978. GROUP C∗-ALGEBRAS WITHOUT THE COMPLETELY BOUNDED APPROXIMATION PROPERTY 23 [Her74] [HK94] [Knu15] [KR97] Carl Herz. Une généralisation de la notion de transformée de Fourier-Stieltjes. Ann. Inst. Fourier (Grenoble), 24(3):xiii, 145–157, 1974. Uffe Haagerup and Jon Kraus. Approximation properties for group C ∗-algebras and group von Neu- mann algebras. Trans. Amer. Math. Soc., 344(2):667–699, 1994. Søren Knudby. Weak amenability and simply connected Lie groups. Preprint, to appear in Kyoto J. Math., 2015. Richard V. Kadison and John R. Ringrose. Fundamentals of the theory of operator algebras. Vol. II, volume 16 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1997. Advanced theory, Corrected reprint of the 1986 original. [LDlS11] Vincent Lafforgue and Mikael De la Salle. Noncommutative Lp-spaces without the completely [Lep68] [Mac52] [Oza12] [Pau84] [Rag72] [RD66] [Wan69] bounded approximation property. Duke Math. J., 160(1):71–116, 2011. Horst Leptin. Sur l'algèbre de Fourier d'un groupe localement compact. C. R. Acad. Sci. Paris Sér. A-B, 266:A1180–A1182, 1968. George W. Mackey. Induced representations of locally compact groups. I. Ann. of Math. (2), 55:101– 139, 1952. Narutaka Ozawa. Examples of groups which are not weakly amenable. Kyoto J. Math., 52(2):333– 344, 2012. Vern I. Paulsen. Every completely polynomially bounded operator is similar to a contraction. J. Funct. Anal., 55(1):1–17, 1984. M. S. Raghunathan. Discrete subgroups of Lie groups. Springer-Verlag, New York-Heidelberg, 1972. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 68. B. Russo and H. A. Dye. A note on unitary operators in C ∗-algebras. Duke Math. J., 33:413–416, 1966. S. P. Wang. The dual space of semi-simple Lie groups. Amer. J. Math., 91:921–937, 1969. DEPARTMENT OF MATHEMATICS AND COMPUTER SCIENCE, UNIVERSITY OF SOUTHERN DENMARK CAMPUSVEJ 55, 5230 ODENSE M, DENMARK
1302.5781	1	1302	2013-02-23T10:02:08	Type III actions on boundaries of $\tilde A_n$ buildings	[ "math.OA" ]	Let $\Gamma$ be a group of type rotating automorphisms of a building $\fX$ of type $\tilde A_n$ and order $q$. Suppose that $\G$ acts freely and transitively on the vertex set of $\fX$. Then the action of $\Gamma$ on the boundary of $\fX$ is ergodic, of type $\tq$ or type $\tqs$ depending on whether $n$ is odd or even.	math.OA	math	TYPE III ACTIONS ON BOUNDARIES OF An BUILDINGS PAUL CUTTING AND GUYAN ROBERTSON Abstract. Let Γ be a group of type rotating automorphisms of a building X of type An and order q. Suppose that Γ acts freely and transitively on the vertex set of X. Then the action of Γ on the boundary of X is ergodic, of type III1/q or type III1/q2 depending on whether n is odd or even. Introduction Let M be a compact Riemannian manifold of negative sectional cur- vature, and let Γ = π1(M). Then Γ acts on the sphere at infinity S of the universal cover M of M. The main result of [S] is that the action of Γ on S is ergodic, amenable and type III1. This applies in particular to a cocompact Fuchsian group in G = PGL(2, R) acting on the circle. A discrete analogue of this result was proved in [RR1]. Namely, let Γ be a free group acting simply transitively on the vertices of a locally finite homogeneous tree T of degree q + 1. Then T is the universal covering space of a graph with fundamental group Γ. It was shown in [RR1] that the action of Γ on the boundary of the tree is ergodic, amenable and of type III1/q. Turning to higher rank spaces of nonpositive curvature, it is known that if Γ is a lattice in G = PGL(n + 1, R) with n ≥ 1 and if Ω = G/B where B is the Borel subgroup of upper triangular matrices in G, then the action of Γ on Ω is ergodic of type III1. Here Ω is the maximal boundary of Furstenberg [Mar, VI.1.7]. A similar result holds more generally for a lattice Γ in any semisimple noncompact Lie group G [Zi, 4.3.15]. The discrete analogue of this construction is obtained by replacing R by a nonarchimedean local field F with residue field of order q. The affine Bruhat-Tits building X of G = PGL(n+1, F) is a building of type Date: January 22, 2001. 1991 Mathematics Subject Classification. Primary 46L80; secondary 58B34, 51E24, 20G25. This research was supported by The University of Newcastle. 1 Typeset by AMS-LATEX. 2 PAUL CUTTING AND GUYAN ROBERTSON An [St]. The vertex set of X may be identified with the homogeneous space G/K, where K is a maximal compact subgroup of G, and G acts on the boundary Ω = G/B, where B is a Borel subgroup of G. The precise higher rank analogue of the setup in [RR1] is as follows. Let Γ be a group of type rotating automorphisms of a building X of type An, and suppose that Γ acts simply transitively on the vertices In view of the fact that A1 buildings are trees, such groups of X. Γ should be regarded as higher rank analogues of free groups. Note however that not every A2 building X is the Bruhat-Tits building of PGL (3, K) where K is a local field [CMSZ, II §8]. Geometrically, an An building X is an n-dimensional contractible simplicial complex in which each codimension one simplex lies on q + 1 maximal simplices (chambers). If n ≥ 2 then the number q is necessarily a prime power and is referred to as the order of the building. The boundary Ω of X is a totally disconnected compact Hausdorff space and is endowed with a natural family of mutually absolutely continuous Borel probability measures. In [RR2] it was proved that, if n = 2 and q ≥ 3, then the action of Γ on Ω is ergodic and of type III1/q2. The purpose of the present article is to remove both these hypotheses and prove the following general result. Theorem 1. Let n ≥ 2 and let X be a locally finite thick An building of order q. Let Γ be a group of type rotating automorphisms of X which acts simply and transitively on the vertices of X. Then the action of Γ on the boundary Ω of X is amenable, ergodic and of type IIIλ, where λ =(1/q 1/q2 if n is odd, if n is even. The proof of this result will be completed in Section 3. In Section 4 we deal with freeness of the action, which is required in order to prove that the associated von Neumann algebra is a factor. In particular, Section 4 removes a gap in the proof of freeness in [RR2]. We therefore obtain the following consequence. Corollary 1. Let Γ and Ω be as above. Then the crossed product von Neumann algebra L∞(Ω) ⋊ Γ is the AFD factor of type IIIλ, where λ = 1/q if n is odd, and λ = 1/q2 if n is even. A simple variation on the arguments leading to Theorem 1 proves the following result: see subsection 3.2. Theorem 2. Let p ≥ 2 be a prime number, let n ≥ 1, and let Ω be the boundary of the affine building of PGL(n + 1, Qp). That is Ω = PGL(n + 1, Qp)/B, where B is the Borel subgroup of upper triangular matrices. Then the action of PGL(n + 1, Q) on Ω is ergodic and of type IIIλ, where 3 λ =(1/p 1/p2 if n is odd, if n is even. Similar results can be stated for linear groups over other local fields, but this is perhaps the most striking case. Note that, in contrast to Theorem 1, PGL(n + 1, Q) is not a lattice in PGL(n + 1, Qp), and its action on the boundary is not amenable. Given an An building X, there is a type map τ defined on the vertices of X such that τ (v) ∈ Z/(n+1)Z for each vertex v ∈ X. Every chamber of X has precisely one vertex of each type. An automorphism α of X is said to be type-rotating if there exists i ∈ {0, 1, . . . , n} such that τ (αv) = τ (v) + i for all vertices v ∈ X. An A1 building is a tree, with two types of vertices, and every automorphism of the tree is type rotating. We shall refer to a group Γ satisfying the hypotheses of Theorem 1 as an An group. In [Ca1] it was shown that there is a 1-1 correspondence between An groups and "triangle presentations". The right Cayley graph of an An group Γ relative to a natural set of generators is the 1-skeleton of the An building X. We shall frequently refer to [Ca3], which lays much of the groundwork for dealing with the higher rank An buildings. Throughout the paper X will denote a thick, locally finite An build- ing, and the vertices of the building will be denoted by X0. If X is associated with the An group Γ then the underlying set of the group Γ will be identified with X0, and the action of Γ on the building will be by left multiplication. The identity of Γ will be denoted by 1 throughout. For x, y ∈ X0, d(x, y) will denote the graph distance between those vertices in the 1-skeleton of X, and x = d(x, 1). Further information on buildings can be found in [Ca2], [St], [Br] and [R]. The first two of these references are introductory, while the last two provide a fuller account of the theory of buildings. 1. Preliminaries This section mainly recalls material from [Ca3], to which we refer for a more complete discussion. An An building is a union of apartments. An apartment is isomorphic to a Coxeter complex of type An. Let Σ denote the Coxeter complex of type An. The n-simplices of Σ are referred to as chambers and can be regarded as forming a tessellation of Rn. The vertex set of Σ can be identified with Zn+1/Z(1, 1, . . . , 1). Two 4 PAUL CUTTING AND GUYAN ROBERTSON vertices [a], [b] ∈ Σ, [a] = a + Z(1, 1, . . . , 1) and [b] = b + Z(1, 1, . . . , 1), are adjacent if there exist representative vectors (a1, a2, . . . , an+1) ∈ [a] and (b1, b2, . . . , bn+1) ∈ [b] such that ai ≤ bi ≤ ai + 1 for all 1 ≤ i ≤ n. The type τ (x) ∈ Z/(n + 1)Z of a vertex [x] = [(x1, x2, . . . , xn+1)] ∈ Σ is given by τ (x) = Xi xi! mod (n + 1) . Each chamber of Σ has precisely one vertex of each type. Let bi = (0, . . . , 0, 1, . . . , 1), where precisely i entries equal 1. Note that each x ∈ Zn+1 can be written as Hence there is a mapping Zn+1/Z(1, 1, . . . , 1) → Zn defined by x = x1(1, 1, . . . , 1) +X(xi+1 − xi)bi. [x] 7→ (x2 − x1, x3 − x2, . . . , xn+1 − xn). This mapping is a canonical group homomorphism between Zn+1/Z(1, 1, . . . , 1) and Zn, and by means of it the vertices of Σ can be coordinatized by Zn. Throughout this paper, if k ∈ Zn, then ki denotes the ith entry of k. 1.1. The Boundary of an An Building. Given an An building X, one can define the boundary of X by means of equivalence classes of sectors. The central concern of this paper is the boundary regarded as a measure space. For a discussion of the geometric structure of the boundary, the reader is referred to [R, Chapter 9, 10]. Let S0 be the simplicial cone in the An Coxeter complex Σ with vertex set coordinatized by Zn +. A subcomplex S of X is called a sector if there is an apartment A containing S and a type-rotating isomorphism φ : A → Σ such that φ(S) 7→ S0. (Recall that the isometry φ is said to be type rotating if there exists j ∈ Z/(n + 1)Z such that, for each vertex v of S, τ (φ(v)) = τ (v) + j (mod n + 1). Note that if a, b are vertices in a sector S of X, and φ : S → S0 is a type preserving isomorphism such that φ(a) = (0, 0, . . . , 0) and φ(b) = (k1, k2, . . . , kn), where ki ∈ Z+, then the ki do not depend on the particular apartment A containing S [Ca3, Lemma 2.3]. Thus, for x ∈ X0 and k ∈ Zn +, one can define a set Sk(x) consisting of those elements y ∈ X0 such that there exists a sector S containing x and y, and a type rotating isomorphism φ : S → S0 such that φ(x) = (0, 0, . . . , 0) and φ(y) = k. Given a sector S and a type rotating isomorphism with φ(S) = S0, the basepoint of S is v = φ−1(0, 0, . . . , 0). If x, y ∈ X and y ∈ Sk(x), with k = (k1, k2, . . . , kn) ∈ Zn +, then the graph distance between x and y is given by d(x, y) =Pi ki. 5 Two sectors S1, S2 are said to be equivalent if S1 ∩ S2 contains a subsector. Let Ω be the set of all such equivalence classes of sectors. Then Ω is called the boundary of X. Given ω ∈ Ω and x ∈ X0, there exists an unique sector with basepoint x which is contained in the equivalence class ω [R, Lemma 9.7]. Denote this sector by [x, ω). Also, for m ∈ Zn m(ω) be the unique element in the intersection Sm(x)∩ [x, ω). What Follows Is Based On [Ca3]. Lemma 1.1. Let ω ∈ Ω, and let x, y ∈ X0. Then there exists m(x, y; ω) ∈ Zn such that k′(ω) where k′ = k + m(x, y; ω) , k(ω) = sy sx +, let sx for all k ∈ Z+ such that ki + mi(x, y; ω) ≥ 0 for 1 ≤ i ≤ n. Proof. (c.f. [CMS, Lemma 2.1].) Since [x, ω) is in the same equiva- lence class as [y, ω), [x, ω) ∩ [y, ω) contains a subsector. Choose u = sx k(ω) = sy k′(ω) ∈ [x, ω) ∩ [y, ω) . +}, and T ′ = {sy k+l(ω); l ∈ Zn Let T = {sx +}. Then T , T ′ are sectors in the equivalence class ω with a common base point, and so by [R, Lemma 9.7], T = T ′. k′+l(ω) are both in Su l (ω) ∩ T , and hence are equal. Thus m(x, y; ω) = k′ − k, and m(x, y; ω) is clearly independent of the choice of u ∈ [x, ω) ∩ [y, ω). (cid:3) It follows that sx k′+l(ω); l ∈ Zn k+l(ω), sy Lemma 1.2. Let x be a vertex of X, and let C be a chamber containing x. Then for ω0 ∈ Ω, there exists an apartment A which contains C and the sector S = [x, ω0). Proof. By [R, Lemma 9.4], given the chamber C and sector [x, ω0), there exists an apartment A containing a subsector S ′ ⊂ [x, ω0) and the chamber C. Note that as x ∈ C, one has x ∈ A. Choose a sector S ′′ in A with base vertex x and parallel to S ′. Then S ′′ is equivalent to [x, ω0) and so S ′′ = [x, ω0), by uniqueness of the sector with base vertex x representing the boundary point ω0. (cid:3) The next lemma is a generalisation of [CMS, Corollary 2.3]. Proposition 1.3. For x, y ∈ X0, and ω ∈ Ω, one has if ki ≥ d(x, y) for 1 ≤ i ≤ n. sx k(ω) ∈ [x, ω) ∩ [y, ω) Proof. Set r = d(x, y), and let k = (r, r, . . . , r). An easy consequence of Lemma 1.1 is that z ∈ [x, ω) implies [z, ω) ⊂ [x, ω), and so it is sufficient to show that sx k ∈ [x, ω) ∩ [y, ω). To proceed inductively, the case d(x, y) = 1 is established first. By Lemma 1.2, there exists an apartment A containing both y and S = 6 PAUL CUTTING AND GUYAN ROBERTSON [x, ω). As S is a sector, there exists a type rotating isomorphism ϕ : A → Σ such that ϕ(S) = S0 with ϕ(x) = 0. Since x, y are adjacent in A, ϕ(y) = (y1, y2, . . . , yn), where yi ∈ {−1, 1, 0}, 0 ≤ i ≤ n. Next, define the type rotating isomorphism φ : A → Σ by φ(z) = ϕ(z) − ϕ(y) . This map takes y to the origin in Σ, and (φ)−1(S0) is a sector. Moreover, for z ∈ sx k(ω), φ(z) = ϕ(z) − ϕ(y) = ((k1 − y1), (k2 − y2), . . . , (kn − yn)). Thus φ(z) ∈ S0 if and only if ki ≥ yi for all 1 ≤ i ≤ n. It follows that φ−1(S0) = [y, ω). Moreover, as (1, 1, . . . , 1) ≥ (y1, . . . , yn), one has that sx d(x, y) = 1. (1,1,...,1)(ω) ∈ [x, ω) ∩ [y, ω). This proves the case for In general, given s ∈ Z+, s > 1, suppose that the statement of the lemma is true for all y′ such that d(x, y′) ≤ s−1, and let y ∈ Sk(x) with d(x, y) = s. Without loss of generality, suppose that k1 ≥ 1 and set k′ = (k1 − 1, k2, . . . , kn). Let z be the unique element in conv(x, y) ∩ Sk′(x) and note that d(z, y) = 1 and d(z, x) = d(y, x) − 1 = s − 1. Hence by the inductive hypothesis a = sx (s−1,s−1,...,s−1)(ω) ∈ [x, ω) ∩ [z, ω) . Then for some t = (t1, t2, . . . , tn) ∈ Zn t (ω) and mi(x, z; ω) = (ti − (s − 1)). As ti ∈ Z+, it follows that ti + 1 ≥ d(y, z) = 1. By the inductive hypothesis, this implies that +, one has a = sz sz (t1+1,...,tn+1)(ω) ∈ [z, ω) ∩ [y, ω) . Writing t′ = t − m(x, z; ω), sz (t1+1,t2+1,...,tn+1)(ω) = sx (t′ 1+1,t′ 2+1,...,t′ n+1) = sx (s,...,s)(ω) ∈ [x, ω) ∩ [y, ω) . The result follows. (cid:3) Definition 1.4. Given y ∈ X0, the topology on Ω based at y is given by the basis of open sets {Ωx y}x∈X0, where Ωx y = {ω ∈ Ω; x ∈ [y, ω)} . The topology so defined is independent of the choice of y. See below +, the boundary Ω can be for details. Note that for y ∈ X0, and k ∈ Zn expressed as the disjoint union Ω = [x∈Sk(y) Ωx y . There is a natural class of Borel measures on Ω. Namely, for a fixed y ∈ X0 and a basic open set Ωx y with x ∈ Sk(y), let 7 νy(Ωx y) = 1 Sk(y) . Sk = Sk(y) is independent of y and its actual value was determined in [Ca3, Corollary 2.7]. Specifically, let q be the order of the An building. Also, for k = (k1, k2, . . . , kn) ∈ Zn +, index the non-zero entries of k by {i : ki ≥ 1} = {j1, . . . , jt}, and set j0 = 0 and jt+1 = n + 1. Then (1.1) Sk = q− Pt v=1 jv(jv+1−jv)(cid:20) n + 1 j1 − j0, . . . , jt+1 − jt(cid:21)q · qPn i=1 i(n+1−i)ki , where [· · · ]q = [n + 1]q/([j1 − j0]q · · · [jt+1 − jt]q), and [k]q = (qk − 1) · · · (q − 1). Unlike the topology on Ω, the value of the measure νy is dependent on the choice of y ∈ X0. However, as shown by the following lemmas, the set of measures {νy}y∈X0 is absolutely continuous. The next lemma is generalized from [CMS, Lemma 2.4]. Lemma 1.5. Let y ∈ Sk(x). Suppose that z ∈ Sl(x) ∩ Sl′(y), where li ≥ d(x, y) for all i. Then Ωz y. Moreover, if m(x, y; ω) = i −li for all ω ∈ Ωz (m1, . . . , mn), as in Lemma 1.1, then mi(x, y; ω) = l′ x x ⊂ Ωz Proof. Let ω ∈ Ωz x. Then z = sx [x, ω) ∩ [y, ω) by Lemma 1.3 and choice of l ∈ Zn Moreover, by the proof of Lemma 1.1, mi(x, y; ω) = l′ l (ω), and so z is an element of +. Thus ω ∈ Ωz y. (cid:3) i − li. Lemma 1.6. The topology on Ω does not depend on the vertex y ∈ X0 chosen in Definition 1.4. For any x, y ∈ X0, the measures νx, νy are mutually absolutely continuous, and the Radon Nikodym derivative of νy with respect to νx is given by (1.2) dνy dνx (ω) = q− Pn i=1 i(n+1−i)mi , for ω ∈ Ω, where mi = mi(x, y; ω). Proof. Let x, y ∈ X0. In view of the preceding results, the proof that topology is independent of the base vertex y proceeds exactly as in the case n = 2 [CMS, Lemma 2.5]. Now fix ω ∈ Ω. Choose k ∈ Zn mi(x, y; ω) ≥ d(x, y). Set z = sx x = Ωz Lemma 1.5 implies that Ωz i=1 i(n+1−i)miSk)−1 = q− Pn x) = (Sk′)−1 = (qPn νy(Ωz + such that ki ≥ d(x, y) and ki + k(ω) = sy k′(ω), where k′ = k + m(x, y; ω). y. Moreover it follows from (1.1) that i=1 i(n+1−i)miνx(Ωz x) . 8 PAUL CUTTING AND GUYAN ROBERTSON Since {Ωz hoods of ω, the result follows. x; z ∈ [x, ω), d(x, z) ≥ d(x, y)} is a basic family of neighbour- (cid:3) Remark 1.7. Equation (1.2) is precisely [Ca3, Equation (1.6)], and its proof is outlined in [Ca3, Section 4]. 1.2. An Groups. Let Π be a finite projective geometry of dimension n and order q. If n > 2 then Π is the Desarguesian projective geometry Π(V ), where V is a vector space of dimension n + 1 over a finite field of order q. Let dim(u) denote the dimension of the subspace u of V . In the Desarguesian case the points and lines of Π are the one- and two- dimensional subspaces of V respectively. We shall extend this notation to the non Desarguesian case, so that an element u of a projective plane Π satisfies dim u = 1 if it is a point and dim u = 2 if it is a line. Let λ be an involution of Π such that dim(λ(u)) = n + 1 − dim(u) mod (n + 1). An An triangle presentation T compatible with λ is defined as follows. Let T be a set of triples {(u, v, w) : u, v, w ∈ Π} which satisfy the following properties. (1) Given u, v ∈ Π, then (u, v, w) ∈ T for some w ∈ Π if and only if λ(u) and v are distinct and incident. (2) If (u, v, w) ∈ T , then (v, w, u) ∈ T . (3) If (u, v, w1) ∈ T and (u, v, w2) ∈ T , then w1 = w2. (4) If (u, v, w) ∈ T , then (λ(w), λ(v), λ(u)) ∈ T . (5) If (u, v, w) ∈ T , then dim(u)+dim(v)+dim(w) ≡ 0 mod n+1. The group associated with this triangle presentation is given by ΓT =(cid:28){av}v∈Π(x)(cid:12)(cid:12)(cid:12)(cid:12) (1)aλ(v) = a−1 v (2)auavaw = 1 for all v ∈ Π for all (u, v, w) ∈ T(cid:29) . The Cayley graph of ΓT , with respect to the generators {au}u∈Π is the 1-skeleton of an An building X and ΓT acts on the vertices of the building in a type rotating manner. Conversely any group Γ acting on an An building in this way arises as Γ = ΓT for some triangle presentation T [Ca1, pp 45 -- 46]. Unless otherwise specified, a generator au of Γ will be identified with the corresponding element u ∈ Π. Remark 1.8. The type rotating hypothesis in the definition of an An group has been removed and the appropriate notion of triangle presentation studied in the Ph.D. thesis of T. Svenson [Sv], thereby generalising the results of [Ca1]. For the rest of this article, the An group Γ will be assumed to act on X by left translation with Γ being identified with the vertex set X0. The identity element 1 of Γ is a preferred vertex of X of type 0, and 9 we write Sk = Sk(1) for k = (k1, k2, . . . , kn) ∈ Zn naturally on the boundary Ω. +. The group Γ acts If u1, u2 are elements of Π we denote by u1 ∨ u2 their join; that is their least upper bound in the lattice of subspaces of Π. If Π = Π(V ) is Desarguesian then u1 ∨ u2 = Π means simply that u1 + u2 = V . On the other hand, if Π is a non Desarguesian plane and u1 is a point and u2 is a line of Π, then u1 ∨ u2 = Π means that u1 and u2 are not incident. By [Ca1, Lemma 2.2], every word in Γ can be expressed uniquely in normal form x = u1u2 . . . ul , where dim(ui) ≤ dim(ui+1) and u−1 where kj = {ui : dim(ui) = j}. i ∨ ui+1 = Π. Moreover, x ∈ Sk, Recall from [Ca1, Proof of Theorem 2.5] that if x ∈ X0 then the projective geometry of neighbours of x is {xu : u ∈ Π} and τ (xu) = τ (x) + dim u mod (n + 1). Moreover, xu and xu′ are adjacent vertices if and only if u and u′ are incident in Π (that is, u ⊂ u′ or u′ ⊂ u). In particular a chamber of X containing the vertex x has the form {x, xu1, xu2, . . . , xun} where dim ui = i and u1 ⊂ u2 ⊂ · · · ⊂ un is a complete flag in Π. For more information on An groups, the reader is referred to [Ca1]. 2. An Ergodic measure preserving subgroup of the full group. The action of an An group Γ on the boundary Ω of the corresponding An building, is measure-theoretically ergodic with respect to each of the measures νy, y ∈ X. For the classification of the action it will be necessary to show that the full group [Γ] (defined below) contains a countable measure preserving subgroup K0 ⊂ [Γ] which acts ergodically on Ω. The following two lemmas are straightforward generalisations of [RR2, Lemma 4.6] and [RR2, Lemma 4.7] respectively. Lemma 2.1. Let K be a group which acts on Ω. If K acts transitively on the collection of sets {Ωx +, then K acts ergodically on Ω. 1 : x ∈ Sk} for every k ∈ Zn Proof. Observe first that K preserves ν1 since ν1(Ωx 1) is independent of x. Suppose that X0 ⊆ Ω is a Borel set which is invariant under K and such that ν1(X0) > 0. It will be shown that ν1(Ω \ X0) = 0, thus establishing the ergodicity of the action. 10 PAUL CUTTING AND GUYAN ROBERTSON Define a new measure µ by µ(X) = ν1(X ∩ X0) for each Borel set X ⊆ Ω. Now, for each g ∈ K, µ(gX) = ν1(gX ∩ X0) = ν1(X ∩ g−1X0) ≤ ν1(X ∩ X0) + ν1(X ∩ (g−1X0 \ X0)) = ν1(X ∩ X0) = µ(X). Similarly, µ(gX) ≤ µ(g−1gX) = µ(X). Therefore µ is K-invariant. 1 = Ωy For each x, y ∈ Sk there exists a g ∈ K such that gΩx 1 by 1). Since Ω is the union of Sk disjoint 1) = µ(Ωy transitivity. Thus µ(Ωx sets Ωx 1, y ∈ Sk, each of which has equal measure, one has that µ(Ωx 1) = c Sk , for each x ∈ Sk, where c = µ(X0) = ν1(X0) > 0. Thus µ(Ωx x ∈ X. 1) = cν1(Ωx 1) for every vertex Since the sets Ωx 1 generate the Borel σ-algebra, it follows that µ(X) = cν1(X) for each Borel set X. Therefore ν1(Ω \ X0) = c−1µ(Ω \ X0) = c−1ν1((Ω \ X0) ∩ X0) = 0, thus proving ergodicity. (cid:3) Lemma 2.2. Assume that K ≤ Aut(Ω) acts transitively on the collec- tion of sets {Ωx +. Then there is a countable subgroup K0 of K which also acts transitively on the collection of sets {Ωx 1 : x ∈ Sm} for every m ∈ Zn 1 : x ∈ Sm} for every Sm, m ∈ Zn +. 1 = Ωx Proof. For each pair x, y ∈ Sm, there exists an element k ∈ K such that kΩy 1. Choose one such element k ∈ K and label it kx,y. Since Sm is finite, there are a finite number of elements kx,y ∈ K for each Sm. There are countably many sets Sm, so the set {kx,y : x, y ∈ Sm, m ∈ Zn +} is countable. Hence the group is countable and satisfies the required condition. K0 =(cid:10)kx,y; x, y ∈ Sm, m ∈ Zn +(cid:11) ≤ K (cid:3) Definition 2.3. Given a group Γ acting on a measure space Ω, define the full group, [Γ], of Γ by [Γ] = {T ∈ Aut(Ω); T ω ∈ Γω for almost every ω ∈ Ω} . The set [Γ]0 of measure preserving maps in [Γ] is then given by [Γ]0 = {T ∈ [Γ]; νy ◦ T = νy, y ∈ X0} . It will be shown that there is a countable group K0 of measure- preserving automorphisms of Ω such that 11 (1) K0 acts ergodically on Ω. (2) K0 ≤ [Γ]. By the Lemmas above and the definition of [Γ], it is enough to find + and x, y ∈ Sk, an automorphism g ∈ Aut(Ω) such that 1) = Ωy Identify a simplex in X with its vertex set, and recall from section 1 and gω ∈ Γω for almost all ω ∈ Ω. for each k ∈ Zn g(Ωx 1.2 that a chamber of X containing the vertex x is of the form {x, xu1, xu2, . . . , xun} . where dim ui = i and u1 ⊂ u2 ⊂ · · · ⊂ un is a complete flag in Π. Lemma 2.4. Let C = {1, p1, p2, . . . , pn} be a chamber in X with base vertex the identity element 1 of Γ, where pi are generators of Γ and dim pi = i. There are q chambers C ′ = {x, p1 . . . , pn} in X meeting C in the face {p1, . . . , pn}. The vertex x opposite 1 in C ∪ C ′ has the normal form x = p1un, where dim un = n and p−1 1 ∨ un = Π. Thus x ∈ S(1,0,...,0,1). Equivalently, x = pnu′ 1, where dim u′ 1 = 1 and pn ∨ (u′ 1)−1 = Π. pj x C ′ ...................................................................................................................................................................................... ...................................................................................................................................................................................... C 1 p1 Figure 1. Proof. Consider the projective geometry of the neighbours of p1. For 2 ≤ i ≤ n there exists ui−1 ∈ Πi−1 such that pi = p1ui−1 and ui−1 ⊂ uj−1 for i ≤ j . Now choose un ∈ Πn such that un−1 ⊂ un and un 6= p−1 1 . There exist q such choices for un. One then has that for all 2 ≤ i ≤ n, ui−1 ⊂ un, and hence pi = p1ui−1 is adjacent to p1un. Thus C ′ = {p1, p2, . . . , pn, p1un} is a chamber of X and p1un is the vertex x opposite 1 in C ∪C ′. Clearly p−1 1 ∨ un = Π, so x = p1un is the normal form expressing x as a word of minimal length. 12 PAUL CUTTING AND GUYAN ROBERTSON It now follows that x ∈ S(1,0,...,0,1), and a similar argument proves the (cid:3) final statement. Lemma 2.5. Let x ∈ Sk and y ∈ Sk′, k, k′ ∈ Zn (k1, . . . , kn) and k′ = (k′ ϕ of Ω such that +, where k = n). Then there exists an automorphism 1, . . . , k′ (1) ϕ ∈ [Γ], the full group of Γ; (2) ϕ is almost everywhere a bijection from Ωx (3) ϕ is the identity on Ω\(Ωx 1 ∪ Ωy 1). 1 onto Ωy 1; Moreover, if k = k′ then ϕ is measure preserving. Proof. Let δ = e1 + en = (1, 0, . . . , 0, 1) and consider the set of all vertices x1 ∈ Sk+δ such that x ∈ conv{1, x1}. For such a vertex x1, one has that x1 ∈ Sδ(x) and Ωx1 1 is a (disjoint) union of sets of the form Ωx1 x , where x1 ∈ Sδ(x) ∩ Sk+δ is constructed using the procedure of Lemma 2.4. x . Thus Ωx 1 = Ωx1 1 = Ωx1 Similarly, Ωy 1 is a disjoint union of sets of the form Ωy1 1 = Ωy1 y , where y1 ∈ Sδ(y) ∩ Sk′+δ. Refer to Figure 2 below. x2vnpi x2vn y1zvnpi x3 x2 Cx C ′ x .............................................................................................................................................................. .............................................................................................................................................................. ............................................................................... ............................................................................... ............................................................................... x x1 ............................................................................... ............................................................................... y1zvn y1z = y2 y3 Cy C ′ y .............................................................................................................................................................. .............................................................................................................................................................. ............................................................................... ............................................................................... ............................................................................... y y1 ............................................................................... ............................................................................... Figure 2. It is therefore enough to show that for every such x1, y1, there is a y which coincides pointwise x . That is, for almost measure preserving bijection ϕ : Ωx1 with the action of Γ almost everywhere on Ωx1 each ω ∈ Ωx1 x , there exists g ∈ Γ such that ϕω = gω. x → Ωy1 Fix such x1, y1. Choose x2 ∈ Sen(x1)∩Sk+δ+en. Also choose vn ∈ Sen such that x2vn ∈ S2en(x1) ∩ Sk+δ+2en . Since y1 ∈ Sk′+δ, it has normal form y1 = u1 . . . ul where ul ∈ Sen . We now show that there exists z ∈ Sen such that (2.1) u−1 l ∨ z = Π and z−1 ∨ vn = Π . 13 To prove the claim, it is necessary to make use of the identification of the generators of Γ with elements of the finite projective space Π. Set Πr = {x ∈ Π : dim x = r} = Ser . Now, u−1 l ∈ Π1, and vn ∈ Πn. Therefore (a): {z ∈ Πn : u−1 l ∨ z 6= Π} = {z ∈ Πn : u−1 l ⊂ z} = 1 + q + q2 + · · · + qn−1. (b): {z ∈ Πn : z−1 ∨ vn 6= Π} = {z ∈ Πn : z−1 ⊂ vn} = 1 + q + q2 + · · · + qn−1. Also, Πn = 1 + q + q2 + · · · + qn > 2(1 + q + · · · + qn−1) . Hence there exists z ∈ Πn such that (2.1) holds. It follows that the word y1zvn = u1 . . . ulzvn is in normal form and hence that y1zvn ∈ Sk′+δ+2en . Moreover, y2 = y1z ∈ Sk′+δ+en. It will now be shown that the chambers Cx, Cy can be constructed which lie as indicated in (the two dimensional) Figure 2. By this it is meant, for example, that if ω ∈ Ω and Cx ⊂ Sx2(ω) then Sx2(ω) ⊂ Sx1(ω) ⊂ Sx(ω). In fact, Cx = x2C and Cy = y1zC, where C is the chamber based at 1, as illustrated in Figure 3 (in two dimensions). vnpi w .................................................................................................................................................................................................................................................................................................. .................................................................................................................................................................................................................................................................................................. C 1 vn Figure 3. The vertices x2vnpi, 2 ≤ i ≤ n, will be constructed from a flag v−1 n ⊂ p2 ⊂ p3 ⊂ · · · ⊂ pn, where pi ∈ Πi. For the following argument, note that if b ∈ Πr−1, where r ≥ 2, then {a ∈ Πr : a ⊃ b} = 1 + q + · · · + qn−r+1 ≥ 1 + q . 14 PAUL CUTTING AND GUYAN ROBERTSON There are at least 1 + q elements p2 ∈ Π2 such that p2 ⊃ v−1 n . By reference to both Lemma 2.4 and the proof of Proposition 2.7 in [Ca3], there is precisely one such p2 such that x2vnp2 < x2vn. (In fact, in that case x2vnp2 ∈ Sk+δ+2en−en−1.) Similarly, there is precisely one p2 ∈ Π2, p2 ⊃ v−1 n such that y1zvnp2 < y1zvn. One can therefore choose p2 ∈ Π2 with p2 ⊃ v−1 n such that x2vnp2 ≥ x2vn and y1zvnp2 ≥ y1zvn. Moreover, since vnp2 is then adjacent to 1, these inequalities are in fact equalities. This process is now continued. There are at least 1 + q elements p3 ∈ Π3 such that p3 ⊃ p2 and at most two of them satisfy either x2vnp3 < x2vn or y1zvnp2 < y1zvn. Thus we may choose p3 ⊃ p2 such that x2vnp3 = x2vn and y1zvnp3 = y1zvn. Continue in this way to obtain a flag v−1 n ⊂ p2 ⊂ p3 ⊂ · · · ⊂ pn such that the vertex set of the chamber C is {1, vn, vnp2, vnp3, . . . , vnpn} . Then Cx = x2C and Cy = y2C. Now choose, by Lemma 2.4, a vertex w (one of q possible) of a chamber C ′ x which meets Cx in the face Cx\{1}. Thus w ∈ Sδ, and x3 = x2w ∈ Sk+2δ+en. Also y3 = y2w ∈ Sk′+2δ+en. (Recall that, by definition, y2 = y1z.) Moreover, y2x−1 2 Cx = Cy. It has now been shown that Ωx3 x2 ⊂ Ωx1 x , Ωy3 y2 ⊂ Ωy1 y and y2x−1 2 Ωx3 x2 = Ωy3 y2 . Therefore one can define the map ϕ on Ωx3 x2 by ϕω = y2x−1 2 ω . Now recall that x ∈ Sk, y ∈ Sk′ and x1 ∈ Sk+δ, y1 ∈ Sk′+δ were fixed, and that x2 ∈ Sen(x2) ∩ Sk+δ+en was chosen. The set Ωx1 x is a disjoint union of sets of the form Ωx3 x2 where x3 ∈ Sδ(x2). Let K denote the number of such sets. This number is independent of the choice of x, x1 and k by [Ca3, Lemma 2.4], (or by the fact that Γ acts simply transitively on X0). The definition ϕω = y2x−1 2 ω in the above choice of Ωx3 x2 therefore leaves ϕ undefined on a proportion (1 − 1 x . However, where ϕ is defined it coincides with the action of an element of Γ, namely y2x−1 2 . Now repeat the process on each of the K − 1 sets of the form Ωx3 x2 where ϕ has not been defined. As before, ϕ can be defined except on K ) of Ωx1 15 a proportion (1 − 1 everywhere except on a proportion (1 − 1 K ) of each such set, and ϕ can therefore be defined K )2 of the original set Ωx1 x . Continuing in this manner, one sees that at the nth step, ϕ has been defined everywhere except on a proportion (1 − 1 K )n of Ωx1 x . Since (1 − 1 K )n → 0 as n → ∞, ϕ is defined almost everywhere on x and satisfies the required properties. If k = k′ then it is clear from (cid:3) Ωx1 the construction that ϕ is measure preserving. Remark 2.6. This result extends [RR2, Proposition 4.9]. Moreover, for n = 2 this proof deals with the case q = 2, which was left open in [RR2]. Thus the hypothesis q ≥ 3 in the main result Theorem 4.10 of [RR2] is not in fact necessary. We can now prove the following. Proposition 2.7. There exists a countable subgroup K0 of [Γ] such that (1) K0 is measure preserving; (2) K0 acts ergodically on Ω. Proof. It suffices to take the group generated by the automorphisms of the form ϕ defined in Lemma 2.5, with k = k′ then use Lemma 2.2 to extract a countable subgroup K0. Finally, Lemma 2.1 shows that the action of K0 is ergodic. (cid:3) Lemma 2.8. Let Γ be a countable group acting on a measure space (Ω, µ). Suppose that the action of the full group [Γ] is ergodic. Then so is the action of Γ. Proof. Let S be a measurable subset of Ω such that µ(gS\S) = 0 for all g ∈ Γ. Let k ∈ [Γ]. It will be shown that µ(kS\S) = 0. For each g ∈ Γ, let Sg = {ω ∈ S : kω = gω} , which is a measurable subset of S. Since k ∈ [Γ] it follows that S = S0 ∪ [g∈Γ Sg , where S0 has measure zero. 16 PAUL CUTTING AND GUYAN ROBERTSON Then µ(kS\S) ≤ Xg∈Γ = Xg∈Γ ≤ Xg∈Γ µ(kSg\S) µ(gSg\S) µ(gS\S) = 0 . Thus µ(kS\S) = 0 for all k ∈ [Γ]. Since the action of [Γ] is ergodic it follows that S is either null or conull with respect to the measure µ. Therefore the action of Γ is ergodic. (cid:3) Corollary 2.9. The action of Γ on Ω is ergodic. Proof. This follows from Proposition 2.7 and Lemma 2.8. (cid:3) 3. Classification of the action of Γ on Ω Having shown that the action of an An group on its boundary Ω is ergodic, we now show that it is type IIIλ, where 0 ≤ λ ≤ 1, and the value of λ depends on the ratio set. Definition 3.1. Let Γ be a countable group of automorphisms of the measure space (Ω, ν). Following Krieger, define the ratio set r(Γ) to be the set of λ ∈ [0, ∞) such that for every ǫ > 0 and Borel set E with ν(E) > 0, there exists a g ∈ Γ and a Borel set F such that ν(F ) > 0, F ∪ gF ⊂ E and for all ω ∈ F . < ǫ dν ◦ g dν (cid:12)(cid:12)(cid:12)(cid:12) (ω) − λ(cid:12)(cid:12)(cid:12)(cid:12) Remark 3.2. The ratio set r(Γ) depends only on the quasi-equivalence class of the measure ν [HO, section I-3, Lemma 14]. It also depends only on the full group in the sense that [H] = [G] =⇒ r(H) = r(G) . Proposition 3.3. Let X be a locally finite, thick An building of order q, and let Γ be a countable group of type rotating automorphisms of X. Fix a vertex O ∈ X0 of type 0, and suppose that for each 0 ≤ i ≤ n there exists an element gi ∈ Γ such that d(giO, O) = 1 and giO is a 17 vertex of type i. Also, suppose that there exists a countable subgroup K of [Γ]0 whose action on Ω is ergodic. Then r(Γ) =(cid:26) {qn : n ∈ Z} ∪ {0} for n odd {q2n : n ∈ Z} ∪ {0} for n even . Proof. By Remark 3.2, it is sufficient to prove the statement for some group H such that [H] = [Γ]. In particular, since [Γ] = [hΓ, Ki] for any subgroup K of [Γ]0, we may assume without loss of generality that K ≤ Γ. Let ν = νO. For gi ∈ Γ as in the statement of the lemma, let x = giO, . If m(O, x; ω) = (m1, m2, . . . , mn) then by and note that νx = ν ◦ g−1 Lemma 1.6, i dν ◦ g−1 i (ω) = (ω) = q− Pn i=1 i(n+1−i)mi . dνx dν dν Then for ω ∈ Ωx O, one has that m(O, x; ω) = (0, . . . , 0, −1, 0, . . . , 0), where the −1 is in the ith place. Thus (3.1) dνx dν (ω) = qi(n+1−i) for ω ∈ Ωx O. Let E ⊂ Ω be a Borel set with ν(E) > 0. Then by the ergodicity of K, there exist k1, k2 ∈ K such that the set F = {ω ∈ E : k1ω ∈ Ωx O and k2g−1 i k1ω ∈ E} has positive measure. Next, let t = k2g−1 since K is measure preserving, i k1 ∈ Γ. By construction, F ∪ tF ⊂ E. Moreover, dν ◦ t dν (ω) = dν ◦ g−1 i dν (k1ω) = dνxi dν (k1ω) = qi(n+1−i) for all ω ∈ F by (3.1), and since kiω ∈ Ωx O. Hence qi(n+1−i) ∈ r(Γ) for 1 ≤ i ≤ n. Since the action of Γ on Ω is ergodic, r(Γ) − {0} forms a group. It is now possible to determine the generator of r(Γ) − {0}. Suppose that n is odd. Then for i ∈ {1, 2}, one has that qn, q2(n−1) ∈ r(Γ). As n, 2(n − 1) are coprime for n odd, and as r(Γ) − {0} forms a group, it follows that q ∈ r(Γ), and hence r(Γ) = {qn : n ∈ Z} for n odd. Suppose that n is even. As before, qn, q2(n−1) ∈ r(Γ). Moreover, as highest common factor of n, 2(n − 1) is 2 for n even, and as r(Γ) forms a group, it follows that q2 ∈ r(Γ). Finally, as i(n + 1 − i) is even for all i if n is even, it follows that for g ∈ Γ, and x = g−1O, dν ◦ g dν = dνx dν = q− P i(n+1−i)mi ∈ {q2n : n ∈ Z} . 18 PAUL CUTTING AND GUYAN ROBERTSON Thus r(Γ) = {q2n : n ∈ Z} for n even. (cid:3) Proposition 3.4. Let Γ be an An group. Then the action of Γ on Ω is amenable. Proof. This is a straightforward generalization of the case n = 2, (cid:3) proved in [RR2, Proposition 3.14]. 3.1. Proof of Theorem 1. This follows from Proposition 2.7, Corol- lary 2.9, Proposition 3.3 and Theorem 3.4. (cid:3) 3.2. Proof of Theorem 2. The proof of Theorem 2 is now easy. Let X be the affine building of G = PGL(n + 1, Qp). By [Br, Proposition VI.9F], the boundary Ω of X is isomorphic to PGL(n + 1, Qp)/B as a topological G-space. The measure µ on G/B is (up to equivalence) the natural quasi-invariant Borel measure on G/B. The vertex set X0 of X is identified with PGL(n + 1, Qp)/PGL(n + 1, Zp), where Zp is the ring of p-adic integers. Let O = PGL(n+1, Zp) ∈ X0. It follows from [St, Proposition 3.1] that PGL(n + 1, Zp) acts transitively on each set Sk(O). Since the vertex set X0 = PGL(n + 1, Qp)/PGL(n + 1, Zp) is a discrete space and Z is dense in Zp it follows that PGL(n + 1, Z) also acts transitively on Sk(O) for each k ∈ Zn +. Moreover PGL(n + 1, Z) stabilizes O. Therefore PGL(n + 1, Z) acts ergodically on Ω by Lemma 2.1. The group PGL(n + 1, Q) also acts transitively on X0. By our previous computation of Radon-Nikodym derivatives and the argument of [RR2, Proposition 4.4] the type of the action is as stated. Note that in this argument there is no need to consider the full group, since PGL(n + 1, Z) is already a measure preserving ergodic subgroup of PGL(n + 1, Q). Thus the proof is considerably simpler than the proof of Theorem 1. 4. Freeness of the action on the boundary A simple modification of [RR2, Proposition 4.12] shows that if F is a (possibly non commutative) local field then the action of PGL(n+1, F) on its Furstenberg boundary Ω is measure-theoretically free. If X is a thick, locally finite affine building of type An, where n ≥ 3, then X is the building of PGL(n + 1, F) for some such local field [R, p137]. All known type-rotating An groups, with n ≥ 3, embed in PGL(n + 1, F) and act upon the building of PGL(n + 1, F) in the canonical way. For such groups Γ, the action on Ω is therefore measure theoretically free. 19 The case of A2 groups is more interesting because the associated A2 building may not be the affine building of a linear group. In fact this is the case for the buildings of many of the groups constructed in [CMSZ]. The boundary action of Γ is nevertheless free. Proposition 4.1. Let X be an A2 building, and let Γ be an A2 group. Then the action of Γ on Ω is measure-theoretically free [RR2, c.f. Proposition 3.10]. This result was stated in [RR2, Proposition 3.10]. The proof given there contains a gap, because the proof of [RR2, Lemma 3.7, Case 2] is not complete. Our purpose is to fill this gap (Lemma 4.3 below). A similar argument applies to the An case, where n ≥ 3. In view of the comments at the beginning of this section, we have confined ourselves to the A2 case. Let O be a fixed vertex in X and let n ∈ Z+. Any sector S based at O has at its base a triangle Tn with apex O consisting of all vertices v in S such that d(v, O) ≤ n. (See Figure 4 below). . . 3n-3 .......................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................................................................................................................ .......................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................... .............................................................................................. .............................................................................................. ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................... .............................................................................................. .................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................. .............................................................................................. .............................................................................................. .............................................................................................. .............................................................................................. .......................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................... ............................................................................................................................................................................................ .............................................................................................. ........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................ .......................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................... ...................................................................................................................................................................................................................................................................................................................................................................................................................................................................................... .......................................................................................................................................................................................................................................................................................... .............................................................................................. .............................................................................................. .............................................................................................. .............................................................................................. 2n-1 . 4 3 2 2n 2n-2 .............................................................................................. 1 C O Figure 4. Consider the set Sn of all such triangles Tn in X. Lemma 4.2. Let Sn be as above. Then Sn = (q2 + q + 1)(q + 1)q3n−3 . Proof. First note that there are (q2 + q + 1)(q + 1) choices for the base chamber C. 20 PAUL CUTTING AND GUYAN ROBERTSON The reason for this is that one edge of C containing O is determined by the number of points P in a finite projective plane Π of order q. There are q2 + q + 1 such points P . See Figure 5 below. L P C ............................................................................................................................................................. ............................................................................................................................................................. O Figure 5. Having chosen the point P , there are precisely q + 1 possible lines L in Π which are incident with P . There are therefore (q2 + q + 1)(q + 1) choices of C. Having chosen C, there are q choices for each of the chambers labeled 1, 2, 3, 4, . . . , (2n − 2) in the figure. Then choose the chamber labeled (2n − 1) (q choices) in Figure 4. This choice then determines the whole shaded region in the figure (which is contained in the convex hull of the chambers already chosen, and hence is uniquely determined). Now choose the chamber labeled 2n (q choices) and continue the process until finally chamber 3n − 3 is chosen. This determines the triangle completely and there are (q2 +q +1)(q +1)q3n−3 possibilities altogether. (cid:3) This demonstrates that for each positive integer n, the boundary Ω of X is partitioned into (q2 + q + 1)(q + 1)q3n−3 sets {ΩT : T ∈ Sn}, where ΩT = {ω ∈ Ω : T ⊂ [O, ω)} . Moreover each of these sets has the same measure [CMS]. The proof of [RR2, Lemma 3.7, case 2.] can now be completed. Lemma 4.3. Let W be a wall of X and let Σ denote the set of boundary points ω ∈ Ω such that for some vertex v, the sector [v, ω) lies in an apartment containing W . Then νO(Σ) = 0 . Proof. By translating to a parallel sector, one can assume that v = O. Also, W is the union of two sector panels, which will be denoted by [O, W +), [O, W −). Given n ∈ Z+, let S + triangles T + W as illustrated below in Figure 6. n , T − n , T ⊥ n , S ⊥ n , S − n denote the subsets of Sn consisting of n respectively, lying in some apartment containing .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... ....................................................................................................................................................................................... T ⊥ ..................................................................................................................................................................................... ..................................................................................................................................................................................... .......... .......... ...................................................................................................................................................................................... T + T − .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... .......... O W [O, W −) 21 W [O, W +) Figure 6. Let S W n = S + n ∪ S − n ∪ S ⊥ n . Then (4.1) Σ ⊂ [T ∈SW n ΩT . The first step is to calculate the number of triangles in S W n . To do this, the number of possible choices for T + ∈ S + n must be determined. Refer to Figure 7 below. cn ............................................... ............................................... .................................................................................................................................................................................................................................................................................................................................... ............................................................................................................................................. ......................................................................................................................................................................................................................................................................................................................................... ........................................................................................................................................................................................................................................................................................................................................................................................ ............................................... ......................................................................................................................................................................................................................................................................................................................................... ............................................... ......................................................................................................................................................................................................................................................................................................................................... ............................................................................................................................................. ............................................... ............................................................................................................................................. ............................................................................................................................................. ............................................................................................................................................. ............................................................................................................................................................................................ ............................................................................................................................................. ............................................................................................................................................. ............................................... ............................................... ............................................... ............................................... ............................................................................................................................................. ........................................................................................................................................................................................................................................................................................................................................................................................ ............................................... ............................................................................................................................................. ............................................... ............................................... ............................................... 2 2 2 2 2 2 2 2 c2 ............................................... ............................................... ............................................................................................................................................. ............................................... ............................................................................................................................................. ............................................................................................................................................. ............................................................................................................................................. ............................................................................................................................................. ............................................... ............................................... ............................................................................................................................................. ............................................... ............................................... ............................................................................................................................................. 1 1 1 1 1 1 1 1 1 1 c1 O [O, W +) Figure 7. There are (q + 1) possible choices for the chamber c1. This choice then determines all the other chambers labeled 1 which lie in the convex hull of c1 and [0, W +). There are then q choices for the chamber c2. This choice now determines all chambers labeled 2 which are in the convex hull of c2 and all the other chambers previously determined. Continue in this way until the chamber cn is reached, thereby deter- n . There are thus (q + 1)qn−1 choices mining the whole triangle T + ∈ S + for T +. Now each triangle T + ∈ S + T − ∈ S − convex hull of S + n , T ⊥ ∈ S ⊥ n determines a unique pair of triangles n subject to the condition that T −, T ⊥ lie in the n ∪ W . Conversely such T − or T ⊥ determine T + 22 PAUL CUTTING AND GUYAN ROBERTSON uniquely. Hence the sets S + It follows that n , S − n , S ⊥ n have the same number of elements. S W n = 3(q + 1)qn−1 . Since the sets ΩT , T ∈ Sn have equal measure and partition Ω, it follows from Lemma 4.2 and equation (4.1) that νO(Σ) ≤ νO  [T ∈SW n ΩT  3(q + 1)qn−1 = (q2 + q + 1)(q + 1)q3n−3 → 0 as n → ∞ . Thus νO(Σ) = 0. (cid:3) References [Br] [Ca1] [Ca2] [Ca3] K. Brown. Buildings. Springer-Verlag, New York 1989. D. I. Cartwright, Groups acting simply transitively on the vertices of a building of type An, Groups of Lie type and their Geometries, W. M. Kantor and L. Di Martino, editors, 43 -- 76. Cambridge University Press, 1995. D. I. Cartwright, A brief introduction to buildings, Harmonic Functions on Trees and Buildings (New York 1995), 45 -- 77, Contemp. Math. 206, Amer. Math. Soc., 1997. D. I. Cartwright, Harmonic functions on buildings of type An, Random Walks and Discrete Potential Theory (Cortona 1997), 104 -- 138, Symposia Mathematica, Vol XXXIX, Cambridge University Press 1999. [CMS] D. I. Cartwright, W. M lotkowski, and T. Steger. Property (T) and A2 groups, Ann. Inst. Fourier 44 (1994), 213 -- 248. [CS] [CMSZ] D. I. Cartwright, A. M. Mantero, T. Steger, and A. Zappa, Groups acting simply transitively on the vertices of a building of type A2, I & II, Geom. Ded. 47 (1993), 143 -- 166 and 167 -- 226. D. I. Cartwright, and T. Steger, A family of An groups, Israel J. Math. 103 (1998), 125 -- 140. A. Connes, Une classification des facteurs de type III., Ann. Scient. Ec. Norm. Sup. 6 (1973), 133 -- 252. T. Hamachi and M. Osikawa, Ergodic Groups of Automorphisms and Krieger's Theorems, Seminar on Mathematical Sciences No. 3, Keio Uni- versity, Japan, 1981. [HO] [C] [Mar] G. A. Margulis, Discrete subgroups of semisimple Lie groups, Springer- [RR1] [RR2] Verlag, Berlin, 1991. J. Ramagge and G. Robertson, Factors from trees, Proc. Amer. Math. Soc. 125 (1997), 2051 -- 2055. J. Ramagge and G. Robertson, Triangle buildings and actions of type III1/q2 , J. Func. Anal. 140 (1996), 472 -- 504. 23 [R] [S] [St] [Su] [Sv] [Zi] M. A. Ronan, Lectures on Buildings, Perspectives in Math. 7, Academic Press, London 1989. R. J. Spatzier, An example of an amenable action from geometry, Ergod. Th. & Dynam. Sys. 7 (1987), 289 -- 293. T. Steger, Local fields and buildings, Harmonic Functions on Trees and Buildings (New York 1995), 79 -- 107, Contemp. Math. 206, Amer. Math. Soc., 1997. V. S. Sunder, An invitation to von Neumann Algebras, Universitext, Springer-Verlag, New York 1987. T. Svenson, Groups acting simply transitively on An buildings, Ph.D. Thesis, University of Sydney, 1998. R. L. Zimmer, Ergodic Theory and Semisimple Groups, Birkhauser, Boston 1985. Mathematics Department, University of Newcastle, Callaghan, NSW 2308, Australia E-mail address: [email protected] E-mail address: [email protected]
1303.4483	2	1303	2013-05-29T06:22:33	Purely infinite partial crossed products	[ "math.OA" ]	Let (A,G,\alpha) be a partial dynamical system. We show that there is a bijective correspondence between G-invariant ideals of A and ideals in the partial crossed product A xr G provided the action is exact and residually topologically free. Assuming, in addition, a technical condition---automatic when A is abelian---we show that A xr G is purely infinite if and only if the positive nonzero elements in A are properly infinite in A xr G. As an application we verify pure infiniteness of various partial crossed products, including realisations of the Cuntz algebras O_n, O_A, O_N, and O_Z as partial crossed products.	math.OA	math	PURELY INFINITE PARTIAL CROSSED PRODUCTS THIERRY GIORDANO1 AND ADAM SIERAKOWSKI2 Abstract. Let (A, G, α) be a partial dynamical system. We show that there is a bijective correspondence between G-invariant ideals of A and ideals in the partial crossed product A ⋊α,r G provided the action is exact and residually topologically free. Assuming, in addition, a techni- cal condition -- automatic when A is abelian -- we show that A ⋊α,r G is purely infinite if and only if the positive nonzero elements in A are prop- erly infinite in A ⋊α,r G. As an application we verify pure infiniteness of various partial crossed products, including realisations of the Cuntz algebras On, OA, ON, and OZ as partial crossed products. 1. Introduction In the theory of operator algebras, the crossed product construction has been one of the most important and fruitful tools both to construct examples and to describe the internal structure of operator algebras (in particular the von Neumann algebras). Partial actions of a discrete group on C∗-algebras and their associated crossed products were gradually introduced in [14] and [35], and since then developed by many authors. Several important classes of C∗-algebras have been realised as crossed products by partial actions, including in particular AF-algebras, Cuntz-Krieger algebras, Bunce-Deddens algebras, among oth- ers (see for example [5, 12, 13, 15, 16, 19, 26]). The description of C∗-algebras as partial crossed products has also proved useful for the computation of their K-theory. In this paper we pursue the study of partial C∗-dynamical systems and their crossed products associated. We begin by recalling (in Section 2 and Appendix A.1) the construction of the partial crossed product A ⋊α,r G associated to a partial action α (a compatible collection of isomorphisms αt : Dt−1 → Dt, t ∈ G of ideals in A). In Section 3, we study the ideal structure of partial crossed products, generalising the results on C∗-dynamical systems obtained by the second author in [41]. Recall that a partial action α on a C∗-algebra A has the the Date: May 29 2013. 1Research supported by NSERC, Canada. 2Communicating author. Research supported by Professor Collins, Handelman and Giordano, Canada. 1 2 THIERRY GIORDANO AND ADAM SIERAKOWSKI intersection property if every nontrivial ideal in A ⋊α,r G intersects A non- trivially. Then (Definition 3.1) we say that a partial C∗-dynamical system (A, G, α) has the residual intersection property if for every G-invariant ideal I in A, the induced partial action of G on A/I has the intersection property. We establish in Theorem 3.2 a one-to-one correspondence between ideals in A ⋊α,r G and G-invariant ideals of A provided -- in fact if and only if -- the partial action α is exact and has the residual intersection property. When the partial action α is minimal (no nontrivial G-invariant ideals in A), then the exactness (any nontrivial G-invariant ideal I in A induces a short exact sequence at the level of reduced crossed products) is automatic. In [19], having defined a topologically free partial action (by partial home- omorphisms) on a locally compact space X, Exel, Laca, and Quigg proved the simplicity of the partial crossed product C0(X) ⋊α,r G under the pres- ence of minimality and topological freeness of α. In [33], Lebedev extended the definition of topological freeness to non-commutative partial actions and showed that a topologically free partial action has always the intersection property. With Lebedev's result, we recover in Corollary 2.9 the theorem of Exel, Laca, and Quigg. Theorem 3.2 allows us also to extend Echterhoff and Laca's work on crossed products: We show that the canonical map J 7→ J ∩ A between ideals in A ⋊α,r G and G-invariant ideals of A restricts to a continuous map from the space of prime ideals of A ⋊α,r G to the space of G-prime ideals in A. This restriction is a homeomorphism provided α is exact and residually topologically free, where residual topological freeness is an ideal related ver- sion of topological freeness. When A is separable and abelian we show that the space of prime ideals of A ⋊α,r G is homeomorphic to the quasi-orbit space of Prim A. In Section 4 we generalise some of the main results in [39], by Rørdam and the second named author, to partial C∗-dynamical systems. In particular we give sufficient conditions for a partial crossed products to be purely infinite in the sense of Kirchberg and Rørdam. One of the keys assumptions of The- orem 4.2 goes back to Elliott's notion of proper outerness (an automorphism α of A is properly outer if kαI − βk = 2 for every α-invariant ideal I in A and any inner automorphism β of I). Based on Kishimoto's work, Olsen and Pedersen proved that an automorphism α of a separable C∗-algebra A is properly outer if and only if inf{kx(aδ)xk : x ∈ B+, kxk = 1} = 0 for every a ∈ A, and every nonzero hereditary C∗-algebra B in A, where δ is the unitary implementing α, i.e, α(a) = δaδ∗ in A ⋊α Z. Moreover Archbold and Spielberg proved that topological freeness of an action α ensures that αt, t 6= e is properly outer. In Proposition 3.10, generalising these results, we prove the equivalence of the following statements for a partial action α on an abelian C∗-algebra A: (i) α is topologically free, (ii) kαtI − idI k = 2 for every αt-invariant ideal I in Dt−1 (and t 6= e), PURELY INFINITE PARTIAL CROSSED PRODUCTS 3 (iii) inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0 for every a ∈ Dt, and every nonzero hereditary C∗-algebra B in A (and t 6= e). Without assuming that A is commutative and under the assumption of property (iii) for actions of G on A/I for any G-invariant ideal I in A and exactness of α we once again obtain a one-to-one correspondence between ideals in A ⋊α,r G and G-invariant ideals in A. However, in this context we can state in Theorem 4.2 sufficient and necessary conditions for pure infiniteness of the partial crossed products: A ⋊α,r G is purely infinite if and only if every nonzero positive element in A is properly infinite in A ⋊α,r G (under presence of ideal property of A). In the second part of Section 4, we study pure infiniteness of partial crossed products A ⋊α,r G when A is abelian. We establish a geometrical condition sufficient -- and sometimes also necessary -- to obtain pure infinite- ness of partial crosses products. Specifically we show in Theorem 4.4 that an exact and residually topologically free partial action on a totally dis- connected locally compact Hausdorff space X gives a purely infinite partial crossed product C0(X) ⋊α,r G provided that every compact and open subset of X is (G, τX )-paradoxical (Definition 4.3). In Section 5, we apply our results to the a variety of know examples of partial crossed products: (i) Hopenwasser constructs a partial action of the semidirect product Qn ×δ Z of n-adic rationals by the integers Z on the Cantor set XC such that the partial crossed product C(XC) ⋊α (Qn ×δ Z) is isomorphic to the Cuntz algebra On. We verify that for this (exact and residually topologically free) action the compact and open subset of XC are (Qn ×δ Z, τXC )-paradoxical. (ii) For a {0, 1}-valued n by n matrix A = [aij] with no zero rows Exel, j = 1,Pj aijsjs∗ Laca, and Quigg realised OA = C∗(s1, . . . , sn : Pj sjs∗ j = s∗ i si) as a partial crossed product C(XA) ⋊α Fn, where Fn denotes the free group of rank n. The algebra OA was defined by Astrid an Huef and Raeburn as an universal analogue of the Cuntz-Krieger algebra. When α is exact and residually topologically free we prove that C(XA) ⋊α Fn is purely infinite if and only if the compact and open subset of XA are (Fn, τXA)-paradoxical. We use graph C∗-algebraic results by Raeburn, Kumjian, Pask, Raeburn, Renault, Mann, Sutherland, Hong and Szymanski among others. (iii) Boava and Exel showed that for each integral domain R with finite quotients R/(m), m 6= 0, the semidirect product K ⋊ K × of K by K\{0} (where K is the field of fractions of R) acts on the Cantor set XR such that the partial crossed product C(XR) ⋊α (K ⋊ K ×) is isomorphic to the regular C∗-algebra A[R] of R. We verify that for this (exact and residually topologically free) action the compact and open subset of XR are (K ⋊ K ×, τXR)-paradoxical. In [1], Abadie described a class of partial crossed products C0(X) ⋊α,r G Morita-Rieffel equivalent to ordinary crossed products. In Section 6, we prove that for such a partial crossed product C0(X) ⋊α,r G if the compact 4 THIERRY GIORDANO AND ADAM SIERAKOWSKI and open subsets in the spectrum of the the original C∗-algebra of the partial crossed product are paradoxical, then the same property must holds for the corresponding (ordinary) crossed product. Much of the work described in this paper was done while the second author held a postdoctoral fellowship at the University of Ottawa. He wishes to thank the members of the Department of Mathematics and Statistics at the University of Ottawa for their warm hospitality. 2. Definition of partial crossed product. Let A be a C∗-algebra and G be a discrete group. We recall the definition of a partial dynamical system and the corresponding partial crossed products (see Appendix A.1 for more details). Definition 2.1. Let A be a C∗-algebra and G be a discrete group. A partial action of G on A, denoted by α, is a collection (Dt)t∈G of closed two-sided ideals of A and a collection (αt)t∈G of ∗-isomorphisms αt : Dt−1 → Dt such that (i) De = A, where e represents the identity element of G; (ii) α−1 (iii) αt ◦ αs(x) = αts(x), ∀ x ∈ α−1 s (Ds ∩ Dt−1) ⊆ D(ts)−1 ; s (Ds ∩ Dt−1). The triple (A, G, α) is called a partial dynamical system. The equivalence between this definition of partial action by Dokuchaev and Exel in [9] and the original definition of McClanahan ([35]) was proven in [37]. For the case when A is abelian we refer to [1, 12, 19]. Definition 2.2. Let (A, G, α) be a partial dynamical system. Let L be the normed ∗-algebra of the finite formal sums Pt∈G atδt, where at ∈ Dt. The operations and the norm in L are given by (atδt)(asδs) = αt(αt−1 (at)as)δts, kXt∈G atδtk = Xt∈G (atδt)∗ = αt−1(a∗ katk. t )δt−1 , Let Bt denote the vector subspace Dtδt of L. The family (Bt)t∈G generates a Fell bundle. The full crossed product A ⋊α G and the reduced crossed product A ⋊α,r G are, respectively, the full and the reduced cross sectional algebras of (Bt)t∈G. Both crossed products are completions of L with respect to a certain C∗-norm. We recall the construction of these crossed products in Appendix A.1. We suppress the canonical inclusion map A → A ⋊α G, a 7→ aδe and view A as a sub-C∗-algebra of A ⋊α G. All ideals (throughout out this paper) are assumed to be closed and two-sided. The set τX denotes the topology of a topological space X. The C ∗-algebra of continuous functions vanishing at infinity on a locally compact Hausdorff space X is denoted by C0(X). Every abelian C ∗-algebra arises in this form. When the algebra is unital then X is compact and we emphasise this fact by writing it as C(X). PURELY INFINITE PARTIAL CROSSED PRODUCTS 5 3. Ideal structure of partial crossed product. Before stating our results on the ideal structure of a partial crossed prod- uct generalising [41] we need a few definitions. Definition 3.1. Let (A, G, α) be a partial dynamical system. Then (i) A closed two-sided ideal I of A is G-invariant provided that αt(I ∩ Dt−1) ⊆ I for every t ∈ G. (ii) The partial action is exact if every G-invariant ideal I of A induces a short exact sequence 0 / I ⋊α,r G / A ⋊α,r G / A/I ⋊α,r G / 0 at the level of reduced crossed products. (iii) The partial action has the residual intersection property if for every G-invariant ideal I of A the intersection of A/I with any nonzero ideal in A/I ⋊α,r G is nonzero. Theorem 3.2. Let (A, G, α) be a partial dynamical system. There is a one- to-one correspondence between ideals in A ⋊α,r G and G-invariant ideals of A if and only if the partial action is exact and has the residual intersection property. Proof. Sufficiency: Suppose that α is exact and has the residual intersection property. Let EA : A ⋊α,r G → A denote the usual conditional expectation on the crossed product (see Appendix A.1). Let Ideal[S] denote the smallest ideal in A ⋊α,r G generated by S ⊆ A ⋊α,r G. Let ϕ denote the map J 7→ J ∩ A from the ideals in A ⋊α,r G into G-invariant ideals in A. Using that Ideal[I] ∩ A = I, for any G-invariant ideals I of A, cf. Proposition A.4, we conclude that ϕ is surjective. To show ϕ is injective it is enough to show that J = Ideal[EA(J )] for every ideal J in A ⋊α,r G. If we have two ideals J 1, J 2 of A ⋊α,r G with the same intersection it then follows that J 1 = Ideal[EA(J 1)∩A] ⊆ Ideal[J 1∩A] = Ideal[J 2∩A] ⊆ Ideal[EA(J 2)] = J 2, using that EA(J 1) ⊆ Ideal[EA(J 1)] = J 1, and J 2∩A ⊆ EA(J 2) (cf. Propo- sition A.4). Fix any ideal J of A ⋊α,r G. (i) J ⊆ Ideal[EA(J )]: Let I denote the (smallest) G-invariant ideal in A generated by EA(J ). By Proposition A.3 we have the commuting diagram 0 0 / I ⋊α,r G ι A ⋊α,r G ρ A/I ⋊α,r G EI / I EA / A EA/I / A/I 0 / 0 For each x ∈ J + we have EA(x) ∈ I and hence that EA/I (ρ(x)) = EA(x) + I = 0. By exactness ker ρ ⊆ Ideal[I] (not true in general). Hence x ∈ ker ρ = Ideal[I] = Ideal[EA(J )]. Since every element in J is a linear combination of positive elements J ⊆ Ideal[EA(J )]. / / / / / / / / / / / / / / / 6 THIERRY GIORDANO AND ADAM SIERAKOWSKI (ii) Ideal[EA(J )] ⊆ J : Let I denote the intersection J ∩ A. By exactness we have a short exact sequence 0 / I ⋊α,r G ι / A ⋊α,r G ρ / A/I ⋊α,r G / 0 With K := I ⋊α,r G ⊆ J we can therefore make the following identifications: ρ(J ) = J /K, A/I = A/(K ∩ A) = (A + K)/K. Suppose that J /K and (A + K)/K has a nonzero intersection. Then there exist j ∈ J and a ∈ A such that j + K = a + K 6= K. Since K ⊆ J it follows that a ∈ J and hence that a ∈ J ∩ A = I ⊆ K. But then a + K = K giving a contradiction. We conclude that ρ(J ) ∩ A/I = 0. The residual intersection property implies that ρ(J ) = 0. By Proposition A.3 we have the commuting diagram 0 0 / I ⋊α,r G ι A ⋊α,r G ρ A/I ⋊α,r G EI / I EA / A EA/I / A/I 0 / 0 For each x ∈ J we have that EA/I (ρ(x)) = EA(x) + I = 0. Hence EA(x) ∈ I = J ∩ A. We conclude that Ideal[EA(J )] ⊆ J . Necessity: Suppose that ϕ is bijective, where ϕ denotes the map J 7→ J ∩ A from the ideals in A⋊α,r G into G-invariant ideals in A. As previously let EA : A ⋊α,r G → A denote the conditional expectation on the crossed product. (iii) Exactness: Fix any G-invariant ideal I of A. By Proposition A.3 we have the commuting diagram 0 0 / I ⋊α,r G ι A ⋊α,r G ρ A/I ⋊α,r G EI / I EA / A EA/I / A/I 0 / 0 By assumption J := ker ρ has the form (J ∩ A) ⋊α,r G, cf. Proposition A.4. This implies that EA/I (ρ(J )) = EA(J ) + I = 0. Hence EA(J ) ⊆ I. By Proposition A.4 we also have that J ∩ A ⊆ EA(J ). We conclude ker ρ = (J ∩ A) ⋊α,r G ⊆ Ideal[EA(J )] ⊆ I ⋊α,r G. (iv) Residual intersection property: Fix any G-invariant ideal I of A and any ideal J of A/I ⋊α,r G with zero intersection with A/I. By Proposition A.3 we have the commuting diagram 0 0 / I ⋊α,r G ι A ⋊α,r G ρ A/I ⋊α,r G EI / I EA / A EA/I / A/I 0 / 0 / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / PURELY INFINITE PARTIAL CROSSED PRODUCTS 7 By assumption J 1 := ρ−1(J ) has the form (J 1 ∩ A) ⋊α,r G, cf. Proposition A.4. Hence EA/I (J ) = EA/I (ρ(J 1)) = EA(J 1) + I = (J 1 ∩ A) + I = EA(J 1 ∩ A) + I = EA/I (ρ(J 1 ∩ A)) ⊆ EA/I (ρ(J 1) ∩ ρ(A)) = EA/I (J ∩ A/I) = J ∩ A/I = 0 By the faithfulness of the conditional expectation (on positive elements) we conclude that J = 0. (cid:3) Remark 3.3. Exactness of a partial action is somehow mysterious because there are no concrete examples of a partial action that is not exact (even though existence has been established). Nevertheless, exactness plays an im- portant role as we have seen above. Here is another result ([40, Proposition 2.2.4]), proved by the second named author, relying heavily on exactness: Let (A, G, α) be a partial dynamical system. The action α is exact and A ⋊α,r G ∼= A ⋊α G if, and only if, A/I ⋊α,r G ∼= A/I ⋊α G for every G-invariant ideal I in A. If A is a C∗-algebra, let Prim A denote its primitive ideal space and A it spectrum. Moreover, if J is a closed two-sided ideal of A, then supp J will denote the subset {I ∈ Prim A : J * I} and AJ = {[π] ∈ A : π(J ) 6= 0}. Following Lebedev in [33], we define the map θt : supp Dt−1 → supp Dt as follows; for any point x ∈ supp Dt−1 such that x = ker π, where [π] ∈ ADt−1 , we set θt(x) = ker (π ◦ αt−1), t ∈ G. Notice that we then have that (θt)t∈G is an partial action of G by partial homeomorphisms of Prim A (see Appendix A.2 for more details). Definition 3.4. Let (A, G, α) be a partial dynamical system. (i) The partial action α is topologically free if for any finite set F in G \ {e} the union [t∈F (cid:8)x ∈ supp Dt : θt(x) = x(cid:9) has empty interior, cf. [33]. (ii) The partial action α is residually topologically free if the induced action of G on A/I is topologically free for every G-invariant ideal I of A. In [33], Lebedev shows that a residually topologically free action has always the residual intersection property (see Appendix A.2 for details). Hence, we have Corollary 3.5. Let (A, G, α) be a partial dynamical system. Suppose that the action is exact and residually topologically free. Then there is a one- to-one correspondence between ideals in A ⋊α,r G and G-invariant ideals in A. 8 THIERRY GIORDANO AND ADAM SIERAKOWSKI There is a useful consequence of Corollary 3.5 regarding G-prime ideals. Recall that a G-invariant ideal I in a C∗-algebra A is called G-prime (resp. prime if G is the trivial group) if for any pair of G-invariant ideals J , K of A with J ∩ K ⊆ I we have J ⊆ I or K ⊆ I, cf. [10]. Let I(A) denote the set of closed two-sided ideals in A. Imposing that the set {J ∈ I(A) : I 6⊆ J } is open for any I ∈ I(A) defines a sub-basis for the Fell-topology on I(A). This topology induces topologies on the set of prime and G-prime ideals in I(A). Corollary 3.6. Let (A, G, α) be a partial dynamical system. The map J 7→ J ∩ A, J ∈ I(A ⋊α,r G), restricts to a continuous map from the space of prime ideals of A ⋊α,r G to the space of G-prime ideals in A. Moreover, if the action of G on A is exact and residually topologically free this restriction is a homeomorphism. Proof. The second statement is contained in [10] for ordinary crossed prod- ucts and it is evident that the proof generalizes. The same hold for the fact that the map J 7→ J ∩ A is continuous. We conclude that the restriction is continuous (provided it is well defined). The fact that it is well defined follows from: The ideal J ∩ A is G-prime for any prime ideal J in A⋊α,r G. The proof is contained in [10] omitting one detail. If I and K are G-invariant ideals in A then (I ⋊α,r G) ∩ (K ⋊α,r G) ⊆ (I ∩ K) ⋊α,r G. To see this we use that the intersection of two closed two- sided ideals equals their product and an approximation argument. It is then sufficient to show that if aδt ∈ I ⋊α,r G and bδt ∈ K ⋊α,r G then (aδt)(bδs) ∈ (IK) ⋊α,r G. But this is evident from (aδt)(bδs) = αt(αt−1 (a)b)δts. (cid:3) Remark 3.7. If a partial action of a discrete group G on a C∗-algebra A is exact then Ti(I i ⋊α,r G) = (Ti Ii) ⋊α,r G for any family (I i) of G-invariant ideals, cf. [40]. G on Prim A define Gx := St∈G for which x∈supp Dt−1 Let x be a fixed element in Prim A. Using the partial action (θt)t∈G of {θt(x)}. The then quasi- orbit space O(Prim A) of Prim A is defined as the quotient space Prim A/ ∼ by the equivalence relation x ∼ y ⇔ Gx = Gy. Recall that for A = C0(X) the partial action α is induced by a partial action θ of G on X, i.e., a collection of open sets (Ut)t∈G and a collection (θt)t∈G of homeomorphisms θt : Ut−1 → Ut such that Ue = X and θst extends θs ◦ θt, cf. [35, 16, 37] and Appendix A.2. Then α is given by αt(f )(x) := f (θt−1(x)), f ∈ C0(Ut−1 ). So, here the ideals are Dt = C0(Ut). As the canonical homeomorphism from Prim A to X is G-equivariant we identify the actions on Prim A and X. The next corollary and is proof is a generalisation of [10, Lemma 2.5 and Corollary 2.6]. PURELY INFINITE PARTIAL CROSSED PRODUCTS 9 Corollary 3.8. Let (A, G, α) be a partial dynamical system with A abelian and separable. Suppose that the action of G on A is exact and residually topologically free. Then the space of prime ideals of A ⋊α,r G is homeomor- phic to the quasi-orbit space of Prim A. Proof. We show that the following map is a homeomorphism: [x] 7→ Ix ⋊α,r G, Ix := \I∈Gx I, [x] ∈ O(Prim A). Well defined: Fix any [x] ∈ TI∈Gx I. It is evident that Ix is a closed two-sided ideal in A. Let π : A → B(H) be the irreducible representation corresponding to x, i.e. x = ker π. Since A = C0(X) for some locally compact Hausdorff space [4, II.6.2.9] ensures that π(f ) = f (y) for some y ∈ X, so x = C0(X \ {y}). Hence θt(x) = ker(π ◦ αt−1) = {f : αt−1(f )(y) = 0} = C0(X \ {θt(y)}). By continuity we obtain that Ix = C0(X \ Gy). This ideal is G-invariant since Gy is G-invariant (and the complement of a G-invariant set is also G- invariant). Let us verify Gy is G-invariant for completeness. Fix any s ∈ G and any θt(y) ∈ Gy ∩ Us−1, where Ds = C0(Us). Since θt(y) ∈ Ut ∩ Us−1 [37, Lemma 1.2] ensures that y ∈ θt−1(Ut ∩ Us−1) = Ut−1 ∩ U(st)−1. In particular θs(θt(y)) = θst(y). We conclude θs(Gy ∩ Us−1) ⊆ Gy making Gy invariant. The fact that Gy is invariant is a simple limit argument used on Gy ∩ Us−1 ⊆ Gy ∩ Us−1. Let us verify that Ix is G-prime. Suppose that U, V are closed G-invariant subsets of X. If C0(X \ U ) ∩ C0(X \ V ) ⊆ Ix it follows that y ∈ U ∪ V , hence either C0(X \ U ) ⊆ Ix or C0(X \ V ) ⊆ Ix. It now follows from Corollary 3.6 that Ix ⋊α,r G is prime. Surjectivity ([10, Lemma 2.5]): Let J be any prime ideal in A ⋊α,r G. By Corollary 3.6 the ideal I = J ∩ A is G-prime. As I is G-invariant I = C0(X \ V ) for some nonempty closed G-invariant set V in X. Let ρ : Prim A → T denote the quotient map into the quasi-orbit space T := O(Prim A). To show that V = Gx for some x ∈ X we verify that F := ρ(V ) is the closure of a single point in T . Notice that T is totally Baire (every intersection of an open and a closed subset is a Baire space) and second countable. For such a space a non-empty closed subset F is the closure of a simple point if and only if F is not a union of two proper closed subsets, cf. [24, Lemma p. 222]. However such two proper closed subsets give raise to two closed G-invariant subsets U1, U2 ( V with U1 ∪ U2 = V and hence two G-invariant ideals I1, I1 ) I with I 1 ∩ I2 = I. This contradicts G- primeness of I. Injectivity: Fix any two prime ideals Ix ⋊r,α G and Iy ⋊r,α G in A ⋊r,α G such that Ix⋊r,α G = I y ⋊r,α G. By Proposition A.4 we obtain C0(X \Gx) = Ix = Iy = C0(X \ Gy). We conclude that x ∼ y. (cid:3) 10 THIERRY GIORDANO AND ADAM SIERAKOWSKI Corollary 3.9 (Lebedev [33]). Let (A, G, α) be a partial dynamical system. Suppose that the action is minimal (i.e, A does not contain any non-trivial G-invariant ideals) and topologically free. Then the crossed product A⋊α,r G is simple. The notation of topological freeness for partial actions is well known. We recall it in Appendix A.2 and show that the following equivalent conditions hold: Proposition 3.10. Let (A, G, α) be a partial dynamical system with A abelian. Then the following properties are equivalent: (i) α is topologically free (ii) kαtK − idKk = 2 for every αt-invariant ideal K in Dt−1 (and t 6= e) (iii) inf{kxαt(x)k : x ∈ K+, kxk = 1} = 0 for every nonzero ideal K in Dt−1 (and t 6= e) (iv) inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0 for every a ∈ Dt, and every nonzero hereditary C∗-algebra B in A (and t 6= e) In [11], Elliott defines an automorphism α of a C∗-algebra A to be properly outer if for any nonzero α-invariant ideal I of A and any inner automorphism β of I, kαI − βk = 2. Proper outerness for dynamical systems has been vastly studied in [31, 36, 3, 40, 41] among others. We will address in one of our upcoming works how one might generalise the notion proper outerness beyond ordinary crossed products. In the rest of the section, we generalize [19], Theorem 2.6, to the non- abelian case using condition (iv) of Proposition 3.10. We first need the following lemma. Lemma 3.11. Let (A, G, α) be a partial dynamical system. Suppose that for every t 6= e, every a ∈ Dt, and every nonzero hereditary C∗-algebra B in A inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0. Then for every b ∈ (A ⋊α,r G)+ and every ε > 0 there exist a positive contraction x ∈ A satisfying kxEA(b)x − xbxk < ε, kxEA(b)xk > kEA(b)k − ε. Proof. Fix b ∈ (A ⋊α,r G)+ and ε > 0. (i) We may assume b ∈ L: Since L is dense in A ⋊α,r G there exist c ∈ (A ⋊α,r G)+ ∩ L such that kc − bk < ε. Find a positive contraction x ∈ A PURELY INFINITE PARTIAL CROSSED PRODUCTS 11 satisfying kxEA(c)x − xcxk < ε, and kxEA(c)xk > kEA(c)k − ε. Then kxEA(b)x − xbxk ≤ kxEA(b)x − xEA(c)xk + kxEA(c)x − xcxk + kxcx − xbxk < 3ε, kEA(b)k < kEA(b − c)k + ε + kxEA(c − b)xk + kxEA(b)xk < kxEA(b)xk + 3ε. (ii) We may assume EA(b) has norm one: If b = 0 then any positive kEA(b)k using EA is faithful. kEA(b)k , contraction x ∈ A works. For b 6= 0 define c := Find a positive contraction x ∈ A satisfying kxEA(c)x − xcxk < and kxEA(c)xk > kEA(c)k − kEA(b)k . Then ε b ε kxEA(b)x − xbxk < ε, kxEA(b)xk > kEA(b)k − ε. (iii) It is enough to show that: For every ε > 0, every b0 ∈ A+, with kb0k = 1, every finite set F ⊆ G \ {e}, and every sequence of elements (bt)t∈F , with bt ∈ Dt, there exist x ∈ A+ such that kxk = 1, kxb0xk > kb0k − ε, kx(btδt)xk < ε, t ∈ F. By (i) − (ii) we may assume that b = b0 + Pt∈F btδt for some finite set F ⊆ G \ {e}, where b0 = EA(b) is positive (since EA is positive) and has norm one. Using the assumption on b0 and the sequence (bt)t∈F choose x ∈ A+ such that kxk = 1, kxb0xk > kb0k − ε, kx(btδt)xk < ε, t ∈ F. We conclude that kxEA(b)xk > kEA(b)k − ε, kxEA(b)x − xbxk = kx(b − b0)xk ≤ Xt∈F kx(btδt)xk < F ε. (iv) Finishing the proof: Fix ε > 0, b0 ∈ A+, with kb0k = 1, a finite set F ⊆ G \ {e}, and a sequence of elements (bt)t∈F , with bt ∈ Dt. Let f : [0, 1] → [0, 1] be a continuous increasing function taking the value zero on [0, 1 − ε] and one on [1 − ε 2 , 1]. Define x0 := f (b0) and set B1 := {x ∈ A : xx0 = x0x = x}. If g : [0, 1] → [0, 1] is a continuous increasing function equal to zero on [0, 1− ε 2 ] and one on [1 − ε 4 , 1] then gf = f g = g and kg(b0)k = 1. We conclude that B1 is nonzero. Using the C∗-norm identify it follows that L = {x ∈ A : x∗x ∈ B1} is a closed left ideal in A. Verifying that B1 = L ∩ L∗ we obtain that B1 is hereditary. Write F := {t1, . . . , tn}. Since bt1 ∈ Dt1 and B1 is nonzero hereditary in A inf{kx(bt1 δt1 )xk : x ∈ (B1)+, kxk = 1} = 0. Select x′ 1k < ε. Let h : [0, 1] → [0, 1] be a continuous increasing function equal to the identity on [0, 1 − ε] 1 ∈ (B1)+ such that kx′ 1k = 1 and kx′ 1(bt1 δt1)x′ 12 THIERRY GIORDANO AND ADAM SIERAKOWSKI and one on [1 − ε 2 , 1]. With x1 := h(x′ 1) we have (since gh = hg = g) that x1 ∈ (B1)+, B2 := {x ∈ B1 : xx1 = x1x = x} ∋ g(x′ kx1k = 1, 1) 6= 0. kx1(bt1 δt1 )x1k < (2 + kbt1 k)ε, Since bt2 ∈ Dt2 and B2 is nonzero hereditary in A inf{kx(bt2 δt2 )xk : x ∈ (B2)+, kxk = 1} = 0. Repeating the procedure above we obtain a sequence of hereditary C∗- algebras B1 ⊇ B2 ⊇ · · · ⊇ Bn, and positive elements of norm one in them, x1, x2, . . . , xn, fulfilling that xi ∈ (Bi)+, kxik = 1, kxi(btiδti )xik < (2 + kbtik)ε. With x := xn we have that xxi = xix = x for 0 ≤ i < n. It follows that x ∈ A+, kxk = 1, kx(bti δti)xk < (2 + kbti k)ε, i = 1, . . . , n. Finally, since f (t)2t ≥ (1 − ε)f (t)2 for every t ∈ [0, 1], we get xx0b0x0x ≥ (1 − ε)xx0x0x, kxb0xk > kb0k − ε. (cid:3) Theorem 3.12. Let (A, G, α) be a partial dynamical system. Suppose that for every t 6= e, every a ∈ Dt, and every nonzero hereditary C∗-algebra B in A inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0. Then every nonzero ideal in A ⋊α,r G has a nonzero intersection with A. Proof. The proof by Exel, Laca and Quigg in [19] for partial crossed product with A abelian generalises to the non-abelian case by means of Lemma 3.11. (cid:3) Corollary 3.13. Let (A, G, α) be a partial dynamical system. Suppose that the action is exact and that for every t 6= e, every G-invariant ideal I in A, every a ∈ Dt/(Dt ∩ I), and every nonzero hereditary C∗-algebra B in A/I inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0. Then there is a one-to-one correspondence between ideals in A ⋊α,r G and G-invariant ideals in A. Proof. Using Theorem 3.12 we obtain that α has the residual intersection property. The desired correspondence follows from Theorem 3.2. (cid:3) PURELY INFINITE PARTIAL CROSSED PRODUCTS 13 4. Pure infiniteness of partial crossed products We recall the definition of purely infinite C∗-algebras with the ideal prop- erty and the notion of paradoxical actions. Let A be a C∗-algebra, and let a, b be positive elements in A. We say that a is Cuntz below b, denoted a - b, if there exists a sequence (rn) in A such that r∗ nbrn → a. More generally for a ∈ Mn(A)+ and b ∈ Mm(A)+ we write a - b if there exists a sequence (rn) in Mm,n(A) with r∗ nbrn → a. For a ∈ Mn(A) and b ∈ Mm(A) let a ⊕ b denote the element diag(a, b) ∈ Mn+m(A). A nonzero positive element a in A is properly infinite if a ⊕ a - a. A C∗-algebra A is purely infinite if there are no characters on A and if for every pair of positive elements a, b in A such that b belongs to the ideal in A generated by a, one has b - a. Equivalently, a C∗-algebra A is purely infinite if every non-zero positive element a in A is properly infinite, cf. [29, Theorem 4.16]. Definition 4.1. A C∗-algebra A has the ideal property if projections in A separate ideals in A, i.e. , whenever I, J are ideals in A such that I * J , then there is a projection in I \ (I ∩ J ). The next result generalises the work in [39] on ordinary crossed products. The proof is included for completeness. Theorem 4.2. Let (A, G, α) be a partial dynamical system. Suppose that the action is exact and that for every t 6= e, every G-invariant ideal I in A, every a ∈ Dt/(Dt ∩ I), and every nonzero hereditary C∗-algebra B in A/I inf{kx(aδt)xk : x ∈ B+, kxk = 1} = 0. Suppose also that A has the ideal property. Then the following statements are equivalent (i) Every nonzero positive element in A is properly infinite in A ⋊α,r G. (ii) The C∗-algebra A ⋊α,r G is purely infinite. (iii) Every nonzero hereditary sub-C∗-algebra in any quotient of A ⋊α,r G contains an infinite projection. Proof. The implications (iii) ⇒ (ii) ⇒ (i) are valid for any partial dynami- cal system, cf. [29]. For (i) ⇒ (iii) let J be any ideal in A ⋊α,r G and let B be any nonzero hereditary sub-C∗-algebra in the quotient (A⋊α,r G)/J . We show that B contains an infinite projection. By assumption on α and Corol- lary 3.13 we have that (A ⋊α,r G)/J ∼= (A/I) ⋊α,r G for I = J ∩ A. Select a nonzero positive element b in B such that kEA/I (b)k = 1. By Lemma 3.11 there exist a positive contraction x ∈ A/I satisfying kxEA/I (b)x − xbxk < 1/4, kxEA/I (b)xk > kEA/I (b)k − 1/4 = 3/4. With a = (xEA/I (b)x − 1/2)+ we claim that 0 6= a - xbx - b. Indeed, the element a is nonzero because kxEA/I (b)xk > 1/2, and a - xbx holds since kxEA/I (b)x − xbxk < 1/2, cf. [38, Proposition 2.2]. By the assumption that A has the ideal property we can find a projection q ∈ A that belongs to the 14 THIERRY GIORDANO AND ADAM SIERAKOWSKI ideal in A generated by the preimage of a in A but not to I. Then q + I belongs to the ideal in A/I generated by a, whence q+I - a - b in A/I ⋊r G (because a is properly infinite by assumption), cf. [29, Proposition 3.5 (ii)]. From the comment after [29, Proposition 2.6] we can find z ∈ A/I ⋊r G such that q + I = z∗bz. With v = b1/2z it follows that v∗v = q + I, whence p := vv∗ = b1/2zz∗b1/2 is a projection in B, which is equivalent to q + I. By the assumption q is properly infinite, and hence so is q + I (since the relation a ⊗ a - a passes to quotients) and p. (cid:3) We now introduce a geometrical condition sufficient for the pure infinite- ness of a large class of partial crossed product C∗-algebras. It is not clear if this geometrical condition is also a necessary one. Definition 4.3. Let (C0(X), G, α) be a partial dynamical system, together with the corresponding collection of homeomorphisms (θt : Xt−1 → Xt)t∈G, inducing α, cf. Appendix A.2. Let E denote a family of subsets of X. A non-empty set V ⊆ X is called (G, E)-paradoxical if there exist sets V1, V2 . . . , Vn+m ∈ E and elements t1, t2, . . . , tn+m ∈ G such that n n+m Vi = [i=1 [i=n+1 Vi = V, Vi ⊆ Xt−1 i , θti(Vi) ⊆ V, θtk (Vk)∩θtl(Vl) = ∅, k 6= l. We let τX denote the topology of X. Theorem 4.4. Let (C0(X), G, α) be a partial dynamical system. Suppose that the action is exact and residually topologically free and that X is totally disconnected. Suppose also that every compact and open subset of X is (G, τX )-paradoxical. Then C0(X) ⋊α,r G is purely infinite. Proof. The proof consist of two parts. First we show that for any (G, τX )- paradoxical, compact and open subset U ⊆ X the projection 1U is properly infinite in C0(X) ⋊α,r G. We then prove that proper infiniteness of such projections is enough to ensure pure infiniteness of C0(X) ⋊α,r G. We do not know if the second part follows from Theorem 4.2. (i) Proper infiniteness: Recall that α is induced by a collection of open sets (Xt)t∈G and a collection (θt)t∈G of homeomorphisms θt : Xt−1 → Xt such that Xe = X and θst extends θs ◦ θt, cf. [35, 16, 37] and Appendix A.2. The partial action α of G on C0(X) corresponding to θ is given by αt(f )(x) := f (θt−1(x)), f ∈ C0(Xt−1 ). So, here the ideals are Dt = C0(Xt). Let (Vi, ti)n+m i=1 denote the system of open sets of U and elements in G wit- i=1 and (hi)n+m nessing the paradoxality of U . Find partitions of unity (hi)n i=n+1 for U relative to the open covers (Vi)n i=n+1, respectively. For each i = 1, . . . , n + m we have that the (compact) support of hi lies in Vi ⊆ Xt−1 . Hence hi ∈ Dt−1 ), for i = 1, . . . , n + m, and the elements and we can define ai := αti(h1/2 i=1 and (Vi)n+m i i i n n+m x = Xi=1 aiδti, y = Xi=n+1 aiδti, p = 1U . PURELY INFINITE PARTIAL CROSSED PRODUCTS 15 Using that (atδt)(asδs) = αt(αt−1(at)as)δts and (atδt)∗ = αt−1 (a∗ follows that t )δt−1 it i (a∗ (aiδti)∗(ajδtj ) = (αt−1 i )δt−1 (αti (αt−1 = αt−1 (a∗ i aj)δt−1 = αt−1 tj = (cid:26) hi if i = j if i 6= j. 0 i i i i i )(ajδtj ) (a∗ i ))aj )δt−1 i gj We obtain that x∗x = y∗y = p and y∗x = 0. Moreover, since 1U (aiδti) = αe(αe(1U )ai)δeti = (1U ai)δti = aiδti, we have that px = x, py = y. This implies that xx∗ + yy∗ ≤ p, hence p is properly infinite, cf. [29]. (ii) Pure infiniteness: This part follows by inspection of the proof of Theorem 4.2. The difference is that we do not assume that every nonzero element a ∈ A/I is properly infinite (in A/I ⋊α,r G) and we can therefore not conclude that the selected nonzero projection q + I in the ideal in A/I generated by a fulfils that q + I - a (in A/I ⋊r G) using [29, Proposition 3.5 (ii)]. Instead we find q ∈ A fulfilling 0 6= q + I - a as follows: Since X is totally disconnected we get that A and A/I have real rank zero and every projection in A/I lifts to a projection in A, cf. [6, Proposition 1.1, Theorem 3.14]. We can therefore select any nonzero projections in the hereditary sub- C∗-algebra of A/I generated by a and lift it to a projections q ∈ A. Since 0 6= q + I is contained in the hereditary sub-C∗-algebra of A/I, q + I - a, [29, Proposition 2.7]. (cid:3) 5. Examples Example 5.1 (The Cuntz algebra On). The Cuntz algebra, denoted by On, n ≥ 2 is the universal C∗-algebra generated by isometries s1, s2, . . . , sn sub- ject to the relation n sis∗ i = 1 Xi=1 Let Qn be the group { p nk : p ∈ Z, k ∈ N∪{0}} of n-adic rationals. We will denote by Qn×δZ the semidirect product: Qn×δZ = {(r, k) : r ∈ Qn, k ∈ Z}. The two group operations are given by: (s, j)(r, k) = ( r nj + s, j + k) and (r, k)−1 = (−nkr, −k). Hopenwasser constructs in [26] a partial action α of Qn on the Cantor set X such that C(X) ⋊α,r G ∼= On. Proposition 5.2. Let X be the Cantor set and let (C(X), Qn ×δ Z, α) be the partial dynamical system described in [26]. Then every clopen subset of X is (Qn ×δ Z, τX)-paradoxical. 16 THIERRY GIORDANO AND ADAM SIERAKOWSKI Proof. Let X denote the Cantor set based on [0, 1] where each n-adic rational r (exempt 0 and 1) is replaced by a pair r−, r+, such that r− is the immediate predecessor of r+ (i.e. there are no non-trivial elements x in X such that r− ≤ x ≤ r+). To simplify notation we will identify 0+ with 0 and 1− with 1. nk )+, ( q Let U be any nonempty clopen subset of X. Recall that sets of the form [( p nl )−] (with p, q ∈ N∪{0}, and k, l ∈ N) form a basis for the topology on X consisting of compact and open sets. It follows that U is a finite disjoint union of such sets (because we can cover U by finitely many such subsets of U and then remove the parts where they intersect). Since a disjoint union of compact open (Qn ×δ Z, τX)-paradoxical sets is again (Qn ×δ Z, τX)-paradoxical we can assume that U = [( p nl )−] for some p, q ∈ N∪{0}, and k, l ∈ N. Rewriting the fractions allows us to assume l = k. Since U is nonempty p < q. Suppressing the index ± we have that U is a disjoint union of sets [ p nk ]. We can therefore assume that q = p + 1. If U ⊆ Xs−1 for some s ∈ Qn ×δ Z it follows that U is (Qn ×δ Z, τX)-paradoxical if, and only if, θs(U ) ⊆ Xs is (Qn ×δ Z, τX)- paradoxical. Using the relations nk ], . . . , [ q−1 nk )+, ( q nk , p+1 nk , q X (− 1 nk ,0) −1 = [ θ(− 1 1 nk , 1], 1 nk(cid:3) p + 1 p nk ,0)(cid:0)(cid:2) nk , θ(0,−k)(cid:0)(cid:2)0, nk (cid:3)(cid:1) = (cid:2) nk(cid:3)(cid:1) = X. 1 p − 1 nk , p nk(cid:3) (if k 6= 0), X(0,−k)−1 = (cid:2)0, we can assume that U = X. It is evident that X is (Qn×δZ, τX)-paradoxical. In fact just like the crossed product associated to the action of the Baumslang- Solitar group on R, cf. [30], we also have here an action that mimics both translation and scaling. By first shrinking two copies of X and then translat- ing one of subsets away from the other, we naturally obtain the paradoxical property of X. (cid:3) Example 5.3 (The Cuntz-Krieger algebra OA). Let A = [aij] be a {0, 1}- valued (n × n)-matrix with no zero rows. We define the algebra OA to be the universal C∗-algebra generated by n partial isometries {si}n i=1 satisfying that sjs∗ j = 1, Xj and Xj aijsjs∗ j = s∗ i si for i = 1, 2, . . . , n. We have chosen to define OA as a universal object, as in [2], because it allows us to identify it with a partial crossed product. If the matrix A satisfies Condition (I) of Cuntz and Krieger, which implies the uniqueness of the C ∗-algebra C ∗({si}) provided that si 6= 0 for every i, we obtain the well known Cuntz-Krieger algebra introduced in [7]. Proposition 5.4. Let X be a compact totally disconnected space and G = Fn be the free group on n generators {g1, g2, . . . , gn}. Let (C(X), G, α) denote the partial dynamical system described by [19] such that C(X) ⋊α,r G ∼= OA. PURELY INFINITE PARTIAL CROSSED PRODUCTS 17 If A is such that the action is exact and residually topologically free, then the following statements are equivalent (i) Every compact and open subset of X is (G, τX )-paradoxical (ii) The C∗-algebra C(X) ⋊α,r G is purely infinite. Proof. The implication (i) ⇒ (ii) follows form Theorem 4.4. For the con- verse implication (ii) ⇒ (i) we need to recall the construction of the par- tial crossed product. An infinite admissible path is an infinite sequence µ = µ1µ2 . . . of generators of G such that A(µj, µj+1) = 1 for every j ∈ N (where we identify A(gi, gj ) with A(i, j)). Let X be the path space of in- finite admissible paths with the relative topology inherited as a closed and i=1{g1, . . . , gn}. Let g±1 t (= k) denote the length of a reduced word t = g±1 in G. An i2 i1 action θ of G on X is called semisaturated if θts = θt ◦ θs for every t, s ∈ G with ts = t + s (i.e. when there is no reduction in the concatenation of the reduced words t and s). By [19]1 we have that C(X) ⋊α,r G ∼= OA, where the partial action is the unique semisaturated partial action of G on X such that hence compact subspace of the infinite product space Q∞ · · · g±1 ik Xg−1 i = {µ ∈ X : A(gi, µ1) = 1} = domain( θgi ), θgi = µ 7→ giµ, i . where giµ means concatenation of gi at the beginning of µ. Using the iso- morphism between OA and C(X) ⋊α,r G, we have that si = 1Xgi δgi, where Xgi = θgi(Xg−1 ), cf. [19, Theorem 7.4] and [17, Theorem 6.5]. In particular sis∗ i = 1Xgi Let U be any compact and open subset of X. Recall that the cylinder sets in X (i.e. the sets consisting of infinite admissible paths with the same finite initial sequence) form a basis for the topology on X. In particular U is a finite disjoint union of cylinder sets. Since a disjoint union of compact open (G, τX )-paradoxical sets is again (G, τX )-paradoxical we can assume that U is a cylinder set. If U ⊆ Xs−1 for some s ∈ G it follows that U is (G, τX )-paradoxical if, and only if, θs(U ) ⊆ Xs is (G, τX )-paradoxical. We can therefore translate U until it has the form U = Xgi for some i = 1, . . . , n. We recall the definition of the C ∗-algebra corresponding to a directed graph, cf. [22]. Let E = (E0, E1, r, s) be a directed graph with count- ably many vertices E0 and edges E1, and range and source functions r, s : E1 → E0, respectively. The C ∗-algebra C ∗(E) is the universal C ∗-algebra generated by families of projections {pv : v ∈ E0} and partial isometries {se : e ∈ E1}, subject to the following relations: (i) pvpw = 0 for v, w ∈ E0, v 6= w. esf = 0 for e, f ∈ E1, e 6= f . (ii) s∗ ese = pr(e) for e ∈ E1. (iii) s∗ e ≤ ps(e) for e ∈ E1. (iv) ses∗ 1The version on arxiv.org have an additional section on this topic. See also [16, p. 55]. 18 THIERRY GIORDANO AND ADAM SIERAKOWSKI (v) pv = X{e∈E1: s(e)=v} ses∗ e for v ∈ E0 such that 0 < s−1(v) < ∞. Let E be the graph corresponding to the matrix A, where the genera- tors for G are the vertices and where we draw an edge from gi to gj when A(i, j) = 1. It follows form [32, Proposition 4.1] that C ∗(E) ∼= OA. Us- ing the isomorphism to identify elements in OA and C ∗(E) we have that si = P{e∈E1 : s(e)=gi} se and se = ss(e)sr(e)s∗ r(e), cf. [34]. This gives us that 1U = 1Xgi = sis∗ i = X{e∈E1 : s(e)=gi} ses∗ e = pgi. Since C ∗(E) is purely infinite, then 1U is properly infinite (since every nonzero positive element in a purely infinite C∗-algebra is properly infinite), but we need a bit more work to obtain the (G, τX )-paradoxical property of U . We will now argue that U = Xgi is a disjoint union of sets of the form Xµ1µ2...µk+1, where there is a loop based at the last vertex µk+1. By a loop we mean a sequence ν1ν2 . . . νl+1, l ≥ 1 of elements in {g1, . . . , gn} such that A(νi, νi+1) = 1 and νl+1 = ν1). If there is a loop based at gi we are done (k = 0). Otherwise we consider all gj in {g1, . . . , gn} such that A(i, j) = 1. As A has no zero rows we have the union Xgi = S{j : A(i,j)=1} Xgigj is non- empty. We now look at each gj to see if there is a loop based at gj. If there is a loop based at gj we keep Xgigj as it is. If not, we consider all gk ∈ {g1, . . . , gn} such that A(j, k) = 1 and rewrite Xgigj as S{k : A(j,k)=1} Xgigjgk . We now look at each gk to see if there is a loop based at gk. If there is a If not, we rewrite Xgigjgk as loop based at gk we keep Xgigjgk as it is. S{l : A(k,l)=1} Xgigjgkgl. We continue this process until U is rewritten into the desired form. This process is finite because we pick a new element in {g1, . . . , gn} every time we do a rewriting (if a vertex appears a second time when doing a rewriting then there must be a loop at that vertex, hence we never started the rewriting at that particular vertex to begin with). Knowing that U is a disjoint union of sets of the form Xµ1µ2...µk+1, where there is a loop based the last vertex µk+1 (and that disjoint union of para- doxical sets is paradoxical), we can assume U = Xµ1µ2...µk+1, with a loop based as at µk+1. With t = (µ1µ2 . . . µk)−1 we have that k (θµ−1 (Xµkµk+1) = θµ−1 Xµk+1 = θµ−1 = θµ−1 = θ(µk−2µk−1µk)−1(Xµk−2µk−1µkµk+1) = θt(Xµ1...µkµk+1) = θt(U ). k µ−1 k−1 k k−1 (Xµk−1µkµk+1)) (Xµk−1µkµk+1) = θ(µk−1µk)−1(Xµk−1µkµk+1) In particular we can assume that U = Xgi, gi ∈ {g1, . . . , gn} with a loop based at gi. As C ∗(E) is purely infinite then cf. [25, Theorem 2.3], the graph E satisfies condition (K) (i.e., no vertex v ∈ E0 lies on a loop, or there are two loops β′, β′′ based at v such that neither β′ nor β′′ is an initial PURELY INFINITE PARTIAL CROSSED PRODUCTS 19 subpart of the other, cf. [32]). In particular there are two distinct finite loops β′, β′′ based at gi. This implies that Xβ′ ∪ Xβ′′ ⊆ Xgi, where the union is disjoint, and where Xβ denotes the set of infinite admissible path starting with the finite sequence β. If we follow the two loops β′ and β′′ we come back to gi, implying the existence of t1, t2 ∈ G such that θt1(Xβ′) = Xgi and θt2(Xβ′) = Xgi. We conclude that U = Xgi is (G, τX )-paradoxical. (cid:3) Example 5.5 (C∗-algebras of integral domains). Let R be an integral domain with the property that the quotient R/(m) is finite, for all m 6= 0 in R. Set R× := R\{0}. Following Boava and Exel [5] we define the regular C∗-algebra A[R] of R as the universal C∗-algebra generated by isometries {sm : m ∈ R×} and unitaries {un : n ∈ R} subject to the relations smsm′ = smm′, unun′ = un+n′ , smun = umnsm, Xl+(m)∈R/(m) ulsms∗ mu−l = 1, for m, m′ ∈ R× and n, n′ ∈ R. Following Boava and Exel [5], let K denote the field of fractions of R, and K × the set K\{0}. Let G be the semidirect product K ⋊ K × = {(u, w) : u ∈ K, w ∈ K ×} equipped with the following operations (u, w)(u′, w′) = (u + u′w, ww′) and (u, w)−1 = (−u/w, 1/w). As in [5] we define a partial order on K × given by w ≤ w′ if w′ = wr for some r ∈ R. Let (w) denote the ideal wR ⊆ K. Let X be the space of all sequences (uw + (w))w∈K × in Qw∈K ×(R + (w))/(w) fulfilling that uw′ + (w) = uw + (w) if w ≤ w′. Then in [5], Boava and Exel prove (see Appendix A.3 for a more detailed description) that where the partial action θ of G on X is defined by C(X) ⋊α,r G ∼= A[R] , X(u,w) = {(uw′ + (w′))w′ ∈ X : uw + (w) = u + (w)}, θ(u,w) = (uw′ + (w′))w′ 7→ (u + wuw−1w′ + (w′))w′. Note that the sets (Xt)t∈G are compact, open and form a basis for the topology on X, cf. [5]. Proposition 5.6. If (C(X), G, α) denotes the partial dynamical system de- scribed in [5] and recalled above, then every compact and open subset of X is (G, τX )-paradoxical. In particular OZ ∼= A[Z] is purely infinite. Proof. Let U be any compact and open subset of X. For w ∈ K × and Cw ⊆ (R + (w))/(w) set w := {(uw′ + (w′))w′ ∈ X : uw + (w) ∈ Cw}. V Cw Boava and Exel showed in [5] that the family of sets (V Cw w )w∈K × is closed under complement, intersections and finite unions. It follows that U = V Cw for some w ∈ K × and Cw ⊆ (R+(w))/(w). Since a disjoint union of compact open (G, τX )-paradoxical sets is again (G, τX )-paradoxical we can assume that Cw contains precisely one element, i.e. U = Xt for some t = (u, w) ∈ G. w 20 THIERRY GIORDANO AND ADAM SIERAKOWSKI Since X(u,w) = ∅ ⇔ u /∈ R + (w), cf. [5, Proposition 4.4], we can assume that u ∈ R + (w). Using that X(u,w) = X(u+w,w) we can assume u ∈ R. Since Cw 6= ∅ it follows from an argument prior [5, Proposition 4.5] that , with Cwr containing at last two elements for some2 r in R. In w = V Cwr V Cw wr particular U = [s+(rw)∈Crw X(s,rw), Crw > 1. Using the relations (with s ∈ R×) X(0,s)−1 = X X(s,1)−1 = X θ(0,s)(X(u,w)) = X(su,sw), θ(s,1)(X(u,w)) = X(s+u,w). it follows that there exist a finite number of elements t1, . . . , tn in G and a finite number of open, pairwise disjoint subsets U1, . . . , Un of U such that U = [i∈{1,...,n} Un, n > 1, θtj (U ) = Uj, j ∈ {1, . . . , n}. This shows that U is (G, τX )-paradoxical. Since G is solvable both the group G and the action α is exact. It was shown in [5, Proposition 4.5] that the action is residually topologically free. By Theorem 4.4 OZ ∼= A[Z] is purely infinite. (cid:3) Corollary 5.7. Let (C(X), G, α) be the partial dynamical system as above with R = Z and G = Q ⋊ Q×. Then every compact and open subset of X is +. In particular ON ∼= C(X) ⋊α,r H is (H, τX)-paradoxical, with H = Q ⋊ Q× purely infinite. 6. Connection to crossed products Abadie showed in [1] that certain partial crossed products are Morita- Rieffel equivalent to ordinary crossed products. Since pure infiniteness is preserved under stable isomorphism (see [29]) one might also expect that the mentioned Morita-Rieffel equivalence maps paradoxical sets in the realm of partial actions into paradoxical sets in the realm of ordinary action. This is precisely the case. Let us first recall the result of Abadie in [1]. Definition 6.1. Let (C0(X), G, α) be a partial dynamical system, together with the corresponding collection of homeomorphisms (θt : Xt−1 → Xt)t∈G, inducing α. The envelope space, denoted by X e, is the topological quotient space (G×X) ∼ , where ∼ is the equivalence relation given by (r, x) ∼ (s, y) ⇔ x ∈ Xr−1s and θs−1r(x) = y. The envelope action, denoted by he, is the action induced in X e by the action he s(t, x) 7→ (st, x). 2In [5] it is stated that any non-invertible r in R will work, which is not true. PURELY INFINITE PARTIAL CROSSED PRODUCTS 21 Theorem 6.2 (Abadie [1]). Let (C0(X), G, α) be a partial dynamical system such that X e is Hausdorff. Let αe denote the action of G on C0(X e) induced by the envelope action. Then C0(X) ⋊α,r G is Morita-Rieffel equivalent to C0(Xe) ⋊αe,r G. We then have: Theorem 6.3. Let (C0(X), G, α) be a partial dynamical system such that X e Hausdorff. If every compact open subset of X is (G, τX )-paradoxical, then every compact open subset of X e is (G, τX e)-paradoxical. Proof. Let U e be a compact open subset of X e. Find a cover of U e consisting of open subsets U e of the form {t}×Ut ∼ , with Ut ∈ τX and t ∈ G. By compactness we can assume the cover is finite. Moreover, we can assume that (i) The union is disjoint: Let ρ denote the canonical surjection G × X 7→ X e, and let X c t , t ∈ G, denote the complement of Xt in X. Fix t ∈ G. For each s ∈ G and x ∈ Xt−1s we have that (t, x) ∼ (s, θs−1t(x)). In particular ∼ ) = Ss∈G{s} × θs−1t(Xt−1s) is open in X, implying that {t}×X ρ−1( {t}×X is open in X e. Since X e is Hausdorff the sets (Xt)t∈G are clopen in X, s−1t is open in X, and {t}×X se can, by compactness, ∼ )t∈F . By passing ∼ of t ∩ U e we t , where the union is disjoint. For each f ∈ F define t } ∈ τX, where ρt : X → X e denotes the (continues) cf. [12, Proposition 3.1]. Hence Ss∈G{s} × X c is therefore clopen in X e. Since X e = St∈G cover U e by finitely many clopen set of the form ( {t}×X to subsets of the sets ( {t}×X clopen subsets of Xe fulfilling that V e ∼ )t∈F we obtain a partition (V e ∼ . With U e t have that U e = St∈F U e Ut := {x ∈ X : ρt(x) ∈ U e composition of x 7→ (t, x) and ρ. It follows easily that U e t )t∈F of F ×X t ⊆ {t}×X := V e {t}×X ∼ ∼ ∼ t = {t}×Ut ∼ . that Ut ⊆ Sm of open subsets of X. Notice that each V e i i ∩ V e t consisting of open subsets of X e and, together with (V e (ii) Each set Ut is compact: Fix t ∈ F . Let (Vi)i be a cover of Ut consisting is open in X e (since t )i is a cover of U e s )s∈F \{t}, a cover of U e. By compactness we can assume i = 1, . . . , m for some m ∈ N. Intersecting with V e i . It follows Ss∈G{s} × θs−1t(Vi ∩ Xt−1s) is open in X). The sets (V e t ⊆ Sm i=1 Vi. Hence Ut is compact. t we have that U e t ⊆ Sm := {t}×Vi i ∩ V e i=1 V e i=1 V e ∼ ∼ ∩ {e}×Vl (iii) U e is paradoxical: Each set of the form {e}×U ∼ , with U compact and open in X, is (G, τX e)-paradoxical: If (Vi, ti)n+m i=1 witness the paradoxicality of U and k 6= l are any natural numbers in 1, . . . , n + m if follows that {e}×Vk ∼ , ti)n+m i=1 witness the paradoxicality of {e}×U ∼ . A translate of a (G, τX e)-paradoxical set is again (G, τX e )-paradoxical. Hence each {t}×Ut is (G, τX e )-paradoxical. A finite disjoint union of (G, τX e )-paradoxical set is again (G, τX e )-paradoxical. We conclude that U e is (G, τX e )-paradoxical. (cid:3) ∼ = ∅ because Vk ∩ Vl = ∅. We obtain that ( {e}×Vi ∼ 22 THIERRY GIORDANO AND ADAM SIERAKOWSKI Remark 6.4. In a very interesting recent preprint [28], Kellerhals, Monod, and Rørdam proved that a countable group is non-supramenable if and only if it admits a free, minimal, purely infinite action on the locally compact non-compact Cantor set. Then they use this characterisation to associate to such dynamical systems stable Kirchberg C-algebras in the UCT class. Based on the examples presented in Section 5, where the opens sets (Xt)t∈G we used to construct the partial crossed products were clopen (en- suring X e is Hausdorff, see [12, Proposition 3.1]), and Proposition 6.3, we obtain another class of examples of dynamical systems whose associated crossed products are stable Kirchberg C-algebras. Appendix A. A.1. Basic definitions. Recall that a Fell bundle over a discrete group G is a collection B = (Bt)t∈G of closed subspaces of a C∗-algebra B, indexed by a discrete group G, satisfying B∗ t = Bt and BtBs ⊆ Bts for all t and s in G (cf. [16]). A section of B is a function ξ : G → St∈G Bt with the property that ξ(t) ∈ Bt for all t ∈ Bt. Let l1(B) denote the Banach ∗-algebra (cf. [35, Proposition 2.1]) consisting of all sections ξ of B with a finite l1-norm. The operations and the norm in l1(B) are given by ξη(t) = Xs∈G ξ(s)η(s−1t), ξ∗(t) = ξ(t−1)∗, kξk1 = Xt∈G kξ(t)k. We define the full cross sectional algebra of B, denoted C∗(B), to be the enveloping C∗-algebra of l1(B) (cf. [21, Section VIII.17.2]). Recall that a right Hilbert A-module X is a right A-module X equipped with a map h·, ·iA : X × X → A that is linear in the second component and for x, y ∈ X, a ∈ A, (i) hx, xiA ≥ 0 with equality only if x = 0; (ii) hx, y · aiA = hx, yiAa; (iii) hx, yiA = hy, xi∗ A; and (iv) X is complete in the norm defined by kxk2 As in [27, Section 1.1.7], let l2(B) denote the right Hilbert Be-module con- A = khx, xiAk. converges unconditionally (i.e. for any ε > 0 there exist a finite set F ⊆ G such that for every finite subsets H ⊆ G containing F one has that sisting of all cross sections ξ of B fulfilling that the series x = Pt∈G ξ(t)∗ξ(t) kx −Pt∈H ξ(t)∗ξ(t)k < ε). We equip l2(B) with the inner product hξ, ηi = Xt∈G ξ(t)∗η(t), ξ, η ∈ l2(B). Let L(l2(B)) denote the C∗-algebra of all adjointable operators on l2(B) (i.e. linear maps T : l2(B) → l2(B) with a linear map T ∗ such that hξ, T ηi = PURELY INFINITE PARTIAL CROSSED PRODUCTS 23 hT ∗ξ, ηi, for all ξ, η ∈ l2(B)). With the ∗-homomorphism ΛB : l1(B) → L(l2(B)) defined (in [20, Proposition 2.6]) by (ΛB(ξ)η)(t) = Xs∈G ξ(s)η(s−1t), ξ ∈ l1(B), η ∈ l2(B), t ∈ G, we define the reduced cross sectional algebra of B, denoted C∗ r(B), to be the sub-C∗-algebra of L(l2(B)) generated by the range of ΛB. Let δs,t, s, t ∈ G denote the Kronecker symbol. Each Bt is a right Hilbert Be-module with the inner product hb, ci = b∗c. Let jt ∈ L(Bt, l2(B)) denote the adjointable operator defined by (jt(bt))(s) = δs,tbt, for bt ∈ Bt and s ∈ G. The adjoint j∗ t ∈ L(l2(B), Bt) is simply the map j∗ t (ξ) = ξ(t), ξ ∈ l2(B). Recall, cf. [16, r(B) and t ∈ G the tth Fourier coefficient of x is Definition 2.7], that for x ∈ C∗ the unique element x(t) ∈ Bt such that (j∗ t ◦ x ◦ je)(a) = x(t)a for all a ∈ Be. The map E : C∗ r(B) → Be, given by x 7→ x(e) is a positive, contractive, conditional expectation. Moreover, E is faithful on positive elements since r(B) (cf. [16, Proposition 2.12]). Let Cc(B) ⊆ l1(B) denote the set of finitely supported sections of B. E(x∗x) = Pt∈G x(t)∗ x(t) (unconditional convergence) for each x ∈ C∗ Let A be a C∗-algebra and G be a discrete group. Recall (see [9]) that a partial action of G on A, denoted by α, is a collection (Dt)t∈G of closed two- sided ideals of A and a collection (αt)t∈G of ∗-isomorphisms αt : Dt−1 → Dt such that s (Ds ∩ Dt−1) ⊆ D(ts)−1 ; (i) De = A, where e represents the identity element of G; (ii) α−1 (iii) αt ◦ αs(x) = αts(x), ∀ x ∈ α−1 The triple (A, G, α) is called a partial dynamical system. Fix a partial dy- namical system (A, G, α). Recall the corresponding Fell bundle B as defined in Definition 2.2: We let L be the normed ∗-algebra of the finite formal sums s (Ds ∩ Dt−1). The family (Bt)t∈G generates the Fell bundle B. It follows that Cc(B) = L. Pt∈G atδt, where at ∈ Dt. We let Bt denote the vector subspace Dtδt of L. Moreover, for x = Pt∈G atδt ∈ Cc(B), we have that E(ΛB(x)) = ae. Lemma A.1 (McClanahan). Let (A, G, α) be a partial dynamical system. For any G-invariant ideal I of A we have a canonical embedding of I ⋊α,r G in A ⋊α,r G. Remark A.2. To best of our knowledge our proof of Lemma A.1 is new. For a different proof using covariant representation we refer to the work in [35]. Since our proof applies to general Fell bundles (and not only the one defining crossed products) we have included it for completeness. Proof. Fix an approximate unit (en) for I, and the Fell bundle E = (Et)t∈G, Et = (Dt ∩ I)δt, equipped with the operations and norm coming from B. For each section ξ ∈ l2(B) let ξn : G → St∈G Et denote the map t 7→ ξ(t)en contained in 24 THIERRY GIORDANO AND ADAM SIERAKOWSKI l2(E). Let ϕ : L(l2(E)) → L(l2(B)) denote the map ϕ(T )ξ = limn T (ξn). For x ∈ Cc(E), ϕ(ΛE(x)) = ΛB(x). We have that ϕ(ΛE(x)) = ΛB(x) and computation kϕ(ΛE(x))k = kΛE(x)k, for any element x = Pt∈G atδt in Cc(E), by direct kϕ(ΛE(x))k2 = kϕ(ΛE(x))ξk2 sup ξ∈Cc(B),kξk≤1 = = = sup ξ∈Cc(B),kξk=1 lim n k Xt,s,r∈G(cid:0)(atδt)ξn(t−1r)(cid:1)∗((asδs)ξn(s−1r))k sup η∈Cc(E),kηk≤1 sup η∈Cc(E),kηk≤1 k Xt,s,r∈G(cid:0)(atδt)η(t−1r)(cid:1)∗((asδs)η(s−1r))k kΛE(x)ηk2 = kΛE(x)k2 The first and last equality above use the fact that the finitely supported sections are dense in l2. The second and fourth equality follows from the definition of the maps ϕ and ΛE. For the third equality we obviously have ≥ after taking the limit. To get the ≤ notice that if we remove "supξ∈Cc(B),kξk≤1 limn" from the left hand side we have ≤ since ξn ∈ {η ∈ Cc(E), kηk ≤ 1} for any n and any ξ ∈ Cc(B) with kξk ≤ 1. The inequality ≤ remains valid when we take the limit and the supremum. I particular we ob- tain that the canonical embedding of Cc(E) into Cc(B), given by atδt 7→ atδt, extends to an embedding ι : I ⋊α,r G → A ⋊α,r G. (cid:3) Let I be a G-invariant ideal of A. In order to introduce the map ρ : A ⋊α,r G → A/I ⋊α,r G we can again turn to [16]. Let F, denote the Fell bundle (Ft)t∈G, Ft = (Dt/(I ∩ Dt))δt, equipped with the operations and norm coming from B. The canonical surjection of Cc(B) onto Cc(F), given by atδt 7→ (at + Dt ∩ I)δt, extends to a surjective ∗-homomorphism ρ : A ⋊α,r G → A/I ⋊α,r G (cf. [16, Proposition 3.11]). In particular, by continuity of ι, ρ, and E, we have the following result: Proposition A.3 (Exel, McClanahan). Let (A, G, α) be a partial dynamical system. For any G-invariant ideal I of A we have the commuting diagram 0 0 / I ⋊α,r G ι A ⋊α,r G ρ A/I ⋊α,r G EI / I EA / A EA/I / A/I 0 / 0 Using that EA is idempotent and identifying I ⋊α,r G with a subset of A ⋊α,r G we have the additional properties: / / / / / / / / / / / PURELY INFINITE PARTIAL CROSSED PRODUCTS 25 Proposition A.4 (Exel, McClanahan). Let (A, G, α) be a partial dynamical system. For any G-invariant ideal I of A and any ideal J of in A ⋊α,r G we have the following identities: (I ⋊α,r G) ∩ A = I, J ∩ A ⊆ EA(J ), Ideal[I] = I ⋊r G, where Ideal[S] denotes the smallest ideal in A ⋊α,r G generated by S ⊆ A ⋊α,r G. A.2. Topological freeness. Let (A, G, α) be a partial dynamical system with ideals (Dt)t∈G and ∗-isomorphisms (αt)t∈G. If we replace the C∗-algebra A by a locally compact Hausdorff space X, the ideals Dt by open sets Xt and the ∗-isomorphisms αt by homeomorphisms θt : Xt−1 → Xt, we obtain a partial action θ of G on the space X. A partial action of a group G on a space X induces naturally a partial action α of G on C0(X). The ideals are C0(Xt) and αt(f ) = f ◦ θt−1. The converse (still in the abelian case) is also true (cf. [1, 23]). Let Prim A denote the primitive ideal space of A with respect to the Jacobson topology. Let A denote the spectrum of A with respect to the topology induced by the surjection κ : A → Prim A, κ([π]) = ker π. Fol- lowing [33] recall how α defines a partial action of G on Prim A and on A: For any ideal J of A we let supp J denote the subset {x ∈ Prim A : J * x}. It is known (see [8, Section 3.2.1] or [40, Section 1.4]) that the mapping x 7→ x ∩ J establishes a homeomorphism supp J ↔ Prim J and supp J is an open set in Prim A. Set also AJ = {[π] ∈ A : π(J ) 6= 0}. Then the mapping [π] 7→ [πJ ] establishes a homeomorphism AJ ↔ J and AJ is an open set in A (see [8, Section 3.2.1] or [40, Section 1.4]). Let us define the mapping τt : ADt−1 → ADt in the following way: For any [π] ∈ ADt−1 we set τt([π]) = [π ◦ αt−1], t ∈ G. The foregoing observations tell us that τt is a homeomorphism. Let us also define the mapping θt : supp Dt−1 → supp Dt in the following way: For any point x ∈ supp Dt−1 such that x = ker π where [π] ∈ ADt−1 we set θt(x) = ker (π ◦ αt−1), t ∈ G. Clearly θt is a homeomorphism. For τt and θt defined in the above described way we have that (τt)t∈G defines a partial action of G on A and (θt)t∈G defines a partial action of G of Prim A. Recall that the action α is called topologically free if for any finite set {t1, . . . , tn}, ti 6= e the set n [i=1(cid:8)x ∈ supp Dt−1 i : θti(x) = x(cid:9) has empty interior. This condition can be also formulated in the follow- ing way: For any finite set {t1, ...tk} ⊆ G and any nonempty open set U there exists a point x ∈ U such that all the points θti(x), i = 1, . . . , k that 26 THIERRY GIORDANO AND ADAM SIERAKOWSKI are defined (⇔ x ∈ supp Dt−1 ) are distinct. Recall that the action α has the intersection property if every nonzero ideal in A ⋊α,r G has a nonzero intersection with A. i Theorem A.5 (Lebedev). Let (A, G, α) be a partial dynamical system. Suppose that the action is topologically free. Then α has the intersection property. Theorem A.6 (Exel, Laca, Quigg). Let (A, G, α) be a partial dynamical system with A abelian. Suppose that the action is topologically free. Then for every b ∈ A ⋊α,r G and every ε > 0 there exist a positive contraction x ∈ A satisfying kxEA(b)x − xbxk < ε, kxEA(b)xk > kEA(b)k − ε. Proposition A.7. Let (A, G, α) be a partial dynamical system with A abelian. Then the following properties are equivalent: (i) α is topologically free, (i′) Ft = {x ∈ Ut−1 : θt(x) = x} has empty interior (t 6= e), (ii) kαtK − idKk = 2 for every αt-invariant ideal K in Dt−1 (and t 6= e), (ii′) kαtK − idKk 6= 0 for every αt-invariant ideal K in Dt−1 (and t 6= e), (iii) inf{kxαt(x)k : x ∈ K+, kxk = 1} = 0 for every nonzero ideal K in Dt−1 (and t 6= e), (iv) inf{kx(btδt)xk : x ∈ B+, kxk = 1} = 0 for every bt ∈ Dt, and every nonzero hereditary C∗-algebra B in A (and t 6= e), (iv′) inf{kx(btδt)xk : x ∈ K+, kxk = 1} = 0 for every bt ∈ Dt, and every nonzero ideal K in A (t 6= e). Proof. The implications (i) ⇔ (i′) follows from [19, Definition 2.1, Lemma 2.2], and the implications (iv) ⇔ (iv′) and (ii) ⇒ (ii′) are evident. Pick any x ∈ C, where C := {y ∈ V : αt(f )(y) − f (y) > 0}. (ii′) ⇒ (ii): Find some function f ∈ K+ = C0(V )+ such that αt(f ) 6= f . Notice x 6= θt−1(x) in V . (cid:0)If f (x) 6= 0 then -- for h := αt−1(f ) ∈ K -- we have that h(θt−1 (x)) = f (x) 6= 0, recalling that αt(f )(x) = f (θt−1(x)). Hence θt−1(x) ∈ V . If f (x) = 0 then αt(f )(x) = f (θt−1(x)) 6= 0. Hence θt−1(x) ∈ V . If x = θt−1(x) then 0 = f (θt−1(x)) − f (x) = αt(f )(x) − h(x) = 2. Hence kαtK − idKk = 2. f (x) > 0.(cid:1) One can now easily find a function h ∈ K such that h(θt−1 (x))− (i′) ⇒ (iii): Assume not (iii). Find some t 6= e and K = C0(V ) in C0(Ut−1) such that inf{kxβ(x)k : x ∈ K+, kxk = 1} > 0, where β = αt. Suppose (i′). Then the Ut−1-open set F c t = {x ∈ Ut−1 : θt(x) 6= x} is dense in Ut−1. As V is Ut−1-open it has a non-empty intersection with F c t . Therefore we can find a non-empty Ut−1-open set Y in V such that θt(Y ) is disjoint from Y . Hence there is a function x in K+ of norm one such that xβ(x) = 0. Contradiction, hence not (i′). (iii) ⇒ (ii′): Assume not (ii′). Then there exists a nonzero αt-invariant ideal K = C0(V ) in C0(Ut−1 ) (for some t 6= e) such that kαtK − idKk = 0. PURELY INFINITE PARTIAL CROSSED PRODUCTS 27 We conclude that αt(x) = x for all x ∈ K. Hence inf{kxαt(x)k : x ∈ K+, kxk = 1} is nonzero implying not (iii). (ii′) ⇒ (i′): Assume not (i′). Find t 6= e and a Ut−1-open nonempty subset V in Ft (recall that Ft is Ut−1-closed but has nonempty interior). Since θt(Ut−1 ) = Ut we see that each x ∈ V ⊆ Ut−1 also belongs to Ut and θt−1(x) = x, x ∈ V . Hence K := C0(V ) is αt-invariant (i.e. αt(f ) = f = αt−1(f ), f ∈ K). We obtain that kαtK − idKk = 0. Hence not (ii′). (iii) ⇒ (iv′): Fix any t 6= e, any bt ∈ Dt, and any nonzero ideal K in A. Define I := K ∩ Dt−1. Suppose I = 0. Fix any x ∈ K. Using that xbt ∈ Dt we get αt−1(xbt) ∈ Dt−1 and αt−1(xbt)x = 0. Hence x(btδt)x = αt(αt−1 (xbt)x)δt = 0 and inf{kxbtδt)xk : x ∈ K+, kxk = 1} = 0. Suppose I 6= 0. Then I is an ideal in Dt−1 and kx(btδt)xk ≤ kαt(αt−1(xbt))αt(x)k ≤ kbtkkxαt(x)k, x ∈ I +. It follows from (iii) that inf{kx(btδt)xk : x ∈ K+, kxk = 1} = 0. (iv′) ⇒ (iii): Fix any t 6= e and nonzero ideal K in Dt−1 . Define I := Dt ∩ K. Suppose I = 0. Fix any x ∈ K. Then x ∈ Dt−1, αt(x) ∈ Dt, αt(x)x ∈ Dt ∩ K, and αt(x)x = 0. Hence inf{kxαt(x)k : x ∈ K+, kxk = 1k} = 0. Suppose I 6= 0. Then I := C0(V ) is a nonzero ideal in A. Pick any x ∈ V . Choose an open set U with compact closure such that x ∈ U ⊆ ¯U ⊆ V. With K := ¯U ⊆ V there exist by Urysohn's Lemma a function h ∈ C0(V ) such that 0 ≤ h ≤ 1, hK = 1. Using (iv) on bt := h ∈ Dt and the nonzero ideal J := C0(U ) in A (and the equality x(btδt)x = xbtαt(x)δt valid for every x ∈ J ⊆ I ⊆ K ⊆ Dt−1) we obtain inf{kxbtαt(x)k : x ∈ J +, kxk = 1k} = 0. Since hU = 1 we have that xbt = x for every x ∈ J+, hence inf{kxαt(x)k : x ∈ J +, kxk = 1k} = 0. Since J ⊆ I ⊆ K also inf{kxαt(x)k : x ∈ K+, kxk = 1k} = 0. (cid:3) 28 THIERRY GIORDANO AND ADAM SIERAKOWSKI A.3. C∗-algebras of an Integral Domain. Let R be an integral domain (i.e. a commutative unital ring without zero divisor). Set R× := R \ {0}. Following Boava and Exel [5] we impose that the quotient R/(m) is finite, for all m 6= 0 in R and let A[R] denote the regular C∗-algebra of R, i.e., the universal C∗-algebra generated by isometries {sm : m ∈ R×} and unitaries {un : n ∈ R} subject to the relations smsm′ = smm′, unun′ = un+n′ , smun = umnsm, Xl+(m)∈R/(m) ulsms∗ mu−l = 1, for every m, m′ ∈ R× and n, n′ ∈ R. As in [5] we let K denote the field of fractions of R and let K × denote the set K\{0}. To form a group one equip the semidirect product K ⋊ K × = {(u, w) : u ∈ K, w ∈ K ×} with the following two group operations (u, w)(u′, w′) = (u + u′w, ww′), (u, w)−1 = (−u/w, 1/w). Boava and Exel showed in [5] that the regular C∗-algebra of R is a partial crossed product by first showing that the algebra A[R] is ∗-isomorphic to a partial group algebra of K ⋊ K × (with certain relations R), and then use that every partial group algebra is a partial crossed product. We will not describe these isomorphisms but will instead focus on the the description of the partial crossed product. We have a partial order on K × given by w ≤ w′ if w′ = wr for some r ∈ R. For w ∈ K let (w) denote the ideal wR ⊆ K. For each pair w ≤ w′ in K × we have a canonical map pw,w′ : (R + (w′))/(w′) → (R + (w))/(w) given by pw,w′(uw′ + (w′)) = uw′ + (w). Boava and Exel proved that the inverse limit lim←−{(R + (w))/(w), pw,w′} is isomorphic to the space X of all sequences (uw + (w))w∈K × ∈ Yw∈K × (R + (w))/(w) fulfilling that uw′ + (w) = uw + (w) if w ≤ w′. When (R + (w))/(w) is given the discrete topology and (R + (w))/(w) the product topology X becomes a compact topological space. Moreover, cf. [5], there is a partial action θ of G on X defined by X(u,w) = {(uw′ + (w′))w′ ∈ X : uw + (w) = u + (w)}, θ(u,w) = (uw′ + (w′))w′ 7→ (u + wuw−1w′ + (w′))w′, for any (u, w) ∈ K ⋊ K ×. The partial actions θ induces a partial dynam- ical system (C(X), K ⋊ K ×, α), and hence also a partial crossed product C(X) ⋊α,r (K ⋊ K ×). Theorem A.8 (Boava-Exel). Let R be a integral domain with finite quo- tients R/(m), m ∈ R×. Suppose that R is not a filed. Then the maps un 7→ 1X δ(n,1), sm 7→ 1X(0,m) δ(0,m) PURELY INFINITE PARTIAL CROSSED PRODUCTS 29 induce a ∗-isomorphism between A[R] and C(X) ⋊α,r (K ⋊ K ×). References 1. Fernando Abadie, Enveloping actions and Takai duality for partial actions, J. Funct. Anal. 197 (2003), no. 1, 14 -- 67. MR 1957674 (2004c:46130) 2. Astrid an Huef and Iain Raeburn, The ideal structure of Cuntz-Krieger algebras, Er- godic Theory Dynam. Systems 17 (1997), no. 3, 611 -- 624. MR 1452183 (98k:46098) 3. R. J. Archbold and J. S. Spielberg, Topologically free actions and ideals in discrete C ∗-dynamical systems, Proc. Edinburgh Math. Soc. (2) 37 (1994), no. 1, 119 -- 124. MR 1258035 (94m:46101) 4. B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122, Springer-Verlag, Berlin, 2006, Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non-commutative Geometry, III. MR 2188261 (2006k:46082) 5. Giuliano Boava and Ruy Exel, Partial crossed product description of the C ∗-algebras associated with integral domains, 2012. 6. Lawrence G. Brown and Gert K. Pedersen, C ∗-algebras of real rank zero, J. Funct. Anal. 99 (1991), no. 1, 131 -- 149. MR MR1120918 (92m:46086) 7. Joachim Cuntz and Wolfgang Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980), no. 3, 251 -- 268. MR 561974 (82f:46073a) 8. Jacques Dixmier, Les C ∗-alg`ebres et leurs repr´esentations, Deuxi`eme ´edition. Cahiers Scientifiques, Fasc. XXIX, Gauthier-Villars ´Editeur, Paris, 1969. MR 0246136 (39 #7442) 9. M. Dokuchaev and R. Exel, Associativity of crossed products by partial actions, en- veloping actions and partial representations, Trans. Amer. Math. Soc. 357 (2005), no. 5, 1931 -- 1952. MR 2115083 (2005i:16066) 10. Siegfried Echterhoff and Marcelo Laca, The primitive ideal space of the C ∗-algebra of the affine semigroup of algebraic integers, Math. Proc. Cambridge Philos. Soc. 154 (2013), 119 -- 126. 11. George A. Elliott, Some simple C ∗-algebras constructed as crossed products with dis- crete outer automorphism groups, Publ. Res. Inst. Math. Sci. 16 (1980), no. 1, 299 -- 311. MR 574038 (81j:46102) 12. R. Exel, T. Giordano, and D. Gon¸calves, Enveloping algebras of partial actions as groupoid C ∗-algebras, J. Operator Theory 65 (2011), no. 1, 197 -- 210. MR 2765764 (2012f:46096) 13. Ruy Exel, The Bunce-Deddens algebras as crossed products by partial automorphisms, Bol. Soc. Brasil. Mat. (N.S.) 25 (1994), no. 2, 173 -- 179. MR 1306559 (95m:46091) 14. Ruy Exel, Circle actions on C ∗-algebras, partial automorphisms, and a generalized Pimsner-Voiculescu exact sequence, J. Funct. Anal. 122 (1994), no. 2, 361 -- 401. MR 1276163 (95g:46122) 15. 16. 17. 18. , Approximately finite C ∗-algebras and partial automorphisms, Math. Scand. 77 (1995), no. 2, 281 -- 288. MR 1379271 (97e:46085) , Amenability for Fell bundles, J. Reine Angew. Math. 492 (1997), 41 -- 73. MR 1488064 (99a:46131) , Partial actions of groups and actions of inverse semigroups, Proc. Amer. Math. Soc. 126 (1998), no. 12, 3481 -- 3494. MR 1469405 (99b:46102) , Exact groups and Fell bundles, Math. Ann. 323 (2002), no. 2, 259 -- 266. MR 1913042 (2003e:46095) 19. Ruy Exel, Marcelo Laca, and John Quigg, Partial dynamical systems and C ∗-algebras generated by partial isometries, J. Operator Theory 47 (2002), no. 1, 169 -- 186. MR 1905819 (2003f:46108) 20. Ruy Exel and Chi-Keung Ng, Approximation property of C ∗-algebraic bundles, Math. Proc. Cambridge Philos. Soc. 132 (2002), no. 3, 509 -- 522. MR 1891686 (2002k:46189) 30 THIERRY GIORDANO AND ADAM SIERAKOWSKI 21. J. M. G. Fell and R. S. Doran, Representations of ∗-algebras, locally compact groups, and Banach ∗-algebraic bundles. Vol. 2, Pure and Applied Mathematics, vol. 126, Academic Press Inc., Boston, MA, 1988, Banach ∗-algebraic bundles, induced repre- sentations, and the generalized Mackey analysis. MR 936629 (90c:46002) 22. Neal J. Fowler, Marcelo Laca, and Iain Raeburn, The C ∗-algebras of infinite graphs, Proc. Amer. Math. Soc. 128 (2000), no. 8, 2319 -- 2327. MR 1670363 (2000k:46079) 23. Daniel Gon¸calves, Produtos cruzados, 2001, Tese de mestrado, Departamento de Matematica, UFSC. 24. Philip Green, The local structure of twisted covariance algebras, Acta Math. 140 (1978), no. 3-4, 191 -- 250. MR 0493349 (58 #12376) 25. Jeong Hee Hong and Wojciech Szyma´nski, Purely infinite Cuntz-Krieger algebras of directed graphs, Bull. London Math. Soc. 35 (2003), no. 5, 689 -- 696. MR 1989499 (2005c:46097) 26. Alan Hopenwasser, Partial crossed product presentations for On and Mk(On) us- ing amenable groups, Houston J. Math. 33 (2007), no. 3, 861 -- 876. MR 2335740 (2008k:46161) 27. Kjeld Knudsen Jensen and Klaus Thomsen, Elements of KK-theory, Mathematics: Theory & Applications, Birkhauser Boston Inc., Boston, MA, 1991. MR 1124848 (94b:19008) 28. Julian Kellerhals, Nicolas Monod, and Mikael Rørdam, Non-supramenable groups act- ing on locally compact spaces, preprint (2013) 29. Eberhard Kirchberg and Mikael Rørdam, Non-simple purely infinite C ∗-algebras, Amer. J. Math. 122 (2000), no. 3, 637 -- 666. MR 1759891 (2001k:46088) 30. Eberhard Kirchberg and Adam Sierakowski, Strong pure infiniteness of crossed prod- ucts. 31. Akitaka Kishimoto, Outer automorphisms and reduced crossed products of simple C ∗- algebras, Comm. Math. Phys. 81 (1981), no. 3, 429 -- 435. MR 634163 (83c:46061) 32. Alex Kumjian, David Pask, Iain Raeburn, and Jean Renault, Graphs, groupoids, and Cuntz-Krieger algebras, J. Funct. Anal. 144 (1997), no. 2, 505 -- 541. MR 1432596 (98g:46083) 33. A. V. Lebedev, Topologically free partial actions and faithful representations of crossed products, Funktsional. Anal. i Prilozhen. 39 (2005), no. 3, 54 -- 63, 96. MR 2174606 (2006f:46068) 34. M. H. Mann, Iain Raeburn, and C. E. Sutherland, Representations of finite groups and Cuntz-Krieger algebras, Bull. Austral. Math. Soc. 46 (1992), no. 2, 225 -- 243. MR 1183780 (93k:46046) 35. Kevin McClanahan, K-theory for partial crossed products by discrete groups, J. Funct. Anal. 130 (1995), no. 1, 77 -- 117. MR 1331978 (96i:46083) 36. Dorte Olesen and Gert K. Pedersen, Applications of the Connes spectrum to C ∗- dynamical systems. III, J. Funct. Anal. 45 (1982), no. 3, 357 -- 390. MR 650187 (83i:46080) 37. John Quigg and Iain Raeburn, Characterisations of crossed products by partial actions, J. Operator Theory 37 (1997), no. 2, 311 -- 340. MR 1452280 (99a:46121) 38. Mikael Rørdam, On the structure of simple C ∗-algebras tensored with a UHF-algebra. II, J. Funct. Anal. 107 (1992), no. 2, 255 -- 269. MR MR1172023 (93f:46094) 39. Mikael Rørdam and Adam Sierakowski, Purely infinite C ∗-algebras arising from crossed products, Ergodic Theory Dynam. Systems 32 (2012), no. 1, 273 -- 293. MR 2873171 (2012m:46063) 40. Adam Sierakowski, Discrete crossed product C ∗-algebras, 2009, PhD thesis published by PhD Series, ISBN 978-87-91927-40-9, University of Copenhagen. 41. , The ideal structure of reduced crossed products, Munster J. Math. 3 (2010), 237 -- 261. MR 2775364 (2012g:46103) PURELY INFINITE PARTIAL CROSSED PRODUCTS 31 Department of Mathematics and Statistics, University of Ottawa, 585 King Edward Ave, K1N6P1 Ottawa, Canada E-mail address: [email protected] The School of Mathematics & Applied Statistics, University of Wollon- gong, Northfields Ave, 2522 NSW, Australia E-mail address: [email protected]
1804.03455	1	1804	2018-04-10T11:16:19	Monic representations of finite higher-rank graphs	[ "math.OA", "math.DS", "math.FA" ]	In this paper we define the notion of monic representation for the $C^$-algebras of finite higher-rank graphs with no sources, and undertake a comprehensive study of them. Monic representations are the representations that, when restricted to the commutative $C^$-algebra of the continuous functions on the infinite path space, admit a cyclic vector. We link monic representations to the $\Lambda$-semibranching representations previously studied by Farsi, Gillaspy, Kang, and Packer, and also provide a universal representation model for nonnegative monic representations.	math.OA	math	Monic representations of finite higher-rank graphs Carla Farsi, Elizabeth Gillaspy, Palle Jorgensen, Sooran Kang, and Judith Packer April 11, 2018 Abstract In this paper we define the notion of monic representation for the C∗-algebras of finite higher-rank graphs with no sources, and undertake a comprehensive study of them. Monic representations are the representations that, when restricted to the commutative C∗-algebra of the continuous functions on the infinite path space, admit a cyclic vector. We link monic representations to the Λ-semibranching representations previously studied by Farsi, Gillaspy, Kang, and Packer, and also provide a universal representation model for nonnegative monic representations. 2010 Mathematics Subject Classification: 46L05, 46L55, 46K10 Keywords and phrases: C∗-algebras, monic representations, higher-rank graphs, k-graphs, Λ-semibranching function systems, coding map, Markov measures, projective systems. Contents 1 Introduction 2 Foundational material 2.1 Higher-rank graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Λ-semibranching function systems and their representations . . . . . . . . . 3 Representations of higher-rank graph C∗-algebras: first analysis 3.1 Λ-projective systems and representations . . . . . . . . . . . . . . . . . . . . 3.2 Projection valued measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Disjoint and irreducible representations . . . . . . . . . . . . . . . . . . . . . 4 Monic representations of finite k-graph algebras 4.1 Λ-semibranching function systems and monic representations . . . . . . . . . 5 A universal representation for non-negative Λ-projective systems 2 4 4 6 8 8 11 13 14 19 23 1 1 Introduction Higher-rank graphs Λ – also known as k-graphs – and their C∗-algebras C∗(Λ) were in- troduced by Kumjian and Pask in [37], building on the work of Robertson and Steger [45, 46]. Generalizations of the Cuntz–Krieger C∗-algebras associated to directed graphs (cf. [11, 12, 23, 38]), k-graph C∗-algebras share many of the important properties of Cuntz and Cuntz–Krieger C∗-algebras, including Cuntz–Krieger uniqueness theorems and realiza- tions as groupoid C∗-algebras. Moreover, the C∗-algebras of higher-rank graphs are closely linked with orbit equivalence for shift spaces [10] and with symbolic dynamics more generally [43, 47, 44], as well as with fractals and self-similar structures [25, 26]. More links between higher-rank graphs and symbolic dynamics can be seen via [3, 4] and the references cited therein. The research presented in the pages that follow develops a non-commutative harmonic analysis for finite higher-rank graphs with no sources. More precisely, we introduce monic representations for the C∗-algebras associated to finite higher-rank graphs with no sources; undertake a detailed theoretical analysis of such representations; and present a variety of examples. Like the Cuntz–Krieger algebras, k-graph C∗-algebras often fall in a class of non-type I, and in fact purely infinite C∗-algebras. The significance of this for representation theory is that the unitary equivalence classes of irreducible representations of k-graph C∗-algebras do not arise as Borel cross sections [29, 30, 15, 21, 22]. In short, for these C∗-algebras, only subfamilies of irreducible representations admit "reasonable" parametrizations. Various specific subclasses of representations of Cuntz and Cuntz–Krieger C∗-algebras have been extensively studied by many researchers, who were motivated by their applicability to a wide variety of fields. In addition to connections with wavelets (cf. [17, 18, 40, 27, 28, 26]), representations of Cuntz–Krieger algebras have been linked to fractals and Cantor sets [48, 35, 25, 26] and to the endomorphism group of a Hilbert space [8, 39]. Indeed, the astonishing goal of identifying both discrete and continuous series of representations of Cuntz (and to some extent Cuntz–Krieger) C∗-algebras, was accomplished in [19, 20, 5], building on the pioneering results of [7]. In the setting of higher-rank graphs, however, the representation theory of these C∗- algebras is in its infancy. Although the primitive ideal space of higher-rank graph C∗-algebras is well understood [9, 34], representations of k-graph C∗-algebras have only been systemati- cally studied in the one-vertex case [13, 50, 14]. This motivated us to undertake the present detailed study of monic representations of k-graph C∗-algebras and their unitary equivalence classes. Despite the similarities between the Cuntz algebras and k-graph C∗-algebras which we have highlighted above, there are fundamental structural differences between them: for example, k-graph C∗-algebras need not be simple, nor is their K-theory known in general. Thus, the extension of results on representations for Cuntz algebras to the k-graph context is not automatic, and we are pleasantly surprised to have obtained such extensions in the pages that follow. The monic representations that we focus on in this paper were inspired in part by the wavelet theory for higher-rank graphs which was developed in [27]. These wavelets relied on the concept of Λ-semibranching function systems, introduced in [27] and further studied in [24]. In this paper, we refine the Λ-semibranching function systems into a crucial technical 2 tool for studying monic representations, namely the Λ-projective systems of Definition 3.1. Monic representations also have strong connections to Markov measures [19, 5] and Nelson's universal representation of an abelian algebra [42]. Indeed, studying monic representations enables us to convert questions about the representation theory of higher-rank graphs into measure-theoretic questions (see Theorems 3.11 and 3.13 below). This paper is organized as follows. We begin with an introductory section which reviews the basic notation and terminology for higher-rank graphs, as well as the Λ-semibranching function systems from [27]. Before turning our attention to a theoretic and systematic analysis of the monic representations of finite k-graph C∗-algebras, Section 3 develops the technical tools we will need for this analysis. The Λ-semibranching function systems of [27] are refined in Section 3.1 into Λ-projective systems, and Section 3.2 analyzes the projection- valued measure P = Pπ on the infinite path space Λ∞ which arises from a representation π of C∗(Λ). Section 3 ends by addressing the question of when representations of k-graph C∗-algebras are disjoint or irreducible, see Theorems 3.11 and 3.13. To be precise, Theorem 3.11 shows that for representations of C∗(Λ) arising from Λ-projective systems, the task of checking when two representations are equivalent reduces to a measure-theoretical problem. Theorem 3.13 characterizes the commutant of such representations, enabling a precise description of when a representation arising from a Λ-projective system is irreducible. Having laid the necessary technical groundwork, we undertake the promised analysis of monic representations of C∗(Λ) in Section 4. This section contains two of the main results of this paper as well as a number of examples of monic representations. Theorem 4.2 establishes that, when Λ is a finite and source-free k-graph, monic representations of C∗(Λ) are always unitarily equivalent to a Λ-projective representation on Λ∞. Theorem 4.5 gives an alternative, measure-theoretic characterization of when a Λ-semibranching function system gives rise to a monic representation. More precisely, Theorem 4.5 proves that a Λ- semibranching representation is monic if and only if the measure-theoretic subsets specified by the Λ-semibranching function system (see Definition 2.7 below) generate the σ-algebra. The final section, Section 5, relates monic representations to Nelson's universal Hilbert space, which we denote H(Λ∞). Theorem 5.7 shows that every monic representation whose associated Λ-projective system consists of positive functions is unitarily equivalent to a sub- representation of the so-called "universal representation" of C∗(Λ) on H(Λ∞) which is de- scribed in Proposition 5.3. In particular, this Theorem establishes that every representation of C∗(Λ) which arises from a Λ-semibranching function system, as in [27], is unitarily equiv- alent to a sub-representation of the universal representation. Acknowledgments The authors would like to thank Daniel Gon¸calves, Janos Englander and Alex Kumjian for helpful discussions. E.G. was partially supported by the Deutsches Forschungsgemeinschaft via the SFB 878 "Groups, Geometry, and Actions" of the Universitat Munster. S.K. was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (#2017R1D1A1B03034697). C.F. and J.P. were partially supported by two individual grants from the Simons Foundation (C.F. #523991; J.P. #316981). P.J. thanks his colleagues in the Math Department at the Uni- 3 versity of Colorado, for making a week-long visit there possible, for support, and for kind hospitality. Progress towards the completion of this manuscript was made by the first three named co-authors while in attendance at the Fields Institute (Toronto) and the Mathemat- ical Congress of the Americas (Montreal) in 2017; we are grateful for their support of our collaboration. C.F. also thanks IMPAN for hospitality during her visits to IMPAN, War- saw, Poland, where some of this work was carried out (grant #3542/H2020/2016/2). This paper was partially supported by the grant H2020-MSCA-RISE-2015-691246-QUANTUM DYNAMICS. 2 Foundational material 2.1 Higher-rank graphs We recall the definition of higher-rank graphs and their C∗-algebras from [37]. Let N = {0, 1, 2, . . .} denote the monoid of natural numbers under addition, and let k ∈ N with k ≥ 1. We write e1, . . . ek for the standard basis vectors of Nk, where ei is the vector of Nk with 1 in the i-th position and 0 everywhere else. Definition 2.1. [37, Definition 1.1] A higher-rank graph or k-graph is a countable small category1 Λ with a degree functor d : Λ → Nk satisfying the factorization property: for any morphism λ ∈ Λ and any m, n ∈ Nk such that d(λ) = m + n ∈ Nk, there exist unique morphisms µ, ν ∈ Λ such that λ = µν and d(µ) = m, d(ν) = n. When discussing k-graphs, we use the arrows-only picture of category theory; thus, ob- jects in Λ are identified with identity morphisms, and the notation λ ∈ Λ means λ is a morphism in Λ. We often regard k-graphs as a generalization of directed graphs, so we call morphisms λ ∈ Λ paths in Λ, and the objects (identity morphisms) are often called vertices. For n ∈ Nk, we write (1) Λn := {λ ∈ Λ : d(λ) = n} With this notation, note that Λ0 is the set of objects (vertices) of Λ, and we will call elements of Λei (for any i) edges. We write r, s : Λ → Λ0 for the range and source maps in Λ respectively. For vertices v, w ∈ Λ0, we define vΛw := {λ ∈ Λ : r(λ) = v, s(λ) = w} and vΛn := {λ ∈ Λ : r(λ) = v, d(λ) = n}. Our focus in this paper is on finite k-graphs with no sources. A k-graph Λ is finite if Λn is a finite set for all n ∈ Nk. We say that Λ has no sources or is source-free if vΛn 6= ∅ for all v ∈ Λ0 and n ∈ Nk. It is well known that this is equivalent to the condition that vΛei 6= ∅ for all v ∈ Λ and all basis vectors ei of Nk. For m, n ∈ Nk, we write m ∨ n for the coordinatewise maximum of m and n. Given λ, η ∈ Λ, we write Λmin(λ, η) := {(α, β) ∈ Λ × Λ : λα = ηβ, d(λα) = d(λ) ∨ d(η)}. (2) 1Recall that a small category is one in which the collection of arrows is a set. 4 If k = 1, then Λmin(λ, η) will have at most one element; this need not be true if k > 1. For finite source-free k-graphs Λ, for each 1 ≤ i ≤ k, we can define the ith vertex matrix (3) Ai ∈ MΛ0(N) by Ai(v, w) = vΛeiw. Observe that the factorization property implies that AiAj = AjAi for 1 ≤ i, j ≤ k. We now describe two fundamental examples of higher-rank graphs which were first men- tioned in the foundational paper [37]. More examples of higher-rank graphs can be found in Section 4.1 below. Example 2.2. (a) For any directed graph E, let ΛE be the category whose objects are the vertices of E and whose morphisms are the finite paths in E. Then ΛE is a 1-graph whose degree functor d : ΛE → N is given by d(η) = η (the number of edges in η). (b) For k ≥ 1, let Ωk be the small category with Obj(Ωk) = Nk, and Mor(Ωk) = {(p, q) ∈ Nk × Nk : p ≤ q}. Again, we can also view elements of Obj(Ωk) as identity morphisms, via the map Obj(Ωk) ∋ p 7→ (p, p) ∈ Mor(Ωk). The range and source maps r, s : Mor(Ωk) → If we define d : Ωk → Nk by Obj(Ωk) are given by r(p, q) = p and s(p, q) = q. d(p, q) = q − p, then one can check that Ωk is a k-graph with degree functor d. Definition 2.3 ([37] Definitions 2.1). Let Λ be a k-graph. An infinite path in Λ is a k-graph morphism (degree-preserving functor) x : Ωk → Λ, and we write Λ∞ for the set of infinite paths in Λ. Since Ωk has a terminal object (namely 0 ∈ Nk) but no initial object, we think of our infinite paths as having a range r(x) := x(0) but no source. For each m ∈ Nk, we have a shift map σm : Λ∞ → Λ∞ given by σm(x)(p, q) = x(p + m, q + m) (4) for x ∈ Λ∞ and (p, q) ∈ Ωk. It is well-known that the collection of cylinder sets Z(λ) = {x ∈ Λ∞ : x(0, d(λ)) = λ}, for λ ∈ Λ, form a compact open basis for a locally compact Hausdorff topology on Λ∞, under reasonable hypotheses on Λ (in particular, when Λ is row-finite: see Section 2 of [37]). If Λ is finite, then Λ∞ is compact in this topology. We also have a partially defined "prefixing map" σλ : Z(r(λ)) → Z(λ) for each λ ∈ Λ: λ(p, q), x(p − d(λ), q − d(λ)), λ(p, d(λ))x(0, q − d(λ)), p < d(λ) < q q ≤ d(λ) p ≥ d(λ)  σλ(x) = λx =(p, q) 7→ 5 Remark 2.4. The factorization rule implies an important property of infinite paths: for any x ∈ Λ∞ and m ∈ Nk, we have x = x(0, m)σm(x). Taking m = pej for an arbitrary p ∈ N reveals that every infinite path must contain infinitely many edges of each color. Moreover, if we take m = (n, n, . . . , n) ∈ Nk for some n ≥ 1, the factorization rule tells us that x(0, m) can be written uniquely as a "rainbow sequence" of edges: where d(f j i ) = ei. x(0, m) = f 1 1 f 1 2 · · · f 1 k f 2 1 · · · f 2 k f 3 1 · · · f n k , For example, suppose Λ is a 2-graph. We can visualize Λ as arising from a 2-colored graph (red and blue edges). Moreover, each infinite path x ∈ Λ∞ can be uniquely identified with an infinite string of alternating blue and red edges (setting blue to be "color 1" and red to be "color 2"). We stress that even finite k-graphs may have nontrivial infinite paths; in an infinite path, the same edge may occur multiple times and even infinitely many times. Now we introduce the C∗-algebra associated to a finite, source-free k-graph Λ. Definition 2.5. Let Λ be a finite k-graph with no sources. A Cuntz–Krieger Λ-family is a collection {tλ : λ ∈ Λ} of partial isometries in a C∗-algebra satisfying (CK1) {tv : v ∈ Λ0} is a family of mutually orthogonal projections, (CK2) tλtη = tλη if s(λ) = r(η), (CK3) t∗λtλ = ts(λ) for all λ ∈ Λ, (CK4) for all v ∈ Λ and n ∈ Nk, we have tv =Pλ∈vΛn tλt∗λ. The Cuntz–Krieger C∗-algebra C∗(Λ) associated to Λ is the universal C∗-algebra generated by a Cuntz–Krieger Λ-family. The condition (CK4) implies that for all λ, η ∈ Λ, we have t∗λtη = X(α,β)∈Λmin(λ,η) tαt∗β. (5) It follows that C∗(Λ) = span{tαt∗β : α, β ∈ Λ, s(α) = s(β)}. 2.2 Λ-semibranching function systems and their representations In [27], separable representations of C∗(Λ) were constructed by using Λ-semibranching func- tion systems on measure spaces. A Λ-semibranching function system is a generalization of the semibranching function systems studied by Marcolli and Paolucci in [40]. As established in [40, 27], Λ-semibranching function systems (and their one-dimensional counterparts) give rise to representations of C∗(Λ), and we provide examples of such representations in Section 4.1 below. Indeed, we build upon the notion of Λ-semibranching function systems in Sections 3 and 4 below to characterize the monic representations of higher-rank graphs. Definition 2.6. [40, Definition 2.1] Let (X, µ) be a measure space. Suppose that, for each 1 ≤ i ≤ N, we have a measurable map σi : Di → X, for some measurable subsets Di ⊂ X. The family {σi}N i=1 is a semibranching function system if the following holds: 6 (a) Setting Ri = σi(Di), we have µ(X \ ∪iRi) = 0, µ(Ri ∩ Rj) = 0 for i 6= j. (b) For each i, the Radon-Nikodym derivative Φσi = d(µ ◦ σi) dµ satisfies Φσi > 0, µ-almost everywhere on Di. A measurable map σ : X → X is called a coding map for the family {σi}N for all x ∈ Di. Definition 2.7. [27, Definition 3.2] Let Λ be a finite k-graph and let (X, µ) be a measure space. A Λ-semibranching function system on (X, µ) is a collection {Dλ}λ∈Λ of measurable subsets of X, together with a family of prefixing maps {τλ : Dλ → X}λ∈Λ, and a family of coding maps {τ m : X → X}m∈Nk, such that (a) For each m ∈ Nk, the family {τλ : d(λ) = m} is a semibranching function system, with i=1 if σ ◦ σi(x) = x coding map τ m. (b) If v ∈ Λ0, then τv = id, and µ(Dv) > 0. (c) Let Rλ = τλ(Dλ). For each λ ∈ Λ, ν ∈ s(λ)Λ, we have Rν ⊆ Dλ (up to a set of measure 0), and τλτν = τλν a.e. (Note that this implies that up to a set of measure 0, Dλν = Dν whenever s(λ) = r(ν)). (d) The coding maps satisfy τ m ◦ τ n = τ m+n for any m, n ∈ Nk. (Note that this implies that the coding maps pairwise commute.) Remark 2.8. We pause to note that condition (c) of Definition 2.7 above implies that Dλ = Ds(λ) and Rλ ⊂ Rr(λ) for λ ∈ Λ. Also, when Λ is a finite 1-graph, the definition In particular, of a Λ-semibranching function system is not equivalent to Definition 2.6. Definition 2.7(b) implies that the domain sets {Dv : v ∈ Λ0} must satisfy µ(Dv ∩ Dw) = µ(Rv ∩ Rw) = 0 for v 6= w ∈ Λ0, but Definition 2.6 does not require that the domain sets Di be mutually disjoint µ-a.e. In fact, Definition 2.7 implies what is called condition (C-K) in Section 2.4 of [5]: up to a measure zero set, Dv = [λ∈vΛm Rλ (6) for all v ∈ Λ0 and m ∈ N, since Rv = τv(Dv) = id(Dv) = Dv. Also notice that in the above decomposition the intersections Rλ ∩ Rλ′, Rλ 6= λ′, have measure zero. This condition is crucial to making sense of the representation of C∗(Λ) associated to the Λ-semibranching function system (see Theorem 2.10 below). As established in Theorem 2.22 of [5], in order to obtain a representation of a 1-graph algebra C∗(Λ) from a semibranching function system, one must also assume that the semibranching function system satisfies condition (C-K). 7 We pause to enumerate some properties of Λ-semibranching function systems, which can be proved by routine computations. Remark 2.9. 1. For any n ∈ Nk and any measurable E ⊆ X, we have τλ(E) and consequently µ ◦ (τ n)−1 << µ (7) (τ n)−1(E) = [λ∈Λn in any Λ-semibranching function system. 2. On Rλ, we have (τλ)−1 = τ n. Therefore, Condition (b) of Definition 2.6 implies that µ ◦ (τλ)−1 << µ on Rλ, and d(µ◦(τλ)−1) dµ is nonzero a.e. on Rλ. As established in [27], any Λ-semibranching function system gives rise to a representa- tion of C∗(Λ) via 'prefixing' and 'chopping off' operators that satisfy the Cuntz-Krieger relations. Intuitively, a Λ-semibranching function system is a way of encoding the Cuntz- Krieger relations at the measure-space level: the prefixing map τλ corresponds to the partial isometry sλ ∈ C∗(Λ). For the convenience of the reader, we recall the formula for these Λ-semibranching representations of C∗(Λ). Theorem 2.10. [27, Theorem 3.5] Let Λ be a finite k-graph with no sources and suppose that we have a Λ-semibranching function system on a measure space (X, µ) with prefixing maps {τλ : λ ∈ Λ} and coding maps {τ m : m ∈ Nk}. For each λ ∈ Λ, define an operator Sλ on L2(X, µ) by Sλξ(x) = χRλ(x)(Φτλ(τ d(λ)(x)))−1/2ξ(τ d(λ)(x)). Then the operators {Sλ : λ ∈ Λ} generate a representation π of C∗(Λ) on L2(X, µ). 3 Representations of higher-rank graph C∗-algebras: first analysis We begin this section by developing the technical tools which we will rely on throughout the paper: Λ-projective systems and projection valued measures. These tools enable us to describe when certain representations of k-graph C∗-algebras are disjoint or irreducible, see Theorems 3.11 and 3.13. 3.1 Λ-projective systems and representations The definition of a Λ-projective system generalizes to the k-graph setting the definition of a monic system in [19] (for the Cuntz algebras ON ) and [5] (in the case of Cuntz–Krieger algebras OA). We have decided to change the name because even for OA, not every monic system gives rise to a monic representation of OA. The word "projective" refers to the cocycle-like Condition (b) of Definition 3.1. Definition 3.1. Let Λ be a finite k-graph with no sources. A Λ-projective system on a measure space (X, µ) is a Λ-semibranching function system on (X, µ), with prefixing maps {τλ : Dλ → Rλ}λ∈Λ and coding maps {τ n : n ∈ Nk} together with a family of functions {fλ}λ∈Λ ⊆ L2(X, µ) satisfying the following conditions: 8 dµ (a) For any λ ∈ Λ, we have 0 6= d(µ◦(τλ)−1) = fλ2; (b) For any λ, ν ∈ Λ, we have fλ · (fν ◦ τ d(λ)) = fλν. Thus, a Λ-projective system on (X, µ) consists of a Λ-semibranching function system plus some extra information (encoded in the functions fλ). We have a certain amount of choice for the functions fλ; we can take positive or negative (or imaginary!) roots of d(µ◦(τλ)−1) for fλ, as long as they satisfy the multiplicativity Condition (b) above. dµ Example 3.2. For any Λ-semibranching function system on (X, µ), there is a natural choice of an associated Λ-projective system; namely, for λ ∈ Λn we define fλ(x) := Φλ(τ n(x))−1/2χRλ(x). (8) Condition (a) is satisfied because of the hypothesis that the Radon–Nikodym derivatives be strictly positive µ-a.e. on their domain of definition. Since the operators Sλ ∈ B(L2(X, µ)) of Theorem 2.10 are given by and [27, Theorem 3.5] establishes that {Sλ}λ∈Λ is a Cuntz–Krieger family, Proposition 3.4 below shows that Equation (8) indeed describes a Λ-projective system. Sλ(f ) = fλ · (f ◦ τ n), Remark 3.3. Observe that Condition (a) of Definition 3.1 forces fλ(x) = 0 a.e. outside of Rλ, since d(µ◦(τλ)−1) is supported only on Rλ. dµ Condition (b) of Definition 3.1 is needed to associate a representation of C∗(Λ) to a Λ-projective system. To be precise, we have: Proposition 3.4. Let Λ be a finite, source-free k-graph. Suppose that a measure space (X, µ) admits a Λ-semibranching function system with prefixing maps {τλ : λ ∈ Λ} and coding maps {τ n : n ∈ Nk}. Suppose that {fλ}λ∈Λ is a collection of functions satisfying Condition (a) of Definition 3.1. Then the maps {τλ}, {τ n} and {fλ}λ form a Λ-projective system on (X, µ) if and only if the operators Tλ ∈ B(L2(X, µ)) given by Tλ(f ) = fλ · (f ◦ τ d(λ)) (9) form a Cuntz–Krieger Λ-family with each Tλ nonzero (and hence give a representation of C∗(Λ)). Proof. If the operators Tλ of Equation (9) form a nontrivial Cuntz–Krieger Λ-family, then it is easily checked that the functions {fλ}λ∈Λ satisfy the hypotheses of Definition 3.1. On the other hand, suppose that (X, µ) admits a Λ-projective system with prefixing maps {τλ}λ∈Λ, coding maps {τ n}n∈Nk, and functions {fλ}λ∈Λ. We will show that the operators {Tλ} of Equation (9) satisfy Conditions (CK1)–(CK4). For (CK1), observe that if v ∈ Λ0, Tv(f ) = fv · (f ◦ τ 0) is supported on Dv = Rv by Condition (a) of Definition 3.1. Moreover, since v = v2 for any v ∈ Λ0, and τv = idDv = τ 0, Condition (b) of Definition 3.1 implies that fv = fv · (fv ◦ τ 0) = f 2 v ⇒ fv = χDv. 9 Consequently, Tv(f ) = χDv · f . Since the sets {Dv = Rv}v∈Λ0 are disjoint (up to a set of measure zero), it follows that {Tv : v ∈ Λ0} is a set of mutually orthogonal projections; in other words, (CK1) holds. For (CK2), fix λ, ν ∈ Λ with s(λ) = r(ν). Since fν(x) = 0 unless x ∈ Rν, we see that TλTν(f )(x) =(0, x 6∈ τλ(Rν) fλ(τλ ◦ τν(y)) · fν(τν(y)) · f (y), x = τλτν(y). On the other hand, Condition (b) of Definition 3.1 implies that if x = τλτν(y), fλ(τλ ◦ τν(y)) · fν(τν(y)) = fλ(x) · fν(τ d(λ)(x)) = fλν(y). This implies that TλTν = Tλν as claimed. To check (CK3), we first compute that T ∗λ f = f ◦ τλ · fλ ◦ τλ · Φλ. Alternatively, T ∗λ f = χDλ · (f ◦ τλ) fλ ◦ τλ . (10) Condition (CK3), and the fact that the operators Tλ are partial isometries, now follow from straightforward calculations. Finally, an easy computation establishes that TλT ∗λ (f ) = χRλ·f for any λ ∈ Λ, from which (CK4) follows. We call the representation given in Equation (9) a Λ-projective representation. The following Proposition enables us to translate a Λ-projective system on (X, µ) to a Λ-projective system on (X, µ′) for any measure µ′ which is equivalent to µ. Proposition 3.5. Let Λ be a finite k-graph with no sources. Suppose we are given a Λ- projective system {τλ : λ ∈ Λ}, {τ n : n ∈ Nk} and {fλ : λ ∈ Λ} on a measure space (X, µ). Let µ′ be a measure equivalent to µ, and set g1(x) = dµ′ dµ′ (x). If we define { fλ}λ∈Λ by dµ (x) and g2(x) = dµ · fλ(x), λ ∈ Λ, (11) fλ(x) = pg1 ◦ τ d(λ)(x) pg1(x) then {τλ : λ ∈ Λ}, {τ n : n ∈ Nk} and { fλ}λ∈Λ give a Λ-projective system on (X, µ′). Moreover, the associated representations {Tλ : λ ∈ Λ} and { Tλ : λ ∈ Λ} of C∗(Λ) on L2(X, µ) and L2(X, µ′) given by Equation (9) of Proposition 3.4 are unitarily equivalent via the unitary U given by U(f )(x) = s dµ dµ′ (x) · f (x), f ∈ L2(X, µ), U−1(h)(x) = sdµ′ dµ (x) · h(x), h ∈ L2(X, µ′). Proof. We leave the verification of this proposition to the reader. Proposition 3.6 below is the analog of Proposition 2.11 of [19] for Λ-projective systems. 10 Proposition 3.6. Let Λ be a finite k-graph with no sources. Suppose we are given two Λ-projective systems on X, with the same prefixing and coding maps {τλ : λ ∈ Λ}, {τ n : n ∈ Nk}, but with different measures µ, µ′ and Λ-projective functions {fλ}λ∈Λ for (X, µ) and {f′λ}λ∈Λ for (X, µ′). Let dµ′ = h2dµ + dν be the Lebesgue-Radon-Nikodym decomposition, with h ≥ 0 and ν singular with respect to µ. Then there is a partition of X into Borel sets X = A ∪ B such that: (a) The function h is supported on A, ν is supported on B, and µ(B) = 0, ν(A) = 0. (b) The sets A, B are invariant under τ n for all n ∈ Nk, i.e., (τ n)−1(A) = A, and (τ n)−1(B) = B. λ << ν and kλ :=q d(ν◦τ −1 dν λ ) (c) We have ν ◦ τ−1 (d) f′λ · h = fλ · (h ◦ τ d(λ)) µ-a.e. on A and f′λ = kλ ν-a.e. on B. Proof. We start by proving (a) and (b) together. Let B be the support of ν, and observe that µ( B) = 0. We observe that the definitions of Λ-semibranching function systems and is supported on B. Λ-projective systems, together with the fact that (τ n)−1( B) =Sλ∈Λn τλ( B), imply that have µ-measure zero. Therefore we can take the orbit B of B under the functions {τ n : n ∈ Nk} and {τλ : λ ∈ Λ}, and B will then have µ-measure zero. Let A := X\B. Then A contains the support of µ, and we can choose h to be supported on A. Moreover, ν(A) = 0. By construction, A and B are invariant under τ n. This establishes (a) and (b). To prove (c), let E be a Borel set with ν(E) = 0. Then ν(E ∩ B) = 0, so the fact that µ vanishes on B implies that µ′(E ∩ B) = 0. We consequently have µ′(τ−1 λ (E ∩ B)) = 0, which means that µ′(τ−1 λ are supported on B, it follows that kλ is supported on B. To see (d), let f be a bounded Borel function supported on A. Then we have λ (E)) = 0. Since B is invariant under τ−1 λ (E)∩ B) = 0, so ν(τ−1 λ and ν and ν ◦ τ−1 (τ n)−1( B) and (τλ)−1( B) f ZA f′λ2 f h2 dµ =ZA f′λ2 f dµ′ =ZA (f ◦ τλ) h2 dµ =ZX f (h2 ◦ τ d(λ))fλ2dµ, =ZA =ZX dµ′ =ZA d(µ′ ◦ τ−1 λ ) (f ◦ τλ) dµ′ d(µ′) (f ◦ τλ) (h2 ◦ τ d(λ) ◦ τλ) dµ =ZX f (h2 ◦ τ d(λ)) d(µ ◦ τ−1 λ ) which implies the first relation. The second relation follows from the fact that µ′B = ν. 3.2 Projection valued measures The second technical tool which underpins our analysis of the monic representations of C∗(Λ) is the projection valued measure associated to a representation of C∗(Λ). Our work in this section is inspired by Dutkay, Haussermann, and Jorgensen [19, 20]. 11 Definition 3.7. Let Λ be a finite k-graph with no sources. Given a representation {tλ}λ∈Λ of a k-graph C∗-algebra C∗(Λ) on a Hilbert space H, we define a projection valued function P on Λ∞ by P (Z(λ)) = tλt∗λ for all λ ∈ Λ. In the proof (Proposition 3.9) that P indeed defines a projection-valued measure on Λ∞, we rely on the following well-known Lemma. Thus, in our application, X = Λ∞ and Fn will be the σ-algebra generated by the cylinder sets Z(λ) with d(λ) = (n, . . . , n). Lemma 3.8 (Kolmogorov Extension Theorem, [36, 49]). Let (X,Fn, νn)n∈N be a sequence of probability measures (νn)n∈N on the same space X, each associated with a σ-algebra Fn; further assume that (X,Fn, νn)n∈N form a projective system, i.e., an inverse limit. Suppose that Kolmogorov's consistency condition holds: νn+1Fn = νn. Then there is a unique extension ν of the measures (νn)n∈N to the σ-algebraWn∈N Fn gener- ated bySn∈N Fn. Proposition 3.9. Let Λ be a finite k-graph with no sources. Given a representation {tλ}λ∈Λ of a k-graph C∗-algebra C∗(Λ) on a Hilbert space H, the function P of Definition 3.7 extends to a projection valued measure on the Borel σ-algebra Bo(Λ∞) of the infinite path space Λ∞. Proof. Recall from the proof of [27, Lemma 4.1] that {Z(λ) : d(λ) = (n, n, . . . , n) for some n ∈ N} generates the topology on Λ∞. Thus, Bo(Λ∞) = lim−→Fn. By Lemma 3.8, it therefore suffices to show that P (Z(λη)) P (Z(λ)) = Xη∈s(λ)Λ(1,...,1) P (Z(λ)) = tλt∗λ = tλ Xη∈s(λ)Λ(1,...,1) tηt∗η t∗λ = Xη∈s(λ)Λ(1,...,1) whenever d(λ) = (n, . . . , n) for some n ∈ N. However, this follows immediately from (CK4): P (Z(λη). (12) We now record some properties of P which we will rely on in the sequel. The equations below are the analogues for k-graphs of Equations (2.7) and (2.8) and (2.13) of [20]. Proposition 3.10. Let Λ be a row-finite, source-free k-graph, and fix a representation {tλ : λ ∈ Λ} of C∗(Λ). (a) For λ, η ∈ Λ with s(λ) = r(η), we have tλP (Z(η))t∗λ = P (σλ(Z(η))), where σλ is the prefixing map on Λ∞ given in Equation (4). 12 (b) For any fixed n ∈ Nk, we have Xλ∈f (η)Λn tλP (σ−1 λ (Z(η)))t∗λ = P (Z(η)); (c) For any λ, η ∈ Λ with r(λ) = r(η), we have tλP (σ−1 (d) When λ ∈ Λn, we have tλP (Z(η)) = P ((σn)−1(Z(η)))tλ. Proof. Straightforward calculation. λ (Z(η))) = P (Z(η))tλ; 3.3 Disjoint and irreducible representations In this section we will derive from the technical results in Section 3.1 important consequences that detail when representations of k-graph C∗-algebras are disjoint or irreducible. In par- ticular, Theorem 3.11 suggests the importance of dealing with Λ-projective systems with non-negative functions fλ. We will focus more exclusively on such Λ-projective systems in Section 5 below. Theorem 3.11. (C.f. Theorem 2.12 of [19]) Let Λ be a finite k-graph with no sources. Suppose we are given two Λ-projective systems on the infinite path space Λ∞ with the standard prefixing and coding maps {σλ : λ ∈ Λ}, {σn : n ∈ Nk}, but associated to different measures µ, µ′ and different Λ-projective families of non-negative functions {fλ}λ∈Λ on (Λ∞, µ), and {f′λ}λ∈Λ on (Λ∞, µ′). Then the two associated representations {Tλ : λ ∈ Λ} and {T ′λ : λ ∈ Λ} of C∗(Λ) given by Equation (9) of Proposition 3.4 are disjoint if and only if the measures µ and µ′ are mutually singular. Proof. If the representations are not disjoint, there exist subspaces Hµ ⊆ L2(Λ∞, µ) and Hµ′ ⊆ L2(Λ∞, µ′), preserved by their respective representations, and a unitary W : Hµ → Hµ′ such that W TλHµ = T ′λHµ′ W, W T ∗λHµ = (T ′λ)∗Hµ′ W. The fact that each operator T ∗λ also preserves Hµ implies that W TλT ∗λHµ = W TλHµT ∗λHµ = T ′λHµ′ W T ∗λHµ = T ′λ(T ′λ)∗Hµ′ W. Moreover, it follows easily from the formulas for Tλ and T ∗λ in Equations (9) and (10) that TλT ∗λ = MχZ(λ) = T ′λ(T ′λ)∗. In other words, the representations of C(Λ∞) given by χZ(λ) 7→ TλT ∗λ and χZ(λ) 7→ T ′λ(T ′λ)∗ (on L2(Λ∞, µ) and L2(Λ∞, µ′) respectively) are multiplication representations. Since W im- plements a unitary equivalence between their subrepresentations on Hµ and Hµ′ respectively, Theorem 2.2.2 of [2] tells us that the measures µ, µ′ cannot be mutually singular. For the converse, assume that the representations are disjoint and that the measures are not mutually singular. Then, use Proposition 3.6 and decompose dµ′ = h2dµ + dν, with the 13 subsets A, B as in Proposition 3.6. Define the operator W on L2(Λ∞, µ′) by W (f ) = f · h if f ∈ L2(A, µ′), and W (f ) = 0 on the orthogonal complement of L2(A, µ′) ⊆ L2(Λ∞, µ′). Since A is invariant under τ n for all n, L2(A, µ′) is an invariant subspace for the representation. To check that W is intertwining, we use part (d) of Proposition 3.6 and the non-negativity condition on {fλ} and {f′λ} to obtain the almost-everywhere equalities TλW (f ) = fλ(h ◦ τ d(λ))(f ◦ τ d(λ)) = f′λ h (f ◦ τ d(λ)) = W T ′λ(f ). Since W intertwines the representations {Tλ}λ∈Λ,{T ′λ}λ∈Λ of C∗(Λ), we must have W = 0; hence h = 0, so µ, µ′ are mutually singular. Remark 3.12. As a Corollary of Theorem 3.11, we see that the examples of Markov measures introduced in [24, Section 4.2] generate representations of C∗(Λ) disjoint from the represen- tation of [27, Theorem 3.5]; see also Example 4.10 below. In fact, these Markov measures are mutually singular with the Perron–Frobenius measure [19], [33]. Theorem 3.13. (C.f. Theorem 2.13 of [19]) Let Λ be a finite k-graph with no sources. Suppose that the infinite path space Λ∞ admits a Λ-projective system on (Λ∞, µ) for some measure µ, with the standard prefixing maps {σλ : λ ∈ Λ} and coding maps {σn : n ∈ Nk} of Definition 2.3. Let {Tλ : λ ∈ Λ} be the associated representation of C∗(Λ). Then: (a) The commutant of the operators {Tλ : λ ∈ Λ} consists of multiplication operators by functions h with h ◦ σn = h, µ-a.e for all n ∈ Nk. (b) The representation given by {Tλ : λ ∈ Λ} is irreducible if and only if the coding maps σn are jointly ergodic with respect to the measure µ, i.e., the only Borel sets A ⊂ Λ∞ with (σn)−1(A) = A for all n are sets of measure zero, or of full measure. Proof. We first observe that the commutant of {Tλ}λ∈Λ is contained in C(Λ∞)′ ⊆ B(L2(Λ∞, µ)) and hence consists of multiplication operators. The proof of part (a) is then a straightforward calculation, and part (b) follows from part (a) and the definition of ergodicity. 4 Monic representations of finite k-graph algebras The first main result of this section, Theorem 4.2, establishes that every monic representation of a finite, strongly connected k-graph algebra C∗(Λ) is unitarily equivalent to a Λ-projective representation of C∗(Λ) on L2(Λ∞, µπ), where the measure µπ arises from the representation. (See Definition 4.1 and Equation (14) below for details.) After proving Theorem 4.2, we examine a variety of examples of representations of C∗(Λ), and identify which representations are monic. This analysis requires our second main result, Theorem 4.5, which provides a measure-theoretic characterization of when a Λ-semibranching representation is monic. Definition 4.1. Let Λ be a finite k-graph with no sources. A representation {tλ : λ ∈ Λ} of a k-graph on a Hilbert space H is called monic if tλ 6= 0 for all λ ∈ Λ, and there exists a vector ξ ∈ H such that spanλ∈Λ{tλt∗λξ} = H. 14 From the projection valued measure P associated to {tλ : λ ∈ Λ} as in Theorem 3.9, we obtain a representation π : C(Λ∞) → B(H): π(f ) =ZΛ∞ f (x)dP (x), which gives, for λ ∈ Λ, π(χZ(λ)) =ZΛ∞ χZ(λ) (x) dP (x) = P (Z(λ)) = tλt∗λ. (13) Since we can view C(Λ∞) as a subalgebra of C∗(Λ) via the embedding χZ(λ) 7→ tλt∗λ, the representation π is often understood as the restriction of the representation {tλ}λ∈Λ to the "diagonal subalgebra" span{tλt∗λ}λ∈Λ. If the representation {tλ}λ is monic, there is a cyclic vector ξ ∈ H for π. This induces we obtain a Borel measure µπ on Λ∞ given by µπ(Z(λ)) = hξ, P (Z(λ))ξi = hξ, tλt∗λξi. (14) Theorem 4.2. Let Λ be a finite k-graph with no sources. If {tλ}λ∈Λ is a monic representation of C∗(Λ) on a Hilbert space H, then {tλ}λ∈Λ is unitarily equivalent to a representation {Sλ}λ∈Λ associated to a Λ-projective system on (Λ∞, µπ), which is associated to the standard coding and prefixing maps σn, σλ of Definition 2.3. Conversely, if we have a representation of C∗(Λ) on L2(Λ∞, µ) which arises from a Λ- projective system associated to the standard coding and prefixing maps σn, σλ, then the rep- resentation is monic. By Example 3.2, this implies that a Λ-semibranching function system on (Λ∞, µ), for any Borel measure µ, gives rise to a monic representation of C∗(Λ). Proof. Suppose that the representation {tλ}λ∈Λ of C∗(Λ) is monic, and let ξ ∈ H be a cyclic vector for C(Λ∞). Note that the map W : C(Λ∞) → H given by W (f ) = π(f )ξ is linear. Moreover, if we think of C(Λ∞) as a dense subspace of L2(Λ∞, µπ), the operator W is isometric: L2 =ZΛ∞ f2 dµπ = hξ, π(f2)ξi = kπ(f )ξk2 = kW (f )k2. kfk2 Therefore W extends to an isometry from L2(Λ∞, µπ) to H. Since W is also onto (because the representation is monic), W is a surjective isometry; that is, W is a unitary. Moreover, for any f ∈ C(Λ∞) and any ϕ ∈ L2(Λ∞, µπ), we have π(f )W (ϕ) = π(f )π(ϕ)ξ = π(f · ϕ)ξ = W (f · ϕ). Thus, unitarity of W implies that W ∗π(f )W acts on L2(Λ∞, µπ) by multiplication by f : W ∗π(f )W = Mf and W Mf W ∗ = π(f ). (15) 15 Now define the operator Sλ = W ∗tλW for λ ∈ Λ. By construction, the operators {Sλ}λ∈Λ also give a representation of C∗(Λ). Moreover, since W is a unitary, SλS∗λ(f ) = W ∗tλt∗λW (f ) = W ∗π(χZ(λ))π(f )ξ = W ∗π(χZ(λ) · f )ξ = W ∗W (χZ(λ) · f ) = χZ(λ) · f. (16) Let 1 denote the characteristic function of Λ∞, and define a function fλ ∈ L2(Λ∞, µπ) by fλ = Sλ1 = W ∗tλξ. We will now show that the functions fλ, combined with the usual coding and prefixing maps {σn, σλ}n,λ on Λ∞, form a Λ-projective system on (Λ∞,Bo(Λ∞), µπ). To that end, we will invoke Proposition 3.10. Since P (Z(ν)) = π(χZ(ν)) for any ν ∈ Λ, and the proof of [27, Lemma 4.1] shows that characteristic functions of cylinder sets densely span L2(Λ∞, µπ), the equalities established in Proposition 3.10 still hold if we replace P (Z(ν)) by π(f ) for any f ∈ L2(Λ∞, µπ). In particular, noting that χ(σn)−1(Z(ν)) = χZ(ν) ◦ σn and χσ−1 λ (Z(ν)) = χZ(ν) ◦ σλ Part (d) of Proposition 3.10 implies that if d(λ) = n, tλπ(f ) = π(f ◦ σn)tλ (17) and Part (c) implies that (18) Let f ∈ L2(Λ∞, µπ) and let n = d(λ). By using Part (d) of Proposition 3.10, Equation t∗λπ(f ) = π(f ◦ σλ)t∗λ. (15), and the fact that W is a unitary, we obtain Sλ(f ) = W ∗tλW (f ) = W ∗tλπ(f )ξ = W ∗π(f ◦ σn)tλξ = W ∗π(f ◦ σn)W W ∗tλξ = (f ◦ σn) · fλ. In order to show that {Sλ}λ∈Λ is a Λ-projective representation, then, Proposition 3.4 tells us that it remains to check that the standard prefixing and coding maps make (Λ∞, µπ) into a Λ-semibranching function system, and that Condition (a) of Definition 3.1 holds for the functions fλ. To establish Condition (a), we work indirectly. Since W is a unitary, we have (for any f ∈ L2(Λ∞, µπ) and any λ ∈ Λn) ZΛ∞ fλ2 · f dµπ = hSλ(1), Sλ(1) · fiL2 = hW ∗tλ(ξ), Mf W ∗tλ(ξ)iL2 = htλξ, W Mf W ∗(tλξ)iH = hξ, t∗λπ(f )tλξiH = hξ, π(f ◦ σλ)ξiH =ZΛ∞ f ◦ σλ dµπ =ZΛ∞ f d(µπ ◦ σ−1 λ ). 16 If E ⊆ Λ∞ is any set for which µπ(E) = 0, then taking f = χE above shows that µπ ◦ σ−1 λ (E) = 0 also – in other words, The uniqueness of Radon-Nikodym derivatives then implies that µπ ◦ σ−1 λ << µπ. (19) fλ2 = d(µπ ◦ σ−1 λ ) d(µπ) . In other words, Condition (a) of Definition 3.1 holds. We now show that fλ 6= 0 a.e. on Z(λ). Define Eλ ⊆ Z(λ) by Eλ := {x ∈ Z(λ) : fλ(x) = 0}. λ (Eλ) = π(χσ−1 Then 0 =REλ fλ2 dµπ = µπ ◦ σ−1 λ (Eλ)) = t∗λπ(χEλ)tλ by Proposition 3.10. By hypothesis, tλ 6= 0, so there exists ζ ∈ H such that tλ(ζ) 6= 0. However, for any ζ, htλ(ζ), π(χEλ)tλ(ζ)i = htλ(ζ), π(χEλ)2π(χZ(λ))tλ(ζ)i = ht∗λπ(χEλ)tλ(ζ), t∗λπ(χEλ)tλ(ζ)i = 0 by the Cuntz–Krieger relations and the fact that π(C(Λ∞)) is abelian. In other words, π(χEλ) is orthogonal to the range projection π(χZ(λ)) of tλ. On the other hand, χEλχZ(λ) = χEλ since Eλ ⊆ Z(λ). It follows that π(χEλ) = 0; equivalently, µπ(Eλ) = 0. In other words, the set Eλ ⊆ Z(λ) of points where fλ = 0 has µπ-measure zero, as claimed. Similarly, for any set F ⊆ Z(s(λ)) such that µπ(F ) = 0, taking f = χσλ(F ) reveals that 0 = µπ(F ) = µπ ◦ σ−1 λ (σλ(F )) =Zσλ(F ) fλ2 dµπ. Since fλ2 > 0 a.e. on Z(λ) ⊇ σλ(F ), we must have µπ ◦ σλ(F ) = 0 and hence µπ ◦ σλ << µπ. Furthermore, the Radon-Nikodym derivative d(µπ◦σλ) is nonzero µπ-a.e. on Z(s(λ)). To see d(µπ ) this, we set λ )(σλ(E)) = 0. Equation (19) therefore implies that µπ(E) = (µπ ◦ σ−1 [24, Proposition 4.1] now implies that the standard prefixing and coding maps make (Λ∞, µ) into a Λ-semibranching function system. Consequently, the functions fλ make {Sλ}λ∈Λ into a Λ-projective representation, which is unitarily equivalent to our initial monic representation by construction. For the converse, suppose that {tλ}λ∈Λ is a representation of C∗(Λ) on L2(Λ∞, µ), for some Borel measure µ, which arises from a Λ-projective system {fλ}λ∈Λ associated to the standard coding and prefixing maps {σn, σλ}n,λ. The computations from Proposition 3.4 establish that tλt∗λ is given by multiplication by χRλ = χZ(λ). Consequently, 1 = χΛ∞ is a cyclic vector for C(Λ∞) ⊆ C∗(Λ). Thus, {tλ}λ∈Λ is monic. 17 and observe that E =(cid:26)x ∈ Z(s(λ)) : µπ(σλ(E)) =ZE d(µπ ◦ σλ) d(µπ) = 0(cid:27) d(µπ ◦ σλ) d(µπ) dµπ = 0. Remark 4.3. In the final section of their paper [5], Bezuglyi and Jorgensen studied the relationship between semibranching function systems and monic representations of Cuntz– Krieger algebras (1-graph C∗-algebras). Theorem 5.6 of [5] establishes that within a specific class of semibranching function systems, which the authors term monic systems, those for which the underlying space is the infinite path space Λ∞ are precisely the systems which give rise to monic representations of the Cuntz–Krieger algebra. The Λ-projective systems studied in Section 3.1 constitute our extension to k-graphs of the monic systems for Cuntz–Krieger algebras. Thus, even in the case of 1-graph algebras (Cuntz–Krieger algebras), our Theorem 4.2 is substantially stronger than Theorem 5.6 of [5]: our Theorem 4.2 gives a complete characterization of monic representations, without the hypothesis that such representations arise from a monic or Λ-projective system. Theorem 4.4. Let Λ be a finite, source-free k-graph, and let {Sλ}λ∈Λ, {Tλ}λ∈Λ be two monic representations of C∗(Λ). Let µS, µT be the measures on Λ∞ associated to these representations as in (14). The representations {Sλ}λ∈Λ, {Tλ}λ∈Λ are equivalent if and only if the measures µS and µT are equivalent and there exists a function h on Λ∞ such that dµS dµT = h2 and f S λ = h ◦ σn h f T λ for all λ ∈ Λ with d(λ) = n. (20) (21) Proof. Suppose {Sλ}λ∈Λ, {Tλ}λ∈Λ are equivalent representations of C∗(Λ). From Theorem 3.11, it follows that the associated measures µS, µT are equivalent. Let W : L2(Λ∞, µS) → L2(Λ∞, µT ) be the intertwining unitary for them. Then the two representations are also equivalent when restricted to the diagonal subalgebra C∗({tλt∗λ : λ ∈ Λ}). By linearity, we can extend the formula from Equation (16) to all of C(Λ∞). It follows that πS, πT are both given on C(Λ∞) by multiplication: πS(φ) = Mφ and πT (φ) = Mφ ∀ φ ∈ C(Λ∞). Since W intertwines a dense subalgebra – namely πS(C(Λ∞)) – of the maximal abelian sub- algebra L∞(Λ∞, µS) ⊆ B(L2(Λ∞, µS)) which consists of multiplication operators, with the dense subalgebra πT (C(Λ∞)) ⊆ L∞(Λ∞, µT ), the unitary W must be given by multiplication by some nowhere-vanishing function h on Λ∞: W (f ) = hf. Moreover, since W is a unitary, ZΛ∞ W (f )2dµT =ZΛ∞ f2h2dµT =ZΛ∞ f2dµS for all f ∈ L2(Λ∞, µS), which implies (20). From the intertwining property Tλ W = W Sλ we obtain, for any f ∈ L2(Λ∞, µS) and any λ with d(λ) = n, that Tλ W (f ) = W Sλ(f ), that is, f T λ (h ◦ σn)(f ◦ σn) = hf S λ (f ◦ σn). Take f = 1 and we obtain that f T λ h ◦ σn h = f S λ 18 as claimed in (21). For the converse, suppose that the measures µS, µT are equivalent and there is a function h on Λ∞ satisfying (20) and (21). Then define W : L2(Λ∞, µS) → L2(Λ∞, µT ) by W f = hf ; it is then straightforward to check that W Sλ = TλW and that W is a unitary. 4.1 Λ-semibranching function systems and monic representations In this section, we discuss several examples of Λ-semibranching function systems and identify which of them give rise to monic representations of C∗(Λ) – or, equivalently, which are unitarily equivalent to Λ-semibranching function systems on the infinite path space. First, we provide another characterization of monic representations. The next theorem shows that a Λ-semibranching system on (X, µ) induces a monic representation of C∗(Λ) if and only if its associated range sets generate the σ-algebra of X. To state our result more precisely, we will denote by (X,F , µ) the measure space associated to L2(X, µ); in particular, F is the standard σ-algebra associated to L2(X, µ). Theorem 4.5. Let Λ be a finite, source-free k-graph and let {tλ}λ∈Λ be a Λ-semibranching representation of C∗(Λ) on L2(X,F , µ) with µ(X) < ∞. Let R be the collection of sets which are modifications of range sets Rλ by sets of measure zero; that is, each element X ∈ R has the form X = Rλ ∪ S or X = Rλ\S for some set S of measure zero. Let σ(R) be the σ-algebra generated by R. The represen- tation {tλ}λ∈Λ is monic, with cyclic vector χX ∈ L2(X,F , µ), if and only if σ(R) = F . In particular, for a monic representation {tλ}λ∈Λ, the set S :=n nXi=1 aitλit∗λiχX n ∈ N, λi ∈ Λ, ai ∈ Co =n nXi=1 aiχRλi n ∈ N, λi ∈ Λ, ai ∈ Co is dense in L2(X,F , µ). Proof. Suppose first that the representation {tλ}λ∈Λ is monic and that χX is a cyclic vector for the representation. As computed in the proof of Theorem 3.4 of [27], we have tλt∗λ(χX) = χRλ. Therefore, our hypothesis that χX is a cyclic vector implies that for any f ∈ L2(X,F , µ), there is a sequence (fj)j, with fj ∈ span{χRλ : λ ∈ Λ}, such that j→∞ZX fj − f2 dµ = 0. lim In particular, (fj) → f in measure. 19 For any σ-algebra T , standard measure-theoretic results [41, Proposition 6] imply that since µ(X) < ∞, convergence in measure among T -a.e. finite measurable functions on (X,T , µ) is metrized by the distance dT (f, g) :=ZΩ f − g 1 + f − g dµ. Moreover, dT makes the space of S-a.e. finite measurable functions into a complete metric space (this can be seen, for example, by combining Proposition 1 and Corollary 7 of [41]). The fact that (fj)j → f in measure in (X,F , µ), and that fj ∈ L2(X, σ(R), µ) for all j, implies that (fj)j is a Cauchy sequence with respect to both dF and dσ(R). Consequently, the limit f of (fj)j must also be a σ(R)-a.e. finite measurable function. In other words, every f ∈ L2(X,F , µ) is in fact in L2(X, σ(R), µ). Since R ⊆ F by construction we must have σ(R) = F , as desired. For the converse, assume σ(R) = F . We begin by observing that R := {finite unions of elements in R} aiχBi n ∈ N, Bi ∈ σ( R), ai ∈ Co is a subalgebra of P(X) – that is, closed under finite unions and complements. Closure under finite unions follows from the definition, while the second claim follows from Equation (6). Moreover, σ(R) = σ( R) = F , so S :=n nXi is dense in L2(X,F , µ). Therefore, the Carath´eodory/Kolmogorov extension theorem implies restricted to R induces a unique (extended) measure on F = σ(R), that the measure µ R which we still call µ. (This is indeed the original measure on L2(X,F , µ) by the uniqueness of the extension.) To show that the vector χX is monic, equivalently that the set S is dense in L2(X,F , µ) equipped with the usual metric dL2(X,F ,µ) coming from the L2 norm, we invoke a standard if (Q, dQ) is a metric space, and if Σ ⊆ Q is a dense subset of fact about metric spaces: (Q, dQ), then any other subset Σ ⊆ Q having the property ∀ǫ > 0, ∀x ∈ Σ, ∃ xǫ ∈ Σ with dQ(x, xǫ) < ǫ is also dense in (Q, dQ). We wish to apply this fact in the setting where (Q, dQ) = (L2(X,F , µ), dL2(X,F ,µ)), with Σ = S, Σ = S. Choose s ∈ S and fix ǫ > 0. Without loss of generality we can assume s =Pn i aiχBi for some n ∈ N, Bi ∈ σ( R), ai 6= 0. Define A :=Pn i ai ∈ R. The Carath´eodory/Kolmogorov extension theorem2 also guarantees that for any B ∈ σ( R) = F and for any ǫ > 0, there exists A B ǫ(cid:17) < ǫ, µ(cid:16) B∆A B ǫ ∈ R with 2See [16] Page 452, Appendix: Measure theory, Exercise 3.1. 20 where ∆ denotes symmetric difference. In other words, for each i, there exists Aǫ that i ∈ R such µ(Bi∆Aǫ i) < ǫ2 A2 , or, equivalently,(cid:16)µ(Bi∆Aǫ i)(cid:17)1/2 < ǫ A . Thus, setting sǫ :=Pn as desired. i aiχAǫ i and using the triangle inequality yields that dL2(X,F ,µ)(s, sǫ) < ǫ, Remark 4.6. Using the characterization of monic representations from Theorem 4.5, it is straightforward to check that the Λ-semibranching function systems detailed in [24, Example 3.5 and Section 4] generate monic representations of C∗(Λ). Similarly, suppose Λ = Λ1×Λ2 is a product k-graph and we have Λi-semibranching function systems on measure spaces (Xi, µi) for i = 1, 2, such that the associated Λi-semibranching representations are monic. Then Theorem 4.5 combines with [24, Proposition 3.4] to tell us that the product Λ-semibranching function system on (X1 × X2, µ1 × µ2) also gives rise to a monic representation of C∗(Λ). We now proceed to analyze several other examples of representations arising from Λ- semibranching function systems and establish which ones are monic representations. Example 4.7. We present here an example of a Λ-semibranching representation on a 1-graph that is not monic. The 1-graph Λ has two vertices v1 and v2 and three edges f1, f2 and f3. f2 f1 v1 v2 f3 Let X be the closed unit interval [0, 1] of R with the usual Lebesgue σ-algebra and measure µ. For v1 and v2, let Dv1 = [0, 1 2, 1]. Also for each edge f ∈ Λ, let Df = Ds(f ), and hence Df1 = Dv1 = [0, 1 2, 1]. Now define prefixing maps for f1, f2 and f3 by 2] and Dv2 = ( 1 2], Df2 = Dv2 = ( 1 2 , 1] and Df3 = Dv2 = ( 1 τf1(x) = − τf2(x) = − 1 2 1 2 x + x + 1 2 1 2 τf3(x) = x 2 1 2(cid:3), for x ∈ Df1 =(cid:2)0, , 1(cid:3), for x ∈ Df2 =(cid:0) 1 for x ∈ Df3 =(cid:0) 1 , 1(cid:3). 2, 1(cid:3). Then the ranges of the prefixing maps are x∈E⊆Dfi( 1 i = 1, 2, i = 3 2 µ(E) µ(E) , µ(E) µ(E) , = inf 2 Then Rf1 =(cid:2) 1 4 , 1 2(cid:3), Rf2 =(cid:2)0, 1 4(cid:1) and Rf3 =(cid:0) 1 mutually disjoint and X = Rf1 ∪ Rf2 ∪ Rf3. For each fi, since Lebesgue measure is regular, the Radon-Nikodym derivative of τfi is given by Φfi(x) = inf x∈E⊆Dfi (µ ◦ τfi)(E) µ(E) =(cid:26) 1 2, 1, i = 1, 2 i = 3. 21 Now define τ 1 : X → X by τ 1(x) = τ−1 f1 (x) τ−1 f2 (x) τ−1 f3 (x) for x ∈ Rf1 for x ∈ Rf2 for x ∈ Rf3 Since the sets Rfi are mutually disjoint, τ 1 is well defined on X. Then τ 1 is the coding map satisfying τ 1 ◦ τfi(x) = x for all x ∈ Dfi. It is a straightforward calculation to check that {τfi : Dfi → Rfi, i = 1, 2, 3} is a semi- branching function system for (X, µ). To see that this this Λ-semibranching function system does not give rise to a monic representation, we argue by contradiction. First, observe that the only finite paths with range v2 are of the form f3f3 · · · f3; and since τf3(x) = x on D3 = (1/2, 1], we have Rf3 = Rf3f3···f3 = (1/2, 1]. Every other finite path λ, having range v1, will satisfy Rλ ⊆ Dv1 = [0, 1/2]. Consequently, R = {Rλ}λ∈Λ does not generate the usual Lebesgue σ-algebra on [0, 1], even after modification by sets of measure zero, since the restriction of R to (1/2, 1] contains no nontrivial measurable sets. Theorem 4.5 therefore implies that the representation of C∗(Λ) associated to this Λ-semibranching function system is not monic, and hence is not equivalent to any representation on L2(Λ∞, µ) arising from a Λ-projective system. Remark 4.8. We observe that since monic representations are multiplicity free, it is easy to construct further examples of non-monic representations by using direct sums of monic representations, see [2] page 54. In order to describe the following example of a Λ-semibranching representation which is monic, we review the concept of a Markov measure (see [19, Section 3.1] or [24, Section 4.2] for more details, or [5] for Markov measures in a more general context). Definition 4.9 (Definition 3.1 of [19]). A Markov measure on the infinite path space Λ∞ ON Λ∞ ON = ∞Yi=1 ZN = {(i1i2 . . . ) : in ∈ ZN , n = 1, 2, . . .}. of the Cuntz algebra ON is defined by a vector λ = (λ0, . . . , λN−1) and an N × N matrix T such that λi > 0, Ti,j > 0 for all i, j ∈ ZN , and if e = (1, 1, . . . , 1)t then λT = λ and T e = e. The Carath´eodory/Kolmogorov extension theorem then implies that there exists a unique Borel measure µ on Λ∞ ON extending the measure µC defined on cylinder sets by: µC(Z(I)) := λi1Ti1,i2 · · · Tin−1,in, if I = i1 . . . in. (22) The extension µ is called a Markov measure on Λ∞ ON . For N = 2, given a number x ∈ (0, 1), we can take T = Tx =(cid:18) x (1 − x) (1 − x) x (cid:19), and λ = (1, 1). The resulting measure will in this case be called µx. Moreover, if x 6= x′, Theorem 3.9 of [19] guarantees that µx, µx′ are mutually singular. 22 Example 4.10. We now consider an example of Λ-semibranching function system which does give rise to a monic representation. Let Λ be the 2-graph below. f2 f1 e v Recall from Remark 2.4 that every infinite path in a 2-graph can be uniquely written as an infinite string of composable edges which alternate in color: red, blue, red, . . . . It follows that the infinite path space of the above 2-graph is homeomorphic to Λ∞ Z2 via the identification O2 ∼=Q∞i=1 efj1efj2efj3 · · · 7→ j1j2j3 · · · . Therefore, the measure µx described above can be viewed as a measure on Λ∞. It is straightforward to check that, as operators on L2(Λ∞, µx), the prefixing operators σe, σf1, σf2 have positive Radon-Nikodym derivatives at any point z ∈ Λ∞. Consequently, the standard prefixing and coding maps make (Λ∞, µx) into a Λ-semibranching function system. The associated representation of C∗(Λ) is therefore monic, by Theorems 4.2 and Theorem 2.10, and Example 3.2. 5 A universal representation for non-negative Λ-projective systems The focus of this section is the construction of a 'universal representation' of C∗(Λ), general- izing the work of [19, Section 4] for the Cuntz algebra setting, such that every non-negative monic representation of C∗(Λ) is a sub-representation of the universal representation. The Hilbert space H(Λ∞) on which our universal representation is defined is the 'universal space' for representations of C(Λ∞), see [42], and also [19, 6, 1, 32]. For the case of ON , this space was also shown to be the 'universal representation space' for monic representations in [19]. We recall the construction of H(Λ∞) below. Definition 5.1. Let Λ be a finite k-graph with no sources, and let Λ∞ be the infinite path space of Λ, endowed with the topology generated by the cylinder sets and the Borel σ-algebra associated to it. Consider the collection of pairs (f, µ), where µ is a Borel measure on Λ∞, and f ∈ L2(Λ∞, µ). We say that two pairs (f, µ) and (g, ν) are equivalent, denoted by (f, µ) ∼ (g, ν), if there exists a finite Borel measure m on Λ∞ such that µ << m, ν << m, and fr dµ dm = gr dν dm We write f √dµ for the equivalence class of (f, µ). in L2(Λ∞, m). 23 Proposition 8.3 of [6] establishes that H(Λ∞) is a Hilbert space, with the vector space structure given by scalar multiplication and and the inner product given by d(µ + ν) fpdµ + g √dν :=(cid:16)fs dµ hfpdµ, g √dνi :=ZΛ∞ + gs dν d(µ + ν)(cid:17)pd(µ + ν), f g(cid:16)s dµ d(µ + ν) s dν d(µ + ν)(cid:17)d(µ + ν). (23) We call H(Λ∞) the universal Hilbert space for Λ∞. The following fundamental property of H(Λ∞) justifies the name 'universal Hilbert space.' Proposition 5.2. ([32, Theorem 3.1], [19], [1]) Let Λ be a finite k-graph with no sources. For every finite Borel measure µ on Λ∞, define Wµ : L2(Λ∞, µ) → H(Λ∞) by Wµ(f ) = f√dµ. Then Wµ is an isometry of L2(Λ∞, µ) onto a subspace of H(Λ∞), which we call L2(µ). We are now ready to present the universal representation πuniv of C∗(Λ) on H(Λ∞). Proposition 5.3. Let Λ be a finite k-graph with no sources. Fix (f, µ) ∈ H(Λ∞). For each λ ∈ Λn, define Suniv ∈ B(H(Λ∞)) by Suniv λ (fpdµ) := (f ◦ σn)qd(µ ◦ σ−1 λ ), where σλ and σn are the standard prefixing and coding maps of Definition 2.3. Then: (a) The adjoint of Suniv λ is given by (Suniv λ )∗(f √dµ) := (f ◦ σλ)pd(µ ◦ σλ). (b) The operators {Suniv λ : λ ∈ Λ} generate a representation πuniv of C∗(Λ) on H(Λ∞), which we call the 'universal representation'. (c) The projection valued measure P on Λ∞ given in Definition 3.7 associated to the uni- versal representation πuniv is given by: where A is a Borel set of the Borel σ-algebra generated by the cylinder sets. P (A)(fpdµ) = (χA · f )pdµ, (24) Proof. The proof is similar to that of Proposition 4.2 of [19], although the details are more involved because of the more complicated k-graph structure. To simplify the notation, in this proof we will drop the superscript univ from Suniv }λ∈Λ gives a representation of C∗(Λ), first we observe that the operators Sλ are well defined; in other λ ) for all words, if f√dµ = g√dν, we must have (f ◦ σn)qd(µ ◦ σ−1 λ ) = (g ◦ σn)qd(ν ◦ σ−1 . To check that {Suniv λ λ λ ∈ Λ. 24 λ = λ = σd(λ) on Z(λ) now implies that, if n := d(λ), is zero off Z(λ), and µ ◦ σ−1 λ (cid:1) on Z(λ). The fact that σ−1 Suppose that f√dµ = g√dν. Observe that µ ◦ σ−1 (cid:0)mZ(s(λ)) ◦ σ−1 (f ◦ σn)qd(µ ◦ σ−1 =(cid:16)g√dν(cid:17)Z(s(λ)) ◦ σ−1 λ ) =(cid:18)(f ◦ σn)qd(µ ◦ σ−1 λ =(cid:18)(g ◦ σn)qd(ν ◦ σ−1 λ It follows that Sλ is well defined. λ )(cid:19)Z(λ) =(cid:16)fpdµ(cid:17)Z(s(λ)) ◦ σ−1 λ )(cid:19)Z(λ) = (g ◦ σn)qd(ν ◦ σ−1 λ λ ). To check the formula for S∗λ given in the statement of the proposition, we compute: hS∗λ(fpdµ), g√dνi = hfpdµ, Sλg√dνi = hfpdµ, (g ◦ σn)qd(ν ◦ σ−1 =ZΛ∞ This integral vanishes off Z(λ), since σ−1 λ fact that (ν ◦ σ−1 λ )Z(λ) = νZ(s(λ)) ◦ σ−1 λ ))s d(ν ◦ σ−1 λ ) d(µ + (ν ◦ σ−1 λ )) (and consequently d(ν ◦ σ−1 to rewrite f (x)(g ◦ σn)(x)s d(µ + (ν ◦ σ−1 dµ λ λ )i d(µ + (ν ◦ σ−1 λ )). λ )) do. We thus use the f ◦ σλ(x)g(x)s d(µ ◦ σλ) hS∗λ(fpdµ), g√dνi =ZZ(s(λ)) Hence S∗λ(f√dµ) = (f ◦ σλ)pd(µ ◦ σλ), which proves (a). d((µ ◦ σλ) + ν)s straightforward computation, analogous to the proof of Proposition 3.4. Checking condition (b), that the operators {Sλ} give a representation of C∗(Λ), is a To see (c), note that Equation (24) follows from the observation that SνS∗ν acts by multiplication by χZ(ν); the fact that disjoint unions of cylinder sets Z(ν) generate the σ- algebra up to sets of measure zero [27, Lemma 4.1] therefore enables us to compute P (A) by linearity, for any Borel set A. dν d((µ ◦ σλ) + ν) d((µ ◦ σλ) + ν). The following two Propositions, which detail additional technical properties of the pro- jection valued measure associated to πuniv, will be used in the proof of Theorem 5.7, the main result of this Section. Proposition 5.4. Let Λ be a finite k-graph with no sources and let H(Λ∞) be the Hilbert space described in Definition 5.1, and let πuniv = {Suniv : λ ∈ Λ} be the universal represen- tation of C∗(Λ) on H(Λ∞) given in Proposition 5.3. (a) For y ∈ H(Λ∞), define a function νy on Λ∞ by νy(Z(λ)) := h(Suniv )∗)y, yi, (Suniv λ λ λ where h·,·i is the inner product given on H(Λ∞) in Equation (23). Then νy gives a measure on Λ∞. (b) Let T be a bounded operator on H(Λ∞). If T commutes with πunivC(Λ∞), then for any x ∈ H(Λ∞) we have νT (x) << νx. 25 (c) For every vector f√dµ ∈ H(Λ∞), we have νf√dµ = f2µ. Proof. As in Equation (14), it is straightforward to see (a). For (b), fix x ∈ H(Λ∞). Then since T commutes with Suniv , we have λ νT (x)(Z(λ)) = h(Suniv )∗)x, T ∗T (x)i. is a partial isometry, the Cauchy-Schwarz inequality then gives )∗)T (x), T (x)i = h(Suniv (Suniv (Suniv λ Since each Suniv λ λ λ λ νT (x)(Z(λ))2 ≤ k(Suniv λ (Suniv λ )∗)xk2 kT ∗T xk2 = νx(Z(λ))2 kT ∗T xk2, which gives that νT (x) << νx. For (c), Equation (24) implies that, for any cylinder set Z(η), νf√dµ(Z(η)) = h(χZ(η)f )pdµ, fpdµi =Z χZ(η) · f2dµ =ZZ(η) f2 dµ. This gives the desired result. We now present an important result which will allow us to derive, in Theorem 5.7, the desired universal property of the representation. Theorem 5.5. Let Λ be a finite k-graph with no sources. Let H(Λ∞) be the universal Hilbert space for Λ∞ and πuniv be the universal representation of C∗(Λ) on H(Λ∞). Then: (a) An operator T ∈ B(H(Λ∞)) commutes with πunivC(Λ∞) if and only if for each finite Borel measure µ on Λ∞ which arises from a monic representation of C∗(Λ) as in Equation (14), there exists a function Fµ in L∞(Λ∞, µ) such that: (i) sup{kFµk : µ arises from a monic representation } < ∞. (ii) If µ << λ then Fµ = Fλ, µ-a.e. (iii) T (f√dµ) = Fµf√dµ for all f√dµ ∈ H(Λ∞) (b) Let H denote the subspace of H(Λ∞) spanned by vectors of the form f√dµ where µ arises from a monic representation. An operator T ∈ B(H(Λ∞)) commutes with πunivH if and only if for every finite Borel measure µ on Λ∞ arising from a monic representation of C∗(Λ), and for each λ ∈ Λ, we have Fµ = Fµ◦σ−1 λ ◦ σλ, µ − a.e. Proof. Recall from Proposition 5.2 the isometry Wµ of L2(Λ∞, µ) onto L2(µ). Throughout the proof, we will assume that the finite Borel measure µ arises from a monic representation. We first claim that if T commutes with πunivC(Λ∞), then T maps L2(µ) into itself. To prove this, let x = f√dµ be in L2(µ), and let T (x) = g√dζ for (g, ζ) ∈ H(Λ∞). Then Proposition 5.4 (b) implies that νT (x) << νx. By Proposition 5.4 (c), we have νx = f2µ, and νT (x) = g2ζ. 26 Therefore g2ζ << µ, so by the Radon–Nikodym theorem there exists h ≥ 0 in L1(Λ∞, µ) such that g2 dζ = h dµ. Then and g gpdζ = g √hpdµ. If g = 0 on some Borel set A, then √h√dµ(A) = 0 also. Therefore, √dµ ∈ L2(µ), gpdζ = √hpdµ, gpdζ =( g√h g 6= 0 g = 0 g 0, which shows that T maps L2(µ) into itself. We now make some computations regarding the relationship between an arbitrary monic representation π and the universal representation πuniv. Note that Equation (24) implies that πuniv(ψ)(f√dµπ) = (ψ · f )√dµπ for any f ∈ L2(Λ∞, µπ). On the other hand, since π is a monic representation, π(χZ(λ))f = χZ(λ) · f ∈ L2(Λ∞, µπ) by Equation (16). Therefore, πuniv(ψ)(fpdµπ) = [π(ψ)(f )]pdµπ. By hypothesis, T commutes with πunivC(Λ∞). Since T preserves L2(µ) for each measure µ arising from a monic representation, there must exist g ∈ L2(µπ) such that T (f√dµπ) = g√dµπ. Consequently, T [π(ψ)f ]pdµπ = T πuniv(ψ)(fpdµπ) = πuniv(ψ)T (fpdµπ) = πuniv(ψ)(gpdµπ) = [π(ψ)(g)]pdµπ = π(ψ)T(cid:16)fpdµπ(cid:17) , so (identifying L2(µπ) ⊆ H(Λ∞) with L2(Λ∞, µπ)) we see that T commutes with π(ψ) for all ψ ∈ C(Λ∞). the multiplication operators {Mf : f ∈ C(Λ∞)}. The fact (cf. [31]) that the maximal abelian subalgebra of B(L2(Λ∞, µ)), for any finite Borel measure µ, is the sub-algebra L∞(Λ∞, µ) too. In other words, there exists a function Fµ in L∞(Λ∞, µ) such that Therefore, we can pull-back T to an operator eT on L2(Λ∞, µ) that commutes with all of consisting of multiplication operators now implies that eT must be a multiplication operator (25) T (fpdµ) = Fµ fpdµ kFµkL∞(µ) ≤ kTk for all f ∈ L2(Λ∞, µ), establishing (iii ). It remains to check the properties of the functions Fµ. One immediately observes that and this implies (i ). To check (ii ), suppose µ << λ. Then, for all f ∈ L2(Λ∞, µ), we have f√dµ = fpdµ/dλ√dλ, and hence T (fpdµ) = T frdµ dλ √dλ! =⇒ Fµfpdµ = Fλfrdµ dλ √dλ 27 Thus, as elements of L2(λ), Fµfpdµ/dλ = Fλfpdµ/dλ for any f ∈ L2(Λ∞, µ), which implies Fµdµ/dλ = Fλdµ/dλ (λ − a.e.). It follows that, for any Borel set A, ZA (Fµ − Fλ) dµ =ZA (Fµ − Fλ) dµ dλ dλ = 0, so Fµ = Fλ, µ-a.e. This proves (ii ). For the converse, assume that T is given on L2(µ) by a function Fµ ∈ L∞(Λ∞, µ) satisfying i.e. T (f√dµ) = Fµf√dµ for all f√dµ ∈ H(Λ∞) such that µ arises from a (i ), (ii ), (iii ), monic representation. Then (i) implies that T is bounded with kTk ≤ supµ{kFµkL∞(µ)}. Since T acts as a multiplication operator on each L2(µ), Part (c) of Proposition 5.3 implies that T commutes with P (A) for all Borel subsets A and therefore T commutes with the restricted universal representation, πunivC(Λ∞), which proves (a). To prove (b), note that if an operator T ∈ B(H(Λ∞)) commutes with the universal representation πuniv of C∗(Λ) on H, then in particular T commutes with πunivC(Λ∞) on H, and hence T (f√dµ) = Fµf√dµ is a multiplication operator on each L2(µ) when the measure µ arises from a monic representation. In particular, T is normal (when restricted to H). Therefore, by the Fuglede–Putnam theorem, TH commutes with πuniv iff T Suniv H = Suniv λ TH for all λ ∈ Λ. Using the formulas for Suniv from Theorem 5.5, we see that TH commutes with πunivH if and only if, for each f√dµ ∈ H and λ ∈ Λn, λ = (Fµ ◦ σn)(f ◦ σn)qdµ ◦ σ−1 Fµ◦σ−1 or equivalently, Fµ◦σ−1 from monic representations of C∗(Λ). Composing with σλ gives the desired result of (b). Definition 5.6. A monic representation {tλ}λ∈Λ of a finite, source-free k-graph Λ is said to be nonnegative if the functions {fλ}λ∈Λ of the associated Λ-projective system on Λ∞ are nonnegative a.e. (f ◦ σn)qdµ ◦ σ−1 = (Fµ ◦ σn) for λ ∈ Λn, (µ ◦ σ−1 λ ) − a.e. for all measures µ arising λ , λ λ λ λ The following result, a consequence of Theorem 5.5, proves that every nonnegative monic }λ∈Λ, justifying the name 'uni- representation is equivalent to a sub-representation of {Suniv versal representation' for {Suniv Theorem 5.7. Let Λ be a finite k-graph with no sources. Let {tλ}λ∈Λ be a nonnegative monic representation of C∗(Λ) on L2(Λ∞, µπ). Let W be the isometry from L2(Λ∞, µπ) onto L2(µπ) given in Proposition 5.2, so that W f = f√dµπ. Then W intertwines {tλ}λ∈Λ with the sub-representation {Suniv }λ. Proof. By Theorem 4.2 and Proposition 3.4, we can assume that tλ is of the form L2(µπ )}λ∈Λ of the universal representation {Suniv }λ∈Λ. λ λ λ λ tλ(f ) = fλ · (f ◦ σd(λ)), rem 5.5, where, since {tλ}λ∈Λ is assumed nonnegative, we may assume fλ =q d(µπ◦(σλ)−1) W (tλf ) = W (fλ(f ◦ σd(λ))) = fλ(f ◦ σd(λ))pdµπ = (f ◦ σd(λ))pfλ2dµπ λ W (f ). dµπ = (f ◦ σd(λ))pd[µπ ◦ (σλ)−1] = Suniv λ (fpdµπ) = Suniv L2(µπ )}λ∈Λ, as claimed. In other words, W intertwines {tλ}λ∈Λ and {Suniv λ . By Theo- 28 References [1] D. Alpay, P.E.T. Jorgensen, and I. Lewkowicz, Markov measures, transfer operators, wavelets and multiresolutions. arXiv:1606.07692. [2] W. Arveson, An invitation to C∗-algebras. Graduate Texts in Mathematics, No. 39. Springer-Verlag, New York-Heidelberg, 1976. x+106 pp. [3] S. Bezuglyi, J. Kwiatkowski and K. Medynets, Aperiodic substitution systems and their Bratteli diagrams. Ergodic Theory Dynam. Systems 29 (2009), 37–72. [4] S. Bezuglyi, J. Kwiatkowski, K. Medynets, and B. Solomyak, Finite rank Bratteli di- agrams: structure of invariant measures. Trans. Amer. Math. Soc. 365 (2013), 2637– 2679. [5] S. Bezuglyi and P.E.T. Jorgensen, Representations of Cuntz-Krieger relations, dynamics on Bratteli diagrams, and path-space measures. Trends in harmonic analysis and its applications, Contemp. Math. 650 (2015), Amer. Math. Soc., Providence, RI, 57–88. [6] S. Bezuglyi and P.E.T. Jorgensen, Infinite-dimensional transfer operators, endomor- phisms, and measurable partitions. arXiv:1702.02657. [7] O. Bratteli and P.E.T. Jorgensen, Iterated function systems and permutation represen- tations of the Cuntz algebra. Mem. Amer. Math. Soc. 139 (1999). x+89 pp. [8] O. Bratteli, P.E.T. Jorgensen, and J. Price, Endomorphisms of B(H). Quantization, nonlinear partial differential equations, and operator algebra, Proc. Sympos. Pure Math. 59 (Cambridge, MA, 1994), Amer. Math. Soc., Providence, RI, 93–138. [9] T.M. Carlsen, S. Kang, J. Shotwell and A. Sims, The primitive ideals of the Cuntz- Krieger algebra of a row-finite higher-rank graph with no sources. J. Funct. Anal. 266 (2014), 2570–2589. [10] T.M. Carlsen, E. Ruiz, and A. Sims, Equivalence and stable isomorphism of groupoids, and diagonal-preserving stable isomorphisms of graph C∗-algebras and Leavitt path al- gebras. Proc. Amer. Math. Soc. 145 (2017), 1581–1592. [11] J. Cuntz, Simple C∗-algebras generated by isometries. Comm. Math. Phys. 57 (1977), 173–185. [12] J. Cuntz and W. Krieger, A class of C∗-algebras and topological Markov chains. Invent. Math. 56 (1980), 251–268. [13] K.R. Davidson, S.C. Power and D. Yang, Atomic representations of rank 2 graph alge- bras. J. Funct. Anal. 255 (2008), 819–853. [14] K.R. Davidson and D. Yang, Representations of higher rank graph algebras. New York J. Math. 15 (2009), 169–198. 29 [15] J. Dixmier, Les C∗-alg`ebres et leurs repr´esentations. Cahiers Scientifiques, Fasc. XXIX. ´Editions Jacques Gabay, Paris, 1964. xi+382 pp. [16] R. Durrett, Probability: Theory And Examples, second edition. Cambridge University Press, Cambridge, MA, 1996. 440 pp. [17] D.E. Dutkay and P.E.T. Jorgensen, Martingales, endomorphisms, and covariant systems of operators in Hilbert space. J. Operator Theory 58 (2007), 269–310. [18] D.E. Dutkay and P.E.T. Jorgensen, Fourier series on fractals: a parallel with wavelet theory. Radon Transforms, Geometry, and Wavelets, Contemp. Math 464 (2008), Amer. Math. Soc., Providence, RI, 75–101. [19] D.E. Dutkay and P.E.T. Jorgensen, Monic representations of the Cuntz algebra and Markov measures. J. Funct. Anal. 267 (2014), 1011–1034. [20] D.E. Dutkay, J. Haussermann, and P.E.T. Jorgensen, Atomic representations of Cuntz algebras. J. Math. Anal. Appl. 421 (2015), 215–243. [21] E. Effros, The Borel space of von Neumann algebras on a separable Hilbert space. Pacific J. Math. 15 (1965), 1153–1164. [22] E. Effros, Transformation groups and C∗-algebras. Ann. of Math. (2) 81 (1965), 38–55. [23] M. Enomoto and Y. Watatani, A graph theory for C∗-algebras. Math. Japon. 25 (1980), 435–442. [24] C. Farsi, E. Gillaspy, P.E.T. Jorgensen, S. Kang, and J. Packer, Representations of higher-rank graph C∗-algebras associated to Λ-semibranching function systems. arXiv:1803.08779. [25] C. Farsi, E. Gillaspy, A. Julien, S. Kang, and J. Packer, Wavelets and spectral triples for fractal representations of Cuntz algebras. Contemp. Math. 687 (2017), Amer. Math. Soc., Providence, RI, 103–133. [26] C. Farsi, E. Gillaspy, A. Julien, S. Kang, and J. Packer, Spectral tripes and wavelets for higher-rank graphs, arXiv:1803.09304. [27] C. Farsi, E. Gillaspy, S. Kang, and J. Packer, Separable representations, KMS states, and wavelets for higher-rank graphs. J. Math. Anal. Appl. 434 (2015), 241–270. [28] C. Farsi, E. Gillaspy, S. Kang, and J. Packer, Wavelets and graph C∗-algebras. Excur- sions in Harmonic Analysis 5 (2016), The February Fourier Talks at the Norbert Wiener Center. Edited by R. Balan, M. Begu´e, J. J. Benedetto, W. Czaja and K. A. Okoudjou. [29] J. Glimm, Locally compact transformation groups. Trans. Amer. Math. Soc. 101 (1961), 124–138. [30] J. Glimm, Families of induced representations. Pacific J. Math. 12 (1962), 885–911. 30 [31] V. Jones, Von Neumann algebras in mathematics and physics. Introduction to Modern Mathematics, 285–321, Adv. Lect. Math. (ALM) 33, Int. Press, Somerville, MA, 2015. [32] P.E.T. Jorgensen, Iterated function systems, representations, and Hilbert space. Inter- nat. J. Math. 15 (2004), no. 8, 813–832. [33] S. Kakutani, On equivalence of infinite product measures. Annals of Math. 49 (1948), 214–224. [34] S. Kang and D. Pask, Aperiodicity and primitive ideals of row-finite k-graphs. Internat. J. Math. 25 (2014), 1450022, 25 pp. [35] K. Kawamura, Pure states on Cuntz algebras arising from geometric progressions. Al- gebr. Represent. Theory 19 (2016), 1297–1319. [36] A.N. Kolmogorov, Foundations of the theory of probability. Chelsea Publishing Com- pany, New York, NY, 1950. [37] A. Kumjian and D. Pask, Higher-rank graph C∗-algebras. New York J. Math. 6 (2000), 1–20. [38] A. Kumjian, D. Pask and I. Raeburn, Cuntz-Krieger algebras of directed graphs. Pacific J. Math. 184 (1998), 161–174. [39] M. Laca, Endomorphisms of B(H) and Cuntz algebras. J. Operator Theory 30 (1993), 85–108. [40] M. Marcolli and A.M. Paolucci, Cuntz-Krieger algebras and wavelets on fractals. Com- plex Anal. Oper. Theory 5 (2011), 41–81. [41] C.C. Moore, Group extensions and cohomology for locally compact groups. III. Trans. Amer. Math. Soc. 221 (1976), 1–33. [42] E. Nelson, Topics in Dynamics. I: Flows. Math. Notes, Princeton University Press, Princeton, NJ, 1969. [43] D. Pask, I. Raeburn, and N. Weaver, A family of 2-graphs arising from two-dimensional subshifts. Ergodic Theory Dynam. Systems 29 (2009), 1613–1639. [44] D. Pask, A. Sierakowski, and A. Sims, Twisted k-graph algebras associated to Bratteli diagrams. Integral Equations Operator Theory 81 (2015), 375–408. [45] G. Robertson and T. Steger, C∗-algebras arising from group actions on the boundary of a triangle building, Proc. London Math. Soc. 72 (1996), 613–637. [46] G. Robertson and T. Steger, Affine buildings, tiling systems and higher rank Cuntz- Krieger algebras, J. Reine Angew. Math. 513 (1999), 115–144. [47] A. Skalski and J. Zacharias, Entropy of shifts on higher-rank graph C∗-algebras. Houston J. Math. 34 (2008), no. 1, 269–282. 31 [48] R.S. Strichartz, Besicovitch meets Wiener–Fourier expansions and fractal measures. Bull. Amer. Math. Soc. (N.S.) 20 (1989), 54–59. [49] R. Tumulka, A Kolmogorov extension theorem for POVMs. Letters in Math. Phys. 84 (2008), 4–46. [50] D. Yang, Endomorphisms and modular theory of 2-graph C∗-algebras. Indiana Univ. Math. J. 59 (2010), 495–520. Carla Farsi, Judith Packer : Department of Mathematics, University of Col- orado at Boulder, Boulder, CO 80309-0395, USA. E-mail address: [email protected], [email protected] Elizabeth Gillaspy : Department of Mathematics, University of Montana, 32 Campus Drive #0864, Missoula, MT 59812-0864, USA. E-mail address: [email protected] Palle Jorgensen : Department of Mathematics, 14 MLH, University of Iowa, Iowa City, IA 52242-1419, USA. E-mail address: [email protected] Sooran Kang : College of General Education, Chung-Ang University, 84 Heukseok- ro, Dongjak-gu, Seoul, Republic of Korea. E-mail address, [email protected] 32
1005.1305	1	1005	2010-05-07T22:29:21	Spectra self-similarity for almost Mathieu operators	[ "math.OA", "math-ph", "math.FA", "math-ph" ]	We determine numerically the self-similarity maps for spectra of the almost Mathieu operators, a two-dimensional fractal-like structure known as the Hofstadter butterfly. The similarity maps each have a horizontal component determined by certain algebraic maps, and vertical component determined by a Mobius transformation, indexed by a semigroup of the matrix group $GL_2(\Z)$. Based on the numerical evidence, we state and prove a continuity result for the similarity maps. We note a connection between the indexing of the similarity maps and Morita equivalence of rotation algebras $A_\theta$, a continuous field of C*-algebras.	math.OA	math	SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Abstract. We determine numerically the self-similarity maps for spectra of the almost Mathieu operators, a two-dimensional fractal-like structure known as the Hofstadter butterfly. The similarity maps each have a horizontal component deter- mined by certain algebraic maps, and vertical component determined by a Mobius transformation, indexed by a semigroup of the matrix group GL2(Z). Based on the numerical evidence, we state and prove a continuity result for the similarity maps. We note a connection between the indexing of the similarity maps and Morita equivalence of rotation algebras Aθ, a continuous field of Calgebras. 0 1 0 2 y a M 7 ] . A O h t a m [ 1 v 5 0 3 1 . 5 0 0 1 : v i X r a Contents Introduction 1. 2. Numerical evidence: Similarity maps 3. Numerical evidence: Vertical similarities from GL2(Z) 4. Numerical evidence: A symmetry with break 5. Generating the GL2(Z) symmetries 6. Horizontal similarity: Interval maps 7. Horizontal similarity: Cubic case 1 (cid:55)→ 1/3 8. Horizontal similarity: Quintic case 1/3 (cid:55)→ 1/5 9. Horizontal similarity: Algebraic curves 10. The similarity maps: General case 11. The similarity maps: Gap labelling 12. The similarity maps: Proof of continuity 13. Three generators for the butterfly similarities 14. Conclusions Acknowledgments References Appendix 1 Appendix 2 Appendix 3 Date: May 7, 2010. 1 2 4 6 10 11 12 14 16 17 20 22 23 28 29 29 30 31 33 34 2 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 1. The Hofstadter butterfly, a fractal-like structure of line spectra. 1. Introduction Looking at the image presented in Figure 1, one immediately notices the striking, repetitive pattern of "butterfly wings" that march off towards the vertical horizons at the top and bottom of the image. This rendering of the so-called Hofstadter butterfly, drawn to high-resolution using a combination of numerical algorithms and PostScript graphics programming, reveals some beautiful symmetries of a fundamentally math- ematical object. The goal of this paper is to specify exactly the symmetries of this image and prove continuity results of the corresponding similarity maps, motivated by numerical evidence collected in our study of the butterfly. SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 3 Although the Hofstadter butterfly looks like a fractal, the image is not created using the usual iterative equations or recursive graphical methods designed to visually render typical fractals (Mandelbrot, 1990). Rather, this image is created from an explicit numerical computation of spectral values of a family of linear operators on Hilbert space. More precisely, this image is constructed as a layered suite of horizontal lines which arise as the spectra of almost Mathieu operators. Each of these operators is succinctly represented as the self-adjoint element (1) hθ = u + u∗ + v + v∗ in the C-algebra Aθ generated by universal unitaries u, v satisfying the commutation rule (2) In Figure 1, the vertical axis is spanned by the parameter θ, with 0 ≤ θ ≤ 1, while the horizontal axis corresponds to spectral values x in the range −4 ≤ x ≤ 4. vu = e2πiθuv. Only rational values of θ are used in the construction of the image in Figure 1. For this reason, the image is properly called the rational Hofstadter butterfly. Extensions to irrational values of θ has been a theme in the long history of study these operators and the physical problems that motivated it, going back at least to a mathematical analysis of Bloch electrons (Brown, 1964). A selection of relevant studies over the years is given in the references, including (Avila and Jitomirskaya, 2006; Bellissard and Simon, 1982; Bellissard, 1990; Choi et al., 1990; Connes, 1994; Goldman, 2009; Hofstadter, 1976; Kaminker and Putnam, 2003; Last, 1994; Puig, 2004; Ypma, 2007). Unlike the irrational case as considered in (Arveson, 1994), rational values of θ = p/q lead to the computation of spectral lines based on finite dimensional eigen- value calculations using tridiagonal q × q matrices (Lamoureux, 1997). In particular, the endpoints of the spectral lines are specified by eigenvalues of these matrices. Numerical algorithms for the tridiagonal eigenvalue computations are rapid and ac- curate, making calculations of the rational butterfly ideal for numerical experiments. Based on these experiments, we deduce a systematic catalogue of the symmetries of the butterfly. Repetition of a geometric object fading out to infinity is characteristic of hyperbolic geometry (Coxeter, 1942), hence it is perhaps no surprise that the matrix group GL2(Z) arises in the observed symmetries. Sections 2 through 5 reveal a semigroup of linear fractional transformations represented by elements of GL2(Z) which act on the vertical parameter θ to generate the family of similarity maps. Equally important are the algebraic curves specifying the horizontal component of the similarities, acting on the horizontal parameter x. Sections 6 through 9 present numerical evidence of the continuous maps taking one level of horizontal spectra 4 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN to another, using an indexing of intervals and a correspondence of image points under polynomial maps. These are precisely the characteristic polynomials of the aforementioned finite dimensional q × q matrices giving the spectra. Uniting the vertical and horizontal components leads to an identification of the general form of the similarity maps, as presented in Section 10. The similarity extends to gap-labelling, a specific method for indexing the butterfly wings based on Chern characters, which is discussed in Section 11. Proof of continuity of the similarity maps is given in Section 12 and generators for the semigroup of similarities is presented in Section 13. "Xenocides, who is ugly, makes ugly poetry," said Aristophanes. With the Hofs- tadter butterfly, we see pretty mathematics making pretty pictures. 2. Numerical evidence: Similarity maps Mirror image symmetry in the horizontal direction suggests an obvious self-similarity of the image in Figure 1 given by a reflection about the line x = 0, namely: (3) (x, θ) (cid:55)→ (−x, θ). This map is continuous, and does map the butterfly to itself, as the spectrum of each hθ is symmetric. In the vertical direction, another mirror image symmetry in the figure suggests a second self-similarity map given by reflection about the line θ = 1/2, namely: (4) (x, θ) (cid:55)→ (x, 1 − θ). Again, this is a continuous map, and again maps spectra properly since the algebra Aθ is isomorphic to A1−θ, with operator hθ mapping onto h1−θ under the isomorphism. Now, a more interesting symmetry is observed mapping the large central butterfly onto the next largest butterfly in the bottom half of the image in Figure 1. Zooming in on this butterfly, as shown in Figure 2, one sees the top of the butterfly at θ = 1/3, the bottom at θ = 0, and the centre of the butterfly at θ = 1/4. This suggest we must find a self-similarity map that, on vertical parameter values θ, will map θ (cid:55)→ θ(cid:48) as (5) 0 (cid:55)→ 0, 1/2 (cid:55)→ 1/4, 1 (cid:55)→ 1/3. No linear map will do, but a linear fractional transformation1 does work, using the map (6) 1i.e. a Mobius transformation θ (cid:55)→ θ(cid:48) = θ 2θ + 1 . SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 5 Figure 2. Similarity to the bottom third of the butterfly. It is reassuring to notice that other obvious lines in the full image map correctly under this map. For instance, the line at θ = 1/3 in the full butterfly in Figure 1 should map to the top of the second largest butterfly in Figure 2, which is at θ(cid:48) = 1/5. And indeed, the LFT here does just that. Similarly we can check for that the line at 1/4 maps to 1/6, the line at 1/5 maps to 1/7, and so on. Checking numerically many of these lines assures us that the linear fractional transformation is the proper choice. Going to a numerical experiment based on these observations, we construct a map from the full image in Figure 1 to subset like Figure 2, using the LFT in the vertical component, and contracting linearly in the horizontal component. Specifically, we map according to the rule (7) (x, θ) (cid:55)→ (x(cid:48), θ(cid:48)) = (1 − .82 ∗ θ)x, θ 2θ + 1 (cid:18) (cid:19) . The coefficient 0.82 was chosen to get the correct width for the top horizontal line. The result is the image shown in Figure 3, which is very much like the lower butterfly image in Figure 2. Observing the slight differences between the image in Figure 1 and the result of the numerical experiment shown in Figure 3, (the difference in the curvature of the - 4401/31/41/51/61/7 1 0 2 16 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 3. A rendering of an approximate similarity to the bottom third, using a linear fractional transformation vertically, linear scaling horizontally. wingtips for instance), we can conclude the whatever the similarity map is, it is only approximately linear. Nevertheless, the basic structure and position of the tiny butterflies are preserved. This serves as numerical evidence that the vertical maps are indeed given by linear fractional transformations. The horizontal component is more complicated, and is discussed in Sections 7, 8, 9. 3. Numerical evidence: Vertical similarities from GL2(Z) Perhaps it is suggestive that two of the self-similarity maps identified so far (verti- cal flip, and the bottom third map) have vertical components give by Mobius trans- formations, (8) Both transformations are specified by a 2 × 2 matrix in GL2(Z), in the form 2θ + 1 1 θ (cid:55)→ 1 − θ and θ (cid:55)→ θ . (9) M = (cid:20) a b c d (cid:21) , yielding a corresponding linear fractional transformation (10) θ (cid:55)→ θ(cid:48) = aθ + b cθ + d . It will be convenient to label these maps by the corresponding matrix, and we note that composition of the LFTs corresponds to matrix multiplication in GL2(Z). A leap of faith suggests looking for symmetries in the butterfly indexed by elements of GL2(Z). Examining Figure 4, we see the bottom of the central core of butterflies, and we can compute a list of matrices that implement the self-similarity map on the SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 7 Figure 4. A sequence of similarities in the lower central core of the butterfly, and matrices in SL2(Z) that implement the vertical compo- nent of the similarity. vertical component. For instance, we can map the whole butterfly onto the central bottom butterfly that tops out at label θ = 1/5. For this, we need an LFT that maps 0 (cid:55)→ 0, 1/2 (cid:55)→ 1/6, 1 (cid:55)→ 1/5. Solving for coefficients a, b, c, d in the LFT, we find the matrix (cid:21) (cid:20) 1 0 4 1 (11) (12) will work to implement an LFT that maps the whole figure into this central bottom butterfly. Continuing down the central core, we see an infinite sequence of butterflies extend- ing to the "horizon" at θ = 0. Some simple calculations analogous to those above produce a sequence of similarity maps indexed by matrices of the form = A2n, for n ≥ 1. (cid:20) 1 0 2n 1 = (cid:21) (cid:21)2n (cid:20) 1 0 1 1 We will see later that this matrix A is a key generator of the similarities. Hopping to the top of butterfly, in Figure 5 we see a central core of butterflies extending to the upper "horizon" at level θ = 1. Here, the self-similarity maps are - 44010211041106110811/31/41/51/61/71/81/98 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 5. A sequence of similarities in the upper central core of the butterfly, and matrices in GL2(Z) that implement the vertical compo- nent of the similarity. obtained using matrices in GL2(Z) of the form (cid:20) 1 − 2n 2n −2n 2n + 1 (cid:21) , for n ≥ 1. Matrices from symmetries on top (Eqn 13) are related to matrices for the symme- tries on the bottom (Eqn 12), via the conjugation (cid:20) 1 − 2n 1 −2n 2n + 1 = (cid:21) (cid:20) −1 1 (cid:21)(cid:20) 1 (cid:21) (cid:20) −1 1 0 1 (cid:21)(cid:20) −1 1 (cid:21) . 0 1 0 2n 1 This is just conjugation with the flip symmetry map of the butterfly, as the matrix (15) induces the linear fractional transformation mapping θ to 1− θ, turning the butterfly upside down. B = 0 1 Attending to some of the butterflies on the side of the image, we find the relevant linear fractional transformations are represented by matrices of the form (13) (14) (16) (cid:20) (cid:21) 0 1−1 n , for n ≤ 3, 1 -1 2 -2 3 -3 4 -4 5 -5 6 -6 7 -7 8 -8 92/33/44/55/66/77/8SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 9 Figure 6. A sequence of similarities on the edge of the butterfly, and matrices in GL2(Z) that implement the similarity. which generate the symmetries on the bottom half of the image shown in Figure 6. On the top of the image, we obtain transformations indexed by matrices of the form (cid:21) 1 n − 1 (cid:20) 1 n − 2 (cid:20) −1 1 (cid:21) = 0 1 (cid:21)(cid:20) (cid:20) 1 n − 2 1 n − 1 (17) (18) Again, we observe there is a relation between these two forms of transformation, through conjugation with the flip, as we have , for n ≥ 3. (cid:21)(cid:20) −1 1 (cid:21) . 0 1 0 1−1 n 1 1 1 2 1 2 1 3 1 3 1 4 1 4 1 5 0 1-13 0 1-14 0 1-15 0 1-161/61/51/41/31/22/33/44/55/610 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Note, however, that these self-similarity maps have a certain discontinuity. In particular, the centre of the large butterfly gets mapped to a "broken" butterfly on these side maps. This tells us the self-symmetry maps need not be continuous. In particular there may be a discontinuity in the horizontal direction as we cross the x = 0 spectral value. Nevertheless, this is a mild discontinuity and is easy to understand. In the con- struction of the Hofstadter butterfly, for θ = p/q with q even, the spectrum consists of q intervals on the real line, two of which touch at the point zero. (Choi et al., 1990) It is these "touching intervals" that are getting broken apart in the above similarity maps. So although there is an apparent discontinuity, it is perhaps better to describe it as the breaking apart of two touching spectral lines. The similarity map should take the single, common endpoint, and map it to two distinct endpoints of two non- overlapping intervals. The details of this double-valued map will be discussed in Section 10. 4. Numerical evidence: A symmetry with break As noted in the previous section, under the similarity map sometimes the butterfly breaks. This suggests we can look for more similarities if we are a bit more open to what a broken butterfly looks like. In Figure 7, we have an example of a really broken butterfly. The butterfly fits and the similarity map is given by the linear fractional transformation with matrix into the region (19) (20) 0 ≤ θ ≤ 1/2, (cid:20) 1 0 1 1 (cid:21) . (cid:20) 1 0 (cid:21) This break (at θ = 1/3) looks pretty bad, but in fact we know from gap labelling theorems that when one plots the gaps an inverse slope 2, there is a discontinuity exactly at θ = 1/3 (Lamoureux, 2010). So in fact this is just confirming the fact that the butterflies are coming from gap labelling. By including this symmetry, we include the matrix A = (21) in our collection of matrices in GL2(Z) generating similarity maps. We observe that all the LFTs seen so far map the θ interval [0, 1] into itself, and are indexed by certain elements of of GL2(Z). This evidence suggests a certain semigroup in GL2(Z) specifies the possible self-similarity maps. A precise statement identifying the semigroup is given in the next section. 1 1 SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 11 Figure 7. An unusual, cracked similarity. Note the basic butterfly structure spanning the bottom half of the full image. 5. Generating the GL2(Z) symmetries The following technical result states that the set of linear fractional transforma- tions which map the interval [0, 1] into itself forms a semigroup, indexed by a 2- generated semigroup of matrices in GL2(Z). Theorem 1. Let G be the semigroup of linear fractional transformations of the form (22) θ (cid:55)→ θ(cid:48) = aθ + b cθ + d , for some ∈ GL2(Z)/{±I}, which map the interval [0, 1] into itself. Then G is generated by the two maps φ(θ) = θ/(θ + 1) and ψ(θ) = 1 − θ, represented by matrices c d (cid:21) (cid:20) a b (cid:20) −1 1 0 1 (cid:21) . (cid:21) (cid:20) 1 0 1 1 (23) A = , and B = The proof of this technical result is given in Appendix 1, and is based on the Euclidean algoritm. We make a few observations about this theorem. refers to the quotient group GL2(Z)/{±I}. The LFTs do not notice a change of sign in the representing matrix, so the theorem - 440 1 0 1 11/21/31/412 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 8. Multiple similarities, indexed by the same matrix in SL2(Z). These two generators correspond to two symmetries for the Hofstadter butterfly. The first generates the map that takes the butterfly to its lower half, mentioned in Section 4. The second is the vertical flip symmetry, mentioned in Section 2. By including the vertical flip in our symmetries, we are able to generate all (vertical components) of the symmetries using only two generating maps. The fact that the Euclidean algorithm is involved suggests that determining the details of a symmetry map will be rather involved -- the factorization into generators is not simple. It is worth mentioning a deep result from C-algebras, that states two rotation algebras Aθ, Aθ(cid:48) are Morita equivalent if and only if there is a linear fractional trans- formation mapping (24) (cid:20) a b c d θ (cid:55)→ θ(cid:48) = aθ + b cθ + d , (cid:21) for some matrix in GL2(Z) (Effros and Shen, 1980). We have no idea what connection this might have with the self-similarity maps we have above, which act on spectra of operators in the algebras Aθ, Aθ(cid:48), not on the algebras themselves. 6. Horizontal similarity: Interval maps Considering now the horizontal component of the similarity maps, a close exami- nation of the apparent symmetries suggests the intervals I1, I2, . . . , Iq in the spectrum at level θ = p/q are mapped onto a subcollection of the intervals I(cid:48) 1, I(cid:48) q(cid:48) in the spectrum at level θ(cid:48) = p(cid:48)/q(cid:48). We note here which subcollections appear. 2, . . . I(cid:48) 2/53/81/32/53/81/32/53/81/3 1 1 2 3 1 1 2 3 1 1 2 3SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 13 Examining Figure 8, notice for the single vertical map with linear fractional trans- formation given by matrix (25) (cid:20) 1 1 2 3 (cid:21) , there are in fact three choices of horizontal maps. Start first at initial θ = 1, which maps to θ(cid:48) = 2/5 under this LFT. The spectrum at level θ = 1 has one interval, while at level θ(cid:48) = 2/5, there are five intervals that we can map to. Only three are obtained. In the left part of the image in Figure 8, the initial interval I1 maps onto interval I(cid:48) 3, and the right image maps onto interval I(cid:48) (26) 1; the middle image maps onto interval I(cid:48) 5. Writing p(cid:48)/q(cid:48) = 2/5, we see that the three interval maps are given as I1 (cid:55)→ I(cid:48) I1 (cid:55)→ I(cid:48) I1 (cid:55)→ I(cid:48) 5 = I(cid:48) 1 = I(cid:48) Checking now at the level θ = 1/2 mapping to θ(cid:48) = 3/8, the two intervals I1, I2 at 5 in the middle; 8 on the right. In this case, we have p(cid:48)/q(cid:48) = 3/8 and we can summarize the 2 on the left; to the two interval I(cid:48) θ = 1/2 map to two interval I(cid:48) and I(cid:48) left, middle and right maps as 1, I(cid:48) 7, I(cid:48) 4, I(cid:48) 3 = I(cid:48) 1p(cid:48)+1, 0p(cid:48)+1, 2p(cid:48)+1. (27) Ik (cid:55)→ I(cid:48) 0p(cid:48)+k, Ik (cid:55)→ I(cid:48) 1p(cid:48)+k, Ik (cid:55)→ I(cid:48) 2p(cid:48)+k, for k = 1, 2. Similarly, at level θ = 0, the single interval I1 maps as 5 = I(cid:48) (28) where p(cid:48)/q(cid:48) = 1/3. The form of the interval map is very consistent. I1 (cid:55)→ I(cid:48) I1 (cid:55)→ I(cid:48) I1 (cid:55)→ I(cid:48) 3 = I(cid:48) 1 = I(cid:48) 0p(cid:48)+1, 1p(cid:48)+1, 2p(cid:48)+1, We go back and check the other similarity maps. In Figure 4, the interval maps are apparently of the form (29) for the vertical similarity given by (30) Ik (cid:55)→ I(cid:48) n·p+k = I(cid:48) n·p(cid:48)+k, (cid:20) 1 0 2n 1 (cid:21) . We can see this by observing the q intervals at level θ = p/q map to the central q intervals of the set of q(cid:48) = 2np + q intervals at level θ(cid:48) = p(cid:48)/q. (31) In Figure 5, the interval maps are of the form n·(q−p)+k = I(cid:48) Ik (cid:55)→ I(cid:48) for the vertical similarity given by n·(q(cid:48)−p(cid:48))+k, (cid:20) 1 − 2n 2n −2n 2n + 1 (cid:21) . (32) Again, the q intervals at level θ = p/q map to the central q intervals of the set of q(cid:48) = 2n(q − p) + q intervals at level θ(cid:48) = p(cid:48)/q. 14 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN In Figure 6, a similarity in the top half, say with matrix (33) (35) (cid:20) 1 1 1 2 (cid:21) , (cid:20) (cid:21) , 0 1−1 3 gives interval maps of the form (34) Ik (cid:55)→ I(cid:48) r·p(cid:48)+k, with r = 1 on the right side of the butterfly (as illustrated), and r = 0 on the left side of the butterfly (not shown in Figure 6). However, a similarity in the bottom half of Figure 6, such as with matrix gives interval maps of the form (36) Ik (cid:55)→ I(cid:48) r·(q(cid:48)−p(cid:48))+k, with r = 1 on the right map, r = 0 on the left map. Generally speaking, it appears there are two types of interval maps, those of the form (37) Ik (cid:55)→ I(cid:48) r·p(cid:48)+k, and those of the form (38) Ik (cid:55)→ I(cid:48) for some choices of non-negative integer r. r·(q(cid:48)−p(cid:48))+k, In the next three sections we determine how the horizontal component of the similarity map behaves within each interval Ik. 7. Horizontal similarity: Cubic case 1 (cid:55)→ 1/3 The vertical component of the self-similarity map is given by linear fractional transformations, as described in Sections 2 through 4. The horizontal component maps intervals to intervals in the spectra, and appears to be nearly linear, as we saw in the construction of Figure 3. We investigate the details of this horizontal map. Again referring to the butterfly list in Figure 1, we see that the spectral line at height θ = 1 is approached by a dense cluster of spectral points a heights θ < 1. Under the LFT mapping θ (cid:55)→ θ(cid:48) = θ/(2θ + 1), the point θ = 1 maps to θ(cid:48) = 1/3, and this dense cluster of spectral points near 1 maps to a dense cluster near 1/3. We would expect by continuity arguments that the horizontal map at θ = 1 should be well-approximated by following how the nearby cluster points map. Setting ra- tional p/q = θ < 1, we have p ≈ q, and under the LFT we have p(cid:48)/q(cid:48) = p/(2 ∗ p + q). SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 15 Figure 9. Map from θ = 1 line to 1/3 lines (middle interval). There are q points near line 1, and they map to the middle third of the 2p + q ≈ 3q points under the LFT. So we do a numerical experiment: Map these cluster points near 1 to corresponding points near 1/3, and plot the result, to obtain Figure 9. We note the curve is almost linear, but not quite. Performing a best fit polynomial approximation2, we find the map is in fact given by (39) x = 6y − y3. We observe that these polynomial maps are known from earlier work (Choi et al., 1990) as characteristic polynomials Pθ for matrices used to compute the spectra of q× q matrices representing the operators hθ in the rotation algebra Aθ, with θ = p/q. Thus, the map we observe is simply (40) where by Pp/q(x) we mean the q-th order characteristic polynomial of the q×q matrix (41) H = U + U∗ + V + V ∗, P1(x) = −P1/3(y), with U the cyclic permutation matrix and V a diagonal matrix with consecutive powers of e2πip/q on the diagonal. Details are in (Choi et al., 1990; Lamoureux and Mingo, 2007). To see the full algebraic curve, we plot the full curve in Figure 10. Observe the central portion of the curve corresponds to the nearly linear section we saw in Figure 9. 2In fact, in our first numerical experiments, we inadvertently computed the inverse of a polyno- mial, with expansion y = x/6 + x3/1296 + x5/93312 + ··· . The second author recognized this as the inverse of a cubic! -4-2024-2-101216 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 10. Algebraic curve P1(x) = −P1/3(y) defines the 1 to 1/3 map. Figure 11. Map from 1/3 lines to 1/5 lines (3 middle intervals). 8. Horizontal similarity: Quintic case 1/3 (cid:55)→ 1/5 We repeat the numerical calculation, using the same central symmetry with LFT given by the matrix (42) (cid:20) 1 0 2 1 (cid:21) . But now we look how the three spectral lines at θ = 1/3 maps to the three central spectral lines at level θ(cid:48) = 1/5. A numerical experiment as in the last section produces the plot in Figure 11. Again, the three segments look nearly linear, but not quite. An inspired guess suggest these segments come from a portion of the algebraic curve (43) P1/3(x) = −P1/5(y), and a comparison of the two plots shows this is indeed the case. -6-4-20246-4-2024-2-112-1-0.50.51SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 17 Figure 12. Algebraic curve P1/3(x) = −P1/5(y) defines the 1/3 to 1/5 map. Note the three nearly linear segments across the centre diagonal, matching Figure 11. It is worth noting that we can restrict the algebraic curve to the segments shown in Figure 12 by plotting only those points (x, y) such that P1/3(x) = −P1/5(y) and P1/3(x) ≤ 4. 9. Horizontal similarity: Algebraic curves We have seen in Sections 7 and 8 that the horizontal maps appear as nearly linear sections of algebraic curves, in the form Pθ(x) = ±Pθ(cid:48)(x(cid:48)) ≤ 4, for the two characteristic polynomials Pθ and Pθ(cid:48), where θ, θ(cid:48) are rational. The full algebraic curves are interesting in themselves, so in this section we present a few plots of the curves and point out some obvious patterns. Referring to Figure 13, we see the algebraic curves Pθ(x) = Pθ(cid:48)(y) for rationals θ = 1/q (cid:55)→ θ(cid:48) = 1/(q + 2), with q odd, have a particularly simple graphs. There is a single connected component, that spirals around the origin, with more spirals as the denominator q increases. We see q nearly linear segments, which are the parts of the graph that define the relevant horizontal map on the fractal. We also see the odd symmetries of the curves, that point (x, y) is in the curve if and only if (−x,−y) is in it too. -4-2024-4-202418 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 13. A sequence of related algebraic curves, for rationals: 1 (cid:55)→ 1/3, 1/3 (cid:55)→ 1/5, 1/5 (cid:55)→ 1/7, 1/7 (cid:55)→ 1/9, 1/9 (cid:55)→ 1/11. Notice the odd number of nearly linear segments down the diagonal of each map. Referring to Figure 14, we see the maps for rationals 1/q (cid:55)→ 1/(q + 2), with q even, have somewhat more complex graphs. There are q/2 connected components, that form something like figure eights around the origin, with more figure eights as the denominator q increases. We see q nearly linear segments, which are the parts of the -6-4-20246-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 19 Figure 14. A sequence of related algebraic curves: 1/2 (cid:55)→ 1/4, 1/4 (cid:55)→ 1/6, 1/6 (cid:55)→ 1/8, 1/8 (cid:55)→ 1/10. Notice the symmetry, and the even number of nearly linear segments down the diagonal of each map. graph that define the relevant horizontal map on the fractal. We also see the even symmetries of the curves, that point (x, y) is in the curve if and only if the points (±x,±y) are in it too. Figure 15 examines the curves for rational θ of the form θ = p/3. In both cases (p = 1/3, 2/3), there is odd symmetry, there are three nearly linear segments, and some spiraling of the curve about the origin. But for the case θ = 2/3, there are also two additional disconnected components to the curve, that seem to have nothing to do with the linear segments that are involved in the fractal map. In Appendix 3, we include more plots of these algebraic curves. We note with even denominator q, we have even symmetry in the plots, there are q nearly linear segments, and we see nested "figure eights" defining the key parts of the curve. For p > 1, there are disconnected components of the curve that seem to have nothing to do with the main part of the curve that defines the fractal map. -4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-202420 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 15. Algebraic curves with denominator 3: 1/3 (cid:55)→ 1/5, 2/3 (cid:55)→ 2/7. Notice the the three nearly linear segments down the diagonal of each map, and the disconnected components on the right graph. We also observe the curves appear to get much more complex with increases in the numerator p, for the parameter θ = p/q. There may be some connection between the various straight line segments in curves of different θ = p/q with the same denominator q, but these patterns are not entirely clear. We leave discussion of these curves and their connection with symmetries to future work. 10. The similarity maps: General case The work in Sections 7 through 9 suggests an obvious candidate for the horizontal maps within spectral intervals: namely, a correspondence of points determined by two characteristic polynomials. Combining this observation with the earlier sections discussing the vertical maps and interval mapping, we can now specify the form of the general similarity maps of the butterfly, as follows: Fix a matrix element M in GL2(Z), represented as (cid:21) (cid:20) a b c d (44) Fix an integer r ≥ 0. The similarity map S = SM,r,+ on the rational butterfly is described as M = . (45) (x, θ) (cid:55)→ S(x, θ) = (x(cid:48), θ(cid:48)), -4-2024-4-2024-4-2024-4-2024SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 21 where the vertical component of the map is given by (46) θ (cid:55)→ θ(cid:48) = aθ + b cθ + d , x ∈ Ik (cid:55)→ x(cid:48) ∈ I(cid:48) r·p(cid:48)+k = Ik(cid:48) with (−1)q+kPθ(x) = (−1)q(cid:48)+k(cid:48) while the horizontal component is given by the interval and polynomial correspon- dences. Specifically, recall the spectrum at level θ = p/q is a set of q intervals Ik, k = 1...q, while at level θ(cid:48) = p(cid:48)/q(cid:48) there are q(cid:48) intervals I(cid:48) k(cid:48), k(cid:48) = 1...q(cid:48). The map from x at level θ to x(cid:48) at level θ(cid:48) is determined by the interval and polynomial conditions (47) The sign choice for the polynomials is made so that ±Pθ(x) and ±Pθ(cid:48)(x(cid:48)) are both monotonically increasing on Ik, I(cid:48) k(cid:48), respectively. These two conditions determine the image point x(cid:48) uniquely, except possibly in one special case. Except for this case (described next), we have determined the similarity maps precisely. The special case to consider is when x = 0, θ = p/q with q even. In this case, the two intervals Iq/2, Iq/2+1 overlap at x = 0 and may map to two disjoint intervals I(cid:48) k(cid:48) and I(cid:48) k(cid:48)+1. Here, we must decide where the point x = 0 maps to, either an endpoint k(cid:48) or to an endpoint of I(cid:48) of I(cid:48) k(cid:48)+1. It is convenient to chose to map to both points, making the function S double-valued at this point (x = 0, θ = p/q). We will see in the next section that the double-valued function is continuous. Pθ(cid:48)(x(cid:48)). There is another set of similarity maps, S = SM,r,−, which differs from the above example by the way the interval maps are selected. For this maps, we chose (48) x ∈ Ik (cid:55)→ x(cid:48) ∈ I(cid:48) r·(q(cid:48)−p(cid:48))+k = Ik(cid:48), which we observed in Section 6 as similarities that can occur. Of course, not all such maps SM,r,± necessarily appear as similarities of the rational butterfly. For instance, from Theorem 1, we know the matrix M must be chosen from the semigroup G ⊂ GL2(Z)/±I described in Theorem 1. The integer r must be small enough that r · p(cid:48) + q ≤ q(cid:48) r · (q(cid:48) − p(cid:48)) + q ≤ q(cid:48) r ≤ min θ cθ + d − 1 aθ + b r ≤ min θ cθ + d − 1 (c − a)θ + d − b in the " -- " case, for all θ = p/q. Equivalently, we have (49) in the "+" case, and (50) (51) in the "+" case, and (52) 22 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN in the " -- " case. But these are the only apparent restrictions on M, r,±. 11. The similarity maps: Gap labelling The spectral gaps forming the butterfly are conveniently labeled using a Diophan- tine equation, where integer parameters (s, t) are fixed, and the k-th gap in the spectrum θ = p/q is identified using the formula k = t ∗ p − s ∗ q. (53) This indexing is known as gap labelling, as described in (Bellissard, 1990; Goldman, 2009; Kaminker and Putnam, 2003; Ypma, 2007), where the parameters (s, t) are related to Chern numbers. These parameters give a convenient way of labelling the "wings" in the butterfly. Note there is a limited range on the indices: for t > 0 we require (54) while for t < 0 we require (55) 0 ≤ s ≤ t − 1, t ≤ s ≤ −1. It is easy to check that the similarity map S = SM,r,± maps labeled gaps to gaps, as stated in the following: Theorem 2. The similarity map S = SM,r,+ maps the gap labeled (s, t) to the gap (s(cid:48), t(cid:48)) according to the formula Also, the similarity map S = SM,r,− maps the gap labeled (s, t) to the gap (s(cid:48), t(cid:48)) according to the formula (cid:21) (cid:21) t(cid:48) (cid:20) s(cid:48) (cid:20) s(cid:48) t(cid:48) (56) (57) r (cid:20) 0 (cid:20) r r (cid:21) (cid:21) . . + + c d (cid:20) a b (cid:20) a b (cid:20) a b c d t (cid:21)(cid:20) s (cid:21)(cid:20) s (cid:21)(cid:20) p t (cid:21) (cid:21) (cid:21) = (ad − bc) = (ad − bc) (cid:20) p(cid:48) (cid:21) Proof: We consider S = SM,r,+. Note the matrix M in the similarity determines the map from angle θ = p/q to θ(cid:48) = p(cid:48)/q(cid:48) according to the matrix formula (58) In the case where the determinant ad − bc is one, we invert to find (59) p = (dp(cid:48) − bq(cid:48)) and q = (−cp(cid:48) + aq(cid:48)). c d = q(cid:48) q . SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 23 k(cid:48) = r · p(cid:48) + k The index r of the similarity tells us how intervals Ik are mapped to intervals Ik(cid:48), where k(cid:48) = r · p(cid:48) + k. Thus the k-th gap is mapped to gap k(cid:48) = r · p(cid:48) + k. With gap k = t ∗ p − s ∗ q, we write (60) (61) (62) (63) (64) from which we read off s(cid:48) = (as + bt) and t(cid:48) = (cs + dt + r), as desired. The other cases (negative determinant, and S = SM,r,−) are similar. = r · p(cid:48) + t ∗ p − s ∗ q = r · p(cid:48) + t(dp(cid:48) − bq(cid:48)) − s(−cp(cid:48) + aq(cid:48)) = (cs + dt + r)p(cid:48) − (as + bt)q(cid:48) = t(cid:48) ∗ p(cid:48) − s(cid:48) ∗ q(cid:48), QED 12. The similarity maps: Proof of continuity We show first continuity along the horizontal direction for the similarity maps described Section 10: Theorem 3. Given two rational parameters θ = p/q, θ(cid:48) = p(cid:48)/q(cid:48), and a correspon- dence between points in pairs of intervals Ik, I(cid:48) k(cid:48) in the spectra Spec(hθ), Spec(hθ(cid:48)) respectively, given by (65) x ∈ Ik (cid:55)→ x(cid:48) ∈ Ik(cid:48) = I(cid:48) r·p(cid:48)+k with (−1)q+kPθ(x) = (−1)q(cid:48)+k(cid:48) Pθ(cid:48)(x(cid:48)), then the correspondence is a continuous bijection on each interval. In particular, this yields a continuous map from the set of points x ∈ Spec(hθ) into x(cid:48) ∈ Spec(hθ(cid:48)), with the possible exception at the point x = 0, when q is even. Proof: The polynomial maps (−1)q+kPθ(x), (−1)q(cid:48)+k(cid:48) Pθ(cid:48)(x(cid:48)) are both continuous and monotonically increasing on the intervals in question, and thus have continuous inverses. So on each interval, we have a continuous bijection. For q odd, the spectrum Spec(hθ) is a disjoint union of the intervals I1...Iq, and so the collection extends to a continuous map into Spec(hθ(cid:48)). In the case where q is even, two of the intervals Iq/2 and Iq/2+1 overlap at the point x = 0. So, except at this point, the correspondence extends to a continuous map on the union of intervals -- that is, it is continuous at Spec(hθ)\0. At the point x = 0, the correspondence maps the one point to (possibly) two points, one at the right endpoint of I(cid:48) r·p(cid:48)+q/2 and the other at the left endpoint of I(cid:48) r·p(cid:48)+q/2+1. This double-valued splitting (although not a singled-valued function) QED is nevertheless continuous. Figure 16 indicates graphically how the bijection works on level θ = 1/3 to θ = 1/5. We note that the correspondences of the form (66) x ∈ Ik (cid:55)→ x(cid:48) ∈ Ik(cid:48) = I(cid:48) r·(q(cid:48)−p(cid:48))+k with (−1)q+kPθ(x) = (−1)q(cid:48)+k(cid:48) Pθ(cid:48)(x(cid:48)), 24 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 16. Plot of cubic polynomial P1/3(x) and quintic P1/5(x(cid:48)). Match the zero crossing, move polynomials up and down, and observe a continuous correspondence of the zero crossings. This gives the con- tinuous correspondence between line spectra. also are continuous, as in the theorem. This covers the similarity maps of the form SM,r,−. The numerical work suggests that piecing together these individual polynomial maps, at the different θ levels, gives a continuous map on the whole butterfly. We now state and prove this as a theorem: Theorem 4. Suppose S = SM,r,± is a similarity of the rational butterfly, as described in Section 10. Then S is a continuous (single-valued) map, except possibly at the points (x = 0, θ = p/q) with q even, where the map is double-valued and continuous. Proof: We consider the case S = SM,r,+. Fix a sequence of points (xn, θn) converg- ing to point (x∗, θ∗) and set (67) By continuity of the linear fractional transformation, the sequence θ(cid:48) n converges to θ(cid:48) ∗ = (aθ∗ + b)/(cθ∗ + d). To show continuity of S, we only need to verify convergence of the x(cid:48) n, θ(cid:48) n) converges to the point (x(cid:48) ∗) = S(x∗, θ∗), which, by compactness, establishes conti- nuity. n. We will show every convergent subsequence of the image points (x(cid:48) n) = S(xn, θn). (x(cid:48) n, θ(cid:48) ∗, θ(cid:48) -4-224-100-5050-50SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 25 n, θ(cid:48) n converges to x(cid:48) n) with limit (x(cid:48) n = θ(cid:48) Take a convergent subsequence of the (x(cid:48) The other case to consider is when there are at most finitely many n with θ(cid:48) o, θ(cid:48) ∗). Suppose in this subsequence, there are infinity many points where θ(cid:48) ∗ (and consequently, θn = θ∗). Applying Theorem 2, by continuity along the spectral lines at θ∗, we can conclude that x(cid:48) ∗. (Except possibly at x∗ = 0, θ∗ = p/q, with q even. Here, the map S may be doubly valued, as discussed earlier.) n = θ(cid:48) ∗. By restricting to the tail of the subsequence, and renumbering the subsequence, WLOG we may assume θn (cid:54)= θ∗ for all n. and points xn ∈ Ikn, x∗ ∈ Ik∗, while in the range we have θ(cid:48) points x(cid:48) and k(cid:48) Set the notation consistently, so in the domain we have θn = pn/qn, θ∗ = p∗/q∗, ∗, and n + kn n, θ(cid:48) ∗ = p(cid:48) ∗/q(cid:48) n = r· p(cid:48) . From the definition of S, we know k(cid:48) o is to be determined (it will equal k(cid:48) ∗). n ∈ I(cid:48) , x(cid:48) k(cid:48) ∗ = r · p(cid:48) ∗ + k∗. The value of k(cid:48) o ∈ I(cid:48) k(cid:48) ∗ ∈ I(cid:48) k(cid:48)∗, x(cid:48) n = p(cid:48) n/q(cid:48) We apply the trace τθn to the spectral projection χ(−∞,xn)(hθn). By Lemma 7 in n o Appendix 2, we have (68) τθn(χ(−∞,xn)(hθn)) = kn − 1 qn + 1 qn F ((−1)kn+qnPθn(xn)), where the function F is an integrated density of states on ho. By continuity of the trace and spectral projection (see (Boca, 2001), Chap. 5), this quantity converges to (69) τθ∗(χ(−∞,x∗)(hθ∗)) = k∗ − 1 q∗ + 1 q∗ F ((−1)k∗+q∗Pθ∗(x∗)). Since the sequence of ratios pn/qn (cid:54)= p∗/q∗ converges to p∗/q∗, we have that qn converges to infinity, so the terms above with 1/qn in them converge to zero. Thus, from the limit of the traces, we conclude (70) convergent sequence x(cid:48) lim n→∞ kn qn n → x(cid:48) k∗ − 1 q∗ = F ((−1)k∗+q∗Pθ∗(x∗)). + 1 q∗ Similarly, by applying the trace to the sequence of spectral projections on the (71) converges to (72) Thus (73) τθ(cid:48) n(χ(−∞,x(cid:48) n)(hθ(cid:48) o, we find n − 1 k(cid:48) q(cid:48) n)) = n + τθ(cid:48)∗(χ(−∞,x(cid:48) o)(hθ(cid:48)∗)) = o − 1 k(cid:48) q(cid:48)∗ + F ((−1)k(cid:48) n+q(cid:48) n(x(cid:48) n)), nPθ(cid:48) F ((−1)k(cid:48) o+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) o)). 1 q(cid:48) n 1 q(cid:48)∗ k(cid:48) q(cid:48) n n o − 1 k(cid:48) q(cid:48)∗ + 1 q(cid:48)∗ = lim n→∞ F ((−1)k(cid:48) o+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) o)). 26 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN With (74) k(cid:48) q(cid:48) n n r · p(cid:48) n + kn q(cid:48) n = = r · (apn/qn + b) + kn/qn cpn/qn + d we take limits to obtain the identity (75) lim k(cid:48) q(cid:48) n n = r · θ(cid:48) ∗ + 1 cθ∗ + d lim kn qn . Replacing with the limiting values computed above, we obtain (76) o − 1 k(cid:48) q(cid:48)∗ o)) = r·θ(cid:48) ∗+ o+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) F ((−1)k(cid:48) cθ∗ + d 1 q(cid:48)∗ 1 q∗ q∗ + + 1 (cid:18) k∗ − 1 (cid:19) F ((−1)k∗+q∗Pθ∗(x∗)) . Multiplying through by q(cid:48) o − 1 + F ((−1)k(cid:48) k(cid:48) (77) ∗ and simplifying yields o+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) o)) = r · p(cid:48) ∗ + k∗ − 1 + F ((−1)k∗+q∗Pθ∗(x∗)). Now, if x∗ is an interior point of the interval Ik∗, this last equation has F taking a fractional value strictly between zero and one. Equating the integer part of the equation gives (78) from which we conclude k(cid:48) point x(cid:48) ∗ are in the same spectral interval I(cid:48) k∗. o − 1 = r · p(cid:48) k(cid:48) ∗, which tells us that the limit point x(cid:48) o = k(cid:48) ∗ + k∗ − 1, Equating the fractional part of the equation gives o and image F ((−1)k(cid:48)∗+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) F ((−1)k(cid:48)∗+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) o)) = F ((−1)k∗+q∗Pθ∗(x∗)), (79) which, from the polynomial correspondence between x∗ and x(cid:48) o)) = F ((−1)k(cid:48)∗+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) (80) ∗)). Since F , composed with polynomial (−1)k(cid:48)∗+q(cid:48)∗Pθ(cid:48)∗(x) is strictly increasing on the interval I(cid:48) o must be the corresponding endpoint of interval I(cid:48) limit point x(cid:48) left endpoint of interval Ik: in this case, equation 77 reduces to Now, when x∗ is at an endpoint of interval Ik, we use gap labeling to show the k∗. Suppose x∗ is the k∗, the equality implies x(cid:48) ∗, as desired. o = x(cid:48) ∗, yields o − 1 + F ((−1)k(cid:48) k(cid:48) (81) so we know that k(cid:48) o equals either k(cid:48) o)) = r · p(cid:48) ∗. We just need to show it is equal to k(cid:48) ∗. With the point x∗ (cid:54)= 0 at a left endpoint, there must be a gap to the immediate left of x∗, which has some label (s, t). By continuity of gap labeling (Prop 11.11 in (Boca, 2001)), the gap with this label forms an open set, so for n sufficiently large, the point xn is to the right of this gap. Hence the interval Ikn is also to the o+q(cid:48)∗Pθ(cid:48)∗(x(cid:48) ∗ − 1 or k(cid:48) ∗ + k∗ − 1 + 0. SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 27 right of the gap. Since the interval number must be bigger than the gap number, we have kn > t ∗ pn − s ∗ qn. kn and its index satisfies (82) kn is in interval I(cid:48) The image point x(cid:48) n = r · p(cid:48) k(cid:48) n + kn n + t ∗ pn − s ∗ qn > r · p(cid:48) = r · p(cid:48) n + t ∗ (dp(cid:48) n − bq(cid:48) n) − s ∗ (−cp(cid:48) n − (as + bt) ∗ q(cid:48) = (cs + dt + r) ∗ p(cid:48) = t(cid:48) ∗ p(cid:48) (83) (84) (85) (86) (87) where the indices (s(cid:48), t(cid:48)) label the image gap, as in Theorem 2. Thus the interval I(cid:48) k(cid:48) is to the right of the gap (s(cid:48), t(cid:48)) and hence so is the point x(cid:48) n. Since this gap is open, the limit point x(cid:48) ko containing point x(cid:48) o is to the right. Consequently, we have o is to the right of the gap, and thus the interval I(cid:48) n − s(cid:48) ∗ q(cid:48) n n + aq(cid:48) n) n n ∗ − 1. ∗ − s(cid:48)q(cid:48) k(cid:48) o > t(cid:48)p(cid:48) (88) This eliminates the possibility that k(cid:48) o = k(cid:48) which we quickly conclude that the point x(cid:48) thus is equal to x(cid:48) ∗. common endpoint of two overlapping intervals shows the image sequence x(cid:48) converge to endpoints of two different, disjoint intervals. Handling a right endpoint is similar. A similar analysis of the case x∗ = 0 at the n could QED ∗ = k(cid:48) ∗ − 1, so we are left with k(cid:48) o = k(cid:48) o is the left endpoint of interval I(cid:48) ∗, from k∗ and It would seem the result on the rational butterfly should extend by continuity to the full butterfly, including irrational values for vertical parameter θ. We state this precisely as a conjecture. Conjecture 5. Given a similarity S = SM,r,± of the rational butterfly, as described in Section 10, there is a unique continuous extension of S to the full butterfly, possibly double-valued at certain points along the vertical line x = 0. We do not have a proof of this result. It seems a density argument and use of the continuous trace on the field of rotation algebras Aθ should lead to the result. It is curious to consider how the polynomial correspondences on the rational spec- tral lines might extend to irrational values of θ, in which case there are no finite polynomials to describe the mapping. 28 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN 13. Three generators for the butterfly similarities There are three similarity maps of the butterfly which apparently generate all the similarities discussed above. There is the horizontal flip H defined as the map There is the vertical flip V defined as the map (89) (90) (91) with matrix B = (cid:21) (cid:20) −1 1 0 1 H(x, θ) = (−x, θ). V (x, θ) = (x, 1 − θ), V = SB,0,+, which can be represented in the form discussed in Section 10 as . Finally, there is the similarity map discussed in Sec- tion 4, taking the butterfly to the bottom half given by S(x, θ) = (x(cid:48), θ (92) where x ∈ Ik in the k-th interval at level θ is mapped to x(cid:48) ∈ I(cid:48) at level θ(cid:48). This map is represented as (93) S = SA,0,+, θ + 1 ), (cid:20) 1 0 1 1 (cid:21) . for matrix A = k in the k-th interval Some elementary calculations show how these three similarities combine alge- braically. First, there are the two obvious identities V 2 = I, H 2 = I, (94) and the commutation relation (95) HV = V H. The horizontal flip almost commutes with similarity S; in fact it just gives a shift in the interval indexing as follows: (96) HSA,0,+H = SA,1,+. This shifting extends to powers of S, so we find (97) (HSA,0,+H)n = SAn,n,+. By composing powers of S and HSH, we obtain all similarities of the form (98) SAn,r,+ for 0 ≤ r ≤ n. We can also verify that such operators combine in the form ,r+r(cid:48),+. ,r(cid:48),+)(SAn,r,+) = SAn+n(cid:48) (SAn(cid:48) (99) SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 29 The vertical flip does not commute with similarity S, but instead introduces the similarity SM,r,− in the mix. Again, a straightforward calculation shows (100) V SAn,r,+V = SBAnB,r,−. As noted in Theorem 1, the matrices A, B generate a large semigroup of linear fractional transformations. From the calculations noted above, the three similarities of the butterfly generate enough similarities to cover this semigroup. Theorem 6. The three similarities H, V, S of the rational butterfly generate a semi- group of continuous (possibly double-valued) similarities of the butterfly, including maps of the form (101) where the matrices M range over the semigroup of elements of GL2(Z)/ ± I repre- senting linear fractional transformations mapping the interval [0, 1] into itself. SM,r,+ and SM,r,− The double-valued character is, of course, from the interval splitting in the case of θ = p/q with q even. There may be other similarities of the butterfly. The point is, we can at least see this large semigroup from GL2(Z) appearing. 14. Conclusions The Hofstadter butterfly, representing spectra of a continuous family of almost Mathieu operators, shows obvious fractal-like symmetry. By investigating these sym- metries numerically, we have catalogued the self-similarity maps of the butterfly using a semigroup of Mobius transformations in the vertical direction, indexed by elements in the matrix group GL2(Z). This semigroup is generated by two matrices. In the horizontal direction, the self-similarity maps are given by algebraic curves determined by characteristic polynomials. Properties of the algebraic curves are demonstrated in a series of plots. These algebraic curves show nearly linear segments, demon- strating an almost linear behaviour in the horizontal component of the self-similarity maps. We proved continuity of the similarity maps on the rational butterfly, possibly double-valued at points with horizontal parameter x = 0. The similarity maps are generated by exactly three continuous symmetries. We conjecture the semigroup of similarities on the rational butterfly extends to a family of continuous similarities on the full butterfly. Acknowledgments This work was supported in part by NSERC Discovery grants of the first two authors, and an NSERC Summer Research award of the third author. Numerical calculations were done in MATLAB (TheMathWorks, 2008) and rendered directly 30 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN in PostScript, following a method similar to those discussed in (Casselman, 2005). Algebraic curves were rendered using Mathematica (WolframResearch, 2002). We would like to thank the organizers of the 2006 BIRS Workshop on Operator Methods in Fractal Analysis, Wavelets, and Dynamical Systems, for encouraging us to present this work in an early form. References Arveson, W. 1994. C-algebras and numerical linear algebra, J. Funct. Anal. 122, no. 2, 333 -- 360. Avila, A. and S. Jitomirskaya. 2006. Solving the ten martini problem, Lect. Notes. in Physics 690, 5 -- 16. Bellissard, J. 1990. Gap labelling theorems for Schrodinger operators, From number theory to physics, pp. 140 -- 150. Bellissard, J. and B. Simon. 1982. Cantor spectrum for the almost Mathieu equation, Journal of Functional Analysis 48, no. 3, 408 -- 419. Boca, F. 2001. Rotation c-algebras and almost Mathieu operators, Fundatia Theta. Brown, E. 1964. Bloch electrons in a uniform magnetic field, Phys. Rev. 133, no. 4A, A1038 -- A1044. Casselman, B. 2005. Mathematical Illustrations, Cambridge University Press. Choi, M. D., G. A. Elliott, and N. Yui. 1990. Gauss polynomials and the rotation algebra, Invent. Math. 99, 225 -- 246. Connes, A. 1994. Noncommutative geometry, Academic Press. Coxeter, H.S.M. 1942. Non-euclidean geometry, University of Toronto Press, Toronto. Effros, E. and C. Shen. 1980. Approximately finite C-algebra and continued fractions, Indiana Univ. Math. J. 12, no. 2, 191 -- 204. Goldman, N. 2009. Characterizing the Hofstadter butterfly's outline with Chern numbers, J. Phys. B: At. Mol. Opt. Phys. 42, no. 055302. Hofstadter, D. R. 1976. Energy levels and wave functions of Bloch electrons in rational and irra- tional magnetic fields, Physical Review B 14, no. 6, 2239 -- 2249. Kaminker, J. and I. Putnam. 2003. A proof of the gap labeling conjecture, Michigan Math J. 51, 537 -- 546. Lamoureux, M. P. 1997. Reflections on the almost Mathieu operator, Integr. equ. oper. theory 28, no. 1, 45 -- 59. . 2010. Drawing butterflies from the almost Mathieu operator. preprint. Lamoureux, M. P. and J. A. Mingo. 2007. On the characteristic polynomial of the almost Mathieu operator, Proc. AMS 135, 3205 -- 3215. Last, Y. 1994. Almost everything about the almost Mathieu operator, Proceedings of the XI-th International Congress of Mathematical physics. Mandelbrot, Benoit B. 1990. The fractal geometry of nature, Spektrum Akademischer Verlag. Puig, Joaquim. 2004. Cantor spectrum for the almost Mathieu operator, Comm. in Math Phys 244, no. 2, 297 -- 309. TheMathWorks. 2008. MATLAB R2008a. WolframResearch. 2002. Mathematica 4.2. Ypma, F. 2007. K-theoretic gap labeling for quasicrystals, Contemporary mathematics: Geometric and topological methods for quantum field theory., pp. 247. (cid:20) a b (cid:21) (cid:20) 1 0 (cid:21) SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 31 Appendix 1 Proof of Theorem 1: We show the corresponding semigroup in GL(2, Z)/{±I} is generated by the two matrices A and B. Suppose the matrix (102) is in the semigroup. In the case that b = 0, the determinant condition gives a = ±d = ±1; the condition that the LFT maps [0, 1] into [0, 1] reduces WLOG to a = d = 1 and c ≥ 0. This leave matrix M in the form M = c d = Ac, c 1 M = (103) with c ≥ 0. Hence the matrix M is generated by A alone. In the case that b (cid:54)= 0, WLOG (by mod-ing out by ±I), we can assume d ≥ b > 0 as we know that 0 maps to b/d ∈ [0, 1] under the LFT. By the determinant condition ad − bc = ±1, we have that the gcd of b, d is one, so we may apply the Euclidean algorithm to obtain a sequence of strictly positive quotients q0, q1, q2, . . . and remainders b = r0, r1, r2, . . . with d = q0r0 + r1 b = q1r1 + r2 r1 = q2r2 + r3 . . . rk−1 = qkrk + rk+1, where rk = gcd(b, d) = 1 and rk+1 = 0. Applying these values to matrix M , we can factor as for some appropriate values of a0, a1, . . . , c0, c1, . . .. Now, the matrix factors with the qj are generated by the A, B matrices, since (104) = Aq−1BA. (cid:21) (cid:21)(cid:20) a1 r2 (cid:20) 0 (cid:21) c1 r1 ··· (cid:21) (cid:21)(cid:20) ak 0 (cid:21) ck 1 , 1 1 qk (cid:21) (cid:20) a b c d = = = (cid:20) 0 (cid:20) 0 (cid:20) 0 1 1 q0 1 1 q0 1 1 q0 (cid:21)(cid:20) a0 r1 (cid:21)(cid:20) 0 (cid:21)(cid:20) 0 (cid:21) (cid:20) 0 1 c0 r0 1 1 q1 1 1 q1 1 q 32 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN In the last factor, (105) we know that ck is non-negative, else the LFT has a pole in [0, 1], which is not allowed. By the determinant condition, we have ak = ±1. In the case ak = 1, we have that the last factor appears as (106) In the case where ak = −1, we may combine the last two factors to observe ck 1 = Ack. (cid:20) 0 1 (cid:21)(cid:20) −1 0 (cid:21) 1 q c 1 (cid:20) ak 0 ck 1 (cid:21) , (cid:21) (cid:20) 1 0 (cid:20) 0 1 (cid:20) 0 1 q = (cid:21)(cid:20) 1 −1 (cid:21)(cid:20) c − 1 1 (cid:21)(cid:20) c − 1 1 (cid:21) c 1 0 1 (cid:21) 1 1 q − 1 = = (Aq−2BA)(BAc). c 1 (107) (109) For q ≥ 2, we are done: these last factors appear as generated by matrices A, B, as desired. In the case q = 1, it easy to check that we must have more than one q-type matrix in the factorization, for if not, we get (cid:21) (cid:20) a b c d (cid:20) 0 = 1 1 q0 (cid:21)(cid:20) −1 0 (cid:21) , c0 1 where q0 = 1, giving an LFT mapping 1 to (c + 1)/c, which is outside the interval [0, 1]. With a second q-type matrix in the factorization, we get (108) M = ··· (Aq−1BA)(A−1BA)(BAck) = ··· (Aq−1B)(AA−1)(BA)(BAck) = ··· AqBAck, QED which puts M in the semigroup generated by A, B. Here is a nice example, to show that the ak = −1 case really does appear. (cid:21) (cid:20) 2 1 3 2 = (cid:20) 0 1 (cid:21)(cid:20) −1 0 (cid:21) 1 2 2 1 The matrix on the left gives an LFT which maps 0 to 1/2, and 1 to 3/5. This simi- larity map actually appears in the Hofstadter butterfly, as we can see by examining Figure 1. We get a1 = −1 in the factorization, which perhaps was not expected. Thus that little factor matrix on the right is not in the semigroup, since it maps 1 to -1/3, and thus does not map interval [0, 1] to itself. SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 33 However, the original matrix on the left is actually in the semigroup, as shown in the proof. Appendix 2 We prove a basic result concerning the trace of certain spectral projections in the rotation algebra Aθ. The machinery for this result is standard, and we borrow heavily from the results in (Boca, 2001). Lemma 7. Let τθ and τo be the (framed) tracial states on rotation algebras Aθ, Ao, respectively, with θ = p/q. Let [a, b] be the k-th interval Ik in the spectrum of the operator hθ = u + u∗ + v + v∗ in algebra Aθ. Then for x ∈ [a, b], (110) τθ(χ[a,x)(hθ)) = 1 q τo(χ[−4,(−1)k+qPθ(x))(ho)), where χ[a,x) is the spectral projection onto interval [a, x). Consequently, (111) τθ(χ(−∞,x)(hθ)) = k − 1 q + 1 q τo(χ[−4,(−1)k+qPθ(x))(ho)). Note: the choice of the sign in front of polynomial Pθ(x) ensures that the map (112) x (cid:55)→ (−1)k+qPθ(x) is monotonic increasing on the interval Ik. In the earlier sections of this paper, it is convenient to write the trace-spectral projection function as the cumulative density F (x) for the operator ho, so we may write k − 1 (113) τθ(χ(−∞,x)(hθ)) = + 1 q q F ((−1)k+qPθ(x)), where (114) is a continuous, strictly increasing function mapping [−4, 4] onto [0, 1]. F (x) = τo(χ[−4,x)(ho)) Proof: We do a direct calculation. The trace on Aθ is obtained from a representa- tion of the algebra in C(T2) ⊗ Mq(C) with generators u = ι1 ⊗ Uo, v = ι2 ⊗ Vo, (115) where the functions ι1, ι2 are the coordinate maps ι1(z1, z2) = z1, ι2(z1, z2) = z2, and the q × q matrices Uo, Vo are the usual cyclic permutation and the diagonal with powers of e2πiθ. The trace in this frame is just (116) τθ = µ2 ⊗ trq, 34 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN where µ2 is the usual Haar measure on T2 and trq is the normalized trace on q × q matrices. (See (Boca, 2001), page 11.) Fix x in Ik. The matrix-valued function h = u + u∗ + v + v∗ has exactly one eigenvalue in the interval Ik, which moves continuously for parameter values (z1, z2) ∈ T2. This eigenvalue contributes to the trace precisely when it lies in the interval [a, x]. That is, we get a contribution precisely when the value of the polynomial (−1)k+qPθ(x) is smaller than zq q. We compute the trace in terms of a characteristic function on the torus T2, normalized by q, so q + zq 1 + z1 2 + z2 τθ(χ[a,x)(hθ)) = (117) q) dµ2. With y = (−1)k+qPθ(x), and noting that the map (z1, z2) (cid:55)→ (zq preserving inverse on the torus, we can write χ((−1)k+qPθ(x)<zq q+zq 2 +z2 1 +z1 T2 1, zq 2) has measure (cid:90) 1 q (cid:90) (cid:90) (cid:90) T2 T2 T2 τθ(χ[a,x)(hθ)) = = = = 1 q 1 q 1 q 1 q χ(y<zq 1 +z1 q+zq 2 +z2 q) dµ2 χ(y<z1+z1+z2+z2) dµ2 χ(Po(y)<z1+z1+z2+z2) dµ2 τo(χ[−4,y)(ho)). where we use the fact that Po(y) = y and so the last integral above yields the trace in Ao applied to the spectral projection obtained by applying χ[−4,y] onto ho. Using the relation y = (−1)k+qPθ(x), we obtain τθ(χ[a,x)(hθ)) = (118) as desired. The trace evaluated on the full interval (−∞, x) picks up an additional contribution of (k − 1)/q from the previous k − 1 intervals in the spectrum. QED τo(χ[−4,(−1)k+qPθ(x))(ho)), 1 q It is worth noting that this trace integral can be computed explicitly, the result involves an integral of arccos(y/2 + cos(θ)). Appendix 3 We conclude with some final plots of the algebraic curves which determine the horizontal component of the similarity maps. Properties to notice are that the curves have even symmetry for denominator q even, odd symmetry for q odd, and in all cases there is a sequence of nearly linear segments along the diagonal line y = x. It SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 35 is also notably that as the numerator p increases, the algebraic curves become much more complicated. Figure 17. Algebraic curves with denominator 5: 1/5 (cid:55)→ 1/7, 2/5 (cid:55)→ 2/9, 3/5 (cid:55)→ 3/11, 4/5 (cid:55)→ 4/13. In each graph, there are five nearly linear segments down the diagonal of the graph. All but the first graph have some disconnected components. .. .. -4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-202436 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 18. Algebraic curves with denominator 7: 1/7 (cid:55)→ 1/9, 2/7 (cid:55)→ 2/11, 3/7 (cid:55)→ 3/13, 4/7 (cid:55)→ 4/15, 5/7 (cid:55)→ 5/17, 6/7 (cid:55)→ 4/19. In each graph, there are seven nearly linear segments down the diagonal of the graph. All but the first graph have some disconnected components. .. .. -4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 37 Figure 19. Algebraic curve with denominator 2, the 1/2 (cid:55)→ 1/4 map. Figure 20. Algebraic curves with denominator 4: 1/4 (cid:55)→ 1/6, 3/4 (cid:55)→ 3/10. In each graph, there are four nearly linear segments down the diagonal of the graph. The right graph has some components which are disconnected from the linear segments. .. .. -4-2024-4-2024-4-2024-4-2024-4-2024-4-202438 MICHAEL P. LAMOUREUX, JAMES A. MINGO, AND SYDNEY R. PACHMANN Figure 21. Algebraic curves with denominator 6: 1/6 (cid:55)→ 1/8, 5/6 (cid:55)→ In each graph, there are six nearly linear segments down the 5/16. diagonal of the graph. The right graph has some components which are disconnected from the linear segments. .. .. -4-2024-4-2024-4-2024-4-2024SPECTRA SELF-SIMILARITY FOR ALMOST MATHIEU OPERATORS 39 Figure 22. Algebraic curves with denominator 8: 1/8 (cid:55)→ 1/10, 3/8 (cid:55)→ 3/14, 5/8 (cid:55)→ 5/18, 7/8 (cid:55)→ 7/22. In each graph, there are eight nearly linear segments down the diagonal of the graph. All but the first graph has some components which are disconnected from the linear segments. (M. Lamoureux) Dept. Mathematics and Statistics, University of Calgary, 2500 University Ave NW, Calgary AB T2N 1N4, Canada E-mail address: [email protected] (J. Mingo) Dept. Mathematics and Statistics, Queens University, Kingston, ON Canada E-mail address: [email protected] (S. Pachmann) Faculty of Engineering, University of Calgary, Calgary AB, Canada E-mail address: [email protected] -4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024-4-2024
1703.03605	1	1703	2017-03-10T10:06:50	Infinite index extensions of local nets and defects	[ "math.OA", "math-ph", "math-ph", "math.QA" ]	Subfactor theory provides a tool to analyze and construct extensions of Quantum Field Theories, once the latter are formulated as local nets of von Neumann algebras. We generalize some of the results of [LR95] to the case of extensions with infinite Jones index. This case naturally arises in physics, the canonical examples are given by global gauge theories with respect to a compact (non-finite) group of internal symmetries. Building on the works of Izumi, Longo, Popa [ILP98] and Fidaleo, Isola [FI99], we consider generalized Q-systems (of intertwiners) for a semidiscrete inclusion of properly infinite von Neumann algebras, which generalize ordinary Q-systems introduced by Longo [Lon94] to the infinite index case. We characterize inclusions which admit generalized Q-systems of intertwiners and define a braided product among the latter, hence we construct examples of QFTs with defects (phase boundaries) of infinite index, extending the family of boundaries in the grasp of [BKLR16].	math.OA	math	Infinite index extensions of local nets and defects Simone Del Vecchio1 and Luca Giorgetti1 1Dipartimento di Matematica, Universit`a di Roma Tor Vergata, Via della Ricerca Scientifica, 1, I-00133 Roma, Italy, [email protected], [email protected] Abstract Subfactor theory provides a tool to analyze and construct extensions of Quantum Field Theories, once the latter are formulated as local nets of von Neumann algebras. We generalize some of the results of [LR95] to the case of extensions with infinite Jones index. This case naturally arises in physics, the canonical examples are given by global gauge theories with respect to a compact (non-finite) group of internal symmetries. Building on the works of Izumi, Longo, Popa [ILP98] and Fidaleo, Isola [FI99], we consider generalized Q-systems (of intertwiners) for a semidiscrete inclusion of properly infinite von Neumann algebras, which generalize ordinary Q-systems introduced by Longo [Lon94] to the infinite index case. We characterize inclusions which admit generalized Q-systems of intertwiners and define a braided product among the latter, hence we construct examples of QFTs with defects (phase boundaries) of infinite index, extending the family of boundaries in the grasp of [BKLR16]. 1 Introduction The study of extensions in relativistic Quantum Field Theory (QFT) is well-motivated in several respects. Gauge theory, for instance, provides examples of extensions where a theory of (anti)commuting fields obeying Bose/Fermi statistics, equipped with a gauge group sym- metry, contains a subtheory generated by gauge invariant (observable) fields. The former can be viewed as an extension of the latter, and similarly any intermediate theory gives rise to a smaller extension of the observable theory. Defects and boundaries can also be described by extensions, where different types of bulk fields (depending on their relative spacetime localiza- tion with respect to a certain "defect" line or hypersurface) generate extensions of a common subtheory which contains, for example, the components of the stress-energy tensor that are conserved across the boundary. Extensions also appear in classification instances of QFTs, where all the theories belonging to a certain family share a common subtheory (dictated, e.g., by spacetime symmetry), hence the classification problem can be turned into a classification of extensions. This is the case, for example, in chiral Conformal Field Theory (CFT) where the Fourier modes of the conformal stress-energy tensor necessarily obey the commutation relations of the Virasoro algebra at a fixed value of the central charge parameter. Lastly, the analysis of extensions can be used to construct new examples of QFTs. Starting from some 1 theory, if one can write it as a non-trivial subtheory extended by a certain family of genera- tors, then new theories can be constructed by suitably manipulating the generators and their commutation relations, compatibly with locality, and leaving the subtheory untouched. All of these different situations and problematics have a model-independent and mathe- matically rigorous formulation in the Algebraic approach to QFT (AQFT ) due to Haag and Kastler, see [Haa96]. Global gauge theories have been tackled since the early works of [DHR69a], [DHR69b], [DR72], culminating in [DHR71], [DHR74] and [DR90] in particular, where it is shown that every theory of local observables arises as gauge group fixed points of a bigger field theory, obeying (anti)commutation relations and equipped with a (global) gauge group symmetry. Both the gauge group and the field extension are intrinsically determined by the local ob- servables (hence dictated by locality, i.e., Einstein's causality). Intermediate extensions of gauge group fixed points (in 3+1 spacetime dimensions) have been studied in [CDR01]. In the chiral CFT setting (1 spacetime dimension) gauge group fixed points (also called orbifold CFTs) appear in [Xu00], [Xu05], [Mug05], and have been generalized to finite hypergroup fixed points by [Bis17] (generalized orbifold CFTs), where the information about gauge in- variance contained in the conditional expectation is expressed by an hypergroup action via completely positive (CP) maps. Intermediate chiral extensions have been analysed by [Lon03] and [Xu14]. Defects and boundaries have been studied with AQFT methods in recent works by [BKL15], [BKLR16], [BR16], see also [BKLR15, Ch. 5] where the main mathematical tools to construct and classify boundary conditions are developed. This analysis of defects and boundaries in QFT has been our initial motivation for the work presented in this article. Lastly, again using extensions, the classification of all chiral CFTs with central charge c < 1 has been achieved in [KL04]. In the Haag-Kastler formalism, a (local) quantum field theory is described by a net of local algebras {O 7→ A(O)}, see [Reh15], [HM06] for self-contained introductions. Local algebras A(O) are assumed to be von Neumann algebras on the vacuum Hilbert space, and they typically turn out to be factors (hyperfinite and of type III 1 in the classification of Connes [Con73], [Haa87]). Hence an extension of QFTs {A ⊂ B} is naturally described by a family of subfactors indexed by spacetime regions O (e.g., double cones in Minkowski space or bounded intervals on the line). It was in the work of Longo and Rehren [LR95] (indeed titled "Nets of subfactors") that it became clear how to use subfactor theory as a tool to classify and construct extensions in QFT. Their main idea is to exploit the notion of Q-system, due to [Lon94], for nets of subfactors, in order to relate coherent families of conditional expectations (which respect the net structure) to coherent families of (dual) canonical endomorphisms [Lon87] (which turn out to be the restrictions to different spacetime regions O of a unique global DHR endomorphism θ of {A}). The family of conditional expectations generalizes the notion of global gauge symmetry, while the DHR endomorphism θ represents the (reducible) vacuum representation of the bigger theory {B} once restricted to {A}. Mathematically speaking, the theory of subfactors plays a prominent role in the panorama of Operator Algebras since the work of Jones [Jon83]. He established a notion of index for subfactors, which is an invariant (hence opened the way to classification questions) and surprisingly quantized for values between 1 and 4 (Jones' rigidity theorem). Since then, the major efforts have been devoted to the study of finite index (finite depth) subfactors and a complete classification has been achieved for subfactors with index at most 5 + 1 4 [JMS14], [AMP15], using techniques of [Pop95a] and [Jon99]. At the same time, the analyses of QFT 2 extensions [LR95] and of theories with defects and boundaries [BKLR16] cover the finite index case only, both being based on the notion of Q-system (which is tightly connected to the existence of conjugate morphisms ¯ι of the inclusion morphism ι : N ֒→ M for a subfactor N ⊂ M, hence to the finiteness of the dimension of ι in the sense of [LR97]). In this article, building on the notion of Pimsner-Popa basis for an inclusion of von Neumann algebras, see [PP86], [Pop95b], and on the characterization of the canonical endo- morphisms given by Fidaleo and Isola in [FI99], we reformulate the results on QFT extensions of [LR95, Sec. 4] in the finite index case, and generalize them to infinite index extensions, see Section 6. This case naturally appears in physical situations, e.g., if we consider global gauge theories with respect to a compact non-finite group of internal symmetries. In order to do so, we first adopt the notion of generalized Q-system, due to [FI99], see Definition 3.1, and then consider more special generalized Q-systems of intertwiners, see Defi- nition 3.7 and 5.8. The latter can be thought, roughly speaking, as C ∗ Frobenius algebra-like objects with possibly infinitely many comultiplications, see Remark 3.8, 5.7 and cf. [BKLR16, Sec. 3.1]. Any semidiscrete inclusion of (properly infinite, with separable predual) von Neumann algebras N ⊂ M, i.e., an inclusion endowed with a faithful normal conditional expectation E : M → N , admits a generalized Q-system. Vice versa, from any generalized Q-system one can (re)construct the bigger algebra M and the conditional expectation E, see [FI99, Thm. 4.1]. An advantage of using generalized Q-systems (in the finite index case as well) is that no factoriality or irreducibility assumption on the inclusion is needed along the way. This enhanced flexibility is particularly desirable in the study of boundary conditions, see com- ments after [BKLR16, Thm. 4.4], where non-irreducible, non-factorial extensions necessarily appear. On the other hand, generalized Q-systems (in the infinite index case) dwell a bit further away from the purely categorical setting of their finite index counterpart. Given a semidiscrete inclusion of von Neumann algebras N ⊂ M (where N is an infinite factor), we show that the existence of a generalized Q-system with the additional intertwining property is actually equivalent to the discreteness of the inclusion in the sense of [ILP98] (but admitting non-irreducible extensions), see Section 5. This characterization relies on strong results of [ILP98] and [FI99], and can be physically interpreted by saying that a semidiscrete extension is discrete if and only if it is generated by charged fields, in the sense of [DR72]. These are elements ψ ∈ M which generate from the vacuum a non-trivial (irreducible) sub- sector ρ ≺ θ of the dual canonical endomorphism θ, i.e., ψn = ρ(n)ψ for every n ∈ N . In Section 6 it is shown that generalized Q-systems of intertwiners indeed induce discrete (finite or infinite index) extensions of QFTs in the sense of [LR95]. Two different ways to obtain the construction are provided: one is a direct generalization of [LR95, Thm. 4.9], the other one exploits an inductive procedure which is somewhat more suitable to be used for the analysis of braided products and boundary conditions in the subsequent sections. In Section 7, we give a general proof of covariance of QFT extensions constructed from covariant nets of local observables. This fact is apparently well known to experts, and clear in many examples, but we could not find a general statement in the literature (on finite index extensions). The key ingredient in our proof is the equivariance of the action of the spacetime symmetry group on the DHR category. More precisely, the mere existence of covariance cocycles, see Definition 7.3, is not sufficient to guarantee covariance. One needs in addition naturality and tensoriality properties of the cocycles. Given two generalized Q-systems of intertwiners (in a C ∗ braided tensor category), one 3 can easily define their braided product in analogy with the case of ordinary Q-systems, see Definition 4.1 and cf. [EP03, Sec. 3], [BKLR16, Sec. 4.9]. In Section 4, we prove the non- trivial statement that the braided product of two generalized Q-systems of intertwiners is again a generalized Q-system of intertwiners, i.e., that the analytical properties defining a Pimsner-Popa basis (as a part of the definition of a generalized Q-system) behave well with respect to the categorical notions of naturality and tensoriality of a braiding in a C ∗ tensor category. Thus, in the QFT setting, we can define the braided product of nets of local observables and construct new examples of irreducible phase boundary conditions with infinite index (infinitely many bulk fields) by taking the direct integral decomposition of the braided product net with respect to its center, see Section 8 and 9. On the other hand, we leave open the questions about universality of the braided product construction and the classification of boundary conditions in the infinite index case, cf. [BKLR16, Sec. 5]. In Section 10, we work out examples of infinite index (discrete) extensions of the chiral U (1)-current algebra [BMT88] and explicitly compute their braided products. These exam- ples show an important difference with the analysis of boundary conditions in the finite index case, namely the center of the braided product may be a continuous algebra, i.e., with no non-trivial minimal projections, hence the irreducible boundary conditions constructed by direct integral decomposition need not be representations of the braided product itself. Notation-wise we work with nets of local algebras {O 7→ A(O)} indexed by partially ordered and directed sets of spacetime regions K, in order to formulate our results, when possible, for arbitrary spacetime dimensions, e.g., in 1D theories on the line, 1+1D or 3+1D theories in Minkowski space. 2 Pimsner-Popa bases E Let N ⊂ M be a unital inclusion of von Neumann algebras with a normal faithful conditional expectation E : M → N . Assume that M acts standardly on a separable Hilbert space H and let N ⊂ M ⊂ M1 := hM, eNi 1 be the Jones basic construction [Jon83], see also [Pop95b, Sec. 1.1.3], [LR95, Sec. 2.2]. Up to spatial isomorphism it can be characterized as follows. Let Ω ∈ H be a cyclic and separating vector for M such that the induced (normal faithful) state ω of M is invariant under E, i.e., ω ◦ E = ω, and set eN := [N Ω], the orthogonal projection on the subspace H0 := N Ω ⊂ H. The projection eN ∈ N ′ ∩ M1 is the Jones projection of N ⊂ M with respect to E, and implements E in the sense that E(m)eN = eN meN , m ∈ M. Moreover, it is uniquely determined up to conjugation with unitaries in M′ [Kos89, App. I]. Definition 2.1. [PP86], [Pop95b]. A Pimsner-Popa basis for N ⊂ M is a family of elements {Mi} ⊂ M, where i runs in some set of indices I, such that E (i) Pi := M ∗ i eN Mi are projections in M1 which are mutually orthogonal, i.e., PiP ∗ i = Pi and PiPj = δi,jPi for every i, j ∈ I. (ii) Pi Pi = 1, where the sum converges (unconditionally) in the strong operator topology. For future reference, we mention the following equivalent characterization of the algebraic properties of Pimsner-Popa bases, see [Pop95b, Sec. 1.1.4]. 1Here hSi denotes the von Neumann algebra generated by a subset S ⊂ B(H). For a pair of subsets S1, S2 ⊂ B(H) we also denote hS1, S2i by S1 ∨ S2. 4 Lemma 2.2. In the notation of Definition 2.1, the conditions (i) and (ii) are respectively equivalent to (i)′ qi := E(MiM ∗ i ) are projections in N (not necessarily mutually orthogonal) and E(MiM ∗ j ) = 0 for every i 6= j, i, j ∈ I. i eNH = H in the Hilbert space topology. (ii)′ Pi M ∗ Proposition 2.3. [Pop95b]. If {Mi} is a Pimsner-Popa basis for N has the following expansion m =Xi M ∗ i E(Mim) E ⊂ M then every m ∈ M unconditionally convergent in the topology generated by the family of seminorms {k · kϕ : ϕ ∈ (M∗)+, ϕ = ϕ ◦ E}, with kmkϕ := ϕ(m∗m)1/2. The expansion is unique if and only if E(MiM ∗ i ) = 1 for every i ∈ I. Remark 2.4. In view of the proposition above, Pimsner-Popa bases {Mi}, or better their adjoints {M ∗ i } can be seen as bases for M as a right pre-Hilbert N -module with the N - valued inner product (M ∗ j ) := E(MiM ∗ j ). i M ∗ The cardinality of a Pimsner-Popa basis {Mi} is not a invariant for N ⊂ M. Indeed, by the following cutting and gluing procedures [Pop95b, Sec. 1.1.4] we obtain other Pimsner- Popa bases: E i ∈ N such thatPj aj i aj∗ i = E(MiM ∗ i ), (1) If, for each i, we take a set of partial isometries aj then {aj∗ i Mi} is also a basis. j ) and E(MkM ∗ (2) If E(Mj M ∗ k ) are orthogonal, then we can replace the pair Mj, Mk in {Mi} by Mj + Mk and we still get a basis. The good notion of dimension of M as an N -module is given by the Jones index of the inclusion N ⊂ M with respect to E, [Jon83], [Kos86]. This guiding idea is supported by the following theorem due to [PP86, Prop. 1.3], [BDH88, Thm. 3.5], [Pop95b, Thm. 1.1.5, 1.1.6], which characterizes the finiteness of the index (and computes its value) by means of Pimsner-Popa bases. E Theorem 2.5. [Pop95b]. N ⊂ M has finite Jones index if and only if it has a Pimsner- Popa basis {Mi} such that Pi M ∗ i Mi is ultraweakly convergent in M. In this case, Pi M ∗ i Mi belongs to the center of M, it holds Xi i Mi = Ind(N ⊂ M) M ∗ E E where Ind(N Pimsner-Popa basis. ⊂ M) denotes the Jones index of E, and the same is true for any other If in addition N is properly infinite, then N ⊂ M has finite Jones index if and only if it has a Pimsner-Popa basis made of one element {M}. Moreover, M can be chosen such that E(M M ∗) = 1. E 5 We are mainly interested in inclusions of properly infinite von Neumann algebras (with separable predual), due to their appearance in QFT, see, e.g., [Kad63], [Lon79]. In this setting, with no finite index or factoriality assumptions, it was shown by Fidaleo and Isola [FI99, Thm. 3.5] that Pimsner-Popa bases made of elements of M always exist. Proposition 2.6. [FI99]. Every inclusion N ⊂ M of properly infinite von Neumann algebras with a normal faithful conditional expectation E : M → N admits a Pimsner-Popa basis {Mi} ⊂ M in the sense of Definition 2.1. E 3 Infinite index and generalized Q-systems (of intertwiners) Q-systems were introduced by R. Longo in [Lon94, Sec. 6]. They provide a way to alge- braically characterize infinite subfactors N ⊂ M with finite index together with a normal faithful conditional expectation E : M → N by means of data pertaining to the smaller factor N . The main technical tool to achieve this characterization is the notion of canonical endomorphism [Lon87] for the inclusion N ⊂ M, namely the homomorphism γ : M → N defined by γ := (jN jM)↾M, where jN := AdJN ,Φ, jM := AdJM,Φ and JN ,Φ, JM,Φ are respec- tively the modular conjugations of N , M with respect to a cyclic and separating vector Φ for N and M. From a categorical perspective, a Q-system is a special C ∗ Frobenius algebra in a strict C ∗ tensor category C with simple unit, cf. [BKLR15, Def. 3.8]. In the more concrete case of subfactors, the category is C = End0(N ), whose objects are the endomorphisms of the factor N with finite dimension in the sense of [LR97]. Here we recall and analyze the more general notion of generalized Q-system, introduced by F. Fidaleo and T. Isola in [FI99, Sec. 5] for a possibly infinite index (semidiscrete or semicompact) inclusion of properly infinite von Neumann algebras. We then introduce the more special notion of generalized Q-system of intertwiners that will be the fundamental object in the subsequent sections, in particular for the applications to QFT. Let N ⊂ M be a unital inclusion of properly infinite von Neumann algebras on a separable Hilbert space H. Denote by C(M,N ) and E(M,N ) respectively the set of all normal and normal faithful conditional expectations of M onto N . We call the inclusion N ⊂ M semidiscrete if E(M,N ) 6= ∅, and semicompact if E(N ′,M′) 6= ∅, or equivalently if E(M1,M) 6= ∅ or E(N ,N1) 6= ∅, where N1 ⊂ N ⊂ M ⊂ M1 denotes the tower of von Neumann algebras obtained by canonical extension and restriction of the original inclusion [LR95, Sec. 2.5 A]. The terminology is adopted from [FI99], [ILP98], [FI95], [HO89]. Recall that a finite index inclusion is both semidiscrete and semicompact, see e.g. [Lon90, Prop. 4.4]. Let End(N ) be the collection of normal faithful unital -endomorphisms of N . The following notion is tailored to describe semidiscrete inclusions of von Neumann algebras N ⊂ M with E ∈ E(M,N ), possibly of infinite index. Definition 3.1. [FI99]. Let N be a properly infinite von Neumann algebra. A generalized Q-system in C = End(N ) is a triple (θ, w,{mi}) consisting of an endomorphism θ ∈ End(N ), an isometry w ∈ HomEnd(N )(id, θ) (i.e., wn = θ(n)w, n ∈ N ), and a family {mi} ⊂ N indexed by i in some set I, such that (i) pi := m∗ i ww∗mi are mutually orthogonal projections in N , i.e., pip∗ j = δi,jpi, such that Pi pi = 1. ("Pimsner-Popa condition") 6 (ii) nw = 0 ⇒ n = 0 if n ∈ N1 := hθ(N ),{mi}i. ("faithfulness condition") Remark 3.2. An analogous definition of generalized Q-system in End(N ), involving an isom- etry x ∈ HomEnd(N )(θ, θ2) instead of w ∈ HomEnd(N )(id, θ), can be given in the semicompact case, see [FI99, Sec. 5]. We shall however be interested in extensions N ⊂ M with a (nor- mal faithful) conditional expectation E ∈ E(M,N ) as they arise in QFT when N = A(O), M = B(O) are local algebras (relative to some spacetime region O) and {A ⊂ B} is an extension of a net of local observables {A} by means of a "field net" {B}. Here E generalizes the notion of an average over a global gauge group action on fields, giving the observables as the gauge invariant part. Theorem 3.3. [FI99]. Let N be a properly infinite von Neumann algebra with separable predual and θ ∈ End(N ). Then the following are equivalent (1) There is a von Neumann algebra N1 such that N1 ⊂ N with E′ ∈ E(N1,N2) 6= ∅, where N2 := θ(N ) ⊂ N1, and θ is a canonical endomorphism for N1 ⊂ N . (2) There is a von Neumann algebra M such that N ⊂ M with E ∈ E(M,N ) 6= ∅, and θ is a dual canonical endomorphism for N ⊂ M, i.e., θ = γ↾N where γ ∈ End(M) is a canonical endomorphism for N ⊂ M. (3) The endomorphism θ is part of a generalized Q-system in End(N ), (θ, w,{mi}), see Definition 3.1. Proof. We may assume that N is in its standard representation on H. The equivalence of (1) and (2) is then obtained by canonical extension and restriction [LR95, Sec. 2.5 A]. The tower of von Neumann algebras reads . . . ⊂ N2 = θ(N ) E′ ⊂ N1 = hθ(N ),{mi}i θ ⊂ N E,γ ⊂ M ⊂ . . . (1) where N2 ⊂ N1 = γ(N ⊂ M) is a spatial isomorphism of inclusions and the relation E′ ◦ γ = γ ◦ E on M gives a bijection between E(N1,N2) and E(M,N ). The equivalence of (1) and (3) is due to [FI99, Thm. 4.1]. In particular, they show that eN2 := ww∗ is a Jones projection for the inclusion N2 ⊂ N1 with respect to E′ := θ(w∗·w) and that N = hN1, eN2i is the associated Jones extension. Hence the condition (i) in Definition 3.1 says that {mi} ⊂ N1 is a Pimsner-Popa basis for N2 ⊂ N1 with respect to E′. The condition (ii) in Definition 3.1 is nothing but faithfulness of E′. i w is a projection, which is equivalent to ww∗mim∗ Remark 3.4. The condition that the pi in Definition 3.1 are (mutually orthogonal) projections j = δi,jpi, does not enter in the proof of (3) ⇒ (1) of Theorem 3.3, onlyPi pi = 1 in N , i.e., pip∗ i ww∗mi is a projection if and is relevant there. We can however always assume it because m∗ i )ww∗ is a only if w∗mim∗ i ) is a projection, because ww∗ = eN2 and n 7→ neN2 is an isomor- projection, i.e., E′(mim∗ phism of N2 onto N2eN2. Hence we can apply a Gram-Schmidt orthogonalization procedure to the {mi} with respect to the operator-valued inner product (mimj) := E′(mjm∗ i ) and choose another basis { mi} such that E′( mj m∗ Remark 3.5. Notice that no factoriality Z(N ) = C1, Z(M) = C1, nor irreducibility N ′ ∩ M = C1 assumptions enter in the proof of Theorem 3.3, [FI99]. In the case of non-irreducible finite index subfactors, E is not necessarily the minimal conditional expectation, see [Hia88], [Lon89, Sec. 5]. i ww∗ = E′(mim∗ i ) = δi,j 1. 7 Proposition 3.6. [FI99]. Let N ⊂ M a semidiscrete inclusion of properly infinite von Neumann algebras and let γ be a canonical endomorphism. The following are equivalent (1) N ⊂ M is irreducible in the sense that N ′ ∩ M = Z(N ). (2) E(M,N ) contains only one element. (3) HomEnd(N )(id, θ) is cyclic as a Z(N )-module, where θ = γ↾N . We now specialize the notion of generalized Q-system (Definition 3.1) by requiring an additional intertwining property of the Pimsner-Popa elements. Definition 3.7. Let N be a properly infinite von Neumann algebra. We call (θ, w,{mi}) a generalized Q-system of intertwiners in C = End(N ) if, in addition to the properties of Definition 3.1, it satisfies mi ∈ HomEnd(N )(θ, θ2) (i.e., miθ(n) = θ2(n)mi, n ∈ N ) for every i ∈ I. In this case we can use string diagrams to denote w and mi as follows w = id • θ = • , mi = θ • i , i ∈ I. θ θ θ At this point, a comparison between the notions of generalized Q-system and "ordinary" Q-system in the finite index setting is due. Remark 3.8. (The finite index case). An infinite subfactor N ⊂ M with E ∈ E(M,N ) can be characterized by an "ordinary" Q-system (θ, w, x) if and only if the Jones index of E is finite, see [Lon94], [LR95, Sec. 2.7]. The algebraic relations defining a Q-system in End0(N ) read as follows: θ ∈ End0(N ), w ∈ HomEnd0(N )(id, θ), x ∈ HomEnd0(N )(θ, θ2) and x∗x ∈ C1. xx∗ = x∗θ(x) = θ(x∗)x, w∗x = θ(w∗)x = 1, x2 = θ(x)x, The conditions in the line above are called respectively unit property, associativity, Frobenius property and specialness, see [BKLR15, Def. 3.8]. It is known that the Frobenius property is a consequence of the other properties [LR97], [BKLR15, Lem. 3.7] and that specialness is not needed to construct the extension N2 = θ(N ) ⊂ N1, i.e., N ⊂ M [BKLR15, Rmk. 3.18]. Moreover, it is an easy exercise to check that ordinary Q-systems are also generalized Q-system of intertwiners with {mi} = {x} (up to a normalization of w and x), in the sense Indeed the Pimsner-Popa condition x∗ww∗x = 1 follows by w∗x = 1, of Definition 3.7. and the faithfulness condition nw = 0 ⇒ n = 0, n ∈ N1 = hθ(N ), xi follows because hθ(N ), xi = θ(N )x = x∗θ(N ) hold, due to θ(w∗)x = 1, associativity and Frobenius property. On the other hand, a finite index inclusion of infinite factors N2 ⊂ N1 with normal faithful conditional expectation E′(·) = θ(w∗ · w), always has a Pimsner-Popa basis of one element, m ∈ N1, such that E′(mm∗) = 1 by Theorem 2.5. The triple (θ, w, m) is a generalized Q-system in the sense of Definition 3.1. The characterizing properties m∗ww∗m = 1, w∗mm∗w = 1 8 are a weaker form of the unit property for ordinary Q-systems, and the Pimsner-Popa expan- sion of Proposition 2.3 gives in particular m2 = E′(m2m∗)m, mm∗ = E′(m(m∗)2)m. If we assume the unit property w∗m = 1 to hold, we get back the associativity m2 = θ(m)m and the Frobenius property mm∗ = θ(m∗)m. If (θ, w,{mi}) is a generalized Q-system (of intertwiners) in C = End(N ), consider the tower of von Neumann algebras . . . ⊂ N2 E′ ⊂ N1 θ ⊂ N E,γ ⊂ M γ1 ⊂ M1 ⊂ . . . as in equation (1), where the Jones extension M1 = hM, eNi of N ⊂ M with respect to E coincides with the canonical extensions, namely hM, eNi = jM(N ′), see [LR95, Sec. 2.5 D], [Lon89, Sec. 3]. Here Ω is a cyclic and separating vector for M as in Section 2 and jM = AdJM,Ω is the associated modular conjugation. Moreover, θ and γ1 are canonical −1(eN2) = eN endomorphisms dual to γ, hence θ = γ↾N , γ = γ1↾M. Then γ1 and {Mi := γ−1(mi)} ⊂ M clearly forms a Pimsner-Popa basis for N ⊂ M with respect to E. Definition 3.9. We call (γ, w,{Mi}) a generalized Q-system (of intertwiners) dual to (θ, w,{mi}). The intertwining relation mi ∈ HomEnd(N )(θ, θ2) is equivalent to Min = θ(n)Mi, n ∈ N . −1(ww∗) = γ1 4 Braided products Suppose additionally that two generalized Q-systems of intertwiners are composed of data belonging to a certain braided tensor subcategory of End(N ), we can consider their braided product as follows Definition 4.1. Let N be a properly infinite von Neumann algebra and C ⊂ End(N ) a C ∗ braided tensor subcategory of End(N ). Let (θA, wA,{mA j }) two generalized Q-systems of intertwiners in C (Definition 3.7), indexed respectively by i ∈ I and j ∈ J. We call i1}) and (θB, wB,{mB the braided product of (θA, wA,{mA where ε mB j i ×± mA (θAθB, wAwB,{mA i ×± ε mB j }) i }) and (θB, wB,{mB j }), indexed by (i, j) ∈ I × J, := θA(ε± i θA(mB j ) θA,θB )mA depending on the ± choice. Here ε+ = ε and ε− = εop denote respectively the braiding of C and its opposite. Equivalently wAwB = • • , θA θB i ×+ mA ε mB j = θA θB i • j • θA θB θA θB , (i, j) ∈ I × J and similarly for mA i ×− ε mB j . 9 Surprisingly, the analytic conditions dictated on generalized Q-systems by subfactor the- ory (e.g. the property characterizing a Pimsner-Popa basis) turn out to be naturally compati- ble with the categorical notion of braiding in a tensor category of endomorphisms. Indeed we have the following proposition which extends the braided product construction, see [BKLR15, Sec. 4.9], to the infinite index case. Proposition 4.2. The braided product of two generalized Q-systems of intertwiners is again a generalized Q-system of intertwiners. Proof. The intertwining properties appearing in Definition 3.7 are easily checked once we write the operators wAwB and mA j , (i, j) ∈ I × J as tensor products and compositions of arrows in the braided tensor category of endomorphisms C ⊂ End(N ), as in the case of ordinary Q-systems [BKLR15, Def. 4.30]. The Pimsner-Popa condition (i) in Definition 3.1 is more lengthy to check. For each i ∈ I i ×± ε mB and j ∈ J, let = θA(mB∗ ε mB := (mA pAB,± i,j i θA((ε± j )mA∗ j )mA∗ = θA(mB∗ i ×± j )∗wAwBwB∗wA∗(mA i ×± θA,θB )∗)θA(wB)wAwA∗θA(wB∗)θA(ε± i θA(θA(wB))wAwA∗θA(θA(wB∗))mA ε mB j ) θA,θB )mA i θA(mB j ) i θA(mB j ) θA,θB )∗wB = θA(wB)(ε± because (ε± tensor category C, or of its opposite ε− = εop, and ε± i wAwA∗mA j )θA(wB)mA∗ = θA(mB∗ θA,id = 1 by convention. Moreover i θA(wB∗)θA(mB j ) θA,id)∗ by naturality of the braiding ε+ = ε in the braided = mA∗ i wAwA∗mA i θA(mB∗ j wBwB∗mB j ) hence we have shown (2) where pA j , j ∈ J are the projections appearing in Definition 3.1 respectively for the two generalized Q-systems. Equation (2) is much more effectively expressed using graphical calculus i , i ∈ I and pB j ) = θA(pB j )pA i pAB,± i,j = pA i θA(pB θA θB • • i • • • • i θA Now one can easily check that pAB,± 1. i,j j • • j θB = • • θA i • • i θA θB j • • j θB • • . are mutually orthogonal projections which sum up to 1 ×± ε N B The faithfulness condition (ii) in Definition 3.1 follows from Lemma 4.3 below. Indeed, let 1 (see below), then nwAwB = nθA(wB)wA = 0 if and only if nθA(wB) = 0 since θA,θB )∗wB = 0 1 we have that n = 0 and n ∈ N A 1 ×± N A by naturality of the braiding. Since Ad(ε± the proof is complete. 1 . Now nθA(wB) = 0 if and only if ε± θA,θB )(N A θA,θB nθA(wB) = ε± 1 ×± ε N B 1 ⊂ N A 1 ) ⊂ N B θA,θB n(ε± ε N B 10 Lemma 4.3. In the notation of Definition 4.1, consider the two towers of von Neumann 1 ⊂ N ⊂ MB respectively associated to the algebras N A two generalized Q-systems (of intertwiners) as in Theorem 3.3. Let 1 ⊂ N ⊂ MA and N B 2 ⊂ N B 2 ⊂ N A then 1 ×± N A ε N B 1 :=(cid:10)θAθB(N ),{mA θA,θB )(N A 1 , Ad(ε± ε mB j }(cid:11) 1 ) ⊂ N B ε N B 1 . i ×± 1 ×± 1 ×± N A ε N B 1 ⊂ N A Proof. The first inclusion follows from the very definitions. To show the second observe that Ad(ε± θA,θB )(θAθB(N )) = θBθA(N ) ⊂ θB(N ). Hence it is enough to check that θA,θB θAθB(ε± ε± θA,θB )θA(ε± θA,θB )mA i θA(mB j )(ε± θA,θB )∗ = θB(ε∓ θB,θA)mB j θB(mA i ), but this follows from repeated application of naturality and tensoriality of the braiding mA i θA(mB j )(ε± θA,θB )∗ = θAθA(mB j )mA i ε∓ θB ,θA where ε∓ θBθB ,θAθA = θA(ε∓ θB,θA)θAθB(ε∓ = θAθA(mB j )ε∓ θB ,θAθAθB(mA θB ,θA)ε∓ i ) = ε∓ θB ,θAθB(ε∓ θB ,θA). θBθB ,θAθAmB j θB(mA i ) Corollary 4.4. (of Proposition 4.2). θAθB ∈ End(N ) is a canonical endomorphism for the inclusion N A ε N B 1 ×± 1 ⊂ N . Moreover, the inclusion 1 ×± θAθB(N ) ⊂ N A ε N B 1 (3) is semidiscrete 2 with (normal faithful) conditional expectation given by EAB ′ := θAθB(wB∗wA∗ · wAwB). Denote by MA ×± the von Neumann algebra appearing in the tower ε MB . . . ⊂ θAθB(N ) ⊂ N A 1 ×± ε N B 1 θAθB ⊂ N γAB ⊂ MA ×± ε MB ⊂ . . . ε N B ε MB dual to θAθB. By definition, γAB obtained as in Theorem 3.3 from the braided product Q-system. We call it the braided product of MA and MB. Here γAB denotes a canonical endomorphism for the inclusion N ⊂ MA ×± ε MB) = 1 ×± 1 . Similarly, γA and γB are respectively canonical endomorphisms dual to θA and N A θB. In order to show that the braided product MA ×± ε MB actually contains MA and MB as subalgebras (see Proposition 4.5 below) we need to consider generalized Q-systems of intertwiners with an additional property, which is a weaker version of the unit property in ordinary Q-systems, namely θ(w∗)x = 1, cf. [BKLR16, Prop. 4.12]. We shall come back to this property in the next section, see Proposition 5.5 and Definition 5.8. ↾N = θAθB and γAB(MA ×± 2By the results of the next section, the inclusion (3) is also discrete in the sense of Definition 5.1. 11 Proposition 4.5. In the notation of Definition 4.1, let (θA, wA,{mA fulfill in addition θA(wA∗)mA and 0 ∈ J. Then the maps 0 = 1 and θB(wB∗)mB i }) and (θB, wB,{mB j }) 0 = 1 for two distinguished labels 0 ∈ I A : MA → MA ×± ε MB, B : MB → MA ×± A := (γAB)−1 ◦ Ad(ε± θA,θB )∗ ◦ θB ◦ γA ε MB, B := (γAB)−1 ◦ θA ◦ γB are embeddings respectively of MA and MB into MA ×± ε MB. Call ιA : N → MA and ιB : N → MB the embeddings of N into MA and MB respectively. Then the two embeddings of N into the braided product coincide, i.e. A ◦ ιA = B ◦ ιB, i and M B the commutation relations among M A MA ×± ε MB are given by B(M B i )B(M B j ), Moreover, MA and MB generate the braided product, i.e. θA,θB A(M A j )A(M A i ) = ε± i ∈ I, j ∈ J. j , as in Definition 3.9, in the braided product (4) (5) (6) (7) Proof. We show first that MA ×± Ad(ε± ε MB =(cid:10)A(MA), B(MB)(cid:11) . ε N B θA,θB )∗ ◦ θB(N A 1 ) ⊂ N A θA(N B 1 ) ⊂ N A 1 ×± ε N B 1 ×± 1 1 (8) ε MB. For the inclusion (8) 0 = 1, hence from which it is clear that A and B are embeddings into MA ×± it is enough to show that θA(mB θA(mB j ) = θA(wA∗)mA j ) ∈ N A 1 ×± 0 θA(mB 1 . By assumption θA(wA∗)mA 0 θA(mB j ) ε N B j ) = θAθB(wA∗)θA(ε± θA,θB )mA 0 ×± 1 ×± ε N B j ∈ N A ε mB 1 = θAθB(wA∗)mA using naturality of the braiding and ε± id,θB = 1. For the inclusion (7), it is enough to observe 1 , cf. [BKLR15, Sec. 4.9], that Adε∓ and consider the previous case interchanging A with B and ± with ∓. Now, (4) is clear. To show the commutation relations among M A j apply first γAB to equation (5). The r.h.s. then reads is an isomorphism between N A 1 and N B i and M B 1 ×∓ ε N B ε N A 1 ×± θA,θB θAθB(ε± θA,θB )(ε± θA,θB )∗θB(mA i )ε± = θAθB(ε± θA,θB )(ε± θA,θB )∗ε± θAθA,θB mA θA,θB θA(mB j ) i θA(mB j ) = θAθB(ε± θA,θB )θA(ε± θA,θB )mA i θA(mB j ). Similarly, one can compute the l.h.s., namely j )(ε± θA(mB θA,θB )∗θB(mA By the intertwining property θAθA(mB we have equation (5). In the previous computations we have shown in particular that θA,θBθB )θAθA(mB j ) and by tensoriality of the braiding θA,θB ) = θA(ε± i = mA i θA(mB j )mA i . i )(ε± j )mA A(M A i )B(M B j ) = (γAB)−1(mA i ×± ε mB j ) from which equation (6) follows. 12 5 The case of discrete inclusions Generalized Q-systems with the additional intertwining property mi ∈ HomEnd(N )(θ, θ2) as in Definition 3.7 can be constructed whenever the inclusion N ⊂ M is discrete (see Definition 5.1 below, cf. [ILP98, Def. 3.7]). The main idea is to look first at elements ψρ ∈ M which generate on N subendomorphisms ρ ≺ θ of the dual canonical endomorphism θ ∈ End(N ) of N ⊂ M from the vacuum (identity representation), namely such that ψρn = ρ(n)ψρ, n ∈ N . Such elements are called charged fields after the work of [DR72] in QFT. In the subfactor setting they can be constructed as in [ILP98, Prop. 3.2]. We generalize the latter construction to the case of non-irreducible, non-factorial extensions (as one needs in the study of defects in QFT, see [BKLR16, Thm. 4.4]), and we show how charged fields can be used, in the discrete case, to define generalized Q-systems of intertwiners. Moreover, we show that a semidiscrete inclusion admitting a generalized Q-system of intertwiners is necessarily discrete. Consider an inclusion N ⊂ M, where N is an infinite factor and M is a properly infinite If E(M,N ) 6= ∅ denote by E ∈ von Neumann algebra on a separable Hilbert space H. P (M1,M) the normal semifinite faithful operator-valued weight dual to E ∈ E(M,N ), see [Kos86], [ILP98], [FI99]. Definition 5.1. [ILP98]. In the above notation, the inclusion N ⊂ M is called discrete if E(M,N ) 6= ∅ (semidiscreteness) and E↾N ′∩M1 is semifinite for some (hence for all) E ∈ E(M,N ). Proposition 5.2. Let N be an infinite factor with separable predual. Then a semidiscrete extension N ⊂ M can be characterized as in Theorem 3.3 by a generalized Q-system of intertwiners in C = End(N ) (Definition 3.7) if and only if it is discrete. Proof. We begin with necessity. Let (θ, w,{mi}) be a generalized Q-system of intertwiners in C = End(N ) and consider the dual generalized Q-system of intertwiners (γ, w,{Mi}) as in Definition 3.9. By definition M ∗ i eN Mi are mutually orthogonal projections in M1 = hM, eNi which Pi M ∗ i eN Mi = 1. On one hand, MeNM ⊂ m E, where m E denotes the domain of E, because E(eN ) = 1 by [Kos86, Lem. 3.1]. On the other hand, M ∗ i eN Mi ∈ N ′ ∩ M1 by the intertwining property of the Mi on N . Hence M ∗ i eN Mi are mutually orthogonal projections which sum up to 1 in the domain of E↾N ′∩M1 ∈ P (N ′ ∩ M1,N ′ ∩ M). This is equivalent to semifiniteness of E↾N ′∩M1 by [FI99, Lem. 3.2], see also [HKZ91, Lem. 2.2], hence to discreteness of N ⊂ M. The same is true if N ⊂ M is an arbitrary semidiscrete inclusion of von Neumann algebras with separable predual. The converse implication relies on deep results on the structure of N ′∩M1 due to [ILP98]. Consider a discrete inclusion N ⊂ M where N is a factor, M a von Neumann algebra, and choose E ∈ E(M,N ). Then M1 = hM, eNi is a factor and N ⊂ M1 a subfactor. By the same argument leading to [ILP98, Prop. 2.8] we get a decomposition of N ′ ∩ M1 as a direct sum of four algebras, where only the first survives by discreteness assumption and because AdJM,Ω(N ′∩M1) = N ′∩M1, cf. comments after [ILP98, Def. 3.7]. In particular N ′∩M1 is a direct sum of type I factors and PN ⊂ PM1P has finite index for every finite rank projection P ∈ N ′ ∩M1 by [ILP98, Lem. 2.7 (ii)]. Now, arguing as in the proof of [FI99, Thm. 3.5] and using [FI99, Lem. 3.2], see also [ILP98, Prop. 3.2 (ii) ⇒ (i)], by discreteness we can write 13 1 =Pi Pi, i ∈ I, where Pi ∈ N ′ ∩ M1 are non-trivial mutually orthogonal projections such that Pi ∈ m E. Each Pi gives rise to a subendomorphism of the dual canonical endomorphism θ ∈ End(N ) of N ⊂ M. Indeed, Pi and eN are infinite projections in M1 because N is an infinite factor, cf. [FI99, Lem. 3.1], hence we can choose partial isometries Wi ∈ M1 such that W ∗ i Wi = Pi, WiW ∗ i = eN . Then WinW ∗ i = ρi(n)eN , n ∈ N defines ρi ∈ End(N ), ρi ≺ θ, because eNM1eN = N eN , cf. [ILP98, Lem. 3.1]. The endomorphism ρi has finite index, i.e., finite dimension [LR97], whenever Pi has finite rank in N ′ ∩ M1, indeed the inclusion PiN ⊂ PiM1Pi is isomorphic to ρi(N ) ⊂ N . Moreover, θ = ⊕iρi. From Wi = eN WiPi we get WiPi ∈ n E because n E is a left ideal, and Wi ∈ m E = n∗ n E. By the push-down lemma [ILP98, Lem. 2.2] generalized to non-factorial inclusions E [FI99, Lem. 3.3] we can write Wi = eN ψi, where ψi ∈ M, ψi := E(Wi). One can check that ψi is a charged field for ρi, indeed ψin = E(WiPin) = E(WinPi) = E(ρi(n)eN Wi) = ρi(n)ψi, n ∈ N i ) = eN ψiψ∗ i eN = WiW ∗ i = eN , hence E(ψiψ∗ and that eN E(ψiψ∗ i ) = 1 because n 7→ neN is an isomorphism of N onto N eN . Moreover Pi = ψ∗ i eN ψi, hence {ψi} is a Pimsner-Popa basis for N ⊂ M with respect to E. In particular, M = hN ,{ψi}i as in the proof of [FI99, Thm. 4.1], see also [Pop95b, Sec. 1.1.4], [ILP98, Lem. 3.8]. Now, chosen a canonical endomorphism γ for N ⊂ M, thanks to [Lon89, Prop. 5.1] there is an isometry w ∈ N such that w ∈ HomEnd(N )(id, θ), where θ := γ↾N , E(m) = w∗γ(m)w for every m ∈ M, and eN = γ−1 1 (ww∗). Define wi := γ(ψ∗ i )w, Mi := wiψi, i ∈ I where wi ∈ N are such that wi ∈ HomEnd(N )(ρi, θ), cf. [LR95, Sec. 5], and Mi ∈ M have the desired intertwining property with θ, namely Min = θ(n)Mi, n ∈ N , cf. Definition 3.9. Observe that wi are non-trivial isometries w∗ i eN ψi = M ∗ i eN Mi because eN ∈ N ′. As a consequence {Mi} is another Pimsner-Popa basis for N ⊂ M with respect to E and E(MiM ∗ i ) = 1 and that Pi = ψ∗ i ) = wiE(ψiψ∗ i wi = E(ψiψ∗ i = wiw∗ i . Setting i )w∗ mi := γ(Mi), i ∈ I we have that mi ∈ N1 := γ(M) = hθ(N ),{mi}i fulfill mi ∈ HomEnd(N )(θ, θ2), pi := γ1(Pi) = m∗ i ww∗mi are mutually orthogonal projections in N which Pi pi = 1, and nw = 0 ⇒ n = 0 for n ∈ N1 follows immediately from faithfulness of E′ := γ ◦ E ◦ γ−1. Hence (θ, w,{mi}) is a generalized Q-system of intertwiners in C = End(N ) associated, in the sense of Theorem 3.3, to the discrete inclusion N ⊂ M. Remark 5.3. With these normalizations for w and ψi, i ∈ I, we have that ρi(n) = E(ψinψ∗ i ), n ∈ N , i.e., ρi is implemented by a single charged field ψi via the conditional expectation E, cf. [LR95, Sec. 5], [BKLR16, Sec. 4.4]. 14 Proposition 5.4. Let N ⊂ M be a discrete inclusion as in Definition 5.1 and {ψi} ⊂ M a Pimsner-Popa bases of charged fields as in the proof of Proposition 5.2. Then for every m ∈ M, the coefficients E(ψim) ∈ N in the Pimsner-Popa expansion (Proposition 2.3) m =Xi ψ∗ i E(ψim) are uniquely determined. i ) = 1 for every Proof. We have already checked in the proof of Proposition 5.2 that E(ψiψ∗ i ∈ I, hence we can apply Proposition 2.3. Proposition 5.5. Let N be an infinite factor with separable predual and N ⊂ M a discrete extension as in Proposition 5.2. Fix a conditional expectation E ∈ E(M,N ) and a canonical endomorphism γ with dual canonical endomorphism θ = γ↾N . Then a generalized Q-system of intertwiners (θ, w,{mi}) can be chosen such that the set of indices I labels the irreducible subsectors [ρi] (necessarily with finite dimension) of [θ], counted with (arbitrary) multiplicity. There is a distinguished label 0 ∈ I, corresponding to one occurrence of the identity sector [id], such that (9) m0 = θ(w) i.e. θ • 0 θ = • . θ θ θ θ Proof. N ′ ∩ M1 is a direct sum of type I factors by discreteness assumption. Hence we can refine the family of orthogonal projections Pi encountered in the proof of the previous proposition such that each Pi is minimal in N ′ ∩ M1 and again in the domain of E, thus each ρi ≺ θ, i ∈ I, is irreducible (with finite index). Every subsector of [θ] arises in this way and θ = ⊕iρi. The second statement follows by observing that the Jones projection eN is minimal in N ′ ∩M1 if and only if N is a factor, if and only if id is irreducible as an object (tensor unit) of End(N ). Now, by [HKZ91, Lem. 2.2, Prop. 2.4] we can assume that P0 = eN and choose W0 = eN , hence ψ0 = E(eN ) = 1 and w0 = M0 = w, i.e., m0 = θ(w). Remark 5.6. In the assumptions of Proposition 5.2, discreteness of the inclusion N ⊂ M implies [θ] = ⊕i[ρi] where [ρi] are irreducible subsectors with finite dimension and counted with (arbitrary) mul- tiplicity in the set of indices I, cf. comments after [ILP98, Def. 3.7]. If in addition N ⊂ M is irreducible, i.e., N ′ ∩ M = C1, then the multiplicity of each [ρi] in [θ] is finite and bounded above by the square of the dimension of [ρi], see [ILP98, Thm. 3.3, App.]. 15 Remark 5.7. The Pimsner-Popa elements {mi} ⊂ N , or equivalently {Mi} ⊂ M (Definition 3.9), constructed from discrete inclusions via charged fields as in Proposition 5.2 have the following additional properties. Compute w∗mi = w∗γ(wiψi) = wiw∗γ(ψi) = wiw∗ i , hence w∗mi = m∗ i w = wiw∗ i (10) i.e. θ • i θ • = • θ • i θ θ w∗ i ρi , = wi θ θ θ θ • i • =: • • i = θ θ • • i • • i θ for every i ∈ I. Consider the spatial isomorphism θ(N ) ⊂ N = γ1(N ⊂ M1) such that θ(N )′ ∩ N = γ1(N ′ ∩ M1), where γ1 is the canonical endomorphism for M ⊂ M1 dual to γ. From γ1(Pi) = wiw∗ i ) we conclude that i = E(MiM ∗ pi = qi where pi = γ1(Pi), and qi = E(MiM ∗ mutually orthogonal projections in N such that Pi qi = 1 as well. i ) are defined in Lemma 2.2. In particular qi, i ∈ I, are 0 = If we consider {Mi} constructed as in Proposition 5.5 we have in addition w∗m0 = w0w∗ ww∗ and i.e. (11) θ(w∗)mi = δi,01 θ • i • θ = δi,0 θ θ for every i ∈ I. Moreover E(Mi) = E(MiM ∗ 0 )M0 = δi,0M0. Definition 5.8. We say that a generalized Q-system of intertwiners (Definition 3.7) is unital, if it satisfies in addition the analogue of equations (9), (10), (11), namely m0 = θ(w), w∗mi = m∗ i w = (w∗mi)2, θ(w∗)mi = δi,01 for every i ∈ I, and for a distinguished label 0 ∈ I. One can easily check that the braided product (Definition 4.1) of two unital generalized Q-systems of intertwiners is again unital. 16 6 Generalized Q-systems of intertwiners for local nets Let {A} = {O ∈ K 7→ A(O)} be a net of infinite von Neumann factors (typically of type III1) over a partially ordered by inclusion and directed set K of open bounded regions O of spacetime (e.g., the set of open proper bounded intervals O = I ⊂ R, or double cones in Minkowski space O ⊂ Rn+1, n ≥ 1). A net is called isotonous if O ⊂ O implies A(O) ⊂ A( O), and local if A(O) and A( O) commute elementwise whenever O ⊂ O′, where O′ denotes the space-like complement of O in Rn+1, n ≥ 1, or the interior of the complement of the interval O = I in R. Definition 6.1. A net {A} as above fulfilling isotony and locality is called a net of local observables, also abbreviated as local net. We refer to [Haa96], [LR95, Sec. 3], [CCG+04, Ch. 5] for more explanations and for the physical motivations behind this notion. Now, let {A} be realized on a separable Hilbert space H0 (vacuum space) and assume the existence of a unit vector Ω0 ∈ H0 (vacuum vector) which is cyclic and separating for each local algebra A(O). In this case, we say that {A} is a standard net on H0 with respect to Ω0 and denote by ω0 := (Ω0 · Ω0) the vacuum state of the net. We say that Haag duality holds for {A} in the vacuum space if A(O′)′ = A(O) for every O ∈ K, where A(O′) is the C ∗-algebra generated by all A( O), O ∈ K, O ⊂ O′. Denote by DHR{A} ⊂ End(A) the category of DHR endomorphisms of the net, see [DHR71], [DHR74], [FRS92], and by A the quasilocal algebra, i.e., the C ∗-algebra generated by {A}. In the following we shall be interested in two distinguished subcategories of the DHR category, namely Definition 6.2. Denote by DHRf{A} and DHRd{A} the full subcategories of DHR{A} whose objects are, respectively, finite-dimensional DHR endomorphisms and (possibly infi- nite, countable) direct sums of those. i ∈ ρ(A)′ ∩A(O) are mutually orthogonal projections and Pi wiw∗ j tPi wiw∗ More precisely the most general object ρ in DHRd{A} arises as follows. Let ρi be a family of at most countably many irreducible finite-dimensional DHR endomorphisms which can be localized in O ∈ K. Let {wi} be a (possibly infinite) Cuntz family of isometries in A(O) such that wiw∗ i = 1. Then ρ = Pi Adwi ρi, where ρi =: Adw∗ ρ and the sum converges elementwise in the strong operator topology because ρ = Pi wiw∗ i . Similarly, the most general arrow t between objects ρ, σ in DHRd{A} can be written as t =Pj vjv∗ i = Pi,j vjtj,iw∗ j twi are arrows from ρi to σj. Remark 6.3. Observe that DHRf{A} ⊂ DHRd{A} ⊂ DHR{A} and each inclusion is full, replete and stable under (finite) direct sums and subobjects. The first two categories are semisimple in the sense that every object can be written as a (possibly infinite) direct sum of irreducible finite-dimensional objects. i where {wi}, {vj} are Cuntz families, respectively, for ρ, σ and tj,i := v∗ i i ρ(·)Pj wjw∗ j = Pi wiρi(·)w∗ The following is the net-theoretic version of Definition 3.7, and generalizes the notion of Q-system for local nets given in [LR95, Sec. 4]. 17 Definition 6.4. Let {A} be a local net fulfilling Haag duality as above. A generalized net Q-system of intertwiners in C = DHR{A} is a triple (θ, w,{mi}) consisting of a DHR endomorphism θ in DHR{A}, an isometry w ∈ HomDHR{A}(id, θ), and a family {mi} ⊂ A indexed by i in some set I, such that (i) pi := m∗ i ww∗mi are mutually orthogonal projections in A, i.e., pip∗ j = δi,jpi, such that Pi pi = 1. of θ and for any other O ∈ K such that O ⊂ O. (ii) aw = 0 ⇒ a = 0 if a ∈ A1(O) := hθ(A(O)),{mi}i for some localization region O ∈ K (iii) mi ∈ HomDHR{A}(θ, θ2), i ∈ I. Remark 6.5. By the localization property of θ and by Haag duality, (θ, w,{mi}) is a gener- alized Q-system of intertwiners in End(A( O)) (Definition 3.7) for every O as above. Indeed, DHR{A} sits into End(A( O)) via the restriction functor as a (full if local intertwiners are global), replete and braided tensor subcategory for every such O, cf. [GR15, Sec. 3]. Remark 6.6. Condition (iii) in Definition 6.4, in view of Proposition 5.2, excludes many interesting infinite index extensions of local nets. Notably the Virasoro net {Virc} in one spacetime dimension, which sits in every conformal (diffeomorphism covariant) net, gives often rise to infinite index semidiscrete but non-discrete extensions if c > 1, see [Reh94], [Car04], [Xu05]. It is however fulfilled in many examples of chiral conformal embeddings with infinite index, see Section 10, as in compact orbifold theories in low and higher dimensions, see [Xu00] and [DR90], and of course in every finite index extension. Definition 6.7. [LR95]. An inclusion of nets is defined by two isotonous nets of von Neumann algebras {A}, {B} over the same partially ordered set of spacetime regions K and realized on the same separable Hilbert space H such that A(O) ⊂ B(O) for every O ∈ K. In this case, we write {A ⊂ B} and call {B} an extension of {A}. The inclusion is called irreducible if A(O)′ ∩ B(O) = C1 for every O ∈ K. The net {B} is relatively local with respect to {A} if B(O) ⊂ A(O′)′ for every O ∈ K. If {A} is local, {B} will be always implicitly assumed to be relatively local with respect to {A}. The inclusion of nets is called standard if there is a vector Ω ∈ H which is standard for {B} on H and for {A} on a subspace H0 ⊂ H. A normal faithful conditional expectation E of {B} onto {A} is a family indexed by O ∈ K of normal faithful conditional expectations EO ∈ E(B(O),A(O)) which respect inclusions, namely such that E O ↾B(O) = EO if O ⊂ O. A normal faithful state ω of {B} is a conditional expectation of {B} onto the trivial net {C} and E as above is called standard if it preserves the standard vector state ω := (Ω · Ω) of the net, namely ωO ◦ EO = ωO for every O. We say that the extension {A ⊂ B} is discrete if A(O) ⊂ B(O) is discrete (Definition 5.1) for every O ∈ K. The following theorem extends the results of [LR95, Thm. 4.9] to the case of infinite index discrete inclusions of nets of von Neumann algebras. Theorem 6.8. Let {A} be a local net fulfilling Haag duality and standardly realized on H0 as in the beginning of this section. Then a generalized net Q-system of intertwiners (θ, w,{mi}) 18 in C = DHR{A} (Definition 6.4) which is also unital (Definition 5.8) gives an isotonous net of von Neumann algebras {B} such that {A ⊂ B} is a discrete standard inclusion of nets with a normal faithful standard conditional expectation. The net {B} is always relatively local with respect to {A}, and it is itself local if and only if θ(εθ,θ)mimj = mjmi for every i, j ∈ I, where ε denotes the DHR braiding. Proof. Let O ∈ K be a localization region of the DHR endomorphism θ, call N := A(O) and θ ≡ θ↾N ∈ End(N ) the restriction of θ to N , and observe that w ∈ N , mi ∈ N for every i ∈ I by Haag duality. From Theorem 3.3 we get N2 ⊂ N1 ⊂ N with a normal faithful conditional expectation E′ := θ(w∗ · w) ∈ E(N1,N2) and such that θ is a canonical endomorphism for N1 ⊂ N . Now N acts standardly on H0 by assumption hence θ = AdΓ on N , where Γ := JN1,ΦJN ,Φ and Φ ∈ H0 is cyclic and separating for N1 and N . Let M := AdΓ∗(N1) be the corresponding canonical extension of N1 ⊂ N with canonical endomorphism γ := AdΓ↾M. Lift accordingly the conditional expectation E := w∗γ(·)w ∈ E(M,N ) and consider the normal faithful E-invariant state ϕ := ω0 ◦ E of M, where ω0 = (Ω0 · Ω0) is the vacuum state of {A}. The operators Mi := γ−1(mi) ∈ M as in Definition 3.9 form a Pimsner-Popa basis for N ⊂ M with respect to E and fulfill Min = θ(n)Mi, (12) i ∈ I, n ∈ N . Now consider the (normal faithful) GNS representation (Hϕ, πϕ, Ωϕ) of M with respect to ϕ = ϕ ◦ E. The inclusion πϕ(N ) ⊂ πϕ(M) on Hϕ with conditional expectation Eϕ given by Eϕ(πϕ(m)) := πϕ(E(m)), m ∈ M, is spatially isomorphic to N ⊂ M on H0 with respect to E. Moreover, (ΩϕEϕ(πϕ(m))Ωϕ) = (Ωϕπϕ(m)Ωϕ), m ∈ M and eN := [πϕ(N )Ωϕ] is the associated Jones projection. By spatial isomorphism we have that {πϕ(Mi)} is a Pimsner- Popa basis for πϕ(N ) ⊂ πϕ(M) with respect to Eϕ, and γϕ given by γϕ(πϕ(n)) := πϕ(γ(n)), n ∈ N , is a canonical endomorphism with dual canonical θϕ := γϕ↾πϕ(N ). In particular, we have a direct sum decomposition Hϕ =Xi πϕ(M ∗ i )eNHϕ where πϕ(M ∗ projections and we let i )eN , i ∈ I, are partial isometries with mutually orthogonal range and domain H0,N ,ϕ := eNHϕ. i )ψi =Xi Every n ∈ N ⊂ M acts by left multiplication on Hϕ, then πϕ(M ∗ πϕ(n)Xi where ψ =Pi πϕ(M ∗ i )ψi, with ψi := eN πϕ(Mi)ψ ∈ πϕ(qi)eNHϕ, is the generic vector of Hϕ. As in the proof of [LR95, Thm. 4.9], this representation of N = A(O) extends to the whole net. Indeed, the linear map i )πϕ(θ(n))ψi πϕ(M ∗ (13) U0 : nΩ0 7→ πϕ(n)Ωϕ, n ∈ N extends to a unitary operator from H0 onto H0,N ,ϕ, due to ϕ ≡ ω0 ◦ E and E(n) = n, n ∈ N , which implements πϕ↾N on the subspace H0,N ,ϕ via adjoint action. For every quasilocal observable a ∈ A and ψi as above, define πϕ(a)Xi πϕ(M ∗ i )ψi :=Xi πϕ(M ∗ i )U0θ(a)U ∗ 0 ψi. 19 One can check that πϕ is a well-defined bounded and locally normal representation of A on Hϕ which extends the GNS representation restricted to N due to equation (13). In this representation, the intertwining relation (12) extends to the net, namely πϕ(Mi)πϕ(a) = πϕ(θ(a))πϕ(Mi), i ∈ I, a ∈ A. (14) To show this, we first check that in the representation on Hϕ we have that e N := [πϕ( N )Ωϕ], where N := A( O), fulfills eN = e N for every O ∈ K (not necessarily O ⊂ O). Indeed, Ωϕ = πϕ(M ∗ 0 w = w∗w = 1 by unitality assumption, and the closed linear span in Hϕ of vectors of the form 0 )πϕ(w)Ωϕ because M ∗ πϕ(a)Ωϕ = πϕ(a)πϕ(M ∗ 0 )πϕ(w)Ωϕ = πϕ(M ∗ 0 )U0θ(a)wU ∗ 0 Ωϕ = πϕ(M ∗ 0 )U0waΩ0 = U0aΩ0 does not depend on whether a ∈ N or a ∈ N by the intertwining property of w on A and because Ω0 is cyclic for every local algebra on H0 by standardness assumption. Hence e NHϕ = H0,N ,ϕ ⊂ Hϕ for every O ∈ K. Now let ψ and ψi be as in equation (13), and assume that a ∈ A( O) for some O ⊂ O. From the l.h.s. of (14) we get πϕ(Mi)πϕ(a)ψ =Xj πϕ(MiM ∗ j )U0θ(a)U ∗ 0 ψj because left multiplication is continuous in the GNS representation. By Proposition 2.3 (valid for arbitrary semidiscrete inclusions) we can write πϕ(Mi)πϕ(Mj )∗ =Xk πϕ(Mk)∗πϕ(lki j ), lki j := E(MkMiM ∗ j ) ∈ N (15) j = w∗γ(MkMiM ∗ j )w = w∗mkmim∗ j w and intertwines θ with θ2 on the whole net by where lki assumption, i.e. lki j ∈ HomDHR{A}(θ, θ2). Recall that the convergence in the r.h.s. of equation (15) is given by the topology induced by η = η(m∗m), m ∈ πϕ(M), with η any normal state on πϕ(M) such that the seminorms kmk2 η = η ◦ Eϕ. Thus Xj 0 ψj =Xj (cid:0)Xk j )(cid:1)U0θ(a)U ∗ j )U0θ(a)U ∗ πϕ(MiM ∗ k )πϕ(lki πϕ(M ∗ 0 ψj and since the vector U0θ(a)U ∗ πϕ(M), we get =Xj,k 0 ψj ∈ e NHϕ = eNHϕ induces a normal Eϕ-invariant state on πϕ(M ∗ k )U0θ2(a)lki j U ∗ 0 ψj = πϕ(θ(a))πϕ(Mi)ψ which is the r.h.s. of (14), for every ψ ∈ Hϕ, thus the equation is proven. We define B(O) := πϕ(M) ≡ hπϕ(N ),{πϕ(Mi)}i, 20 the crucial point is to extend the construction to bigger regions and define accordingly a co- herent family of normal faithful standard conditional expectations with respect to a common cyclic and separating vector. Let O ∈ K be such that O ⊂ O and set N := A( O), B( O) := hπϕ( N ),{πϕ(Mi)}i, clearly B(O) ⊂ B( O) holds by isotony of {A}. onto πϕ( N )eN , and because of Now, Ω0 is separating for every N , thus πϕ(a) 7→ πϕ(a)eN is an isomorphism of πϕ( N ) eNB( O)eN ⊂ eNhB( O), eNieN = eNhπϕ( N ), πϕ(M), eN ieN = πϕ( N )eN provided O ⊂ O, we can define by eN beN = Eϕ(b)eN , b ∈ B( O) a conditional expectation of B( O) onto πϕ( N ) (for arbitrarily big region O) which extends the one previously given on B(O) ≡ πϕ(M). Eϕ is clearly normal and fulfills (ΩϕEϕ(·)Ωϕ) = (Ωϕ · Ωϕ), while faithfulness remains to be checked, together with the separating property of Ωϕ for B( O) if O ⊂ O and cyclicity for B( O) if O ⊂ O, where B( O) in this second case is defined below. For an arbitrary region O ∈ K, set B( O) := hπϕ( N ),{πϕ(u)πϕ(Mi)}i, (16) where u ∈ HomDHR{A}(θ, θ) is a unitary charge transporter (in A) and θ is DHR localizable in O, 3. In order to show the desired properties of Ωϕ with respect to these new local algebras we need to introduce more GNS representations. Namely, let N := A( O) (θ ≡ θ↾ N , w := uw,{ mi := uθ(u)miu∗}) be a generalized Q-system of intertwiners in End( N ), see Remark 6.5, and perform the same construction as above on the GNS Hilbert space H ϕ of some canonical extension N ⊂ M with E ∈ E( M, N ) and state ϕ := ω0 ◦ E of M. Consider )e NH ϕ π ϕ( Mi ∗ H ϕ =Xi and extended as before to an isometric operator into H ϕ. Then the linear map defined by ∗ i )U0u∗φi, φi ∈ H0 U0 : nΩ0 7→ π ϕ(n)Ω ϕ, H0 → H0, N , ϕ ≡ e NH ϕ, UXi ) U0φi :=Xi πϕ(M ∗ π ϕ( Mi n ∈ N sends U Ω ϕ = U π ϕ( M0 ∗ ) U0 wΩ0 = πϕ(M ∗ 0 )U0u∗ wΩ0 = Ωϕ because u∗ w = w, it extends to a unitary operator from H ϕ onto Hϕ, and fulfills U π ϕ( M)U ∗ = hπϕ( N ),{πϕ(u)πϕ(Mi)}i ≡ B( O). (17) 3Notice that we assume DHR sectors to be localizable in every region in K. 21 Indeed, if n ∈ N then U π ϕ(n)U ∗Xi =Xi πϕ(M ∗ πϕ(M ∗ i )U0φi = U π ϕ(n)Xi i )U0u∗ θ(n)uφi = πϕ(n)Xi ∗ π ϕ( Mi ) U0uφi πϕ(M ∗ i )U0φi because u∗ θ(n)u = θ(n), and U π ϕ( Mi)U ∗Xj πϕ(M ∗ j )U0φj =Xj,k πϕ(M ∗ k )U0u∗lki j uφj where the coefficients lki j One can compute θ(u∗)u∗lki j u = lki j , hence := E( Mk Mi M ∗ j ) = w mk mi m∗ j w ∈ N are analogous to those in (15). =Xj,k πϕ(M ∗ k )U0θ(u)lki j φj = πϕ(u)πϕ(Mi)Xj πϕ(M ∗ j )U0φj which proves (17). By considering this unitary intertwiner for every region O we obtain that Ωϕ is cyclic and separating for every B( O) on Hϕ, in particular Eϕ is faithful over every O. The extension {B} does not depend on the specific choice of unitary charge transporter u made in equation (16). Indeed, by Haag duality any two of them u, v differ by uv∗ ∈ A( O). Also, it depends on the choice of the initial localization region O for θ and of the extended vacuum state ϕ only up to unitary isomorphism. Relative locality of {B} with respect to {A} is always guaranteed by the localization properties of θ while the statement about locality of {B} follows by the very definition of the DHR braiding. Indeed, uMivMj = vMjuMi where u and v are unitaries in A which transport the localization region of θ to two mutually space-like regions (respectively left and right localized in low dimensions) if and only if εθ,θMiMj = θ(u∗)v∗uθ(v)MiMj = MjMi for every i, j ∈ I, and the proof is complete. Remark 6.9. If the Pimsner-Popa expansion appearing in equation (15) comes from an irre- ducible subfactor, or from a finite index inclusion, and if the unital generalized Q-system of intertwiners is defined from charged fields Mi = wiψi, i ∈ I, see Proposition 5.5 and Remark 5.6, then the sum over k (a priori convergent in the GNS topology) is finite by Frobenius reciprocity among finite-dimensional endomorphisms of N , [LR97, Lem. 2.1]. Indeed, in this case one has lki In other words, we have a unital -algebra of charged intertwiners with possibly infinitely j ) ∈ HomEnd(N )(ρj, ρkρi). j = wkρk(wi)E(ψkψiψ∗ j and E(ψkψiψ∗ j )w∗ many generators {Mi} but finite ("discrete") fusion rules, cf. [LR04, App. A]. In this section we assumed that a generalized net Q-system of intertwiners (Definition 6.4) was given and we have shown how to associate to it a relatively local net extension. In Section 5, given a discrete inclusion of von Neumann algebras, we have seen how to construct a generalized Q-system of intertwiners (Definition 3.7). But when do generalized Q-systems of intertwiners exist for discrete relatively local net extensions? In the finite index setting, in [GL92] it is proven that the DHR category restricted to finite index endomorphisms of a chiral CFT is a full and replete subcategory of the category of endomorphisms of a local algebra ("local intertwiners are global"). This of course single- handedly carries over the theory of ordinary Q-systems to such nets. In [LR95] it is shown 22 in the presence of a coherent conditional expectation, with that actually less is needed; finite index arguments it can be shown that a Q-system in the DHR category does exist, see [LR95, Cor. 3.8, Cor. 3.7 and Lem. 4.1]. On the other hand, when we consider also infinite-dimensional irreducible sectors, in general it is not true that the category of DHR endomorphisms is a full subcategory of the category of endomorphisms of a local algebra, see [Wei08] for a counter-example using {Virc}, c > 2. In the absence of these features, we can make some additional assumptions. Proposition 6.10. Let {A ⊂ B} be a discrete relatively local inclusion of nets with standard conditional expectation E (Definition 6.7) and let {A} be a Haag dual net of type III factors. Let γ be a canonical endomorphism for {A ⊂ B} as in [LR95, Cor. 3.3] and θ ∈ DHR{A} be its dual canonical endomorphism as in [LR95, Cor. 3.8] (localized in O0 ∈ K). Moreover, let w ∈ A(O0) as in [LR95, Cor. 3.7] such that E = w∗γ(·)w. Assume that one of the following two conditions is fulfilled (i) {A ⊂ B} is irreducible and θ = ⊕iρi in DHR{A} where each ρi is an irreducible DHR subendomorphism (localized in O0) and [ρi] has finite dimension in DHR{A}. (ii) HomDHR{A}(θ, θ) = HomEnd(A(O))(θ, θ) for every O ∈ K, O0 ⊂ O, or equivalently θ = ⊕iρi in DHR{A} and HomDHR{A}(ρi, ρj) = HomEnd(A(O))(ρi, ρj) for every i, j ∈ I. Then there is a Pimsner-Popa basis {ψi} for the inclusion A(O0) global charged fields, i.e. E ⊂ B(O0) consisting of ψia = ρi(a)ψi, a ∈ A. E i = pi. i )wψi we have another Pimsner-Popa basis for A(O0) Setting Mi := γ(ψ∗ ⊂ B(O0) (cf. Propo- sition 5.2), and (θ, w,{mi}), where mi := γ(Mi) ∈ A(O0), is a unital generalized net Q- system of intertwiners in DHR{A} (Definition 6.4 and 5.8). Proof. Assume (i) and let pi ∈ HomDHR{A}(θ, θ) be the projections which determine the decomposition θ = ⊕iρi, together with orthogonal isometries wi ∈ HomDHR{A}(ρi, θ)∩A(O0) such that wiw∗ For O ∈ K, let eA = [A(O)Ω] be the Jones projection, and B1(O) = hB(O), eAi the Jones extension for the inclusion A(O) ⊂ B(O). By standardness assumption the Jones projections agree for every O ∈ K. Denote by EO ∈ P (B1(O),B(O)) the operator-valued weight dual to be respectively its domain and definition ideal. By the same E↾B(O) and let m EO arguments leading to [LR95, Thm. 3.2, Cor. 3.3], the formula and n EO x ∈ B1(O) γ1(x)eA = v1xv∗ 1, where v1 ∈ B1(O) is an isometry such that v1v∗ 1 = eA, γ1(v1) = w, allows to extend the dual canonical endomorphism γ1 : B1(O0) → A(O0) to a map between the quasilocal algebras B1 and A, such that γ1↾B1(O) is the dual canonical endomorphism for the inclusion A(O) ⊂ B(O) for every O ∈ K, O0 ⊂ O (i.e., γ1↾B1(O) is a canonical endomorphism for B(O) ⊂ B1(O), γ1↾B(O) = γ↾B(O) and γ1(B1(O)) = A(O)). By the discreteness assumption and [ILP98, Prop. 2.8] we have that, for every O ∈ K, B1(O)∩A(O)′ is a direct sum of type I factors, which by irreducibility of A(O) ⊂ B(O) and [ILP98, Thm. 3.3] are finite dimensional. 23 Pi := γ−1 ρi↾A(O) has finite dimension by hypothesis, and thus Pi ∈ n EO support of Pi in B1(O) ∩ A(O)′. EO↾zPi (B1(O)∩A(O)′)zPi as a sum of minimal projections in n EO dimensional. Now, let 1 (pi) is a finite sum of minimal projections in B1(O) ∩ A(O)′, O0 ⊂ O, since . Indeed, let zPi be the central is semifinite since zPi can be written , and thus finite since zPi(B1(O) ∩ A(O)′)zPi is finite 1 (ww∗ . Let ψi,O := EO(Wi). Exactly as in Since Wi = eAWiPi and Pi ∈ n EO the proof of Proposition 5.2, and using wi = γ(ψ∗ i,O)w, we get that ψi,O is a charged field for ρi on A(O), i.e., ψi,Oa = ρi(a)ψi,O, a ∈ A(O), and the collection {ψi,O} is a Pimsner-Popa basis for the inclusion A(O) ⊂ B(O). For any O ∈ K, O0 ⊂ O, by the push-down lemma we have Wi := γ−1 , we have Wi ∈ m EO i ) ∈ B1(O0). Thus applying EO to the above formula, we obtain ψi,O = ψi,O0. The rest of the proof follows exactly as in Proposition 5.2. eAψi,O = Wi = eAψi,O0. Assuming (ii), the proof proceeds along similar lines. By discreteness assumption, one can take projections pi ∈ θ(A(O0))′ ∩ A(O0) which lay in n E′ = γ ◦ E↾B(O0) ◦ γ−1, and which give a local (in this case also global) decomposition of θ = ⊕iρi into DHR subendomorphisms. Now, let wi such that pi = wiw∗ i and define Pi, Wi and ψi,O0 as before such that Wi = eAψi,O0. To conclude it is enough to observe that for any O ∈ K, O0 ⊂ O, eAψi,O0 by construction and ψi,O0 ∈ B(O0) ⊂ B(O), we have Pi ∈ n EO and also Pi ∈ B1(O) ∩ A(O)′, because θ(A(O0))′ ∩ A(O0) = θ(A(O))′ ∩ A(O) for every such O. , because Pi = ψ∗ , where E′ i,O0 O0 O0 These assumptions are verified, e.g., for compact group orbifolds [DR72], [Rob74, Thm. 4.3], and for theories with a good behaviour with respect to the scaling limit [DMV04, Cor. 6.2] in 3+1D, and of course for strongly additive CFTs in 1D, i.e., Haag dual nets on R [GLW98, Lem. 1.3]. 6.1 Construction of extensions: an alternative way In this section we present an alternative proof of Theorem 6.8 which we feel somewhat more intuitive and which lends itself to describing the braided product of nets in a more direct way. The basic idea is that a generalized net Q-system of intertwiners (θ, w,{mi}) in DHR{A} (Definition 6.4), assuming Haag duality of {A}, induces a family of Q-systems (one for every local algebra) each one of which characterizes a local extension. The algebraic structure of the extended net, including a distinguished conditional expectation, is then captured with a coherent inductive procedure and the spatial features of the net are completely determined by the vacuum state. More precisely, let O0 ∈ K be a reference localization region for θ and choose a unitary charge transporter uO ∈ HomDHR{A}(θ, θO) for every O ∈ K, where θO := AduO θ is localized in O. For every O ∈ K, we obtain a generalized Q-system (of intertwiners) in End(A(O)) (Definition 3.7) by setting (θO, wO,{miO}) := (AduO θ, uOw,{uOθ(uO)miu∗ O}). 24 The faithfulness condition appearing in Definition 3.1 is verified since, for every O ∈ K such that O0 ∪ O ⊂ O, we have where A1(O) ⊂ uOhθ(A( O)), miiu∗ O = AduO (A1( O)), A1(O) := hθO(A(O)), uOθ(uO)miu∗ Oi. With this data, we now construct an inductive generalized sequence of net extensions {{B O}, O ∈ K} of {A}, indexed by O ∈ K, and defined only on regions O ∈ K, O ⊂ O, which we will then patch together. For a fixed O ∈ K, θ O = Adu O θ is a canonical endomorphism for A1( O) ⊂ A( O) by Theorem 3.3, and thus can be implemented on A( O) by a unitary Γ O, namely Similarly, for every O ⊂ O, we have θ O(x) = Γ OxΓ∗ O , x ∈ A( O). θO(x) = AduO (θ(x)) = AduOu∗ O Γ O (x) , x ∈ A( O). Now, fixed O ∈ K, we define an isotonous net {B O} := {O ∈ K, O ⊃ O 7→ B O(O)} by setting B O(O) := Ad(uOu∗ O Γ O)∗(A1(O)) = hA(O),{uO AdΓ∗ O u O (mi)}i. Remark 6.11. The dual canonical endomorphism θ for an extension of nets {A ⊂ B}, [LR95, Cor. 3.8], is not implemented globally by unitaries. This is clear since by [LR95, Prop. 3.4] the embedding homomorphism ι of {A} into {B} is equivalent to θ as a representation and thus would imply the inclusion to be trivial. Of course it is possible to find unitaries which implement θ locally but the choice of these unitaries in non-unique. The above coherent choice guarantees the isotony of the net {B O}. The next step is the construction of an inductive family of embeddings ι O1, O2 of {B O1} into {B O2} with O1 ⊂ O2. This is straightforward. Proposition 6.12. Let O1 ⊂ O2, O1, O2 ∈ K. The map Γ O1 ι O1, O2 := AdΓ∗ O2 u O2 u∗ O1 is an embedding of {B O1} into {B O2}, i.e., it sends local algebras B O1 B O2 Proof. Follows by easy direct computation. (O) and acts as the identity map on A(O) for every O ∈ K, O ⊂ O1. (O) onto local algebras The collection of nets {{B O}, O ∈ K} and maps {ι O1, O2 , O1 ⊂ O1, O1, O2 ∈ K} forms an inductive system. We can thus take the inductive limit of the C ∗-algebras B O( O) from which we obtain a C ∗-algebra B. The subalgebras B(O) := ιO(BO(O)), where ιO is the embedding of BO(O) into B, are W ∗-algebras since it is easy to see that they have a predual. Thus we have obtained an isotonous net of W ∗-algebras, {B}. Now we see that from the data of the Q-system we can also define a consistent family of conditional expectations. Proposition 6.13. There is a normal faithful conditional expectation from {B} to {A}, i.e., EO : B(O) → A(O) for every O ∈ K, such that EO2 ↾B(O1) = EO1 if O1 ⊂ O2. 25 Proof. First define a coherent conditional expectation on {B O} for O ∈ K. By Theorem 3.3 we have a conditional expectation E′ O·uOw) for the inclusion θO(A(O)) ⊂ A1(O). O can be lifted to a conditional expectation E OO for the inclusion A(O) ⊂ B O(O), O ⊂ O, E′ since the two inclusions are isomorphic via AduOu∗ O . Computing explicitly, we get O := θO(w∗u∗ Γ O E OO = w∗u∗ OΓ O · Γ∗ Ou Ow which shows that we indeed have a consistent family of conditional expectations on {B O}. Now, to show that these expectations lift to the inductive limit net {B}, it is enough to check that E O2O ↾ι O1, O2 = E O1O for O1 ⊂ O2, but this is a trivial computation. (B O1 (O)) If ω0 is the vacuum state of {A}, let ω := ω0 ◦ E, where E is the consistent conditional expectation of the inclusion {A ⊂ B} defined above, lifted to the quasilocal C ∗-algebra B. We call ω the vacuum state of {B} and the GNS representation induced by ω the vacuum representation. We denote by {A ⊂ BQ} and {BQ} the extension constructed in this way, in its vacuum representation. Remark 6.14. It is not hard to check that the construction of the net {BQ} and its conditional expectation EQ onto {A} does not depend on the choice of the family of unitary charge transporters uO, nor on the choice of Γ O. Up to now, we have seen that we can build (discrete, relatively local) extensions of nets {A ⊂ BQ} associated to generalized net Q-systems of intertwiners in DHR{A}. A natural question to ask is if this procedure insures that, if the Q-system comes from a given extension {A ⊂ B}, the induced extension will be unitarily equivalent to the starting one. The answer is affirmative when the generalized net Q-system is constructed as in Proposition 6.10. Lemma 6.15. Let {A ⊂ B} be as in Proposition 6.10, assuming either (i) or (ii), and let θ = ⊕imiρi be a decomposition of θ into irreducibles in DHR{A}, where [ρi] 6= [ρj], mi is the multiplicity of [ρi] in [θ], and ρi, θ are localized in O ∈ K. Then Hρi(O) := {ψ ∈ B(O), ψa = ρi(a)ψ, a ∈ A} is isomorphic as a Hilbert space to HomDHR{A}(ρi, θ) via the map Φ : ψ 7→ γ(ψ∗)w. Proof. Note that E(ψ1ψ∗ 2) ∈ HomDHR{A}(ρi, ρi), ψ1, ψ2 ∈ Hρi(O), is an inner product for Hρi(O) since ρi is irreducible in DHR{A}. We have seen in Proposition 6.10 that there is a collection {ψj} ⊂ Hρi(O), j = 1, . . . , mi (with mi possibly infinite), which is orthonormal with respect to the above inner product, and which is mapped via Φ : ψ 7→ γ(ψ∗)w onto an 2) = Φ(ψ1)∗Φ(ψ2), the map Φ : ψ 7→ orthonormal basis of HomDHR{A}(ρi, θ). Since E(ψ1ψ∗ γ(ψ∗)w is an isomorphism of Hρi(O) onto HomDHR{A}(ρi, θ). Proposition 6.16. Let {A ⊂ B} and (θ, w,{mi}) be as in Proposition 6.10, assuming either (i) or (ii). Then the inclusion {A ⊂ BQ} obtained from (θ, w,{mi}) is unitarily equivalent to {A ⊂ B}. Proof. We first show that A1(O) = γO(B(O)), A1(O) := hθO(A(O)), uOθ(uO)miu∗ Oi 26 where γO := AduO γ is a canonical endomorphism for the inclusion A(O) ⊂ B(O). To see ∗uOwi this, let wi ∈ A(O) be isometries such that wi wi is a unitary and ρi,O := Ad wi ∗uOwi ρi is localized in O. Using Lemma 6.15, we have that O. Then we have that wi ∗ = uOwiw∗ i u∗ Hρi(O0) = Hρi( O) = wi ∗uOwiHρi,O ( O) = wi ∗uOwiHρi,O (O) for every O ∈ K with O0 ∪ O ⊂ O. Consequently uOMi = uOwiψi = uOwiw∗ O)(uOwiψi) i wiψi = (uOwiw∗ i u∗ ∗)(uOwiψi) = wi ψi ∈ B(O) = ( wi wi from which the claim easily follows. Thus it is clear that for a fixed O ∈ K, the map π O : B( O) → BQ ( O), π O := AdΓ∗ γ O, is an isomorphism of von Neumann algebras, which maps B(O) onto BQ (O) and which lifts to a representation π of the net {B}. To show that this representation is unitarily equivalent to the vacuum representation, it is enough to show by the GNS theorem that O O O EQ ◦ π = E with EQ and E respectively the conditional expectations of {A ⊂ BQ} and {A ⊂ B}, but this is clear using E = w∗γ(·)w. Remark 6.17. Note that in our second construction of the net {BQ}, the only instance where the intertwining property of the mi was used is to make sure that miO = uOθ(uO)miu∗ O ∈ A(O) (precisely by the intertwining property of the mi and Haag duality of {A}, cf. Remark 6.5). If a generalized Q-system does not have the intertwining property then the isotonous, rel- atively local net extension {BQ} can still be defined in the same way, although the conditional expectation EQ cannot be defined on regions O 6⊃ O0 since in general miO = uOθ(uO)miu∗ need not be in A(O), and thus (θO, wO,{miO}) is not a priori a generalized Q-system in End(A(O)). O 7 Covariance of extensions In this section we show how spacetime covariance (e.g., Mobius covariance in 1D or Poincar´e covariance in 3+1D) extends from {A} to {B}, where {A} is a local covariant net over a directed set of spacetime regions K, and {A ⊂ B} is an extension with the properties implied by Theorem 6.8. This fact is common knowledge among experts, cf. [KL04, Rmk. 4.3] for irreducible extensions of {Virc}, c < 1, [BMT88, Sec. 3C] for time translation covariance in extensions of the chiral U (1)-current, and [MTW16, Sec. 6] for more recent examples of diffeomorphism covariant extensions of the U (1)-current in 1+1D. See also [DR90, Sec. 6], [DR89, Thm. 8.4] for the covariance of canonical field extensions in 3+1D. In this section, see Theorem 7.7, we give a general proof of covariance for extensions of nets, with finite or infinite index (of discrete type as in Theorem 6.8). The proof essentially relies on tensoriality and naturality properties of the action of the spacetime symmetry group (implemented by covariance cocycles) on the DHR category. Hence we formulate it in a C ∗ tensor categorical language, cf. [Tur10, App. 5] due to M. Muger. But first we need a few definitions. Let P be a (pathwise) connected and simply connected group of spacetime symmetries (e.g., P = gMob the universal covering of the Mobius group acting on R (actually on R = 27 R ∪ {∞}), or P = P↑ + the universal covering of the proper orthochronous Poincar´e group acting on R3+1). Assume that P contains a distinguished (n + 1)-parameter subgroup, n ≥ 0, of "spacetime translations" (e.g., the rotations inside gMob, or the four-dimensional spacetime translations inside P↑ +). The following definition describes Poincar´e covariant theories on Minkowski space and Mobius covariant theories on the real line at the same time, cf. [GL92, Sec. 8], [BGL93, Sec. 1], [CKL08, Sec. 3]. Definition 7.1. An isotonous net {A} of von Neumann algebras realized on H0 over a directed set of spacetime regions K is called covariant with respect to P if there is a strongly continuous unitary representation g 7→ U (g) of P on H0 such that U (g)A(O)U (g)∗ = A(gO), O ∈ K, g ∈ UO where UO ⊂ P denotes the (pathwise) connected component of the identity e in P of the set {g ∈ P : gO ∈ K}. We always assume that UO is a non-trivial neighborhood of e for every O ∈ K (i.e., K is "locally stable" under the action of P), and that if O1,O2 ∈ K and g ∈ UO1 ∩ UO2 then there is O ∈ K such that O1 ∪ O2 ⊂ O and g ∈ UO (i.e., K is "P-stably directed").4 Concerning spectral properties, we assume that the generators of the spacetime transla- tion subgroup (energy-momentum operators) have positive joint spectrum, and that there is a P-invariant unit vector Ω0 ∈ H0 (vacuum vector) which is cyclic for hA(O), U (g) : O ∈ K, g ∈ Pi. Remark 7.2. Assume first that P preserves K, i.e., gO ∈ K for every g ∈ P, O ∈ K (e.g., if K is the set of all double cones in Minkowski space and P is the universal covering of the Poincar´e group, or if K is the set of all open proper bounded intervals of R and P is the translation- dilation subgroup of the Mobius group), or equivalently UO = P for every O ∈ K. Consider then a local net {A} over K as in Definition 6.1, fulfilling Haag duality and covariant with respect to P as in Definition 7.1. Denoted by αg := AdU (g) the adjoint action on B(H0), we have an action of P on the net {A} (which extends to an action by normal -automorphisms of the quasilocal algebra A), and another action of P on DHR endomorphisms ρ in DHR{A} given by gρ := αgρα−1 g . Observe that gρ is again DHR and localizable in gO if ρ is localizable in O. Moreover, gt := αg(t) ∈ HomDHR{A}(gρ, gσ), g ∈ P, if t ∈ HomDHR{A}(ρ, σ). In other words, we have an action of P on the category DHR{A} (as a strict C ∗ braided tensor category) by autoequivalences (actually automorphisms), which is also strict in the terminology of [Tur10, App. 5]. Indeed, one can easily check that g(ρ × σ) = gρ × gσ, where ρ × σ = ρσ (composition of endomorphisms of A), and g id = id for every g ∈ P and ρ, σ in DHR{A}. Also, g(hρ) = ghρ and eρ = ρ if e is the identity in P. On the other hand, if not every g ∈ P, O ∈ K fulfill gO ∈ K (e.g., if K is the set of all open proper bounded intervals in R and P is the universal covering of the Mobius group) then αg, g ∈ P, are not always automorphisms of the quasilocal algebra A and the previous global statements have to be replaced with local ones by specifying local algebras and spacetime regions. For instance, gρ = αgραg−1, for a fixed g ∈ P, is well defined on every A(O), O ∈ K, such that g−1O ∈ K, and it is an endomorphisms of A(O) if ρ is and endomorphisms of A(g−1O) (e.g., if ρ is DHR localizable in g−1O). Similarly, the intertwining relation for 4These assumptions are the abstraction of the geometric properties which are needed in this section. They are fulfilled, e.g., by all the examples of spacetime symmetries P acting on directed sets of bounded regions K mentioned above. 28 gt between gρ and gσ, if t ∈ HomDHR{A}(ρ, σ), must be intended locally. generality we give the following definition, cf. [Lon97, Sec. 2, App. A], [Tur10, App. 5]. In this level of Definition 7.3. Let {A} be a local net realized on H0 as in Definition 6.1, fulfilling Haag duality and covariant with respect to a group of spacetime symmetries P as in Definition 7.1. Let αg := AdU (g), g ∈ P, and let C ⊂ DHR{A} be a full and replete tensor subcategory, closed under finite direct sums and subobjects. We say that P has an equivariant action on C (and write CP = C) if there is a map where g ∈ P, ρ is an object of C, such that g, ρ 7→ z(g, ρ) (i) z(·, ρ) is a strongly continuous unitary valued map in B(H0), z(g, id) = 1 for every g ∈ P, z(e, ρ) = 1 for every ρ in C, and z(gh, ρ) = αg(z(h, ρ))z(g, ρ) for every g, h ∈ P and ρ in C. ("cocycle identity") (ii) (iii) (iv) (v) z(g, ρ)ρ(a)z(g, ρ)∗ = αgρ αg−1(a), a ∈ A( O) if ρ is DHR localizable in O0 ∈ K, g ∈ UO0 is such that g−1 ∈ UO0, and O ∈ K is such that O0 ∪ gO0 ⊂ O, g−1 ∈ U O. 5 ("local intertwining property") if ρ is DHR localizable in O0 ∈ K, g ∈ UO0, and O ∈ K is such that O0 ∪ gO0 ⊂ O. z(g, ρ) ∈ A(O) αg(t) = z(g, σ)tz(g, ρ)∗ if ρ and σ are DHR localizable respectively in O1 and O2 ∈ K, g ∈ UO1 ∩ UO2, and t ∈ HomDHR{A}(ρ, σ). ("naturality of cocycles") if ρ, σ, g are as in (iv). ("tensoriality of cocycles") z(g, ρσ) = z(g, ρ)ρ(z(g, σ)) (vi) Adz(g,ρ) ρ is DHR localizable in gO0 ∈ K, if ρ, g are as in (iii). ("global localization property") 5The existence of at least one O with these properties is guaranteed because K is P-stably directed by assumption (Definition 7.1). 29 Remark 7.4. In the case that gO ∈ K for every g ∈ P, O ∈ K, we have a (global) action of P on C ⊂ DHR{A} (as a strict C ∗ braided tensor category), see [Tur10, Def. 1.2]. Then the equivariance of the action as in Definition 7.3, cf. [Tur10, Def. 2.1], says that the map z defines a tensor natural transformation (isomorphism) between the trivial action ι of P on C by autoequivalences and the action defined by α. Naturality is automatic because P is considered as a discrete tensor category, i.e., the only morphisms are the identity morphisms, while tensoriality is encoded in the cocycle identity (i). The properties (iv) and (v) above say that z(g,·) is a natural tensor transformation (unitary isomorphism) between tensor functors ιg and αg for every g ∈ P. Lemma 7.5. In the assumptions of Definition 7.3, let z(·,·) be a map fulfilling the properties (i) and (ii), then the following holds as well (ii)′ z(g, ρ)ρ(αg (a))z(g, ρ)∗ = αg(ρ(a)), a ∈ A(O), g ∈ UO if ρ is DHR localizable in O ∈ K. The same is true if a ∈ A( O) and g ∈ U O for any O ∈ K. In other words, the unitaries Uρ(g) := z(g, ρ)∗U (g), g ∈ P, implement the covariance of ρ with respect to P (cf. [CKL08, Sec. 4.2]) and give a strongly continuous unitary representation of P on H0. Proof. Let a ∈ A(O), g ∈ UO and assume that O ∈ K is a localization region of ρ. Also, let V ⊂ UO be a symmetric neighborhood of e, e.g., V := UO ∩ U −1 O . Consider the set of all elements V(e) ⊂ UO that can be joined to e by a V-chain in UO, namely those g ∈ UO such that there are x1, . . . , xn ∈ UO, n ≥ 1, with x1 = e, xn = g, and xj+1x−1 j ∈ V for every j = 1, . . . , n − 1, cf. [BP01, Def. 19, 144]. By a standard argument, V(e) is open and closed in UO, hence V(e) = UO by connectedness. Then every g ∈ UO can be written as g = g1g2 ··· gn where gj ∈ V, and in addition gj ··· gn ∈ UO for every j = 1, . . . , n, n ≥ 1. Just set gj := xj+1x−1 j , j = 1, . . . , n − 1, and gn := e. Now, by the cocycle identity (i) we have z(g1g2 ··· gn, θ) = αg1···gn−1(z(gn, θ))··· αg1(z(g2, θ))z(g1, θ) (18) Thus, z(g1, θ)ρ(αg(a))z(g1, θ)∗ = αg1(ρ(αg2···gn(a))) because g1, g1 and we want to compute its adjoint action on ρ(αg(a)), a ∈ A(O). −1 ∈ V ⊂ UO, hence equality (ii) holds on every A( O), O ∈ K, such that O ∪ g1O ⊂ O and g−1 1 ∈ U O. Moreover, 1 g = g2 ··· gn ∈ UO, hence we can assume that gO ⊂ O, g−1 1 ∈ UgO = UOg−1 because g−1 by enlarging O if necessary, because K is P-stably directed by assumption (Definition 7.1). Continuing, z(g2, θ)ρ(αg2···gn(a))z(g2, θ)∗ = αg2(ρ(αg3···gn(a))) because g2, g2 −1 ∈ V ⊂ UO, hence we can choose O ∈ K as in (ii) such that g−1 2 ∈ U O and again further assume that g2 ··· gnO ⊂ O. Indeed, g−1 2 because g3 ··· gn ∈ UO. By finite iteration we get the first claim. By the cocycle identity (i), the unitaries Uρ(g) := z(g, ρ)∗U (g), g ∈ P, form a repre- sentation of P. Indeed, Uρ(g)Uρ(h) = z(g, ρ)∗αg(z(h, ρ))∗U (gh) = Uρ(gh) and z(g−1, ρ) = αg−1(z(g, ρ)∗), hence also Uρ(g−1) = Uρ(g)∗, follow from z(e, ρ) = 1. We want to show that it implements the covariance of ρ. 2 ∈ Ug2···gnO = UOg−1 n ··· g−1 Let a ∈ A( O), g ∈ U O for an arbitrary O ∈ K. Define V := U O ∩ U −1 and consider W := V ∩ V, or any other symmetric neighborhood W of e such that W ⊂ UO ∩ U O. By the O 30 same argument as above, we have g ∈ U ( O) = W(e), i.e., we can write g = g1 ··· gn, where gj ∈ W and gj ··· gn ∈ U O for every j = 1, . . . , n, n ≥ 1, and Uρ(g)ρ(a)Uρ(g)∗ = Uρ(g1)··· Uρ(gn)ρ(a)Uρ(gn)∗ ··· Uρ(g1)∗. Now, gn ∈ W ⊂ UO ∩U O hence there is O1 ∈ K such that O ∪ O ⊂ O1, gn ∈ UO1, moreover ρ is localized in O1, a ∈ A(O1), thus by the first claim we get Uρ(gn)ρ(a)Uρ(gn)∗ = ρ(αgn(a)). Continuing, gn−1 ∈ UO ∩Ugn O and we can repeat the previous argument on O2 ∈ K such that O ∪ gn O ⊂ O2, gn−1 ∈ UO2, to get Uρ(gn−1)ρ(αgn(a))Uρ(g∗ n−1) = ρ(αgn−1gn(a)). By finite iteration we get Uρ(g)ρ(a)Uρ(g)∗ = ρ(αg(a)) where a ∈ A( O), g ∈ U O for an arbitrary O ∈ K, completing the proof. With similar arguments one can extend naturality and tensoriality of cocycles to (almost all) g ∈ P, namely Lemma 7.6. In the assumptions of Definition 7.3, let z(·,·) be a map fulfilling the properties (i) and (iv), then the following holds as well (iv)′ αg(t) = z(g, σ)tz(g, ρ)∗, g ∈ P. If z(·,·) fulfills (i), (ii), (iii) and (v), then it fulfills also (v)′ z(g, ρσ) = z(g, ρ)ρ(z(g, σ)), where O2 ∈ K is a DHR localization region of σ. g ∈ UO2, Proof. Let O1,O2 ∈ K be respectively localization regions of ρ, σ. To prove the first state- ment, write g ∈ P as g = g1 ··· gn, n ≥ 1, where gj ∈ UO1 ∩ UO2, j = 1, . . . , n. Then make use of equation (18) and apply (iv) at each step. To prove the second statement, write g ∈ UO2 as before, and assume in addition that gj ··· gn ∈ UO2, j = 1, . . . , n, cf. proof of Lemma 7.5. Then make again use of equation (18) for z(g, ρσ) and apply (v) for each gj ∈ UO1 ∩ UO2. Repeated use of Lemma 7.5 gives the desired conclusion. Notice that z(g, σ) belongs to the quasilocal algebra A because of assumption (iii), hence one can safely apply the endomorphisms ρ. Now we show that the properties (mainly tensoriality and naturality) of covariance co- cycles expressed by the equivariance of the action of spacetime symmetries on the DHR category ensure covariance of the extended nets constructed as in Theorem 6.8. Theorem 7.7. Let {A} be a local net fulfilling Haag duality, standardly realized on H0, and covariant with respect to a group of spacetime symmetries P (Definition 7.1). Assume in addition that either P acts transitively on K (i.e., for every O1,O2 ∈ K there is g ∈ UO1 such that gO1 = O2), or P preserves K (i.e., gO ∈ K for every g ∈ P, O ∈ K), 6. Then an extension {B} of {A} constructed as in Theorem 6.8 from a unital generalized net Q-system of intertwiners (θ, w,{mi}) is automatically covariant, provided that P has an equivariant action on a tensor subcategory C ⊂ DHR{A} (i.e., CP = C) which contains θ (Definition 7.3). 6This assumption is needed to obtain covariance of {B} on all the regions in K, cf. footnote after equation + preserving double cones in R3+1. (16). Examples are gMob acting transitively on bounded intervals in R, or P ↑ 31 Proof. Let (θ, w,{mi}), i ∈ I be a generalized net Q-system of intertwiners in DHR{A} and construct the extension {A ⊂ B} as in Theorem 6.8. Here O ∈ K is a fixed localization region for θ and N = A(O). In the following we denote by H := Hϕ the Hilbert space of {B}, we identify H0 = eNH and a ∈ A, Mi ∈ B(O) with their images under πϕ in B(H). Thus i H0 where every ψ ∈ H can be written as ψ = Pi M ∗ i ψi, with ψi ∈ qiH0. Moreover, we have Mia = θ(a)Mi for every a ∈ A, i ∈ I. Having full control of the Hilbert space thanks to the Pimsner-Popa condition, we can set H =Xi M ∗ U (g)ψ :=Xi M ∗ i z(g, θ)∗U (g)ψi for every g ∈ P, ψ ∈ H, where U implements the covariance of {A} on H0 and z(·, θ) is the covariance cocycle of θ given by equivariance. By definition of αg and by the cocycle identity we have z(g, θ)∗U (g)z(h, θ)∗U (h) = z(g, θ)∗αg(z(h, θ))∗U (g)U (h) = z(gh, θ)∗U (gh), hence U is a representation of P on H, which is strongly continuous and unitary as one can easily check. In order to show that U implements covariance of {B} with respect to P, take first a ∈ A(O), where O is as above, take g ∈ UO, see Definition 7.1, ψ ∈ H, and compute M ∗ U (g)a U (g)∗ψ =Xi i z(g, θ)∗αg(θ(a))z(g, θ)ψi =Xi M ∗ =Xi i z(g, θ)∗U (g)θ(a)U (g)∗z(g, θ)ψi M ∗ i θ(αg(a))ψi = αg(a)ψ, where the third equality follows from Lemma 7.5. Take now Mi ∈ B(O), i ∈ I, then U (g)Mi U (g)∗ψ =Xj,k M ∗ k z(g, θ)∗U (g)lki j U (g)∗z(g, θ)ψj where the coefficients lki cocycles, see the property (iv) in Definition 7.3, and because lki have αg(lki the property (v) in Definition 7.3, we have that z(g, θ2) = z(g, θ)θ(z(g, θ)), hence j ∈ N = A(O) are those given in equation (15). By naturality of j ∈ HomDHR{A}(θ, θ2), we j z(g, θ)∗. Moreover, by tensoriality of cocycles, see j U (g)∗ = z(g, θ2)lki j ) ≡ U (g)lki =Xj,k M ∗ k θ(z(g, θ))lki j ψj = z(g, θ)Miψ. The global localization property (vi) in Definition 7.3 implies that z(g, θ) is a unitary charge transporter for θ from O to gO ∈ K, hence z(g, θ)Mi ∈ B(gO) by definition (16) of the local algebras, and we conclude U (g)B(O) U (g)∗ = B(gO). Now, covariance for arbitrary regions O ∈ K and g ∈ U O follows either by transitivity of P on K (trivially), or because P preserves K, in which case U O = P and we can meaningfully write AdUθ(g) θ(u) = θ(αg(u)), u ∈ HomDHR{A}(θ, θ) and θ(z(g, θ)) in A, cf. Lemma 7.5, 7.6. 32 0 wΩ0, indeed U (g)Ω = M ∗ Positivity of the energy-momentum spectrum holds because Uθ has positive spectrum, indeed θ is a (possibly infinite) direct sum of covariant endomorphisms fulfilling the spectrum condition, see [DHR74, Thm. 5.2]. The P-invariance of the vacuum vector Ω := Ωϕ follows from Ω = M ∗ 0 z(g, θ)∗αg(w)Ω0 = Ω by naturality (iv) of the action of g on w ∈ HomDHR{A}(id, θ) and because z(g, id) = 1. Thus the extended net {B} is covariant as in Definition 7.1. Remark 7.8. If the quasilocal algebra A together with the elements of the Pimsner-Popa basis Mi, i ∈ I, form a -algebra of charged intertwiners in the sense of Remark 6.9, one can try to define covariance of the extension B (at the -algebra level) by postulating αg := αg on A and αg(Mi) := z(g, θ)Mi. In this case as well, naturality and tensoriality of the cocycle z guarantee that αg is -multiplicative. 0 z(g, θ)∗U (g)wΩ0 = M ∗ Next, we show how equivariance holds, in the sense of Definition 7.3, for the action of some typical spacetime symmetry groups on the DHR category in different dimensions. More precisely, we consider here the subcategory C = DHRd{A} of DHR{A} (Definition 6.2) which is relevant for finite index or infinite index discrete extensions treated in Theorem 6.8. Example 7.9. (Mobius covariant nets in 1D). Let P = gMob the universal covering of the Mobius group and K = {open proper bounded intervals I ⊂ R}. Consider a local P-covariant net {A} over K as in Definition 6.1, 7.1, fulfilling Haag duality on R, namely A(I ′)′ = A(I), I ∈ K, I ′ = R r ¯I. By locality and P-covariance we have U (Rot2π) = 1 [GL96, Thm. 1.1], hence {A} is automatically Mob-covariant and we can extend it to a net { A} over the open proper intervals of S1, see [CKL08, Prop. 16, Cor. 17]. The extension coincides with the one given by A(I) := A(S1 r ¯I)′ if I ⊂ S1 contains the point at infinity in its closure and A(I) := A(I) otherwise, see [KLM01, Lem. 49]. Moreover, every endomorphism ρ in DHR{A} extends to a representation {πI , I ⊂ S1} of { A} on H0 such that πI = ρ↾A(I) if I is identified to a bounded interval of R via the Cayley map, see [KLM01, Prop. 50]. The Bisognano-Wichmann property [GL96, Prop. 1.1] and strong additivity [GLW98, Lem. 1.3] ensure that finite-dimensional DHR endomorphisms are covariant (with positive energy) with respect to P, see [GL92, Thm. 5.2]. For every ρ in DHRf{A}, following [GL92, Prop. 8.2], we can define by equation (18) the cocycle z(·, ρ). The definition is well posed (in B(H0)) by [GL92, Eq. (8.5)] because any two chains in P are homotopic by simple connectedness of P, see [BP01, Def. 45, Lem. 46] for more details. Thus the properties (i), (ii), (iii) of Definition 7.3 hold, see also (ii)′ of Lemma 7.5. The property (vi) holds by additivity [FJ96, Sec. 3] while (iv) and (v) can be derived from the results of [Lon97]. Indeed, let ρ and σ in DHR{A} and choose a common localization interval I ∈ K. Let I1 ∈ K be such that ¯I ⊂ I1 and I2, I3 ∈ K such that {Ii, i = 1, 2, 3} is a partition of S1 obtained by removing three distinct points and counterclockwise ordered. Let V ⊂ UI ∩U −1 I be an arbitrarily small symmetric neighborhood of e in P, whose elements g can be written as products of dilations ΛIi associated to Ii, i = 1, 2, 3 such that in addition Λj Iij ··· Λn map I inside I1 for Iij every j = 1, . . . , n and g = Λ1 . Thus at each step we can consider I as a subinterval of either I1, or S1 r ¯I2, or S1 r ¯I3. Observe that the dilations with respect to any such partition of S1 into three intervals generate P, see [GLW98, Lem. 1.1]. Now, with this choice of V, cf. [Lon97, Lem. 2.2], for every g ∈ V we have Ii1 ··· Λn Iin and Λj Iin αg(t) = z(g, σ)tz(g, ρ)∗ 33 if t ∈ HomDHR{A}(ρ, σ), and z(g, ρσ) = z(g, ρ)ρ(z(g, σ)). Building suitable V-chains in P and reasoning as in the proof of Lemma 7.6, one can show the properties (iv) and (v)7 in their global formulation (iv)′ and (v)′ of Lemma 7.6. Now, if ρ is in DHRd{A} let {wi} be a (possibly infinite) Cuntz family of isometries i ∈ ρ(A)′ ∩ A(I) are mutually orthogonal ρ =: ρi are irreducible DHR endomorphisms of finite- in A(I), for I ∈ K big enough, such that wiw∗ projections, Pi wiw∗ dimension. For every g ∈ P i = 1 and Adw∗ i z(g, ρ) :=Xi αg(wi)z(g, ρi)w∗ i converges in the strong operator topology and extends the definition given in DHRf{A} by [Lon97, Prop. 1.3, Eq. (1.13)]. Let C := DHRd{A}, the unitaries z(g, ρ), g ∈ P, ρ in C form again a cocycle map, as one can check directly on each direct summand of ρ = ⊕iρi, hence the action of P on C is equivariant in the sense of Definition 7.3. Example 7.10. (Poincar´e covariant nets in 3+1D). Let P = P↑ + the universal covering of the Poincar´e group and K = {double cones O ⊂ R3+1}. Consider a local P-covariant net {A} over K as in Definition 6.1, 7.1, fulfilling Haag duality on R3+1. Assume that {A} fulfills in addition the Bisognano-Wichmann property on wedges, see [BW75], and that local intertwiners between finite-dimensional DHR endomorphisms are global intertwiners, cf. [Rob74, Thm. 4.3], [DMV04, Cor. 6.2]. Due to the fact that Lorentz boosts with respect to different wedges generate P = P↑ +, we can make again use of the results of [GL92], [Lon97], in a different geometrical situation, to draw analogous conclusions. Namely, the action of P of C := DHRd{A}, which in this case is globally defined, see Remark 7.2, is again equivariant in the sense of Definition 7.3. 8 Braided product of nets In this section we apply the braided product construction to nets of von Neumann algebras and show that it enjoys some remarkable properties, in analogy to the finite index case, which allows one to extract boundary quantum field theories as in [BKLR16]. Such field theories with transparent boundaries will be discussed in the next section. EL ER Denote by {A ⊂ BR} two discrete relatively local extensions of the same local net {A} (Definition 6.7) constructed from unital generalized net Q-systems of intertwiners (θL, wL,{mL j }) in DHR{A} as in Theorem 6.8. By Proposition 4.2 and again Theorem 6.8 we know that there is a braided product extension {A ⊂ BL ×± ε BR} such that ⊂ BL}, {A i }), (θR, wR,{mR A ⊂ ⊂ ⊂ ⊂ BL BR BL ×± ε BR 7Tensoriality also follows by observing that P = gMob is perfect, see, e.g., [Lon97, App. A], hence the unitary representation Uρσ which implements covariance of ρσ in DHRf {A} is unique. 34 where ε is the DHR braiding. Of course we have the analogous of Proposition 4.5, which we rewrite below to establish notation. We prefer to think of the net {BL ×± ε BR} as the one constructed in Section 6.1. Let ε BR) O} be the inductive family of nets indexed by O ∈ K (see Section 6.1), and let {(BL ×± Γ O be a unitary that implements (θLθR) O := θL O on A( O), namely θR O OθR θL O(x) = AduL O θL(uR O ) θLθR(x) = AdΓ O (x), x ∈ A( O). Proposition 8.1. The maps L, O : BL( O) → (BL ×± ε BR) O( O) , L, O := AdΓ∗ O ◦ Ad(ε± )∗ ◦ θR O ◦ γL O ,θR O θL O O ◦ γR lift to embeddings 8 of {BL} and {BR} into the braided product net {BL ×± R, O : BR( O) → (BL ×± ε BR) O( O) , R, O := AdΓ∗ O ◦ θL O ε BR} L : BL → BL ×± R : BR → BL ×± ε BR ε BR. Proof. It is enough to show that L/R, O(BL/R(O)) ⊂ (BL×± L/R, O1 = (L/R, O2)↾BL/R( O1) for O1 ⊂ O2, but these follow from elementary calculations. Proposition 8.2. Let ELR denote the distinguished conditional expectation from {BL×± to {A} obtained as in Proposition 6.13, then ε BR) O(O) for O ⊂ O and ι O1, O2◦ ε BR} ELR(L(bL)R(bR)) = EL(bL)ER(bR), bL ∈ BL, bR ∈ BR. Proof. Using θL(wR∗ θL,θR = wR∗ )ε±∗ θR(wR∗ θLθR(wL∗ and wR∗ θL,θR = θR(wR∗ ε± ) we get )ε±∗ θL,θRγRγL(bL)ε± γRγL(bL)θL(wR∗ = θLθR(wL∗ θL,θRγLγR(bR)θL(wR)wL) γR(bR)wR)wR) = θLθR(EL(bLER(bR))) = θLθR(EL(bL)ER(bR)) from which the proposition follows. For the rest of the section, we assume that {A ⊂ BL}, {A ⊂ BR} are as in Proposition 6.10, so that the generalized Q-systems (θL, wL,{mL j }) are induced by Pimsner- σj} ⊂ BR(O), where ρi ≺ θL, σj ≺ θR Popa bases of global charged fields {ψL are respectively irreducible DHR subendomorphisms with finite dimension, localized in O ∈ K. Proposition 8.3. i }), (θR, wR,{mR ρi} ⊂ BL(O),{ψR {L({ψL ρi})R({ψR σj})} is a Pimsner-Popa basis for ι(A(O)) expansion (Proposition 2.3). ELR ⊂ (BL×± ε BR)(O) and it gives a unique Pimsner-Popa 8The same is true in the representation employed in the first proof of Theorem 6.8, but it is more lengthy to check. 35 Proof. The first statement is immediate. To prove the second statement, it is enough to show that ELR(L(ψL which follows directly from a calculation analogous to the proof of Proposition 8.2. σj )R(ψR∗ σ′ h )L(ψL∗ ρ′ k ρi )R(ψR )) = δρi,ρ′ k δσj ,σ′ h In the following, with abuse of notation, we shall often suppress the above embeddings ε BR as well. L/R, and ι : A → BL ×± Remark 8.4. In [BKLR16] it is shown that the extension associated to the braided product of two "ordinary" Q-systems is characterized algebraically in the following way. Let N ⊂ MA and N ⊂ MB be two finite index inclusions and let (θA, wA, xA), (θB, wB, xB) be the associated Q-systems. Denote by ιA/B the respective inclusion maps and by θA = ⊕iρi, θB = ⊕jσj the irreducible decompositions of the dual canonical endomorphisms. Then it is known [BKLR15, Thm. 3.11] that MA (resp. MB) is finitely generated by N and {ψA ρi} (resp. {ψB ε MB can be completely characterized as the -algebra freely generated by MA and MB, modulo the relations In this case, the braided product MA ×± σj} are charged fields. σj}), where {ψA ρi}, {ψB ιA(n) = ιB(n), n ∈ N , ψB σj ψA ρiψB σj . ρi = ε± ρi,σj ψA In the discrete (infinite index) case this is no longer true since the extensions are not finitely generated by N and the charged fields. We have to settle for a weaker form of this result, valid for pairs of irreducible extensions, that will nevertheless prove to be useful in Section 10. Let BL ⊂ BL be the -algebra generated by ιL(A) and the charged fields {ψL ε BR and {R(ψR ε BR be the -algebra generated by L(ιL(N )) = R(ιR(N )), {L(ψL ilarly, let BR ⊂ BR be the -algebra generated by ιR(A) and the charged fields {ψR Let BL×R ⊂ BL ×± BL ×± Lemma 8.5. Suppose in addition that {A ⊂ BL} and {A ⊂ BR} are irreducible exten- sions. Then BL×R is isomorphic to the -algebra freely generated by BL and BR, modulo the relations ε BR. Then we have the following σj )} ⊂ BL ×± ρi} ⊂ BL. Sim- σj} ⊂ BR. ρi)} ⊂ ιL(a) = ιR(a), ρi = ε± ψR σj ψL ρi,σj ψL a ∈ A, ρiψR σj . relations can be written as a finite sum x = P nρi,σj ψL Proof. An arbitrary element x of the free -algebra generated by BL and BR modulo these σj in a unique way. The same is true for any element in BL×R by Proposition 8.3 and Remark 6.9, thus the expansion yields an isomorphism. ρiψR In [BKLR16] it was shown that the center of the braided product extension is an object of great interest since it contains all the information on transparent boundary conditions between the two starting quantum field theories. We here show that in the discrete case some relevant structural features are retained, in particular that the center of the braided product extension agrees with the relative commutant, which will be useful in the next section for the construction of irreducible phase boundaries from the central decomposition of the braided product. The expansion in terms of the Pimsner-Popa basis of charged fields (Proposition 8.3) can be used to characterize the relative commutant. 36 Lemma 8.6. For every x ∈ (BL ×± ε BR)(O) we have x ∈ (BL ×± ε BR)(O) ∩ A(O)′ ⇔ x = Xρi,σj σj x) ∈ HomEnd0(A(O))(id, ρiσj). with rρi,σj := ELR(ψL Proof. It is enough to use the uniqueness of the expansion in Proposition 8.3 ρiψR ψR σj ∗ ∗ ψL ρi rρi,σj ∗ ∗ ψL ρi ψR σj ρiψR ELR(ψL nx =Xi,j =Xi,j σj nx) =Xi,j σj xn) =Xi,j for every n ∈ A(O), thus rρi,σj n = ρi(σj(n)) rρi,σj . ELR(ψL ρiψR ψR σj ψL ρi ∗ ∗ ψR ρi ∗ ψL σj ∗ ρiσj(n)ELR(ψL ρiψR σj x) ψR σj ∗ ∗ ψL ρi ELR(ψL ρiψR σj x)n = xn As in the finite index case, cf. [BKLR16, Prop. 4.19], the center of the braided product of two local extensions coincides with the relative commutant of {A} in the braided product. Proposition 8.7. Suppose in addition that {A ⊂ BL} and {A ⊂ BR} are local extensions, then (BL ×± ε BR)(O) ∩ A(O)′ = (BL ×± ε BR)(O)′. Proof. Let us first verify that the von Neumann algebra Z generated by r∗ rρi,σj ∈ HomEnd0(A(O))(id, ρiσj), is contained in the center. For every n ∈ A(O) we have ε BR)(O) ∩ (BL ×± ρi,σj ψL ρiψR σj , with r∗ ρi,σj ψL ρiψR σj nψL ρ′ k ψR σ′ t = nr∗ ρi,σj ρi(ε± ρ′ k,σj )ε± ρ′ k,ρi kρi(ε± ρ′ σ′ t,σj )ρ′ k(ε± σ′ t,ρi )ψL ρ′ k ψR σ′ t ψL ρiψR σj by direct computation and using locality of {BL} and {BR}, i.e. ψL ρ′ k ρi = ε± ψL ψL ρiψL ρ′ k , ρi,ρ′ k ψR σ′ t σj = ε± ψR σj ,σ′ t ψR σj ψR σ′ t , cf. Theorem 6.8. Now, it is easy to see that ρi,σj ρi(ε± r∗ ρ′ k,σj )ε± ρ′ k,ρi kρi(ε± ρ′ σ′ t,σj )ρ′ k(ε± σ′ t,ρi ) = ρ′ kσ′ t(r∗ ρi,σj ) from which r∗ ρi,σj ψL ρiψR σj is contained in the center of the braided product. For brevity, in the following we denote B ∩ A′ = (BL ×± ε BR)(O)′, and consider the inclusions ε BR)(O) ∩ (BL ×± (BL ×± ε BR)(O) ∩ A(O)′ and B ∩ B′ = Z ⊂ B ∩ B′ ⊂ B ∩ A′. If we take the GNS representation of B ∩ A′ with respect to the vacuum Ω, we get a cyclic and separating vector for B∩A′ and for B∩B′ as well, by Lemma 8.6. Now we check that the canonical conjugations of B ∩ A′ and B ∩ B′ with respect to Ω agree. This holds because the Tomita operator S of (B ∩ A′, Ω), i.e., the closure of the operator S0 : xΩ → x∗Ω, x ∈ B ∩ A′, is an extension of the Tomita operator of (B ∩ B′, Ω). Since the latter is continuous and defined on all the GNS Hilbert space (because B ∩ B′ is abelian), the two operators agree and coincide with the respective canonical conjugations. Thus J(B ∩ A′)J = (B ∩ A′)′ ⊂ (B ∩ B′)′ = J(B ∩ B′)J ⊂ J(B ∩ A′)J from which the result follows. 37 Lastly, as an application of Theorem 7.7, we show covariance of the braided product net. Proposition 8.8. Let {A} be a local net, covariant with respect to P as in the assumptions of Theorem 7.7. Let {BL} and {BR} be two extensions of {A} constructed as in Theorem 6.8 from unital generalized net Q-systems of intertwiners (θL, wL,{mL i }) and (θR, wR,{mR j }). Assume that P acts equivariantly on two tensor subcategories CL and CR of DHR{A} which contain respectively θL and θR. Then the braided product net {BL ×± ε BR} is also covariant with respect to P. Moreover, the embeddings L and R given in Proposition 8.1 are covariant as representations, namely L ◦ αL g ◦ R(bR), where g (bL) = αLR bL/R ∈ BL/R(O), for every g ∈ UO and O ∈ K. Proof. P acts equivariantly on θLθR and on the (full, replete) tensor subcategory D generated in DHR{A} by θL and θR. Indeed, the cocycle given by z(g, θLθR) = z(g, θL)θL(z(g, θR)), g ∈ P, is manifestly natural and tensor in D, hence we can apply Theorem 7.7. i )) and αLR j )) = R(αR i )) = L(αL j )) by direct computation using naturality of cocycles. g ◦ L(bL) and R ◦ αR The second statement follows from αLR g (M R g g (bR) = αLR (L(M L g (M L (R(M R g 9 Applications to phase boundaries in QFT The main application in QFT for the braided product of ordinary Q-systems in [BKLR16] is the construction and classification of phase boundary QFTs. A boundary is simply a time-like hypersurface of codimension 1 in Minkowski spacetime Rn+1, n ≥ 1, or a point in R. Perhaps the simplest type of boundary QFT is a system in a one-sided box. Namely, on one side of the boundary (the side of the box) there is a physical system described by bulk fields, while on the other side there is no physical content. This situation is usually referred to as a hard boundary, or reflective boundary. In the following we will be concerned with phase boundaries, also called transmissive boundaries, which describe QFTs sharing some distinguished chiral fields across the boundary (for example the stress-energy tensor) but the field content may in general be different on the two opposite sides. If the common fields which are not affected by the presence of the boundary include the stress-energy tensor, then the bulk fields may be defined by covariance on all Minkowski spacetime. Of course they do not represent physically meaningful quantities when they are transported to the opposite side of the boundary. In any case, this observation is crucial for the meaningfulness of the following definition. Let {A} be a local net and let {A ⊂ BL}, {A ⊂ BR} be two local extensions (see Definition 6.7). Let ιL and ιR be the corresponding embeddings. ML and MR denote the two portions of Minkowski spacetime determined by the boundary. Definition 9.1. A phase boundary condition (for short phase boundary) between two local extensions {A ⊂ BL}, {A ⊂ BR} is a pair of locally normal representations πL and πR of the nets {BL} and {BR}, respectively, on a common Hilbert space H, with the following properties. They agree when restricted to the common subnet A, namely πL ◦ ιL = πR ◦ ιR and, for O1 ⊂ ML, O2 ⊂ MR, and O1,O2 in relative space-like position, πL(BL(O1)) and πR(BR(O2)) commute, i.e., they respect locality across the boundary. 38 A phase boundary is called irreducible if the inclusions πL/R ◦ ιL/R(A(O)) ⊂ πL(BL(O)) ∨ πL(BR(O)) are irreducible for every O ∈ K. In the present setting, we show that the braided product can be decomposed over its center (in general as a direct integral) and its components give rise to irreducible phase boundaries, in analogy to the finite index case. Remark 9.2. A prominent feature of the finite index case is that the phase boundaries found within the braided product net by central decomposition do exhaust the set of all possible irreducible phase boundaries modulo unitary equivalence. The proof of the latter heavily relies on the finiteness of the index since this insures that the braided product construction can be completely determined algebraically as the free ∗-algebra generated by the starting nets {BL} and {BR} modulo relations as in Remark 8.4. This makes the braided product a universal object in the sense that every irreducible phase boundary condition arises a representation of the former [BKLR16, Prop. 5.1]. In the infinite index setting this is no longer the case as we will see in Section 10. For ease of exposition, we state the results for chiral CFTs (and thus phase boundaries in 1D) although the analysis can be extended to greater generality without difficulty. More- over, in order to avoid inconvenient technicalities with disintegration theory, we assume that the starting local extensions have the split property, [DL84]. This assumption is not too restrictive since most interesting models in QFT have this property, in particular all chiral diffeomorphism covariant models [MTW16]. Let {A} be a local conformal net (Mobius covariant, see Definition 7.1) on R over a separable Hilbert space and satisfying Haag duality on R. Exactly as in the notation of [KLM01, Prop. 55], for I, I ∈ K (here K is the set of open proper bounded intervals of R), I ⊂⊂ I means that ¯I ⊂ I. If {A} has the split property, then, for each pair of intervals I ⊂⊂ I, there is an intermediate type I factor A(I) ⊂ N (I, I) ⊂ A( I) and we denote by K(I, I) the compact operators of N (I, I). IQ is the set of intervals with rational endpoints and A is the separable C ∗-subalgebra of A generated by all K(I, I) with I ⊂⊂ I, I, I ∈ IQ. Proposition 9.3. [KLM01]. Let π be a locally normal representation of A. Then π↾A is a representation of A and π↾K(I, I) is non-degenerate for every pair of intervals I ⊂⊂ I. Conversely, if σ is a representation of A such that σ↾K(I, I) is non-degenerate for all intervals I, I ∈ IQ, I ⊂⊂ I, there exists a unique locally normal representation σ of A that extends σ. Moreover, equivalent representations of A correspond to equivalent representations of A. Now, let {BL} and {BR} be local conformal nets extending {A} as in Definition 6.7. As- sume that {BL/R} have the split property and that {A ⊂ BL/R} are discrete irreducible ex- tensions with corresponding unital generalized net Q-systems of intertwiners (θL, wL,{mL i }), (θR, wR,{mR Define KL(I, I), KR(I, I), and the separable C ∗-algebras BL, BR as above. Using the last proposition, we want to show that the embedding homomorphisms L and R into the braided product (Proposition 8.1) can be decomposed as representations with respect to the j }) given by global charged fields as in Proposition 6.10. 39 center of the braided product. Note that by Proposition 8.7 and 8.8 the centers of the local algebras of the braided product agree, namely we have Z((BL ×± ε BR)(I)) = Z((BL ×± ε BR)(J)) = Z(BL ×± ε BR) for every I, J ∈ K. Proposition 9.4. Let L R ↾BL ∼=Z ⊕ ↾BR ∼=Z ⊕ X X L λ dµ(λ) R λ dµ(λ) be the disintegration of the restrictions of the embeddings L, R to the separable C ∗- ε BR) ∼= λ lift to locally normal represen- subalgebras BL and BR with respect to the center of the braided product Z(BL ×± L∞(X, dµ). Then, for dµ-almost every λ ∈ X, the L λ and R tations of the quasilocal C ∗-algebras BL and BR respectively. Proof. To prove the assertion, by the above proposition, it is enough to show that there is a dµ-null set E such that L λ ↾KR(I, I)) is non-degenerate for every λ /∈ E and I, I ∈ IQ, I ⊂⊂ I. This is easily checked, because for fixed I, I, L ) is non-degenerate by Proposition 9.3 and consequently L λ ↾KL(I, I)) is also non- degenerate for dµ-almost every λ ∈ X. Since I, I ∈ IQ, I ⊂⊂ I are countable, the statement follows. λ ↾KL(I, I) (resp. R λ ↾KL(I, I) (resp. L (resp. R ↾KR(I, I) ↾KL(I, I) Proposition 9.5. Let λ ∈ X r E as above. Then (1) R λ ◦ ιR = L λ (mL λ (mR λ ◦ ιL. i ) = L j )L λ (ε± (2) R θL,θR)L λ (mL i )R λ (mR j ). (3) If U (g) = RX Uλ(g)dµ is the disintegration of the representation of the universal cov- ering of the Mobius group (given by Proposition 8.8) with respect to the center of the braided product, then Uλ(g)L Uλ(g)R λ (BL(J)) Uλ(g)∗ = L λ (BR(J)) Uλ(g)∗ = R λ (BL(gJ)) λ (BR(gJ)). (4) Let Ω =RX Ωλdµ, then Uλ(g)Ωλ = Ωλ and Ωλ is cyclic for λ (BR(J)). λ (BL(J)) ∨ R L _J λ (BR(I)) is a factor. λ (ιL(A(I))) ⊂ L (5) L λ (BL(I)) ∨ R (6) The inclusion L λ (BL(I)) ∨ R λ (BR(I)) is irreducible. Proof. Most of these assertions are trivial and follow from standard techniques in disintegra- tion theory. Covariance, i.e., point (3), follows by Proposition 8.8, Example 7.9 and by the fact that U (g) ∈ Z(BL ×± ε BR)′ by the expansion in Lemma 8.6. 40 Remark 9.6. Proposition 9.5 shows that the braided product construction of two net exten- sions {A ⊂ BL}, {A ⊂ BR} with the required properties induces, via central decomposition, a family of irreducible phase boundaries (L λ ) indexed by the spectrum X (up to a measure zero set) of the center of BL ×± Of course, depending on whether {BL} and {BR} are interpreted to be theories respec- tively on the left and on the right of the boundary, or vice versa, one has to take the braided product with the correct sign, namely with ε+ or ε−. ε BR and living on the Hilbert space Hλ, λ ∈ X. λ , R 10 An example with the U (1)-current In this section we work out concretely the braided product between local extensions of the U (1)-current net. We will see examples where the center of the braided product net is a continuous algebra and therefore the direct integral representation as in Proposition 9.4 does not reduce to a direct sum. This shows in particular that the braided product is not a universal object in the sense of [BKLR16, Prop. 5.1]. This behaviour is expected, since, as in the finite index case, phase boundary conditions for orbifold theories should be determined by their gauge group, see [BKLR16, Sec. 6.2]. We will show the manifestation of this fact in at least one example. For the definition of the U (1)-current {AU (1)} we refer to [BMT88], [GLW98], [Lon08], and to [DV17, Ch. 12] for more detailed calculations. Let I be a proper interval of S1 r {1} and let f ∈ C ∞(S1, R) with support contained in I. Define the net representation {ρf,J}J first on Weyl operators W (g) in the following way ρf,J (W (g)) := ei R f (θ)g(θ) dθ 2π W (g) for g ∈ C ∞(S1, R) with support in a proper interval J of S1 r{1}. These above defined maps are locally unitarily implemented: let I0 be a proper interval of S1 r {1} disjoint from I and J, and let f0 ∈ C ∞(S1, R) with support in I0 and such that RS1 f = RS1 f0. Define LI→I0 as a primitive of f0 − f , namely L′ I→I0 = f0 − f . It is an easy calculation to show that W (LI→I0)W (g)W (LI→I0)∗ = ρf,J (W (g)). Thus the maps {ρf,J}J can be extended in a unique way to the local von Neumann algebras and they determine a locally normal representation of {AU (1)}, which is clearly DHR. More- over these representations are classified up to unitary equivalence by the value RS1 f which is usually referred to as the charge, thus yielding a continuous family of irreducible DHR sectors. We now compute explicitly the braiding operator for the irreducible DHR representations described above. Let ρf be localized in the interval I. If I is an interval disjoint from I and I < I, take f ∈ C ∞(S1, R) with support in I and with same charge as f , i.e., RS1 f =RS1 f . If we denote by u I := W (LI→ I) ∈ HomDHR{A}(ρf , ρ f ) the charge transporter between ρf and ρ f , by definition the braiding operator ε+ ρf ,ρf is obtained by Performing the computation we get ρf ,ρf = u∗ ε+ ρf (u I ). I (u I ) = e−iπQ2 ρf ,ρf = ei R f LI→ I u∗ ε+ I where Q is the charge of the DHR sector ρf . In particular ε+ with N ∈ N. ρf ,ρf = 1 if and only if Q = √2πN 41 10.1 Buchholz-Mack-Todorov extensions We here quickly review the local extensions of the U (1)-current net constructed in [BMT88]. Let ρf be a DHR automorphism of the U (1)-current net localized in the interval I as above, such that ε+ ρf ,ρf = 1. To shorten notation denote ρ = ρf . Any such automorphism gives a local extension of the net by a crossed product with the group Z which acts on the net as powers of ρ. Let H :=Mk∈Z Hk with Hk = H (= vacuum Hilbert space of the U (1)-current net) and let π be a representation of the quasilocal C ∗-algebra AU (1) of the net restricted to R ∼= S1 r {1}, defined as π : AU (1) → B( H) π(a) :=Mk∈Z ρk(a) Denote by U the shift operator on H, i.e., U{ξk}k∈Z = {ξk+1}k∈Z for ξ ∈ H. It is clear that the shift operator U implements the localized automorphism ρ in this representation In other words U is a charged field for ρ. U π(a)U ∗ = π(ρ(a)). Definition 10.1. The BMT (Buchholz-Mack-Todorov) extension {Bρ} = {AU (1) ⋊ρ Z} is the net given by Bρ(I) := hπ(A(I)), Ui Bρ(J) := hπ(A(J)), π(uJ )Ui . It is an easy matter to check that this definition is well posed and the net is isotonous (it follows directly from Haag duality of the U (1)-current net on R, i.e., strong additivity). Locality of BMT extensions {Bρ} follows from ε+ ρ,ρ = 1, cf. Theorem 6.8. The inclusion {A ⊂ Bρ} is clearly discrete and irreducible. The DHR automorphisms of the U (1)-current extend to representations of the net {Bρ}, and the DHR sectors of BMT extensions were already classified in [BMT88]. We recall these facts to establish the notation. ρ,ρ = ε− ρ,σ = ε− Proposition 10.2. [BMT88]. For every DHR automorphism σ of AU (1) there are two locally normal representations σ± of Bρ such that σ±(π(a)) = π(σ(a)), a ∈ AU (1). Moreover, σ+ = σ− if and only if σ+ (or equivalently σ−) is a DHR representation of the net {Bρ}, if and only ρ,σ. Otherwise σ± have solitonic localization (they are localizable in half-lines). In if ε+ particular, there are 2N inequivalent DHR automorphisms of the net {Bρ}, where Q = √2πN is the charge of ρ. Proof. The automorphisms σ± can be defined by α-induction of σ for the extension {AU (1) ⊂ Bρ}, [LR95, Prop. 3.9], but we here describe them explicitly since we will need them in the following. We first define the action of σ± on the -algebra B generated by π(AU (1)) and the shift U . Define σ±(π(a)) := π(σ(a)) 42 σ±(U n) := π(ε± ρn,σ)U n To check that this is a well defined endomorphism of the -algebra it is enough to check that σ±(U ∗) = σ±(U )∗, σ±(U )σ±(π(a))σ±(U )∗ = σ±(π(ρ(a))). The first relation is an immediate consequence of naturality of the braiding, for the second we have σ±(U )σ±(π(a))σ±(U )∗ = π(ε± ρ,σ)π(ρ(σ(a)))π(ε± = π(ε± ρ,σ)U π(σ(a))U ∗π(ε± ρ,σ)∗ = π(σ(ρ(a))). ρ,σ)∗ Now, observe that for a fixed proper bounded interval J of R, the endomorphism (σ)± restricted to Bρ(J) ∩ B is locally implemented by the unitary π(u I ) := π(W (LI→ I)) where I is a proper bounded interval where I < J if we consider σ+ and J < I if we consider σ−, i.e., σ±(b) = Adπ(u I )(b) for every b ∈ Bρ(J) ∩ B. Since Bρ(J) ∩ B is ultraweakly dense in Bρ(J), the endomorphism can be extended in a unique way consistently on every local algebra. Regarding the localization of σ±, if J < I σ+(π(uJ )U ) = π(σ(uJ ))π(ε+ ρ,σ)U = π(σ(uJ ))π(σ(uJ ))∗π(uJ )U = π(uJ )U Similarly for I < J we have the same result for σ−. Thus they are localizable in half-lines, a priori, and also DHR if and only if ε+ ρ,σ = ε− ρ,σ. 10.2 Braided product of BMT extensions Let {BρL},{BρR} be two local BMT extensions of the U (1)-current net given by two DHR automorphisms ρL and ρR as in the previous section. We would like to construct the braided product of two such nets in a concrete fashion. Let Ml∈Z H = M(l,h)∈Z2 H ιL : BρL →Ml∈Z Let ιL be the solitonic representation of BρL defined on the above Hilbert space as follows where H is the vacuum Hilbert space of the U (1)-current net, and ΩAU (1) is the vacuum vector. We denote Ω = { Ωl,h}(l,h)∈Z2 , with Ωl,h := δl,0δh,0ΩAU (1). B( H) ⊂ B( M(l,k)∈Z2 H) ιL :=Ml∈Z B( H) ⊂ B( M(l,k)∈Z2 H) ιR :=Mh∈Z ιR : BρR →Mh∈Z and similarly for ιR ρR ρL 43 Define ε : M(l,h)∈Z2 H → M(l,h)∈Z2 H ε{ξl,h}(l,h)∈Z2 := {ε± L,ρh ρl R ξl,h}(l,h)∈Z2 and twist the representation ιR by ε Observe that ιR := Adε(ιR(·)). ιL ◦ πL = ιR ◦ πR where πL and πR are the inclusion maps of AU (1) into BρL and BρR respectively, explicitly ρl L(ρh R(a))) = (ιR ◦ πR)(a) (ιL ◦ πL)(a) = Mk,l∈Z ρh R(ρl L(a)) = Adε(Mk,l∈Z for every a ∈ AU (1). Let U ∈ BρL(I) and V ∈ BρR(I) be the charged fields for the DHR automorphisms ρL and ρR respectively. Then Proposition 10.3. ιR(V )ιL(U ) = ιL(πL(ε± ρL,ρR)) ιL(U )ιR(V ) Proof. By direct computation. Proposition 10.4. Let {BρL} and {BρR} two local BMT extensions as above. The net of von Neumann algebras defined by B±(I) :=(cid:10)ιL ◦ πL(AU (1)(I)), ιL(U ), ιR(V )(cid:11) , B±(J) :=(cid:10)ιL ◦ πL(AU (1)(J)), ιL ◦ πL(uJ )ιL(U ), ιR ◦ πR(vJ )ιR(V )(cid:11) , where uJ and vJ are unitary charge transporters respectively for ρL and ρR between intervals I and J (i.e. the endomorphisms AduJ ρL and AdvJ ρR are localized in J), is unitarily equivalent to the braided product net, i.e. { B±} ∼= {BρL ×± ε BρR}. Proof. By Lemma 8.5, Proposition 10.3 and the relation ιL ◦ πL = ιR ◦ πR, we know that there exists a surjective homomorphism of -algebras φ : BL×R → B ε BρR is defined as in Lemma 8.5 and B ⊂ B± is the -algebra generated where BL×R ⊂ BρL×± by ιL ◦ πL(AU (1)) and ιL(U ), ιR(V ). By the GNS theorem for -algebras, see e.g. [KM15, Sec. 1.3], in order to show that φ is implemented by a unitary it is enough to check that ω0 ◦ ELR = ( Ωφ(·) Ω), where Ω is the vacuum vector of { B±}. This is clear since, for x = Pi,j L/R(xi,j)L(U i)R(V j) ∈ BL×R, we have ω0 ◦ ELR(x) = (ΩAU (1), x0,0ΩAU (1)) = ( Ω, φ(·) Ω). 44 By considering the braided product of a local BMT extension with itself (as concretely constructed in the previous proposition by taking ρL = ρR = ρ) we give examples where the center of the braided product is a continuous algebra, more specifically L∞(S1, dµ). Proposition 10.5. Let {Bρ} be the BMT extension obtained from a DHR automorphism ε Bρ) ∼= ρ and let {Bρ ×± L∞(S1, dµ) with dµ the Lebesgue measure on the circle. ε Bρ} be the braided product extension with itself. Then Z(Bρ ×± Proof. Recall that the center of the braided product is the same as the relative commutant Z(Bρ ×± ε Bρ) = Z((Bρ ×± ε Bρ)(J)) = (Bρ ×± ε Bρ)(J)′ ∩ ι(AU (1)(J)) ε Bρ) V −iU ixi for any proper bounded interval J of R. Thus Lemma 8.6 provides an expansion for elements x ∈ Z(Bρ ×± x =Xi∈Z with xi ∈ HomDHR{AU (1)}(id, ρiρ−i = id) ∼= C. It is easy to see that there is an isomorphism between the -algebra generated by the {U iV −i}i and the -algebra generated by the characters of the circle. This same map is also an isomorphisms of pre-Hilbert spaces with inner product on one side induced by the vacuum state ω = ω0 ◦ ELR, where ELR is the standard expectation of the braided product net (Proposition 8.2) and ω0 the vacuum state for {AU (1)}, and on the other side the usual L2(S1, dµ) inner product. Thus let B denote the -algebra generated by the {U iV −i}i, ¯Bk·kω its Hilbert completion and let Char(S1) be the *-algebra generated by characters of the circle. ¯Bk·kω ∼= L2(S1, dµ) as Hilbert spaces and let W be the unitary which implements the isomorphism. If πω is the GNS representation of B induced by the state ω = ω0◦ E (on the Hilbert space ¯Bk·kω ), and if πdµ is the GNS representation of Char(S1), we have AdW πω = πdµ. Hence the isomorphism extends to the ultraweak closure, and Z(Bρ ×± ε Bρ)(I) ∼= πω(Z(Bρ ×± ε Bρ)) = πω(B)′′ ∼= πdµ(Char(S1))′′ ∼= L∞(S1, dµ) concluding the proof. We thus have an example of an uncountable family of (one-dimensional) irreducible phase boundaries, parametrized by S1, obtained from the braided product construction. This is obviously in contrast with the finite index case, where the relative commutant is necessarily finite-dimensional. But the difference from the finite index case is actually greater than this: we have an example where the relative commutant is not a discrete algebra. This means that the disintegration in Proposition 9.5 that yields irreducible phase boundaries is not a direct sum. Moreover it is not true, in contrast with the finite index case, that every irreducible phase boundary condition comes from a representation of the braided product extension, see [BKLR16, Prop. 5.1, Cor. 5.3], due to the absence of non-trivial minimal central projections. Similarly, one can construct examples where the braided product is itself an irreducible extension and thus it yields a unique irreducible phase boundary. It is not hard to see that this is the case for the braided product of two local BMT extensions of the U (1)-current whose generating DHR automorphisms ρf1, ρf2 have charges RS1 f1 and RS1 f2 with irrational quotient. The claim simply follows from the expansion given in Lemma 8.6 and observing that, in this case, the dual canonical endomorphisms of the BMT extensions θ1 and θ2 have only one irreducible subendomorphism in common: the identity. 45 11 Conclusions Index theory provides an elegant and effective machinery to classify and construct extensions of von Neumann algebras and local nets. When this framework is not fully applicable (infi- nite index case), we have seen that under some physically meaningful structural hypotheses (semidiscreteness, discreteness) some of these results can be suitably generalized. The price to pay is abandoning the purely categorical setting of finite index Q-systems by the emer- gence of analytical conditions. At the same time, these analytical conditions (convergence of projections, faithfulness of expectations) provide a way to control infinite objects (gauge groups, representation categories, sets of generating fields) exploiting techniques of Operator Algebras in their application to QFT. In particular, we have introduced the notion of generalized Q-system of intertwiners (in the category of localizable superselection sectors DHR{A}) for a local net {A}, and we have shown that from this data a net extension of {A}, in the spirit of [LR95], can be constructed. At the level of properly infinite inclusions, we have seen that the existence of generalized Q- systems of intertwiners is equivalent to the inclusion to be of discrete type. When passing from subfactors to inclusions of local nets as in [LR95] this matter is more subtle, and we provided sufficient conditions to guarantee the existence of generalized Q-systems of intertwiners for nets, which cover most interesting examples in low and higher spacetime dimensions. We leave open the question on whether these conditions are always verified by discrete QFT extensions. The notion of generalized Q-system of intertwiners lends itself to generalize the definition of braided product between ordinary Q-systems. After proving that the analytic properties of generalized Q-systems of intertwiners turn out to be compatible with the purely algebraic definition of the braided product, we explore some properties of the resulting net extension, showing that it retains some features of its finite index counterpart. In particular, in the case of chiral CFTs, we have seen that its central decomposition can yield uncountable families of irreducible phase boundaries with infinite index. An important issue left open is the classification of all phase boundary conditions among two CFTs. In particular, one would like to understand if, in analogy with [BKLR16], all the boundary conditions arise in the disintegration of the center of the braided product. Although the discrete case covers many physical examples, e.g., every orbifold construc- tion by a compact group, the setting of greatest generality for irreducible inclusions of local CFTs (at least assuming the existence of a vacuum vector) is semidiscreteness. Generalized Q-systems do always exist for semidiscrete extensions of properly infinite von Neumann al- gebras [FI99]. An issue that would be worth analyzing further is if methods similar to those explored in this paper can be generalized to treat extensions of local nets which are semidis- crete but not discrete [Car04], [Xu05]. It would also be interesting to extend the analysis of discrete inclusions to the case of non-separable Hilbert spaces, given that good candidates for such extensions in QFT already appear in [Cio09], [MTW16]. Lastly, we mention that one can easily construct discrete non-finite local extensions which are not compact group orbifolds by taking tensor products of local nets, 9. It would also be worth investigating which kind of extensions can arise from braided products of compact group orbifolds, given that, by the arguments of our last section, one can construct extensions whose generating fields have the commutation relations of non-commutative tori. 9We thank Y. Tanimoto for pointing out this interesting fact. 46 Acknowledgements. Supported by the European Research Council (ERC) through the Advanced Grant QUEST "Quantum Algebraic Structures and Models", and by PRIN-MIUR. We are indebted to R. Longo for proposing us the problem investigated in this work, and to M. Bischoff and K.-H. Rehren for many discussions and suggestions, and for their motivating interest. We also thank I. Khavkine and Y. Tanimoto for useful comments and criticism. L.G. wishes to thank K.-H. Rehren for an invitation to Gottingen (Institut fur Theoretis- che Physik, Georg-August-Universitat) and W. Yuan for an invitation to Beijing (Academy of Mathematics and Systems Science, Chinese Academy of Sciences), where this work has been presented, and thanks them for hospitality and for useful conversations in both these occasions. References [AMP15] N. Afzaly, S. Morrison, and D. Penneys. The classification of subfactors with index at most 5 1 4 . preprint arXiv:1509.00038, 2015. [BDH88] M. Baillet, Y. Denizeau, and J.-F. Havet. Indice d'une esp´erance conditionnelle. Compo- sitio Math., 66:199 -- 236, 1988. [BGL93] R. Brunetti, D. Guido, and R. Longo. Modular structure and duality in conformal quantum field theory. Comm. Math. Phys., 156:201 -- 219, 1993. [Bis17] M. Bischoff. Generalized orbifold construction for conformal nets. Rev. Math. Phys., 29:1750002 1 -- 53, 2017. [BKL15] M. Bischoff, Y. Kawahigashi, and R. Longo. Characterization of 2D rational local con- formal nets and its boundary conditions: the maximal case. Doc. Math., 20:1137 -- 1184, 2015. [BKLR15] M. Bischoff, Y. Kawahigashi, R. Longo, and K.-H. Rehren. Tensor categories and endo- morphisms of von Neumann algebras. With applications to quantum field theory, Springer Briefs in Mathematical Physics, Vol. 3. Springer, Cham, 2015. [BKLR16] M. Bischoff, Y. Kawahigashi, R. Longo, and K.-H. Rehren. Phase Boundaries in Algebraic Conformal QFT. Comm. Math. Phys., 342:1 -- 45, 2016. [BMT88] D. Buchholz, G. Mack, and I. Todorov. The current algebra on the circle as a germ of local field theories. Nucl. Phys., B, Proc. Suppl., 5:20 -- 56, 1988. [BP01] [BR16] [BW75] [Car04] V. Berestovskii and C. Plaut. Covering group theory for topological groups. Topology Appl., 114:141 -- 186, 2001. M. Bischoff and K.-H. Rehren. The hypergroupoid of boundary conditions for local quan- tum observables. preprint arXiv:1612.02972, 2016. J. J. Bisognano and E. H. Wichmann. On the duality condition for a Hermitian scalar field. J. Mathematical Phys., 16:985 -- 1007, 1975. S. Carpi. On the representation theory of Virasoro nets. Comm. Math. Phys., 244:261 -- 284, 2004. [CCG+04] A. Connes, J. Cuntz, E. Guentner, N. Higson, J. Kaminker, and J. E. Roberts. Non- commutative geometry, Lecture Notes in Mathematics, Vol. 1831. Springer-Verlag, Berlin; Centro Internazionale Matematico Estivo (C.I.M.E.), Florence, 2004. Lectures given at the C.I.M.E. Summer School held in Martina Franca, September 3 -- 9, 2000, Edited by S. Doplicher and R. Longo. 47 [CDR01] R. Conti, S. Doplicher, and J. E. Roberts. Superselection theory for subsystems. Comm. Math. Phys., 218:263 -- 281, 2001. [Cio09] F. Ciolli. Massless scalar free field in 1 + 1 dimensions. I. Weyl algebras products and superselection sectors. Rev. Math. Phys., 21:735 -- 780, 2009. [CKL08] [Con73] S. Carpi, Y. Kawahigashi, and R. Longo. Structure and classification of superconformal nets. Ann. Henri Poincar´e, 9:1069 -- 1121, 2008. A. Connes. Une classification des facteurs de type III. Ann. Sci. ´Ecole Norm. Sup., 6:133 -- 252, 1973. [DHR69a] S. Doplicher, R. Haag, and J. E. Roberts. Fields, observables and gauge transformations. I. Comm. Math. Phys., 13:1 -- 23, 1969. [DHR69b] S. Doplicher, R. Haag, and J. E. Roberts. Fields, observables and gauge transformations. II. Comm. Math. Phys., 15:173 -- 200, 1969. [DHR71] S. Doplicher, R. Haag, and J. E. Roberts. Local observables and particle statistics. I. Comm. Math. Phys., 23:199 -- 230, 1971. [DHR74] S. Doplicher, R. Haag, and J. E. Roberts. Local observables and particle statistics. II. Comm. Math. Phys., 35:49 -- 85, 1974. [DL84] S. Doplicher and R. Longo. Standard and split inclusions of von Neumann algebras. Invent. Math., 75:493 -- 536, 1984. [DMV04] C. D'Antoni, G. Morsella, and R. Verch. Scaling algebras for charged fields and short- distance analysis for localizable and topological charges. Ann. Henri Poincar´e, 5:809 -- 870, 2004. [DR72] [DR89] [DR90] [DV17] [EP03] [FI95] [FI99] [FJ96] S. Doplicher and J. E. Roberts. Fields, statistics and non-abelian gauge groups. Comm. Math. Phys., 28:331 -- 348, 1972. S. Doplicher and J. E. Roberts. Endomorphisms of C ∗-algebras, cross products and duality for compact groups. Ann. of Math., 130:75 -- 119, 1989. S. Doplicher and J. E. Roberts. Why there is a field algebra with a compact gauge group describing the superselection structure in particle physics. Comm. Math. Phys., 131:51 -- 107, 1990. S. Del Vecchio. Extensions in Quantum Field Theory: Q-systems and defects for infinite index inclusions. PhD thesis, Universit`a degli Studi di Roma Tor Vergata, Facolt`a di Scienze Matematiche Fisiche e Naturali, 2017. D. E. Evans and P. R. Pinto. Subfactor realisation of modular invariants. Comm. Math. Phys., 237:309 -- 363, 2003. F. Fidaleo and T. Isola. On the conjugate endomorphism in the infinite index case. Math. Scand., 77:289 -- 300, 1995. F. Fidaleo and T. Isola. The canonical endomorphism for infinite index inclusions. Z. Anal. Anwendungen, 18:47 -- 66, 1999. K. Fredenhagen and M. Jorss. Conformal Haag-Kastler nets, pointlike localized fields and the existence of operator product expansions. Comm. Math. Phys., 176:541 -- 554, 1996. [FRS92] K. Fredenhagen, K.-H. Rehren, and B. Schroer. Superselection sectors with braid group statistics and exchange algebras. II. Geometric aspects and conformal covariance. Rev. Math. Phys., SI1 (Special Issue):113 -- 157, 1992. [GL92] D. Guido and R. Longo. Relativistic invariance and charge conjugation in quantum field theory. Comm. Math. Phys., 148:521 -- 551, 1992. 48 [GL95] [GL96] D. Guido and R. Longo. An algebraic spin and statistics theorem. Comm. Math. Phys., 172:517 -- 533, 1995. D. Guido and R. Longo. The conformal spin and statistics theorem. Comm. Math. Phys., 181:11 -- 35, 1996. [GLW98] D. Guido, R. Longo, and H.-W. Wiesbrock. Extensions of conformal nets and superselec- tion structures. Comm. Math. Phys., 192:217 -- 244, 1998. [GR15] L. Giorgetti and K.-H. Rehren. Braided categories of endomorphisms as invariants for local quantum field theories. preprint arXiv:1512.01995 (to appear in Comm. Math. Phys.), 2015. [Haa87] U. Haagerup. Connes' bicentralizer problem and uniqueness of the injective factor of type III1. Acta Math., 158:95 -- 148, 1987. [Haa96] R. Haag. Local quantum physics. Springer Berlin, 1996. [Hia88] F. Hiai. Minimizing indices of conditional expectations onto a subfactor. Publ. Res. Inst. Math. Sci., 24:673 -- 678, 1988. [HKZ91] H. Halpern, V. Kaftal, and L. Zsid´o. Finite weight projections in von Neumann algebras. [HM06] [HO89] Pacific J. Math., 147:81 -- 121, 1991. H. Halvorson and M. Muger. math-ph/0602036, 2006. Algebraic quantum field theory. preprint arXiv R. H. Herman and A. Ocneanu. Index theory and Galois theory for infinite index inclusions of factors. C. R. Acad. Sci. Paris S´er. I Math., 309:923 -- 927, 1989. [ILP98] M. Izumi, R. Longo, and S. Popa. A Galois correspondence for compact groups of au- tomorphisms of von Neumann algebras with a generalization to Kac algebras. J. Funct. Anal., 155:25 -- 63, 1998. [JMS14] V. F. R. Jones, S. Morrison, and N. Snyder. The classification of subfactors of index at most 5. Bull. Amer. Math. Soc., 51:277 -- 327, 2014. [Jon83] V. F. R. Jones. Index for subfactors. Invent. Math., 72:1 -- 25, 1983. [Jon99] V. F. R. Jones. Planar Algebras. I. preprint arXiv math.QA/9909027, 1999. [Kad63] [KL04] R. V. Kadison. Remarks on the type of von Neumann algebras of local observables in quantum field theory. J. Mathematical Phys., 4:1511 -- 1516, 1963. Y. Kawahigashi and R. Longo. Classification of local conformal nets. Case c < 1. Ann. Math., 160:493 -- 522, 2004. [KLM01] Y. Kawahigashi, R. Longo, and M. Muger. Multi-interval subfactors and modularity of representations in conformal field theory. Comm. Math. Phys., 219:631 -- 669, 2001. [KM15] [Kos86] [Kos89] [Lon79] I. Khavkine and V. Moretti. Algebraic QFT in curved spacetime and quasifree Hadamard states: an introduction. In Advances in algebraic quantum field theory, Math. Phys. Stud., 191 -- 251. Springer, Cham, 2015. H. Kosaki. Extension of Jones' theory on index to arbitrary factors. J. Funct. Anal., 66:123 -- 140, 1986. H. Kosaki. Characterization of crossed product (properly infinite case). Pacific J. Math., 137:159 -- 167, 1989. R. Longo. Notes on algebraic invariants for noncommutative dynamical systems. Comm. Math. Phys., 69:195 -- 207, 1979. [Lon87] R. Longo. Simple injective subfactors. Adv. Math., 63:152 -- 171, 1987. 49 [Lon89] [Lon90] [Lon94] [Lon97] [Lon03] [Lon08] R. Longo. Index of subfactors and statistics of quantum fields. I. Comm. Math. Phys., 126:217 -- 247, 1989. R. Longo. Index of subfactors and statistics of quantum fields. II. Correspondences, braid group statistics and Jones polynomial. Comm. Math. Phys., 130:285 -- 309, 1990. R. Longo. A duality for Hopf algebras and for subfactors. I. Comm. Math. Phys., 159:133 -- 150, 1994. R. Longo. An analogue of the Kac-Wakimoto formula and black hole conditional entropy. Comm. Math. Phys., 186:451 -- 479, 1997. R. Longo. Conformal subnets and intermediate subfactors. Comm. Math. Phys., 237:7 -- 30, 2003. R. Longo. Lecture notes on conformal nets. Part II. Nets of von Neumann algebras. Lecture notes, available at http://www.mat.uniroma2.it/longo/Home.html, 2008. [LR95] R. Longo and K.-H. Rehren. Nets of subfactors. Rev. Math. Phys., 7:567 -- 597, 1995. [LR97] R. Longo and J. E. Roberts. A theory of dimension. K-Theory, 11:103 -- 159, 1997. [LR04] R. Longo and K.-H. Rehren. Local fields in boundary conformal QFT. Rev. Math. Phys., 16:909 -- 960, 2004. [MTW16] V. Morinelli, Y. Tanimoto, and M. Weiner. Conformal covariance and the split property. preprint arXiv:1609.02196 (to appear in Comm. Math. Phys.), 2016. [Mug05] M. Muger. Conformal Orbifold Theories and Braided Crossed G-Categories. Comm. Math. Phys., 260:727 -- 762, 2005. [Pop95a] S. Popa. An axiomatization of the lattice of higher relative commutants of a subfactor. Invent. Math., 120:427 -- 445, 1995. [Pop95b] S. Popa. Classification of subfactors and their endomorphisms, CBMS Regional Conference Series in Mathematics, Vol. 86. Amer. Math. Soc., Providence, RI, 1995. [PP86] M. Pimsner and S. Popa. Entropy and index for subfactors. Ann. Sci. Ecole Norm. Sup, 19:57 -- 106, 1986. [Reh94] K.-H. Rehren. A new view of the Virasoro algebra. Lett. Math. Phys., 30:125 -- 130, 1994. [Reh15] [Rob74] [Tur10] K.-H. Rehren. Algebraic conformal quantum field theory in perspective. In: Advances in Algebraic Quantum Field Theory, R. Brunetti et al., eds., Mathematical Physics Studies, 331 -- 364. Springer International Publishing, 2015. J. E. Roberts. Some applications of dilatation invariance to structural questions in the theory of local observables. Comm. Math. Phys., 37:273 -- 286, 1974. V. Turaev. Homotopy quantum field theory, EMS Tracts in Mathematics, Vol. 10. Eu- ropean Mathematical Society (EMS), Zurich, 2010. Appendix 5 by Michael Muger and Appendices 6 and 7 by Alexis Virelizier. [Wei08] M. Weiner. Restricting positive energy representations of Diff +(S 1) to the stabilizer of n points. Comm. Math. Phys., 277:555 -- 571, 2008. [Xu00] F. Xu. Algebraic orbifold conformal field theories. Proc. Nat. Acad. Sci. U.S.A., 97:14069, 2000. [Xu05] F. Xu. Strong additivity and conformal nets. Pacific J. Math., 221:167 -- 199, 2005. [Xu14] F. Xu. On intermediate conformal nets. J. Reine Angew. Math., 692:125 -- 151, 2014. 50
1802.09964	3	1802	2019-10-10T15:26:19	On the vanishing cohomology problem for cocycle actions of groups on II$_1$ factors	[ "math.OA" ]	We prove that any free cocycle action of a countable amenable group $\Gamma$ on any II$_1$ factor $N$ can be perturbed by inner automorphisms to a genuine action. This {\em vanishing cohomology} property, that we call $\mathcal V\mathcal C$, is also closed to free products with amalgamation over finite groups. But beyond this no other examples of $\mathcal V\mathcal C$-groups are known. In turn, by considering special cocycle actions $\Gamma \curvearrowright N$ in the case $N$ is the hyperfinite II$_1$ factor $R$, respectively the free group factor $N=L(\Bbb F_\infty)$, we exclude many groups from being $\mathcal V\mathcal C$. We also show that any free action $\Gamma \curvearrowright R$ gives rise to a free cocycle $\Gamma$-action on the II$_1$ factor $R'\cap R^\omega$ whose vanishing cohomology is equivalent to Connes' Approximate Embedding property for the II$_1$ factor $R\rtimes \Gamma$.	math.OA	math	ON THE VANISHING COHOMOLOGY PROBLEM FOR COCYCLE ACTIONS OF GROUPS ON II1 FACTORS SORIN POPA University of California, Los Angeles Abstract. We prove that any free cocycle action of a countable amenable group Γ on any II1 factor N can be perturbed by inner automorphisms to a genuine action. Besides being satisfied by all amenable groups, this universal vanishing cohomology property, that we call V C, is also closed to free products with amalgamation over finite groups. While no other examples of V C-groups are known, by considering special cocycle actions Γ y N in the case N is the hyperfinite II1 factor R, respectively the free group factor N = L(F∞), we exclude many groups from being V C. We also show that any free action Γ y R gives rise to a free cocycle Γ-action on the II1 factor R′ ∩ Rω whose vanishing cohomology is equivalent to Connes' Approximate Embedding property for the II1 factor R ⋊ Γ. 0. Introduction A cocycle action of a group Γ on a II1 factor N is a map σ : Γ → Aut(N ) which is multiplicative modulo inner automorphisms of N , σgσh = Ad(vg,h)σgh, ∀g, h ∈ Γ, with the unitary elements vg,h ∈ U(N ) satisfying the cocycle relation vg,hvgh,k = σg(vh,k)vg,hk, ∀g, h, k ∈ Γ. If Γ is a free group Fn, for some 1 ≤ n ≤ ∞, then any cocycle Γ-action on any II1 factor N can obviously be perturbed by inner automorphisms {Ad(wg)}g of N so that to become a "genuine" action, i.e., such that σ′ g = Ad(wg)σg is a group morphism, in fact so that the stronger condition vg,h = σg(w∗ g wgh, ∀g, h, holds true. We obtain in this paper several results towards identifying the class VC of all countable groups Γ that satisfy this universal vanishing cohomology property. Thus, we first prove that any free product of amenable groups amalgamated over a common finite subgroup is in the class VC. Then we show that if a group Γ has h)w∗ Supported in part by NSF Grant DMS-1700344 1 Typeset by AMS-TEX 2 SORIN POPA an infinite subgroup which either has relative property (T), or has non-amenable centralizer, then Γ is not in VC. To prove that all amenable groups lie in VC we use subfactor techniques to reduce the problem to the case N is the hyperfinite II1 factor R, where vanishing cohomology holds due to results in ([Oc85]). To exclude groups from being in VC we apply W∗-rigidity results to two types of cocycle actions that are "hard to untwist": the ones arising from t-amplifications of Bernoulli actions on N = R introduced in [P01]; and the ones considered in [CJ84], arising from normal inclusions F∞ ֒→ Fn with Fn/F∞ = Γ, which give cocycle Γ-actions on N = L(F∞). Untwisting cocycle actions on II1 factors is a basic question in non-commutative ergodic theory and very specific to this area. Besides its intrinsic interest, the prob- lem occurs in the classification of group actions on II1 factors ([C74], [J80], [Oc85], [P01a]) and, closely related to it, in the classification of factors through unique crossed-product decomposition (as in [C74] for amenable factors, or [P01a], [P03], [P06a], [IPeP05], [PV12] for non-amenable II1 factors). Another aspect, which goes back to ([CJ84]) and is important in W∗-rigidity, relates non-vanishing cohomology for certain cocycle Γ-actions on L(F∞) to non-embeddability of L(Γ) into L(Fn). From an opposite angle, which offers a new point of view much emphasized here, vanishing cohomology results for cocycle actions are relevant to embedding prob- lems, such as finding unusual group factors that embed into L(F2) and Connes Approximate Embedding conjecture. To describe the results in this paper in more details we need some background and notations. Let us first note that cocycle actions are more restrictive than outer actions, which are maps σ : Γ → Aut(N ) that only require σgσhσ−1 gh ∈ Inn(N ), ∀g, h ∈ Γ. It has in fact been shown in ([NT59]) that there is a scalar 3-cocycle νσ ∈ H 3(Γ) associated to an outer action σ. If σ is free, i.e., σg 6∈ Inn(N ), ∀g 6= e, then νσ is trivial if and only if σ is a cocycle action. Thus, if we view the vanishing cohomology problem as a question about lifting a 1 to 1 group morphism σ : Γ → Out(N ) to a group morphism into Aut(N ), then the problem is not well posed unless one requires νσ ≡ 1, i.e., that σ defines a cocycle action. Like for genuine actions, one can associate to a cocycle action Γ yσ N a tracial crossed product von Neumann algebra N ⋊ Γ, with the freeness of σ equivalent to the condition N ′ ∩ N ⋊ Γ = C1. Thus, if σ is free then N ⊂ M = N ⋊ Γ is an irreducible inclusion of II1 factors with the normalizer of N in M generating M as a von Neumann algebra (N is regular in M ). Conversely, any irreducible regular inclusion of II1 factors N ⊂ M arises this way, from a crossed product construction involving a free cocycle action (cf. [J80]). The crossed product framework allows an alternative formulation of vanishing cohomology. Thus, if M = N ⋊ Γ denotes the crossed product II1 factor associated VANISHING COHOMOLOGY 3 with the free cocycle action (σ, v) of Γ on N , and we let {Ug}g ⊂ M denote the canonical unitaries implementing σ on N , then the existence of wg ∈ U(N ) such that vg,h = wgσg(wh)w∗ gh, ∀g, h (i.e., vanishing cohomology for v) amounts to U ′ g), g ∈ Γ, is a genuine action (i.e., weak vanishing cohomology for v) amounts to the weaker condition that {U ′ g = wgUg being a Γ-representation. While the condition that σ′ g}g is a projective Γ-representation. g = Ad(U ′ Given a II1 factor N , we denote by VC(N ) (respectively VCw(N )) the class of countable groups Γ with the property that any free cocycle action of Γ on N satisfies the strong form (respectively weak form) of the vanishing cohomology. Also, we denote by VC (respectively VCw) the class of countable groups Γ with the property that any free cocycle action of Γ on any II1 factor N satisfies the strong form (respectively weak form) of the vanishing cohomology. The class VC contains all finite groups by ([J80], [Su80]) and all groups with polynomial growth by ([P89]). The first main result in this paper, which we prove in Section 2, shows that in fact VC contains all countable amenable groups. Since by [J80] all 1-cocycles for actions of finite groups are co-boundaries, this allows to deduce that, more than just containing the free groups, all amalgamated free products of amenable groups over finite groups belong to VC. 0.1. Theorem. The class VC contains all countable amenable groups. Also, if {Γn}n is a sequence of groups in VC and K ⊂ Γn is a common finite subgroup, n ≥ 1, then Γ1 ∗K Γ2 ∗K ... ∈ VC. To prove the first part of this result we show that any cocycle action σ of a countable amenable group Γ on a separable II1 factor N can be perturbed by inner automorphisms to a cocycle action σ′ that leaves invariant an irreducible hyperfinite −1 are implemented subfactor R ⊂ N with the additional property that σ′ by unitaries in R, ∀g, h, with σ′ still free when restricted to R (see Theorem 2.1). This reduces the vanishing cohomology problem to the case N = R, where one can apply the vanishing cohomology result in ([Oc85]) to finish the proof. hσ′ gh gσ′ To prove the existence of a "large" R ⊂ N that's normalized by inner perturba- tions of σ we use an idea introduced in ([P89]; cf. also 5.1.5 in [P91]), of translating the problem into the question of whether there exists a sub-inclusion of hyperfinite factors inside the "diagonal subfactor" N ≃ N σ ⊂ M σ associated with σ, so that to have a non-degenerate commuting square satisfying a strong smoothness condi- tion on higher relative commutants. This subfactor problem was solved in [P89] in the case Γ is finitely generated with trivial Poisson boundary (e.g., with poly- nomial growth; see [KV83]), by constructing R as a limit of relative commutants P ′ n ∩ N of factors in a tunnel N ⊃ N1..., obtained by iterating the downward basic construction (in the spirit of [P91], [P93]). 4 SORIN POPA However, that construction depends crucially on the trivial Poisson boundary condition on Γ. We use here the amenability of Γ alone to construct a more elab- orate decreasing sequence of subfactors Pn ⊂ N with P ′ n ∩ N ր R "large" in N , obtained through reduction/induction in Jones tunnels. In fact, this method allows us to obtain the existence of strongly smooth embedding of hyperfinite subfactors into any finite index subfactor N ⊂ M with standard invariant GN⊂M amenable (in the sense of [P91], i.e., with its graph ΓN⊂M satisfying the Kesten-type condition kΓN⊂M k2 = [M : N ]; see also [P93], [P94a], [P97] for other equivalent definitions). We in fact show that given any amenable C∗-category G of endomorphisms of a II∞ factor N (viewed here as an outer action of an abstract rigid C∗-tensor cate- gory), there exists a "large" approximately finite dimensional (AFD) II∞ subfactor R⊗B(ℓ2N) in N that's normalized by G, modulo inner automorphisms (see Theorem 2.12). In Section 5 we use the strong solidity of free group factors ([OP07]) to prove that in order for a group Γ to satisfy the property that any of its actions on II1 factors normalizes a hyperfinite subfactor (modulo inner automorphisms), Γ must necessarily be amenable. The problem of whether this dichotomy still holds for subfactor standard invariants and rigid C∗-tensor categories, remains open. In turn, in Sections 3 and 4 we obtain a series of obstruction criteria for groups to belong to the classes VC(R), VC(L(F∞)), VC, summarized in the following: 0.2. Theorem. 1◦ If a countable group Γ has an infinite subgroup which either has the relative property (T), or has non-amenable centralizer in Γ, then Γ 6∈ VCw(R). 2◦ Assume a countable group Γ satisfies one of the following: (a) Γ does not have Haagerup property (e.g., it contains an infinite subset with relative property (T)); (b) The Cowling-Haagerup invariant Λ(Γ) is larger than 1; (c) Γ has an infinite subgroup with non-amenable centralizer; (d) Γ has an infinite amenable subgroup with non-amenable normalizer. Then Γ 6∈ VCw(L(F∞)), in particular Γ 6∈ VCw. To prove the restrictions on VC(R) we use the t-amplifications of Bernoulli ac- tions on R introduced in [P01a] and results obtained there and in ([P06a]) through In turn, to get restrictions on VCw(L(F∞)), we deformation-rigidity arguments. use the Connes-Jones (CJ) cocycles associated with surjective group morphisms π : FS → (Γ, S), extending the map assigning the free generators of FS to a set of generators S ⊂ Γ. As shown in [CJ84], if Γ is infinite, non-free, then kerπ ≃ F∞ and the inclusion L(F∞) = N ⊂ M = L(FS) is of the form N ⊂ N ⋊(σπ ,vπ) Γ, for a free cocycle action (σπ, vπ). The vanishing of the cocycle vπ implies that L(Γ) embeds into L(FS), hence 2◦ above follows from results in ([CJ84], [P01b], [O03], [P06b], [OP07]). The CJ-cocycles seem the "most difficult to untwist", in the sense that if all such VANISHING COHOMOLOGY 5 cocycle actions of Γ on L(F∞) untwist, then Γ ought to be in VC. In particular, this would show that VC = VC(L(F∞)). Since untwisting a CJ-cocycle for Γ implies that Γ is in the class W∗ leq(F2) of groups whose von Neumann algebra embeds into L(F2), one has VC ⊂ VC(L(F∞)) ⊂ W∗ leq(F2). Very little is actually known about the class W∗ leq(F2), which is extremely interesting by itself. Any success in proving VC property for some "exotic" group Γ, would provide embeddings L(Γ) ֒→ L(F2). In the final part of the paper we discuss a connection between vanishing coho- mology phenomena and Connes Approximate Embedding conjecture, on whether any separable II1 factor M embeds in the ultrapower Rω of the hyperfinite II1 factor R. Thus, we notice that any free action of a group Γ on R (such as the "non-commutative" Bernoulli action Γ y R⊗Γ ≃ R), gives rise to a free cocycle action of Γ on the centralizer Rω = R′ ∩ Rω of R in Rω. We deduce that this cocycle untwists if and only M = R ⋊σ Γ satisfies the conjecture. Acknowledgement. I am very grateful to Damien Gaboriau, Vaughan Jones, Jesse Peterson and Stefaan Vaes for many useful discussions related to this paper. I am also grateful to the referee for his/her many pertinent questions that led to what I believe to be a much improved final version. 1. Preliminaries and notations. For general background on II1 factors we refer the reader to ([AP17]; also [T79], [BrO08] for general theory of operator algebras and von Neumann algebras). 1.1. Cocycle actions and crossed products. Given a II1 factor N , we denote by Aut(N ) the group of automorphisms of N . An automorphism θ of N is inner if there exists u in the unitary group of N , U(N ), such that θ(x) = Adu(x) = uxu∗, ∀x ∈ N . We denote by Inn(N ) ⊂ Aut(N ) the group of all such inner automorphisms and by Out(N ) the quotient group Aut(N )/Inn(N ). Given a discrete group Γ, an action of Γ on N is a group morphism σ : Γ → Aut(N ). We will use the notation Γ yσ N to emphasize such an action. More generally, a cocycle action σ of Γ on N is a map σ : Γ → Aut(N ) with the property that there exists v : Γ × Γ → U(N ) such that: (1.1.1) (1.1.2) σe = id and σgσh = Ad vg,hσgh, ∀g, h ∈ G vg,hvgh,k = σg(vh,k)vg,hk, ∀g, h, k ∈ Γ. The cocycle action σ is free if σg cannot be implemented by unitary elements in N, ∀g 6= e, in other words if the factoring of σ through the quotient map Aut(N ) → Out(N ) is 1 to 1. All cocycle actions (in particular all actions) that we will consider in this paper are assumed to be free. 6 SORIN POPA Following ([KT02]), a map σ : Γ → Aut(N ) that's a 1 to 1 group morphism when factored through the quotient map Aut(N ) → Out(N ), is called an outer Γ-action (an alternative terminology for such σ is Q-kernel, notably used in [J80], [Oc85]). Thus, an outer action satisfies (1.1.1) above, but not necessarily (1.1.2). As shown in ([NT59]), if σ is an outer Γ-action, then one can associate to it a scalar 3-cocycle νσ ∈ H 3(Γ), with the property that νσ ≡ 1 if and only if σ is a cocycle action, and which one calls the H 3(Γ)-obstruction of σ. If σ is a cocycle action, then a map v satisfying (1.1.2) is called a 2-cocycle for σ. The 2-cocycle is normalized if vg,e = ve,g = 1, ∀g ∈ G. Note that any 2-cocycle satisfies ve,e ∈ C. Thus any 2-cocycle v can be normalized by replacing it, if necessary, by v′ e,e vg,h, g, h ∈ Γ. All 2-cocycles considered from now on will be normalized. g,h = v∗ Also, when given a cocycle action σ, we will sometimes specify from the beginning the 2-cocycle it comes with, thus considering it as a pair (σ, v). Note that the 2-cocycle v is unique modulo perturbation by a scalar 2-cocycle µ. More precisely, v′ : Γ × Γ → U(N ), with v′ e,e = 1, satisfies conditions (1.1.1), (1.1.2) if and only if v′ = µv for some scalar valued function µ : Γ × Γ → T satisfying µe,e = 1 and (1.1.3) µg,hµgh,k = µh,kµg,hk, ∀g, h, k ∈ Γ Let us recall the definition of the crossed product von Neumann algebra associated with a cocycle action (σ, v) of Γ on N , denoted N ⋊(σ,v) Γ (or simply N ⋊σ Γ if σ is a genuine action). So let H denote the Hilbert space ⊕h(L2N )h ≃ ℓ2(Γ, L2N ), which we view as the space of ℓ2-summable formal series Ph Uhξh, where {Ug}g∈Γ are here "indeterminates" (labels) and the "coefficients" ξh belong to L2N . We define a ∗-algebra structure on the subspace H0 ⊂ H of finitely supported sums with "bounded" coefficients ξg = xg ∈ N , and at the same time a Hilbert H0-bimodule structure on H, as follows. First, we let Ue act on both left and right on H as the identity idH and identify N with UeN acting left-right on H by x(Ph Uhξh)y = Ph Uh(σ−1 h (x)ξhy). Then we let Ug · Ph Uhξh = Ph vg,hUghξh, which by the change of variables h′ = gh and "moving" vg,h from left to right ac- cording to the above multiplication by N rule, is equal to Ph Uhσ−1 h (vg−1h,h)ξg−1h. We also let (Ph Uhξh) · Ug = Ph vh,gUhgσ−1 g (ξh) which by similar considerations is equal to Ph Uhσ−1 g,g−1 and (Ugx)∗ = x∗U ∗ g,g−1σg(x∗). The ∗-algebra H0 has a trace given by τ (Ph Uhxh) = τN (xe) which recovers for elements in H0 the H-scalar product, i.e., if X, Y ∈ H0 then hX, Y iH = τ (Y ∗X). g (ξhg−1). Finally, we let U ∗ g = Ug−1 v∗ h (vhg−1,g)σ−1 g = Ug−1v∗ It is easy to verify that the left multiplication by Ug give unitary operators on H, the left multiplication by N = N Ue ⊂ H0 on H gives a representation of N as VANISHING COHOMOLOGY 7 a von Neumann algebra, and altogether left multiplication by elements in H0 are bounded operators on H that give a ∗-representation M0 of H0 in B(H), with the trace τ being implemented by the vector state given by 1 = Ue1 ∈ H. The crossed product von Neumann algebra N ⋊(σ,v) Γ is by definition the weak operator closure of M0 in B(H). It is a finite von Neumann algebra with faithful normal state τ (X) = hX1, 1iH, ∀X ∈ M . One clearly has a natural identification between the standard representation (or standard Hilbert M -bimodule) L2(M, τ ) and H. The cocycle action (σ, v) is free if σg is an outer automorphism, ∀g 6= e. It is well known (and immediate to check) that (σ, v) is free if and only if N ′∩N ⋊(σ,v)Γ = C1. So in this case M = N ⋊σ Γ is a II1 factor with the normalizer NM (N ) = {u ∈ U(M ) uN u∗ = N } of N in M generating M (i.e., with N regular in M ). Conversely, if N ⊂ M is a regular, irreducible inclusion of II1 factors and we denote Γ = NM (N )/U(N ), with Ug ∈ NM (N ), g ∈ Γ, a lifting of Γ, Ue = 1, and we denote σg = Ad(Ug)N , vg,h = UghU ∗ g , then (σ, v) is a free cocycle action of Γ on N , with N ⊂ N ⋊(σ,v) Γ naturally isomorphic to N ⊂ M (see e.g., [J80]). 1.2. Cocycle conjugacy of cocycle actions. The cocycle actions (σi, vi) of Γ on Ni, i = 1, 2, are cocycle conjugate if there exists an isomorphism θ : N1 ≃ N2 and a map w : Γ → U(N2) such that the following conditions are satisfied: h U ∗ (1.2.1) (1.2.2) θσ1(g) θ−1 = Ad(wg)σ2(g), ∀g. θ(v1(g, h)) = wgσ2(g)(wh)v2(g, h) w∗ gh, ∀g, h. The cocycle actions σ1, σ2 are outer conjugate (or weakly cocycle conjugate) if condition (1.2.1) is satisfied. Note that this is equivalent to σ1, σ2 composed with the quotient map Aut(N ) → Out(N ) being conjugate by an element in Out(N ). Similarly, two outer actions σ1, σ2 → Aut(N ) are outer conjugate, if there exists θ ∈ Aut(N ) such that σ1(g) = θσ2(g)θ−1 modulo Inn(N ), for all g ∈ Γ. The actions σ1, σ2 are conjugate if there exists an isomorphism θ : N1 ≃ N2 such that conditions (1.2.1) is satisfied with w = 1. We then write σ1 ∼ σ2. We recall here the following well known observation (see e.g., [J80]), which trans- lates cocycle conjugacy of free cocycle actions into the isomorphism of the associated crossed-product inclusions of factors. Proposition. Let (σi, vi) be a cocycle action of the discrete group Γi on the II1 factor Ni, i = 1, 2. If there exists a ∗-isomorphism Φ : N1⋊(σ1,v1)Γ1 ≃ N2⋊(σ2,v2)Γ2 such that Φ(N1) = N2, then σ1 and σ2 are cocycle conjugate. More precisely, there 8 SORIN POPA exists a group isomorphism γ : Γ1 → Γ2, and unitaries wg ∈ U(N2), for all g ∈ Γ1, such that: (i) Φσ1(g)Φ−1 = Adwg σ2(γ(g)), for all g ∈ Γ1, (ii) Φ(v1(g, h)) = wgσ2(γ(g)(wh)v2(γ(g), γ(h))w∗ Conversely, if Φ : N1 ≃ N2 is a ∗-isomorphism, γ : Γ1 ≃ Γ2 is a group isomor- phism, and there exist unitaries wg ∈ U(N2) for all g ∈ G1 such that (i), (ii) are satisfied, then Φ can be extended to an isomorphism N1 ⋊(σ1,v1) Γ1 ≃ N2 ⋊(σ2,v2) Γ2 (hence, to an isomorphism of the associated inclusions). gh, for all g, h ∈ Γ1. 1.3. 1-cocycles for actions. Assume σ is a genuine action of Γ on the II1 factor N . A map w : Γ → U(N ) satisfying condition (1.3.1) wgσg(wh) = wgh, ∀g, h is called a 1-cocycle for σ. Such a 1-cocycle for σ is a coboundary (or it is trivial) if there exists a unitary element v ∈ U(N ) such that wg = v∗σg(v), ∀g. (Clearly, such maps wg do satisfy the 1-cocycle condition (1.3.1)). The map w is called a weak 1-cocycle if it satisfies the relation (1.2.1) modulo the scalars, i.e., (1.3.1') wgσg(wh)w∗ gh ∈ T1, ∀g, h ∈ Γ Note that this is equivalent to Ad(wg)σg being an action. Note also that if w is a weak 1-cocycle then µg,h = wgσg(wh)w∗ gh is a scalar 2-cocycle for Γ, i.e., µ ∈ H 2(Γ). Also, cocycle conjugacy of two (genuine) actions σi : Γ → Aut(Ni), i = 1, 2, amounts to conjugacy of σ1 and σ′ 2(g) = Ad(wg)σ2(g), g ∈ Γ, for some 1-cocycle w for σ2. 2, where σ′ A (weak) 1-cocycle w is weakly trivial (or weak cobouboundary) if there exists a unitary element v ∈ U(N ) such that vwgσg(v)∗ ∈ T1, ∀g. Two (weak) 1-cocycles w, w′ of the action σ are equivalent if there exists a unitary element v ∈ N such that w′ g = vwgσg(v)∗, ∀g ∈ Γ (resp. modulo scalars). Thus, a weak 1-cocycle is weakly trivial iff it is equivalent to a scalar valued weak 1-cocycle (N.B.: these are just plain scalar functions on Γ). Note that the scalar valued genuine 1-cocycles are just characters of Γ. Two free actions σ1, σ2 of Γ on N are cocycle conjugate iff σ1 is conjugate to σ′ 2, where σ′ 2(g) = Ad(wg)σ2(g), ∀g ∈ Γ, for some 1-cocycle w for σ2. We also mention here a well known result from [J80], showing that any 1-cocycle of an action of a finite group Γ is co-boundary. This property is actually specific to finite groups: we use a result in [P01a] to deduce that if Γ is infinite, then there exist free ergodic actions Γ y R which admit non-trivial 1-cocycles. (N.B. In the particular case when Γ is amenable, this fact can be derived from [Oc] as well). VANISHING COHOMOLOGY 9 Proposition. 1◦ Let Γ yσ N be a free action of a finite group Γ on a II1 factor N . If w is a 1-cocyle for σ, then there exists u ∈ U(N ) such that wg = uσg(u∗), ∀g ∈ Γ. 1+λ , λ 2◦ Let (N0, ϕ0) be a copy of the 2 by 2 matrix algebra with the state given by weights { 1 1+λ }, for some 0 < λ < 1, and Γ be an infinite group. Let Γ y (N , ϕ) = ⊗g(N0, ϕ0)g be the Bernoulli Γ-action with base (N0, ϕ0). Let Γ yσ N = Nϕ ≃ R be the corresponding Connes-Størmer Bernoulli action. Let B ⊂ N be an atomic von Neumann subalgebra of the form ⊕nBn, with Bn ≃ Mkn×kn (C) having minimal projections of trace λmn , with m1 < m2 < ..... Then there exists a 1- cocycle w for σ such that σ′ g = Ad(wg)σg, g ∈ Γ, has B as its fixed point algebra. If B 6= C, then any such 1-cocycle is not a co-boundary. Proof. 1◦ This is (Corollary 2.16 in [CT76; see also [J80]). We include here the proof, for completeness, which is based on Connes well known "2 by 2 matrix trick". Thus, let σ be the action of Γ on N = M2×2(N ) = N ⊗ M2×2(C) given by If {eij 1 ≤ i, j ≤ 2} is a matrix unit for M2×2(C) ⊂ N , then σg = σg ⊗ id. If Q ⊂ N denotes the fixed point algebra wg = e11 + wge22 is a cocycle for σ. of the action σ′ g = Ad( wg)σ, then e11, e22 ∈ Q and the existence of a unitary element u ∈ N satisfying wg = uσg(u∗), ∀g, is equivalent to the fact that e11, e22 are equivalent projections in Q. But the fixed point algebra of any free action of a finite group on a II1 factor is a II1 factor. Thus, e11, e22 are equivalent in Q and w follows co-boundary. n }1≤j≤kn ∈ N be isometries such that V j = N and {V i ∗ ∈ N , ∗ 1 ≤ i, j ≤ kn} be the matrix units τ (V j of Bn. Let also πn be the trivial representation of Γ of multiplicity kn. Then by n)∗, (Theorem 3.2 in [P01a]), wg = Pn Pi,j πn(g)i,jV i g ∈ Γ, defines a 1-cocycle for σ and σ′ g = Ad(wg)σg has B as its fixed point algebra. n )∗ = Pn Pi V i nσg(V j nσg(V i 2◦ For each n ≥ 1, let {V j n V j n ) = λmn , V j n N V j n ∗ ∗ nV j n n V j n Since the fixed point algebra of an action is a conjugacy invariant of the action and σ is mixing (thus ergodic), it follows that σ, σ′ are not conjugate, in particular there exists no u ∈ U(N ) such that σ′ g = Ad(u)σgAd(u∗), ∀g, a relation that amounts to wg = uσg(u∗) modulo scalars, ∀g. (cid:3) 1.4. Vanishing cohomology and property VC. The 2-cocycle v for the cocycle action σ vanishes (or it is a coboundary) if there exists a map w : Γ → U(N ) such that we = 1 and v = ∂w, i.e.: (1.4.1) vg,h = (∂w)g,h def= σg(w∗ h)w∗ gwgh, ∀g, h ∈ Γ. The 2-cocycle v weakly-vanishes (or it is a weak coboundary) if there exists w : 10 SORIN POPA Γ → U(N ) such that we = 1 and v = ∂w modulo scalars, i.e.: (1.4.2) wgσg(wh)vg,hw∗ gh ∈ C1, ∀g, h ∈ Γ. Note that this is equivalent to (1.4.2′) (Ad wgσg) (Ad whσh) = Ad wghσgh, ∀g, h i.e., to σ′ g def= Ad wgσg being a "genuine" action. {wg}g ⊂ N such that U ′ U ′ gh, ∀g, h ∈ Γ). In turn, the vanishing of v is equivalent to the existence of unitary elements g = wgUg ∈ M = N ⋊(σ,v) Γ give a representation of Γ (i.e., h = U ′ gU ′ Given a II1 factor N , we denote by VC(N ) the class of countable groups Γ for which any free cocycle action (σ, v) of Γ on N has the property that the 2-cocycle v vanishes (or is coboundary) and by VCw(N ) the class of groups Γ for which any free cocycle action (σ, v) of Γ on N has the property that v is a weak-coboundary. We denote by VC (respectively VCw) the class of countable groups Γ with the property that Γ ∈ VC(N ) (resp. Γ ∈ VCw(N )) for any II1 factor N . If Γ ∈ VC then we also say that Γ has property VC or that it is a VC-group. We are especially interested in identifying the VC and VCw groups, i.e., groups that have the most "universal" vanishing cohomology property. Other classes of interest will be VC(N ) for N equal to the hyperfinite II1 factor R and for N equal to the free group factor L(F∞). This is because R and L(F∞) are the most interesting "non-commutative probability spaces". Also, any countable group Γ has "many" free actions on these factors, in fact both of them have a lot of generalized sym- metries (notably L(F∞), on which by [PS01] any "group-like" object admits free actions). Also, both factors admit many cocycle actions that are "hard to untwist" (cf. [CJ84], [P01a] and Section 3 and 4 below). (N.B. It should be noticed that by the way we have defined VC(N ), if a factor N has only inner automorphisms, i.e., Out(N ) = {1}, like the examples in [IPeP05], then any Γ belongs to VC(N )!) We'll now show that the class VC is closed to amalgamated free products over finite subgroups and that vanishing cohomology for cocycle actions of countable groups is essentially a "separability" property: 1.5. Proposition. 1◦ if {Γn}n≥0 ⊂ VC(N ) (respectively VCw(N )) for some II1 factor N and K ⊂ Γn is a common finite subgroup, n ≥ 0, then Γ0 ∗K Γ1 ∗K Γ2 ∗K ... ∈ VC(N ) (respectively VCw(N )). Also, if {Γn}n ⊂ VC (resp. VCw), then Γ0 ∗K Γ1 ∗K .... ∈ VC (resp. VCw). 2◦ Let N be a II1 factor, Γ ⊂ Out(N ) a countable group with a lifting {σg}g∈Γ ⊂ Aut(N ) and denote vg,h ∈ U(N ) a set of unitaries satisfying Ad(vg,h) = σgσhσ−1 gh , VANISHING COHOMOLOGY 11 g, h ∈ Γ. There exists a separable II1 subfactor Q ⊂ N that contains the countable set {vg,h g, h ∈ Γ} and is normalized by σ, with σg outer, ∀g ∈ Γ. 3◦ VC (respectively VCw) coincides with the class of countable groups Γ with the property that Γ ∈ VC(N ) (resp. Γ ∈ VCw(N )) for any separable II1 factor N . Proof. 1◦ Assume Γn ∈ VC(N ). Let (σ, v) be a free cocycle action of G = ∗KΓn on N and denote M = N ⋊σ G with Ug, g ∈ G the corresponding canonical unitaries. Since Γn ∈ VC(N ), there exist unitaries {wn g = wn gU 0 g , g ∈ Γ0, we may assume w0 g g ∈ Γn} in N such that U n g Ug, g ∈ Γn, give left regular representations of Γn. Replacing Ug by w0 g = 1, ∀g ∈ Γ0. k = wn But then for each n ≥ 1, U n k Uk, k ∈ K, for some 1-cocyles wn : K → U(N ) for the restriction to K of the Γn-action σn implemented by U n g , g ∈ Γn. By ([J80]; see Proposition 1.3 above) any 1-cocycle of a free action of a finite group vanishes. Hence, there exists vn ∈ U(N ) (with v0 = 1) such that wn n), n, k ∈ K. But then the unitaries {v∗ equivalently U n g σg(vn) g ∈ Γn, n ≥ 0} generate inside M a copy of the left regular representation of G = ∗K Γn implementing a G-action on N that gives an inner perturbation of the initial cocycle G-action σ. k = vnσn(k)(v∗ nU n k = vnUkv∗ 2◦ We construct recursively an increasing sequence of separable von Neumann subalgebras Qn, n ≥ 0, such that Q0 ⊃ {vg,h g, h ∈ Γ} and for each m ≥ 1 we have m∩N (x) = τ (x)1, ∀x ∈ (Qm−1)1; m∩N ⋊Γ(Ug) = 0, ∀g 6= e; (a) EQ′ (b) EQ′ (c) Qm ⊃ ∪gσg(Qm−1), where Ug ∈ N ⋊ Γ are the canonical unitaries implementing σ. Assume we have constructed these algebras up to m = n. Since N ′ ∩ N ⋊ Γ = C, by using (Theorem 0.1 in [P13]) we can get a Haar unitary v = (vk)k ∈ N ω that's free independent to Qn−1 ∪ {Ug}g. Thus, if we take Q0 n to be the von Neumann algebra generated by Qn−1 and {vk}k, then we already have (a) and (b) satisfied for Qn = Q0 n by the von Neumann algebra generated by Qn = ∪gσg(Q0 n, and then we can replace this "initial" Q0 n), to have (c) satisfied as well. w Finally, if we define Q = ∪nQn , then Q is clearly separable, condition (c) insures that Q is normalized by σ, condition (a) shows that EQ′∩N (∪nQn) ∈ C1, implying that Q is a factor, while condition (b) shows that σg is outer on Q, ∀g 6= e. 3◦ This part is now trivial by 2◦. 1.6. Remarks 1◦ As we will see in Sections 3 and 4, it is in general not true that if Γi are in VC then their amalgamated free product over a common (infinite) (cid:3) 12 SORIN POPA amenable subgroup H ⊂ Γi, Γ = Γ1 ∗H Γ2 is in VC. For instance, Z2 ⋊ SL(2, Z) does not even belong to VCw(R) (see Theorem 3.2). 2◦ The classes VC may satisfy other general permanence properties. For instance, it may be true that Γ ∈ VC implies Γ0 ∈ VC for any subgroup Γ0 ⊂ Γ (or at least when [Γ : Γ0] < ∞). However, the obvious idea for a proof, which is to "co-induce" a given cocycle action Γ0 yσ0 N0 to a set of automorphisms {σg g ∈ Γ} on N = N ⊗Γ/Γ0 doesn't work when [Γ : Γ0] = ∞, because an infinite tensor product of inner automorphisms may become outer, so the σg's may in fact not give a cocycle action of Γ. When the index of Γ0 in Γ is finite, then σ defined this way does give a cocycle action of Γ on N , but it is not immediate of how to "bring down to N0" the vanishing of the cohomology for σ to the vanishing of the cohomology for the initial Γ0 yσ0 N0. 0 2. Groups with the property VC In this section we prove a vanishing cohomology result for arbitrary free cocycle actions of countable amenable groups on arbitrary II1 factors. To do this, we'll first show that any amenable subgroup Γ ⊂ Aut(N )/Inn(N ) can be lifted to a set {σg g ∈ Γ} ⊂ Aut(N ) normalizing a "large" hyperfinite subfactor of N (see Theorem 2.1). As it happens, this property, which is interesting by itself, characterizes the amenability of the group Γ. Indeed, we will show in Section 5 that any non-amenable group admits a free action on N = L(F∞) that cannot be perturbed to a cocycle action that normalizes a hyperfinite subfactor of N . Once we prove that any cocycle action σ of an amenable group Γ on N normalizes (modulo inner perturbation) a hyperfinite II1 factor R ⊂ N , we reduce the vanishing cohomology problem to the case N = R, where by a well known result of Ocneanu [Oc85] free cocycle actions of amenable groups can indeed be "untwisted" to genuine actions. The fact that R is "large in N " assures that by untwisting σ on R we have untwisted it as an action on N as well. To show that σ normalizes up to Inn(N ) a "large hyperfinite subfactor of N ", we reduce the problem to a statement about commuting squares of subfactors, as fol- lows. Assume Γ is generated by a finite set e ∈ F = F −1 ⊂ Γ and consider the locally trivial subfactor obtained by the diagonal embedding N σ,F := {Pg∈F σg(x)egg x ∈ N } ⊂ MF ×F (N ) =: M σ,F , where {egh}g,h∈F ⊂ MF ×F (C) are the matrix If Q ⊂ R is an inclusion of factors with Q ⊂ N σ,F , units (see 5.1.5 in [P91]). R ⊂ M σ,F , egg ∈ R, ∀g ∈ F , and (Q ⊂ R) ⊂ (N σ,F ⊂ M σ,F ) makes a non- degenerate commuting square, then Q ⊂ R is itself locally trivial and there exist unitary elements wg ∈ N such that wgeeg ∈ R. If one denotes Q0 ⊂ N the image of Q under the isomorphism N σ,F ≃ N , then this amounts to Q0 being invariant VANISHING COHOMOLOGY 13 g = Adwg ◦ σg, ∀g ∈ F . Moreover, if Q′ ∩ R = N σ,F ′ to σ′ iff σg is outer, ∀g ∈ F . This observation applied to N σ,F ⊂ M σ,F are the factors in the tower for N σ,F ⊂ M σ,F ), in combination with (5.1.5 in [P91]), shows that if all the higher relative commutants in the Jones towers for Q ⊂ R and N σ,F ⊂ M σ,F coincide, then σ′ implements an outer action of Γ on Q0. ∩ R, then σ′ (where M σ,F n n is outer gQ0 So all we need to do is to produce an inclusion of hyperfinite factors Q ⊂ R inside N σ,F ⊂ M σ,F , making a commuting square and having same higher relative commutants. We will solve this commuting square problem by only using that N σ,F ⊂ M σ,F has amenable graph. Thus, we will in fact prove that any finite index inclusion of separable II1 factors N ⊂ M with amenable standard invariant GN⊂M contains an inclusion of hyperfinite factors (Q ⊂ R) ⊂ (N ⊂ M ), that makes a non-degenerate commuting square and has identical higher relative commutants in the associated Jones tower (in particular same standard invariant), in fact even satisfies the strong smoothness condition Q′ ∩ Rn = Q′ ∩ Mn = N ′ ∩ Mn, ∀n (see Theorem 2.10 below). We'll obtain Q ⊂ R as an inductive limit of relative commutants P ′ n ∩ M of a decreasing sequence of finite index subfactors Pn ⊂ N that come from repeated downward basic constructions of subfactors M p′ ⊂ p′Mnp′ obtained by appropriate induction/reduction in the Jones tower N ⊂ M ⊂ M1 ⊂ ..., with choices "dictated" by the local characterization of the amenability of ΓN⊂M in ([P97], Theorem 6.1). n ∩ N ⊂ P ′ 2.1. Theorem. Let N be a II1 factor and σ : Γ → Aut(N ) an outer action of a countable amenable group Γ on N , with H 3(Γ)-obstruction νσ and with vg,h ∈ U(N ) satisfying σgσh = Ad(vg,h)σgh, ∀g, h ∈ Γ. Then there exist {wg}g ⊂ U(N ) and a hyperfinite subfactor R ⊂ N such that if we denote σ′ g = Ad(wg)σg and g,h = wgσg(wh)vg,hw∗ v′ gh, g, h ∈ Γ, then we have: g(R) = R and v′ g,h ∈ R, ∀g, h ∈ Γ. (2.1.1) σ′ (2.1.2) {σ′ gR }g is an outer action of Γ on R with same H 3(Γ)-obstruction as σ. If in addition (σ, v) is a free cocycle action of Γ on N , then {wg}g can be chosen R, v′) is a free cocycle action of Γ on R. Moreover, if N is separable, so that (σ′ then one can choose σ′, v′, R so that to also have R′ ∩ N = C. Let us show right away how Theorem 2.1 combined with Ocneanu's Theorem in [Oc85] can be used to derive the vanishing cohomology result for cocycle actions of arbitrary amenable groups: 14 SORIN POPA 2.2. Theorem. Let N be a II1 factor, Γ a countable amenable group and (σ, v) a free cocycle action of Γ on N . Then there exist unitary elements {wg ∈ U(N ) g ∈ Γ} such that σ′ g = Ad(wg) ◦ σg, g ∈ Γ, is a genuine action of Γ on N , in fact such that moreover we have vg,h = σg(w∗ gwgh, ∀g, h ∈ Γ. h)w∗ Proof of 2.2. By Theorem 2.1, there exist unitary elements w0 hyperfinite II1 subfactor R ⊂ N such that: g ∈ N , g ∈ Γ, and a (2.2.1) σ0 (2.2.2) v0 (2.2.3) σ0 g := Ad(w0 g,h := wgσg(wh)vg,hw∗ gR is outer, ∀g 6= e. g)σg leaves R invariant, ∀g ∈ Γ; gh belongs to R, ∀g, h ∈ Γ. But then, (σ0 R, v0) implements a free cocycle action of the countable amenable group Γ on R, so by Ocneanu's Theorem [Oc85] the 2-cocycle v0 is co-boundary on R, i.e., there exist unitary elements w1 gh, ∀g, h ∈ Γ. This shows that wg = w1 (cid:3) g,h = σ0 g satisfy the required condition. g ∈ R such that v0 ∗)w1 g g(w1 h ∗w1 gw0 2.3. Corollary. The class VC contains all countable amenable groups and is closed to free products with amalgamation over finite subgroups, i.e., if {Γn}n ⊂ VC and K ⊂ Γn is a common finite subgroup, then Γ0 ∗K Γ1 ∗K .... ∈ VC. For the rest of this section, we will use concepts and notations from [J83] (such as the basic construction, the Jones tower of factors, etc), as well as from ([PiP84], [P91], [P93], [94a], [94b], [P97]). In particular, we will often use as framework the symmetric enveloping (abbreviated SE) inclusion M ⊗M of N ⊂ M , M op op ⊂ M ⊠ eN introduced in [P94b] (cf. also the extended version of the paper, [P97]). We begin by recalling some properties relating Jones tower/tunnel of a subfactor with its symmetric enveloping inclusion. It will be useful for the reader to keep in mind that if M ⊂ M1 ⊂ ... is the Jones tower of factors associated with a subfactor of finite index N ⊂ M , then M L2(Mn)M =M L2(M1)⊗M ....⊗M L2(M1)M (n-times tensor/M product). Also, if one denotes by {Hk}k∈K the list of irreducible Hilbert M -bimodules appearing in {L2(Mn)}n, then for any m ≥ 1 and any nonzero 1 ∩ Mm we have M L2(M1)M ⊂M L2(p′Mmp′)M , and thus we projection p′ ∈ M ′ recover all {Hk}k in the tensor powers M L2(p′Mmp′)n⊗M M , n ≥ 1. Equivalently, if M ֒→ p′Mmp′ = M 0 and M 0 ⊂ M 0 1 ⊂ ... is its Jones tower, then {Hk}k coincides with the list of irreducible Hilbert M -bimodules appearing in L2(M 0 n), n ≥ 1 (see [P91] and [Bi97] for basics of subfactor theory). 2.4. Lemma. Let N ⊂ M be an extremal inclusion of II1 factors of index [M : N ] = λ−1 < ∞, T = M ⊗M = S its SE inclusion of II1 factors M op op ⊂ M ⊠ eN VANISHING COHOMOLOGY 15 and ... ⊂ Nm ⊂ ... ⊂ N ⊂e−1 M ⊂eN =e0 M1 ⊂e1 ...Mm ⊂em .... a tunnel-tower for N ⊂ M inside S. 1◦ If en −n ∈ Mn+1 ⊂ S denotes the projection of trace [M : N ]−n−1 obtained −n −n) = S. Thus, the map that acts as −n gives a natural identification between as a scalar multiple of the word of maximal length in e−n, ..., e0, ...., en, then en op implements EM Nn , EM N op n the identity on M ∨M M ∨ M M and one has vN (T, en and takes eNn to en and T ⊂ S. op op op ⊂ M ⊠ eNn 2◦ Let p ∈ P(N ′ op op n ∩ M ) and p of S. Then Nn+1pp tiautomorphism sic construction for (V ⊂ U ) = (Nn+1pp f = τ (p)−1pp = τ (p)−1pen op op en −npp op op ⊂ pM pp ∈ M ′ ∩ Mn+1 ⊂ S its image under the an- is a ba- ), with Jones projection . Moreover U ⊠ is eV ⊂ pM pp f p −np = τ (p)−1p Mn+1pp ⊂ pp U op op op op op op op op op op op M naturally embedded into S as the von Neumann subalgebra generated by pM pp pp p equal to pp and the amplification by τ (p)2 of (T ⊂ S), with eV 7→ f . , n ∩ N , then this latter algebra is actually ) and f . If in addition p ∈ N ′ Spp , thus giving a natural identification between (U ⊗U ⊂ U ⊠ eV U op op op op 3◦ Let p ∈ P(N ′ n ∩ N ) be as in the last part of 2◦. Let P ⊂ N be a subfactor op = T ⊂ S1 = M ⊠ eP such that P ⊂ M is a downward basic construction for M ≃ M p and denote M ⊗M an orthonormal basis of N over P , then e = Pj mjeP m∗ is a projection of trace λ = [M : N ]−1 in S1 that implements both EM N op and satisfies vN (T, e) = S1, thus giving an identification between T ⊂ S and T ⊂ S1, with eN 7→ e. op eP m j op N and EM its SE inclusion. If {mj}j ⊂ N is j = Pj m Mn+1p ⊂ p op j M op ∗ op op op Proof. Part 1◦ is (Proposition 2.9(a) in [P97]), the first part of 2◦ is (Lemma 2.8.(c) and 2.9(c) in [P97]), while the first part of 3◦ is (Proposition 2.10 in [P97]). To prove the last part of 2◦, note that with the notation U = pM pp op op op op op U Spp ⊂ pp k⊗H′ we have U L2(S0)U ⊂ U L2(pp S0 = U ⊠ eV bimodules. Then notice that by (Theorem 4.5 in [P97]), U L2(pp )U = ⊕k∈KH′ of all distinct ir- reducible Hilbert M -bimodules in ∪nL2(Mn). Since the list of irreducible U - bimodules in the Jones tower for Nnp ⊂ pM p contains all {H′ k}k∈K (because p ∈ N ′ )U ⊂ U L2(S0)U as well. Thus, the inclusion U L2(S0)U ⊂ U L2(pp k}k∈K denotes the reduction by pp )U is an equality, forcing S0 = S as well. n ∩ N ), it follows that U L2(pp k , where {H′ Spp Spp Spp Spp op op op op op op op and )U as Hilbert op op The last part of 3◦ follows by taking again into account (Lemma 2.8. (c) in [P97]), part 2◦ above and the remark before the statement of the lemma, then using the 16 SORIN POPA same "exhaustion by bimodules" argument used above (based on Theorem 4.5 in [P97]). (cid:3) 2.5. Definition. Let N ⊂ M be an inclusion of II1 factors with finite index and denote by N ⊂ M ⊂ M1 ⊂ M2 ⊂ ... its Jones tower. A subfactor P in N is said to be (N ⊂ M )-compatible if there exist n ≥ 1 and a non-zero projection p′ ∈ M ′ 1 ∩Mn such that P p′ ⊂ M p′ ⊂ p′Mnp′ is a basic construction. Let us note right away that this property is in some sense "hereditary": 2.6. Lemma. With N ⊂ M as above, assume P ⊂ N is (N ⊂ M )-compatible. If a subfactor Q ⊂ P is (P ⊂ M )-compatible, then Q ⊂ N is (N ⊂ M )-compatible. Proof. Let P ⊂ N ⊂ M ≃ M p′ ⊂ M1p′ ⊂ p′Mnp′ be a basic construction, for some n ≥ 1 and a non-zero projection p′ ∈ M ′ 1 ∩ Mn. Note that if we denote V = M p′ ⊂ p′Mnp′ = U then, given any m ≥ 1, its associated Jones tower of factors up to m, V1 ≃ P ⊂ M ≃ V ⊂ U ⊂ U1 ⊂ ...Um, can be realized (up to isomorphism) by inducing/reducing in the initial tower M ⊂ M1 ⊂ M2... ⊂ Mk, for some large enough k, with the projections involved p′ ij ∩ Mij+1 , where i0 = p′, i1 = n, and i0 < i1 < .... This shows that if Q ⊂ P is (P ⊂ M )- i0 = 1, p′ compatible, then one obtains a basic construction Q ⊂ M ≃ M q′ ⊂ q′Mk0 q′, for some appropriate k0 and q′ ∈ M ′ ij . But this means that Q ⊂ N is (N ⊂ M )-compatible. (cid:3) 1 ∩ Mk0 obtained as a product of such p′ ij ∈ M ′ For the reader's convenience, we recall here two of the equivalent definitions of amenability for "group-like" objects arising from subfactors, that we have intro- duced and studied in ([P91], [P93], [P94b], [P97]), and that we need hereafter. The standard invariant GN⊂M of an extremal inclusion of factors with finite Jones index N ⊂ M is amenable if its principal graph ΓN⊂M satisfies the Kesten- type condition kΓN⊂M k2 = [M : N ]. This very first definition of amenability was proposed in [P91], and we will also take it to be the definition of amenability for the various abstractizations of standard invariants: a standard λ-lattice G as in [P04b] (or a planar algebra as in [J99]) is amenable if its graph ΓG satisfies the condition kΓGk2 = λ−1. An alternative notion of amenability was introduced in [P93], by requiring the following Følner-type condition on G = GN⊂M : let {dk}k∈K denote the standard weights of ΓN⊂M (resp. ΓG ), obtained for instance as dim(M HkM )1/2, where {Hk}k∈K is the list of irreducible M − M Hilbert bimodules appearing at even levels in G, indexed by the set K of left vertices of the bipartite graph ΓG = ΓN⊂M ; G satisfies the Følner condition if for any ε > 0 there exists a finite set F ⊂ K such that if one denotes by ∂F = {k ∈ K \ F ∃k0 ∈ F with (ΓGΓt G)kk0 6= 0} (the boundary of F ) then P k∈∂F d2 k < ε P k∈F d2 k. VANISHING COHOMOLOGY 17 These two conditions were shown equivalent in ([P97] Theorem 5.3; the result had already been announced in [P93] and [P94b]). Several other equivalent amenabil- ity conditions were in fact introduced and studied in [P97], notably a local finite dimensional approximation property ([P97], Theorem 6.1) which will be crucial for the proof of Theorem 2.10 below. The Kesten and the Følner-type conditions provide in particular equivalent def- initions of amenability for a finitely generated rigid C∗-tensor category of Hilbert bimodules G (as defined for instance in [PV14]), having {Hk}k∈K as irreducible morphisms. One then defines amenability for an arbitrary (possibly infinitely gen- erated) rigid C∗-tensor category by requiring that any finitely generated subcate- gory is amenable (see the detailed definitions in the paragraphs preceding Theorem 2.12). Due to its various equivalent characterizations, amenability in this context is ⊂ = S associated with an extremal inclusion N ⊂ M coincides with the a very "robust" property. For instance, since the SE inclusion T = M ⊗M M ⊠ eN M op op op op M ⊂ M ⊠ eNn associated with Nn ⊂ M , for any n ≥ 0 SE inclusion M ⊗M (e.g., by Lemma 2.4.1◦ above), it follows from (Theorem 5.3 in [P97]) that GN⊂M is amenable iff GMi⊂Mj is amenable for some (and thus all) i < j. Note that this can also be deduced from the fact that ΓMi⊂Mj is a alternate product of ΓN⊂M and its transpose, or of ΓM⊂M1 and its transpose, j − i times, which shows that one always have kΓMi⊂Mj k = kΓN⊂M kj−i (cf. 1.3.5 in [P91]), implying that kΓN⊂M k2 = [M : N ] iff kΓMi⊂Mj k2 = [Mj : Mi]. Moreover, if p′ is a non-zero projection in M ′ i ∩ Mj, for some i < j in Z, then by (Coroally 6.6 (ii) in [P97]) ΓMi⊂Mj amenable (which we already know is equivalent to ΓN⊂M being amenable) implies V = Mip′ ⊂ p′Mjp′ = U has amenable graph as well. Also, note that if j ≥ i + 1 and p′ ∈ M ′ i+1 ∩ Mj, then by Lemma 2.4 and the above argument, one conversely has that V ⊂ U amenable implies N ⊂ M amenable. We record all these facts in the following: 2.7. Proposition. Let N ⊂ M be an extremal inclusion of factors with finite index and {Mi}i∈Z be a tunnel/tower of factors for N ⊂ M . If GN⊂M is amenable, then for any i < j in Z and any non-zero projection p′ ∈ M ′ i ∩ Mj, the inclusion Mip′ ⊂ p′Mjp′ has amenable standard invariant. If in addition j ≥ i + 1 and p′ ∈ M ′ i+1 ∩ Mj, then conversely GMip′⊂p′Mj p′ amenable implies GN⊂M amenable. 2.8. Lemma. Let N ⊂ M be a finite index extremal inclusion of II1 factors with . Given any (N ⊂ M )- amenable standard invariant and SE factor S = M ⊠ eN M op 18 SORIN POPA compatible subfactor P ⊂ N and any ε > 0, there exists a (P ⊂ M )-compatible subfactor Q ⊂ P such that kE(Q′∩M )′∩S(x) − EM ′∩S(x)k2 ≤ εkxk, ∀x ∈ P ′ ∩ S. Proof. Let us first note a few Facts needed in the proof. Fact 1. It is sufficient to show that there exists a compatible subfactor Q ⊂ P with the property that kEQ′∩M )′∩S(f ) − τ (f )1k2 ≤ ε/([M : P ] + 1)5/2, where f ∈ S is the Jones projection for P ⊂ M ⊂ hM, P i, viewed inside S (cf. 2.4.3◦ above). Indeed, because if {mj}j ⊂ M is an orthonormal basis of M over P with [M : j f }j is , f i is of P ]+1 many elements of norm ≤ [M : P ]1/2 (cf. [PiP84]), then {[M : P ]1/2m an orthonormal basis of P ′∩S over M ′∩S = M the form x = Pj[M : P ]1/2m j , where y has operator norm majorized by [M : P ]1/2kxkkmjf k = [M : P ]1/2kxk, thus giving the estimates and any x ∈ P ′∩S = hM ∗ x) ∈ M op op j = [M : P ]1/2EM op (f m j op op op j f y op op op kE(Q′∩M )′∩S(x) − EM ′∩S(x)k2 = k X [M : P ]1/2m op j (E(Q′∩M )′∩S(f ) − τ (f )1)y op j k2 j ≤ [M : P ]1/2 X kmjkky op j kkE(Q′∩M )′∩S(f ) − τ (f )1k2 j ≤ [M : P ]3/2([M : P ] + 1)kxkε/([M : P ] + 1)5/2 < εkxk. Fact 2. By Proposition 2.7 above, GN⊂M amenable implies GP ⊂M amenable. Fact 3. By (Theorem 6.1 in [P97]), if P ⊂ M is an extremal inclusion of factors with amenable standard invariant then for any δ > 0 there exists n ≥ 1 and a projection p ∈ P ′ 2 < δτ (p), where ... ⊂ Pn ⊂ Pn−1 ⊂ ...P ⊂ M ⊂f0=eP hM, P i denotes a Jones tunnel and basic construction for P ⊂ M . n∩M )p′∩phM,P ip(f0p) − τ (f0)pk2 n ∩ P such that kEp(P ′ Let us now proceed with the proof of the lemma. By Fact 2, P ⊂ M is amenable so we can apply Fact 3 to (P ⊂ M ⊂f0 hM, P i), to get an n ≥ 1 and a projection p ∈ P ′ n ∩ P such that (2.8.1) kEp(P ′ n∩M )p′∩phM,P ip(f0p) − τ (f0)pk2 < εkpk2/([M : P ] + 1)2 By amplifying by α = τ (p)−1 the inclusions of factors (2.8.2) Pnp ⊂ pP p ⊂ pM p ⊂f0p phM, P ip VANISHING COHOMOLOGY 19 using partial isometries in P , we obtain inclusions of factors (2.8.3) (Pnp)α = Q ⊂ P ⊂ M ⊂f0 hM, P i having same relative commutants as (2.8.2). Thus, by (2.8.1), it follows that Q satisfies (2.8.4) kEQ′∩hM,P i(f0) − τ (f0)k2 < ε/([M : P ] + 1)2, By the way it is defined, Q ⊂ P is an (P ⊂ M )-compatible subfactor, and thus (N ⊂ M )-compatible as well, while by Fact 1 and (2.8.4) we have kE(Q′∩M )′∩S(x)− EM ′∩S(x)k2 ≤ εkxk, for all x ∈ P ′ ∩ S. (cid:3) 2.9. Lemma. Let N ⊂ M be a finite index extremal inclusion of II1 factors with N ⊂ M ⊂ M1.... ր M∞ its Jones tower of factors and S its SE factor. If B ⊂ M is a diffuse von Neumann subalgebra, then B 6≺M∞ M ′ ∩ M∞ and B 6≺S M Proof. By [P03], in order to prove B 6≺M∞ M ′ ∩ M∞, it is sufficient to prove that for any finite set F in a given total subset X of M∞, there exist un ∈ U(B) such that limn kEM ′∩M∞(y∗unx)k2 = 0, ∀x, y ∈ F . Taking X = ∪mMm, it is sufficient to show this for any m and any finite F ⊂ Mm. But by [P03] this amounts to M 6≺ M ′ ∩Mm, which is trivial since B is diffuse and M ′ ∩Mm is finite dimensional. , which is total in S by (Section 4 in [P97]). Thus, if F ⊂ X is finite then we 2 ∈ F , with may assume F ⊂ MmM x1, y1 ∈ Mm, x2, y2 ∈ M , and we take un ∈ U(B), then we get the estimate we use the same criterion, but with X = ∪mMmM for some large m so if x = x1x To prove B 6≺S M op . op op 2 , y = y1y op op op kEM op (y ∗ op 2 y1 ∗unx1x op 2 )k2 ≤ kx2kky2kkEM ′∩Mm(y∗ 1unx1)k2. This shows that it is actually sufficient to check the criterion for F ⊂ Mm, which (cid:3) amounts again to M 6≺ M ′ ∩ Mm as before. 2.10. Theorem. Let N ⊂ M be a finite index extremal inclusion of separable II1 factors with amenable standard invariant. There exists a sub-inclusion of hyperfinite factors (Q ⊂ R) ⊂ (N ⊂ M ) making a non-degenerate commuting square that's strongly smooth, i.e., Q′ ∩ Rn = Q′ ∩ Mn = N ′ ∩ Mn and R′ ∩ Rn = R′ ∩ Mn = M ′ ∩ Mn, ∀n, where N ⊂ M = M0 ⊂e0 M1 ⊂ ... is the Jones tower for N ⊂ M and Rn = vN (R, e0, ..., en−1), n ≥ 1, the tower for Q ⊂ R. Moreover, Q ⊂ R can be obtained as Q = ∪nP ′ n ∩ M = R, for some decreasing sequence of (N ⊂ M )-compatible subfactors M ⊃ N ⊃ P0 ⊃ P1.... n ∩ N ⊂ ∪nP ′ 20 SORIN POPA Proof. We split this proof into several parts. Fact 1. There exists a sequence of (N ⊂ M )-compatible subfactors .... ⊂ Pn ⊂ ...P0 ⊂ N ⊂ M such that if we define Q = ∪mP ′ m ∩ M = R and Rn = ∪mP ′ m ∩ Mn, n ≥ 1, then (Q ⊂ R) ⊂ (N ⊂ M ) is a non degenerate commuting square of II1 factors, Q ⊂ R ⊂e0 R1 ⊂e1 R2... is its Jones tower and Q′ ∩ Rn = N ′ ∩ Mn, R′ ∩ Rn = M ′ ∩ Mn, ∀n. m ∩ N ⊂ ∪mP ′ To see this, let M ∨ M op ⊂ M ⊠ eN op M = S be the SE inclusion of factors associated with N ⊂ M . By applying recursively Lemma 2.8, we obtain a sequence of subfactors M ⊃ N = P0 ⊃ P1... such that for each n ≥ 1 we have op n−1 ∩ S. n ∩ N , R = ∪nP ′ n ∩ M . If we denote S0 = ∪nP ′ (a) Pn ⊂ Pn−1 is (Pn−1 ⊂ M )-compatible (thus also (N ⊂ M )-compatible). (b) kE(P ′ Let Q = ∪nP ′ n∩M )′∩S(x) − EM ′∩S(x)k2 ≤ 2−nkxk, ∀x ∈ P ′ n ∩ S , then by property (b) above, it follows that R′∩S0 = M . In particular, R is a II1 factor. By the definitions of Q, R, it follows that (Q ⊂ R) ⊂ (N ⊂ M ) is a commuting square, with e0 = eN implementing the conditional expectation of R onto Q and Q = {e0}′ ∩ R. From (b), one also gets ER′∩S(e0) = λ1. This implies that the algebra R0 1 := spRe0R has support 1 and thus any orthonormal basis {mj}j of R over Q must "fill up the identity", i.e., Pj mje0m∗ j = 1. Hence, (Q ⊂ R) ⊂ (N ⊂ M ) is in ′ ∩ S0 = {e0}′ ∩ M fact a non-degenerate commuting square. Moreover, R0 , 1 implying that R0 1e0 is a II1 factor as well, and R0 n := vN (R, e0, ..., en−1), n ≥ 1, is the Jones tower for Q ⊂ R. 1 is a II1 factor. Thus, Q ≃ Qe0 = e0R0 = N At the same time, if for each n ≥ 1 we define Rn = ∪mP ′ m ∩ Mn, then both this 1 ⊂ ... make (non-degenerate) commuting n = Rn and Rn, n−1 be each other's commutant in S0, ∀n ≥ 1. Note that this also implies that sequence and the sequence Q ⊂ R ⊂ R0 squares with N ⊂ M ⊂ M1 ⊂ ..., with R0 N for the downward continuation of these towers we have Q′ ∩ S0 = M op 1 . n ⊂ Rn. This forces R0 op op op So for the higher relative commutants, we have the equalities Q′ ∩ Rn = (Q′ ∩ S0) ∩ Rn = (M op 1 ∩ Mn) ∩ Rn = (N ′ ∩ Mn) ∩ Rn ⊂ Q′ ∩ Rn = M finishing the proof of Fact 1. op 1 ∩ Mn ∩ (N op n−1)′ = N ′ ∩ Mn, Fact 2. Assume ... ⊂ P1 ⊂ P0 ⊂ N ⊂ M are as in Fact 1. If un ∈ U(Pn), n ≥ 0, n is an (N ⊂ M )-compatible subfactor 0 ⊂ N ⊂ M is a sequence of factors still and we define P n in P n−1 satisfying the conditions in the statement of Fact 1. n...u∗ n−1 ⊂ ...P 1 0, then P n 1 ⊂ P 0 n = u0...unPnu∗ n−1 and ...P n n ⊂ P n−1 VANISHING COHOMOLOGY 21 ′ ∩ N ⊂ P n n n ∩ Mk}n, {P n n Indeed, for each k the systems of commuting squares of algebras {P ′ n ∩ N ⊂ ′ ∩ Mk}n (standard λ-lattices P ′ n ∩ M ⊂ ...P ′ n....u∗ in the sense of [P94a]) are isomorphic via the map Φ(x) = limn u0...unxu∗ 0, n ∩ Mk. Thus, Φ implements an isomorphism between R′ ∩ Ri ⊂ Q′ ∩ Ri x ∈ ∪nP ′ and R0′ ∩ R0 ′ ∩ Mi, i ≤ n. Since the isomorphism Φ leaves N ′ ∩ Mi = Q′ ∩ Ri and M ′ ∩ Mi = R′ ∩ Ri fixed, by equality of dimensions via Φ it follows that N ′ ∩ Mi = Q′ ∩ Ri, M ′ ∩ Mi = R′ ∩ Ri, ∀i. i ⊂ Q0′ ∩ R0 ′ ∩ M ⊂ ...P n n i , where R0 i = ∪nP n n Fact 3. Assume ... ⊂ P1 ⊂ P0 ⊂ N ⊂ M are as in Fact 1. Then there exist integers k0 = 0 < k1 < ... and unitaries vn ∈ U(Pkn−1 ), n ≥ 0, such that if we ′ ∩ M = R, then define P n kn Q′ ∩ Mm = Q′ ∩ Rm, R′ ∩ Mm = R′ ∩ Rm, ∀m. 1 and let Q = ∪nP n kn ′ ∩ N ⊂ ∪nP n kn = v1...vnPkn v∗ n...v∗ To show this, let {bk}k ⊂ (∪nMn)1 be a k k2-dense sequence. We choose recur- sively km > km−1, vm ∈ U(Pkm−1 ) such that (F3) kE(P m km ′∩P m−1 km−1 )′∩Mm (bj) − EP m−1 km−1 ′ ∩Mm (bj)k2 ≤ 2−m, ∀j ≤ m. Assume we made this construction up to m = n. Due to Lemma 2.9, (Theorem 0.1 (a) in [P13]) implies that if A0 ⊂ P n is a finite dimensional abelian von Neumann kn subalgebra with all minimal projections of sufficiently small trace, then there exists ′∩Mn+1 (bj)k2 < 2−n−1, ∀j ≤ n + 1. u ∈ U(P n kn For each m ≥ kn denote P n 0. Since B = ∪jP n is diffuse j (because it contains a Jones sequence of λ = [M : N ]−1 projections, which generate a copy of the hyperfinite II1 factor by [J83]), we may assume A0 ⊂ B, and hence ) such that kEuA0u′∩Mn+1 (bj) − EP n n...v∗ m = v0...vnPmv∗ ′ ∩ P n kn kn kEuBu′∩Mn+1(x) − EP n kn ′∩Mn+1 (bj)k2 < kEuA0u′∩Mn+1(bj) − EP n kn ′∩Mn+1 (bj)k2 < 2−n−1. Since B is a "limit" of P n j ′ ∩ P n kn , for j sufficiently large we'll still have kE(uP n j u∗ ′∩P n kn )′∩Mn+1 (bj) − EP n kn ′∩Mn+1 (bj)k2 < 2−n−1, ∀j ≤ n + 1. We choose such a large j and put kn+1 = j. Letting vn+1 = v∗ P n+1 kn+1 n + 1. 1uv1...vn ∈ P n , kn u∗, we see that (F3) is satisfied for m = = v1...vn+1Pn+1v∗ 1 = uP n n+1...v∗ n...v∗ If we now define Q = ∪nP n kn ′ ∩ M = R, then condition (F3) implies Q′ ∩ Mn ⊂ Q′ ∩ Rn, ∀n, while Fact 2 implies we have Q′ ∩ Rn = N ′ ∩ Mn. ′ ∩ N ⊂ ∪nP n kn kn+1 22 SORIN POPA The calculations for the relative commutants of R are similar, thus finishing the proof. (cid:3) Note that the case "Γ finitely generated" of Theorem 2.1 already follows from Theorem 2.10 above, due to the observation we made just before stating Theorem 2.1. But deriving from this the case "Γ infinitely generated" is problematic, as applying it for "larger and larger" finitely generated subgroups of Γ would involve in the limit multiplying infinitely many perturbing unitaries. To deal with this problem we'll use a similar trick, but with a "diagonal embedding" of N ≃ N σ into the algebra of matrices "of size Γ" over N , M σ ≃ N ⊗B(ℓ2Γ). Lemma 2.11. Let {σg}g∈Γ ⊂ Aut(N ) be an outer action of a group Γ on a II1 factor N , with σe = idN , and vg,h ∈ U(N ) satisfying σgσh = Ad(vg,h)σgh, ve,h = ve,g = 1, ∀g, h ∈ Γ. Define M σ = N ⊗B(ℓ2Γ) and N σ = {Pg σg(x)e0 gg x ∈ N } ⊂ M σ, where {e0 gg ′}g,g ′∈Γ ⊂ B(ℓ2Γ) are the usual matrix units. Let R ⊂ M σ be a subfactor, with the property that {e0 gg}g ⊂ R and let Q ⊂ R ∩ N σ be a common subfactor such that Qe0 gg = e0 ggRe0 gg, ∀g ∈ Γ. 1◦ Let Q0 ⊂ N be the unique subfactor with Q0eee = Qeee. There exist wg ∈ U(N ), we = 1, such that σ′ g := Ad(wg) ◦ σg satisfy σ′ g(Q0) = Q0, ∀g ∈ Γ. 2◦ If in addition Q′ ∩ R = N σ ′ ∩ R = {e0 is outer, ∀g ∈ Γ, g 6= e, and v′ g,h = wgσg(wh)vg,hw∗ gh normalize Q0, ∀g, h. gg}′′ g , then σ′ gQ0 3◦ Let e ∈ Fi = F −1 e0 gg. Then the inclusions of II1 factors (Qq0 i ր Γ be the net of finite symmetric subsets, and denote q0 i = i M σq0 Pg∈Fi i ) make a non-degenerate commuting square (with respect to the trace preserving con- ditional expectations), ∀i. i ) ⊂ (N σq0 i ⊂ q0 i ⊂ q0 i Rq0 4◦ Denote M σ 1 := M σ⊗B(ℓ2Γ) ≃ M ⊗B(ℓ2Γ)⊗B(ℓ2Γ) and consider the em- gg, where σ = σ ⊗ 1 and e1 gg, then 1 given by j1(x) = P−1 g σg(x)e1 gg ′}g,g ′ are the matrix units of the 2nd copy of B(ℓ2Γ). If q1 i = Pg∈Fi bedding j1 : M σ ֒→ M σ {e1 N σ ⊂ q0 i )q1 i M σq0 i M σq0 i ≃ j1(q0 i ⊂ q0 i q1 i M σ 1 q0 i q1 i is a basic construction. 5◦ With the above notations let Ri i q1 1 ⊂ q0 i q1 i M σ i )′ ∩ Ri 1 = (N σq0 1 is a basic construction. If (Qq0 Ri {σ′ gQ0 0 ∩ N = C), then v′ Q′ H 3(Γ)-obstruction as σ. i Rq0 i ֒→j1 i , ∀i, then }g is an outer action of Γ on Q0. If in addition Q′ ∩ N σ = C (equivalently, }g ⊂ Out(Q0) has same g,h ∈ Q0, ∀g, h ∈ Γ, and {σ′ i be so that Qq0 i q1 i M σ i ⊂ q0 1 q0 i q1 i )′ ∩ q0 1 q0 i q1 gQ0 Proof. 1◦ Identifying N = N ⊗ 1, let j0 : N ≃ N σ be the isomorphism defined by the property j0(x)e0 ee, x ∈ N . Thus, Q0 = j−1 ee = xe0 0 (Q). VANISHING COHOMOLOGY 23 ee = e0 Since R is a (necessarily semifinite) factor and e0 trace, there exist partial isometries e′ ∀g, and e′ elements wg ∈ N such that e′ Aut(N ), g ∈ Γ. Note that for any x ∈ N we have σ′ addition x ∈ Q0, then e′ Thus, σ′ gg belong to R and have same gg, ee. From the way M σ is defined, it follows that there exist unitary g = Ad(wg) ◦ σg ∈ ∗. If in ee. eg, ∀g, with we = 1. Let σ′ eg ∈ R with left support e0 ee which is equal to Qe0 ee, right support e0 ∗ lies in e0 ee = Q0e0 eg = wge0 egj(x)e′ eg egj(x)e′ eg ee = e′ g(x)e0 eeRe0 gσ′ h = Ad(v′ g,h)σ′ gh, ∀g, h ∈ Γ. Since σ′ g normalize Q0, it follows g(Q0) = Q0. We clearly have σ′ that v′ g,h normalize Q0, ∀g, h. 2◦ Note that if g, g′ ∈ Γ, then e0 is equivalent to σ′ g 6= σ′ this to g′ = e, proves 2◦. g ′ in Out(N ) (resp. σ′ gg(N σ′∩M σ)e0 gQ0 g ′g ′ = 0 (resp. e0 gg(Q′∩R)e0 g ′g ′ = 0) in Out(Q0)). Applying 6= σ′ g ′ Q0 3◦ The fact that the inclusions make a commuting square is obvious from 1◦ above. 4◦ This is in (Section 5.1.5 in [P91]). 5◦ Note that (Qq0 i )′ ∩ q0 q0 i Rq0 already know that σ′ i q1 i M σq0 gQ0 i = (N σq0 i )′ ∩ Ri is outer, ∀g 6= e. 1 = (N σq0 i )′ ∩ i , ∀i, which implies Q′ ∩ R = N σ ′ ∩ M σ. So by 2◦, we i , ∀i, implies (Qq0 i )′ ∩ q0 i M σ 1 q0 i q1 i q1 i q1 i Rq0 i ֒→ Ri i ⊂ q0 i ⊂ q0 Fix g, h ∈ Γ and let Fi be sufficiently large so that g, h, gh ∈ Fi. By taking into 0 and 1, as well as (Section 5.1.5 in [P97]), we see like in the proof gh from the gh in Out(N ) iff g,h) implements g,h ∈ Q0, have the same scalar (cid:3) account the form of the basic constructions N σq0 Qq0 of 2◦ above that the equality of relative commutants multiplied by e0 left and by e0 σ′ gQ0 inner automorphisms on Q0. If in addition Q′ ∀g, h ∈ Γ. But if {v′ H 3(Γ)-obstruction. h not equal to σ′ in Out(Q0). This shows that Ad(v′ g,h}g,h ⊂ Q0, then σ′ (thus σ) and σ′ hhfrom the right implies that σ′ 0 ∩ N = C, then this forces v′ gge1 not equal to σ′ i ֒→ q0 i M σq0 0M σq0 σ′ hQ0 eee1 ghQ0 i q1 i q1 gσ′ Q0 Proof of Theorem 2.1. By Proposition 1.5.2◦, we may assume N is separable. Also, without loss of generality, we may assume σe = idN . Like in Lemma 2.11, we denote {e0 and let M σ := N ⊗B(ℓ2Γ), j0 : N ֒→ M σ given by j0(x) = Pg σg(x)e0 j0(N ) ⊂ M σ . Choose an increasing sequence of finite sets e ∈ Fn = F −1 exhaust Γ and denote q0 gh}g,h∈Γ ⊂ B(ℓ2Γ) the canonical matrix units gg, N σ = n ⊂ Γ that n = Pg∈Fn e0 gg. As in Lemma 2.11.5◦, and with the notations established there, the embedding of factors N ≃j0 N σ ⊂ M σ implements (by induction/reduction) a sequence of n ≃ N Fn whose basic construction inclusions of II1 factors N ≃ N σq0 nM σq0 n ⊂ q0 24 SORIN POPA identifies, via j1(·)q1 n, with q0 nq1 nM σ 1 q0 nq1 n. n ⊂ q0 nM σq0 For each n, the subfactor N σq0 n is a locally trivial extremal inclusion of II1 factors with standard graph given by the Cayley graph of the subgroup Γn = hFni ⊂ Γ (see 5.1.5 in [P91]), which is amenable. Using this, we apply Theorem 2.10 to construct recursively a sequence of subfactors of finite index N ⊃ P0 ⊃ P1 ⊃ ... such that if we denote Q0 = ∪nP ′ n ∩ N , R = ∪nj0(Pn)′ ∩ M σ, then R is an irreducible type II subfactor in M σ containing the finite projections {e0 gg}g, with Q := j0(Q0) ⊂ N σ ∩ R a II1 subfactor satisfying the condition Q′ ∩ R = Q′ ∩ M σ = N σ ′ ∩ M σ = {e0 ggRegg, ∀g. By Lemma 2.11.5◦ the inclusion (Qq0 n) is a non-degenerate commuting square and our construction will show that Rn n is the basic construction algebra for the subfactor (Qq0 n ⊂ q0 n), ∀n, where j = j1 ◦ j0. The desired conclusions will then follow from 2.11.5◦. gg g ∈ Γ}′′ and with Qe0 n ⊂ q0 n ⊂ q0 1 := ∪mj(Pm)′ ∩ q0 gg = e0 nM σq0 n) ⊂ (N σq0 nRq0 nRq0 nM σ 1 q0 nq1 nq1 Let T r denote the semifinite trace τ ⊗ T rB(ℓ2Γ) ⊗ T rB(ℓ2Γ) on M σ 1 and k k2,T r the corresponding L2-norm. Let Y = {yn}n ⊂ (M σ 1 , T r) be a k k2,T r-dense sequence and denote by Yn = {j1(q0 j ykj1(q0 j 1 ≤ j, k ≤ n}. We construct the decreasing sequence of subfactors Pm ⊂ N recursively, such that P0 = N and for m ≥ 1 1 )1 ∪ L2(M σ j )q1 j )q1 1 m)′∩M σ mM σq0 m)-compatible. (y) − Ej1(N σ )′∩M σ m is (N σqm ⊂ q0 (i) (j0(Pm) ⊂ N σ)q0 (y))k2,T r < 2−m/Fm, ∀y ∈ Ym. (ii) kEj1((j0(Pm)′∩N σ )q0 If we made this construction up to m = n, then by applying Theorem 2.10 to the inclusion of II1 factors j0(Pn)q0 n+1 (which has amenable graph by Proposition 2.7, but this is trivial here, because this subfactor is in fact locally trivial, with standard graph given by a Cayley graph of an amenable subgroup of Γ, see 5.1.5 in [P97]) we get a subfactor Pn+1 ⊂ Pn so that its image via j0( · )q0 is (N σq0 m = n + 1. n+1 n+1)-compatible and such that (ii) above is satisfied for n+1 ⊂ q0 n+1M σq0 n+1M σq0 n+1 ⊂ q0 1 With the sequence {Pn}n this way constructed, define Q0, Q and R as explained gg, ∀g, by construction. gg}g and satisfies Qe0 above. Then R contains {e0 gg = e0 ggRe0 Moreover, condition (ii) shows that j(Q)j1(q0 n)q1 n ′ ∩ j1(q0 n)q1 nM σ 1 j1(q0 n)q1 n = j1(q0 n)q1 n(j1(N σ)′ ∩ M σ 1 )j1(q0 n)q1 n = j(Q)′ ∩ Rn 1 . From all this, it also follows that Q′ isomorphic to R by ([MvN43]). 0 ∩ N = C and that Q0 is AFD, and thus Thus, 2.11.5◦ applies, to conclude that there exist {wg}g ⊂ U(N ) such that σ′ g = Ad(wg)σg normalizes the irreducible hyperfinite II1 subfactor Q0 ⊂ N , ∀g, VANISHING COHOMOLOGY 25 v′ g,h = wgσg(wh)vg,hw∗ obstruction. gh belong to Q0, ∀g, h, and σ, σ′ Q0 have the same H 3(Γ) (cid:3) We end this section by noticing that the same argument we used in the proof of Theorem 2.1 above shows that any "amenable family" G of endomorphisms (or Hilbert bimodules) over a II∞ factor N normalizes a "large" AFD subfactor R⊗B(ℓ2N) ⊂ N , on which it "acts faithfully". To state this result, we need to clar- ify the terminology and fix some notations. Also, the reader should recall that for a properly infinite factor N (of type II∞ in our case) one has Connes' well known correspondence between a Hilbert N -bimodule N HN and a N -endomorphisms θH (see [C80]; see also Sec. 1.1 in [P86]). Via this correspondence, the adjoint opera- tion on endomorphisms, θ 7→ ¯θ, is given by ¯θH = θ ¯H and tensor product H⊗N H′ corresponds to composition θH ◦ θH′ (see e.g., 1.3 in [P86]). It is this "nice" cor- respondence between Hilbert-bimodules and endomorphisms for properly infinite factors that imposes using the framework of II∞ (i.e., infinite amplification of a II1 factor), rather than II1 factors, as algebras on which a category "acts" (the II1 framework would instead require considering morphisms between amplifications of the II1 factor, see 1.1 in [P86]). For us here, if N is a given II∞ factor, a concrete C∗-tensor category G of endo- morphisms on N is a family of classes (mod perturbation by inner automorphisms of N ) of unital endomorphisms of N that contains the idN (thus the class of in- ner automorphisms) and satisfies the following properties: (i) it is closed to the if θ ∈ G then ¯θ ∈ G; (ii) it is closed under composition, adjoint operation, i.e. i.e, if θ, θ′ ∈ G then θ ◦ θ′ ∈ G; (iii) Each θ ∈ G dilates the trace T r = T rN by a finite scalar 1 ≤ d(θ) < ∞, i.e., T r ◦ θ = d(θ)T r, with the image subfactor θ(N ) ⊂ N having Jones index given by the formula [N : θ(N )] = d(θ)2; (iv) it is closed to "direct sum and subtraction", in the following sense: (a) if θ, θ′ ∈ G and v, v′ ∈ N are isometries such that vv∗ + v′v′∗ = 1 then the endomorphism N ∋ x 7→ θ ⊕ θ′(x) := Ad(v)θ(x) + Ad(v′)θ′(x) ∈ N belongs to G; (b) if θ ∈ G, 0 6= p ∈ P(θ(N )′ ∩ N )) and v ∈ N is an isometry with range p, then Ad(v∗)θ ∈ G. We denote Irr(G) = {θ ∈ G θ(N )′ ∩ N = C}, the family of irreducible elements in G, which we label by the set K = KG, as {θk}k∈K, with e ∈ K so that θe = idN and with the adjoint operation implemented by k 7→ ¯k (thus θ¯k = ¯θk). It is easy to see from the above properties of G that any θ ∈ G decomposes as a finitely supported direct sum ⊕k∈K nk θk of irreducible endomorphisms in G, with finite multiplicities nk ≥ 0 (thus 0 < Pk nk < ∞). A subset e ∈ F0 = ¯F0 ⊂ K generates G, if any θ ∈ G can be obtained from F0 by applying consecutively finitely many times the operations (i), (ii), (iv). If G is generated by a finite such set F0 ⊂ K, one denotes by ΓG,F0 (or simply ΓG 26 SORIN POPA when F0 is clear from the context) the Cayley-type bipartite graph (or matrix with non-negative integer entries) (akk′ )k,k′∈K with akk′ equal to the multiplicity of θk′ in (Pi∈F0 θi) ◦ θk. We say that such a family G is amenable if it satisfies the Følner-type condition we mentioned before Proposition 2.7, i.e., if given any finite e ∈ F0 = ¯F0 ⊂ K, and any ε > 0, there exists a finite subset F in the subcategory hF0i ⊂ G generated d2 by F0 ⊂ K, such that P k, where ∂F0 F is the boundary of F in hF0i, i.e., the set of all k′ ∈ KhF0i \ F such that ak′k 6= 0 for some k ∈ F , where (akk′)k,k′ = ΓhF0i is the graph of hF0i. By (Theorem 5.3 in [P97]), this condition is equivalent to a Kesten-type condition, requiring that kΓhF0ik2 = (Pk∈F0 dk)2 for any finite e ∈ F0 = ¯F0 ⊂ K. d2 k′ < ε P k∈F k′∈∂F0 F If G is a concrete C∗-tensor category of endomorphisms on N and Q0 ⊂ N is a II∞ subfactor, then we say that G faithfully normalizes Q0, if the following condi- tions are satisfied: (1) the trace T r is semi finite on Q0; (2) for each endomorphism θ ∈ G there exists θ′ ∈ G in the same class as θ such that θ′(Q0) ⊂ Q0, and given any other θ′′ ∈ G in the same class as θ that satisfies θ′′(Q0) ⊂ Q0, there exists w0 ∈ U(Q0) with Ad(w0) ◦ θ′′ ; (3) If θ ∈ G is so that θ(Q0) ⊂ Q0, then [Q0 : θ(Q0)] = d(θ)2 = [N : θ(N )] and θ is irreducible non-inner on N iff θQ0 irreducible and non-inner on Q0. = θ′ Q0 Q0 Note that a "concrete C∗-tensor category G of endomorphisms on a factor N " is an analogue of a "group of outer automorphisms Γ ⊂ Out(N )". If one still denotes by G the underlying abstract C∗-tensor category, as defined for instance in [NeTu13], then such an object can as well be viewed as an "outer action of G by endomorphisms on N ". With this interpretation, if Q0 ⊂ N is faithfully normalized by G, then the restriction map θ 7→ θQ0 is an isomorphism of the underlying abstract rigid C∗-tensor categories. We say that such a restriction is strongly smooth if in addition θ(Q0)′ ∩ N = θ(N )′ ∩ N , ∀θ ∈ G. 2.12. Theorem. Let N be a II∞ factor and G an amenable countably generated concrete C∗-tensor category of endomorphisms on N , as defined above. Then N contains an AFD II∞ subfactor R ⊂ N that's faithfully normalized by G, in the above sense. Moreover, if N is separable, then R can be chosen so that to satisfy the strong smoothness condition θ(N )′ ∩ N = θ(R)′ ∩ N = θ(R)′ ∩ R, for all θ ∈ G with θ(R) ⊂ R. Proof. The proof uses the same ideas and follows exactly the same steps as the proof of Theorem 2.1. Note first that, as in the proof of Theorem 2.1, it is clearly sufficient to prove the case when N is separable. Next, note that we have a version for endomorphisms of Lemma 2.11.1◦, as VANISHING COHOMOLOGY 27 follows: θk(x)e0 Fact 1. Let N be a properly infinite von Neumann factor and θ = {θk}k∈K0 be a set of endomorphisms of N , with e ∈ K0 and θe = idN . Define Mθ = N ⊗B(ℓ2K0) and N θ = {Pk∈K0 kk′}k,k′∈K0 ⊂ B(ℓ2K0) are the usual matrix units. Let R ⊂ Mθ be a subfactor, with the property that {e0 kk are mutually equivalent infinite projections in R. Let Q ⊂ ee = e0 R ∩ N θ be a common subfactor such that Qe0 ee. Then there exists a unique subfactor Q0 ⊂ N such that Q0e0 ee = Qe0 ee and there exist unitary elements wk ∈ N , k ∈ K0, we = 1, such that θ′ kk x ∈ N } ⊂ Mθ, where {e0 k = Ad(wk) ◦ θk satisfies θ′ kk}k ⊂ R and e0 k(Q0) ⊂ Q0, ∀k. eeRe0 The proof if the same so we omit it. Let j0 : N ≃ N θ ⊂ Mθ denote the isomorphism satisfying j0(x)e0 ee = we0 ee, x ∈ N . Thus, Q0 = j−1 0 (Q). In order to make the ideas more transparent, let us first give an argument for the case when G is generated by a finite subset of irreducible endomorphisms {θk}k∈K0, with e ∈ K0 = ¯K0 ⊂ K, by exploiting the remark before the statement of Theorem 2.1. So let N be a II1 factor such that N ∞ = N , let M θ,K0 be the amplification of dk, {pk}k∈K0 ⊂ M θ,K0 a partition of 1 with projections having traces N by Pk∈K0 τ (pk) proportional to dk and consider its subfactor N θ,K0 = {Pk∈K0 θk(x)pk x ∈ N ≃ peM θ,K0pe}, where θk are endomorphisms chosen in their class so that θk(pe) = pk, k ∈ K0. Note that N θ,K0 ⊂ M θ,K0 is an extremal inclusion of II1 dk)2 and relative commutant generated by {pk}k. We factors with index (Pk∈K0 denote j0 : N = peN pe ≃ N θ,K0 the identification. Fact 2. With the above notation, assume (Q ⊂ R) ⊂ (N θ,K0 ⊂ M θ,K0) is a non- degenerate commuting square of factors such that {pk}k ⊂ R and peRpe = Qpe. Let (R ⊂ R1 ⊂ ...) ⊂ (M θ,K0 ⊂ M1 ⊂ ...) be the associated tower of commuting squares, obtained by iterating the basic construction. If Q′ ∩ Rn = N θ,K0 ∩ Mn, ∀n, and one denotes by Q0 = j−1 0 ⊂ N ∞ = N is faithfully normalized by G. 0 (Q) ⊂ N , then Q0 = Q∞ ′ Indeed, by Fact 1 we know that θk can be taken so that Q = {Pk∈K0 θk(x)pk x ∈ Q0} and thus so that θk(Q0) ⊂ Q0. After ∞-amplification of the tower of com- muting squares of factors, the iterated basic construction gives rise to consecutive products of endomorphisms of N = N ∞, θkn ◦ θkn−1 ◦ ...θk1, for ki ∈ K0, which all take Q0 = Q∞ 0 ⊂ N ∞ = N into itself. The condition on the compatibility of higher relative commutants amounts to Q0 being faithfully normalized by G. Now, if G is amenable and finitely generated by its irreducible endomorphisms indexed by K0 ⊂ K as above, then N θ,K0 ⊂ M θ,K0 has amenable graph, so The- orem 2.10 applies to provide a strongly smooth non-degenerate commuting square (Q ⊂ R) ⊂ (N θ,K0 ⊂ M θ,K0), with Q and R hyperfinite II1 factors. By Fact 2, Q 28 SORIN POPA gives rise to a subfactor Q0 ≃ Q∞ that's faithfully normalized by G and satisfies the required strong smoothness condition. This proves the statement in the case G is finitely generated. To deal with the infinitely generated case, we take the inclusion N θ ⊂ Mθ in Fact 1 with K0 = K = KG and {θk}k∈K = Irr(G). We further embed Mθ = N ⊗B(ℓ2K) into Mθ kk′ }k,k′∈K are the matrix units in the 2nd copy of B(ℓ2K). Note that N ∋ x ∈ j1(j0(x)) ∈ Mθ 1 is given by j1(j0(x)) = Pk,k′∈K θk(θk′(x))e1 1 := N ⊗B(ℓ2K)⊗B(ℓ2K) by x 7→ j1(x) = Pk θk(x)e1 kk, where θk = θk ⊗ 1. kk, where {e1 Note that Fact 1 already implies that if (Q ⊂ R) ⊂ (N θ ⊂ Mθ) is a subinclusion ee, then there are 0 (Q) satisfies θk(Q0) ⊂ Q0, ∀k ∈ K. Then note of factors with {e0 representants θk such that Q0 = j−1 the following: ∩ Mθ ⊂ Q′ ∩ R and Qe0 k = N θ′ ee = e0 eeRe0 k′k′ e0 kk}′′ Fact 3. Assume the above inclusion (Q ⊂ R) ⊂ (N θ ⊂ Mθ) is so that: 1 = N θ′ ∩ Mθ i = Pk∈Ki 1; (b) there exist finite subsets e ∈ Ki = ¯Ki ր K such that (a) Q′ ∩ Mθ if we denote q0 i ) is a non- degenerate commuting square (with respect to the T r-preserving expectations). Then Q0 = j−1 0 (Q) ⊂ N is faithfully normalized by G, with its restriction to Q0 strongly smooth. e0 kk then (Qq0 i ) ⊂ (N θq0 i ⊂ q0 i ⊂ q0 i Rq0 i Rq0 The proof of this fact is identical to the proof of 2.11.5◦ above and we leave it as an exercise. θk(x)pk; as well as the embedding M θ,Kn ֒→jn From this point on, the proof of Theorem 2.12 is similar to the proof of Theorem 2.1. Thus, we let e ∈ Kn = ¯Kn ր K be finite sets, denote Dn = Pk∈Kn dk, let {pk}k∈K ⊂ N be mutually orthogonal projections of trace T r(pk) = dk such that 1 − Pk∈K is an infinite projection. By perturbing if necessary each θk by an inner automorphism, we may assume θk(pe) = pk, ∀k. We also choose θ′ k to be in the same class as θk so that θ′ k′ (Pk∈K pk), k′ ∈ K, are mutually orthogonal. For each n consider the embedding N = peN pe ֒→jn 0 (x) = := (M θ,Kn)Dn ≃ Pk∈Kn N D2 n , by x 7→ jn Note that these are extremal embeddings of II1 factors with finite index = D2 n and that jn is a basic construction, ∀n. Moreover, for each n > m, the inclusion for m is obtained from the one for n by induc- tion/reduction. We can thus apply the same iterative construction of subfactors of finite index N ⊃ P1 ⊃ P2..... as in the proof of Theorem 2.1, so that if one denotes Q0 = ∪mP ′ , then (jn ) is a basic construction of non-degenerate commuting squares that are strongly smooth, ∀n. 0 M θ,Kn := N Dn by x 7→ jn 1 (M θ,Kn) ⊂ M θ,Kn m ∩ N , Rn 1 (Rn 0 = ∪mjn 0 ) ⊂ Rn 1 (M θ,Kn) ⊂ M θ,Kn 0 (Pm)′ ∩ M θ,n, Rn 0 (Pm))′ ∩ M θ,n 1 (x) = Pk∈Kn 1 pk and pn 1 (Q0)) ⊂ jn e = Pk∈Kn 0 (N )) ⊂ jn k , where pn 0 (N )) ⊂ jn 1 = ∪mjn 1 M θ,Kn 1 ) ⊂ (jn 1 (jn θ′ k(x)pn k = θ′ k(pn e ). 1 (jn 1 1 1 (jn 1 1 (jn VANISHING COHOMOLOGY 29 But then, it is easy to see that after appropriate amplifications, Q0 ⊂ N gives rise to a II∞ subfactor in Q0 ⊂ N that satisfies the conditions in Fact 3, and is thus faithfully normalized by G, with the strongly smooth condition satisfied. (cid:3) 2.13. Remarks. 1◦ The first result about normalizing a "large" hyperfinite sub- factor R modulo inner perturbations was obtained in [P83], in the case Γ = Z, with R being constructed "by hand", using Rokhlin towers and an iterative procedure. Shortly after, the question of whether any cocycle action of Z2 on an arbitrary II1 factor can be untwisted was asked in [CJ84]. While we realized at that time that if a similar normalization result could be proved for Γ = Z2 then the problem would reduce to the case N = R, where vanishing cohomology was just shown in [Oc85], we could not extend the arguments in [P83] from Z to Z2, despite much effort. Several years later in [P89], we were able to solve this problem by using tools from subfactor theory. However, the argument in [P89] could only cover groups Γ that have a finite set of generators S ⊂ Γ with respect to which Γ has trivial Poisson boundary (e.g., groups with polynomial growth, in particular Z2), depending cru- cially on this condition. In retrospect, it is quite surprising that in fact any outer action of any amenable group Γ on any II1 factor N normalizes (modulo Inn(N )) an irreducible hyperfinite subfactor, a property which turns out to characterize the amenablility of Γ. 2◦ Both in the case of outer action σ of a group Γ on N (or more generally of an outer action of a category G on N ) there may be an AFD subalgebra R ⊂ N (resp. R ⊂ N ) that's normalized by Γ (resp. G) but on which the resulting "outer action" has either more relations among generators (resulting into an outer action of a quotient of Γ, resp. G), or fewer relations (resulting into an outer action of some Γ, resp. G, whose quotient is Γ, resp. G). For instance, let Fn y R be a free action of the free group with n ≥ 2 generators on the hyperfinite II1 factor and Fn → Γ a surjective map on a (non-free) group Γ with n generators (e.g. Γ = Zn) and H its kernel. If we let N = R ⋊ H then this gives rise to a free cocycle action (σ, v) of Γ on N , which normalizes R, but on which it generates a free action of Fn. Similarly, one can take Γ y N = N0⊗R an action that's trivial on R but free on N . This normalizes R but it implements on it the action of the trivial group. 3. Non-vanishing cohomology for amplifications of actions 3.1. Definition ([P01a]). Let Γ yσ N be a free action of a group Γ on a type II1 factor N . Let p ∈ N be a projection and for each g ∈ Γ choose a partial isometry wg ∈ N such that wgw∗ g = p, w∗ g ∈ Aut(pN p) g (x) = wgσg(x)w∗ g, x ∈ pN p. Then σp is a free cocycle action of Γ on pN p, by σp with 2-cocycle vp gh, g, h ∈ Γ. Moreover, up to cocycle conjugacy g wg = σg(p) and we = p. Define σp g,h = wgσg(wh)w∗ 30 SORIN POPA (σp, vp, pN p) only depends on τ (p) = t, thus defining a free cocycle action (σt, vt) of Γ on N t, called the amplification of σ by t. If in addition {wg}g ⊂ N satisfy wgσg(wh) = wgh, ∀g, h ∈ Γ, then w is called a generalized 1-cocycle (of support t) for σ, while if this equality holds modulo scalars then it is called a weak generalized 1-cocycle for σ (of support t). Note that the vanishing (resp. weak-vanishing) of the cocycle vt amounts to being able to chose the partial isometries {wg g ∈ Γ} ⊂ N so that w be a generalized 1-cocycle (resp. weak generalized 1-cocycle) for σ. 3.2. Theorem. Let Γ be a countable group and Γ yσ R = R⊗Γ the non- commutative Bernoulli Γ-action with base R0 ≃ R. Let 0 < t < 1 and denote by σt the free cocycle action obtained by amplifying σ by t, with its 2-cocycle de- noted vt. Assume one of the following properties holds true: (a) Γ contains an infinite subgroup with the relative property (T); (b) Γ contains an infinite subgroup with non-amenable centralizer. Then the cocycle vt is not weak-vanishing. 0 Proof. Assume by contradiction that the cocycle vt is weak-vanishing. If we are under assumption (a), i.e., if Γ has an infinite subgroup with the relative property (T), then (Corollary 4.10 in [P01a]) shows that the support t of the generalized weak 1-cocycle wg must be 1, a contradiction. If in turn we are under assumption (b), with H ⊂ Γ an infinite subgroup with centralizer H ′ = {g ∈ Γ gh = hg, ∀h ∈ H} non amenable, then let (α, β) be the s-malleable deformation of R⊗R that commutes with the double action σ = σ ⊗ σ : Γ y R = R⊗R = (R0⊗R0)⊗Γ, as in ([P01a]). Denote M = R ⊗ Γ, M0 = R ⋊ H, M = R⊗R ⋊σ Γ and M0 = R⊗R ⋊σ H, with M identified as the subfactor R ⊗ 1 ∨ {Ug}g in M , where {Ug}g ∈ M are the canonical unitaries implementing σ. Since H ′ is non-amenable, the action Ad(wgUg) of H ′ on (p ⊗ 1) M0(p ⊗ 1) has spectral gap relative to (p ⊗ 1)M0(p ⊗ 1). Arguing like in ([P06a]), this implies that whUh and αs(wh)Uh, h ∈ H, are uniformly k k2-close, for s ∈ R with s small. Equivalently, the map ξ 7→ (wg ⊗ 1)σg(ξ)αs(w∗ g ⊗ 1) gives a unitary representation πs of the group H on the Hilbert space (p⊗1)L2(R ⊗R)αs(p⊗1) which for s small has the vector ξ0 = (p ⊗ 1)αs(p ⊗ 1) almost fixed by πs(h), uniformly in h ∈ H. We are thus in exactly the same situation as in the case when H ⊂ Γ is rigid. So the proof of (Corollary 4.10 in [P01a]) applies to conclude that the support t of the generalized weak 1-cocycle {wh h ∈ H} for σ must necessarily be equal to 1, a contradiction. (cid:3) 3.3. Corollary. If Γ is a group that contains an infinite subgroup which either has relative property (T), or has non-amenable centralizer in Γ, then Γ 6∈ VCw(R). VANISHING COHOMOLOGY 31 3.4. Theorem. Let Γ be a countable group and Γ yρ L(FΓ) the action imple- mented by the translation from the left by elements h ∈ Γ on the set {ag}g∈Γ of generators of the free group FΓ, ρh(ag) = ahg, ∀h, g ∈ Γ. Let 1 > t > 0 and de- note by ρt the free cocycle action obtained by amplifying ρ by t, with its 2-cocycle denoted vt. Assume one of the following properties is satisfied: (a) Γ contains an infinite subgroup with the relative property (T); (b) Γ contains an infinite subgroup with non-amenable centralizer; (c) Γ contains an infinite amenable group with non- amenable normalizer. Then the cocycle vt is not weak-vanishing. Proof. The assumptions (a) and (b) lead to exactly the same argument as in the proof of Theorem 3.2 above, by using the "free s-malleable deformation" of L(FΓ), like in the proof of (Theorem 6.1 in [P01a]). So let us assume we are under the assumption (c) and let H ⊂ Γ be an infinite amenable subgroup with its normalizer G ⊂ Γ non-amenable. Note that FΓ ⋊ Γ is naturally isomorphic to Γ ∗ Z. Denote N = L(FΓ), M = N ⋊ Γ = L(FΓ) ⋊ Γ = L(Γ ∗ Z). Let p ∈ P(N ) be a projection of trace t < 1 and assume (ρp, vp, pN p) has weak vanishing cohomology. Thus, if {Ug}g∈Γ ⊂ M denote the canonical unitaries implementing ρ on N = L(FΓ), then there exist partial isometries wg ∈ N of left support p such that U ′ g = wgUg ⊂ U(pN p) gives a projective representation of Γ with scalar 2-cocycle g g ∈ G}′′ gives an embedding of Lµ(H) ⊂ Lµ(G) into pM p. µ. In particular, {U ′ Note that Lµ(H) is amenable diffuse, Lµ(G) has no amenable direct summand and Lµ(H) is regular in Lµ(G). Thus, by (Corollary 1.7 in [I13]) it follows that Lµ(G) ≺M L(Γ). But Lµ(G) ≺M L(Γ) implies that the "free" malleable deformation (α, β) of M ⊂ M = L(FΓ ∗ FΓ) ⋊ Γ is uniform on {U ′ g g ∈ G} (because L(Γ) is in the fixed point algebra of the malleable deformation and Lµ(G) is subordinated to L(Γ)). Thus, for each s ∈ R, the G-representation ξ 7→ (wg ∗ 1)σg(ξ)αs(w∗ g ∗ 1) gives a unitary representation πs of the group G on the Hilbert space (p ⊗ 1)L2(N ∗ N )αs(p ⊗ 1) which for s small has the vector ξ0 = (p ∗ 1)αs(p ∗ 1) almost fixed by πs(g), uniformly in g ∈ G. As in the proof of (Theorem 4.1 in [P01a]) this shows that we must necessarily have t = 1, a contradiction. (cid:3) 3.5. Remark. Let (N0, τ0) be an arbitrary tracial von Neumann algebra 6= C and Γ a group that satisfies one of the conditions (a) or (b) in Theorem 3.2. If Γ y N = N ⊗Γ is the Bernoulli action with base N0 and 0 < t < 1, then by using the s-malleable deformation of N = N ⊗Γ in ([I06]) in combination with the argument in the proof of Theorem 3.2, one obtains that the t-amplification (σt, vt, N t) of σ has non-vanishing cohomology. Similarly, by using the s-malleable 0 0 32 SORIN POPA deformation in [IPeP05] one can obtain a generalization of Theorem 3.4 for the free-Bernoulli action Γ yρ N = N ∗Γ 0 . 4. Non-vanishing cohomology for cocycle actions on L(F∞) 4.1. Definition ([CJ84]). Let Γ be an infinite countable group with a set of generators S ⊂ Γ and assume (Γ, S) 6= (FS, S). Let π : FS → Γ be the unique group morphism taking the free generators S of FS onto S ⊂ Γ (so our assumption is equivalent to π not being 1 to 1). Then kerπ ≃ F∞ and FS has infinite conjugacy classes relative to kerπ. Thus, the inclusion of II1 factors L(kerπ) = N ⊂ M = L(FS) is irreducible and regular, with NM (N )/U(N ) = Γ. Consequently, N ⊂ M is a crossed product inclusion of the form L(F∞) ⊂ L(F∞) ⋊(σπ ,vπ) Γ, for some free cocycle action (σπ, vπ) of Γ on L(F∞), that we'll call the Connes-Jones cocycle action associated with π, with vπ its 2-cocycle, called the CJ-cocycle. We summarize here some straightforward consequences of results from ([CJ84], [P01a], [O03], [P06b], [OP07]), but which are stated in a manner pertaining to our vanishing cohomology problem: 4.2. Theorem. Let (Γ, S), π : FS → Γ, Γ y(σπ,vπ) L(F∞) be as in 4.1. Assume Γ satisfies one of the following conditions: (a) it does not have Haagerup property (e.g., it contains an infinite subset with relative property (T)); (b) it has Cowling- Haagerup invariant Λ(Γ) larger than 1; (c) it has an infinite subgroup with non- amenable centralizer; (d) it has an infinite amenable subgroup with non-amenable normalizer. Then the CJ-cocycle vπ is not weak-vanishing. Thus, if Γ satisfies any of these properties then Γ 6∈ VCw(L(F∞)). Proof. If vπ is weak-vanishing, then we can choose representatives {Ug g ∈ Γ} in NM (N ) such that UgUh = µg,hUgh, ∀g, h ∈ Γ, for some scalar 2-cocycle µ ∈ H 2(Γ). g ∈ L(FS)op ≃ L(FS) satisfy Thus, Lµ(Γ) ⊂ M = L(FS). Also, the unitaries U op UgUh = µg,hUgh, implying that {Ug ⊗ U op g ≃ L(Γ) embeds into L(FS)⊗L(FS). g }′′ Since FS has Haagerup property and Λ(FS) = 1 (cf [H79]), the II1 factor L(FS) has Haagerup property and Λ(L(FS)) = 1, implying that L(FS)⊗L(FS) and all its von Neumann subalgebras have these properties as well (see e.g. [P01b] and [OP07]). In particular, L(Γ) has these properties, thus Γ has Haagerup property and Λ(Γ) = 1. If in turn Γ has an infinite subgroup H with non-amenable centralizer H ′, then by (Remark 3◦ in §4 of [P06b]) Lµ(Γ) would contain a diffuse von Neumann subalgebra B with non-amenable centralizer. But then B ⊂ M = L(FS) has non-amenable centralizer, contradicting the solidity of free group factors ([O03]). If Γ has an infinite amenable subgroup H ⊂ Γ with non-amenable normalizer G ⊂ Γ, then the normalizer of B = Lµ(H) in M = L(FS) contains all {Ug}g∈G. VANISHING COHOMOLOGY 33 Since {Ug}′′ group factors ([OP07]). g ≃ Lµ(G) is non-amenable, this contradicts the strong solidity of free (cid:3) 4.3. Notations. Given two discrete groups Γ, Λ, one writes Γ ≤W ∗ Λ whenever L(Γ) can be embedded (tracially) into L(Λ). Denote W∗ leq(Λ) the class of all groups Γ that can be subordinated this way to Λ, i.e., W∗ leq(Λ) = {Γ Γ ≤W ∗ Λ}). 4.4. Corollary. One has VC ⊂ VC(L(F∞)) ⊂ W∗ leq(F2). Thus, if Γ ∈ VC, then Γ has Haagerup property, Cowling-Haagerup constant equal to 1, any infinite subgroup of Γ has amenable centralizer, and any infinite amenable subgroup has amenable normalizer. 4.5. Remarks 1◦ The above criteria for Γ not to be in W∗ leq(F2) (and thus not in VC(L(F∞)) ⊃ VC) are in fact not stated in their optimal form. Thus, the results in [OP07] show that in order for Γ not to be in W∗ leq(F2), it is enough that L(Γ) is not strongly solid, i.e., that L(Γ) merely has a diffuse amenable von Neumann subalgebra whose normalizer in L(Γ) generates a non-amenable von Neumann al- gebra. The list in 4.2 is also not exhaustive. For instance, by [Pe07] it follows that if Γ ∈ W∗ leq(F2), then L(Γ) needs to be L2-rigid, while a result of Ozawa (see e.g. [BrO08]) shows that Γ needs to be exact. One should also note that in the proof of Theorem 4.2 above we showed that an embedding Lµ(Γ) ֒→ L(F2), for some µ ∈ H 2(Γ), gives rise to an embedding L(Γ) ֒→ L(F2 × F2), by simply doubling the canonical unitaries {ug g ∈ Γ} ⊂ Lµ(Γ) ⊂ L(F2), i.e., by taking L(Γ) ≃ {ug ⊗ uop g g ∈ Γ}′′ ⊂ L(F2)⊗L(F2) = L(F2 × F2). 2◦ It is reasonable to expect that VC = VC(L(F∞)), and more specifically that the CJ-cocycles are in some sense the "worse possible", ie., if any such cocycle vanishes for some group Γ, then Γ ∈ VC. We also believe that VC = VCw. 3◦ Given a group Λ, denote by ME(Λ) the class of groups Γ that are measure equivalent (ME) to Λ and by MEleq(Λ) the class of groups Γ that have a free m.p. action which can be realized as a sub equivalence relation of a free ergodic m.p. Λ-action. It would be interesting to explore the possible correlations between the classes VC, W∗ leq(F2), MEleq(F2), etc. In this respect, one should point out that while W∗ leq(F2), MEleq(F2) are obviously "hereditary" classes (i.e., if Γ belongs to any of them, then all subgroups of Γ belong too), we could not prove such hereditarity for VC (cf. Remark 1.6.2◦). See also Section 7 in [PeT10] for more comments on MEleq(F2) and its relations to W∗ leq(F2). One should also note that MEleq(F2) consists of groups Γ that are ME to either Z = F1, F2, or F∞, i.e., MEleq(F2) = ME(Z) ∪ ME(F2) ∪ ME(F∞) (cf. [G04], [Hj04]). 34 SORIN POPA 4◦ We do not know of any examples of groups in VC, W∗ leq(F2) other than amalgamated free products of amenable groups over finite groups. The approach in 4.2 indicates that these two classes may coincide (perhaps with MEleq(F2) as well). An intriguing class of groups that are known from [G04] to belong to MEleq(F2) (in fact, even to ME(F2)), are the free products of finitely many copies of F2 with amalgamation over the subgroup Z ⊂ F2 generated by the commutator aba−1b−1 (where a, b are the generators of F2). Gaboriau conjectured that in fact any amal- gamated free product of Fki , 2 ≤ ki ≤ ∞, i ∈ I (with I finite or I = N), over some Z ֒→ Fki which is maximal abelian in the corresponding Fki, ∀i, is in MEleq(F2). Thus, according to the above speculations, the groups Fk1 ∗Z Fk2 ∗Z ..., with Z maximal abelian in each Fki , should belong to VC as well. However, we were not able to prove this for any such example, except of course the case when any subgroup Z is freely complemented in Fki . Related to this, we pose here the following Question: Let Fn yσ N be a free action of a free group of rank n on a II1 factor N and let W denote the set of all 1-cocycles w for σ (i.e., maps w : Fn → U(N ) satisfying wgσg(wh) = wgh, ∀g, h ∈ Fn). Let Z ⊂ Fn be a maximal abelian subgroup, generated by some element g ∈ Fn. Is it then true that the set {wg w ∈ W} coincides with the unitary group U(N )? Taking into account the way the 1-cocycles w ∈ W are constructed, from n- tuples of unitaries in N that are taken as perturbations of the canonical unitaries U1, ..., Un ∈ N ⋊σ Fn that implement σa1 , ...., σan (where a1, ..., an are the generators of Fn), it immediately follows that this question has an affirmative answer in the case Z is freely complemented in Fn. It is also trivial to see that if the statement holds true, then all of the above Gaboriau groups Γ = Fk1 ∗Z Fk2 ∗Z ... lie in VC. It is not known whether these groups are in W∗ leq(F2) either. In fact, deciding that a group Γ satisfies L(Γ) ֒→ L(F2) is at least as interesting as deciding that it has property VC. So the fact that VC ⊂ W∗ leq(F2) gives another strong motivation for proving the universal vanishing cohomology property for various groups. The above question can also be stated for free measure preserving actions on the probability measure space, Fn y (X, µ), by simply replacing N by A = L∞(X, µ) throughout that statement. Besides answering this question, it would be interesting to know if an affirmative answer would imply Gaboriau's conjecture that the groups Fk1 ∗Z Fk2 ∗Z ... belong to MEleq(F2). 5◦ It has been conjectured by Peterson and Thom (see end of Sec. 7 in [PeT10]) that if two amenable von Neumann subalgebras B1, B2 of the free group factor L(F2) have diffuse intersection, then B1 ∨ B2 should follow amenable. There has been an accumulation of evidence towards this fact being true (e.g., [P81], [Ju06], [Pe07]). For us here, this would imply that if Γ ∈ W∗ leq(F2) is generated by amenable VANISHING COHOMOLOGY 35 subgroups Γ1, Γ2 ⊂ Γ with H = Γ1 ∩ Γ2 infinite, then Γ must be amenable. Thus, if Γ = Γ1 ∗H Γ2 then Γ 6∈ W∗ leq(F2), unless either H is finite, or [Γ1 : H] ≤ 2, [Γ2 : H] ≤ 2. In particular, VC(L(F∞)) should not contain such groups either. 6◦ We expect that VC(R) is equal to VC(L(F∞)). This fact suggests various new statements in deformation-rigidity for factors arising from Bernoulli actions. For instance, it should be possible to prove that if a group Γ does not have Haagerup property, or if it has an infinite amenable subgroup with non-amenable normalizer, then some of the W∗-rigidity results in [P01a], [P03], [P06a] should be true. Com- bining this conjecture with the Peterson-Thom conjecture and remark 5◦ above, this also suggests that VC(R) doesn't contain any non-amenable group Γ that can be generated by amenable subgroups Γ1, Γ2 with H = Γ1 ∩ Γ2 infinite (in particular Γ = Γ1 ∗H Γ2). But the obstruction in this case should be of a completely different nature. The II1 factors arising from Bernoulli actions of such groups (and more generally from non-amenable groups Γ generated by n ≥ 2 amenable groups with infinite intersection) may actually have additional W∗-rigidity properties, providing a new class of factors on which deformation-rigidity techniques should be tested, in the spirit of ([P01a], [P03], [P06a], [IPeP05], [PV12]). 7◦ Note that the W ∗-algebra version of von Neumann's conjecture on whether any non-amenable group Λ contains a copy of F2, amounts to whether for any non- amenable Λ one has F2 ≤W∗ Λ. Note also that by the Gaboriau-Lyons result in [GL07] one indeed has F2 ≤W∗ Z ≀ Λ for any non-amenable Λ, while by [OP07] it follows that if Z ≀ Λ ≤W∗ F2 then Λ must be amenable. 5. A related characterization of amenability We prove in this section that the "normalization" property for cocycle Γ-actions in Theorem 2.1 can only be true when the group Γ is amenable. More precisely, for any non-amenable Γ we exhibit examples of embeddings Γ ⊂ Out(L(F∞)) which admit no lifting to Aut(L(F∞)) that normalizes a hyperfinite subfactor of L(F∞). 5.1. Theorem. Let Γ be a countable group and Γ yσ L(FΓ) the action imple- mented by the translation from the left by elements h ∈ Γ on the set {ag}g∈Γ of generators of the free group FΓ. Let M = L(FΓ) ⋊ Γ with {Ug g ∈ Γ} ⊂ M the canonical unitaries implementing σ. The following conditions are equivalent: (a) Γ is amenable. (b) L(FΓ) contains a hyperfinite subfactor R with R′ ∩ M = C for which there g = wgUg, g ∈ Γ, normalize R, exist {wg}g∈Γ ⊂ U(L(FΓ)) with the property that U ′ implement a free Γ-action on it and satisfy U ′ gh, ∀g, h ∈ Γ. gU ′ h = U ′ 36 SORIN POPA (c) L(FΓ) contains a diffuse AFD von Neumann subalgebra B for which there exist {wg}g∈Γ ⊂ U(L(FΓ)) with the property that Ad(wg) ◦ σg normalize B, ∀g. Proof. By Theorem 2.2 we have (a) ⇒ (b), while (b) ⇒ (c) is trivial. To see that (c) ⇒ (a), note first that one has a natural identification between M = L(FΓ) ⋊σ Γ and L(Γ) ∗ L(Z). Then note that by [OP07], the von Neumann algebra B1 generated by the normalizer of B in L(FΓ) then B1 is amenable, thus AFD by [C75]. As in part (b), denote U ′ g = wgUg ∈ M , g ∈ Γ. Note that {U ′ g}g normalize B1 and implement a cocycle action of Γ on B1, i.e., one has U ′ gh(U ′ We first show that Γ non-amenable implies P = B1 ∨ {U ′ g g ∈ Γ} is non- amenable. We will prove this by contradiction, by showing that if P is amenable, then Γ follows amenable. h)−1 ∈ B1, ∀g, h ∈ Γ. gU ′ geBU ′ g To see this, let P ⊂ hP, eB1i be the basic construction for B1 ⊂ P and note that ∗ ∈ hP, eBi, g ∈ Γ, are mutually orthogonal projections summing up fg := U ′ to 1 and generating an atomic abelian von Neumann subalgebra A ⊂ hP, eB1i natu- rally isomorphic to ℓ∞Γ. Moreover, {Ad(U ′ g)}g normalizes A ≃ ℓ∞Γ implementing on it the Γ-action by left translation. If P is amenable then there exists a state ϕ on B(L2P ) that has P in its centralizer. Then ψ = ϕA is a state on A ≃ ℓ∞Γ that's invariant to {AdU ′ g}g, i.e., to left translations by Γ, showing that Γ is amenable. Now, if P ⊂ M ≃ L(Γ) ∗ L(Z) is non-amenable, then by (Corollary 1.7 in [I13]) it follows that P ≺M L(Γ), in particular B ≺M L(Γ). But by applying (Corollary 2.3 in [P03]), it is trivial to see that L(FΓ) (an algebra on which {Ug}g∈Γ acts) has no diffuse von Neumann subalgebra that can be subordinated ≺M to L(Γ) = {Ug}′′ g , a contradiction. (cid:3) 5.2. Theorem. Let Γ be a countable group with a set of generators S ⊂ Γ and corresponding surjective morphism π : FS → Γ, with kernel kerπ ≃ F∞. Let L(F∞) = N ⊂ M = L(FS) be the associated irreducible, regular inclusion of free group factors, which satisfies NM (N )/U(N ) ≃ Γ, and denote by Γ yσπ N the corresponding cocycle Γ-action. The following conditions are equivalent: (a) Γ is amenable. (b) N contains a hyperfinite subfactor R ⊂ N with R′ ∩ M = C for which g = wgUg, g ∈ Γ, normalize R, gh, ∀g, h ∈ Γ. there exist {wg}g ⊂ U(N ) with the property that U ′ h = U ′ implement a free action on it, and satisfy U ′ gU ′ (c) N contains a diffuse AFD von Neumann subalgebra B such that ∀g ∈ Γ, ∃wg ∈ U(N ) with the property that Adwg ◦ σg normalizes B. Proof. Theorem 2.2 shows that (a) ⇒ (b) and (b) ⇒ (c) is trivial. If (c) holds but we assume Γ is non-amenable, then by (Theorem 3.2.4 in [P86]; see also the direct VANISHING COHOMOLOGY 37 argument in the proof of Theorem 5.2 above) the von Neumann algebra generated by B and its normalizer in M = L(FS) is non-amenable, contradicting the strong solidity of the free group factors ([OP07]). (cid:3) 5.3. Remark. The dichotomy amenable/non-amenable in the above results can probably be extended to cover the converse to Theorem 2.10 as well. Thus, it should be true that if G is a non-amenable standard λ-lattice, then there exists an extremal inclusion of separable II1 factors N ⊂ M with standard invariant equal to G in which one cannot embed with non-degenerate strongly smooth commuting squares any inclusion of hyperfinite II1 factors Q ⊂ R (as before, strongly smooth commuting square inclusion (Q ⊂ R) ⊂ (N ⊂ M ) means that it is non-degenerate and satisfies Q′ ∩ Rn = Q′ ∩ Mn = N ′ ∩ Mn, R′ ∩ Rn = R′ ∩ Mn = M ′ ∩ Mn, ∀n). Similarly, it should be true that if G is a non-amenable countable rigid C∗-tensor category, then G admits an action (by endomorhisms) on a II∞ factor M ∞ which has no hyperfinite II∞ subfactors R∞ with normal expectation that are left invariant by G (modulo inner perturbations) and on which G acts outerly as a C∗-tensor category. In both statements, the obvious candidate for a proof is the canonical inclusion N G(L(F∞)) ⊂ M G(L(F∞)), from ([P94a]), which has L(F∞) as "initial data" and G as standard invariant. Note that the resulting factors N = N G(L(F∞)), M = M G(L(F∞)) were in fact shown to be isomorphic to L(F∞) in ([PS01]). If one assumes by contradiction that there does exist a hyperfinite inclusion Q ⊂ R with G as standard invariant and which can be embedded with strongly smooth com- muting square into N ⊂ M and one takes the associated SE inclusions, then de- formation/rigidity arguments in the style of the proofs of 5.1 and 5.2 above should contradict the non-amenability of G. A study case is when G is the Temperley-Lieb- Jones λ-lattice Gλ of index λ−1 > 4. One difficulty in proving such a result is that so far (relative) strong solidity results can only say something about normalizers of diffuse AFD von Neumann subalgebras, while in the case of an acting standard λ-lattice G (or of an acting rigid C∗-tensor category) one generally has to deal with quasi-normalizers (see [BHV15] for related results). 6. Vanishing cohomology and Connes Embedding conjecture In this section, we'll show that Connes Approximate Embedding (CAE) conjec- ture for factors of the form R ⋊ Γ can be reformulated as a vanishing cohomology problem for a certain cocycle action of Γ. Thus, let ω be an (arbitrary) non-principal ultrafilter on N and denote by Rω the corresponding ω-ultrapower II1 factor, with R ⊂ Rω viewed as constant sequences. 6.1. Proposition. 1◦ Rω := R′ ∩ Rω is a II1 factor whose centralizer in Rω is 38 SORIN POPA equal to R, i.e., R′ ω ∩ Rω = R. 2◦ Given any θ ∈ Aut(R) there exists a unitary element Uθ ∈ NRω (R) such that θ)R = θ, then θ = vUθ = Uθv′ for some v, v′ ∈ U(Rω). Moreover, if U ∈ NRω (R) and one Ad(Uθ)R = θ. If U ′ U ′ denotes θ = Ad(U )R ∈ Aut(R), then U ∈ U(Rω)Uθ. θ ∈ NRω (R) is another unitary satisfying Ad(U ′ 3◦ If θ, Uθ are as in 2◦ above, then Ad(Uθ)Rω implements an element θω ∈ Out(Rω) and an element θω = Ad(Uθ)R∨Rω ∈ Out(R ∨ Rω), with θ ∈ Aut(R) outer iff θω outer and iff θω outer. 4◦ The application Out(R) ∋ θ 7→ θω ∈ Out(R ∨ Rω) is a 1 to 1 group morphism whose image has trivial scalar 3-cocycle, with corresponding cocycle crossed product II1 factor (R ∨ Rω) ⋊ Out(R) equal to the von Neumann algebra generated in Rω by R ∨ Rω and {Uθ θ ∈ Aut(R)} (thus equal to NRω (R)′′ as well). 5◦ Any free action Γ yσ R gives rise to a free cocycle action σω of Γ on R∨Rω, by ω : Γ × Γ → U(Rω). σω(g) = Ad(Uσ(g))R∨Rω , g ∈ Γ, with corresponding 2-cocycle vσ Proof. 1◦ This is a particular case of (Theorem 2.1 in [P13]). 2◦ This is well known (see e.g., [C74]) and is due to the fact that any automor- phism of R is approximately inner. 3◦ Since Uθ normalizes R, it also normalizes its relative commutant R′∩Rω = Rω, and therefore R ∨ Rω as well. If the automorphism θω it implements on Rω is inner, say implemented by some v ∈ U(Rω), then v∗Uθ ∈ R′ ω ∩ Rω = R, implying that Ad(Uθ) is inner on R, i.e., θ is inner. Similarly, if Ad(Uθ) is inner on R, then it is inner on Rω. Since R ∨ Rω ≃ R⊗Rω with Ad(Uθ) splitting as a tensor product of its restrictions to R, Rω, one also has that this automorphism is inner iff both restrictions are inner. 4◦ The II1 factor R∨Rω has trivial relative commutant in Rω and so if we denote by N the unitaries in its normalizer that leave R (and thus also Rω) invariant, then G = N /U(R)U(Rω) is a discrete group implementing a cocycle action on R ∨ Rω, with (R ∨ Rω) ∨ N ≃ (R ∨ Rω) ⋊ G. Also, from the construction of the map of Aut(R) ∋ θ 7→ Uθ ∈ N and part 3◦, we see that this map implements an isomorphism Out(R) ≃ G. 5◦ This part is trivial from 3◦ above. (cid:3) 6.2. Definition. A II1 factor M (respectively a group Γ) has the CAE property if it can be embedded into Rω (respectively into the unitary group of Rω). Note that by a result in [R06], Γ has a faithful representation into U(Rω) iff Rω contains a copy of the left regular representation of Γ, equivalently L(Γ) ֒→ Rω. Thus, Γ has the CAE property iff L(Γ) has the CAE property. VANISHING COHOMOLOGY 39 6.3. Theorem. Let Γ yσ R be a free action of a countable group Γ on the hyperfinite II1 factor R. The II1 factor R ⋊σ Γ has the CAE property if and only if the U(Rω)-valued 2-cocycle vσ ω vanishes, i.e., iff there exist unitary elements {Ug g ∈ Γ} ⊂ NRω (R) that implement σ on R and satisfy UgUh = Ugh, ∀g, h ∈ Γ. Proof. Let M = R ⋊σ Γ with {Ug g ∈ Γ} denoting the canonical unitaries If M is embeddable into Rω, then by using the fact that any implementing σ. two copies of the hyperfinite II1 factor in Rω are conjugated by a unitary element in Rω, it follows that we may assume the hyperfinite subfactor R in M = R ⋊ Γ coincides with the algebra of constant sequences in Rω, with the action σ on it being implemented by {Ug}g ⊂ M ⊂ Rω. By Proposition 6.1.5◦ above, this means the 2-cocycle vσ ω vanishes. Conversely, if vσ ω vanishes, then we clearly have R ⋊σ Γ ֒→ Rω. (cid:3) 6.4. Corollary. Let Γ be a countable group and Γ yσ R⊗Γ the non-commutative Bernoulli Γ-action with base R. Let H be an ICC amenable group (such as the group S∞ of finitely supported permutations of N, or the lamp-lighter group Z/2Z ≀ Z). Then H ≀ Γ is a CAE group iff vσ ω vanishes. 6.5. Remarks. 1◦ It has been shown in [HaS16] that if two groups H, Γ are sofic, then their wreath product H ≀ Γ is sofic as well, so in particular it is CAE. Taking H to be an (arbitrary) amenable ICC group H, for which by Connes Theorem one has L(H) ≃ R, it follows that the crossed product II1 factor R ⋊σ Γ = L(H ≀ Γ) is CAE, where Γ yσ R⊗Γ ≃ R is the non-commutative Bernoulli Γ-action with base ≃ R as in 6.4. Thus, the corresponding cocycle vσ ω vanishes. Equivalently, one can choose Uσ(g) ∈ NRω (R) so that to be a representation of Γ. One can in fact show that these unitaries can be taken so that to also normalize the ultrapower of the Cartan subalgebra, i.e., {Ug}g ⊂ NRω (R) ∩ NRω (Dω), where D = D⊗Γ , D0 being the Cartan subalgebra of the base. 0 2◦ Given any Γ ∈ VC, the wreath product group S∞ ≀ Γ is CAE by Corollary 6.4 above, and thus Γ is a CAE group. However, one already knows this, since we have seen in Section 4 that VC is contained in W∗ leq(F2), and L(F2) ֒→ Rω. But while the class VC has a lot of restrictions on it (cf. Sections 3 and 4 in this paper), the class of CAE groups is manifestly huge, in fact it may well be that all groups are CAE. 3◦ Part 4◦ of Proposition 6.1 above naturally leads to the following: def = θω of Out(R) on R ∨ Rω have 6.6. Problem. Does the cocycle action θ 7→ π(θ) vanishing 2-cohomology? Can π be perturbed by inner automorphisms to a genuine action? Is it true that H 2(Out(R)) = 1 ? 40 SORIN POPA Due to its "huge size" and properties (e.g., all torsion free elements are conju- gate in Out(R), by [C74]), one should have H 3(Out(R)) = 1. If this is the case, then perturbing the cocycle action π by inner automorphisms to a genuine action would be equivalent to untwisting its 2-cocyle. If the CAE conjecture turns out to hold true, then Theorem 6.3 implies that the restriction of π above to any count- able subgroup Γ ⊂ Out(R) that implements a genuine action on R has vanishing cohomology. Nevertheless, even if this is the case, the entire cocyle action π cannot probably be untwisted to a genuine action. One way to prove this would be to show that its restriction to a certain countable subgroup cannot be untwisted. This amounts to saying that some free cocycle action (θ, vθ) of a countable group Γ on R cannot be untwisted when viewed as an action on R ∨ Rω. So to start with, this means (θ, vθ) cannot be untwisted as a cocycle action on R. But the only exemples of cocycle actions (θ, vθ) on R that we know to be "un- twistable" are the ones provided by Theorem 3.2, which are amplifications (σt, vt) of the Bernoulli Γ-actions Γ yσ R⊗Γ (as defined in [P01a]). But if π can be untwisted on σ(Γ), then all these cocycle actions can actually be untwisted on R∨Rω. Indeed, this is because any countable subgroup in the normalizer of R ∨ Rω in Rω has a "huge" non-separable type II1 fixed point algebra. As noted in [P01a], if a genuine action has II1 fixed point algebra, then all its amplifications can be untwisted. Thus, if π can be untwisted when restricted to σ(Γ) ⊂ Out(R) then it can also be untwisted when restricted to σt(Γ) ⊂ Out(R). All this shows that the Problem 6.6. may lead to some interesting logic-related considerations, especially when taken together with the CAE conjecture. References [AP17] C. Anantharaman, S. Popa: "An introduction to II1 factors", www.math.ucla.edu/∼popa/Books/ [Bi97] D. Bisch: Bimodules, higher relative commutants and the fusion algebra associ- ated to a subfactor, The Fields Institute for Research in Mathematical Sciences Communications Series, Vol. 13 (1997), 13-63. [BoHV15] R. Boutonnet, C. Houdayer, S. Vaes: Strong solidity of free Araki-Woods factors, to appear in Amer. J. Math, arXiv:1512.04820 [BrO08] N. Brown, N. Ozawa: "C∗-algebras and finite dimensional-approximations", Amer. Math. Soc. Grad. Studies in Math. 88, 2008. [C74] A. Connes: Outer conjugacy classes of automorphisms of factors, Ann. Ecole Norm. Sup., 8 (1975), 383-419. [C75] A. Connes: Classification of injective factors, Ann. of Math., 104 (1976), 73-115. VANISHING COHOMOLOGY 41 [C80] A. Connes: Correspondences, handwritten notes, 1980. [CJ84] A. Connes, V.F.R. Jones: Property T for von Neumann algebras, Bull. London Math. Soc. 17 (1985), 57-62. [CT76] A. Connes, M. Takesaki: The flow of weights of factors of type III, Tohoku Math. Journ. 29 (1977), 473-575. [D92] K. Dykema: Interpolated free group factors, Pac. J. Math. 163 (1994), 123-135. [G04] D. Gaboriau: Examples of groups that are measure equivalent to the free group, Ergodic Theory Dynam. Systems 25 (2005), 1809-1827. [GL07] D. Gaboriau, R. Lyons: A Measurable-Group-Theoretic Solution to von Neu- mann's Problem, Invent. Math., 177 (2009), 533-540. [H79] U. Haagerup: An example of a non-nuclear C∗-algebra which has the metric approximation property, Invent. Math. 50 (1979), 279-293. [HaS16] B. Hayes, A. Sale: The wreath product of two sofic groups is sofic, arXiv:1601. 03286 [Hj04] G. Hjorth: A lemma for cost attained, Ann. Pure Applied Logic bf 143 (2006), 87-102. [I06] A. Ioana: Rigidity results for wreath product II1 factors, J. Funct. Anal. 252 (2007), 763-791. [I13] A. Ioana: Cartan subalgebras of amalgamated free product II1 factors. (Appendix by A. Ioana and S. Vaes), Ann. Sci. Ec. Norm. Super. 48 (2015), 71-130. [IPeP05] A. Ioana, J. Peterson, S. Popa: Amalgamated Free Products of w-Rigid Factors and Calculation of their Symmetry Groups, Acta Math. 200 (2008), 85-153. [J80] V. F. R. Jones: Actions of finite groups on the hyperfinite type II1 factor, Mem. Amer. Math. Soc., 237, 1980. [J81] V. F. R. Jones: A converse to Ocneanu's theorem, Journal of Operator Theory 10 (1983), 61-64. [J83] V.F.R. Jones: Index for subfactors, Invent. Math. 72 (1983), 1-25. [J99] V.F.R. Jones: Planar Algebras, math.OA/9909027 [Ju06] K. Jung: Strongly 1-bounded von Neumann algebras, Geom. Funct. Anal. 17 (2007), 1180-1200. [KV83] V. A. Kaimanovich, A. M. Vershik: Random Walks on Discrete Groups: Bound- ary and Entropy, Annals of Probability, 11 (1983), 457-490. [KT02] Y. Katayama, M. Takesaki: Outer actions of a countable discrete amenable group on an AFD factor, in "Advances in quantum dynamics" (South Hadley, MA, 2002), 163-171, Contemp. Math., 335, AMS, Providence, RI, 2003. [NT59] M. Nakamura, Z. Takeda: On the extension of finite factors I, II, Proc. Japan Acad., 44 (1959), 149-154, 215-220. [NeTu13] S. Neshveyev, L. Tuset: "Compact Quantum Groups and their Representation Categories", Cours Sp´ecialis´e, Vol 20, Soci´et´e Math. de France, Paris (2013). 42 SORIN POPA [Oc85] A. Ocneanu: "Actions of discrete amenable groups on factors", Springer Lecture Notes No. 1138, Berlin-Heidelberg-New York 1985. [O03] N. Ozawa: Solid von Neumann algebras, Acta Math. 192 (2004), 111-117. [OP07] N. Ozawa, S. Popa: On a class of II1 factors with at most one Cartan subalgebra, Annals of Mathematics 172 (2010), 101-137 (math.OA/0706.3623) [Pe07] J. Peterson: L2-rigidity in von Neumann algebras, Invent. math. 175 (2009), 417-433. [PeT10] J. Peterson, A. Thom: Group cocycles and the ring of affiliated operators, Invent. Math. 185 (2011), 561-592. [PiP84] M. Pimsner, S. Popa: Entropy and index for subfactors, Ann. Sci. Ec. Norm. Super. 19 (1986), 57-106. [PiP88] M. Pimsner, S. Popa: Iterating the basic construction, Trans. AMS, 310 (1988), 127-134. [P81] S. Popa: Maximal injective subalgebras in factors associated with free groups, Advances in Math., 50 (1983), 27-48. [P83] S. Popa, Hyperfinite subalgebras normalized by a given automorphism and related problems, in "Proceedings of the Conference in Op. Alg. and Erg. Theory" Busteni 1983, Lect. Notes in Math., Springer-Verlag, 1132, 1984, pp 421-433. [P86] S. Popa: Correspondences, INCREST Preprint 56/1986, www.math.ucla.edu/ ∼popa/preprints.html [P89] S. Popa: Sous-facteurs, actions des groupes et cohomologie, C.R. Acad. Sci. Paris 309 (1989), 771-776. [P91] S. Popa: Classification of amenable subfactors of type II, Acta Mathematica, 172 (1994), 163-255. [P93] S. Popa: Approximate innerness and central freeness for subfactors: A classifi- cation result, in "Subfactors", (Proc. Tanegouchi Symposium in Operator Alge- bras), Araki-Kawahigashi-Kosaki Editors, World Scientific 1994, pp 274-293. [P94a] S. Popa: An axiomatization of the lattice of higher relative commutants of a subfactor, Invent. Math., 120 (1995), 427-445. [P94b] S. Popa, Symmetric enveloping algebras, amenability and AFD properties for subfactors, Math. Res. Letters, 1 (1994), 409-425. [P97] S. Popa: Some properties of the symmetric enveloping algebras with applications to amenability and property T, Documenta Mathematica, 4 (1999), 665-744. [P01a] S. Popa: Some rigidity results for non-commutative Bernoulli shifts, J. Fnal. Analysis 230 (2006), 273-328 (MSRI preprint No. 2001-005). [P01b] S. Popa: On a class of type II1 factors with Betti numbers invariants, Ann. of Math 163 (2006), 809-899 (math.OA/0209310; MSRI preprint 2001-024). [P03] S. Popa: Strong Rigidity of II1 Factors Arising from Malleable Actions of w- Rigid Groups I, Invent. Math., 165 (2006), 369-408. VANISHING COHOMOLOGY 43 [P06a] S. Popa: On the superrigidity of malleable actions with spectral gap, J. Amer. Math. Soc. 21 (2008), 981-1000 (math.GR/0608429). [P06b] S. Popa: On Ozawa's Property for Free Group Factors, Int. Math. Res. No- tices (2007) Vol. 2007, article ID rnm036, 10 pages, doi:10.1093/imrn/rnm036 published on June 22, 2007 (math.OA/0608451) [P13] S. Popa: Independence properties in subalgebras of ultraproduct II1 factors, Jour- nal of Functional Analysis 266 (2014), 5818-5846 (math.OA/1308.3982) [PS01] S. Popa, D. Shlyakhtenko: Universal properties of L(F∞) in subfactor theory, Acta Mathematica, 191 (2003), 225-257. [PSV15] S. Popa, D. Shlyakhtenko, S. Vaes: Cohomology and L2-Betti numbers for sub- factors and quasi-regular inclusions, math.OA/1511.07329 , to appear in the International Mathematical Research Notices. [PV12] S. Popa, S. Vaes: Unique Cartan decomposition for II1 factors arising from arbitrary actions of free groups, Acta Mathematica, 194 (2014), 237-284 [PV14] S. Popa, S. Vaes: Representation theory for subfactors, λ-lattices and C∗-tensor categories, Commun. Math. Phys. 250 (2015), 1239-1280. [R91] F. Radulescu: The weak closure of the group algebras associated to free groups are stably isomorphic, Comm. Math. Physics 156 (1993), 17-36. [R06] F. Radulescu: The von Neumann algebra of the non-residually finite Baumslag group ha, b ab3a−1 = b2i embeds into Rω. Hot topics in operator theory, 173- 185, Theta Ser. Adv. Math., 9, Theta, Bucharest, 2008. [Su80] C.E. Sutherland: Cohomology and extensions of von Neumann algebras , Publ. RIMS, 16 (1980), 135-176. [T79] M. Takesaki: "Theory of operator algebras" I., Springer-Verlag, New York- Heidelberg, 1979. Math.Dept., UCLA, Los Angeles, CA 90095-1555, [email protected]
1612.05549	1	1612	2016-12-16T16:47:22	Construction of a new class of quantum Markov fields	[ "math.OA", "math.FA", "math.PR" ]	In the present paper, we propose a new construction of quantum Markov fields on arbitrary connected, infinite, locally finite graphs. The construction is based on a specific tessellation on the considered graph, that allows us to express the Markov property for the local structure of the graph. Our main result concerns the existence and uniqueness of quantum Markov field over such graphs.	math.OA	math	CONSTRUCTION OF A NEW CLASS OF QUANTUM MARKOV FIELDS LUIGI ACCARDI,1 FARRUKH MUKHAMEDOV,2 ∗ and ABDESSATAR SOUISSI3 Abstract. In the present paper, we propose a new construction of quantum Markov fields on arbitrary connected, infinite, locally finite graphs. The con- struction is based on a specific tessellation on the considered graph, that allows us to express the Markov property for the local structure of the graph. Our main result concerns the existence and uniqueness of quantum Markov field over such graphs. 1. Introduction One of the basic open problem in quantum probability is to develop a theory of quantum Markov fields, which are conventionally quantum Markov processes with multi-dimensional index set. Here Quantum Markov fields are noncommutative extensions of the classical Markov fields (see [4, 8, 11]). On the other hand, these quantum fields can be considered as extensions of quantum Markov chains [1, 7] to general graphs. In [3, 10] the first attempts to construct quantum analogues of classical Markov chains have been carried out. In [3] quantum Markov fields were considered over integer lattices, unfortunately there was not given any non trivial examples of such fields. In [5, 6], quantum Markov chains (fields) on the tree like graphs (like Cayley tree) have been constructed and investigated, but the proposed construc- tion does not work for general graphs. A main aim of the present paper is to provide a construction of new class of quantum Markov fields on arbitrary connected, infinite, locally finite graphs. The construction is based on a specific tessellation on the considered graphs, it allows us to express the Markov property for the local structure of the graph. Our main result is the existence and uniqueness of quantum Markov field over such graphs. We note that even in the classical case, the proposed construction gives a new ways to define Markov fields (see [13, 14]). 2. Graphs Let G = (V, E) be a ( non-oriented simple ) graph, that is, L is a nonempty set and E is identified as a subset of an ordered pairs of V , i.e. E ⊂ {{x, y} x, y ∈ E, x 6= y} Copyright 2016 by the Tusi Mathematical Research Group. Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz. ∗Corresponding author. 2010 Mathematics Subject Classification. Primary 46L53; Secondary 60J99, 46L60, 60G50, 82B10. Key words and phrases. Quantum Markov field, graph, tessellation, construction. 1 2 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI Elements of V and E are called, respectively, vertices and edges. Two vertices x and y are said to be nearest neighbors if there exist an edge joining them (i.e. {x, y} ∈ E) and we denote them by x ∼ y. For any vertex y ∈ V we denote its nearest neighbors by Ny := {x ∈ V y ∼ x} (2.1) Notice that x /∈ Nx. The set {y}∪Ny is said to be interaction domain or plaquette at y. If for every x ∈ V one has Nx < ∞ then the graph is called locally finite. An edge path or walk joining two vertices x and y is a finite sequence of edges x = x0 ∼ x1 ∼ . . . xd−1 ∼ xd = y. In this case d is the length of the edge path. The graph is said to be connected if every two disjoint vertices can be joined by an edge path. In the sequel, we assume that the graph G is infinite, connected and locally finite. Note that in this case the set V is automatically countable. Now for any nonempty Λ ⊂ V we associate its following parts: • complement: • boundary: • interior : Λc := V \ Λ ∂Λ := {x ∈ Λ ∃y ∈ Λc; x ∼ y} ◦ Λ := Λ \ ∂Λ • external boundary: ~∂Λ := {y ∈ Λc ∃x ∈ Λ; x ∼ y} • closure: Λ := Λ ∪ ~∂Λ By F we denote a net of all finite subsets of V , i.e. F := {Λ ⊂ V V < ∞} where · denotes the cardinality of a set. (2.2) (2.3) (2.4) (2.5) (2.6) (2.7) 3. Tessellations on graphs In this section we propose a tessellation on the considered graphs, which will play a key role in the construction. Therefore, the resulting quantum Markov field will depend also on the tessellation. Fix a "root" y1 ∈ V and define by induction the following sets: Having defined V0,n, put V0,1 := {y1} Vn := [y∈V0,n ({y} ∪ Ny) V0,n+1 := V0,n ∪ ~∂Vn Define the following set of vertices: V0 := [n≥1 V0,n. (3.1) (3.2) (3.3) (3.4) QUANTUM MARKOV FIELDS 3 From now on, elements of V0 will be called vertices, any other element of V belongs to some plaquette at a certain element of V0. Notice that with in this construction, for every n, the inner boundary ∂Vn of each Vn contain no vertex: ∂Vn ∩ V0 = ∅ Since V = +∞ and, by assumption, V is connected, one has Vn+1 ≥ V n + 1 ≥ Vn + 2, V0,n+1 ≥ V0,n + 1 It follows that, if Λ is any finite set, there exists N ∈ N such that Λ ⊆ VN Therefore {Vn} is an exhaustive sequence of finite subsets recovering the all the vertices set V . One can check that and ~∂Vn V0 := {y1} ∪ [n≥1 V = [y∈V0 {y} ∪ Ny (3.5) (3.6) Remark 3.1. (i) For each x ∈ V \ V0, there exists at least one y ∈ V0 such that x belongs to the plaquette at y. (ii) Each y ∈ V0 belongs to its plaquette (i.e. the plaquette {y} ∪ Ny) but no other one with center in V0. The set V0 given by (3.5) (or equivalently the family {V0,n; n = 1, 2, · · · } ) is called tessellation on the graph G. 4. Quantum Markov Fields In this section we propose a definition for backward Markov fields, for the same graph G = (V, E) with the given tessellation {V0,n : n = 1, 2, · · · }. The map (4.1) defines a bundle on V whose fiber is a finite dimensional Hilbert space Hx. Denote Ax := B(Hx), x ∈ V . Define for any finite subset Λ ⊂ V the algebra x ∈ V −→ Hx " state space on x " then one get on a canonical way, the quasi-local algebra defined as the closure of the local algebra AΛ. where F is given by (2.7). Ax. Ax AΛ :=Ox∈Λ AV :=Ox∈V AV,loc := [Λ∈F (4.2) (4.3) (4.4) 4 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI Analogously, one can define for any subset Λ′ ⊂ V , the algebra AΛ′ :=Nx∈Λ′ Ax. Notice that for Λ ⊂ Λ′ ⊂ V one can see AΛ as C ∗-subalgebra of AΛ′ through the following embedding AΛ ≡ AΛ ⊗ 1IΛ′\Λ ⊂ AΛ′ (4.5) Definition 4.1. Consider a triplet C ⊂ B ⊂ A of unital C ∗-algebras. Recall [2] that a quasi-conditional expectation with respect to the given triplet is a com- pletely positive (CP), unital linear map E : A → B such that E(ca) = cE(a), for all a ∈ A, c ∈ C. We give the definition of general of backward quantum Markov field which is independent of the tessellation. Definition 4.2. A state ϕ on AV is said to be backward quantum Markov field if for any sequence {Λn}∞ n=0 of finite subsets of V satisfying Λn ⊂⊂ Λn+1, there exists a pair (ϕΛ0, {EΛn,Λn+1}∞ n=0}) with ϕΛ0 is a state on AΛ0 and EΛn,Λn+1 is a quasi-conditional expectation with respect to the triplet AΛn ⊂ A ¯Λn ⊂ AΛn+1 such that ϕ = lim n→∞ ϕΛ0 ◦ EΛ0,Λ1 ◦ · · · ◦ EΛn,Λn+1 (4.6) where the limit is taken in the weak--topology. Remark 4.3. In Definition 4.2, the condition Λn ⊂⊂ Λn+1 for every n ∈ N implies that Λn ↑ V and the limit state obtained by the right side of the equation (4.6) is defined on the full algebra AV , If ϕ is a backward quantum Markov field in the sense of Definition 4.2, then it satisfy Definition 4.2 of [5] for any increasing sequence {Λn}∞ n=0 of finite subsets of V such that ¯Λn = Λn+1, to get such a sequence of subsets, we consider Λ0 ∈ F , and for ∈ n ≥ 1 put Λn = ¯Λn−1. Clearly one has Λn ⊂⊂ Λn+1 and Λn ↑ V . Now we introduce a class of backward quantum Markov field that depends on the tessellation {V0,n, n = 1, 2, · · · } Definition 4.4. A state ϕ on AV is said to be backward quantum Markov field w.r.t. the tessellation {V0,n, n = 1, 2, · · · }, ( or V0-backward quantum Markov field ) if for any sequence {Λn}∞ n=0 of finite subsets satisfying Λn ⊂⊂ Λn+1, (4.7) there exists a pair (ϕΛ0, {EΛn,Λn+1}∞ n=0}) with ϕΛ0 is a state on AΛ0 and EΛn,Λn+1 is a quasi-conditional expectation with respect to the triplet AΛn ⊂ A ¯Λn ⊂ AΛn+1 such that ~∂Λ ∩ V0 = ∅ where the limit is taken in the weak--topology. ϕ = lim n→∞ ϕΛ0 ◦ EΛ0,Λ1 ◦ · · · ◦ EΛn,Λn+1 Now we fix the following product state ϕ0 :=Ox∈V ϕ0 x (4.8) (4.9) QUANTUM MARKOV FIELDS 5 on the algebra AV , where ϕ0 x is a state on Ax for every x ∈ V . Denote for Λ ⊂ V , ϕ0 Λ :=Ox∈Λ ϕ0 x (4.10) which is the restriction of the state ϕ0 V to AΛ. We aim to construct a quantum Markov field on the algebra AV through a perturbation of the product state ϕ0 V . 5. Construction of conditional density amplitudes It is well known [2] that quasi-conditional expectations are more convenient than Umegaki conditional expectations (see definition (5.1)) to express the non- commutative Markov property. In what follows, we will perturb ϕ-conditional expectations (see [2]) to get quasi-conditional expectations using a commuting set of operators with the considered tessellation. For any ordered pair y ∈ V0 and x ∈ Ny, let be given an operator K(x,y) ∈ A{x,y} such that it is invertible and the C ∗-subalgebra K = { K ∗ {x,y} , K{x,y} : y ∈ V0, x ∈ Ny} C ∗ (5.1) is commutative. Definition 5.1. A Umegaki conditional expectation is a norm one projection from a C ∗-algebra onto its C ∗-subalgebra. Definition 5.2. Let A1, A2 be two C ∗-algebras with units respectively I1 and I2 and let A = A1 ⊗A2. An element K ∈ A is called a conditional density amplitude w.r.t. a state ϕ on I1 ⊗ A2, if one has Eϕ(K ∗K) = I1 (5.2) where Eϕ is the Umegaki conditional expectation from A onto A1 ⊗ I2 defined by the linear extension of Eϕ(a1 ⊗ a2) = ϕ(I1 ⊗ a2)a1 ⊗ I2. (5.3) An operator K is also called a conditional density amplitude for the ϕ-conditional expectation Eϕ. For each x ∈ V by E0 the algebra A onto the algebra A{x}c defined on localized elements a =Nz∈V az = {x}c we denote the Umegaki conditional expectation from ax ⊗ a{x}c by: {x}c(ax ⊗ a{x}c) = ϕ0 E0 x(ax)a{x}c. One can prove the following facts. Lemma 5.3. For every pair of vertices (x, y) ∈ V 2 one has [E0 {x}c, E0 {y}c] = 0 (5.4) (5.5) 6 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI For any Λ ∈ F by virtue of Lemma 5.3 we define E0 Λc := Yx∈Λ E0 {x}c. (5.6) Lemma 5.4. For any Λ ∈ F the map E0 expectation from AV onto AΛc such that for aΛ ∈ AΛaΛc ∈ AΛc one has Λc given by (5.6) is a Umegaki conditional Λc(aΛ ⊗ aΛc) = ϕ0 E0 Λ(aΛ)aΛc (5.7) Remark 5.5. The map E0 Λc can be defined, through the equation (5.7), for an arbitrary part (not necessarily finite) Λ of V and it is still a Umegaki conditional expectation from AV onto AΛc Proposition 5.6. Let y ∈ V0, the operator is invertible. BNy := E0 {y}c (cid:12)(cid:12)(cid:12)(cid:12) Yx∈Ny Proof. Let us consider B{y}∪Ny :=(cid:12)(cid:12)(cid:12)Qx∈Ny ator B{y}∪Ny is positive definite, then σ(B{y}∪Ny ) ⊂]0, keK{y}∪Ny k]. Since the spec- ∈ A{y}∪Ny and denote its spec- trum by σ(B{y}∪Ny ), which is a closed subset of the complex field. Since the oper- 2 K{x,y}(cid:12)(cid:12)(cid:12)(cid:12)  ∈ ANy K{x,y}(cid:12)(cid:12)(cid:12) trum does not contain zero then there is ε > 0 such that σ(B{y}∪Ny ) ⊂ [ε, kBk], therefore B{y}∪Ny ≥ ε1I. This yields E0 {y}c is invertible. (cid:3) {y}c(B) ≥ ε1I, which means BNy = E0 (5.8) 2 In the sequel, we assume that for every y ∈ V0 the operator BNy belongs to the commutant K′ (w.r.t. AV ) of the algebra K (see (5.1)). Note that under this condition the operators B±1/2 also belong to K′. Ny Lemma 5.7. The operator is a ϕ0 {y}-conditional density amplitude in the algebra A{y}∪Ny . Proof. Using the commutativity of the algebra K we obtain E0 {y}c(cid:16)K ∗ {y}∪Ny K{y}∪Ny(cid:17) = E0 K{y}∪Ny := Yx∈Ny eK{x,y} {y}c Ny  Yx∈Ny B−1/2 {y}c Yx∈Ny )∗E0 Ny )∗BNy B−1/2 Ny = (B−1/2 = (B−1/2 = 1I. Ny Ny  B−1/2 (5.9) K{x,y} Ny   B−1/2  K{x,y} ∗ Yy∈Ny  K{x,y}  B−1/2 K ∗ {x,y} Ny This completes the proof. (cid:3) QUANTUM MARKOV FIELDS 7 Now, for each Λ ∈ F , we define ~∂0Λ := [y∈∂Λ∩V0 Ny (5.10) By construction the family {K ∗ {y}∪Ny , K{y}∪Ny : x ∼ y ∈ V0} is commutative, therefore the following operator is well defined KΛ∪~∂0Λ := Yy∈Λ∩V0 K{y}∪Ny ∈ AΛ∪~∂0Λ ⊆ A ¯Λ for every Λ ∈ F . (5.11) Remark 5.8. 1. In general, it is possible that AΛ∪~∂0Λ is a proper sub-algebra of A ¯Λ. Since, by construction of the tessellation, the set Λ ∪ ~∂0Λ cannot contain elements of V0. 2. If Λ ∩ V0 = ∅, we convent that KΛ∪~∂0Λ = 1I. Theorem 5.9. For any Λ ∈ F , the operator KΛ∪~∂0Λ defined by (5.11) is a con- ditional density amplitude for the Umegaki conditional expectation E0 (Λ∩V0)c. Proof. Since the family {K{y}∪Ny , K ∗ one can write {y}∪Ny K ∗ Λ∪~∂0ΛKΛ∪~∂0Λ = Yy∈Λ∩V0 y ∈ Λ ∩ V0} is commutative, then K ∗ {y}∪Ny K{y}∪Ny (5.12) and using the following property of the tessellation: for disjoint elements y and z of V0 the plaquette at y does not contain z, we conclude that K{y}∪Ny is localized in {z}c. Then by Lemma 5.4 one gets {z}c(K ∗ E0 {y}∪Ny K{y}∪Ny ) = K ∗ {y}∪Ny K{y}∪Ny then after a small iteration, we obtain E0 (Λ∩V0)c(cid:16)K ∗ Λ∪~∂0ΛKΛ∪~∂0Λ(cid:17) = Yy∈Λ∩V0 E0 {y}c(K ∗ {y}∪Ny K{y}∪Ny ). By Lemma 5.7, one has E0 {y}c (K ∗ {y}∪Ny K{y}∪Ny ) = 1I, hence one gets E0 (Λ∩V0)c(cid:16)K ∗ Λ∪~∂0ΛKΛ∪~∂0Λ(cid:17) = 1I. The following auxiliary results can be easily proved. Lemma 5.10. For every Λ1 ⊂ Λ2 ⊂ V , one has: E0 Λ1 ◦ E0 Λ2 = E0 Λ1 Lemma 5.11. For Λ, Λ′ ⊂f in V with ¯Λ ∩ Λ′ = ∅, one has: K(Λ∪Λ′)∪~∂(Λ∪Λ′) = KΛ∪ ~∂0ΛKΛ′∪~∂0Λ′ Theorem 5.12. For Λ0 ⊆ ¯Λ0 ⊂ Λ, one has: (cid:3) (5.13) (5.14) 8 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI (i) For z ∈ V0 ∩ (Λ \ ¯Λ0) {z}c(K ∗ E0 Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ) = K ∗ (Λ\{z})∪~∂0(Λ\{z})aΛ0K(Λ\{z})∪~∂0(Λ\{z}) for every aΛ0 ∈ AΛ0; (ii) (Λ\ ¯Λ0)∩V0(K ∗ E0 Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ) = K ∗ ¯Λ0∪~∂0(¯Λ0)aΛ0K ¯Λ0∪~∂0(¯Λ0) (5.15) (5.16) for every aΛ0 ∈ AΛ0. Proof. (i) For a general Λ0, if z ∈ (Λ\ ¯Λ0)∩V0, then Nz can intersect ~∂Λ0, but not Λ0. Therefore, K{z}∪Nz and aΛ0 are localized on disjoint parts, so they commute. Due to the commutativity of {K{y}∪Ny , K ∗ y ∈ Λ ∩ V0} it follows from (5.14) that {y}∪Ny E0 {z}c(K ∗ Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ) K ∗ {y}∪Ny aΛ0 Yy∈Λ∩V0 K{y}∪Ny ) = E0 = E0 = E0 {z}c( Yy∈Λ∩V0 {z}c(cid:16)(K ∗ {z}c(cid:0)K ∗ {z}∪Nz and by Lemma 5.7 one has {z}∪Nz K{y}∪Ny ) × (K ∗ (Λ\{z})∪~∂0(Λ\{z})aΛ0K(Λ\{z})∪~∂0(Λ\{z}))(cid:17) (Λ\{z})∪~∂0(Λ\{z})aΛ0K(Λ\{z})∪~∂0(Λ\{z}) K{y}∪Ny(cid:1) K ∗ {z}c(cid:0)K ∗ E0 {z}∪Nz K{y}∪Ny(cid:1) = 1I Hence, we get {z}c(K ∗ E0 Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ) = K ∗ (Λ\{z})∪~∂0(Λ\{z})aΛ0K(Λ\{z})∪~∂0(Λ\{z}) (ii) Iterating the procedure of (5.15) to cover all z ∈ (Λ \ ¯Λ0) ∩ V0 one finds E0 (Λ\Λ0)∩V0(cid:16)K ∗ Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ(cid:17) = (cid:18) Yz∈(Λ\ ¯Λ0)∩V0 E0 {z}c(cid:19)(cid:16)K ∗ Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ(cid:17) = K ∗ ¯Λ0∪~∂0(¯Λ0)aΛ0K ¯Λ0∪~∂0(¯Λ0) This completes the proof. (cid:3) Remark 5.13. Keeping the notations of Theorem 5.12, if ~∂Λ0 ∩ V0 = ∅ then using the same argument, one gets E0 Λ\Λ0(K ∗ Λ∪~∂0ΛaΛ0KΛ∪~∂0Λ) = K ∗ Λ0∪~∂0Λ0 aΛ0KΛ0∪~∂0Λ0 for every a0 ∈ AΛ0. 6. Main result In this section, we prove a main result of the paper. First we need an auxiliary result. QUANTUM MARKOV FIELDS 9 Proposition 6.1. Let Λ1, Λ2 ∈ F with Λ1 ⊂⊂ Λ2. Define EΛ1,Λ2(a) = E0 (Λ2\ ¯Λ1)c(cid:16)K ∗ (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)aK(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) for a ∈ AV . Then EΛ1,Λ2 is a quasi-conditional expectation w.r.t. AΛ1 ⊂ A ¯Λ1 ⊂ AΛ2. (6.1) the triplet Proof. The map EΛ1,Λ2 is clearly linear and valued in A ¯Λ1. Unitality: using commutativity of the family {E{z}c z ∈ (Λ2 \ Λ1) ∩ V0} (by Lemma 5.3), one can write 0 )c ◦ E0 and using Theorem 5.9 for Λ = Λ2 \ ¯Λ1 we obtain (Λ2\Λ1)c = E0 E0 ((Λ2\ ¯Λ1)∩V c ((Λ2\Λ1)∩V0)c then using (6.2) one finds E0 (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)K(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) = 1I ((Λ2\ ¯Λ1)∩V0)c(cid:16)K ∗ (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)K(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) = E0 ((Λ2\ ¯Λ1)∩V c E0 (Λ2\ ¯Λ1)c(cid:16)K ∗ 0 )c(1I) = 1I (6.2) hence, Complete positivity: One can check that for any y ∈ V0 the map EΛ1,Λ2(1I) = 1I. a 7→ E{y}c(a) := E0 {y}c (K ∗ {y}∪Ny aK{y}∪Ny ) is completely positive. Now using the commutativity of the set {K{y}∪Ny , y ∈ V0} one gets EΛ1,Λ2 = E0 (Λ2\ ¯Λ1)∩V0 ◦(cid:18) Yy∈(Λ2\ ¯Λ1)∩V0 E{y}c(cid:19). Hence EΛ1,Λ2 is the composition of completely positive maps, therefore, it is so. ∈ A¯(Λ2\Λ1) then it commutes Let a ∈ AΛ2, c ∈ AΛ1, while K ∗ (Λ2\Λ1)∪~∂0(Λ2\Λ1) with c, then using the fact that (Λ2\Λ1)c(cd) = cE0 E0 (Λ2\Λ1)c(d) for every d ∈ A, one gets EΛ1,Λ2(ca) = E0 (Λ2\ ¯Λ1)c(cid:16)K ∗ (Λ2\ ¯Λ1)c(cid:16)cK ∗ (Λ2\ ¯Λ1)c(cid:16)K ∗ (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)caK(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)aK(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) (Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)aK(Λ2\ ¯Λ1)∪~∂0(Λ2\ ¯Λ1)(cid:17) = E0 = cE0 = cEΛ1,Λ2(a). Hence, EΛ1,Λ2 is a quasi-conditional expectation w.r.t. the given triplet. This completes the proof. (cid:3) Now we pass to our main result. 10 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI Theorem 6.2. For each Λ ∈ F define the state eϕΛ on A by Then the net {eϕΛ}Λ∈F converges in the weak-*-topology, moreover the limiting state ϕ is a backward Markov field on AV w.r.t. the tessellation V0. eϕΛ(a) := ϕ0(K ∗ Λ∪~∂0ΛaKΛ∪~∂0Λ) Proof. First we prove the existence of the limit. Due to the density argument, it is sufficient to establish the existence of the limit in the local algebra AV,loc. (6.3) Let a ∈ AV,loc then a ∈ AΛ0 for some Λ0 ∈ F . For Λ ∈ F with Λ0 ⊂⊂ Λ, we and by Theorem 5.12 one gets = ϕ0 ◦ E eϕΛ(a) = ϕ0(cid:16)K ∗ (Λ\ ¯Λ0)c(cid:16)K ∗ Λ∪~∂0ΛaKΛ∪~∂0Λ(cid:17) (Λ\ ¯Λ0)c(cid:16)K ∗ Λ∪~∂0ΛaKΛ∪~∂0Λ(cid:17) = K ∗ E Λ∪~∂0ΛaKΛ∪~∂0Λ(cid:17) ¯Λ0∪~∂0 ¯Λ0 aK ¯Λ0∪~∂0 ¯Λ0, have so ¯Λ0∪~∂0 ¯Λ0 aK ¯Λ0∪~∂0 ¯Λ0 ) eϕΛ(a) = ϕ0(K ∗ = eϕ ¯Λ0(a). As Λ → V , we find that Λ0 ⊂⊂ Λ up to some order, hence the net {eϕ(a)}Λ∈F ;Λ0⊂⊂Λ is stationary. This means that (6.4) lim Λ→V ;Λ0⊂⊂ΛeϕΛ(a) = ϕ ¯Λ0(a) =: ϕ(a) therefor the limit exist on the local algebra, and yet it exists on the full algebra AV . Now we establish that the state ϕ is a quantum Markov field. Let {Λn n ∈ N}n∈N be a family of subset of F satisfying Λn ⊂⊂ Λn+1, ~∂Λn ∩ V0 = ∅ Let EΛn,Λn+1 be given by (6.1). Then, for a ∈ AVn, we have eϕΛn ◦ EΛn,Λn+1(a) Λn∪~∂0Λn = ϕ0(cid:16)K ∗ = ϕ0(cid:16)K ∗ EΛn,Λn+1(a)KΛn∪~∂0Λn(cid:17) (Λn+1\ ¯Λn)c(cid:16)K ∗ E0 Λn∪~∂0Λn (Λn+1\ ¯Λn)∪~∂0(Λn+1\ ¯Λn)aK(Λn+1\ ¯Λn)∪~∂0(Λn+1\ ¯Λn)(cid:17) KΛn∪~∂0Λn(cid:17) Since KΛn∪~∂0Λn expectation from AV onto A(Λn+1\ ¯Λn)c then one finds ∈ A ¯Λn ⊂ A(Λn+1\ ¯Λn)c and E0 (Λn+1\ ¯Λn)c is a Umegaki conditional eϕΛn ◦ EΛn,Λn+1(a) (Λn+1\ ¯Λn)c(cid:16)K ∗ = ϕ0E0 Λn∪~∂0Λn K ∗ (Λn+1\ ¯Λn)∪~∂0(Λn+1\ ¯Λn)aK(Λn+1\ ¯Λn)∪~∂0(Λn+1\ ¯Λn)KΛn∪~∂0Λn(cid:17) and by the assumption (4.7) one has ~∂Λn ∩ V0 = ∅ QUANTUM MARKOV FIELDS 11 then ¯Λn ∩ V0 = Λn ∩ V0 and KΛn∪~∂0Λn = Yy∈Λn∩V0 K{y}∪Ny = Yy∈ ¯Λn∩V0 K{y}∪Ny = K ¯Λn∪~∂0 ¯Λn . From Lemma 5.14 it follows that KΛn+1∪~∂0Λn+1 = K ¯Λn∪ ¯Λ1K(Λn+1\ ¯Λn)∪~∂0(Λn+1\ ¯Λn) then we obtain eϕΛn ◦ EΛn,Λn+1(a) = ϕ0 ◦ E0 Hence, by construction one gets ϕ0 V = ϕ0 V ◦ E (Λn+1\ ¯Λn)c (Λn+1\ ¯Λn)c(K ∗ Λn+1∪~∂0Λn+1 aKΛn+1∪~∂0Λn+1 ) so eϕΛn ◦ EΛn,Λn+1(a) = ϕ0 Now iterating the equation (6.5), we obtain V (K ∗ Λn+1∪~∂0Λn+1 aKΛn+1∪~∂0Λn+1 ) = eϕΛn+1(a) (6.5) eϕn = eϕΛ0 ◦ EΛ0,Λ1 ◦ · · · ◦ EΛn−1,Λn ϕV = lim ϕΛ0 ◦ EΛ0,Λ1 ◦ · · · ◦ EΛn−1,Λn . This completes the proof. (cid:3) therefore where ϕΛ0 = eϕΛ0⌈AΛ0 The provided construction allows us to produce a lot of interesting examples of quantum Markov fields on arbitrary connected, infinite, locally finite graphs. Note that the construction is based on a specific tessellation on the considered graph, it allows us to express the Markov property for the local structure of the graph. We note that even in the classical case, the proposed construction gives other ways to define Markov fields different to the existing ones (see [13]). This construction opens new perspectives in the theory of phase transitions in the scheme of quantum Markov fields (comp. [9]). References 1. L. Accardi, On the noncommutative Markov property, Funct. Anal. Appl. 9 (1975) 1 -- 8. 2. L. Accardi, C. Cecchini, Conditional expectations in von Neumann algebras and a Theorem of Takesaki, J. Funct. Anal. 45 (1982), 245 -- 273. 3. L. Accardi, F. Fidaleo, Quantum Markov fields, Infin. Dim. Analysis, Quantum Probab. Related Topics 6 (2003) 123 -- 138. 4. R.L. Dobrushin, Description of Gibbsian Random Fields by means of conditional probabil- ities, Probab. Theory and Appl. 13(1968) 201 -- 229. 5. L. Accardi, H. Ohno, F. Mukhamedov, Quantum Markov fields on graphs, Infin. Dim. Analysis, Quantum Probab. Related Topics 13(2010), 165 -- 189. 6. M. Fannes, B. Nachtergaele, R.F. Werner, Ground states of VBS models on Cayley trees, J. Stat. Phys. 66 (1992) 939 -- 973. 7. M. Fannes, B. Nachtergaele, R.F. Werner, Finitely correlated states on quantum spin chains, Commun. Math. Phys. 144 (1992) 443 -- 490. 8. H.-O. Georgi, Gibbs measures and phase transitions, de Gruyter Studies in Mathematics vol. 9, Walter de Gruyter, Berlin, 1988. 12 L. ACCARDI, F. MUKHAMEDOV and A. SOUISSI 9. F. Mukhamedov, A. Barhoumi, A. Souissi, Phase transitions for Quantum Markov Chains associated with Ising type models on a Cayley tree, J. Stat. Phys. 163 (2016) 544 -- 567. 10. V. Liebscher, Markovianity of quantum random fields, Proceedings Burg Conference 15 -- 20 March 2001, W. Freudenberg (ed.), World Scientific, QP -- PQ Series 15 (2003) 151 -- 159 11. C. Preston, Gibbs states on countable sets, Cambridge University Press, London, 1974. 12. U.A. Rozikov, Gibbs measures on Cayley trees, World Scientific, Singappore, 2013. 13. A. Spataru, Construction of a Markov field on an infinite tree, Advance in Math. 81(1990), 105 -- 116. 14. S. Zachary, Countable state space Markov random fields and Markov chains on trees, Ann. Prob. 11 (1983) 894 -- 903. 1 Centro Vito Volterra, Universita di Roma "Tor Vergata", Roma I-00133, Italy E-mail address: [email protected] 2Department of Mathematical Sciences, College of Science, The United Arab Emirates University, P.O. Box 15551, Al Ain, Abu Dhabi, UAE E-mail address: [email protected] 3.1 Department of Mathematics, Faculty of Sciences of Tunis, University of Tunis El-Manar, 1060 Tunis, Tunisia. 3.2Preparatory Institute for Scientific and Technical Studies, La Marsa, Carthage University, Tunisia0 E-mail address: [email protected]; [email protected]
1706.08833	2	1706	2018-05-04T12:46:25	Quantum Symmetries of Graph C*-Algebras	[ "math.OA", "math.FA" ]	The study of graph C-algebras has a long history in operator algebras. Surprisingly, their quantum symmetries have never been computed so far. We close this gap by proving that the quantum automorphism group of a finite, directed graph without multiple edges acts maximally on the corresponding graph C-algebra. This shows that the quantum symmetry of a graph coincides with the quantum symmetry of the graph C*-algebra. In our result, we use the definition of quantum automorphism groups of graphs as given by Banica in 2005. Note that Bichon gave a different definition in 2003; our action is inspired from his work. We review and compare these two definitions and we give a complete table of quantum automorphism groups (with respect to either of the two definitions) for undirected graphs on four vertices.	math.OA	math	QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS SIMON SCHMIDT AND MORITZ WEBER Abstract. The study of graph C ∗-algebras has a long history in operator alge- bras. Surprisingly, their quantum symmetries have never been computed so far. We close this gap by proving that the quantum automorphism group of a finite, directed graph without multiple edges acts maximally on the corresponding graph C ∗-algebra. This shows that the quantum symmetry of a graph coincides with the quantum symmetry of the graph C ∗-algebra. In our result, we use the definition of quantum automorphism groups of graphs as given by Banica in 2005. Note that Bichon gave a different definition in 2003; our action is inspired from his work. We review and compare these two definitions and we give a complete table of quantum automorphism groups (with respect to either of the two definitions) for undirected graphs on four vertices. 8 1 0 2 y a M 4 ] . A O h t a m [ 2 v 3 3 8 8 0 . 6 0 7 1 : v i X r a Introduction Symmetry constitutes one of the most important properties of a graph. It is captured by its automorphism group Aut(Γ) := {σ ∈ Sn σε = εσ} ⊆ Sn, where Γ = (V, E) is a finite graph with n vertices and no multiple edges, ε ∈ Mn({0, 1}) is its adjacency matrix, and Sn is the symmetric group. In modern mathematics, notably in operator algebras, symmetries are no longer described only by groups, but by quantum groups. In 2005, Banica [1] gave a definition of a quan- tum automorphism group of a finite graph within Woronowicz's theory of compact matrix quantum groups [19]. In our notation, G+ aut(Γ) is based on the C ∗-algebra C(G+ aut(Γ)) := C(S+ = C ∗(uij, i, j = 1, . . . , n uij = u∗ n )/huε = εui ij = u2 ij,Xl uil = 1 =Xl ulj, RBan hold), Date: May 7, 2018. 2010 Mathematics Subject Classification. 46LXX (Primary); 20B25, 05CXX (Secondary). Key words and phrases. finite graphs, graph automorphisms, automorphism groups, quantum automorphisms, graph C ∗-algebras, quantum groups, quantum symmetries. The second author was partially funded by the ERC Advanced Grant NCDFP, held by Roland Speicher. This work was part of the first author's Master's thesis. This work was also supported by the DFG project Quantenautomorphismen von Graphen. 1 2 SIMON SCHMIDT AND MORITZ WEBER where S+ n is Wang's quantum symmetric group [17] and RBan are the relations Xk uikεkj =Xk εikukj. Earlier, in 2003, Bichon [5] defined a quantum automorphism group G∗ aut(Γ) via C(G∗ aut(Γ)) := C ∗(uij, i, j = 1, . . . , n uij = u∗ ij = u2 ij,Xl uil = 1 =Xl ulj, RBic hold), where RBic are the relations Xk uikεkj =Xk εikukj, us(e)s(f )ur(e)r(f ) = ur(e)r(f )us(e)s(f ) for e, f ∈ E, and r : E → V and s : E → V are range and source maps respectively. We immediately see that Aut(Γ) ⊆ G∗ aut(Γ) ⊆ G+ aut(Γ) holds, in the sense that there are surjective ∗-homomorphisms: aut(Γ)) C(G+ uij → 7→ aut(Γ)) C(G∗ uij → 7→ C(Aut(Γ)) (σ 7→ σij) Relatively little is known about these two quantum automorphism groups of graphs and we refer to Section 3.4 for an overview on all published articles in this area. Graph C ∗-algebras in turn are well-established objects in operator algebras. They emerged from Cuntz and Krieger's work [8] in the 1980's and they developed to be one of the most important classes of examples of C ∗-algebras, see for instance Raeburn's book for an overview [14]. Given a finite graph Γ = (V, E) the associated graph C ∗-algebra C ∗(Γ) is defined as C ∗(Γ) := C ∗(pv, v ∈ V, se, e ∈ E pv = p∗ v = p2 v, pvpw = 0 for v 6= w, s∗ ese = pr(e), Xe∈E s(e)=v ses∗ e = pv, if s−1(v) 6= ∅). A natural question is then: What is the quantum symmetry group of the graph C ∗-algebra and is it one of the above two quantum automorphism groups of the underlying graphs? The answer is: It is given by the one defined by Banica. Note however, that Bichon's definition has its justification in other contexts such as in [4, 6] or in the recent work by Speicher and the second author [15]. Moreover, Bichon's work [5] inspired us how to formulate our main theorem, see also Remark 4.2. QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 3 1. Main result Intuitively speaking, our main result is that the quantum symmetry of a finite graph without multiple edges coincides with the quantum symmetry of the associ- ated graph C ∗-algebra. In other words, the following diagram is commutative: finite graphs Γ7→C ∗(Γ) / graph C ∗-algebras ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ aut(Γ) Γ7→G+ s❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤❤ C ∗(Γ)7→QSym(C ∗(Γ)) quantum symmetry groups More precisely, we have the following result. Main Theorem. Let Γ be a finite graph with n vertices V = {1, ..., n} and m edges E = {e1, ..., em} having no multiple edges. The maps α : C ∗(Γ) → C(G+ aut(Γ)) ⊗ C ∗(Γ), pi 7→ sej 7→ n m Xk=1 Xl=1 uik ⊗ pk, 1 ≤ i ≤ n, us(ej)s(el)ur(ej)r(el) ⊗ sel, 1 ≤ j ≤ m, and β : C ∗(Γ) → C(G+ aut(Γ)) ⊗ C ∗(Γ), pi 7→ sej 7→ n m Xk=1 Xl=1 uki ⊗ pk, 1 ≤ i ≤ n, us(el)s(ej )ur(el)r(ej ) ⊗ sel, 1 ≤ j ≤ m define a left and a right action of G+ aut(Γ) on C ∗(Γ), respectively. Moreover, when- ever G is a compact matrix quantum group acting on C ∗(Γ) in the above way, we have G ⊆ G+ aut(Γ) of Γ is the quantum symmetry group of C ∗(Γ), see also Remark 4.1. aut(Γ). In this sense, the quantum automorphism group G+ We also provide some tools for comparing and dealing with the two definitions of quantum automorphism groups of graphs, G+ aut(Γ), notably depending on the complement Γc of Γ, see Section 3.5. Moreover, we provide a list of all Aut(Γ), G+ aut(Γ) for undirected graphs Γ on four vertices, having no multiple edges and no loops, see Section 3.6. aut(Γ) and G∗ aut(Γ) and G∗ / s 4 SIMON SCHMIDT AND MORITZ WEBER 2. Preliminaries 2.1. Graphs. We fix some notations for graphs used throughout this article. A graph Γ = (V, E) is finite, if the set V of vertices and the set E of edges are finite. We denote by r : E → V the range map and by s : E → V the source map. A graph is undirected if for every e ∈ E there is a f ∈ E with s(f ) = r(e) and r(f ) = s(e); it is directed otherwise. Elements e ∈ E with s(e) = r(e) are called loops. A graph without multiple edges is a directed graph, where there are no e, f ∈ E, e 6= f , such that s(e) = s(f ) and r(e) = r(f ). For a finite graph Γ = (V, E) with V = {1, . . . , n}, its adjacency matrix ε ∈ Mn(N0) is defined as εij := #{e ∈ E s(e) = i, r(e) = j}. Here N0 = {0, 1, 2, . . .}. Throughout this article we restrict to finite graphs having no multiple edges. If Γ = (V, E) is a directed graph without multiple edges, we denote by Γc = (V, E′) the complement of Γ, where E′ = (V × V )\E. Within the category of graphs having no loops, the complement Γc is defined using E′ = (V × V )\(E ∪ {(i, i); i ∈ V }). 2.2. Automorphism groups of graphs. For a finite graph Γ = (V, E) without multiple edges, a graph automorphism is a bijective map σ : V → V such that (σ(i), σ(j)) ∈ E if and only if (i, j) ∈ E. In other words, εσ(i)σ(j) = 1 if and only if εij = 1. The set of all graph automorphisms of Γ forms a group, the automorphism group Aut(Γ). We can view Aut(Γ) as a subgroup of the symmetric group Sn, if Γ has n vertices: Aut(Γ) = {σ ∈ Sn σε = εσ} ⊆ Sn 2.3. Graph C ∗-algebras. The theory of Graph C ∗-algebras has its roots in Cuntz and Krieger's work [8] in 1980. Nowadays, it forms a well-developed and very active part of the theory of C ∗-algebras, see [14] for an overview or [9] for recent devel- opments. For a finite, directed graph Γ = (V, E) without multiple edges, the graph C ∗-algebra C ∗(Γ) is the universal C ∗-algebra generated by mutually orthogonal pro- jections pv, v ∈ V and partial isometries se, e ∈ E such that (i) s∗ ese = pr(e) for all e ∈ E (ii) and pv =Pe∈E : s(e)=v ses∗ It follows immediately, that s∗ e for every v ∈ V with s−1(v) 6= ∅. esf = 0 for e 6= f andPv∈V pv = 1 hold true in C ∗(Γ). 2.4. Compact matrix quantum groups. Compact matrix quantum groups were defined by Woronowicz [18, 19] in 1987. They form a special class of compact quantum groups, see [12, 16] for recent books. A compact matrix quantum group G is a pair (C(G), u), where C(G) is a unital (not necessarily commutative) C ∗-algebra which is generated by uij, 1 ≤ i, j ≤ n, the entries of a matrix u ∈ Mn(C(G)). k=1 uik ⊗ ukj Moreover, the -homomorphism ∆ : C(G) → C(G) ⊗ C(G), uij 7→ Pn must exist, and u and its transpose ut must be invertible. QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 5 Example 2.1. As an example, S+ n = (C(S+ group given by quantum symmetric group n ), u) as defined by Wang [17] in 1998. It is the compact matrix quantum consider the n n C(S+ n ) := C ∗(uij uij = u∗ ij = u2 ij, uil = 1 = uli for all 1 ≤ i, j ≤ n). Xl=1 Xl=1 One can show that the quotient of C(S+ n ) by the relations that all uij commute is exactly C(Sn). Moreover, the symmetric group Sn may be viewed as a compact matrix quantum group Sn = (C(Sn), u), where uij : Sn → C are the evaluation maps of the matrix entries. This justifies the name "quantum symmetric group". If G = (C(G), u) and H = (C(H), v) are compact matrix quantum groups with u ∈ Mn(C(G)) and v ∈ Mn(C(H)), we say that G is a compact matrix quantum subgroup of H, if there is a surjective -homomorphism from C(H) to C(G) mapping generators to generators. In this case we write G ⊆ H. As an example: Sn ⊆ S+ n . The compact matrix quantum groups G and H are equal as compact matrix quantum groups, writing G = H, if we have G ⊆ H and H ⊆ G. 2.5. Actions of quantum groups. Let G = (C(G), u) be a compact matrix quantum group and let B be a C ∗-algebra. A left action of G on B is a unital -homomorphism α : B → C(G) ⊗ B such that (i) (∆ ⊗ id) ◦ α = (id ⊗ α) ◦ α (ii) and α(B)(C(G) ⊗ 1) is linearly dense in C(G) ⊗ B. A right action is a unital -homomorphism β : B → C(G) ⊗ B with (i) ((F ◦ ∆) ⊗ id)) ◦ β = (id ⊗ β) ◦ β (ii) and β(B)(C(G) ⊗ 1) is linearly dense in C(G) ⊗ B, where F is the flip map F : C(G) ⊗ C(G) → C(G) ⊗ C(G), a ⊗ b 7→ b ⊗ a. Note that in some articles (for instance in [17]), the property (ii) is replaced by (ii') (ε ⊗ id) ◦ α = id (iii') and there is a dense -subalgebra of B, such that α restricts to a right coaction of the Hopf -algebra on the -subalgebra. One can show that (ii') and (iii') are equivalent to (ii), see [13]. 2.6. Quantum symmetry group of n points. According to Wang's work [17], we know that S+ n (from Example 2.1) is the quantum symmetry group of n points in the sense that (i) S+ n acts from left and right on i = p2 C ∗(p1, . . . , pn pi = p∗ i ,Xl k=1 uik ⊗ pk and β(pi) :=Pn pl = 1) k=1 uki ⊗ pk by α(pi) :=Pn 6 SIMON SCHMIDT AND MORITZ WEBER (ii) and S+ n is maximal with these actions, i.e. any other compact matrix quan- tum group with actions defined as α and β is a compact matrix quantum subgroup of S+ n . See also [11] for similar questions around quantum symmetries. 3. Quantum automorphism groups of graphs Wang's work in the 1990's was the starting point of the investigations of quan- tum symmetry phenomena for discrete structures (within Woronowicz's framework). Note that n points may be viewed as the totally disconnected graph on n vertices. A decade later, Banica and Bichon extended Wang's approach to a theory of quantum automorphism groups of finite graphs. In the sequel, we restrict to finite graphs having no multiple edges. 3.1. Bichon's quantum automorphism group of a graph. In 2003, Bichon [5] defined a quantum automorphism group as follows. Definition 3.1. Let Γ = (V, E) be a finite graph with n vertices V = {1, ..., n} and m edges E = {e1, ..., em}. The quantum automorphism group G∗ aut(Γ) is the compact matrix quantum group (C(G∗ aut(Γ)) is the universal C ∗-algebra with generators uij, 1 ≤ i, j ≤ n and relations aut(Γ)), u), where C(G∗ uij = u∗ ij, n uijuik = δjkuij, ujiuki = δjkuji, 1 ≤ i, j, k ≤ n, n uil = 1 = uli, 1 ≤ i ≤ n, (3.1) (3.2) (3.3) (3.4) (3.5) Xl=1 Xl=1 us(ej)iur(ej)k = ur(ej)kus(ej)i = 0, uis(ej)ukr(ej) = ukr(ej)uis(ej) = 0, us(ej)s(el)ur(ej)r(el) = ur(ej)r(el)us(ej)s(el), ej ∈ E, (i, k) /∈ E, ej ∈ E, (i, k) /∈ E, ej, el ∈ E. In the original definition of Bichon, there is actually another relation which is implied by the others: (3.6) m Xl=1 us(el)s(ej )ur(el)r(ej ) = 1 = m Xl=1 us(ej)s(el)ur(ej)r(el), ej ∈ E Indeed, Relations (3.6) are implied by Relations (3.2), (3.3) and (3.4): us(el)s(ej )ur(el)r(ej ) = m Xl=1 n Xi,k=1 uis(ej)ukr(ej) = n Xi=1 uis(ej)! n Xk=1 ukr(ej)! = 1 3.2. Banica's quantum automorphism group of a graph. Two years later, Banica [1] gave the following definition. QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 7 Definition 3.2. Let Γ = (V, E) be a finite graph with n vertices and adjacency ma- trix ε ∈ Mn({0, 1}). The quantum automorphism group G+ aut(Γ) is the compact ma- trix quantum group (C(G+ aut(Γ)) is the universal C ∗-algebra with generators uij, 1 ≤ i, j ≤ n and Relations (3.1), (3.2) together with aut(Γ)), u), where C(G+ (3.7) uε = εu, which is nothing but Pk uikεkj =Pk εikukj. 3.3. Link between the two definitions. It is easy to see ([10, Lemma 3.1.1] or [15, Lemma 6.7]) that Banica's definition may be expressed as: C(G+ aut(Γ)) = C ∗(uij Relations (3.1) – (3.4)) We thus have Aut(Γ) ⊆ G∗ aut(Γ) ⊆ G+ aut(Γ) in the sense of compact matrix quantum subgroups, see Section 2.4. Equality holds, if C(G∗ aut(Γ)) are commutative. Moreover, note that (see Example 2.1): aut(Γ)) and C(G+ C(S+ n ) = C ∗(uij Relations (3.1) and (3.2)) Example 3.3. As an example, let Γ be the complete graph (i.e. E = V ×V ). Then: Aut(Γ) = G∗ aut(Γ) = Sn, G+ aut(Γ) = S+ n For its complement Γc (i.e. E = ∅), we have: Aut(Γc) = Sn, G∗ aut(Γc) = G+ aut(Γc) = S+ n 3.4. Review of the literature on quantum automorphism groups of graphs. At the moment there are only few articles regarding quantum automorphism groups of graphs. Some results are the following. In [6], Bichon defined the hyperoctahedral quantum group and showed that this group is the quantum automorphism group of some graph. Banica computed the Poincar´e series of G+ aut(Γ) for homogenous graphs with less than eight vertices in [1]. Banica, Bichon and Chenevier considered circulant graphs having p vertices for p prime in [3]. They proved G+ aut(Γ) = Aut(Γ) if the graph Γ does fulfill certain properties. Banica and Bichon investigated G+ aut(Γ) for vertex-transitive graphs of order less or equal to eleven in [2]. They also computed G+ aut(Γ) for the direct product, the Cartesian product and the lexicographic product of specific graphs. Chassaniol also studied the lexicographic product of graphs in [7]. In her PhD thesis [10], Fulton studied undirected trees Γ such that Aut(Γ) = Z2 × Z2 × ... × Z2, where we have k kopies of the cyclic group Z2 = Z/2Z. She proved Aut(Γ) = G∗ aut(Γ) for k ≥ 2. See also [4] for links to quantum isometry groups. aut(Γ) for k = 1 and Aut(Γ) 6= G∗ aut(Γ) = G+ aut(Γ) = G+ 8 SIMON SCHMIDT AND MORITZ WEBER 3.5. Comparing with the complement of the graph. As can be seen from Sec- tion 3.4, the theory of quantum automorphism groups of graphs is still in its infancy. We now provide some basic results on the link between G∗ aut(Γc). Note that while we have aut(Γ) and G∗ Aut(Γ) = Aut(Γc) and G+ aut(Γ) = G+ aut(Γc) for all graphs Γ (using εΓc = A − εΓ for the adjacency matrices, with A ∈ Mn({1}) the matrix filled with units, and uA = A = Au by Relation (3.2)), we may have for instance when Γ is the complete graph, see Example 3.3. G∗ aut(Γ) 6= G∗ aut(Γc), Lemma 3.4. If G∗ aut(Γ) ⊆ G∗ aut(Γc), then G∗ aut(Γ) = Aut(Γ). Proof. Relation (3.5) in C(G∗ C(G∗ (3.4) and (3.5) in C(G∗ implies that uik and ujl commute in aut(Γ)) whenever (i, j) /∈ E and (k, l) /∈ E. Together with Relations (3.3), (cid:3) aut(Γ)) this yields commutativity of all generators. aut(Γc)) Lemma 3.5. If G∗ aut(Γc), then G∗ aut(Γ) = Aut(Γ). aut(Γc) = G+ aut(Γ) ⊆ G+ Proof. We have G∗ aut(Γc) and apply Lemma 3.4. (cid:3) The next lemma shows that the quantum automorphism groups of a graph without aut(Γc) = G∗ aut(Γ) = G+ loops does not change if we add those. Lemma 3.6. Let Γ = (V, E) be a finite graph without loops. Consider Γ′ = (V, E′) with E′ = E ∪ {(i, i), i ∈ V }. It holds (i) G+ (ii) G∗ aut(Γ) = G+ aut(Γ) = G∗ aut(Γ′), aut(Γ′). Proof. For (i), we use εΓ′ = 1 + εΓ, where 1 is the identity matrix in Mn({0, 1}). Thus, uεΓ = εΓu is equivalent to uεΓ′ = εΓ′u. For (ii), all we need to check is that uis(ej)uir(ej) = uir(ej)uis(ej) is fulfilled in (cid:3) aut(Γ)) for all i ∈ V , ej ∈ E, which is true due to Relation (3.1). C(G∗ 3.6. Quantum automorphism groups on four vertices. For a small number of vertices of undirected graphs, a complete classification of G∗ aut(Γ) is possible. For n ∈ {1, 2, 3}, we have C(S+ aut(Γ) = G+ aut(Γ). For n = 4, we now provide a complete table for graphs having no loops. We restrict to undirected graphs in order to keep it simple. We need the following lemma to compute the quantum automorphism groups. aut(Γ) and G+ n ) = C(Sn), hence Aut(Γ) = G∗ Lemma 3.7. Let Γ = (V, E) be a finite graph with V = {1, ..., n} and let ej ∈ E. Let q ∈ V with s−1(q) = ∅. For the generators of C(G+ aut(Γ)) it holds uqs(ej) = 0 = us(ej)q. QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS Proof. By Relations (3.2) and (3.4), we get uqs(ej) = uqs(ej) n Xi=1 uir(ej)! = n Xi=1 uqs(ej)uir(ej) = 0, because (q, i) /∈ E for all i ∈ V . Likewise, we get us(ej)q = 0. In the following, D4 denotes the dihedral group defined as 9 (cid:3) D4 := hx, y x2 = y2 = (xy)4 = ei, H + 2 denotes the hyperoctahedral quantum group defined by Bichon in [6] and Z2 denotes the cyclic group Z/2Z. The quantum group \Z2 ∗ Z2 = (C ∗(Z2 ∗ Z2), u) is understood as the compact matrix quantum group with matrix p 1 − p 1 − p 0 0 p 0 0 0 0 q 0 0 1 − q 1 − q q     where C ∗(Z2 ∗ Z2) is seen as the universal unital C ∗-algebra generated by two pro- jections p and q. Recall that Aut(Γ) = Aut(Γc) and G+ aut(Γc), where Γc is the complement of Γ within the category of graphs having no loops. Parts of the following table were also computed in [2] and [6]. aut(Γ) = G+ Theorem 3.8. Let Γ be an undirected graph on four vertices having no loops and no multiple edges. Then: Γ • • • • • • • • • • • • • • • • • • • • • • • • Γc • • ❅❅ • • • • ❅❅ • • • • ❅❅ • • • • • • • • ❅❅ • • • • • • (1) (2) (3) (4) (5) (6) Aut(Γ) G∗ aut(Γc) G∗ aut(Γ) G+ aut(Γ) S4 S4 S+ 4 S+ 4 Z2 × Z2 Z2 × Z2 \Z2 ∗ Z2 \Z2 ∗ Z2 Z2 D4 S3 Z2 Z2 D4 S3 Z2 Z2 H + 2 S3 Z2 Z2 H + 2 S3 Z2 Proof. For every row of the table, we compute G+ G∗ graphs as follows: aut(Γ). We then obtain G∗ aut(Γ) = aut(Γc) by using Lemma 3.5. We label the points of the aut(Γ) and we show G+ 1 3 • • • • 2 4 10 SIMON SCHMIDT AND MORITZ WEBER (1) Obvious, see Example 3.3. (2) Let (uij)1≤i,j≤4 be the generators of C(G+ aut(Γ)). Lemma 3.7 yields u31 = u32 = u41 = u42 = u13 = u23 = u14 = u24 = 0. With Relations (3.2) we deduce u11 1 − u11 1 − u11 0 0 u11 0 0   0 0 0 0 u33 1 − u33 1 − u33 u33 .   u = Thus G+ aut(Γ) = \Z2 ∗ Z2. Since uijukl = ukluij holds aut(Γ) = G+ C(G+ aut(Γ)), we get G∗ for aut(Γ). (i, k), (j, l) ∈ {(1, 2), (2, 1)} in (3) Lemma 3.7 yields u14 = u24 = u34 = u41 = u42 = u43 = 0. This implies 3 = S3, aut(Γ) is commutative and hence G+ aut(Γ) ⊆ S+ G+ thus G+ (4) Let ∆ and ∆′ be the comultiplication maps of G+ aut(Γ) = Aut(Γ) = Z2. 2 , respectively. We first show that these two quantum groups coincide as compact quantum groups, i.e. there is a ∗-isomorphism aut(Γ) and H + aut(Γ) = G∗ ϕ : C(H + 2 ) → C(G+ aut(Γ)) such that ∆′ ◦ ϕ = (ϕ ⊗ ϕ) ◦ ∆. Step 1: The map ϕ exists and we have ∆′ ◦ ϕ = (ϕ ⊗ ϕ) ◦ ∆. From εu = uε we get u11 u12 u13 u14 u12 u11 u14 u13 u31 u32 u33 u34 u32 u31 u34 u33 u =   .   Define v11 := u11 − u12, v12 := u13 − u14, v21 := u31 − u32 and v22 := u33 − u34. One can compute that vij, i, j = 1, 2 fulfill the relations of C(H + 2 ) and with the universal property we get a -homomorphism ϕ : C(H + aut(Γ)). Since ∆′ ◦ ϕ = (ϕ ⊗ ϕ) ◦ ∆ also holds, we get that G+ aut(Γ) is a quantum 2 ) → C(G+ QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 11 subgroup of H + 2 . Step 2: The map ϕ is a ∗-isomorphism. Let (vij)i,j=1,2 be the generators of C(H + 2 ). Define u11 := u22 := u13 := u24 := u31 := u42 := u33 := u44 := v2 11 + v11 2 v2 12 + v12 2 v2 21 + v21 2 v2 22 + v22 2 , , , , u12 := u21 := u14 := u23 := u41 := u32 := u34 := u43 := v2 11 − v11 2 v2 12 − v12 2 v2 21 − v21 2 v2 22 − v22 2 , , , . One can show that the (uij)1≤i,j≤4 fulfill the relations of C(G+ versal property now gives us a -homomorphism ϕ′ : C(G+ and ϕ′ is the inverse of ϕ and vice versa. aut(Γ)). The uni- aut(Γ)) → C(H + 2 ) Step 3: We have G+ We have seen in Step 1, that aut(Γ) = G∗ aut(Γ). u11 = u22, u31 = u42, u12 = u21, u32 = u41, u13 = u24, u33 = u44, u14 = u23, u34 = u43 and therefore we get uijukl = u2 kl = ukluij for all (i, k), (j, l) ∈ E. Thus G+ aut(Γ) = G∗ aut(Γ). (5) We conclude as in (3). (6) Some direct computations using εu = uε and Relations (3.2) show u33 1 − u33 1 − u33 u =   0 0 u33 0 0 0 0 0 0 u33 1 − u33 1 − u33 u33 .   (cid:3) Thus G+ aut(Γ) is commutative. 12 SIMON SCHMIDT AND MORITZ WEBER 4. Proof of the main result We now prove the main result of this article (see Section 1) for a finite graph Γ with vertices V = {1, . . . , n} and edges E = {e1, . . . , em} having no multiple edges. Remark 4.1. We define the quantum symmetry group QSym(C ∗(Γ)) of C ∗(Γ) to be the maximal compact matrix quantum group G acting on C ∗(Γ) by α : C ∗(Γ) → C(G) ⊗ C ∗(Γ) and β : C ∗(Γ) → C(G) ⊗ C ∗(Γ) as defined in the statement of our main theorem. We thus have to show that G+ aut(Γ) acts on C ∗(Γ) via α and β (see Sections 4.1 and 4.2) and that it is maximal with these actions (see Section 4.3). 4.1. Existence of the maps α and β. In order to prove that α : C ∗(Γ) → C(G+ aut(Γ)) ⊗ C ∗(Γ) n pi 7→ p′ i := uik ⊗ pk, 1 ≤ i ≤ n m Xk=1 Xl=1 sej 7→ s′ ej := us(ej)s(el)ur(ej)r(el) ⊗ sel, 1 ≤ j ≤ m defines a ∗-homomorphism, all we have to show is that the relations of C ∗(Γ) hold ej . We may then use the universal property of C ∗(Γ). The proof for the for p′ existence of β is analogous. i and s′ 4.1.1. The p′ over, using pkpl = δklpk and Relations (3.1), we have i are mutually orthogonal projections. Obviously, p′ i = (p′ i)∗ holds. More- p′ ip′ j = n Xk,l=1 uikujl ⊗ pkpl = n Xk=1 uikujk ⊗ pk = δijp′ i. are partial isometries with (s′ 4.1.2. The s′ ej (see Section 2.3) and Relations (3.1), we have ej )∗s′ ej = p′ r(ej ). Using s∗ elsei = δilpr(ei) m (s′ ej )∗s′ ej = = ur(ej)r(el)us(ej)s(el)us(ej)s(ei)ur(ej)r(ei) ⊗ s∗ elsei ur(ej)r(ei)us(ej)s(ei)ur(ej)r(ei) ⊗ pr(ei). m Xl,i=1 Xi=1 By Relations (3.3) we have ur(ej)j ′us(ej)i′ur(ej)j ′ = 0 for (i′, j′) /∈ E. This yields ur(ej)r(ei)us(ej)s(ei)ur(ej)r(ei) ⊗ pr(ei) = m Xi=1 n Xi′,j ′=1 ur(ej)j ′us(ej)i′ur(ej)j ′ ⊗ pj ′. ej (s′ get for v ∈ V with s−1(v) 6= ∅: 4.1.3. We have Pj: s(ej )=v s′ Xj∈{1,...,m} s(ej)=v s′ ej (s′ ej )∗ = Xj∈{1,...,m} s(ej )=v m = Xl=1 Xi∈{1,...,m} r(ei)=r(el) uvs(el)ur(ej)r(el)ur(ej)r(ei)uvs(ei) ⊗ sels∗ ei m Xi,l=1 ur(ej)r(el)  uvs(ei) ⊗ sels∗ ei uvs(el) s(ej )=v  Xj∈{1,...,m} ur(ej)r(el) = Xq∈V (v,q)∈E uqr(el) QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 13 Using Relations (3.2), we obtain Pn n (s′ ej )∗s′ ej = ur(ej)j ′us(ej)i′ur(ej)j ′ ⊗ pj ′ = i=1 us(ej)i′ = 1 and thus Xi′,j ′=1 ur(ej)j ′ ⊗ pj ′ = p′ r(ej). n Xj ′=1 ej )∗ = p′ v for s−1(v) 6= ∅. Using Relations (3.1), we Now, Xj∈{1,...,m} s(ej)=v and for q ∈ V with (v, q) /∈ E we have uvs(el)uqr(el) = 0 by Relations (3.4). Thus, for any l ∈ {1, . . . , m}, we have using Relations (3.2) and hence: uvs(el) Xq∈V (v,q)∈E uqr(el) = uvs(el)Xq∈V uqr(el) = uvs(el) s′ ej (s′ ej )∗ = Xj∈{1,...,m} s(ej )=v m Xl=1 Xi∈{1,...,m} r(ei)=r(el) uvs(el)uvs(ei) ⊗ sels∗ ei Since Γ has no multiple edges by assumption, r(ei) = r(el) and s(ei) = s(el) implies ei = el. We thus infer using Relations (3.1): s′ ej (s′ ej )∗ = Xj∈{1,...,m} s(ej )=v m Xl=1 uvs(el) ⊗ sels∗ el Now, for V ′ := {q ∈ V s−1(q) 6= ∅}, we have, using the relations in C ∗(Γ): m Xl=1 uvs(el) ⊗ sels∗ el = Xq∈V ′ Xl∈{1,...,m} s(el)=q uvq ⊗ sels∗ el = Xq∈V ′ uvq ⊗ pq 14 SIMON SCHMIDT AND MORITZ WEBER Since we know that uvq = 0 for q /∈ V ′ by Lemma 3.7, we finally conclude s′ ej (s′ ej )∗ = Xj∈{1,...,m} s(ej)=v n Xq=1 uvq ⊗ pq = p′ v. This settles the existence of α. 4.2. The map α is a left action and β is a right action. We only prove this claim for α, the proof for β being analogous. 4.2.1. (∆ ⊗ id) ◦ α = (id ⊗ α) ◦ α holds and α is unital. Using Relations (3.3), this is straightforward to check. It remains to show that S := span α(C ∗(Γ))(C(G+ aut(Γ)) ⊗ 1) is dense in C(G+ aut(Γ)) ⊗ C ∗(Γ), which we will do in the sequel. 4.2.2. The elements 1 ⊗ pl, 1 ⊗ sel and 1 ⊗ s∗ el (3.2) we infer: are in S. Using Relations (3.1) and S ∋ n Xi=1 α(pi)(uil ⊗ 1) = n n Xi=1 Xj=1 uijuil ⊗ pj = n Xi=1 uil ⊗ pl = 1 ⊗ pl Moreover, for el ∈ E we get, using Relations (3.1) and V ′ := {v ∈ V s−1(v) 6= ∅}: Xv∈V ′ Xj∈{1,...,m} s(ej )=v α(sej )(ur(ej)r(el)uvs(el) ⊗ 1) = Xv∈V ′ Xj∈{1,...,m} s(ej )=v = Xv∈V ′   Xk∈{1,...,m} r(ek)=r(el) m uvs(ek)ur(ej)r(ek)ur(ej)r(el)uvs(el) ⊗ sek! Xk=1 uvs(ek) uvs(el) ⊗ sek  ur(ej)r(el)   Xj∈{1,...,m} s(ej)=v We proceed similar to Step 4.1.3. By Relations (3.4), we know uqr(el)uvs(el) = 0 for (v, q) /∈ E. Thus, by Relations (3.1) and (3.2) and using that Γ has no multiple edges, we obtain: Xv∈V ′ Xj∈{1,...,m} s(ej )=v α(sej )(ur(ej)r(el)uvs(el) ⊗ 1) QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 15 r(ek)=r(el)   Xk∈{1,...,m}   Xk∈{1,...,m} r(ek)=r(el) uvs(el) ⊗ sel uqr(el)! uvs(el) ⊗ sek  uvs(ek) n Xq=1 uvs(ek)uvs(el) ⊗ sek  = Xv∈V ′ = Xv∈V ′ = Xv∈V ′ Finally, Lemma 3.7 yields uvs(el) = 0 for v /∈ V ′. Hence, using Relations (3.2): S ∋ Xv∈V ′ Xj∈{1,...,m} s(ej )=v α(sej )(ur(ej)r(el)us(ej)s(el) ⊗ 1) = uis(el) ⊗ sel = 1 ⊗ sel n Xi=1 Define V ′′ := {v ∈ V r−1(v) 6= ∅}. Similar to the computations above, we get S ∋ Xv∈V ′′ Xj∈{1,...,m} s(ej)=v α(s∗ ej )(us(ej)s(el)ur(ej)r(el) ⊗ 1) = 1 ⊗ s∗ el. 4.2.3. If 1 ⊗ x, 1 ⊗ y ∈ S, then also 1 ⊗ xy ∈ S. The remainder of the proof of Step 4.2 consists in general facts for actions of compact matrix quantum groups. We may write 1 ⊗ x ∈ S and 1 ⊗ y ∈ S as l 1 ⊗ x = α(zi)(wi ⊗ 1), 1 ⊗ y = Xi=1 α(tj)(vj ⊗ 1) k Xj=1 for some zi, tj ∈ C ∗(Γ) and wi, vj ∈ C(G+ aut(Γ)). Therefore 1 ⊗ xy = = = l l Xi=1 Xi=1 Xi=1 l α(zi)(wi ⊗ 1)(1 ⊗ y) α(zi)(1 ⊗ y)(wi ⊗ 1) α(zitj)(vjwi ⊗ 1) ∈ S k Xj=1 4.2.4. S is dense in C(G+ all monomials w in pi, sej , s∗ aut(Γ)) ⊗ C ∗(Γ). Summarizing, we get that 1 ⊗ w ∈ S for ej , 1 ≤ i ≤ n, 1 ≤ j ≤ m. Since α is unital, we also have: C(G+ aut(Γ)) ⊗ 1 ⊆ α(C ∗(Γ))(C(G+ aut(Γ)) ⊗ 1) ⊆ S Xk=1 Xi,l=1 Xi=1 Xi=1 16 SIMON SCHMIDT AND MORITZ WEBER We conclude that S is dense in C(G+ aut(Γ)) ⊗ C ∗(Γ), which settles Step 4.2. 4.3. The quantum group G+ aut(Γ) acts maximally on C ∗(Γ). For proving the maximality, let G = (C(G), u) be another compact matrix quantum group acting on C ∗(Γ) by α′ : C ∗(Γ) → C(G) ⊗ C ∗(Γ) and β′ : C ∗(Γ) → C(G) ⊗ C ∗(Γ) in the way G+ aut(Γ) acts on C ∗(Γ) via α and β. We want to show that there is a - homomorphism C(G+ aut(Γ)) → C(G) sending generators to generators. Thus, we need to compute that the generators uij of C(G) fulfill the relations of C(G+ aut(Γ)). 4.3.1. The Relations (3.1) hold in C(G). The equation n n uik ⊗ pk = α′(pi) = α′(pi)∗ = u∗ ik ⊗ pk Xk=1 yields uij = u∗ ij after multiplying from the left with 1 ⊗ pj. We also have n n n ujiuki ⊗ pi = ujiukl ⊗ pipl = α′(pj)α′(pk) = δjkα′(pj) = δjkuji ⊗ pi from which we infer ujiuki = δjkuji. Using β′, we also obtain uijuik = δjkuij. 4.3.2. The Relations (3.2) hold in C(G). From n n 1 ⊗ pk = 1 ⊗ 1 = α′(1) = Xk=1 Xi=1 i=1 uik = 1, and likewise Pn we deduce Pn 4.3.3. The Relations (3.3) hold in C(G). Using s∗ and Relations (3.1) in C(G), we obtain for any j: α′(pi) = i=1 uki = 1 using β′. n Xk=1 n Xi=1 uik! ⊗ pk elset = δltpr(el) (see Section 2.3) n Xq=1 ur(ej)q ⊗ pq = α′(pr(ej)) = α′(s∗ ej sej ) m ur(ej)r(el)us(ej)s(el)us(ej)s(et)ur(ej)r(et) ⊗ s∗ elset ur(ej)r(el)us(ej)s(el)ur(ej)r(el) ⊗ pr(el) = = m Xl,t=1 Xl=1 Multiplication with 1 ⊗ pk yields: ur(ej)k = Xl∈{1,...,m} r(el)=k ur(ej)kus(ej)s(el)ur(ej)k If r−1(k) = ∅, then ur(ej)k = 0 and hence us(ej)iur(ej)k = ur(ej)kus(ej)i = 0 for all i ∈ V . QUANTUM SYMMETRIES OF GRAPH C ∗-ALGEBRAS 17 Otherwise, if r−1(k) 6= ∅, we use Relations (3.1) and (3.2) in C(G) and get ur(ej)kus(ej)s(el)ur(ej)k = ur(ej)k = u2 r(ej )k = ur(ej)kus(ej)iur(ej)k n Xi=1 ur(ej)kus(ej)iur(ej)k = 0. Xi∈V (i,k) /∈E Xl∈{1,...,m} r(el)=k and therefore Since ur(ej)kus(ej)iur(ej)k = (us(ej)iur(ej)k)∗us(ej)iur(ej)k holds, the above is a vanishing sum of positive elements – and hence each summand vanishes. This yields us(ej)iur(ej)k = 0 for all (i, k) /∈ E. 4.3.4. The Relations (3.4) hold in C(G). The argument is analogous to the one for proving Relations (3.3) when replacing α′ by β′. The proof of the main theorem is complete. Remark 4.2. Let Γ be a finite graph with n vertices V = {1, ..., n} and m edges E = {e1, ..., em}. aut(Γ) is the quantum symmetry group of Γ in his sense, where In [5], Bichon showed that G∗ βV : C(V ) → C(G∗ aut(Γ)) ⊗ C(V ), gi 7→ βE : C(E) → C(G∗ aut(Γ)) ⊗ C(E), fj 7→ uki ⊗ gk, us(el)s(ej )ur(el)r(ej ) ⊗ fl, n m Xk=1 Xl=1 define actions of G∗ aut(Γ) on C(V ) and C(E), respectively. Those actions inspired us, how an action of a compact matrix quantum group on C ∗(Γ) should look like. However, note that edges in the commutative C ∗-algebra C(E) of continuous func- tions on E are represented as projections unlike in the case of C ∗(Γ). Therefore, the quantum symmetry group of C ∗(Γ) is G+ aut(Γ). On the other hand, if we consider the quotient of C ∗(Γ) by the relations se = s∗ e, its quantum symmetry group is G∗ aut(Γ). Indeed, selfadjointness of se yields aut(Γ) rather than G∗ us(ej)s(el)ur(ej)r(el) ⊗ sel = α(sej ) = α(sej )∗ = m Xl=1 ur(ej)r(el)us(ej)s(el) ⊗ sel, m Xl=1 from which we obtain Relations (3.5) by multiplication with (1 ⊗ s∗ ei) from the left. 18 SIMON SCHMIDT AND MORITZ WEBER References [1] Teodor Banica. Quantum automorphism groups of homogeneous graphs. J. Funct. Anal., 224(2):243–280, 2005. [2] Teodor Banica and Julien Bichon. Quantum automorphism groups of vertex-transitive graphs of order ≤ 11. J. Algebraic Combin., 26(1):83–105, 2007. [3] Teodor Banica, Julien Bichon, and Gaetan Chenevier. Graphs having no quantum symmetry. Ann. Inst. Fourier, pages 955–971, 2007. [4] Jyotishman Bhowmick, Debashish Goswami, and Adam Skalski. Quantum isometry groups of 0-dimensional manifolds. Trans. Amer. Math. Soc., 363(2):901–921, 2011. [5] Julien Bichon. Quantum automorphism groups of finite graphs. Proc. Amer. Math. Soc., 131(3):665–673, 2003. [6] Julien Bichon. Free wreath product by the quantum permutation group. Algebr. Represent. Theory, 7(4):343–362, 2004. [7] Arthur Chassaniol. Quantum automorphism group of the lexicographic product of finite reg- ular graphs. J. Algebra, pages 23–45, 2016. [8] Joachim Cuntz and Wolfgang Krieger. A class of C ∗-algebras and topological Markov chains. Invent. Math., 56(3):251–268, 1980. [9] Søren Eilers, Gunnar Restorff, Efren Ruiz, and Adam Sørensen. The complete classification of unital graph C ∗-algebras: Geometric and strong. arXiv:1611.07120, 2016. [10] Melanie Fulton. The quantum automorphism group and undirected trees. PhD Thesis, Vir- ginia, 2006. [11] Debashish Goswami and Jyotishman Bhowmick. Quantum isometry groups. Infosys Science Foundation Series. Springer, New Delhi, 2016. Infosys Science Foundation Series in Mathe- matical Sciences. [12] Sergey Neshveyev and Lars Tuset. Compact quantum groups and their representation cate- gories, volume 20 of Cours Sp´ecialis´es [Specialized Courses]. Soci´et´e Math´ematique de France, Paris, 2013. [13] Piotr Podle´s. Symmetries of quantum spaces. Subgroups and quotient spaces of quantum SU(2) and SO(3) groups. Comm. Math. Phys., 170(1):1–20, 1995. [14] Iain Raeburn. Graph algebras, volume 103 of CBMS Regional Conference Series in Mathe- matics. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. [15] Roland Speicher and Moritz Weber. Quantum groups with partial commutation relations. arXiv:1603.09192, 2016. [16] Thomas Timmermann. An invitation to quantum groups and duality. EMS Textbooks in Math- ematics. European Mathematical Society (EMS), Zurich, 2008. [17] Shuzhou Wang. Quantum symmetry groups of finite spaces. Comm. Math. Phys., 195(1):195– 211, 1998. [18] S. L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys., 111(4):613–665, 1987. [19] S. L. Woronowicz. A remark on compact matrix quantum groups. Lett. Math. Phys., 21(1):35– 39, 1991. Saarland University, Fachbereich Mathematik, 66041 Saarbrucken, Germany E-mail address: [email protected], [email protected]
1710.06123	1	1710	2017-10-17T06:51:30	A theorem for random Fourier series on compact quantum groups	[ "math.OA", "math.FA" ]	Helgason showed that a given measure $f\in M(G)$ on a compact group $G$ should be in $L^2(G)$ automatically if all random Fourier series of $f$ are in $M(G)$. We explore a natural analogue of the theorem in the framework of compact quantum groups and apply the obtained results to study complete representability problem for convolution algebras of compact quantum groups as an operator algebra.	math.OA	math	A THEOREM FOR RANDOM FOURIER SERIES ON COMPACT QUANTUM GROUPS SANG-GYUN YOUN Abstract. Helgason showed that a given measure f P M pGq on a compact group G should be in L2pGq automatically if all random Fourier series of f are in M pGq. We explore a natural analogue of the theorem in the framework of compact quantum groups and apply the obtained results to study complete representability problem for convolution algebras of compact quantum groups as an operator algebra. 1. Introduction The random Fourier series theory has been extensively studied for a long period of time [PZ32], [Kah93], [SZ54], [Hun51], [Dud67], [Fer75], [Mar78] and [MP78] and most of the results turned out to be valid for Banach space-valued functions on compact groups [FTR66], [Kwa76] and [MP81]. Moreover, it has been found that these studies can be applied to various topics of harmonic analysis [Pis77], [Pis78] and [Rid75]. A theorem of Helgason [Hel57], which is a subject of this paper, is one of the re- sults of the theory of random Fourier series. It is an improvement of the Littlewood's work on circle [Lit26]. Throughout this paper, we denote by IrrpGq a maximal family of mutually in- equivalent irreducible unitary representations of a compact group G and by U the Upnπq where Updq is the group of unitary matrices of size d product group źπPIrrpGq and nπ is the dimension of π P IrrpGq. Also, we denote by f „ ÿπPIrrpGq nπtrppfpπqπpxqq Fourier coefficients of f P MpGq. Theorem 1.1 (S.Helgason, [Hel57]). Let G be a compact group and fix a measure the Fourier-Stieltjes series of f P MpGq, where ppfpπqqπPIrrpGq is the sequence of f P MpGq. Suppose that there exists µU P MpGq such that µU „ ÿπPpG nπtrpUπpfpπqπpxqq for any U " pUπqπPpG P U. Then ÿπPpG nπtrppfpπqpfpπqq ă 8. After the works of S.Woronowicz, which are [Wor87a], [Wor87b] for compact quantum groups, the theory of topological quantum groups has been greatly ele- vated [KV00], [Kus01], [KV03] and harmonic analysis on quantum groups has also been vigorously studied in [Cas13], [FHL`17], [JMP14], [JPPP17] and [You18]. In particular, progress has been made in research on random Fourier series [Wan17]. Key words and phrases. Random Fourier series, compact quantum group, non self-adjoint operator algebra. 2010 Mathematics Subject Classification. Primary 46L89, Secondary 20G42, 43A30. The author is supported by TJ Park Science Fellowship. 1 The main purpose of this paper is to generalize Theorem 1.1 in the framework of compact quantum groups in a natural manner through mobilizing well-known studies on stochastic behavior of some special vector-valued random variables(e.g. Rademacher and Gaussian variables), namely the non-commutative Khintchine in- equalities [LP86], [LPP91] and (co-)type studies [TJ74], [Fac87]. The Fourier analysis on compact quantum groups has been developed in respect of traditional philosophy of the Fourier analysis on compact groups to large extent. However, a visible difference appears in the Fourier expansion. In fact, our random Fourier series of f P MpGq will be described as fU „ ÿαPIrrpGq dαtrpUαpfpαqQαuαq for any U " pUαqαPIrrpGq P U. We will explain the details of the above expansion in Section 2. As in the classical setting, f P L2pGq will imply that all of random fourier series fU are in L2pGq Ď L1pGq by the Plancherel identity on compact quantum groups (Proposition 2.1). Accordingly what we have to do is to demonstrate the converse direction and main results of this paper are as follows. Theorem 1.2. In the above notation, let us suppose that fU P MpGq for all U " pUαqαPIrrpGq P U. (1) If G is of Kac type, then we have (2) For general compact quantum groups, we have ÿαPIrrpGq ÿαPIrrpGq dα nα nαtrppfpαqpfpαqq ă 8. trpQαpfpαqpfpαqq ă 8. Moreover, we will apply the obtained results to a problem determining whether the associated convolution algebra L1pGq can be completely isomorphic to a closed subalgebra of BpHq, in which H is a Hilbert space (See Section 4). In this case, we will say that L1pGq is completely representable as an operator algebra. Corollary 1.3. Let G " pA, ∆q be a compact quantum group. Then L1pGq is completely representable as an operator algebra if and only if A is finite dimensional. 2. Preliminaries 2.1. Compact quantum group. A compact quantum group is a pair G " pA, ∆q where A is a unital C-algebra and ∆ : A Ñ Abmin A is a unital ´homomorphism satisfying (1) p∆ b idq ∆ " pid b ∆q ∆ and (2) spant∆paqpb b 1q : a, b P Au and spant∆paqp1 b bq : a, b P Au are dense in A bmin A. Let us describe the basic representation theory and the Schur's orthogonality re- lation on compact quantum groups. A finite dimensional unitary representation of ui,k b uk,j and puuqi,j " G is u " pui,jq1ďi,jďn P MnpAq such that ∆pui,jq " u k,iuk,j " δi,j1A " j,k " puuqi,j for all 1 ď i, j ď n. We say that a uni- tary representation u P MnpAq is irreducible if tX P Mn : Xu " uXu " C Idn and a maximal family of mutually inequivalent ui,ku nÿk"1 denote by uα " puα nÿk"1 i,jq1ďi,jďnα(αPIrrpGq unitary irreducible representations of G. nÿk"1 2 Every compact quantum group allows the analogue of the Haar measure, namely the Haar state, which is the unique state h on A such that pidbhqp∆paqq " hpaq1A " ph b idqp∆paqq for all a P A and hp1Aq " 1. is unitary and trpQαq " trpQ´1 i,jqq1ďi,jďnα . We call dα " trpQαq " trpQ´1 α q the quantum dimension of α P IrrpGq. For any α, β P IrrpGq, 1 ď i, j ď nα and 1 ď s, t ď nβ, we have the Schur's orthogonality relation as follows. For each α P IrrpGq, there exists a unique mathrix Qα P Mnα such that Q α q where uα :" ppuα α uαQ ´ 1 α 1 2 2 hppuβ s,tquα i,jq " δα,βδj,tpQ´1 trpQαq α qi,s and hpuβ s,tpuα i,jqq " δα,βδi,spQαqj,t trpQαq . Moreover, we may assume that the matrices Qα are diagonal [Daw10]. Also, we say that G is of Kac type if Qα " Idnα for all α P IrrpGq or equivalently if the Haar state h is tracial. We define the space of matrix coefficients as PolpGq " span uα i,j : α P IrrpGq, 1 ď i, j ď nα( and then the Haar state h is faithful on PolpGq. We denote by L2pGq the closure of PolpGq with respect to the pre-inner product xa, byL2pGq " hpbaq for all a, b P PolpGq and denote by Λ : PolpGq ãÑ L2pGq, a ÞÑ Λpaq the natural embedding. Also, we define CrpGq as the norm-closure of PolpGq in BpL2pGqq under the GNS representation rιpaqspΛpxqq :" Λpaxq for all a, x P PolpGq and L8pGq as weak - closure of CrpGq in BpL2pGqq. The haar state h extends to L8pGq as a normal faithful state and we will denote it by h again. We consider the predual L1pGq of L8pGq and the dual MpGq of CrpGq respectively. Then we have a contractive embedding L8pGq ãÑ L1pGq, x ÞÑ hpxq and the isometric formal identity from L1pGq into MpGq. Note that PolpGq is dense in L1pGq. 2.2. Fourier analysis on compact quantum groups. It is known that every sup Mnα : αPIrrpGq}Xα} ă 8,.- quantum group structure under the generalized Pontrjagin duality. Although we will not mention its quantum group structure, we will specifically explain some of compact quantum group admits the dual discrete quantum group pG with a natural non-commutative ℓp-spaces on the discrete quantum group pG. The associated von Neumann algebra of pG is given by ℓ8ppGq "$&%pXαqαPIrrpGq P źαPIrrpGq with the norm››pXαqαPIrrpGq››ℓ8ppGq " sup We define the Fourier-Stieltjes transform by F : MpGq " CrpGq Ñ ℓ8ppGq, µ ÞÑ ppµpαqqαPIrrpGq with pµpαq " pµppuα nαÿi,j"1 dαtrppµpαqQαuαq " ÿαPIrrpGq f " ÿαPIrrpGq to follow the standard notations of Fourier analysis on compact groups. Then the Fourier expansion of µ P MpGq is given by dαppµpαqQαqi,j uα Indeed, if f P PolpGq, then we have µ „ ÿαPIrrpGq αPIrrpGq}Xα}Mnα j,iqqq1ďi,jďnα , dαtrppfpαqQαuαq. 3 . j,i. The associated ℓ2 and ℓ1 spaces on pG are ℓ2ppGq "$&%pXαqαPIrrpGq P ℓ8ppGq : ÿαPIrrpGq ℓ1ppGq "$&%pXαqαPIrrpGq P ℓ8ppGq : ÿαPIrrpGq ››pXαqαPIrrpGq››ℓ2ppGq "p ÿαPIrrpGq ››pXαqαPIrrpGq››ℓ1ppGq " ÿαPIrrpGq with the norm structures , dαtrpQαX αXαq ă 8,.- dαtrpXαQαq ă 8,.- dαtrpQαX α Xαqq 1 2 , dαtrpXαQαq respectively. When we restrict the domain of the Fourier-Stieltjes transform to L2pGq, we have the isometry F : L2pGq Ñ ℓ2ppGq [Wan17], which is called the Plancherel identity. 2.3. Random Fourier series. From now on, we call formal series ÿαPIrrpGq dαtrpUαpfpαqQαuαq the random Fourier series of f P MpGq for U " pUαqαPIrrpGq P U. The main question of this paper is to find out when all of the random Fourier series are in MpGq simultaneously. It can be seen that one direction on our question can be solved simply from the Plancherel identity. Proposition 2.1. If f P L2pGq, all of random fourier series fU are in L2pGq. Proof. }fU}2 L2pGq " ÿαPIrrpGq " ÿαPIrrpGq α Uαpfpαqq dαtrpQαpfpαqU dαtrpQαpfpαqpfpαqq " }f}2 L2pGq . (cid:3) Lemma 2.2. For µ „ ÿαPIrrpGq dαtrppµpαqQαuαq P MpGq and f P PolpGq, Proof. For f " ÿαPIrrpGq xµ, f yMpGq,CrpGq " ÿαPIrrpGq dαppfpαqQαqi,j uα nαÿi,j"1 f " ÿαPIrrpGq nαÿi,j"1 dαtrppµpαqQαpfpαqq. j,i P PolpGq, dαppfpαqQαqi,jpuα j,iq 4 and xµ, f yMpGq,CrpGq " ÿαPIrrpGq " ÿαPIrrpGq " ÿαPIrrpGq " ÿαPIrrpGq " ÿαPIrrpGq j,iqyMpGq,CrpGq dαppfpαqQαqi,jxµ,puα dαppfpαqQαqi,jrpµpαqsi,j dαppfpαqQαqi,jrpµpαqsj,i nαÿi,j"1 nαÿi,j"1 nαÿi,j"1 dαtrppfpαqQαpµpαqq dαtrppµpαqQαpfpαqq. (cid:3) for all U " pUαqαPIrrpGq P U if and only if Proposition 2.3. Fix a family of matrices pXαqαPIrrpGq. Then dαtrpUαXαQαuαq P MpGq µU „ ÿαPIrrpGq µB „ ÿαPIrrpGq for all B " pBαqαPIrrpGq P ℓ8ppGq. Φ : ℓ8ppGq Ñ MpGq, B " pBαqαPIrrpGq ÞÑ µB „ ÿαPIrrpGq dαtrpBαXαQαuαq P MpGq In this case, the map dαtrpBαXαQαuαq, is automatically a bounded map by the closed graph theorem. Proof. One direction is trivial. On the other direction, pick B " pBαqαPIrrpGq P Ballpℓ8ppGqq and take an arbitrary f P PolpGq. Then xµB, f yMpGq,CrpGq " ÿαPIrrpGq Xα " ř4 For each α P IrrpGq, there exist unitaries vj j"1 vj 2 by considering the following decomposition α dαtrpBαXαQαpfpαqq. α with j " 1, , 4 such that Xα " X α ` Xα 2 ` i Xα ´ X α 2i ": h1 ` ih2 and hj ` ib1 ´ hj2 2 hj ´ ib1 ´ hj2 2 ` hj " p q for j " 1, 2. We define Vj " pvj xµB, f yMpGq,CrpGq " αqαPIrrpGq P U for j " 1, , 4 and then 1 1 2 2 dαtrpvj Since PolpGq is dense in CrpGq, we get the conclusion αAαQαpfpαqq " 4ÿj"1xµVj , f yMpGq,CrpGq. 4ÿj"1 ÿαPIrrpGq µB " ř4 j"1 µVj 2 5 P MpGq. 2.4. Vector valued probabilistic methods. In this subsection, we gather some probabilistic techniques, namely non-commutative Khintchine inequality and cotype 2 condition, to be used in this study. First of all, let us write down the non-commutative Khintchine inequality in the context of the present compact quantum group setting. For more details, see [Pis12]. (cid:3) For a sequence pxjqn j"1 Ď L1pGq, we set ~pxjqn j"1~1 " sup!ÿxxj, yjyL1pGq,L8pGq : yj P L8pGq, max!››pyjqn j"1››C ,››pyjqn 2›››››L8pGq j"1››R "›››››p j"1››C "›››››p nÿj"1 nÿj"1 j"1 be an independent and identically distributed(in short, i.i.d.) sequece of real valued Gaussian random variables with mean zero and variance 1. Then a part of the Khintchine inequality for non-commutative L1-spaces is as follows: where››pyjqn 2›››››L8pGq and››pyjqn Let pgjqn yjy j q y j yjq . 1 1 j"1››R) ď 1) Theorem 2.4 (Theorem 9.1, [Pis12]). If G is of Kac type, there exists a universal constant C ą 0 such that 1 C ~pxjqn dPpwq ď ~pxjqn j"1~1 j"1~1 ďż ››››› j"1 Ď L1pGq. nÿj"1 gjpwqxj››››› for any finite set pxjqn Also, we will make use of a cotype study on non-commutative L1-spaces. In general, we say that a Banach space is of cotype 2 if there is a constant K ą 0 such that for all finite sets pxjqn j"1 Ď B we have ż ››››› nÿj"1 ǫjpwqxj›››››B dPpwq ě Kp nÿj"1}xj}2q 1 2 , where pǫjqj is a family of i.i.d. Rademacher variables. It is known that every predual of von Neumann algebra is of cotype 2 [TJ74] and it is possible to replace the Rademacher varialbes ǫj to the Gaussian variables gj [LT13]. Theorem 2.5. There exists a universal constant K ą 0 such that ż ››››› nÿj"1 gjpwqxj›››››L1pGq j"1 Ď L1pGq. for any finite sets pxjqn dPpwq ě Kp nÿj"1 }xj}2 1 2 L1pGqq 3. The main results 3.1. Affirmative answer on Kac cases. In this subsection, we will make use of random matrices Gn " p dent real valued Gaussian random variables with mean zero and variance 1. i,j"1 P Mn where pgn i,j"1 is a family of indepen- i,jqn gn 1 ?n i,jqn Lemma 3.1 (Proposition 1.5, Chapter 5, [MP81]). There exist universal constants C1, C2 ą 0, which are independent of n, such that C1 ďż }Gnpwq} dPpwq ď C2, where }} is the operator norm of matrices. 6 Theorem 3.2. Let G be a compact quantum group of Kac type and suppose that all random fourier series fU „ ÿαPIrrpGq for all U " pUαqαPIrrpGq P U. Then we have nαtrpUαpfpαquαq of f P MpGq are in MpGq nαtrppfpαqpfpαqq ă 8. ÿαPIrrpGq K }B}ℓ8ppGq ě›››››› Proof. By Proposition 2.3, there exists a universal constant K ą 0 such that fB „ ÿαPIrrpGq rally takes into account a family of random matrices pGnαqαPS " pp nαtrpBαpfpαquαq››››››MpGq for any B " pBαqαPIrrpGq P ℓ8ppGq. Now we fix a finite subset S Ď IrrpGq and natu- in ℓ8ppGq. Then from the estimate above and Theorem 2.4, we have i,jpwqppfpαquαqj,i›››››L1pGq αPS }Gnα} dPpwq ěż ›››››ÿαPS Kż sup nαÿi,j"1 ?nαgα i,jqnα gα 1 ?nα dPpwq i,j"1qαPS (3.1) 1 ě " C ~p?nαppfpαquαqj,iqα,i,j~1 i,j qα,i,jÿxxα i,j, yα 1 C sup pyα i,jyL1pGq,L8pGq i,jqα,i,j Ď L8pGq where xα such that i,j " ?nαppfpαquαqj,i and the supremum runs over all pyα max$&% ›››››pÿαPS ,›››››pÿαPS 2›››››L8pGq 2›››››L8pGq For arbitrary pApαqqαPS such that ÿαPS nαÿi,j"1pyα nαÿi,j"1 i,jqyα i,jq yα i,jpyα i,jqq 1 1 ,.- ď 1. nαtrpApαqApαqq ď 1, we can find that the family pyα i,jq " p?nαrpApαquαqj,isq satisfies the condition above. Indeed, nαÿi,j"1pyα ÿαPS nαpApαquαqj,irpApαquαqj,is i,jqyα nαpApαquαqj,irpuαqApαqsi,j i,j " ÿαPS " ÿαPS " ÿαPS " ÿαPS " ÿαPS nαÿi,j"1 nαÿi,j"1 nαtrpApαquαpuαqApαqq nαtrpApαqApαqq1A nαtrpApαqApαqq1A 7 and ÿαPS nαÿi,j"1 i,jpyαq yα i,j " ÿαPS " ÿαPS " ÿαPS " ÿαPS " ÿαPS so that we have nαrpApαquαqj,ispApαquαqj,i nαpApαquαqj,irpuαqtApαqtsi,j nαÿi,j"1 nαÿi,j"1 nαtrpApαquαpuαqtApαqtq nαtrpApαqApαqtq1A nαtrpApαqApαqq1A, max$&% ›››››pÿαPS ď pÿαPS nαÿi,j"1pyα i,jqyα i,jq nαtrpApαqApαqqq 1 2›››››L8pGq 1 2 ď 1 ,›››››pÿαPS nαÿi,j"1 yα i,jpyα i,jqq 1 2›››››L8pGq ,.- Finally, by Lemma 3.1, we have KC2 ě " " " " sup nα nαÿi,j"1x?nαppfpαquαqj,i,?nαrpApαquαqj,isyL1pGq,L8pGq nαÿi,j"1 nαÿi,j nαtrppfpαqApαqq nαÿk,l"1 pfpαqj,kApαqj,lxuα nαÿk,l"1 nα pfpαqj,kApαqj,l l,iqyL1pGq,L8pGq k,i,puα δk,l nα 1 sup sup 1 C 1 C 1 C pApαqqαPS ÿαPS pApαqqαPS ÿαPS pApαqqαPS ÿαPS pApαqqαPS ÿαPS C pÿαPS nαtrppfpαqpfpαqqq 1 C sup 1 2 since the supremum runs over all pApαqqαPS such that ÿαPS Now we reach the conclusion since the finite set S is chosen to be arbitrary. nαtrpApαqApαqq ď 1. (cid:3) Remark 3.3. In the category of compact groups, it seems that some experts already have their own way to understand Theorem 1.1 through the Khintchine inequality. Indeed, the arguments below (3.1) are simplified as follows. 8 dPpwq dPpwqdx ?nπgπ ?nπgπ i,jpwqppfpπqπpxqqj,i›››››L1pGq ż ›››››ÿπPS nπÿi,j"1 "żGż ÿπPS i,jpwqppfpπqπpxqqj,i nπÿi,j"1 ÁżGpÿπPS nπppfpπqπpxqqj,i nπÿi,j"1 "żGpÿπPS nπtrppfpπqπpxqπpxqpfpπqqq " pÿπPS nπtrppfpπqpfpπqqq 2 dx 2 dx 2 . q 2 1 1 1 Remark 3.4. Let G be a compact quantum group of Kac type. If we suppose that ÿαPIrrpGq nαtrpX α Xαq ă 8, then the multiplier is bounded for all 2 ď p ď 8 since LppGq Ď L2pGq for such p. The converse is also true thanks to Theorem 3.2. Indeed, boundedness of Φ implies that the multiplier Φ : LppGq Ñ ℓ1ppGq, f ÞÑ pXαpfpαqqαPIrrpGq, Ψ : ℓ8ppGq Ñ MpGq,pAαqαPIrrpGq ÞÑ ÿαPIrrpGq dαtrpAαX αuαq is bounded. Then, by Proposition 2.3 and Theorem 3.2, we can conclude that the space of such multipliers from LppGq into ℓ1ppGq can be identified with L2pGq inde- pendently of the choice of 2 ď p ď 8. 3.2. A partial answer on general cases. In spite of the complete answer on the Kac case, the proof of Theorem 3.2 could not be applied for the general case for now. However, using the cotype 2 property of L1pGq, we are able to achieve a convincing conclusion for non-Kac cases. Lemma 3.5. (1) For any α P IrrpGq and any 1 ď i, j ď nα, we have Aikpuα α qi,k2q αqj,jp ď pQ ´ 1 2 . 1 2 2 1 (2) For any α P IrrpGq and any 1 ď i, j ď nα, we have Proof. (1) Since Q is a unitary representation, we have nαÿk"1pAαQ nαÿk"1pBQ p ´ 1 ě pQ 2 α qj,j dα 1 2 αqi,k2q 1 2 . 2 2 1 ´ 1 α α uαQ ››››› k,jq›››››L8pGq nαÿk"1 ››››› k,j›››››L1pGq nαÿk"1pBQαqi,kuα α qj,j››››› nαÿk"1 " ››››› nαÿk"1pAQ nαÿk"1 pAQ nαÿk"1 pAQ α qi,k2q α qi,k2q α qi,krQ Ai,kpuα ď p pQ "p 2 p ´ 1 ´ 1 ´ 1 ´ 1 2 . 2 2 2 1 1 2 1 2 α uQ ´ 1 k,jq›››››L8pGq α sk,j›››››L8pGq nαÿk"1rQ α uQ ´ 1 2 2 1 9 2 α s k,jrQ 1 2 α uQ ´ 1 2 α sk,jq 1 2 (2) Define wi,k " pAQ p1q, we can see that ´ 1 k,j›››››L1pGq nαÿk"1pBQαqi,kuα x 2 ď1 ››››› nαÿk"1pBQαqi,kuα přnα α qj,j " pQ k"1 wi,k2q sup ě ´ 1 2 1 " pQ ´ 1 2 α qj,j 2 1 2 ´ 1 α qj,j k,j,pQ Ai,kpuα nαÿk"1 nαÿk"1pBQαqi,kwi,kpQ αqk,khppuα nαÿk"1 2 " pQ αqi,k pBQ dα α qj,j dα wi,k nαÿk"1 pBQ ´ 1 p 2 2 1 1 1 sup k"1 wi,k2q 1 2 ď1 sup 1 1 2 ď1 k"1 wi,k2q αqi,k2 pBQ d2 α 2 přnα přnα nαÿk"1 k,jqyL1pGq,L8pGq k,jquα k,jq 2 α qi,k for all 1 ď i, k ď nα for convenience. Then, by ´ 1 2 q " pQ α qj,jp αqi,k2q Theorem 3.6. Suppose that all random Fourier series fU „ ÿαPIrrpGq are in MpGq for all U " pUαqαPIrrpGq P U. Then we have 2 2 1 (cid:3) dαtrpUαpfpαqQαuαq Proof. As in the proof of Theorem 3.2, there exists a universal constant K ą 0, which is independent of the choice of finite subset S Ď IrrpGq, such that dPpwq gα j,ipwq where pgα zero and variance 1. Let vα Theorem 2.5, we have that ÿαPIrrpGq dα nα dα?nα nαÿi,j"1 j,ipwqqαPIrrpGq,1ďi,jďnα is a i.i.d. trpQαpfpαqpfpαqq ă 8. k,j›››››L1pGq K ěż ›››››ÿαPS nαÿk"1ppfpαqQαqi,kuα i be the i-th row vector of the matrix pfpαqQ nα››››› k,j››››› nαÿi,j"1 nαÿk"1ppfpαqQαqi,kuα nαÿi,j"1 i }2 α qj,j }vα trpQαpfpαqpfpαqqq K ě K 1pÿαPS ě K 1pÿαPS " pÿαPS pQ´1 dα nα L1pGq d2 α nα 2 . q q 2 1 2 1 2 1 1 2 family of Gaussian variables with mean α . Then, by Lemma 3.5 and Since S Ď IrrpGq is arbitrary, we can conclude that ÿαPIrrpGq dα nα trpQαpfpαqpfpαqq ă 8. (cid:3) Remark 3.7. Although Theorem 3.6 does not give the complete conclusion for non-Kac cases, we can expect that the quantity gets very close to the desired quantity of dα. At least in certain cases, nα is significantly smaller than the quantum dimension dα. Indeed, if the function α ÞÑ nα has subexponential growth and if G 10 dα nα is of non-Kac type, then the function α ÞÑ dα is of exponential growth [DPR16]. In particular, the Drinfeld-Jimbo deformations Gq with 0 ă q ă 1 are important non-Kac type compact quantum groups. In this case, the function α ÞÑ nα is of subexponential growth. Let us clarify what Remark 3.7 implies for concrete example G " SUqp2q. Then we have Corollary 3.8. Let G " SUqp2q with 0 ă q ă 1 and suppose that all random fourier series fU „ ÿαPIrrpGq dαtrpUαpfpαqQαuαq are in MpGq for all U " pUαqαPIrrpGq P U. ÿαPIrrpGq α trpQαpfpαqpfpαqq ă 8 f or each ǫ ą 0. Proof. It is known that IrrpGq is identified with t0u Y N and nk " k ` 1 for all k P IrrpGq " t0u Y N. Also, dk " q´k ` q´k`2 ` ` qk ě q´k for all k ě 0. Therefore, for each ǫ ą 0, d1´ǫ d1´ǫ k ÿkě0 dk nk dǫ nk k k ` 1 q´ǫk trpQkpfpkqpfpkqq ď ÿkě0 ď ÿkě0 p1 ´ qǫq2 ÿkě0 dk nk ď 1 trpQkpfpkqpfpkqq trpQkpfpkqpfpkqq trpQkpfpkqpfpkqq ă 8. dk nk (cid:3) 3.3. A remark on central forms. While Theorem 3.6 clearly contains insufficient points to reach the conclusion, we conjecture that the ultimate conclusion might turn out to be positive. The reason is that if the starting measure f P MpGq has a specific form, the so-called central form, it can offset the deficit considerably. Throughout this subsection, we will consider a sub von Neumann algebra N generated by tχαuαPIrrpGq in L8pGq. Then the restriction of Haar state to N is a finite tracial state since hpχαχβq " hpχ Lemma 3.9. Let G be a compact quantum group satisfying , χαq " hpχβχαq for all α, β P IrrpGq. hpχαq ą 0. Then there exists a universal constant K satisfying the following: for any sequence aα2 ď 1, there exists pTiqi Ď PolpGq X K BallpL8pGqq αχβq " δα,β " δβ,α " hpχ αPIrrpGq inf β paαqαPIrrpGq with ÿαPIrrpGq hpTiχ such that lim i αq ě aα for all α P IrrpGq. Proof. By [Theorem 5, [LP97]], there exists a universal constant K ą 0 and for any cαχα such sequence paαqαPIrrpGq, there exists T P K BallpNq such that T „ ÿαPIrrpGq with cα ě aα for all α P IrrpGq. Lastly, since A " PolpGq X N is a unital -algebra and dense in N under the weak -topology, Ball(Aq is dense in BallpNq by the Kaplansky's density theorem. Thus, we obtain a net pTiqi Ď PolpGq X K BallpL8pGqq satisfying cα " hpT χ αq " lim i hpTiχ αq ě aα for all α P IrrpGq. 11 (cid:3) αPIrrpGq hpχαq ą 0 Theorem 3.10. Let G be a compact quantum group such that inf ÿαPIrrpGq MpGq for all U " pUαqαPIrrpGq P U. Then we have dαtrpUαpfpαqQαuαq are in aud suppose that all random Fourier series fU „ ÿαPIrrpGq trppfpαqq2 ă 8. Proof. Let paαqαPIrrpGq be a sequence with ÿαPIrrpGq aα2 ď 1. Then we have a fam- ily tTiui Ď PolpGq X K BallpL8pGqq coming from Lemma 3.9. Let us write Ti as ÿαPIrrpGq Idnα for all α P IrrpGq. Therefore, using αχα. Then pTipαqQα " Proposition 2.3, there exists a universal constant K 1 ą 0 such that ci α dα ci 8 ą K 1 ě sup " sup U U U ci sup " sup i yL1pGq,L8pGq i xfU , T sup i ÿαPIrrpGq dαtrpUαpfpαqQαpTipαqq sup i ÿαPIrrpGq αtrpUαpfpαqq α trppfpαqq i ÿαPIrrpGqci cα trppfpαqq ě ÿαPIrrpGq aα trppfpαqq. ě ÿαPIrrpGq " sup The conclusion can be reached as stated above since paαq is arbitrary. inf Corollary 3.11. Let G be a compact quantum group such that (cid:3) hpχαq ą 0 cαχα P MpGq. If all random Fourier series αPIrrpGq and fix a central element f „ ÿαPIrrpGq fU „ ÿαPIrrpGq have dαtrpUαpfpαqQαuαq are in MpGq for all U " pUαqαPIrrpGq P U, then we ÿαPIrrpGq dαtrpQαpfpαqpfpαqq " ÿαPIrrpGq cα2 ă 8. Proposition 3.12. The quantum groups listed in the following list are known to satisfy the condition of Corollary 3.11: ‚ Free orthogonal quantum groups O` ‚ Quantum automorphism group GautpB, ψq with a δ-form ψ ‚ Drinfeld-Jimbo q deformations Gq with 0 ă q ă 1 F Proof. Let G1 and G2 be compact quantum groups. If there exists a bijective map Φ : IrrpG1q Ñ IrrpG2q such that Φpπ1 b π2q " Φpπ1q b Φpπ2q and Φp'n i"1πiq " 'n i"1Φpπiq for all πi P IrrpG1q and n ě 1, then the map Φ extends to a bijective -homomorphism Φ : N1 Ñ N2 where Nj is the weak -closure of span χj each j " 1, 2. Also, recall that the Haar states hj are tracial on Nj respectiely. α(αPIrrpGj q in L8pGjq for 12 Moreover, Φ is trace-preserving by repeating the proofs of [Proposition 6.7, [Wan17]] and [Lemma 4.7, [You18]]. Therefore, we can conclude that On the other hand, in [Pri75], it is shown that αq " h2pχ2 h1pχ1 Φpαqq for all α P IrrpGq. inf πPIrrpGq}χπ}L1pGq ą 0 for any compact connected Lie group. The existence of the bijective map Φ : IrrpGq Ñ IrrpGq is known for pG, Gq " N , SUp2qq, pGautpB, ψq, SOp3qq and pGq, Gq. Refer to [Bra13], [NT13] and [Tim08] pO` for details. (cid:3) 4. Application to complete representability problem In the category of Banach algebras, C-algebras are in a special position. One of typical examples that are not C-algebras is the convolution algebra L1pGq of a locally compact group. Moreover, it turned out that L1pGq is isomorphic to a closed subalgebra of BpHq for some Hilbert H as Banach algebras if and only if G is finite. This fact is based on the study of the Arens irregularity for convolution algebras L1pGq [You73], which shows L1pGq is Arens regular only if G is finite. In the framework of locally compact quantum groups, the Fourier algebra ApGq is understood as a dual object of the convolution algebra L1pGq in view of the generalized Pontrjagin duality. Very recently, it has shown that the Fourier algebra is completely isomorphic to a closed subalgebra of BpHq for some Hilbert space if and only if G is finite ([LY17] for discrete groups and [LSS16] for general cases). Here, complete isomorphism is the natural isomorphism in the category of operator spaces. Some terminologies of operator space theory required in this section will be explained in the below. To my knowledge, prior to [LY17], all conclusions in this direction were based on Arens regularity studies. In the quantum group setting, the situation is similar. Thanks to [HNR12] investigating Arens irregularity for convolution algebras L1pGq of locally compact quantum groups, we know that the convolution algebra L1pGq is completely isomorphic to a closed subalgebra of BpHq if and only if L8pGq is finite dimensional for G a co-amenable compact quantum group. Our contribution in this direction is the following fact, which removes the addi- tional condition of co-amenability in the category of compact quantum groups. Theorem 4.1. Let G " pA, ∆q be a compact quantum group. Then L1pGq is completely representable as an operator algebra, i.e. L1pGq is completely isomorphic to a closed subalgebra of BpHq for some Hilbert space H if and only if A is finite dimensional. An abstarct operator space is a vector space E equipped with a matrical norm structure on MnpEq " Mn b E for which (1) }v1 ' v2}Mm`npEq ď maxt}v1}m ,}v2}nu for all v1 P MmpEq and v2 P for all a P Mm,n, b P Mn,m and (2) }avb}MmpEq ď }a}Mm,n }v}MnpEq }b}Mn,m MnpV q and v P MnpEq. We say that a linear map L : E Ñ F between the two operator spaces E and F is completely bounded if sup nPN }idn b L : MnpEq Ñ MnpFq} ă 8. The two operator spaces that are noted in this section are the convolution algebra L1pGq and the extended Haagerup tensor product space L8pGq beh L8pGq. Each 13 spaces has a natural operator space structure. Let us present equivalent description of those operator spaces. First of all, for F " pfi,jqn i,j"1 P MnpL1pGqq, the natural matricial norm structure is given by }F}MnpL1pGqq " sup mPN sup X ››pxfi,j, xk,lyL1pGq,L8pGqqn i,j"1 , m k,l"1››Mmn where the supremum runs over all X " pxk,lqm identification MmpL8pGqq Ď MmpBpL2pGqqq -- BpL2pGq ' ' L2pGqq. k,l"1 P BallpMmpL8pGqqq under the Secondly, ai b bi : ÿi aia i and ÿi i bi converges+ b with respect to the weak -topologies and the matricial norm structure comes from the completely isometric embedding into CBσpBpL2pGqq, BpL2pGqqq. More pre- i,j"1 P Mn b pL8pGq beh L8pGqq is ai,j k b bi,j k qn L8pGq beh L8pGq "#ÿi cisely, the matricial norm for pÿk s,t"1›››››pÿk given by T "pT s,tqm sup mPN sup ai,j k T s,tbi,jqn i,j"1 m s,t"1›››››MmnpBpL2pGqqq , where the supremum runs over all T " pT s,tq1ďs,tďm P BallpMmpBpL2pGqqqq. Note that the restricted comultiplication ∆PolpGq ∆ : L8pGq Ñ L8pGqbL8pGq and it induces the natural Banach algebraic structure on L1pGq. For f1, f2 P L1pGq, we define the convolution product of f1 and f2 by extends to a normal -homomorphism xf1 f2, ayL1pGq,L8pGq " pf1 b f1qp∆paqq for all a P L8pGq. Moreover, the convolution product extends to a completely contractive map m " category of operator spaces. Refer to [Pis03] or [ER00] for details. The result of [Ble95], which actually covers general completely contractive Banach algebras, is written as follows in our setting: ∆ : L1pGqpbL1pGq Ñ L1pGq where pb is the projective tensor product in the Proposition 4.2. The convolution algebra L1pGq is completely isomorphic to a closed subaglebra of BpHq for some Hilbert space H if and only if the comultiplica- tion ∆ : L8pGq Ñ L8pGq beh L8pGq is completely bounded. Lemma 4.3. Given B " pBαqαPIrrpGq P Ballpℓ8ppGqq, define an operator TB P p,jqpBαqp,i. Then TB is a α puαqBαsj,i " j,i ÞÑ rQ´1 α qj,jpuα nαÿp"1pQ´1 BpL2pGqq by uα contraction. Proof. Take v " ÿαPIrrpGq nαÿi,j"1 TBv " ÿαPIrrpGq cα i,j uα j,i. Then nαÿj,p"1r nαÿi"1 i,jpQ´1 cα α qj,jpBαqp,ispuα p,jq. 14 Also, put Cα " pcα i,jq1ďi,jďnα . Then }TBv}2 L2pGq " ÿαPIrrpGq " ÿαPIrrpGq " ÿαPIrrpGq ď ÿαPIrrpGq " ÿαPIrrpGq nαÿj,p"1 nαÿi"1 nαÿj,p,i,i1"1pQ´1 trpQ´1 α C α B i,jpQ´1 cα α qj,jpBαqp,i2pQαqj,j dα α qj,j cα i,jpBαqp,icα i1,jpBαqp,i1 αBαCαq 1 dα α Cαq 1 dα α C trpQ´1 nαÿi,j"1pQ´1 α qj,jcα i,j2 1 dα " }v}2 L2pGq . 1 dα (cid:3) Proof of Theorem 4.1. One direction is trivial. For the other direction, suppose that ∆ : L8pGq Ñ L8pGq beh L8pGq is completely bounded. Then, for any j,i P PolpGq, f " ÿαPIrrpGq nαÿk"1 nαÿi,j"1 dαppfpαqQαqi,j uα j,k b uα nαÿi,j"1 dαppfpαqQαqi,j uα ∆pfq " ÿαPIrrpGq For each pBαqαPIrrpGq P Ballpℓ8ppGqq, pick TB of Lemma 4.3. Then p∆pfqqpTBq x :" ÿαPIrrpGq k,i P L8pGq beh L8pGq. α qk,kpBαqp,iuα nαÿi,j,k,p"1 p,kq P L2pGq. dαppfpαqQαqi,jpQ´1 j,kpuα sends 1A P L2pGq to Then }∆}cb }f}L8pGq ě }x}L2pGq ě hpxq " ÿαPIrrpGq " ÿαPIrrpGq 1 dα pBαqj,i nαÿi,j,k"1 dαppfpαqQαqi,j nαtrppfpαqQαBαq. nα p mined by Since Bα is arbitrarily chosen, we have a bounded map Φ : CrpGq Ñ ℓ1ppGq, f ÞÑ dα pfpαqqαPIrrpGq and it induces the bounded dual map Φ : ℓ8ppGq Ñ MpGq deter- xf, ΦpAqyCrpGq,MpGq " ÿαPIrrpGq IdnαqQαuαq, is bounded since for any f P PolpGq we have Now, we can find that the map Ψ : ℓ8ppGq Ñ MpGq, A " pApαqqαPIrrpGq ÞÑ µA „ ÿαPIrrpGq nαtrppfpαqQαApαqq. dαtrpAαp nα dα xf , ΨpAqyCrpGq,MpGq " xf, ΦpAqyCrpGq,MpGq by Lemma 2.2. 15 Finally, Proposition 2.3 and Theorem 3.6 say that ÿαPIrrpGq dα nα trpQα n2 α d2 α Idnαq " ÿαPIrrpGq nα ă 8, so that A should be finite dimensional. (cid:3) Acknowledgement. The author is grateful to Professor Hun Hee Lee for his encouragement and to Professor Gilles Pisier for his helpful comments, particularly on Remark 3.3. References [Ble95] David P Blecher. A completely bounded characterization of operator algebras. Mathe- matische Annalen, 303(1):227 -- 239, 1995. [Bra13] Michael Brannan. Reduced operator algebras of trace-perserving quantum automor- phism groups. Doc. Math., 18:1349 -- 1402, 2013. [Cas13] Martijn Caspers. The lp -fourier transform on locally compact quantum groups. J.Operator Theory, 2013. [Daw10] Matthew Daws. Operator biprojectivity of compact quantum groups. Proceedings of the American Mathematical Society, 138(4):1349 -- 1359, 2010. [DPR16] Alessandro D'Andrea, Claudia Pinzari, and Stefano Rossi. Polynomial growth for com- pact quantum groups, topological dimension and -regularity of the fourier algebra. Preprint, arXiv:1602.07496, 2016. [Dud67] Richard M Dudley. The sizes of compact subsets of hilbert space and continuity of [ER00] [Fac87] [Fer75] gaussian processes. Journal of Functional Analysis, 1(3):290 -- 330, 1967. Edward G.. Effros and Zhong-Jin Ruan. Operator spaces. Clarendon Press, 2000. Thierry Fack. Type and cotype inequalities for non commutative l p-spaces. Journal of Operator Theory, pages 255 -- 279, 1987. Xavier Fernique. Des resultats nouveaux sur les processus gaussients. In S´eminaire de Probabilit´es IX Universit´e de Strasbourg, pages 318 -- 335. Springer, 1975. [FHL`17] Uwe Franz, Guixiang Hong, Francois Lemeux, Michael Ulrich, and Haonan Zhang. Hypercontractivity of heat semigroups on free quantum groups. Journal of Operator Theory, 77(1):61 -- 76, 2017. [FTR66] Alessandro Fig`a-Talamanca and Daniel Rider. A theorem of littlewood and lacunary [Hel57] series for compact groups. Pacific Journal of Mathematics, 16(3):505 -- 514, 1966. Sigurdur Helgason. Topologies of group algebras and a theorem of littlewood. Transac- tions of the American Mathematical Society, 86(2):269 -- 283, 1957. [HNR12] Zhiguo Hu, Matthias Neufang, and Zhong-Jin Ruan. Module maps over locally compact quantum groups. Studia Mathematica, 211:111 -- 145, 2012. [Hun51] GA Hunt. Random fourier transforms. Transactions of the American Mathematical Society, 71(1):38 -- 69, 1951. [JMP14] Marius Junge, Tao Mei, and Javier Parcet. Smooth fourier multipliers on group von neumann algebras. Geom. Funct. Anal, 24(6):1913 -- 1980, 2014. [JPPP17] Marius Junge, Carlos Palazuelos, Javier Parcet, and Mathilde Perrin. Hypercontractiv- [Kah93] [Kus01] [KV00] [KV03] [Kwa76] [Lit26] [LP86] [LP97] ity in group von neumann algebras. 2017. Jean-Pierre Kahane. Some random series of functions, volume 5. Cambridge University Press, 1993. Johan Kustermans. Locally compact quantum groups in the universal setting. Internat. J. Math., 12(3):289 -- 338, 2001. Johan Kustermans and Stefaan Vaes. Locally compact quantum groups. Ann. Sci. ´Ecole Norm. Sup. (4), 33(6):837 -- 934, 2000. Johan Kustermans and Stefaan Vaes. Locally compact quantum groups in the von Neumann algebraic setting. Math. Scand., 92(1):68 -- 92, 2003. S Kwapie´n. A theorem on the rademacher series with vector valued coefficients. Prob- ability in Banach spaces, pages 157 -- 158, 1976. JE Littlewood. On the mean values of power series. Proceedings of the London Mathe- matical Society, 2(1):328 -- 337, 1926. Fran¸coise Lust-Piquard. In´egalit´es de khintchine dans cp (1¡ p¡). CR Acad. Sci. Paris, 303:289 -- 292, 1986. Fran¸coise Lust-Piquard. On the coefficient problem: a version of the kahane -- katznelson -- de leeuw theorem for spaces of matrices. journal of functional analysis, 149(2):352 -- 376, 1997. 16 [LPP91] Fran¸coise Lust-Piquard and Gilles Pisier. Non commutative khintchine and paley in- [LSS16] [LT13] [LY17] equalities. Arkiv for matematik, 29(1):241 -- 260, 1991. Hun Hee Lee, Ebrahim Samei, and Nico Spronk. Similarity degree of Fourier algebras. J. Funct. Anal., 271(3):593 -- 609, 2016. Michel Ledoux and Michel Talagrand. Probability in Banach Spaces: isoperimetry and processes. Springer Science & Business Media, 2013. Hun Hee Lee and Sang-Gyun Youn. New deformations of convolution algebras and fourier algebras on locally compact groups. Canad.J.Math., 69(2):434 -- 452, 2017. [Mar78] Michael B Marcus. Continuity and the central limit theorem for random trigonometric series. Probability Theory and Related Fields, 42(1):35 -- 56, 1978. [MP78] Michael B Marcus and Gilles Pisier. Necessary and sufficient conditions for the uniform convergence of random trigonometric series. 1978. [NT13] [MP81] Michael B Marcus and Gilles Pisier. Random Fourier Series with Applications to Har- monic Analysis.(AM-101), volume 101 of Annals of Mathematics Studies. Princeton University Press; University of Tokyo Press, 1981. Sergey Neshveyev and Lars Tuset. Compact quantum groups and their representation categories, volume 20 of Cours Sp´ecialis´es [Specialized Courses]. Soci´et´e Math´ematique de France, Paris, 2013. G Pisier. Sur l'espace de banach des s´eries de fourier al´eatoires presque surement con- tinues. S´eminaire Analyse fonctionnelle (dit, pages 1 -- 33, 1977. Gilles Pisier. Ensembles de sidon et processus gaussiens. CR Acad. Sci. Paris, 286(15):A671 -- A674, 1978. Gilles Pisier. Introduction to operator space theory, volume 294 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2003. Gilles Pisier. Grothendiecks theorem, past and present. Bulletin of the American Math- ematical Society, 49(2):237 -- 323, 2012. JF Price. On local central lacunary sets for compact lie groups. Monatshefte fur Math- ematik, 80(3):201 -- 204, 1975. REAC Paley and A Zygmund. On some series of functions,(3). In Mathematical Pro- ceedings of the Cambridge Philosophical Society, volume 28, pages 190 -- 205. Cambridge University Press, 1932. Daniel Rider. Randomly continuous functions and sidon sets. Duke Math. J, 42(4):759 -- 764, 1975. Raphael Salem and Antoni Zygmund. Some properties of trigonometric series whose terms have random signs. Acta Mathematica, 91(1):245 -- 301, 1954. [Rid75] [Pri75] [PZ32] [Pis77] [Pis78] [Pis03] [Pis12] [SZ54] [Tim08] Thomas Timmermann. An invitation to quantum groups and duality. EMS Textbooks in Mathematics. European Mathematical Society (EMS), Zurich, 2008. From Hopf al- gebras to multiplicative unitaries and beyond. N Tomczak-Jaegermann. On the rademacher averages and the moduli of convexity and smoothness of the schatten classes sp. Studia Math, 50:163 -- 182, 1974. Simeng Wang. Lacunary fourier series for compact quantum groups. Communications in Mathematical Physics, 349(3):895 -- 945, 2017. [Wan17] [TJ74] [Wor87a] Stanis law L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys., 111(4):613 -- 665, 1987. [Wor87b] Stanis law L. Woronowicz. Twisted SUp2q group. An example of a noncommutative [You73] [You18] differential calculus. Publ. Res. Inst. Math. Sci., 23(1):117 -- 181, 1987. NJ Young. The irregularity of multiplication in group algebras. The Quarterly Journal of Mathematics, 24(1):59 -- 62, 1973. Sang-Gyun Youn. Hardy-littlewood inequalities on compact quantum groups of Kac type. Analysis & PDE, 11(1):237 -- 261, 2018. DOI 10.2140/apde.2018.11.237. Sang-Gyun Youn : Department of Mathematical Sciences, Seoul National University, San56-1 Shinrim-dong Kwanak-gu, Seoul 151-747, Republic of Korea E-mail address: [email protected] 17
1208.5939	2	1208	2013-06-11T14:08:19	Approximation properties for noncommutative Lp-spaces associated with lattices in Lie groups	[ "math.OA", "math.FA", "math.GR" ]	In 2010, Lafforgue and de la Salle gave examples of noncommutative Lp-spaces without the operator space approximation property (OAP) and, hence, without the completely bounded approximation property (CBAP). To this purpose, they introduced the property of completely bounded approximation by Schur multipliers on Sp and proved that for p < 4/3 and p > 4 the groups SL(n,Z), with n \geq 3, do not have it. Since for 1 < p < \infty the property of completely bounded approximation by Schur multipliers on Sp is weaker than the approximation property of Haagerup and Kraus (AP), these groups were also the first examples of exact groups without the AP. Recently, Haagerup and the author proved that also the group Sp(2,R) does not have the AP, without using the property of completely bounded approximation by Schur multipliers on Sp. In this paper, we prove that Sp(2,R) does not have the property of completely bounded approximation by Schur multipliers on Sp for p < 12/11 and p > 12. It follows that a large class of noncommutative Lp-spaces does not have the OAP or CBAP.	math.OA	math	APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES ASSOCIATED WITH LATTICES IN LIE GROUPS TIM DE LAAT Abstract. In 2010, Lafforgue and de la Salle gave examples of noncommu- tative Lp-spaces without the operator space approximation property (OAP) and, hence, without the completely bounded approximation property (CBAP). To this purpose, they introduced the property of completely bounded approx- imation by Schur multipliers on S p, denoted APSchur p,cb , and proved that for p ∈ [1, 4 3 ) ∪ (4, ∞] the groups SL(n, Z), with n ≥ 3, do not have the APSchur p,cb . Since for p ∈ (1, ∞) the APSchur p,cb is weaker than the approximation property of Haagerup and Kraus (AP), these groups were also the first examples of exact groups without the AP. Recently, Haagerup and the author proved that also the group Sp(2, R) does not have the AP, without using the APSchur p,cb . In this pa- p,cb for p ∈ [1, 12 per, we prove that Sp(2, R) does not have the APSchur 11 ) ∪ (12, ∞]. It follows that a large class of noncommutative Lp-spaces does not have the OAP or CBAP. 1. Introduction Let M be a finite von Neumann algebra with normal faithful trace τ . For 1 ≤ p < ∞, the noncommutative Lp-space Lp(M, τ ) is defined as the completion of M with p , and for p = ∞, we put L∞(M, τ ) = M respect to the norm kxkp = τ ((x∗x) with operator norm. In [23], Kosaki showed that noncommutative Lp-spaces can be realized by interpolating between M and L1(M, τ ). This leads to an operator space structure on them, as described by Pisier [27] (see also [20]). 2 ) p 1 An operator space E is said to have the completely bounded approximation property (CBAP) if there exists a net Fα of finite-rank maps on E such that supα kFαkcb < C for some C > 0, and limα kFαx − xk = 0 for every x ∈ E. The infimum of all possible C's is denoted by Λ(E). If Λ(E) = 1, we say that E has the completely contractive approximation property (CCAP). An operator space E is said to have the operator space approximation property (OAP) if there exists a net Fα of finite-rank maps on E such that limα k(idK(ℓ2) ⊗Fα)x − xk = 0 for all x ∈ K(ℓ2) ⊗min E. Here K(ℓ2) denotes the space of compact operators on the Hilbert space ℓ2. The CBAP goes back to De Canni`ere and Haagerup [5], and the OAP was defined by Effros and Ruan [9]. By definition, the CCAP implies the CBAP, which in turn implies the OAP. Recall that a lattice in a Lie group G is a discrete subgroup Γ of G such that G/Γ has finite invariant measure. In this paper, we consider noncommutative Lp-spaces of the form Lp(L(Γ)), where L(Γ) is the group von Neumann algebra of a lattice Γ in a connected simple Lie group G. Such a von Neumann algebra L(Γ) is finite and 2010 Mathematics Subject Classification: Primary: 46B28; Secondary: 22D25, 46L07. The author is supported by the Danish National Research Foundation through the Centre for Symmetry and Deformation. 1 2 TIM DE LAAT has canonical trace τ : x 7→ hxδ1, δ1i, where δ1 ∈ ℓ2(Γ) is the characteristic function of the unit element 1 ∈ Γ. It was proved by Junge and Ruan [20, Proposition 3.5] that if Γ is a weakly amenable (countable) discrete group, then for p ∈ (1,∞), the noncommutative Lp- space Lp(L(Γ)) has the CBAP. Recall that connected simple Lie groups of real rank zero are amenable. By the work of Cowling and Haagerup [6] and Hansen [17], all connected simple Lie groups of real rank one are weakly amenable. This implies that for every p ∈ (1,∞) and every lattice Γ in a connected simple Lie group G of real rank zero or one, the noncommutative Lp-space Lp(L(Γ)) has the CBAP. The existence of noncommutative Lp-spaces without the CBAP follows from the work of Szankowski [29]. The first concrete examples were given recently by Laf- forgue and de la Salle [24]. They proved that for all p ∈ [1, 4 3 )∪(4,∞] and all lattices Γ in SL(n, R), where n ≥ 3, the space Lp(L(Γ)) does not have the OAP (or CBAP). They also proved analogous results for lattices in Lie groups over nonarchimedean fields. In their work, the failure of the OAP for the aforementioned noncommu- tative Lp-spaces follows from the failure of a certain approximation property for the groups SL(n, R). This property, called the property of completely bounded approximation by Schur multipliers on Sp (see Section 2.6), denoted APSchur p,cb , was introduced by Lafforgue and de la Salle exactly to this purpose. p,cb is weaker than the AP. In this way, the APSchur Other approximation properties for groups (see [3]), e.g., amenability, weak amenability, and the approximation property of Haagerup and Kraus (AP) (see [14]), are related to the APSchur p,cb . It is well-known that amenability of a group G (strictly) implies weak amenability, which in turn (strictly) implies the AP. For p ∈ (1,∞), the APSchur p,cb gave rise to the first example of an exact group without the AP, namely SL(3, Z). Recently, Haagerup and the author proved that also Sp(2, R) does not have the AP [15], in a more direct way than Lafforgue and de la Salle did for SL(3, R). Indeed, the APSchur p,cb was not used in the proof. On the other hand, as was mentioned earlier, the method of Lafforgue and de la Salle also gives information about approximation properties of certain noncommutative Lp-spaces. For this, it is actually crucial to use the APSchur p,cb . Haagerup and the author also proved that all connected simple Lie groups with finite center and real rank greater than or equal to two do not have the AP, building on the failure of the AP for both SL(3, R) and Sp(2, R). The following are the main results of this article. 11 ) ∪ (12,∞], the group Sp(2, R) does not have the Theorem 3.1. For p ∈ [1, 12 APSchur p,cb . Theorem 4.3. Let p ∈ [1, 12 11 ) ∪ (12,∞], and let Γ be a lattice in a connected simple Lie group with finite center and real rank greater than or equal to two. Then Lp(L(Γ)) does not have OAP (or CBAP). The paper is organized as follows. In Section 2, we recall some preliminary results, and we make a study of Schur multipliers on Schatten classes corresponding to (compact) Gelfand pairs, which provides us with suitable tools for our proof. In Section 3, we prove Theorem 3.1, and in Section 4, we prove Theorem 4.3. 2. Preliminaries 2.1. Schur multipliers on Schatten classes. This section partly follows the exposition of [24, Section 1]. More details can be found there. APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 3 p 2 ) 1 n = Sp(ℓ2 For p ∈ [1,∞] and a (separable) Hilbert space H, let Sp(H) denote the pth Schatten class on H. Recall that S∞(H) is the Banach space K(H) of compact operators (with operator norm) on H, and for p ∈ [1,∞), the space Sp(H) consists of the operators T on H such that kTkp = Tr((T ∗T ) p < ∞, where Tr denotes the (semifinite) trace on B(H). In this way, Sp(H) is a Banach space for all p ∈ [1,∞]. n) and Sp = Sp(ℓ2). Note that the space S2(H) We use the notation Sp corresponds to the Hilbert-Schmidt operators on H. Schatten classes can be realized by interpolating between certain noncommuta- tive Lp-spaces in the semifinite setting. Indeed, we have Sp(H) = Lp(B(H), Tr). Noncommutative Lp-spaces in the semifinite setting can be defined analogously to the finite case, which was described in Section 1. For details, see [28]. The natural operator space structure on Sp(H) follows from [27]. For our purposes, the following characterization of the completely bounded norm of a linear map between Schatten classes is important. Recall that Sp(H)⊗ Sp(K) (algebraic tensor product) embeds naturally into Sp(H⊗K) (Hilbert space tensor product). Let T : Sp(H) −→ Sp(H) be a bounded linear map, and let K = ℓ2. Then T is completely bounded if the map T ⊗ idSp extends to a bounded linear map on Sp(H ⊗ ℓ2), and we have kTkcb = kT ⊗ idSp k = supn∈N kT ⊗ idSp n k (see [28, Lemma 1.7]). A linear map T : Mn(C) −→ Mn(C) of the form [xij ] 7→ [ψij xij ] for some matrix ψ ∈ Mn(C) is called a Schur multiplier on Mn(C). More precisely, the operator T is called the Schur multiplier on Mn(C) with symbol ψ, and it is also denoted by Mψ. In what follows, we need more general notions of Schur multipliers. Let (X, µ) be a σ-finite measure space. Let k ∈ L2(X × X, µ ⊗ µ). It is well-known that the map Tk : L2(X, µ) −→ L2(X, µ) defined by (Tkf )(x) = is a Hilbert-Schmidt operator on L2(X, µ). Conversely, if RX k(x, y)f (y)dµ(y), T ∈ S2(L2(X, µ)), then T = Tk for some k ∈ L2(X × X, µ ⊗ µ). In this way, we can identify S2(L2(X, µ)) with L2(X × X, µ ⊗ µ), and we see that every Schur multiplier on S2(L2(X, µ)) comes from a function ψ ∈ L∞(X × X, µ⊗ µ) acting by multiplication on L2(X × X, µ ⊗ µ). Definition 2.1. Let p ∈ [1,∞], and let ψ ∈ L∞(X×X, µ⊗µ). The Schur multiplier with symbol ψ is said to be bounded (resp. completely bounded) on Sp(L2(X, µ)) if it maps Sp(L2(X, µ)) ∩ S2(L2(X, µ)) into Sp(L2(X, µ)) (by Tk 7→ Tψk), and if this map extends (necessarily uniquely) to a bounded (resp. completely bounded) map Mψ on Sp(L2(X, µ)). The norm of such a bounded multiplier ψ is defined by kψkM Sp(L2(X,µ)) = kMψk, and its completely bounded norm by kψkcbM Sp(L2(X,µ)) = kMψkcb. The spaces of multipliers and completely bounded multipliers are denoted by M Sp(L2(X, µ)) and cbM Sp(L2(X, µ)), respectively. It follows that for every p ∈ [1,∞] and ψ ∈ L∞(X × X, µ ⊗ µ), we have kψk∞ ≤ kψkM Sp(L2(X,µ)) ≤ kψkcbM Sp(L2(X,µ)). q = 1, we have kψkM Sp(L2(X,µ)) = kψkM Sq(L2(X,µ)). By interpolation and duality we have that whenever 2 ≤ p ≤ q ≤ ∞, then kψkM Sp(L2(X,µ)) ≤ kψkM Sq(L2(X,µ)). These results also hold for the completely bounded norm. p + 1 If 1 Lemma 2.2. ([24, Lemma 1.5 and Remark 1.6]) The Schur multiplier correspond- ing to ψ ∈ L∞(X × X, µ⊗ µ) is completely bounded on Sp(L2(X, µ)) if and only if the Schur multiplier corresponding to ψ(x, ξ, y, η) = ψ(x, y) is completely bounded 4 TIM DE LAAT on Sp(L2(X × Ω, µ ⊗ ν)), where (Ω, ν) is a σ-finite measure space, and kψkcbM Sp(L2(X,µ)) = k ψkcbM Sp(L2(X×Ω,µ⊗ν)). If L2(Ω, ν) is infinite-dimensional, these norms equal k ψkM Sp(L2(X×Ω,µ⊗ν)). Lemma 2.3. ([24, Theorem 1.19]) Let (X, µ) be a locally compact space with a σ-finite Radon measure µ, and let ψ : X × X −→ C be a bounded continuous function. Let 1 ≤ p ≤ ∞. The following are equivalent: (1) we have ψ ∈ M Sp(L2(X, µ)) with kψkM Sp(L2(X,µ)) ≤ C, (2) for every finite set F = {x1, . . . , xn} ⊂ X such that F ⊂ supp(µ), the Schur multiplier given by (ψ(xi, xj ))i,j is bounded on Sp(ℓ2(F )) with norm smaller than or equal to C. In particular, The analogous statement holds in the completely bounded case. the norm and the completely bounded norm of the multiplier only depend on the support of µ, and if this support does not have any isolated points, then the norm and the completely bounded norm coincide. 2.2. Schur multipliers on locally compact groups. For a locally compact group G and a function ϕ ∈ L∞(G), we define the function ϕ ∈ L∞(G × G) by ϕ(g, h) = ϕ(g−1h). The notation ϕ will be used without further mentioning. In what follows, we will consider continuous functions ϕ : G −→ C such that ϕ is a (completely bounded) Schur multiplier on Sp(L2(G)). 2.3. KAK decomposition for Lie groups. Recall that every connected semi- simple Lie group G with finite center can be decomposed as G = KAK, where K is a maximal compact subgroup (unique up to conjugation) and A is an abelian Lie group such that its Lie algebra a is a Cartan subspace of the Lie algebra g of G. The dimension of a is called the real rank of G and is denoted by RankR(G). The KAK decomposition is in general not unique. However, after choosing a set of positive roots and restricting to the closure A+ of the positive Weyl chamber A+, we still have G = KA+K. Moreover, if g = k1ak2, where k1, k2 ∈ K and a ∈ A+, then a is unique. For more details, see [18], [21]. 2.4. Gelfand pairs and spherical functions. Let G be a Lie group with compact subgroup K. We denote the (left) Haar measure on G by dx and the normalized Haar measure on K by dk. A function ϕ : G −→ C is said to be K-bi-invariant if ϕ(k1gk2) = ϕ(g) for all g ∈ G and k1, k2 ∈ K. Note that for ϕ ∈ C(G), the continuous function defined by ϕK(g) =RKRK ϕ(kgk′)dkdk′ is K-bi-invariant. By abuse of notation, we denote the space of K-bi-invariant compactly supported continuous functions on G by Cc(K\G/K). This space can be considered as a subalgebra of the convolution algebra Cc(G). If this subalgebra is commutative, then the pair (G, K) is said to be a Gelfand pair. Equivalently, if G is a Lie group with compact subgroup K, then (G, K) is a Gelfand pair if and only if for every irreducible unitary representation π of G on a Hilbert space Hπ, the space Hπe consisting of K-invariant vectors, i.e., Hπe = {ξ ∈ H ∀k ∈ K : π(k)ξ = ξ}, is at most one-dimensional. Also, the pair (G, K) is a Gelfand pair if and only if the representation L2(G/K) is multiplicity free. Let (G, K) be a Gelfand pair. A function h ∈ C(K\G/K) is called a spherical function if the functional χ on Cc(K\G/K) given by χ(ϕ) = RG ϕ(x)h(x−1)dx defines a nontrivial character, i.e., χ(ϕ ∗ ψ) = χ(ϕ)χ(ψ) for all ϕ, ψ ∈ Cc(K\G/ APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 5 K). Spherical functions arise as the matrix coefficients of K-invariant vectors in irreducible representations of G. It is possible to consider Gelfand pairs in more general settings than Lie groups, e.g., in the setting of locally compact groups (see [7],[11]). 2.5. Schur multipliers on compact Gelfand pairs. Let G and K be Lie groups such that (G, K) is a Gelfand pair, and let X = G/K denote the homogeneous space (with quotient topology) corresponding with the canonical (transitive) action of G. It follows that K is the stabilizer subgroup of a certain element e0 ∈ X. In this section we consider Schur multipliers on the Schatten classes Sp(H), where H = L2(G) or L2(X). To this end, it is natural to look at multipliers on G that are K-bi-invariant. Denote by D the space K\G/K as a topological space, and denote by f : K\G/K −→ D, KgK 7→ ξ the corresponding homeomorphism. It follows that every function ϕ in C(K\G/K) induces a continuous function ϕ0 on D such that ϕ(g) = ϕ0(ξ) for all g ∈ G, where ξ is the image under the homeomorphism f . A Gelfand pair (G, K) is called compact if G is a compact group. In this section, all Gelfand pairs are assumed to be compact, unless explicitly stated otherwise. For compact groups every representation on a Hilbert space is equivalent to a unitary representation, every irreducible representation is finite-dimensional, and every unitary representation is the direct sum of irreducible ones. For an irreducible unitary representation π of G on a Hilbert space Hπ, let Pπ =RK π(k)dk denote the projection onto Hπe (see Section 2.4), and let GK denote the space of equivalence classes of the irreducible unitary representations π of G such that Pπ 6= 0. Lemma 2.4. Let (G, K) be a compact Gelfand pair, and let X = G/K be the corresponding (compact) homogeneous space. Then L2(X) = ⊕π∈ GKHπ. Let hπ denote the spherical function corresponding to the equivalence class π of representations. Then for every ϕ ∈ L2(K\G/K) we have This lemma follows from the Peter-Weyl theorem applied to a compact homo- geneous space (see, e.g., [19, Section V.4]). The decomposition of ϕ (and hence ϕ0) is stated explicitly in [32, Proposition 9.10.4]. Lemma 2.5. Let (G, K) be a (not necessarily compact) Gelfand pair, and let X = G/K denote the corresponding homogeneous space. Choose e0 ∈ X so that K is its stabilizer subgroup. Let ϕ ∈ C(K\G/K). Then there exists a continuous function ψ : X × X −→ C such that for all g, h ∈ G, ϕ(g−1h) = ψ(ge0, he0). Proof. If ge0 = g′e0 for g, g′ ∈ G, then g−1g′ ∈ K, and hence g′ = gk for some k ∈ K. Hence, by the K-bi-invariance of ϕ, we know that ϕ(g−1h) depends only on the pair (ge0, he0) ∈ X × X, so there exists a function ψ : X × X −→ C such that ϕ(g−1h) = ψ(ge0, he0). Since X = G/K is equipped with the quotient topology, this function is continuous. (cid:3) ϕ = Xπ∈ GK where cπ = hϕ, hπi. Moreover, denoting by h0 responding to hπ, we have ϕ0 =Pπ∈ GK cπ dimHπhπ, cπ(dimHπ)h0 π. π the (spherical) function on D cor- 6 TIM DE LAAT Lemma 2.6. Let (G, K) be a compact Gelfand pair. If ϕ : G −→ C is a continuous K-bi-invariant function such that ϕ ∈ cbM Sp(L2(G)) (see Section 2.2) for some p ∈ [1,∞], then kψkcbM Sp(L2(X)) = k ϕkcbM Sp(L2(G)), where ψ : X × X −→ C is as defined in Lemma 2.5. If K is an infinite group, then these norms are equal to k ϕkM Sp(L2(G)). Proof. By [25, Lemma 1.1], the quotient map G −→ G/K has a Borel cross section. Let Y denote the image of this cross section. The result now follows directly from Lemma 2.2 by putting Ω = K, so that G = Y × K as a measure space by the map (y, k) 7→ yk for y ∈ Y and k ∈ K. (cid:3) We can now prove a decomposition result for Schur multipliers on Sp(L2(G)) coming from K-bi-invariant functions. Proposition 2.7. Let (G, K) be a compact Gelfand pair, suppose that K has infinitely many elements, and let p ∈ [1,∞). Let ϕ : G −→ C be a continuous K-bi-invariant function such that ϕ ∈ M Sp(L2(G)). Then 1 p   Xπ∈ GK cπp(dimHπ)  ≤ k ϕkM Sp(L2(G)), 1 It follows that kT1kSp(L2(X)) = 1. where cπ and Hπ are as in Lemma 2.4. Proof. As before, let (Tkf )(x) = RG k(x, y)f (y)dy. Then T1 is the projection on C1 ∈ L2(X). It is sufficient to prove that (Pπ∈ GK cπp(dimHπ)) p ≤ kTψkSp(L2(X)), where ψ is as before. Indeed, we have kTψkSp(L2(X)) = kTψkSp(L2 (X)) kT1kSp(L2 (X)) ≤ kψkM Sp(L2(X)), which is smaller than or equal to kψkcbM Sp(L2(X)) = k ϕkM Sp(L2(G)) by Lemma 2.6 under the assumption that K is an infinite group. By Lemma 2.4, we have ϕ = Pπ∈ GK cπ dim Hπhπ. By [19, Theorem V.4.3], it follows that the operator PHπ = dimHπTh′ π is the projection onto Hπ, where h′π : X × X −→ C denotes the function induced by hπ (see Lemma 2.5). Since L2(X) decomposes as a direct sum of Hilbert spaces, we have kTψkp Sp(L2(X)) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xπ∈ GK = Xπ∈ GK p π(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) cπ dimHπTh′ cπp Tr(PHπp) = Xπ∈ GK Sp(L2(X)) cπp dimHπ. (cid:3) Lemma 2.8. Let G be a locally compact group with compact subgroup K. For p ∈ [1,∞], let ϕ ∈ C(G) be such that ϕ ∈ M Sp(L2(G)). Then the continuous function ϕK defined by ϕK(g) = RKRK ϕ(kgk′)dkdk′ induces an element ϕK of M Sp(L2(G)), and k ϕKkM Sp(L2(G)) ≤ k ϕkM Sp(L2(G)). The analogous statement holds in the completely bounded case. Proof. Let νn be a sequence of finitely supported probability measures on K point- wise converging to the Haar measure µ. Let ϕn : G −→ C be defined by ϕn(g) = RKRK ϕ(kgk′)dνn(k)dνn(k′). Each ϕn is a convex combination of functions kϕk′ of the form kϕk′ (g) = ϕ(kgk′), where k, k′ ∈ K are fixed. Hence, ϕK is an element APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 7 of the pointwise closure of conv{kϕk′ k, k′ ∈ K}. One easily checks that for all k, k′ ∈ K, we have kk ϕk′kM Sp(L2(G)) = k ϕkM Sp(L2(G)). Hence, by Lemma 2.3, we have ϕK ∈ M Sp(L2(G)), and k ϕKkM Sp(L2(G)) ≤ k ϕkM Sp(L2(G)). The result for the completely bounded case follows in an analogous way. (cid:3) 2.6. The property APSchur p,cb . In this section we recall the definition of the APSchur p,cb , as given by Lafforgue and de la Salle in [24]. First, recall that the Fourier algebra A(G) (see [10]) consists of the coefficients of the left-regular representation of G. More precisely, ϕ ∈ A(G) if and only if there exist ξ, η ∈ L2(G) such that for all x ∈ G we have ϕ(x) = hλ(x)ξ, ηi. With the norm kϕkA(G) = min{kξkkηk ∀x ∈ G ϕ(x) = hλ(x)ξ, ηi}, it is a Banach space. Definition 2.9. ([24, Definition 2.2]) Let G be a locally compact Hausdorff second countable group, and let 1 ≤ p ≤ ∞. The group G is said to have the property of completely bounded approximation by Schur multipliers on Sp, denoted APSchur p,cb , if there exists a constant C > 0 and a net ϕα ∈ A(G) such that ϕα → 1 uniformly on compacta and supα k ϕαkcbM Sp(L2(G)) ≤ C. The infimum of these C's is denoted by ΛSchur p,cb (G). The following result is a key property of the APSchur p,cb (see [24, Theorem 2.5]). Theorem 2.10. Let G be a locally compact Hausdorff group, and let Γ be a lattice in G. Then for 1 ≤ p ≤ ∞, we have ΛSchur p,cb (Γ) = ΛSchur p,cb (G). Lafforgue and de la Salle also proved that for a discrete group Γ and p ∈ (1,∞), it follows that ΛSchur p,cb (Γ) ∈ {1,∞}. Since a semisimple Lie group G has lattices [1], we conclude by the above proposition that for such a group, it also follows that ΛSchur p,cb (G) ∈ {1,∞} for p ∈ (1,∞). Proposition 2.11. Let G be a locally compact Hausdorff group. The APSchur p,cb satisfies the following properties: (1) for p = ∞ (or p = 1, by the third statement of this proposition), the group G p,cb (G) = Λ(G), p,cb if and only if it is weakly amenable, and ΛSchur has the APSchur where Λ(G) denotes the Cowling-Haagerup constant of G; p,cb (G) = ΛSchur q,cb (G); 2,cb (G) = 1; q = 1, then ΛSchur q,cb (G); (2) for every locally compact group, ΛSchur (3) if p, q ∈ [1,∞] such that 1 p + 1 (4) if 2 ≤ p ≤ q ≤ ∞, then ΛSchur p,cb (G) ≤ ΛSchur (5) if H is a closed subgroup of G and 1 ≤ p ≤ ∞, then ΛSchur p,cb (G); (6) if G has a compact subgroup K, and if ϕα is a net in A(G) converging to 1 uniformly on compacta such that supα k ϕαkcbM Sp(L2(G)) ≤ C, then there exists a net ϕα in A(G)∩C(K\G/K) such that supα k ϕαkcbM Sp(L2(G)) ≤ C that converges to 1 uniformly on compacta. p,cb (G) = (7) if K is a compact normal subgroup of G and 1 ≤ p ≤ ∞, then ΛSchur p,cb (H) ≤ ΛSchur ΛSchur p,cb (G/K); (8) if G1 and G2 are locally isomorphic connected (semi)simple Lie groups with finite centers, then for p ∈ [1,∞], we have ΛSchur p,cb (G1) = ΛSchur p,cb (G2); Proof. The first statement is clear. The second through the fifth statement are covered in [24, Section 2]. The sixth statement follows from Lemma 2.8. By com- bining the sixth statement and Lemma 2.6, the seventh statement follows. The fact that the net on the group converges uniformly on compacta if and only if the net 8 TIM DE LAAT on the quotient does, is straightforward (see [6]). For the eighth statement, note that the center is a normal subgroup of a group. Using the seventh statement and the fact that the adjoint groups G1/Z(G1) and G2/Z(G2), where Z(Gi) denotes the center of Gi, are isomorphic, we obtain the result. (cid:3) 2.7. Approximation properties for noncommutative Lp-spaces. The oper- ator space structure on a noncommutative Lp-space Lp(M, τ ) can be obtained by considering this space as a certain interpolation space (see [23]). Indeed, the pair of spaces (M, L1(M, τ )) becomes a compatible couple of operator spaces, and for 1 < p < ∞ we have the isometry Lp(M, τ ) ∼= [M, L1(M, τ )] 1 . By [28, Lemma 1.7], we know that for a linear map T : Lp(M, τ ) −→ Lp(M, τ ), its completely bounded n[Lp(M )]k. Us- norm kTkcb corresponds to supn∈N k idSp ing [28, Corollary 1.4] and the fact that S1 n ⊗ L1(M ) = L1(M ⊗ Mn), we obtain that Sp n[Lp(M )] = Lp(M ⊗ Mn), which implies that kTkcb = supn∈N kT ⊗ id : Lp(M ⊗ Mn) −→ Lp(M ⊗ Mn)k. In Section 1 of this article, we recalled the definition of the CBAP, CCAP and OAP. It was shown by Junge and Ruan [20] that if Γ is a discrete group with the AP (of Haagerup and Kraus), and if p ∈ (1,∞), then Lp(L(Γ)) has the OAP, where L(Γ) denotes the group von Neumann algebra of Γ. Lafforgue and de la Salle related the AP for groups and the OAP for noncommutative Lp-spaces to the APSchur p,cb . n[Lp(M )] −→ Sp p n ⊗ T : Sp Lemma 2.12. ([24, Corollary 3.12]) If Γ is a countable discrete group with the AP, and if p ∈ (1,∞), then ΛSchur Lemma 2.13. ([24, Corollary 3.13]) If p ∈ (1,∞) and Γ is a countable discrete group such that Lp(L(Γ)) has the OAP, then ΛSchur p,cb (Γ) = 1. p,cb (Γ) = 1. One of the main results of Lafforgue and de la Salle is the following. Theorem 2.14. ([24, Theorem E]) Let n ≥ 3. For p ∈ [1, 4 group SL(n, R) does not have the APSchur p,cb . 3 ) ∪ (4,∞], the (exact) As a consequence, the group SL(n, R) does not have the AP, and for p ∈ [1, 4 3 )∪ (4,∞] and a lattice Γ in SL(n, R), the noncommutative Lp-space Lp(L(Γ)) does not have the OAP or CBAP. 3. The group Sp(2, R) In this section, we prove the following theorem. The proof is along the same lines as the proof of the failure of the AP for Sp(2, R) in [15] (and for some details we will refer to that article), but obtaining sufficiently sharp estimates for Schur multipliers on Schatten classes is technically more involved. Theorem 3.1. For p ∈ [1, 12 APSchur p,cb . 11 ) ∪ (12,∞], the group Sp(2, R) does not have the In this section, we write G = Sp(2, R). Recall that G is defined as the Lie group where G := {g ∈ GL(4, R) gtJg = J}, J =(cid:18) 0 −I2 I2 0 (cid:19) . APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 9 Here I2 denotes the 2 × 2 identity matrix. The maximal compact subgroup K of G is isomorphic to U(2) and explicitly given by K =(cid:26)(cid:18) A −B A + iB ∈ U(2)(cid:27). Let A+ = {D(α1, α2) = diag(eα1, eα2 , e−α1, e−α2) α1 ≥ α2 ≥ 0}. G = KA+K. B A (cid:19) ∈ M4(R)(cid:12)(cid:12)(cid:12)(cid:12) It follows that p,cb For p = 1 and ∞, the APSchur is equivalent to weak amenability (as mentioned in Proposition 2.11), and the failure of weak amenability for G was proved in [13]. Therefore, we can restrict ourselves to the case p ∈ (1,∞). As follows from Pro- position 2.11, it suffices to consider approximating nets consisting of K-bi-invariant functions. The following result gives a certain asymptotic behaviour of continuous K-bi-invariant functions ϕ for which the induced function ϕ is a Schur multiplier on Sp(L2(G)). From this, it follows that the constant function 1 cannot be approx- imated pointwise (and hence not uniformly on compacta) by a K-bi-invariant net in A(G) in such a way that the net of associated multipliers is uniformly bounded in the M Sp(L2(G))-norm. This implies Theorem 3.1. Proposition 3.2. Let p > 12. There exist constants C1(p), C2(p) (depending on p only) such that for all ϕ ∈ C(K\G/K) for which ϕ ∈ M Sp(L2(G)), the limit ϕ∞ = limkαk→∞ ϕ(D(α1, α2)) exists, and for all α1 ≥ α2 ≥ 0, ϕ(D(α1, α2)) − ϕ∞ ≤ C1(p)k ϕkM Sp(L2(G))e−C2(p)kak2 , 1 + α2 2. where kαk2 =pα2 Remark 3.3. Note that Proposition 3.2 is stated in terms of the M Sp(L2(G))- norm rather than the cbM Sp(L2(G))-norm. However, we have k.kM Sp(L2(G)) ≤ k.kcbM Sp(L2(G)), which shows that Proposition 3.2 is indeed sufficient to prove The- orem 3.1. Moreover, by [24, Theorem 1.18], the claims are equivalent for non- discrete groups. For the proof of Proposition 3.2, we will identify two Gelfand pairs in G and describe certain properties of their spherical functions. Consider the group U(2), which contains the circle group U(1) as a subgroup via the embedding U(1) ֒→(cid:18) 1 0 U(1) (cid:19) ⊂ U(2). 0 Let K1 denote the copy of U(1) in G under the identification of U(2) with K. It goes back to Weyl [31] that (U(2), U(1)) is a Gelfand pair (see, e.g., [21, Theorem IX.9.14]). The homogeneous space U(2)/ U(1) is homeomorphic to the complex 1-sphere S1 C ⊂ C2 and the double coset space U(1)\ U(2)/ U(1) is homeomorphic to the closed unit disc D ⊂ C by the map U(1)(cid:18) u11 u12 u21 u22 (cid:19) U(1) 7→ u11. The spherical functions for (U(2), U(1)) can be found in [22]. By the homeomorph- ism U(1)\ U(2)/ U(1) ∼= D, they can be considered as functions of one complex variable in the closed unit disc. They are indexed by the integers l, m ≥ 0 and explicitly given by hl,m(cid:18) u11 u12 u21 u22 (cid:19) = h0 l,m(u11), 10 TIM DE LAAT where in the point z ∈ D, the function h0 l,m(z) =( zl−mP (0,l−m) zm−lP (0,m−l) h0 m l l,m is explicitly given by (2z2 − 1) (2z2 − 1) l ≥ m, l < m. n Here P (α,β) denotes the nth Jacobi polynomial. These spherical functions satisfy a certain Holder continuity condition, as is stated in the following lemma (see [15, Corollary 3.5]). The proof of this Lemma makes use of recent results by Haagerup and Schlichtkrull [16]. Lemma 3.4. For all l, m ≥ 0, and for θ1, θ2 ∈ [0, 2π), we have Here C > 0 is a uniform constant. Combining the two, we get ≤ C(l + m + 1) 3 4 θ1 − θ2, ≤ 2C(l + m + 1)− 1 4 . ≤ 2 3 4 Cθ1 − θ2 1 4 . h0 h0 (cid:12)(cid:12)(cid:12)(cid:12) l,m(cid:18) eiθ1 √2(cid:19) − h0 l,m(cid:18) eiθ1 (cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0 l,m(cid:18) eiθ1 (cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0 ϕ(u) = ϕ(cid:18) u11 u12 l,m(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) l,m(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) l,m(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) u21 u22 (cid:19) = ϕ0(u11), h0 Let ϕ : U(2) −→ C be a U(1)-bi-invariant continuous function. Then u ∈ U(2), u11 ∈ D, for some continuous function ϕ0 : D −→ C. By Lemma 2.4, we know that L2(X) = ⊕l,m≥0Hl,m, where X = U(2)/ U(1) ∼= S C 1 . It is known that dimHl,m = l + m + 1, so, by Proposition 2.7, we get ϕ0 = cl,m(l + m + 1)h0 l,m, ∞ Xl,m=0 as above by ϕ(g, h) = ϕ(g−1h). for certain cl,m ∈ C. Moreover, by the same proposition, we obtain that if p ∈ (1,∞), then (Pl,m≥0 cl,mp(l + m + 1)) p ≤ k ϕkM Sp(L2(U(2))), where ϕ is defined Lemma 3.5. Let p > 12, and let ϕ : U(2) −→ C be a continuous U(1)-bi-invariant function such that ϕ is an element of M Sp(L2(U(2))). Then ϕ0 satisfies 1 1 2p 8− 3 p + 1 ≤ C(p)k ϕkM Sp(L2(U(2)))θ1 − θ2 √2(cid:19) − ϕ0(cid:18) eiθ2 ϕ0(cid:18) eiθ1 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) for θ1, θ2 ∈ [0, 2π). Here, C(p) is a constant depending only on p. Proof. Let p, q ∈ (1,∞) be such that 1 √2(cid:19) − ϕ0(cid:18) eiθ2 ϕ0(cid:18) eiθ1 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) cl,m(l + m + 1)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) = Xl,m≥0 q  ≤ cl,mq(l + m + 1) (l + m + 1)(cid:12)(cid:12)(cid:12)(cid:12)  Xl,m≥0  Xl,m≥0  ≤ k ϕkM Sq(L2(U(2))) l,m(cid:18) eiθ1 (l + m + 1)(cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0 q = 1. Then for θ1, θ2 ∈ [0, 2π), l,m(cid:18) eiθ2 l,m(cid:18) eiθ1 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0 p l,m(cid:18) eiθ2 l,m(cid:18) eiθ1 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0  p l,m(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)   Xl,m≥0 h0 h0 h0 1 p . 1 1 p APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 11 Note that k ϕkM Sq(L2(U(2))) = k ϕkM Sp(L2(U(2))). If we look at the terms of the last sum, we get, using Lemma 3.4 and the fact that min{x, y} ≤ xεy1−ε for x, y > 0 and ε ∈ (0, 1), that h0 l,m(cid:18) eiθ1 (l + m + 1)(cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − h0 ≤ min{C p(l + m + 1)1+ 3 ≤ 2p(1−ε)C pθ1 − θ2pε(l + m + 1)1+pε− 1 for ε ∈ (0, 1). Hence, the sum converges for 0 < ε < 1 for p > 12. Hence, if p > 12, and putting ε = 1 4 − 3 2 ( 1 l,m(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 4 pθ1 − θ2p, 2pC p(l + m + n)1− 1 4 p} p 4 p 4 − 3 p ) = 1 p . Such an ε only exists 8 − 3 2p , then ϕ0(cid:18) eiθ1 (cid:12)(cid:12)(cid:12)(cid:12) √2(cid:19) − ϕ0(cid:18) eiθ2 √2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) for some constant C(p) depending only on p. ≤ C(p)k ϕkM Sp(L2(U(2))θ1 − θ2 1 8− 3 2p (cid:3) For α ∈ R consider the map K −→ G defined by k 7→ DαkDα, where Dα = diag(eα, 1, e−α, 1). Lemma 3.6. Let ϕ : G −→ C be a continuous K-bi-invariant function such that ϕ ∈ M Sp(L2(G)) for some p ∈ (1,∞), and for α ∈ R, let ψα : K −→ C be defined by ψα(k) = ϕ(DαkDα). Then ψα is K1-bi-invariant and satisfies k ψαkM Sp(L2(U(2)) ≤ k ϕkM Sp(L2(G)). Proof. Using the fact that the group elements Dα commute with K1, it follows that for all k ∈ K and k1, k2 ∈ K1 ⊂ K2, ψα(k1kk2) = ϕ(Dαk1kk2Dα) = ϕ(k1DαkDαk2) = ϕ(DαkDα) = ψα(k), so ψα is K1-bi-invariant. The second part follows by the fact that DαKDα is a subset of G and by applying (cid:3) Lemma 2.3. From the fact that ψα is K1-bi-invariant, it follows that ψα(u) = ψ0 ψ0 α : D −→ C is a continuous function. Suppose that α1 ≥ α2 ≥ 0, and let D(α1, α2) be as defined above. If we find an element of the form DαkDα in KD(α1, α2)K, we can relate the value of a K-bi- invariant multiplier ϕ to the value of the multiplier ψα that was just defined. This only works for certain α1, α2 ≥ 0. It turns out to be sufficient to consider certain candidates for k, namely the ones of the form α(u11), where (1) u =(cid:18) a + ib √1 − a2 − b2 −√1 − a2 − b2 a − ib (cid:19) with a2 + b2 ≤ 1. For a proof of the following result, see [15, Lemma 3.9]. Lemma 3.7. Let α ≥ 0 and β ≥ γ ≥ 0. If u ∈ K is of the form (1) with respect to the identification of K with U(2), then DαuDα ∈ KD(β, γ)K if and only if (2) ( sinh β sinh γ = sinh2 α(1 − a2 − b2), sinh β − sinh γ = sinh(2α)a. 12 TIM DE LAAT Consider the second Gelfand pair sitting inside G, namely the pair of groups (SU(2), SO(2)). Both groups are subgroups of U(2), so under the embedding into G, they give rise to compact Lie subgroups of G. The subgroup corres- ponding to SU(2) will be called K2, and the one corresponding to SO(2) will be called K3. The group K3 commutes with the group generated by the elements D′α = diag(eα, eα, e−α, e−α), where α ∈ R. The subgroup SU(2) ⊂ U(2) consists of matrices of the form c + id u =(cid:18) a + ib −c + id with a, b, c, d ∈ R such that a2 + b2 + c2 + d2 = 1. By [4, Theorem 47.6], the pair (SU(2), SO(2)) is a Gelfand pair. This also follows from [12, Chapter 9]. The homogeneous space SU(2)/ SO(2) is the sphere S2, and the spherical functions on the double coset space [−1, 1] are indexed by n ≥ 0, and given by the Legendre polynomials a − ib (cid:19) , Pn(2(a2 + c2) − 1) = Pn(a2 − b2 + c2 − d2). Note that the double cosets of SO(2) in SU(2) are labeled by a2 − b2 + c2 − d2. We use the following estimate (see [15, Lemma 3.11]). Lemma 3.8. For all non-negative integers n, and x, y ∈ [− 1 4 √n , Pn(x) − Pn(y) ≤ Pn(x) + Pn(y) ≤ 2 , 1 2 ], Combining the two, we get Pn(x) − Pn(y) ≤(cid:12)(cid:12)(cid:12)(cid:12) Z y x P ′n(t)dt(cid:12)(cid:12)(cid:12)(cid:12) ≤ 4√nx − y. Pn(x) − Pn(y) ≤ 4x − y 1 2 for x, y ∈ [− 1 on [− 1 2 , 1 2 , 1 2 ] with exponent 1 2 . 2 ], i.e., the Legendre polynomials are uniformly Holder continuous Let ϕ : SU(2) −→ C be a SO(2)-bi-invariant continuous function. Then c + id a − ib (cid:19) = ϕ0(2(a2 + c2) − 1) = ϕ0(a2 − b2 + c2 − d2), ϕ(u) = ϕ(cid:18) a + ib −c + id where u ∈ U(2), u11 ∈ D, and where ϕ0 : D −→ C is some continuous function. By Lemma 2.4, we know that L2(X) = ⊕n≥0Hn, where X = SU(2)/ SO(2) ∼= S2. It is known that dimHn = 2n + 1, so, by Proposition 2.7, we get ϕ0 = cn(2n + 1)Pn, ∞ Xn=0 1 ϕ(g, h) = ϕ(g−1h). for certain cn ∈ C. Moreover, by the same proposition, we obtain that if p ∈ (1,∞), then (Pn≥0 cnp(2n + 1)) p ≤ k ϕkM Sp(L2(SU(2))), where ϕ is defined as above by Lemma 3.9. Let p > 4, and let ϕ : SU(2) −→ C be a continuous SO(2)-bi-invariant function such that ϕ ∈ M Sp(L2(SU(2))). Then ϕ0 satisfies ϕ0(δ1) − ϕ0(δ2) ≤ C(p)kϕkM Sp(L2(SU(2))δ1 − δ2 2 ]. Here C(p) is a constant depending only on p. 2 , 1 1 4− 1 p for δ1, δ2 ∈ [− 1 APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 13 Proof. Let p, q ∈ (1,∞) be such that 1 p + 1 q = 1, and let δ1, δ2 ∈ [− 1 2 , 1 2 ]. Then 1 ϕ0(δ1) − ϕ0(δ2) = Xn≥0 ≤ cnq(2n + 1) Xn≥0  ≤ k ϕkM Sq(L2(SU(2))) Xn≥0 (2n + 1)Pn(δ1) − Pn(δ2)p  cn(2n + 1)Pn(δ1) − Pn(δ2) q  Xn≥0 (2n + 1)Pn(δ1) − Pn(δ2)p  1 p . 1 p Note that k ϕkM Sq(L2(SU(2))) = k ϕkM Sp(L2(SU(2))). If we look at the terms of the last sum, we get, using Lemma 3.8 and the fact that min{x, y} ≤ xεy1−ε for x, y > 0 and ε ∈ (0, 1), that (2n + 1)Pn(δ1) − Pn(δ2)p ≤ min{4p(2n + 1)n− p 2 , 4p(2n + 1)n p 2 δ1 − δ2p} ≤ 4p(3n)1+pε− p 2 δ1 − δ2pε 2 − 2 p ) = 1 for ε ∈ (0, 1). Hence, the sum converges for ε ∈ (0, 1 for p > 4. Hence, if p > 4, and putting ε = 1 p ). Such an ε only exists 4 − 1 ϕ0(δ1) − ϕ0(δ2) ≤ C(p)k ϕkM SpL2(U(2))δ1 − δ2 p , we have 2 − 2 4 − 1 p , 2 ( 1 1 where C(p) is a constant depending only on p. (cid:3) For α ∈ R consider the map K −→ G defined by k 7→ D′αkvD′α, where D′α = diag(eα, eα, e−α, e−α) and v ∈ Z(K) is chosen to be the matrix in K that in the U(2)-representation of K is given by (3) v = 1√2 (1 + i) 0 0 (1 + i) ! . 1√2 Given a K-bi-invariant multiplier on G, this map gives rise to a K3-bi-invariant multiplier on K. We state the following result, but omit its proof, as it is similar to the one of Lemma 3.6. Lemma 3.10. Let ϕ : G −→ C be a continuous K-bi-invariant function such that ϕ ∈ M Sp(L2(G)) for some p ∈ (1,∞), and for α ∈ R let χα : K −→ C be defined by χα(k) = ϕ(D′αkvD′α). Then χα is K3-bi-invariant and satisfies k χαkM Sp(L2(K)) ≤ k ϕkM Sp(L2(G)). Consider the restriction χα = χαK2, which is a K3-bi-invariant multiplier on α(a2− b2 + c2− d2), where u ∈ K2, and where a, b, c, d K2. It follows that χα(u) = χ0 are as before, and k χαkM Sp(L2(K2)) ≤ k ϕkM Sp(L2(G)). Suppose that α1 ≥ α2 ≥ 0 and let D(α1, α2) be as defined above. Again, if we find an element of the form D′αuvD′α in KD(α1, α2)K, where now u has to be an element of SU(2), we can relate the value of a K-bi-invariant multiplier ϕ to the value of the multiplier χα. This again only works for certain α1, α2 ≥ 0. Consider a general element of SU(2), u =(cid:18) a + ib −c + id a − ib (cid:19) , c + id with a2 + b2 + c2 + d2 = 1. For a proof of the following, see [15, Lemma 3.15]. 14 TIM DE LAAT Lemma 3.11. Let α ≥ 0 and β ≥ γ ≥ 0, and let u, v ∈ K be of the form as in (1) and (3) with respect to the identification of K with U(2). Then D′αuvD′α ∈ KD(β, γ)K if and only if ( sinh2 β + sinh2 γ = sinh2(2α), 2 sinh2(2α)r, sinh β sinh γ = 1 where r = a2 − b2 + c2 − d2. Now we can combine the results that we obtained for both Gelfand pairs. Lemma 3.12. Let β ≥ γ ≥ 0. Then the equations (4) sinh2(2s) + sinh2 s = sinh2 β + sinh2 γ, sinh(2t) sinh t = sinh β sinh γ have unique solutions s = s(β, γ), t = t(β, γ) in the interval [0,∞). Moreover, (5) , β 4 s ≥ γ 2 . t ≥ A proof of this Lemma can be found in [15, Lemma 3.16]. α2 α1 = α2 (2s, s) α1 = 2α2 (2t, t) (β, γ) α1 The figure above shows the relative position of (β, γ), (2s, s) and (2t, t) as in Lemma 3.13 and Lemma 3.14 below. Note that (β, γ) and (2s, s) lie on a path in the (α1, α2)-plane of the form sinh2 α1 + sinh2 α2 = constant, and (β, γ) and (2t, t) lie on a path of the form sinh α1 sinh α2 = constant. Lemma 3.13. For p > 4, there exists a constant C3(p) > 0 (depending only on p) such that whenever β ≥ γ ≥ 0 and s = s(β, γ) is chosen as in Lemma 3.12, then for all ϕ ∈ C(K\G/K) for which ϕ ∈ M Sp(L2(G)), ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ C3(p)e− β−γ 4 ( 1 p )k ϕkM Sp(L2(G)). 4− 1 Proof. Assume first that β − γ ≥ 8. Let α ∈ [0,∞) be the unique solution to sinh2 β + sinh2 γ = sinh2(2α), and observe that 2α ≥ β ≥ 2, so in particular α > 0. Define r1 = 2 sinh β sinh γ sinh2 β + sinh2 γ ∈ [0, 1], APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 15 and a1 =(cid:0) 1+r1 2 (cid:1) and let 2 . Furthermore, put 1 1 2 and b1 =(cid:0) 1−r1 2 (cid:1) u1 =(cid:18) a1 + ib1 v = 1√2 0 0 (1 + i) 0 a1 − ib1 (cid:19) ∈ SU(2), (1 + i) ! , 1√2 0 as previously defined. We now have 2 sinh β sinh γ = sinh2(2α)r1, and a2 1 = r1, so by Lemma 3.11, we have D′αu1vD′α ∈ KD(β, γ)K. Let s = s(β, γ) be as in Lemma 3.12. Then s ≥ 0 and sinh2(2s) + sinh2 s = sinh2 β + sinh2 γ = sinh2(2α). Put 1 − b2 0 0 r2 = 2 sinh(2s) sinh s sinh2(2s) + sinh2 s ∈ [0, 1], a2 − ib2 (cid:19) ∈ SU(2), u2 =(cid:18) a2 + ib2 2 and b2 = (cid:0) 1−r2 2 (cid:1) α(r2) ≤ C(p)r1 − r2 α(r1) − χ0 2 − b2 1 1 1 where a2 = (cid:0) 1+r2 2 (cid:1) 2 = r2, it follows again by Lemma 3.11 that D′αu2vD′α ∈ KD(2s, s)K. Now, let χα(u) = ϕ(D′αuvD′α) for u ∈ K2 ∼= SU(2). Then by Lemma 3.9 and Lemma 3.10, it follows that 2 . Since a2 and χα(u1) − χα(u2) = χ0 provided that r1, r2 ≤ 1 of ϕ, we get 2 . Hence, under this assumption, using the K-bi-invariance 4 − 1 p k ϕkM Sp(L2(G)), 1 4− 1 sinh2 β ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ C(p)r1 − r2 (6) Note that r1 ≤ 2 sinh β sinh γ r1 ≤ 2 eγ (1−e−2γ ) cosh s ≤ 2e−s. By Lemma 3.12, equation (5), we obtain that r2 ≤ 2e− β 2e−2 ≤ 1 have proved that sinh β . Hence, using β ≥ γ + 8 ≥ γ, we get sinh 2s = 4 ≤ , we 4 ≤ 2e 2 . In particular, (6) holds, and since r1 − r2 ≤ max{r1, r2} ≤ 2e eβ (1−e−2β ) ≤ 2eγ−β. In particular, r1 ≤ 2e−8 ≤ 1 2 . Similarly, r2 ≤ 2 sinh s p k ϕkM Sp(L2(G)). = 2 sinh γ γ−β γ−β 1 4 ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ C(p)2 (7) under the assumption that β ≥ γ + 8. If γ ≤ β < γ + 8, we get from kϕk∞ ≤ k ϕkM Sp(L2(G)) that ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ 2k ϕkM Sp(L2(G)). It follows that p )k ϕkM Sp(L2(G)) p e γ−β 4 ( 1 4− 1 1 4− 1 ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ C3(p)e γ−β 4 ( 1 4− 1 p )k ϕkM Sp(L2(G)) for all (β, γ) with β ≥ γ ≥ 0, if for all p ∈ (1,∞), we put C3(p) = max{ C(p)2 1 4− 1 p , 2e (cid:3) 1 2}. Lemma 3.14. For p > 12, there exists a constant C4(p) > 0 (depending only on p) such that whenever β ≥ γ ≥ 0 and t = t(β, γ) is chosen as in Lemma 3.12, then for all ϕ ∈ C(K\G/K) for which ϕ ∈ M Sp(L2(G)), ϕ(D(β, γ)) − ϕ(D(2t, t)) ≤ C4(p)e− γ 4 ( 1 p )k ϕkM Sp(L2(G)). 4− 3 Proof. Let β ≥ γ ≥ 0. Assume first that γ ≥ 2, and let α ≥ 0 be the unique 2 sinh2 α, and observe that α > 0, solution in [0,∞) to the equation sinh β sinh γ = 1 because β ≥ γ ≥ 2. Put sinh β − sinh γ a1 = sinh(2α) ≥ 0. 16 TIM DE LAAT Since sinh(2α) = 2 sinh α cosh α ≥ 2 sinh2 α, we have = In particular, a1 ≤ 1 sinh β sinh γ = sinh2 α(1 − a2 4γ ≤ 1 . 1 4 sinh γ sinh β 2 sinh2 α sinh β a1 ≤ sinh(2α) ≤ 8 . Put now b1 =q 1 1 − b2 a1 − ib1 ! ∈ SU(2). u1 = a1 + ib1 − 1√2 1√2 1 = 1 1) and sinh β − sinh γ = sinh(2α)a1. Let 1. Then 1 − a2 2 − a2 1 − b2 2 . Hence, By Lemma 3.12, we have sinh(2t) sinh t = sinh β sinh γ = 1 By Lemma 3.7, we have Dαu1Dα ∈ KD(β, γ)K. by (5), we have t ≥ γ we get that the number 2 sinh2 α. Moreover, 2 ≥ 1. By replacing (β, γ) in the above calculation with (2t, t), a2 sinh(2t) − sinh t sinh(2α) ≥ 0, Hence, we can put b2 =q 1 satisfies Then 1 4 sinh 1 ≤ 1 4 . 1 4 sinh t ≤ 2 and a2 ≤ 2 − a2 u2 = a2 + ib2 1√2 a2 − ib2 ! . − 1√2 sinh(2t) sinh t = sinh2 α(1 − a2 2 − b2 2), sinh(2t) − sinh t = sinh(2α)a2, and u2 ∈ SU(2). Hence, by Lemma 3.7, Dαu2Dα ∈ KD(2t, t)K. Put now θj = arg(aj + ibj) = π 2 for j = 1, 2, and since d dt sin−1 t = 1√1−t2 ≤ √2 for t ∈ [0, 1√2 ], it follows that 2 − sin−1(cid:16) aj√2(cid:17) for j = 1, 2. Since 0 ≤ aj ≤ 1 sin−1(cid:18) a1√2(cid:19) − sin−1(cid:18) a2√2(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) θ1 − θ2 ≤(cid:12)(cid:12)(cid:12)(cid:12) ≤ a1 − a2 ≤ max{a1, a2} ≤ max(cid:26) 1 ≤ 4 sinh t(cid:27) 4 sinh γ 1 1 , , 4 sinh γ 2 because t ≥ γ θ1 − θ2 ≤ e− γ 3.6, the function ψα(u) = ϕ(DαuDα), u ∈ U(2) ∼= K satisfies 2 . Since γ ≥ 2, we have sinh γ 2 . Note that aj = 1√2 2 . Hence, eiθj for j = 1, 2, so by Lemma 3.5 and Lemma 2 (1 − e−γ) ≥ 1 4 e 2 = 1 2 e γ γ (8) ψα(u1) − ψα(u2) ≤ C(p)θ1 − θ2 4 − 3 4 ( 1 ≤ C(p)e− γ 1 2p k ψαkM Sp(L2(K)) 8− 3 p )k ϕkM Sp(L2(G)). Since Dαu1Dα ∈ KD(β, γ)K and Dαu2Dα ∈ KD(2t, t)K, it follows that ϕ(D(β, γ)) − ϕ(D(2t, t)) ≤ C(p)e− γ 4 ( 1 4− 3 p )k ϕkM Sp(L2(G)) APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 17 for all γ ≥ 2. For γ satisfying 0 < γ ≤ 2, we can instead use that kϕk∞ ≤ k ϕkM Sp(L2(G)). Hence, for all p ∈ (1,∞) putting C4(p) = max{ C(p), 2e 8}, we obtain 1 ϕ(D(β, γ)) − ϕ(D(2t, t)) ≤ C4(p)e− γ 4 ( 1 4− 3 p )k ϕkM Sp(L2(G)) (cid:3) for all β ≥ γ ≥ 0. For a proof of the following lemma, see [15, Lemma 3.19]. Lemma 3.15. Let s ≥ t ≥ 0. Then the equations (9) sinh2 β + sinh2 γ = sinh2(2s) + sinh2 s, sinh β sinh γ = sinh(2t) sinh t, have a unique solution (β, γ) ∈ R2 for which β ≥ γ ≥ 0. Moreover, if 1 ≤ t ≤ s ≤ 3t 2 , then (10) β − 2s ≤ 1, γ + 2s − 3t ≤ 1. Lemma 3.16. For all p > 12, there exists a constant C5(p) > 0 such that whenever s, t ≥ 0 satisfy 2 ≤ t ≤ s ≤ 6 5 t, then for all ϕ ∈ C(K\G/K) for which ϕ ∈ M Sp(L2(G)), ϕ(D(2s, s)) − ϕ(D(2t, t)) ≤ C5(p)e− s 8 ( 1 4 − 3 p )k ϕkM Sp(L2(G)). Proof. Choose β ≥ γ ≥ 0 as in Lemma 3.15. Then by Lemma 3.13 and Lemma 3.14, we have for p > 12, ϕ(D(2s, s)) − ϕ(D(β, γ)) ≤ C3(p)e− β−γ ϕ(D(2t, t)) − ϕ(D(β, γ)) ≤ C4(p)e− γ 4 ( 1 4− 1 p )k ϕkM Sp(L2(G)), p )k ϕkM Sp(L2(G)). 4− 3 4 ( 1 Moreover, by (10), β − γ ≥ (2s − 1) − (3t − 2s + 1) = 4s − 3t − 2 ≥ s − 2, s − 2 2 Hence, since s ≥ 2, we have min{e−γ, e−(β−γ)} ≤ e− s−2 from Lemma 3.13 and Lemma 3.14 with C5(p) = e γ ≥ 3t − 2s − 1 ≥ s − 2s − 1 = 5 2 . 2 . Thus, the lemma follows (cid:3) 16 (C3(p) + C4(p)). 1 Lemma 3.17. For p > 12, there exists a constant C6(p) > 0 such that for all ϕ ∈ C(K\G/K) for which ϕ ∈ M Sp(L2(G)), the limit c∞(ϕ) = limt→∞ ϕ(D(2t, t)) exists, and for all t ≥ 0, ϕ(D(2t, t)) − c∞(ϕ) ≤ C6(p)e− t 8 ( 1 4− 3 p )k ϕkM Sp(L2(G)). ϕ(D(2u, u)) − ϕ(D(2u + 2γ, u + γ)) ≤ C5(p)e− u Proof. By Lemma 3.16, we have for u ≥ 5 and γ ∈ [0, 1], that 4− 3 (11) since u ≤ u + γ. Let s ≥ t ≥ 5. Then s = t + n + δ, where n ≥ 0 is an integer and δ ∈ [0, 1). Applying equation (11) to (u, γ) = (t + j, 1), j = 0, 1, . . . , n − 1 and (u, γ) = (t + n, δ), we obtain p )k ϕkM Sp(L2(G)), 8 ( 1 ϕ(D(2t, t)) − ϕ(D(2s, s)) ≤ C5(p) Xj=0  ≤ C5(p)′e− t n e− t+j 8 ( 1 4− 3 p )  k ϕkM Sp(L2(G)) 8 ( 1 4− 3 p )k ϕkM Sp(L2(G)), 18 TIM DE LAAT Therefore, c∞(ϕ) = limt→∞ ϕ(D(2t, t)) exists, and where C′5(p) = C5(p)P∞j=0 e− j ϕ(D(2t, t))−c∞(ϕ) = lim 8 ( 1 4− 3 p ). Hence, (ϕ(D(2t, t)))t≥5 is a Cauchy net. s→∞ ϕ(D(2t, t))−ϕ(D(2s, s)) ≤ C′5(p)e− t 8 ( 1 4 − 3 p )k ϕkM Sp(L2(G)) for all t ≥ 5. Since kϕk∞ ≤ k ϕkM Sp(L2(G)), we have for all 0 ≤ t < 5, ϕ(D(2t, t)) − c∞(ϕ) ≤ 2k ϕkM Sp(L2(G)). 5 Hence, the lemma follows with C6(p) = max{C′5(p), 2e Proof of Proposition 3.2. Let ϕ ∈ C(K\G/K) be such that ϕ ∈ M Sp(L2(G)), and let (α1, α2) = (β, γ), where β ≥ γ ≥ 0. Assume first β ≥ 2γ. Then β − γ ≥ β 2 , so by Lemma 3.12 and Lemma 3.13, there exists an s ≥ β 4− 1 ϕ(D(β, γ)) − ϕ(D(2s, s)) ≤ C3(p)e− β p )k ϕkM Sp(L2(G)). 4 such that 32}. 8 ( 1 (cid:3) By Lemma 3.17, ϕ(D(2s, s)) − c∞(ϕ) ≤ C6(p)e− s ≤ C6(p)e− β 8 ( 1 4 − 3 32 ( 1 p )k ϕkM Sp(L2(G)) 4− 3 p )k ϕkM Sp(L2(G)). Hence, ϕ(D(β, γ)) − c∞(ϕ) ≤ (C3(p) + C6(p))e− β 32 ( 1 4− 3 p )k ϕkM Sp(L2(G)). Assume now that β < 2γ. Then, by Lemma 3.12 and Lemma 3.14, we obtain that there exists a t ≥ γ 4 such that 2 > β ϕ(D(β, γ)) − ϕ(D(2t, t)) ≤ C4(p)e− β 8 ( 1 4− 3 p )k ϕkM Sp(L2(G)), and again by Lemma 3.17, ϕ(D(2t, t)) − c∞(ϕ) ≤ C6(p)e− t ≤ C6(p)e− β 8 ( 1 4− 3 32 ( 1 p )k ϕkM Sp(L2(G)) 4− 3 p )k ϕkM Sp(L2(G)). Hence, ϕ(D(β, γ)) − c∞(ϕ) ≤ (C4(p) + C6(p))e− β 32 ( 1 4− 3 p )k ϕkM Sp(L2(G)). Combining these results, and using that kαk2 = pβ2 + γ2 ≤ √2β, it follows that for all β ≥ γ ≥ 0, ϕ(D(β, γ)) − c∞(ϕ) ≤ C1(p)e−C2(p)kαk2k ϕkM Sp(L2(G)), 4 − 3 ( 1 where C1(p) = max{C3(p) + C6(p), C4(p) + C6(p)} and C2(p) = 1 32√2 proves the proposition. p ). This (cid:3) The values p ∈ [1, 12 11 ) ∪ (12,∞] give sufficient conditions for Sp(2, R) to fail the p,cb . We would like to point out that the set of these values might be bigger. APSchur APPROXIMATION PROPERTIES FOR NONCOMMUTATIVE Lp-SPACES 19 4. Noncommutative Lp-spaces without the OAP In the previous section we proved that Sp(2, R) does not have the APSchur p,cb for 11 ) ∪ (12,∞]. By Lemma 2.13, this directly implies the following theorem. 11 ) ∪ (12,∞], and let Γ be a lattice in Sp(2, R). Then p ∈ [1, 12 Theorem 4.1. Let p ∈ [1, 12 the noncommutative Lp-space Lp(L(Γ)) does not have the OAP (or CBAP). Combining Theorem 3.1 and Theorem 2.14, this implies the following result. Theorem 4.2. Let p ∈ [1, 12 11 )∪ (12,∞], and let G be a connected simple Lie group with finite center and real rank greater than or equal to two. Then G does not have the APSchur p,cb . Proof. Let G be a connected simple Lie group with finite center and real rank greater than or equal to two. By Wang's method [30], we may assume that G is the adjoint group, so that G has a connected semisimple subgroup H with real rank 2. Such a subgroup is closed, as was proved in [8]. It is known that H has finite center and is locally isomorphic to either SL(3, R) or Sp(2, R) [2], [26]. Since the APSchur p,cb passes to closed subgroups and is preserved under local isomorphisms (see Proposition 2.11), we conclude that G does not have the APSchur for p ∈ p,cb 11 )∪ (12,∞], since both SL(3, R) and Sp(2, R) do not have the APSchur [1, 12 p,cb for such p. (cid:3) Combining this result with Proposition 2.10 and Lemma 2.13, we obtain the main theorem of this article. Theorem 4.3. Let p ∈ [1, 12 11 ) ∪ (12,∞], and let Γ be a lattice in a connected simple Lie group with finite center and real rank greater than or equal to two. Then Lp(L(Γ)) does not have OAP (or CBAP). Acknowledgements I thank Uffe Haagerup and Magdalena Musat for valuable discussions and useful suggestions and remarks. References 1. A. Borel, Harish-Chandra, Arithmetic subgroups of algebraic groups, Ann. of Math. (2) 76 (1962), 485 -- 535. 2. A. Borel, J. Tits, Groupes r´eductifs, Inst. Hautes ´Etudes Sci. Publ. Math. No. 27 (1965), 55 -- 150. 3. N.P. Brown, N. Ozawa, C ∗-Algebras and Finite-Dimensional Approximations, Graduate Stud- ies in Mathematics, 88, American Mathematical Society, Providence, RI, 2008. 4. D. Bump, Lie Groups, Graduate Texts in Mathematics, 225, Springer-Verlag, New York, 2004. 5. J. de Canni`ere, U. Haagerup, Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups, Amer. J. Math. 107 (1985), no. 2, 455 -- 500. 6. M. Cowling, U. Haagerup, Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one, Invent. Math. 96 (1989), no. 3, 507 -- 549. 7. G. van Dijk, Introduction to Harmonic Analysis and Generalized Gelfand Pairs, Studies in Mathematics, 36, de Gruyter, Berlin, 2009. 8. B. Dorofaeff, Weak amenability and semidirect products in simple Lie groups, Math. Ann. 306 (1996), no. 4, 737 -- 742. 9. E. Effros, Z.-J. Ruan, On approximation properties for operator spaces, Internat. J. Math. 1 (1990), no. 2, 163 -- 187. 20 TIM DE LAAT 10. P. Eymard, L'alg´ebre de Fourier d'un groupe localement compact, Bull. Soc. Math. France 92 (1964), 181 -- 236. 11. J. Faraut, Analyse harmonique sur les paires de Guelfand et les espaces hyperboliques, In: Ana- lyse Harmonique, Les Cours du CIMPA, Nice, 1982, 315 -- 446. 12. J. Faraut, Analysis on Lie groups, Cambridge Studies in Advanced Mathematics, 110, Cam- bridge University Press, Cambridge, 2008. 13. U. Haagerup, Group C ∗-algebras without the completely bounded approximation property, unpublished manuscript (1986). 14. U. Haagerup, J. Kraus, Approximation properties for group C ∗-algebras and group von Neu- mann algebras, Trans. Amer. Math. Soc. 344 (1994), no. 2, 667 -- 699. 15. U. Haagerup, T. de Laat, Simple Lie groups without the Approximation Property, Duke Math. J. 162 (2013), no. 5, 925 -- 964. 16. U. Haagerup, H. Schlichtkrull, Inequalities for Jacobi polynomials, Ramanujan J. (to appear). 17. M.L. Hansen, Weak amenability of the universal covering group of SU(1,n), Math. Ann. 288 (1990), 445 -- 472. 18. S. Helgason, Differential Geometry, Lie Groups and Symmetric Spaces, Pure and Applied Mathematics, 80, Academic Press, New York, 1978. 19. S. Helgason, Groups and Geometric Analysis, Pure and Applied Mathematics, 113, Academic Press, Orlando, 1984. 20. M. Junge, Z.-J. Ruan, Approximation properties for noncommutative Lp-spaces associated with discrete groups, Duke Math. J. 117 (2003), no. 2, 313 -- 341. 21. A.W. Knapp, Lie Groups Beyond an Introduction, Birkhauser, Boston, 1996. 22. T. H. Koornwinder, The addition formula for Jacobi polynomials II. The Laplace type in- tegral representation and the product formula, Report TW 133/72, Mathematical Centre, Amsterdam 1972. 23. H. Kosaki, Applications of the complex interpolation method to a von Neumann algebra: non-commutative Lp-spaces, J. Funct. Anal. 56 (1984), 29 -- 78. 24. V. Lafforgue, M. de la Salle, Noncommutative Lp-spaces without the completely bounded ap- proximation property, Duke. Math. J. 160 (2011), no. 1, 71 -- 116. 25. G.W. Mackey, Induced representations of locally compact groups. I., Ann. of Math. (2) 55 (1952), 101 -- 139. 26. G.A. Margulis, Discrete subgroups of semisimple Lie groups, Springer-Verlag, Berlin, 1991. 27. G. Pisier, The operator Hilbert space OH, complex interpolation and tensor norms, Mem. Amer. Math. Soc. 122 (1996), no. 585. 28. G. Pisier, Non-commutative vector valued Lp-spaces and completely p-summing maps, Ast´erisque, No. 247, Soci´et´e Math´ematique de France, Paris, 1998. 29. A. Szankowski, On the uniform approximation property in Banach spaces, Israel J. Math. 49 (1984), 343 -- 359. 30. S.P. Wang, The dual space of semi-simple Lie groups, Amer. J. Math. 91 (1969), 921 -- 937. 31. H. Weyl, The Theory of Groups and Quantum Mechanics, Methuen and Co., Ltd., London, 1931. 32. J.A. Wolf, Harmonic Analysis on Commutative Spaces, Mathematical Surveys and Mono- graphs, no. 142, American Mathematical Society, Providence, RI, 2007. Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, DK-2100 Copenhagen Ø, Denmark E-mail address: [email protected]
1209.1864	1	1209	2012-09-10T02:00:37	Conditions $C_p$, $C'_p$, and $C"_p$ for $p$-operator spaces	[ "math.OA" ]	Conditions $C$, $C'$, and $C"$ were introduced for operator spaces in an attempt to study local reflexivity and exactness of operator spaces (Effros and Ruan, 2000). For example, it is known that an operator space $W$ is locally reflexive if and only if $W$ satisfies condition $C"$ (Effros and Ruan, 2000) and an operator space $V$ is exact if and only if $V$ satisfies condition $C'$ (Effros and Ruan, 2000). It is also known that an operator space $V$ satisfies condition $C$ if and only if it satisfies conditions $C'$ and $C"$ (Effros and Ruan, 2000, and Han, 2007). In this paper, we define $p$-operator space analogues of these definitions, which will be called conditions $C_p$, $C'_p$, and $C"_p$, and show that a $p$-operator space on $L_p$ space satisfies condition $C_p$ if and only if it satisfies both conditions $C'_p$ and $C"_p$. The $p$-operator space injective tensor product of $p$-operator spaces will play a key role.	math.OA	math	CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES JUNG-JIN LEE Abstract. Conditions C, C ′, and C ′′ were introduced for operator spaces in an attempt to study local reflexivity and exactness of operator spaces [ER00, Chapter 14]. For example, it is known that an operator space W is locally reflexive if and only if W satisfies condition C ′′ [ER00, Theorem 14.3.1] and an operator space V is exact if and only if V satisfies condition C ′ [ER00, Theorem 14.4.1]. It is also known that an operator space V satisfies condition C if and only if it satisfies conditions C ′ and C ′′ [ER00, Lemma 14.2.1], [Han07, Theorem 5]. In this paper, we define p-operator space analogues of these definitions, which will be called conditions Cp, C ′ p, and C ′′ p , and show that a p-operator space on Lp space satisfies condition Cp if and only if it satisfies both conditions C ′ p and C ′′ p . The p-operator space injective tensor product of p-operator spaces will play a key role. 1. Introduction to p-Operator Spaces A concrete operator space V is defined to be a closed subspace of B(H), where B(H) denotes the space of all bounded linear operators on a Hilbert space H. For each n ∈ N, the matrix algebra Mn(B(H)) with entries in B(H) can be identified with B(H ⊕ · · · ⊕ H ) via matrix multiplication   Tij     h1 ... hn   = } j=1 T1jhj n {z ...   Pn Pn j=1 Tnjhj   , [Tij] ∈ Mn(B(H)), hj ∈ H, and this gives rise to a norm k · kn on Mn(V ), which we denote by Mn(V ). It is then easy to verify that = max{kukn, kvkm}. the following two properties (called Ruan's axioms) hold: D∞: for u ∈ Mn(V ) and v ∈ Mm(V ), we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   u 0 0 v   (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)n+m M: for u ∈ Mm(V ), α ∈ Mn,m(C), and β ∈ Mm,n(C), we have kαuβkn ≤ kαkkukmkβk, where kαk is the norm of α as a member of B(ℓm 2 , ℓn 2 ), and similarly for β. An abstract operator space is a Banach space X together with a family of norms k·kn defined on Mn(X) satisfying the conditions D∞ and M above. In [Rua88], Ruan showed that these two concepts coincide and after Ruan's characterization, operator space theory has really been taken off and quickly developed Date: November 16, 2018. The author was supported by Hutchcroft Fund, Department of Mathematics and Statistics, Mount Holyoke College. 1 2 JUNG-JIN LEE into an active research area in modern analysis. Many important applications have been found in some related areas. For example, let G be a locally compact group. It is well known that G is amenable if and only if the convolution algebra L1(G) is amenable as a Banach algebra [Joh72]. We consider another Banach algebra called the Fourier algebra A(G) which consists of all coefficient functions of the left regular representation λ of G, i.e., A(G) = {ω(·) = hλ(·)ξ, ηi : ξ, η ∈ L2(G)}. By [Eym64], A(G) is a commutative Banach algebra with respect to pointwise multiplication and can be regarded as the predual of V N (G), the group von Neumann algebra of G. If G is abelian, then its dual group G is also abelian and we have the isometric isomorphism A(G) ∼= L1( G), and this suggests a relationship between the amenability of G and the amenability (as a Banach algebra) of A(G). Indeed, if A(G) is amenable, then G is amenable. In the opposite direction, Johnson showed that the Banach algebra A(G) fails to be amenable even in the case of very simple compact groups, such as SU (2, C) [Joh94]. In [Rua95], Ruan studied the operator amenability of A(G) which can be regarded as the amenability of A(G) in the category of operator spaces, and proved that a locally compact group G is amenable if and only if A(G) is operator amenable. This suggests that A(G) is better viewed as an operator space, and motivated by this observation, there has been some research [Daw10, ALR10] to study Fig`a-Talamanca- Herz Algebra Ap(G), which can be regarded as an Lp space generalization of the Fourier algebra A(G) (The reader is referred to [FT65, Her71] for more details on Ap(G)), in the framework of Lp space generalization of operator spaces. This leads to the definition of p-operator spaces we will give below. Throughout this paper, we let 1 < p < ∞. Definition 1.1. Let SQp denote the collection of subspaces of quotients of Lp spaces. A Banach space X is called a concrete p-operator space if X is a closed subspace of B(E) for some E ∈ SQp, where B(E) denotes the space of all bounded linear operators on E. Let Mn(X) denote the linear space of all n × n matrices with entries in X. For a concrete p-operator space X ⊆ B(E) and for each n ∈ N, define a norm k · kn on Mn(X) by identifying Mn(X) as a subspace of B(ℓn p (E)), and let Mn(X) denote the corresponding normed space. The norms k · kn then satisfy D∞: for u ∈ Mn(X) and v ∈ Mm(X), we have ku ⊕ vkn+m = max{kukn, kvkm}. Mp: for u ∈ Mm(X), α ∈ Mn,m(C), and β ∈ Mm,n(C), we have kαuβkn ≤ kαkkukmkβk, where kαk is the norm of α as a member of B(ℓm p , ℓn p ), and similarly for β. Remark 1.2. When p = 2, these are Ruan's axioms and 2-operator spaces are simply operator spaces because the SQ2 spaces are exactly Hilbert spaces. As in operator spaces, we can also define abstract p-operator spaces. CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 3 Definition 1.3. An abstract p-operator space is a Banach space X together with a sequence of norms k · kn defined on Mn(X) satisfying the conditions D∞ and Mp above. Thanks to the following theorem by Le Merdy, we do not distinguish between concrete p-operator spaces and abstract p-operator spaces, so from now on we will merely speak of p-operator spaces. Theorem 1.4. [LeM96, Theorem 4.1] An abstract p-operator space X can be isometrically embedded in B(E) for some E ∈ SQp in such a way that the canonical norms on Mn(X) arising from this embedding agree with the given norms. Note that a linear map u : X → Y between p-operator spaces X and Y induces a map un : Mn(X) → Mn(Y ) by applying u entrywise. We say that u is p-completely bounded if kukpcb := supn kunk < ∞. Similarly, we define p-completely contractive, p-completely isometric, and p-completely quotient maps. We write CBp(X, Y ) for the space of all p-completely bounded maps from X into Y , and to turn the mapping space CBp(X, Y ) into a p-operator space, we define a norm on Mn(CBp(X, Y )) by identifying this space with CBp(X, Mn(Y )). Using Le Merdy's theorem, one can show that CBp(X, Y ) itself is a p-operator space. In particular, the p-operator dual space of X is defined to be CBp(X, C). The next lemma by Daws shows that we may identify the Banach dual space X ′ of X with the p-operator dual space CBp(X, C) of X. Lemma 1.5. [Daw10, Lemma 4.2] Let X be a p-operator space, and let ϕ ∈ X ′, the Banach dual of X. Then ϕ is p-completely bounded as a map to C. Moreover, kϕkpcb = kϕk. If E = Lp(µ) for some measure µ and X ⊆ B(E) = B(Lp(µ)), then we say that X is a p-operator space on Lp space. These p-operator spaces are often easier to work with. For example, let κX : X → X ′′ denote the canonical inclusion from a p-operator space X into its second dual. Contrary to operator spaces, κX is not always p-completely isometric. Thanks to the following theorem by Daws, however, we can easily characterize those p-operator spaces with the property that the canonical inclusion is p- completely isometric. Proposition 1.6. [Daw10, Proposition 4.4] Let X be a p-operator space. Then κX is a p-complete contraction. Moreover, κX is a p-complete isometry if and only if X ⊆ B(Lp(µ)) p-completely isometrically for some measure µ. Conditions C, C ′, and C ′′ for operator spaces were introduced and studied in [ER00, Chapter 14] and [Han07] and they play an important role in understanding local reflexivity and exactness of operator spaces. For example, it is known that an operator space is locally reflexive if and only if it satisfies condition C ′′ [ER00, Theorem 14.3.1]. It is also known that an operator space is exact if and only if it satisfies condition C ′ [ER00, Theorem 14.4.1]. In this paper, we define p-operator space analogues of 4 JUNG-JIN LEE these conditions, which will be called conditions Cp, C ′ space satisfies condition Cp if and only if it satisfies both conditions C ′ p, and C ′′ p , and show that a p-operator space on Lp p and C ′′ p . 2. Tensor Product of p-Operator Spaces In this section, we recall basic properties of tensor products on p-operator spaces studied in [Daw10, ALR10]. We mainly focus on p-projective tensor product and p-injective tensor product. Definition 2.1. Let X, Y be p-operator spaces. Let X ⊗ Y denote the algebraic tensor product of X and Y . For u ∈ Mn(X ⊗ Y ), let kuk∧p = inf{kαkkvkkwkkβk : u = α(v ⊗ w)β}, where the infimum is taken over r, s ∈ N, α ∈ Mn,r×s, v ∈ Mr(X), w ∈ Ms(Y ), and β ∈ Mr×s,n. Daws defined and studied the p-projective tensor product [Daw10]. Note that k · k∧p gives the algebraic tensor product X ⊗ Y a p-operator space structure [Daw10, Proposition 4.8]. Furthermore, k · k∧p is the largest subcross p-operator space norm on X ⊗ Y in the sense that kx ⊗ yk ≤ kxkrkyks for all x ∈ Mr(X) and all y ∈ Ms(Y ) [Daw10, Proposition 4.8]. The p-operator space projective tensor product is defined to be the completion of X ⊗ Y with respect to this norm and is denoted by X ∧p ⊗ Y . Remark 2.2. (a) One can show that p-operator space projective tensor product is commutative, i.e., X ∧p ⊗ Y = ∧p ⊗ X p-completely isometrically. Y (b) By universality of the Banach space projective tensor product π ⊗ [BLM04, A.3.3], we have for all u ∈ X ⊗ Y . kuk∧p ≤ kukπ Let V, W , and Z be p-operator spaces, and let ψ : V × W → Z be a bilinear map. Define bilinear maps ψr,s;t,u by ψr,s;t,u : Mr,s(V ) × Mt,u(W ) → Mr×t,s×u(Z), (v, w) 7→ (ψ(vi,j , wk,l)), and let ψr;s = ψr,r;s,s. Finally define kψkjpcb = sup{kψr;sk : r, s ∈ N}. We say that ψ is jointly p-completely bounded (respectively, jointly p-completely contractive) if kψkjpcb < ∞ (respectively, kψkjpcb ≤ 1). The space of all jointly p-completely bounded maps from V × W to Z will be denoted by CBp(V × W, Z) and this space can be turned into a p-operator space in the same way as for CBp(V, W ). Here we collect some results on the p-projective tensor product for convenience. CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 5 Proposition 2.3. [Daw10, Proposition 4.9] Let X, Y , and Z be p-operator spaces. Then we have natural p-completely isometric identifications CBp(X ∧p ⊗ Y, Z) = CBp(X × Y, Z) = CBp(X, CBp(Y, Z)). In particular, (X ∧p ⊗ Y )′ = CBp(X, Y ′). As in operator spaces, the p-operator space projective tensor product is projective in the following sense: Proposition 2.4. [Daw10, Proposition 4.10] Let X, X1, Y , and Y1 be p-operator spaces. If u : X → X1 and v : Y → Y1 are p-complete quotient maps, then u ⊗ v extends to a p-complete quotient map u ⊗ v : ∧p ⊗ Y → X1 ∧p ⊗ Y1. X We now briefly introduce the p-operator space injective tensor product. Definition 2.5. Let X, Y be p-operator spaces. Regarding the algebraic tensor product X ⊗ Y as a subspace of CBp(X ′, Y ), we define the p-operator space injective tensor product X completion of X ⊗ Y in CBp(X ′, Y ). ∨p ⊗ Y to be the To be precise, for u = [uij] ∈ Mn(X ⊗ Y ) with uij = PNij tensor product norm kuk∨p is defined by k=1 xij k ⊗ yij k , the p-operator space injective kuk∨p =kukMn(CBp(X ′,Y )) = kukCBp(X ′,Mn(Y )) (2.1) = sup  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Nij  Xk=1  fst(xij k  k )yij  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mmn(Y ) : m ∈ N, f = [fst] ∈ Mm(X ′)1  , where Mm(X ′)1 denotes the closed unit ball of Mm(X ′) = CBp(X, Mm). Proposition 2.6. Suppose that X, X1, Y , and Y1 are p-operator spaces. Given p-complete contractions ϕ : X → X1 and ψ : Y → Y1, the mapping extends to a p-complete contraction ϕ ⊗ ψ : X ⊗ Y → X1 ⊗ Y1 ϕ ⊗ ψ : X ∨p ⊗ Y → X1 ∨p ⊗ Y1. 6 JUNG-JIN LEE Proof. Since ϕ ⊗ ψ = (idX1 ⊗ ψ) ◦ (ϕ ⊗ idY ), it suffices to show that ϕ ⊗ idY and idX1 ⊗ ψ extend to p-complete contractions. Let u = [uij] ∈ Mn(X ⊗ Y ). Let us write uij = PNij Since k xij k ⊗ yij k for each uij. Nij (ϕ ⊗ idY )n(u) =  Xk  ϕ(xij k  k ) ⊗ yij  ∈ Mn(X1 ⊗ Y ), (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mmn(Y ) k  k ))yij  from (2.1) it follows that k(ϕ ⊗ idY )n(u)k∨p = sup  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Nij  Xk=1  gst(ϕ(xij : m ∈ N, g = [gst] ∈ Mm(X ′ . 1)1  Define hst = gst ◦ ϕ for 1 ≤ s, t ≤ m, then h = [hst] = g ◦ ϕ ∈ Mm(X ′)1 and we have To show that idX1 ⊗ ψ is also p-completely contractive, let v = [vij ] ∈ Mn(X1 ⊗ Y ). Writing vij = k(ϕ ⊗ idY )n(u)k∨p ≤ kuk∨p. PNij k wij k ⊗ yij k , we have On the other hand, Nij kvk∨p = sup  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  Xk=1  k(idX1 ⊗ ψ)n(v)k∨p = sup  = sup  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (2.2) fst(wij k  k )yij  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mmn(Y ) : m ∈ N, f = [fst] ∈ Mm(X ′ 1)1  . : m ∈ N, f = [fst] ∈ Mm(X ′ fst(wij k ) k )ψ(yij  Nij  Xk=1  ψmn  Nij  Xk=1  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mmn(Y1) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mmn(Y1) k   k )yij   ≤kψkpcbkvk∨p. fst(wij : m ∈ N, f = [fst] ∈ Mm(X ′ 1)1  1)1  (cid:3) Remark 2.7. (a) By definition of the Banach space injective tensor product ǫ ⊗, we have kukǫ = kukB(X ′,Y ) ≤ kukCBp(X ′,Y ) = kuk∨p for every u ∈ X ⊗ Y . (b) Let u ∈ Mn(X ⊗ Y ). If Y ⊆ B(Lp(ν)) for some measure ν, then by Definition 2.5 and [Daw10, Theorem 4.3, Proposition 4.4] kuk∨p = sup{kψ(ϕst(uij ))kMrmn : m, k ∈ N, ϕ = [ϕst] ∈ Mm(X ′)1, ψ ∈ Mk(Y ′)1} CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 7 = sup{k(ϕ ⊗ ψ)n(u)k : m, k ∈ N, ϕ ∈ Mm(X ′)1, ψ ∈ Mk(Y ′)1}. (c) Let F : X ⊗ Y → Y ⊗ X denote the "flip", that is, F (P xi ⊗ yi) = P yi ⊗ xi. If Y ⊆ B(Lp(ν)) for some measure ν, then by (b) above, for every u ∈ Mn(X ⊗ Y ), we get kuk∨p = sup{k(ϕ ⊗ ψ)n(u)k : m, k ∈ N, ϕ ∈ Mm(X ′)1, ψ ∈ Mk(Y ′)1}. On the other hand, if X ⊆ B(Lp(µ)) for some measure µ as well, then kFn(u)k∨p = sup{k(ψ ⊗ ϕ)n(Fn(u))k : m, k ∈ N, ϕ ∈ Mm(X ′)1, ψ ∈ Mk(Y ′)1} and it follows that X ∨p ⊗ Y = Y ∨p ⊗ X p-completely isometrically. (d) Mr ∨p ⊗ Ms is p-completely isometrically isomorphic to Mrs. This follows immediately from [ALR10, Theorem 3.2]. At this moment, we do not know whether the p-operator space injective tensor product is injective, that is, if u : X → X and v : Y → Y are p-completely isometric injections, then we do not know ∨p ⊗ Y . But if whether u ⊗ v always extends to a p-completely isometric injection u ⊗ v : X ∨p ⊗ Y → X we assume that all the p-operator spaces under consideration are on Lp space, then we can show that ∨p ⊗ Y is a p-complete isometry as in the following proposition. This fact supports ∨p ⊗ Y → X u ⊗ v : X that the terminology p-injective tensor product is still reasonable. Proposition 2.8. Let µ1, µ2 be measures. For i = 1, 2, suppose Xi ⊆ Yi ⊆ B(Lp(µi)). Then p-completely isometrically. ∨p ⊗ X2 ⊆ Y1 ∨p ⊗ Y2 X1 Proof. For i = 1, 2, let ϕi : Xi ֒→ Yi denote the (p-completely isometric) inclusion. Since ϕ1 ⊗ ϕ2 = (ϕ1 ⊗ idY2 ) ◦ (idX1 ⊗ ϕ2), by Remark 2.7 (c) above, it suffices to show that idX1 ⊗ ϕ2 : X1 ∨p ⊗ X2 → X1 ∨p ⊗ Y2 is p-completely isometric. Note that the following diagram commutes: ∨p ⊗ X2 X1 idX1 ⊗ϕ2 / X1 ∨p ⊗ Y2 CBp(X ′ 1, X2) / CBp(X ′ 1, Y2) Since X1 ∨p ⊗ X2 ⊆ CBp(X ′ 1, X2), X1 ∨p ⊗ Y2 ⊆ CBp(X ′ isometrically, we conclude that idX1 ⊗ ϕ2 is p-completely isometric. 1, Y2), and CBp(X ′ 1, X2) ⊆ CBp(X ′ 1, Y2) p-completely (cid:3) _ / _ / 8 JUNG-JIN LEE 3. Conditions C ′ p, C ′′ p , and Cp for p-Operator Spaces In this section, we define conditions C ′ p, C ′′ p , and Cp for p-operator spaces and prove the main result. Throughout the section, µ and ν will denote measures. Lemma 3.1. Let V and W be p-operator spaces. Then the bilinear mapping Ψ : V ′ × W ′ → (V ∨p ⊗ W )′, (f, g) 7→ f ⊗ g is jointly p-completely contractive and hence the canonical mapping Ψ : V ′ ∧p ⊗ W ′ → (V ∨p ⊗ W )′ is p- completely contractive. Proof. We identify [fij] ∈ Mr(V ′) with an operator F ∈ CBp(V, Mr), and likewise [gkl] ∈ Ms(W ′) with G ∈ CBp(W, Ms). We have the identification Mrs((V ∨p ⊗ W )′) = CBp(V ∨p ⊗ W, Mrs). Let H be the map [fij ⊗ gkl] : V ∨p ⊗ W → Mrs. Then by Proposition 2.6 and Remark 2.7 (d) we have the commutative diagram ∨p ⊗ W V F ⊗G / Mrs H ;✇✇✇✇✇✇✇✇✇ ∼= ∨p ⊗ Ms Mr with kF ⊗ Gkpcb ≤ kF kpcbkGkpcb, and it follows that k[fij ⊗ gkl]k = kHkpcb ≤ kF ⊗ Gkpcb ≤ kF kpcbkGkpcb as required. (cid:3) Lemma 3.2. Let V and W be p-operator spaces. Then k · k∨p is a subcross matrix norm. In particular, for every u ∈ Mn(V ⊗ W ), we have kuk∨p ≤ kuk∧p. Proof. Just to fix notation, we identify Mr(V ) ⊗ Mq(W ) with Mrq(V ⊗ W ) by (vij ) ⊗ (wkl) 7→ (vij ⊗ wkl)(i,k),(j,l) where we have the ordering (1, 1) ≤ (1, 2) ≤ · · · ≤ (1, q) ≤ (2, 1) ≤ · · · ≤ (r, q). Hence Ir ⊗w ∈ Mr ⊗Mq(W ) = Mrq(W ) is identified with a block matrix in Mr(Mq(W )) which has r copies of w down the diagonal and 0 elsewhere. Applying axiom D∞ repeatedly hence shows that kIr ⊗ wkrq = kwkq. Then, for α ∈ Mr, the matrix α⊗w ∈ Mr⊗Mq(W ) = Mrq(W ) is the product (α⊗Iq)(Ir ⊗w) which has norm at most ∨p ⊗ W ). This kαkrkwkq by axiom Mp. Now let v ∈ Mr(V ) and w ∈ Mq(W ), and consider v ⊗ w ∈ Mrq(V tensor induces the operator T ∈ CBp(V ′, Mrq(W )) given by T (f ) = (f (vij )wkl)(i,k),(j,l) = (f (vij )) ⊗ w. For f = (fab) ∈ Mn(V ′), we see that Tn(f ) = hhf, vii ⊗ w ∈ Mnrq(W ), which by the previous paragraph has norm at most khhf, viiknrkwkq ≤ kf knkvkrkwkq. Hence kT kpcb ≤ kvkrkwkq as required. (cid:3) Let V and W be p-operator spaces and fix ϕ ∈ (V ∨p ⊗ W )′. For v0 ∈ V , we define a bounded linear functional v0 ϕ on W by v0 ϕ(w) = ϕ(v0 ⊗ w), w ∈ W. / ; CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 9 In general, when v0 = [vij ] ∈ Mr(V ) and ϕ = [ϕkl] ∈ Mn((V ∨p ⊗ W )′), we define v0 ϕ = [vij ϕkl] ∈ Mrn(W ′). Similarly, for w0 ∈ W , we define ϕw0 ∈ V ′ by ϕw0 (v) = ϕ(v ⊗ w0), v ∈ V. As in v0 ϕ above, we can extend the definition of ϕw0 for w0 ∈ Mr(W ) and ϕ ∈ Mn((V ∨p ⊗ W )′). Define a linear map ΦR V,W : V ⊗ W ′′ → (V ∨p ⊗ W )′′ by ΦR V,W (v ⊗ w′′)(ϕ) = hvϕ, w′′iW ′,W ′′ , v ∈ V, w′′ ∈ W ′′, ϕ ∈ (V Similarly, define a linear map ΦL V,W : V ′′ ⊗ W → (V ∨p ⊗ W )′′ by ΦL V,W (v′′ ⊗ w)(ϕ) = hϕw, v′′iV ′,V ′′ , v′′ ∈ V ′′, w ∈ W, ϕ ∈ (V ∨p ⊗ W )′. ∨p ⊗ W )′. Lemma 3.3. The map ΦR V,W ) defined above extends to a p-completely contractive map ΦR V,W : V ∧p ⊗ W ′′ → (V V,W (respectively, ΦL ∨p ⊗ W )′′ (respectively, ΦL V,W : V ′′ ∧p ⊗ W → (V ∨p ⊗ W )′′). Proof. Consider the bilinear map Φ : V × W ′′ → (V ∨p ⊗ W )′′ given by (v, w′′) 7→ (ϕ 7→ hvϕ, w′′iW ′,W ′′ ), then we get and Φr;s : Mr(V ) × Ms(W ′′) → Mrs((V ∨p ⊗ W )′′), ([vij ], [wkl ′′]) 7→ [Φ(vij , wkl ′′)] k[Φ(vij, wkl ′′)] = sup n (cid:26)khhΦr;s(v, w′′), ϕiik : ϕ ∈ Mn((V ∨p ⊗ W )′), kϕk ≤ 1(cid:27) . Since hhΦr;s(v, w′′), ϕii = hhvϕ, w′′ii, we have khhΦr;s(v, w′′), ϕiik = khhvϕ, w′′iik ≤ kvϕkMrn(W ′) · kw′′kMs(W ′′) and the result follows because ∨p ⊗ is a subcross matrix norm and hence kvϕkMrn(W ′) = supm {khhvϕ, wiikMrnm : w ∈ Mm(W ), kwk ≤ 1} = supm {khhϕ, v ⊗ wiikMrnm : w ∈ Mm(W ), kwk ≤ 1} ≤ kϕk · kvk ≤ kvk. Remark 3.4. Let α be a general subcross matrix norm. (a) We have a natural p-complete contraction V ∧p ⊗ W → V ⊗α W and the adjoint gives a contraction (V ⊗α W )′ → CBp(V, W ′) ⊆ B(V, W ′) given by ϕ 7→ Lϕ, hLϕ(v), wi = ϕ(v ⊗ w), ϕ ∈ (V ⊗α W )′ v ∈ V, w ∈ W. (cid:3) 10 JUNG-JIN LEE (b) Using the natural p-complete contraction V regarded as a member in (V ∧p ⊗ W )′. ∧p ⊗ W → V ⊗α W , each member in (V ⊗α W )′ can be (c) We can define ΦR V,W : V ⊗ W ′′ → (V ⊗α W )′′ and ΦL V,W : V ′′ ⊗ W → (V ⊗α W )′′ for a general subcross norm α and Lemma 3.3 remains valid if ∨p ⊗ is replaced by ⊗α. Let Ψ : V ′ ∧p ⊗ W ′ → (V ∨p ⊗ W )′ denote the canonical map, and consider the following commutative diagram V ⊗ W ′′ ΦR V,W ∨p ⊗ W )′′ / (V Ψ′ / (V ′ ∧p ⊗ W ′)′ , CBσ p,F (V ′, W ′′) ι / CBp(V ′, W ′′) where CBσ from V ′ to W ′′ and ι denotes the inclusion map. This commutative diagram shows that ΦR p,F (V ′, W ′′) denotes the space of all weak∗-continuous p-completely bounded finite rank maps V,W is one-to- one, so one can equip V ⊗ W ′′ with the p-operator space norm inherited from (V ∨p ⊗ W )′′, which will be denoted by, following the notation in [ER00], V p (or V has p) if this induced norm coincides with the p-operator space injective tensor product norm for property C ′ ∨p ⊗ : W ′′. We say that V satisfies condition C ′ every W ⊆ B(Lp(ν)). Similarly, the following diagram V ′′ ⊗ W ΦL V,W ∨p ⊗ W )′′ / (V Ψ′ / (V ′ ∧p ⊗ W ′)′ CBσ p,F (W ′, V ′′) ι / CBp(W ′, V ′′) is also commutative, ΦL V,W is one-to-one, and one can hence equip V ′′ ⊗ W with the p-operator space ∨p ∨p ⊗W . We say that V satisfies condition C ′′ ⊗ W )′′, which will be denoted by V ′′ : p p ) if this induced norm coincides with the injective tensor product norm for every norm inherited from (V (or V has property C ′′ W ⊆ B(Lp(ν)). / / / / / / CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 11 In order to define condition Cp for p-operator spaces, we need the natural map from V ′′ ⊗ W ′′ to ∨p ⊗ W )′′. To do this, let α be a general subcross matrix norm on V ⊗ W and consider the diagram (V (3.1) ∧p ⊗ W ′′)′′ (V (V ⊗α W ) ′′′′ P / (V ⊗α W )′′ , V ′′ ⊗ W ′′ ΦL V,W ′′ 8rrrrrrrrrrrr &▲▲▲▲▲▲▲▲▲▲ V ′′ ,W & ΦR (ΦR V,W )′′ &◆◆◆◆◆◆◆◆◆◆◆ 8qqqqqqqqqq V,W )′′ (ΦL (V ′′ ∧p ⊗ W )′′ where P is the restriction mapping and (ΦR V,W )′′ and (ΦL V,W )′′ are from Remark 3.4 (c). Consider the following p-complete contraction: ∧p ⊗ W )′ ∼= CBp(V, W ′) (V adj −−−−→ CBp(W ′′, V ′) ∼= (V ∧p ⊗ W ′′)′. For ϕ ∈ (V ∧p ⊗ W )′, let ϕ∧ ∈ (V ∧p ⊗ W ′′)′ denote the image of ϕ under this map. Then we have ϕ∧(v ⊗ w′′) = hvϕ, w′′iW ′,W ′′ = ΦR V,W (v ⊗ w′′)(ϕ), v ∈ V, w′′ ∈ W ′′. Moreover, ϕ∧ is weak-continuous in the second variable. Similarly, we also consider the p-complete contraction ∧p ⊗ W )′ ∼= CBp(W, V ′) (V adj −−−−→ CBp(V ′′, W ′) ∼= (V ′′ ∧p ⊗ W )′ and define ∧ϕ, and then we get that ∧ϕ(v′′ ⊗ w) = hϕw, v′′iV ′,V ′′ = ΦL V,W (v′′ ⊗ w)(ϕ), v′′ ∈ V ′′, w ∈ W, and that ∧ϕ is weak-continuous in the first variable. Remark 3.5. Let α be a general subcross matrix norm. By Remark 3.4 (b), we can still define ϕ∧ ∈ ∧p ⊗ W ′′)′ for any ϕ ∈ (V ⊗α W )′. Similarly, we can define ∧ϕ ∈ (V ′′ ∧p ⊗ W )′ for any ϕ ∈ (V ⊗α W )′. (V The next result follows by Remarks 3.4 and 3.5, and the same argument as in the proof of [Han07, Theorem 1]. Theorem 3.6. Let V and W be p-operator spaces. Let α be a subcross matrix norm on V ⊗ W and denote by V ⊗α W the resulting normed space. Then the following are equivalent. (a) There exists a separately weak-continuous extension Φ : V ′′ ⊗ W ′′ → (V ⊗α W )′′ of the natural inclusion ι : V ⊗ W → (V ⊗α W )′′. & 8 / 8 12 JUNG-JIN LEE (b) The following diagram commutes V ′′ ⊗ W ′′ ΦL V,W ′′ 8rrrrrrrrrrrr &▲▲▲▲▲▲▲▲▲▲ V ′′ ,W & ΦR (ΦR V,W )′′ &◆◆◆◆◆◆◆◆◆◆◆ 8qqqqqqqqqq V,W )′′ (ΦL (V ⊗α W ) ′′′′ P / (V ⊗α W )′′ . ∧p ⊗ W ′′)′′ (V (V ′′ ∧p ⊗ W )′′ (c) For every ϕ ∈ (V ⊗α W )′, two functionals (∧ϕ)∧ and ∧(ϕ∧) coincide on V ′′ ⊗ W ′′. (d) For every ϕ ∈ (V ⊗α W )′, Lϕ : V → W ′ is weakly compact, where hLϕ(v), wi = ϕ(v ⊗ w), v ∈ V , w ∈ W . Theorem 3.7. Let V ⊆ B(Lp(µ)) and W ⊆ B(Lp(ν)). For every ϕ ∈ (V ∨p ⊗ W )′, Lϕ is weakly compact, where Lϕ is as in Theorem 3.6 (d). Proof. Without loss of generality, we may assume kϕk(= kϕkpcb) ≤ 1. Let ΦV (respectively ΦW ) denote the embedding ΦV : V ֒→ B(Lp(µ)) (respectively, ΦW : W ֒→ B(Lp(ν))). By Proposition 2.8 and [ALR10, Theorem 3.2], we have p-completely isometric embeddings ∨p ⊗ W ֒→ B(Lp(µ)) ∨p ⊗ B(Lp(ν)) ֒→ B(Lp(µ × ν)). V Consider the diagram below: C 9s s s s s ∨p ⊗ W V ϕ ΦV ⊗ΦW B(Lp(µ)) ∨p ⊗ B(Lp(ν)) s s s s B(Lp(µ × ν)) s s s s ϕ s s By Hahn-Banach Theorem, ϕ extends to ϕ : B(Lp(µ × ν)) → C. Applying the same technique as in the proof of [ALR10, Theorem 3.6], we can find a measure space (Ω, Σ, θ) together with two vectors ξ ∈ Lp(θ), η ∈ Lp′(θ), and a unital p-completely contractive homomorphism π : B(Lp(µ × ν)) → B(Lp(θ)) such that ϕ(·) = hπ(·)ξ, ηi. Define T : B(Lp(µ)) → B(Lp(ν))′ by hT (x), yi = ϕ(x ⊗ y), x ∈ B(Lp(µ)), y ∈ B(Lp(ν)). & 8 / 8 / / _ _ 9 CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 13 Then it is easy to check that the following diagram is commutative: V ΦV Lϕ W ′ (ΦW )′ B(Lp(µ)) T / B(Lp(ν))′ Define R : B(Lp(µ)) → Lp(θ) and S : B(Lp(ν)) → Lp′(θ) by R(x) = π(x ⊗ 1)ξ, x ∈ B(Lp(µ)), and S(y) = (π(1 ⊗ y))′η, y ∈ B(Lp(ν)), then the diagram is commutative, because B(Lp(µ)) T $❏❏❏❏❏❏❏❏❏ R Lp(θ) B(Lp(ν))′ 9ttttttttt S ′ hS ′R(x), yi = hR(x), S(y)i = hπ(x ⊗ 1)ξ, (π(1 ⊗ y))′ηi = hπ(x ⊗ y)ξ, ηi = ϕ(x ⊗ y) = hT (x), yi. Combining these two commutative diagrams, we finally have Lϕ = (ΦW )′S ′RΦV , that is, Lϕ is factorized through a reflexive Banach space Lp(θ), so Lϕ is a weakly compact operator [Meg98, Propositions 3.5.4 and 3.5.11]. (cid:3) Corollary 3.8. Let V, W be p-operator spaces on Lp space. Then there exists a (necessarily unique) separately weak-continuous extension Φ : V ′′ ⊗ W ′′ → (V ∨p ⊗ W )′′ of the natural inclusion ι : V ⊗ W → (V ∨p ⊗ W )′′. Proof. Combine Theorem 3.6 and Theorem 3.7. Uniqueness follows from separate weak*-continuity. (cid:3) Now we are ready to define condition Cp for p-operator spaces. Let Φ be as in Corollary 3.8. The following commutative diagram V ′′ ⊗ W ′′ Φ ∨p ⊗ W )′′ / (V Ψ′ / (V ′ ∧p ⊗ W ′)′ CBσ p,F (V ′, W ′′) ι / CBp(V ′, W ′′) shows that Φ is injective. Thus we can equip V ′′ ⊗ W ′′ with the p-operator space structure induced by Φ, ∨p ⊗ : W ′′. We say that V ⊆ B(Lp(µ)) satisfies condition Cp (or has property which will be denoted by V ′′ : Cp) if the map Φ is isometric with respect to the injective tensor product norm for every W ⊆ B(Lp(ν)). / / _ / O O / / $ 9 / / / 14 JUNG-JIN LEE Proposition 3.9. Suppose that V ⊆ B(Lp(µ)). Then V satisfies condition Cp if and only if V satisfies both condition C ′ p and C ′′ p . Proof. Suppose that V satisfies condition Cp and W ⊆ B(Lp(ν)). By Proposition 2.8 and [Daw10, Theorem 4.3], we have a p-completely isometric embedding V ∨p ⊗ W ′′ ⊆ V ′′ ∨p ⊗ W ′′ and the bottom row in the following commutative diagram ∨p ⊗ : W ′′ V ∨p ⊗ W ′′ V V ′′ : ∨p ⊗ : W ′′ / V ′′ ∨p ⊗ W ′′ is isometric. Therefore the top row is also isometric and hence V satisfies condition C ′ p. That V satisfies condition C ′′ p can be proved using a similar argument. On the other hand, if V satisfies condition C ′′ p , we get V ′′ ∨p ⊗ W ′′ = V ′′ : ∨p ⊗ : W ′′ ֒→ (V ∨p ⊗ W ′′)′′. If V also satisfies condition C ′ p, then ∨p ⊗ W ′′ = V ∨p ⊗ : W ′′ ֒→ (V ∨p ⊗ W )′′, V and hence we have isometric inclusion V ′′ ∨p ⊗ W ′′ ֒→ (V ∨p ⊗ W )′′′′. Since V ′′ ∨p ⊗ W ′′ ⊂ (V ∨p ⊗ W )′′ and (V ∨p ⊗ W )′′ ֒→ (V ∨p ⊗ W )′′′′ isometrically, the inclusion V ′′ ∨p ⊗ W ′′ ⊆ (V ∨p ⊗ W )′′ must be isometric. (cid:3) 4. Acknowledgement The author would like to thank the reviewer for his/her valuable comments, especially the ones that led to Remarks 3.4 and 3.5, to improve the quality of the paper. References [ALR10] Guimei An, Jung-Jin Lee, and Zhong-Jin Ruan. On p-approximation properties for p-operator spaces. Journal of Functional Analysis, 259:933 -- 974, 2010. [BLM04] David P. Blecher and Christian Le Merdy. Operator algebras and their modules -- an operator space approach, volume 30 of London Mathematical Society Monographs. New Series. The Clarendon Press Oxford University Press, Oxford, 2004. Oxford Science Publications. [Daw10] Matthew Daws. p-operator spaces and Fig`a-Talamanca-Herz algebras. J. Opeator Theory, 63:47 -- 83, 2010. [ER00] E. Effros and Z.-J. Ruan. Operator Spaces. Oxford Science Publications, 2000. [Eym64] P. Eymard. L'algebre de Fourier d'un groupe localement compact. Bull. Soc. Math. France, 92:181 -- 236, 1964. / / _ _ / CONDITIONS Cp, C ′ p, AND C ′′ p FOR p-OPERATOR SPACES 15 [FT65] A. Fig`a-Talamanca. Translation invariant operators in Lp. Duke Math. J., 32:495 -- 501, 1965. [Han07] K. H. Han. An operator space approach to condition C. J. Math. Anal. Appl., 336:569 -- 576, 2007. [Her71] C. Herz. The theory of p-spaces with an application to convolution operators. Trans. Amer. Math. Soc., 154:69 -- 82, 1971. [Joh72] B. Johnson. Cohomology in Banach algebras. Mem. Amer. Math. Soc., 127, 1972. [Joh94] B. Johnson. Non amenablity of the Fourier algebra of a compact group. J. London Math. Soc., 50:361 -- 374, 1994. [LeM96] Christian LeMerdy. Factorization of p-completely bounded multilinear maps. Pacific Journal of Mathematics, 172:187 -- 213, 1996. [Meg98] Robert E. Megginson. An Introduction to Banach Space Theory. Springer, 1998. [Rua88] Z.-J. Ruan. Subspaces of C ∗-algebras. J. Funct. Anal., 76:217 -- 230, 1988. [Rua95] Z.-J. Ruan. The operator amenability of A(G). Amer. J. Math., 117:1449 -- 1474, 1995. Department of Mathematics and Statistics, Mount Holyoke College, South Hadley, MA 01075, USA E-mail address, Jung-Jin Lee: [email protected]
1001.0730	1	1001	2010-01-05T16:20:31	Operator algebras associated to integral domains	[ "math.OA" ]	We study operator algebras associated to integral domains. In particular, with respect to a set of natural identities we look at the possible nonselfadjoint operator algebras which encode the ring structure of an integral domain. We show that these algebras give a new class of examples of semicrossed products by discrete semigroups. We investigate the structure of these algebras together with a particular class of representations.	math.OA	math	OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS BENTON L. DUNCAN Abstract. We study operator algebras associated to integral do- mains. In particular, with respect to a set of natural identities we look at the possible nonselfadjoint operator algebras which en- code the ring structure of an integral domain. We show that these algebras give a new class of examples of semicrossed products by discrete semigroups. We investigate the structure of these algebras together with a particular class of representations. Recently, in [4] and [8] the notion of a regular C ∗-algebra associated to an integral domain was introduced as a generalization of a con- struction due to Cuntz, [3]. In these papers the authors associate a C ∗-algebra by representing ℓ2(R) and viewing the operators given by the regular representation R acting on ℓ2(R). In addition they show that the C ∗-algebras so constructed are universal with respect to a col- lection of identities that encode information about the integral domain. Now one can view an integral domain as an additive group together with an action on the additive group given by multiplication by nonzero elements of the ring. This suggests that an important viewpoint for studying operator algebras associated to integral domains is through the use of crossed products. More importantly, since crossed products are well understood many of the significant results can be made brief through the technology of crossed products. For this paper we wish to investigate the operator algebras with slightly less restrictive identities imposed by only natural ring-theoretic constraints. This gives rise to operator algebras with a more natural crossed product structure. However since the multiplication in a ring need not act as automorphisms on the additive group, crossed products are not entirely appropriate. To avoid this we use the nonselfadjoint operator algebras where possible. This goes back to a construction of Arveson and Josephson [1] which was generalized by Peters in [10]. This semicrossed product is a nonselfadjoint operator algebra which encodes 2000 Mathematics Subject Classification. 47L74, 47L40. Key words and phrases. integral domains, semicrossed products. 1 2 BENTON L. DUNCAN the same dynamics as the crossed product but does not require that the action on a topological space be via homeomorphisms. While one may worry that we lose too much information when we lose the ∗-structure of the C ∗-algebra, in recent work [5] it was shown that the semicrossed products of Peters in fact encode the action of a continuous self map on a topological space in a manner which is unique up to conjugacy of the map. This is even true when the map is not a homeomorphism and hence unlike with C ∗-algebras the nonselfadjoint operator algebras can be used as a topological invariant. It is these motivating examples which have led us to study the semi- crossed product algebras in the context of integral domains. In this paper we have defined the universal operator algebra associated to an integral domain (note the different conditions we require from those of Cuntz and Li). We then study the situation in the case that our inte- gral domain is a field. Here the semicrossed product and the crossed product coincide and we can use standard results for crossed products to prove facts about the algebra. After viewing the case of the integral domain being a field we focus on the situation where this may not be true. Here the semicrossed product technology is necessary, however similar results carry through. After defining the requisite notion of semicrossed product and proving some first results in the context of in- tegral domains we prove some results which show that the algebras thus defined are distinct from the algebras studied by Peters. In the last section we analyze what we call unitary representations of an integral domain R. We show that every such representation factors through a regular unitary representation. We describe some standard notation we intend to use. If R is an integral domain we write Q(R) for the field of quotients. We write (R, +) for the additive group on R. This group is a locally compact an element a in (R, +). discrete group. We denote the Pontryagin dual of this group by (dR, +) and denote byba the element in the Pontryagin dual corresponding to The author would like to thank Jim Coykendall and Sean Sather- Wagstaff for helpful discussions about and examples of integral do- mains. 1. Universal algebras of integral domains Let R be an integral domain. Given a Hilbert space H we define an isometric representation of R to be {Sr ∈ B(H) : r ∈ R×} a collection of isometries together with unitaries {U n ∈ B(H) : n ∈ R} that satisfy the relations: OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS 3 (1) SrSt = Srt for all r, t ∈ R×, (2) U nU m = U m+n for all m, n ∈ R, and (3) U nSr = SrU rn for all r ∈ R×, n ∈ R. If the Sr are unitaries for all r ∈ R× then we say that the represen- tation is a unitary representation. We present first two examples: Example. (The regular representation of an integral domain) Let H be equal to ℓ2(R), with eq denoting the characteristic function of {q} ⊆ R. Define operators U n and Sr as follows: U n Xq∈R Sr Xq∈R ζqeq! =Xq∈R ζqeq! =Xq∈R ζqeq+n ζqerq. It is not difficult to see that {U n} and {Sr} give rise to an isometric representation of R. Example. (The regular unitary representation of an integral domain) Let K be equal to ℓ2(Q(R)), with eq denoting the characteristic func- ) = eq for every q ∈ Q(R)). r An important point to notice is that H ⊆ K and further H is tion of {q} ⊆ Q(R). We use the same formulas to define fU n and eSr. Notice this time however that for all r, eSr is onto and hence a unitary. (To see this notice that eSr(e q an invariant subspace for the collections {eSr} and {fU n}. Further Sr = PHeSrH for all r and U n = PHfU nH. For this reason we call this representation the regular unitary representation of R. We let A(R) be the norm closed operator algebra generated by uni- taries {un : n ∈ R} and isometries {sr : r ∈ R×} which is universal for isometric representations of R. We will denote the elements of A(R) with lower case letters to distinguish from a representation of R for which we will use upper case letters. We notice some initial facts about the algebra A(R). Lemma 1. A(R) is unital with u0 = s1. Proof. Let {U r, Sr} be an isometric representation of R. Then notice that U0 is an idempotent unitary, hence 1 = U ∗ 0 U0. Then U0 = 1 · U0 = U ∗ 0 U0 = 1. A similar argument yields the same result for S1. Since this is true for an arbitrary isometric representation of R, the result follows for A(R). (cid:4) 0 = U ∗ 0 U 2 4 BENTON L. DUNCAN Lemma 2. If r is invertible then sr is a unitary with s∗ r = sr−1 Proof. This follows from the previous lemma since SrSr−1 = S1 = Sr−1Sr, for any isometric representation of R. (cid:4) It follows that if R is a field then A(R) is a C ∗-algebra. In addition, in this case, the regular representation is a unitary representation. In fact we have the following: Proposition 1. R is a field if and only if every isometric representa- tion is a unitary representation. Proof. This comes from the fact that for the regular representation (Sr)∗ is in the algebra if and only if r is invertible. (cid:4) Finally we can see that A(R) is functorial for ring monomorphisms since any isometric representation of a ring R2 will give rise to an isometric representation of a ring R1 if there is a ring monomorphism from R1 into R2.. Proposition 2. A(R) is functorial in the sense that if π : R1 → R2 is a unital ring monomorphism then there is an induced completely contractive representation π : A(R1) → A(R2) 2. The universal C ∗-algebra for a field We now analyze the case where R is a field. Here any isometric representation is, in fact, a unitary representation. We let (R, +) de- note the additive group in R. Notice that R× acts on C ∗(R, +) as ∗-automorphisms via the mapping αλ(U n) = U λn where U n is the uni- tary in C ∗(R, +) corresponding to n ∈ (R, +). This allows us to rewrite C ∗(R) as a crossed product. Proposition 3. Let R be a field, then A(R) ∼= C ∗(R, +) ⋊ R×. Proof. We begin by noting (see [2, II.10.3.10]) that since C ∗(R, +) is unital and R× is discrete we have C ∗(R, +) ⊂ C ∗(R, +) ⋊ R× via a representation π0 and there is a natural map ρ0 : R× → C ∗(R, ×) ⋊R×. Together (π0, ρ0) give rise to a covariant representation of the triple (C ∗(R, +), R×, α). Now analyzing the covariance conditions that define C ∗(R, +) ⋊ R× we see that ρ0(r)π0(n)ρ0(r)∗ = π0(rn). Now for each n, π0(n) is a unitary and for each r, ρ0(r) is a unitary and hence the natural covariant representation π0 ⋊ ρ0 gives rise to a unitary representa- tion of R, and hence there is a completely contractive representation ι : A(R) → C ∗(R, +) ⋊ R×. OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS 5 Next notice that any unitary representation of A(R) gives rise to a covariant representation of (C ∗(R, +), R×, α) and hence kxk ≤ kι(x)k so that ι is faithful. (cid:4) Since C ∗(R, +) and R× are both abelian we have that the universal norm on the crossed product C ∗(R, +) ⋊R× coincides with the reduced norm [11, Theorem 7.13]. In addition, we have that the algebra A(R) is nuclear [11, Corollary 7.18], when R is a field. We next analyze the regular representation of A(R) where R is a field. Proposition 4. Let R be a field, then the regular representation of A(R) is faithful. Proof. This follows from analyzing the regular representation of the algebra C ∗(R, +) ⋊α R×, which is faithful since R× is amenable. In effect, we take the left regular representation of C ∗(R, +) acting on L2(R, +) and add to this the action of R× via ∗-automorphisms. This is exactly the construction of the regular representation of R and hence the two coincide. (cid:4) Other facts about A(R) can also be explained using the crossed prod- uct machinery. Proposition 5. For a field R the algebra A(R) is not simple. dual of the locally compact abelian group (R, +). Notice that the ∗- Proof. Notice that C ∗(R, +) = C(dR, +) where (dR, +) is the Pontryagin automorphisms αλ induce a homeomorphism cαλ on (dR, +) which has a fixed point for each λ; in particular, cαλ(b0) =b0) for all λ. It follows that there is a nontrivial invariant ideal in C ∗(R, +) for the action by R× and hence, see [11, Section 3.5] there is a nontrivial induced ideal in C ∗(R, +) ⋊ R×. (cid:4) We can, however completely describe the ideal structure of A(R) in the case of a field. Theorem 1. A(R) is ∗-isomorphic to C ⊕ A where A is a simple C ∗- Proof. We use the nontrivial invariant ideal from the previous propo- sition. algebra. In fact, A is ∗-isomorphic to C0((dR, +) \ {b0}) ⋊ R×. In particular, since C ∗(R, +) = C(dR, +) let π be the multi- plicative linear function given on C(dR, +) by evaluation at b0. Fur- ther, if we let cαλ be the induced homeomorphism on (dR, +) given by the ∗-automorphism αλ for all λ ∈ R×. The induced represen- tation, π is a multiplicative linear functional and hence has range C. Hence, A(R) ∼= C ⊕ ker π. We now wish to describe π. So let 6 BENTON L. DUNCAN σ : C0(dR, +) → C0((dR, +) \ {0}) be the restriction mapping. Further, if λ ∈ R× then cαλ, the homeomorphism on dR, + induced by the au- tomorphism αλ, then the range of σ is invariant under cαλ and hence there is a map τ : C0((dR, +) \ {0}) ⋊ R× → ker π. But since R× acts transitively on (dR, +) the crossed product C0((dR, +) \ {0}) ⋊ R× is simple and hence the map τ must be an isomorphism. (cid:4) 3. Semicrossed products for discrete semigroups The preceding construction suggests that for non-field integral do- mains the crossed product may be replaced by a semicrossed product. We quickly outline the relevant construction referring to [10] for moti- vation and to [7] for more information about this semicrossed product. Given a compact Hausdorff space X we say that a semigroup S acts on X via continuous maps if for each s ∈ S there is a continuous map τs : X → X with τs ◦ τt = τst. If S is unital with identity 0 we will assume that τ0 is the identity map. Say that a pair (π, St) is an isometric covariant representation of (X, S, τs) if π is a representation of C(X) on a Hilbert space H and for each t ∈ S, St is an isometry in B(H) such that Stπ(f (x)) = π(f (τt(x)))St for all x ∈ X. It is not hard to see that given the triple (X, S, τs), there is a non- trivial isometric covariant representation. The construction follows in the same manner as in [10], we only outline the idea here. Let H = ℓ2(X, S) where this latter Hilbert space is sequences indexed over elements of S with entries from x, with canonical basis {es}. Define π : C(X) → B(H) by π(f (x)) = (f (τs(x)))s∈S. Then set St(es) = est and extend by linearity. Then (π, St) is an isometric covariant repre- sentation of (X, S, τs). We say that the universal operator algebra generated by all isometric covariant representations of (X, S, τs) is the semicrossed product of X by S via τ . We denote this algebra as C(X) ⋊τ S. As examples notice that if α is a single endomorphism of a C ∗-algebra then we are in the situation described in Peter's original work [10], where the semigroup is Z+. For an example on the opposite end of the spectrum we can view the semicrossed product of [6] as a semicrossed product where the monoid is the free monoid on n generators. Returning to an integral domain R, we let αr be the ∗-endomorphism of C ∗(R, +) induced by the group endomorphism given by left multipli- cation by r. This gives a map from R× into the set of ∗-endomorphisms of C ∗(R, +). Notice that since we are in an integral domain each of these endomorphisms is injective. However they are only surjective when r is a unit. OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS 7 Proposition 6. The algebra A(R) is completely isometrically isomor- phic to C ∗(R, +) ⋊α R×. Proof. We will show that any isometric representation of R gives rise to an isometric covariant representation of the pair (C ∗(R, +) ⋊α R×) and vice-versa, hence the two algebras will be completely isometrically isomorphic. So let {U n : n ∈ R} and {Sr ∈ R×} be an isometric representation of R. Then the map n 7→ U n gives rise to a representation of C ∗(R, +), call it π. Further the isometries Sr will satisfy SrU nS ∗ r = U rn and hence the pair (π, {Sr}) will be an isometric covariant representation of (C ∗(R, +), R×, α). Let (π, {Sr}) be an isometric covariant representation of the alge- bra (C ∗(R, +), R×, α). Then {π(un)} is a collection of unitaries and {Sr} is a collection of isometries that trivially satisfy the first two con- ditions of an isometric representation of R. Further SrU nS ∗ r = U rn so that SrU n = U rnSr for all r ∈ R× so that we have an isometric representation of R. (cid:4) Notice that in the case of the algebra A(R) for an integral domain R the semigroup R× will always be commutative with no torsion. In addition the semigroup R× can be viewed as a spanning cone for the group Q(R)× (see [9, Page 60] for the definition of a spanning cone for a group). We can actually improve our characterization of A(R) as a semi- crossed product by looking at a more tractable semigroup. Let U(R) denote the group of units in R. Now R× is a commutative monoid which contains U(R) as a normal submonoid. We let M(R) denote the monoid R×/U(R). Notice that R× ⊆ Q(R)× and further that M(R) ⊆ Q(R)×/U(R), this latter group we call G(R). For u ∈ U(R) let αu : C ∗(R, +) → C ∗(R, +) be the ∗-automorphism induced by the automorphism of (R, +) that corresponds to left multi- plication by u. Next for r ∈ M(R) define a covariant representation of the triple (C ∗(R, +), U(R), α) by βr(U n) = U nr, ρ(Su) = Su. This covariant representation induces a ∗-endomorphism of C ∗(R, +) ⋊α U(R). Hence, β gives rise to a map from M(R) into the set of ∗- endomorphisms of C ∗(R, +) ⋊α U(R). Theorem 2. The algebra A(R) is completely isometrically isomorphic to (C ∗(R, +) ⋊α U(R)) ⋊β M(R), and the diagonal algebra A(R) ∩ A(R)∗ = (C ∗(R, +) ⋊α U(R)). 8 BENTON L. DUNCAN Proof. We will show that any isometric representation of R gives rise to an isometric covariant representation of the pair (C ∗(R, +) ⋊α U(R), M(R)) and vice-versa, hence the two algebras will be completely isometrically isomorphic. So let {U n : n ∈ R} and {Sr : r ∈ R×} be an isometric representation of R. Then define a covariant representation of (C ∗(R, +), U(R), α) by n 7→ U n and r 7→ Sr for all r ∈ U(R). This yields a representation π of C ∗(R, +) ⋊α U(R). Next notice that Sλπ(U n) = π(U λn)Sλ for all λ ∈ M(R), and SλSr = SrSλ for all r ∈ U(R), λ ∈ M(R). Hence we have an isometric covariant representation of (C ∗(R, +) ⋊α U(R), M(R), β). Finally we take an isometric covariant representation (π, S) of the triple (C ∗(R, +) ⋊α U(R), M(R), β). Define U n = π(n) and Sr = π(Sr) if r ∈ U(R), else Sr = Sr. This gives rise to an isometric representation of R. The last result follows from Corollary 2 of [7]. (cid:4) Notice that if R is not a field U(R) does not act transitively on the nonunital subalgebra of C ∗(R, +), as in the case of a field. In fact we have the following fact. Corollary 1. The diagonal is isomorphic to C ⊕ A where Further, A is simple if and only if R is a field. A ∼= C0((dG, +) \b0) × U(R). Proof. That A is not simple follows from the fact that U(R) does not act transitively on C0( \C ∗(G, +) \ 0) unless R× = U(R). (cid:4) We now prove some other facts about the relationship between the integral domain R and the structure of the algebra A(R). Proposition 7. A(R) ∼= C(X) ⋊ S, where S is a monoid with no nontrivial invertible elements if and only if the identity of R is the only unit. Further if U(R) = {1} then M(R) will not be finitely generated. Proof. If the identity of R is the only unit, then we have C ∗(R, +) × U(R) ∼= C ∗(R, +) and M(R) has no nontrivial invertible elements, else M(R) ∩ U(R) 6= {1}. Notice that if A(R) ∼= C(X) × S where S is a monoid with no nontrivial invertible elements then the diagonal algebra A(R)∩A(R)∗ = C(X). However, if x ∈ U(R) with x 6= 1, then x 6∈ M(R) and hence OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS 9 Sx ∈ C ∗(R, +) × U(R) = A(R) ∩ A(R)∗. But notice that SxU n 6= U nSx unless x = 1 and hence C ∗(R, +) × U(R) is not commutative. Assume now that U(R) = {1} and M(R) is finitely generated by the set {x1, x2, · · · , xn}, then 1 + x1x2 · · · xn 6∈ U(S) so 2 · · · xkn 1 + x1x2 · · · xn = xk1 1 xk2 n with at least one kj 6= 0. We will assume without loss of generality that k1 6= 0. Then 1 = x1(x2x3 · · · xn − xk1−1 n ) which implies 1 that x1 is a unit yielding a contradiction. (cid:4) xk2 2 · · · xkn As a corollary we have the following. Proposition 8. If R is a unique factorization domain and A(R) ∼= A × Z+ where A is a C ∗-algebra, then A is not commutative and U(R) is not trivial. Proof. We will assume that A is commutative and hence U(R) is trivial. In particular A = C ∗(R, +) = C0(dR, +). Now let x1 and x2 be two irreducible elements of M(R), then define two two-dimensional nest representations of A(R) by 0 f (b1)(cid:21) , for f ∈ C0(dR, +) πi(f ) =(cid:20)f (bxi) πi(Sxi) =(cid:20)0 1 0 0(cid:21) 0 πi(Sr) = 0 for r 6= xi. It follows from the description of the two-dimensional nest represen- tations of A × Z+, see [5], that x1 = x2 which contradicts the fact that M(R) must be infinitely generated. (cid:4) It follows that if R is a unique factorization domain A(R) is never a semicrossed product in the sense of [10] and hence this collection of algebras presents a unique type of semicrossed product. 4. Unitary representations of R Finally we wish to analyze the unitary representations of A(R). Lemma 3. There is a canonical completely contractive representation i : A(R) → C ∗(Q(R)). Proof. As R ⊆ Q(R) the inclusion map provides an isometric represen- tation of R inside C ∗(Q(R)) and hence the induced map on A(R) is completely contractive. (cid:4) 10 BENTON L. DUNCAN Proposition 9. Let π : A(R) → B(H) be a unitary representation. There exists a ∗-representation τπ : C ∗(Q(R)) → C ∗(π(A(R)) which is onto and satisfies τπ ◦ i(x) = π(x) for all x ∈ A(R). Proof. Since π is a unitary representation we know that π(sr) = Tr is a unitary for all r ∈ R×. For all h p q . We also define π(cid:16)u[ p q ](cid:17) = TqV pT ∗ TpT ∗ unitaries TpT ∗ the induced representation τπ will be the required ∗-representation. q . We need only show that the q satisfy the relations for Q(R) and hence qi ∈ Q(R)× we define π(cid:16)s[ p q ](cid:17) = q and TqV pT ∗ Notice first that TpTq = TqTp and T ∗ p T ∗ q = T ∗ q T ∗ p since π is a unitary representation of A(R). It then follows that q TqTpT ∗ q TpTqT ∗ q Tp q = T ∗ = T ∗ = T ∗ TpT ∗ q q for all q, p ∈ R×. It follows that π(cid:16)s[ p1 ] and [ p2 q2 Next notice that V pTq = TqV pq and T ∗ ] in Q(R)×. [ p1 q1 q1 ](cid:17) π(cid:16)s[ p2 q2 ](cid:17) = π(cid:16)s[ p1p2 q1q2 ](cid:17) for all Tq1V p1T ∗ q1Tq2V p2T ∗ q V p = V pqT ∗ q2 = Tq1Tq2V p1q2V q1p2T ∗ = Tq1q2V p1q2+q1p2Tq1q2. q and hence q1T ∗ q2 q1 Next we have that In other words π(cid:16)u[ p1 π(cid:16)u[ p1 ](cid:17) π(cid:16)u[ p2 ](cid:17) π(cid:16)s[ p2 q1 q2 q2 ]+[ p2 q2 ](cid:17). q1Tp2T ∗ q2 q1 ](cid:17) = π(cid:16)u[ p1 ](cid:17) = Tq1V p1T ∗ = π(cid:16)s[ p2 = Tp2T ∗ = Tp2T ∗ q2 q2T ∗ q1 q1 q1q2 q2Tq1Tq2V p1p2T ∗ q2Tq1q2V p1p2T ∗ ][ p2 q2 ](cid:17) π(cid:16)u[ p1 ](cid:17) . 1 ](cid:17) = TpT ∗ Hence the C ∗-algebra generated by the Tp and V n satisfies the relations for Q(R). We call the induced representation τπ. Finally we note that τπ ◦ i(sp) = τπ(cid:16)s[ p p ∈ Q× and τπ ◦ i(un) = τπ(cid:0)u[ n 1 ](cid:1) = T1VnT ∗ hence τπ ◦ i(x) = π(x) for all x ∈ A(R). 1 = Tp for all 1 = Vn for all n ∈ R and (cid:4) It would follow that if every isometric representation of R dilated to a unitary representation (as for example the regular representation does), then we could identify the C ∗-envelope of A(R) as a crossed product, since the canonical representation i would be completely isometric. We OPERATOR ALGEBRAS ASSOCIATED TO INTEGRAL DOMAINS 11 do not think this is likely since this does not even work in the case of C(X) ⋊α Z+ where α is a non-surjective continuous mapping, see [6]. References [1] W. Arveson and K. Josephson. Operator algebras and measure preserving au- tomorphisms. II J. Functional Analysis 4 (1969) 100 -- 134. [2] B. Blackadar, Operator algebras. Theory of C ∗-algebras and von Neumann algebras. Encyclopaedia of Mathematical Sciences, 122. Operator Algebras and Non-commutative Geometry, III. Springer-Verlag, Berlin, 2006. [3] J. Cuntz, C ∗-algebras associated with the ax + b-semigroup over N K-theory and noncommutative geometry, 201 -- 215, EMS Ser. Congr. Rep. Eur. Math. Soc. Zrich, 2008. [4] J. Cuntz and X. Li, The regular C ∗-algebra of an integral domain, preprint, 2008. [5] K. Davidson and E. Katsoulis, Isomorphisms between topological conjugacy algebras, J. reine angew. Math. 621, (2008), 29 -- 51. [6] K. Davidson and E. Katsoulis, Operator algebras for multivariable dynamics, to appear in Mem. Amer. Math. Soc. , 2007. [7] B. Duncan and J. Peters, Dynamics and semicrossed products by discrete semigroups, preprint, 2010. [8] X. Li, Ring C ∗-algebras, preprint, 2009. [9] V. Paulsen, Completely bounded maps and operator algebras Cambridge Uni- versity Press, Cambridge, 2002. [10] J. Peters, Semicrossed products of C ∗-algebras, J. Funct. Anal. 59 (1984) 498 -- 534. [11] D. Williams, Crossed products of C ∗-algebras American Math. Soc. Provi- dence, 2007. Department of Mathematics, North Dakota State University, Fargo, North Dakota, USA E-mail address: [email protected]
1902.02572	1	1902	2019-02-07T11:29:59	Subfactors and Hecke groups	[ "math.OA", "math.GT", "math.NT" ]	We study a relation between the Hecke groups and the index of subfactors in a von Neumann algebra. Such a problem was raised by V. F. R. Jones. We solve the problem using the notion of a cluster C*-algebra.	math.OA	math	SUBFACTORS AND HECKE GROUPS ANDREY GLUBOKOV1 AND IGOR NIKOLAEV2 Abstract. We study a relation between the Hecke groups and the index of subfactors in a von Neumann algebra. Such a problem was raised by V. F. R. Jones. We solve the problem using the notion of a cluster C ∗-algebra. The following problem can be found in [Jones 1991] [5, p.24]: 1. Introduction "Consider the subgroup Gλ of SL2(R) generated by ( 1 λ 0 1 ) and (cid:0) 0 1 −1 0(cid:1). For what values of λ > 0 is it discrete? Answer: λ = 2 cos(cid:0) π n(cid:1) , n = 3, 4, . . . or λ ≥ 2. (...) We have been unable to find any direct connection between this result and Theorem 3.1 (Jones Index Theorem). It is a tantalizing situation." The aim of our note is to solve the problem in terms of the cluster C∗-algebras [6, Section 4.4.3]. To give an idea, let D = {z = x + iy ∈ C r ≤ z ≤ R} be an annulus in the complex plane. Consider the Schottky uniformization of D, i.e. D ∼= CP 1/AZ, (1.1) where CP 1 := C ∪ {∞} is the Riemann sphere and A ∈ SL2(C)/ ± I is a matrix acting on the CP 1 by the Mobius transformation. It follows from [Glubokov & Nikolaev 2018] [3] and Section 2.2, that the index of subfactors in a von Neumann algebra coincides with the square of trace of matrix A, i.e. tr2 (A) ∈ [4,∞) [ {4 cos2(cid:16) π n(cid:17) n ≥ 3}. (1.2) To solve the Jones Problem, we prove in Section 3 that D is a ramified double cover of the orbifold H/Gλ, where H := {x + iy ∈ C y > 0} is the Lobachevsky half-plane and the group Gλ acts on H by the linear fractional trasformations. Since such a cover takes the square root of the moduli parameter tr2 (A) of D, we conclude that λ = tr (A). In other words, the Jones Index Theorem (1.2) is equivalent to the following well-known result: Theorem 1.1. ([Hecke 1936] [4, Satz 1,2 & 6]) The Gλ is a discrete subgroup of SL2(R) if and only if λ ∈ [2,∞) [ {2 cos(cid:16) π n(cid:17) n ≥ 3}. (1.3) 2010 Mathematics Subject Classification. Primary 46L37; Secondary 20H10. Key words and phrases. Hecke group, subfactors, cluster C ∗-algebra. 1 2 A.GLUBOKOV AND I.NIKOLAEV Remark 1.2. The group Gλ appears in the study of the Riemann zeta function [Hecke 1936] [4]. In particular, Hecke's Theorem says that the space of automorphic functions corresponding to Gλ is (i) infinite-dimensional, if λ > 2 or (ii) finite- dimensional, if λ ∈ {2 cos(cid:0) π [Hecke 1936] [4]. On the other hand, cases (i) and (ii) follow from the Sherman- Zelevinsky Theorem for the cluster C∗-algebras of rank 2, see [Glubokov & Nikolaev 2018] [3]. n(cid:1) n ≥ 3}. The proof of this fact is purely analytic The article is organized as follows. Section 2 contains a brief review of prelimi- nary results. Theorem 1.1 is proved in Section 3. 2. Preliminaries The cluster C∗-algebras and their K-theory are covered in [6, Section 4.4.3]. A correspondence between the cluster C∗-algebra of an annulus D and the Jones Index Theorem was established in [Glubokov & Nikolaev 2018] [3]. The Hecke groups were introduced in [Hecke 1936] [4]. 2.1. Cluster C∗-algebras. A cluster algebra A (x, B) of rank n is a subring of the field of rational functions in n variables depending on a cluster of variables x = (x1, . . . , xn) and a skew-symmetric matrix B = (bij) ∈ Mn(Z); the pair (x, B) is called a seed. A new cluster x′ = (x1, . . . , x′k, . . . , xn) and a new skew-symmetric matrix B′ = (b′ij) is obtained from (x, B) by the exchange relations: n xmax(bik ,0) xkx′k = i Yi=1 b′ij = (−bij xmax(−bik,0) i , + n Yi=1 bij + bikbkj +bikbkj 2 if i = k or j = k otherwise. (2.1) The seed (x′, B′) is said to be a mutation of (x, B) in direction k, where 1 ≤ k ≤ n; the algebra A (x, B) is generated by cluster variables {xi}∞i=1 obtained from the initial seed (x, B) by the iteration of mutations in all possible directions k. The Laurent phenomenon says that A (x, B) ⊂ Z[x±1], where Z[x±1] is the ring of the Laurent polynomials in variables x = (x1, . . . , xn) depending on an initial seed (x, B). The A (x, B) is a commutative algebra with an additive abelian semigroup consisting of the Laurent polynomials with positive coefficients. Thus the algebra A (x, B) is a countable abelian group with an order satisfying the Riesz interpolation property, i.e. a dimension group. A cluster C∗-algebra A(x, B) is an AF-algebra, such that K0(A(x, B)) ∼= A (x, B), (2.2) where ∼= is an isomorphism of the dimension groups [6, Section 4.4.3]. 2.2. Schottky uniformization of D. Consider the Riemann surface D (an an- nulus) defined by the formula (1.1). We shall use the Schottky uniformization of D by the loxodromic transformations. Namely, let CP 1 := C ∪ {∞} be the Riemann sphere and consider the Mobius transformation of CP 1 given by the formula: z 7−→ k z, where z ∈ CP 1 and k 6= 1. (2.3) SUBFACTORS AND HECKE GROUPS 3 It is easy to see, that (2.3) can be written in the matrix form: A = √k 0 0 1√k! ∈ P SL2(C) := SL2(C)/ ± I. It is well known, that (2.3) is a loxodromic transformation if and only if ∈ C \ [0, 4]. The Schottky uniformization of D is given by the formula: k tr2 (A) = (k + 1)2 D ∼= CP 1/AZ. (2.4) (2.5) (2.6) 2.3. Admissible values of tr2 (A). Recall that the moduli space of the annulus D is given by the formula: TD =(cid:26)t = R r t > 1(cid:27) . (2.7) We consider a cluster C∗-algebra A(D) associated to a canonical triangulation of D [Fomin, Shapiro & Thurston 2008] [1, Example 4.4]. It follows from the Sherman- Zelevinsky Theorem for the algebra A(D), that the admissible values of the "index" (t+1)2 t must belong to the set: see [Glubokov & Nikolaev 2018] [3] for the proof. [4,∞) [ {4 cos2(cid:16) π n(cid:17) n ≥ 3}, (2.8) We set k = t in the formulas (2.3) - (2.5). Comparing (2.5) and (2.8), one gets Remark 2.1. It follows from (2.5) that A is a loxodromic transformation if and only tr2 (A) ∈ [4,∞) [ {4 cos2(cid:16) π if tr2 (A) ∈ (4,∞). The case tr2 (A) ∈ {4 cos2(cid:0) π elliptic transformation A of order n. Finally, that case tr2 (A) = 4 gives a parabolic transformation A. Note that for the elliptic and parabolic transformations, the values of parameter t in (2.7) are the n-th roots of unity. n(cid:1) n ≥ 3} corresponds to an n(cid:17) n ≥ 3}. (2.9) 3. Proof of theorem 1.1 We split the proof in a series of lemmas. Lemma 3.1. Let p1, p2 ∈ CP 1 be two points on the Riemann sphere, such that p1 6= p2. Then there exists a double covering map p : CP 1 → CP 1 (3.1) ramified over the points p1 and p2. Proof. (i) Recall that the necessary condition for the existence of p is given by the Riemann-Hurwitz formula: χ(CP 1) = 2χ(CP 1) − 2 Xi=1 (ei − 1), (3.2) 4 A.GLUBOKOV AND I.NIKOLAEV where χ(CP 1) is the Euler characteristic and ei is the degree of the map z 7→ zei in the ramification point pi. Since χ(CP 1) = e1 = e2 = 2, we conclude that the condition (3.2) is satisfied. (ii) The sufficient condition for the existence of p can be verified directly using [Gersten 1987] [2, Theorem 1.5]. We leave it as an exercise to the reader. Lemma 3.1 is proved. (cid:3) Remark 3.2. We assume further that in lemma 3.1 we have p1 = 0 and p2 = ∞. Our assumption is not restrictive, since any two points p1, p2 ∈ CP 1 can be put into such a position by a Mobius transformation. ρ D(tr2(A)) ∞ D D(λ) ep H/Gλ Figure 1. Covering map ρ : D → H/Gλ. Lemma 3.3. Consider a Riemann surface: (3.3) where D0 is a disk containing point 0 ∈ CP 1. Then there exists a double covering map H := CP 1 \ {D0,∞}, (3.4) Proof. Let p : CP 1 → CP 1 be the double covering map ramified at 0 and ∞, see lemma 3.1 and remark 3.2. Recall that ρ : D → H . D ∼= CP 1 \ {D0, D∞}, (3.5) where D∞ is a disk containing point ∞ ∈ CP 1. We use a homotopy to contract D∞ to the point ∞, and set the map ρ ≡ p. Comparing (3.3) and (3.5), we conclude that ρ is the required double covering map. Lemma 3.3 is proved. Lemma 3.4. H ∼= H/Gλ, where λ ∈ [2,∞) S {2 cos(cid:0) π n(cid:1) n ≥ 3}. (cid:3) SUBFACTORS AND HECKE GROUPS 5 Proof. It is well known, that the Hecke orbifold {H/Gλ λ > 2} is a topological sphere S2 with a hole D(λ) of radius λ/2, one elliptic fixed point e2 of order 2 and one puncture c, see e.g. [Schmidt & Sheingorn 1995] [7, p. 255]. The elliptic point e2 = i is a fixed point of the matrix(cid:0) 0 1 −1 0(cid:1) having order 2 in the group SL2(Z)/±I. Since S2 ∼= CP 1, we use (3.3) to identify D(λ) ≡ D0 and e2 ≡ ∞. Thus one gets H ∼= H/Gλ. Using lemma 3.3, we obtain a double covering map ρ ramified in the points e2 and 0 ∈ D(λ), i.e. ρ : D → H/Gλ. Remark 3.5. Notice that ρ−1(D(λ)) is a disk and ρ−1(e2) is a regular point of D. The corresponding ramification points are shown in Figure 1. 1 To determine admissible values of the moduli parameter λ, observe that λ = π∂D(λ), where ∂D(λ) is the boundary of the disk D(λ). Observe that the local map at the point 0 ∈ D(λ) is given by the formula z 7→ z2. Therefore using the polar coordinates, we conclude that: (3.6) (3.7) 1 λ2 = π (cid:12)(cid:12)ρ−1(∂D(λ))(cid:12)(cid:12) is a moduli parameter of the Riemann surface D. But according to (1.2) any such a parameter must coincide with the tr2 (A), where A is the matrix in the Schottky uniformization (1.1). Taking positive values of the square root, one gets from (1.2) λ = tr (A) ∈ [2,∞) [ {2 cos(cid:16) π n(cid:17) n ≥ 3}. Lemma 3.4 is proved. (3.8) (cid:3) Remark 3.6. For the sake of brevity, lemma 3.4 is proved for the continuous moduli λ ∈ (2,∞). The case of the discrete moduli is treated likewise, see remark 2.1. Theorem 1.1 follows from lemma 3.4. References 1. S. Fomin, M. Shapiro and D. Thurston, Cluster algebras and triangulated surfaces, I. Cluster complexes, Acta Math. 201 (2008), 83-146. 2. S. M. Gersten, On branched covers of the 2-sphere by the 2-sphere, Proc. Amer. Math. Soc. 101 (1987), 761-766. 3. A. Glubokov and I. Nikolaev, Jones Index Theorem revisited, arXiv:1801.05510 4. E. Hecke, Uber die Bestimmung Dirichletscher Reihen durch ihre Funktionalgleichung, Math. Annalen 112 (1936), 664-699. 5. V. F. R. Jones, Subfactors and Knots, CBMS Series 80, AMS, 1991. 6. I. Nikolaev, Noncommutative Geometry, De Gruyter Studies in Mathematics 66, De Gruyter, Berlin, 2017. 7. T. A. Schmidt and M. Sheingorn, On the infinite volume Hecke surfaces, Compositio Math. 95 (1995), 247-262. 1 Department of Mathematics, Ave Maria University, 5050 Ave Maria Blvd., Ave Maria, FL 34142, United States. E-mail address: [email protected] 2 Department of Mathematics and Computer Science, St. John's University, 8000 Utopia Parkway, New York, NY 11439, United States. E-mail address: [email protected]
1506.02308	2	1506	2015-07-14T22:37:23	The classification of simple separable unital locally ASH algebras	[ "math.OA" ]	Let $A$ be a simple separable unital locally approximately subhomogeneous C*-algebra (locally ASH algebra). It is shown that $A\otimes Q$ can be tracially approximated by unital Elliott-Thomsen algebras with trivial $\textrm{K}_1$-group, where $Q$ is the universal UHF algebra. In particular, it follows that $A$ is classifiable by the Elliott invariant if $A$ is Jiang-Su stable.	math.OA	math	THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS GEORGE A. ELLIOTT, GUIHUA GONG, HUAXIN LIN, AND ZHUANG NIU Abstract. Let A be a simple separable unital locally approximately subhomogeneous C- algebra (locally ASH algebra). It is shown that A ⊗ Q can be tracially approximated by unital Elliott-Thomsen algebras with trivial K1-group, where Q is the universal UHF algebra. In par- ticular, it follows that A is classifiable by the Elliott invariant if A is Jiang-Su stable. 1. Introduction Recently, several major steps have been taken in the classification of what might be called "well behaved" separable amenable simple unital C-algebras. The phenomenon of well behavedness itself was explicitly noticed only relatively recently, by Toms and Winter (see [7]) who conjectured that within this class of C-algebras several properties were equivalent, and that the algebras in this robust subclass (and only these) were the ones that could be classified by means of what might be called the naive Elliott invariants -- the ordered K0-group together with the class of the unit, the simplex of tracial states paired naturally with it, and the K1-group. (Other invariants, such as the Cuntz semigroup, might then be helpful for more general classes of amenable C- algebras.) Breakthroughs in the understanding of the robustness of this class were made by Matui and Sato in [20] and [21]. Perhaps the most striking development in the direction of actually proving isomorphism has been the technique -- sometimes referred to as the Winter deformation technique -- introduced by Winter in [33] (with refinements by Lin in [17] and Lin and Niu in [19]), through which a class of (separable, amenable, simple unital) C-algebras which are well behaved in the sense of absorbing tensorially the Jiang-Su algebra Z, and are also known to satisfy the UCT, can be classified in terms of the (naive) Elliott invariant if this is true for the subclass of algebras absorbing the universal Glimm UHF C-algebra Q. Using this, Gong, Lin, and Niu, in [9] -- following on earlier work in this direction (see e.g. [14], [22], [23], [24], [15], [33], [17], [19], [16]) -- have achieved a classification of finite algebras in this well behaved class which is close to being definitive -- it is simple to describe and it exhausts completely the possible values of the invariant. (The complementary class, the infinite algebras in the well behaved class under consideration, were dealt with some time ago by Kirchberg and Phillips -- [11], [12], [25].) Unfortunately, while it is believed that all well behaved finite separable amenable simple unital C-algebras may be ASH algebras (inductive limits of subalgebras of full matrix algebras over commutative C-algebras) -- and, indeed, that the algebra need not be assumed to be well Date: July 16, 2018. 1 THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 2 behaved if in addition matrix algebras over it are also finite in the Murray-von Neumann sense (i.e., if the algebra is stably finite) -- the class considered by Gong, Lin, and Niu does not on the face of it include the class of all well behaved -- "Jiang-Su stable" -- simple unital ASH algebras. Using the recent result of Santiago, Tikuisis, and the present authors ([5]) that any Jiang-Su stable simple unital ASH algebra has finite nuclear dimension (one of the Toms-Winter well behavedness properties -- the important concept of nuclear dimension was introduced by Winter and Zacharias in [34]) -- and also using the notion of non-commutative cell complex introduced in [5] in the proof of this result -- , the present note shows that indeed such an algebra (Jiang-Su stable simple unital ASH) belongs to the class dealt with by Gong, Lin, and Niu. (Even if the ASH algebra has slow dimension growth, so that by [28] and [31] it is indeed Jiang-Su stable, it is not known to belong to the class studied by Gong, Lin, and Niu -- the class of rationally tracially approximately point -- line algebras -- see below.) Acknowledgements. The research of G. A. E. was supported by a Natural Sciences and En- gineering Research Council of Canada (NSERC) Discovery Grant, the research of G. G. was supported by an NSF Grant, the research of H. L. was supported by an NSF Grant, and the research of Z. N. was supported by a Simons Foundation Collaboration Grant. Z. N. also thanks Aaron Tikuisis for his comments on the early version of the paper. 2. Noncommutative cell complexes Definition 2.1 (Definition 2.1 of [5]). The class of noncommutative cell complexes (NCCC) is the smallest class C of C-algebras such that (1) every finite dimensional algebra is in C; and (2) if B ∈ C, n ∈ N, ϕ : B → Mk(C(Sn−1)) is a unital homomorphism, and A is given by the pullback diagram A Mk(C(Dn)) f →f Sn−1 B ϕ / Mk(C(Sn−1)), then A ∈ C. The reason we consider NCCCs is as follows: Theorem 2.2 (Theorem 2.15 of [5]). Let A be a unital subhomogeneous C-algebra. Then A can be locally approximated by sub-C-algebras which are NCCCs. Definition 2.3. Let A be an NCCC, and fix an NC cell complex decomposition of A with length l (in the sense that A is built from a finite dimensional C-algebra A0 by attaching l noncommutative cells). Assume that A0 = Ms1(C) ⊕ · · · ⊕ Msr (C) for some natural number r, and denote the attached cell at the i-th step by Mki(C(Dni)). Consider the spaces {pt}, ..., {pt} , Dn1, ..., Dnl, } r {z / / / THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 3 and denote them by X1, X2, ..., Xm, where m = r + l. Denote the matrix orders of the corre- sponding C-algebras by Then there is a standard embedding d1, ..., dr, dr+1, ..., dm. Π : A → mMi=1 Mdi(C(Xi)). Denote by Πi : A → Mdi(C(Xi)), i = 1, ..., m, the projection of Π onto the i-th direct summand. Lemma 2.4. Let A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)) be an NCCC, and let τ ∈ T(A). Then there is a decomposition τ (fB, fD) = ατB(fB) + βZDn\Sn−1 tr(fD(x))dµ(x), (fB, fD) ∈ A, where τD ∈ T(B), µ is a probability measure on Dn \ Sn−1, tr is the standard trace of Mk(C), and α, β ∈ [0, 1] with α + β = 1. Moreover, α and β are unique and τB (or µ) is unique if α (or β) is not zero. Proof. The uniqueness part of the lemma is clear. Let us show the existence part. Consider the restriction of τ to I := Mk(C0(Dn \ Sn−1)) ⊆ A. Then it is a trace with norm at most one, and thus there is β ∈ [0, 1] and a probability measure µ on Dn \ Sn−1 such that χη : [0, 1] ∋ x 7→  x ∈ [1 − η/2, 1], 1, linear, x ∈ [1 − η, 1 − η/2], 0, x ∈ [0, 1 − η]; and Then a direct calculation shows that (2.1) τ (f, g) = τ (f, gη), (f, g) ∈ A, η ∈ (0, 1). gη(x) = g(x)χη(kxk). It is clear that τ is self-adjoint; let us show that it is positive. Let (f, g) ∈ A be positive. Define If δ = 0, let us show that τ (f, g) = 0. Indeed, in this case, one has δ = inf{τ (f, gη) : η ∈ (0, 1)}. τ ((f, g)) = τ (f, gη) + τ (0, g − gη) Define a linear map τ : A → C by τ ((0, g)) = βZDn\Sn−1 τ ((f, g)) = τ (f, g) − βZDn\Sn−1 For each g ∈ Mk(C(Dn)) and any η ∈ (0, 1), define tr(g(x))dµ(x), g ∈ Mk(C0(Dn \ Sn−1)). tr(g(x))dµ(x), (f, g) ∈ A. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 4 and hence (2.2) τ ((f, g)) = sup{τ (0, g − gη) : η ∈ (0, 1)} = sup{βZDn\Sn−1 trx(g − gη)dµ(x) : η ∈ (0, 1)}. Note that for any η ∈ (0, 1). τ ((f, g)) = τ (f, g) − βZDn\Sn−1 = τ (f, g) − βZDn\Sn−1 +βZDn\Sn−1 tr(gη(x))dµ(x), tr(g(x))dµ(x) trx(g − gη)dµ(x) and since µ is a probability measure, the integral βRDn\Sn−1 tr(gη(x))dµ(x) is arbitrarily small if η is small enough. By (2.2), one has that τ ((f, g)) = 0. If δ > 0, since µ is a probability measure, there is η ∈ (0, 1) such that βZDn\Sn−1 tr(gη(x))dµ(x) < δ/2, and therefore τ ((f, g)) = τ ((f, gη)) = τ (f, gη) − βZDn\Sn−1 tr(gη(x))dµ(x) ≥ −δ/2 = δ/2 > 0. Therefore, one always has τ ((f, g)) ≥ 0, and so τ is a positive linear functional. Therefore τ is a (positive) trace of A. Note that τ (I) = 0, and therefore τ factors through A/I ∼= B, and hence in fact is a trace of B. Therefore, there are α ∈ [0, 1] and τB ∈ T(B) such that τ (f, g) − βZDn\Sn−1 tr(g(x))dµ(x) = τ (f, g) = ατB(f ), (a, b) ∈ A, as desired. (cid:3) Corollary 2.5. Let A be an NCCC with a given decomposition with length l. Then any trace τ of A has a decomposition τ = α1τ1 + α1µ1 + · · · + αmµm, where m = rank(K0(A0)) + l, µi is a probability measure on Dni \ Sni−1 if Xi = Dni, and µi is the Dirac measure if Xi consists of a point, αi ∈ [0, 1] and α1 + α2 + · · · + αm = 1. Moreover, the coefficients αi are unique. Definition 2.6. Let A be an NCCC with a given decomposition, and let τ ∈ T(A). Referring to Corollary 2.5, define αi(τ ) = αi. Lemma 2.7. Let A be a noncommutative cell complex (NCCC). Then the K-groups of A are finitely generated (as abelian groups). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 5 Proof. The statement is true if A is finite dimensional. Assume the statement is true for non- commutative complexes with length at most l. Let A be a noncommutative complex with length l + 1. Write where B is a noncommutative complex with length l. Then there is a short exact sequence A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)), 0 / Mk(C0(Rn)) / A / B / 0, and the corresponding six-term exact sequence is K0(C0(Rn)) / K0(A) / K0(B) If n is odd, one has and K1(B) K1(A) K1(C0(Rn)). 0 / K0(A) / K0(B) / · · · , 0 / Z/mZ / K1(A) / K1(B) / 0, for some positive integer m. By the inductive hypothesis, the groups K0(B) and K1(B) are finitely generated, and therefore the groups K0(A) and K1(A) are finitely generated. If n is even, a similar argument shows that K0(A) and K1(A) are finitely generated. Therefore, the K-groups of A are always finitely generated. Hence by induction, the statement holds for all noncommutative cell complexes. (cid:3) In general, the positive cone of K0(A) might not be finitely generated; for instance, the positive cone of K0(C(S2)) is which is not finitely generated. (On the other hand, consider the image {(m, n) ∈ Z2 : m > 0} ∪ {(0, 0)}, G := ρ(K0(A)) ⊆ Aff(T(A)), with respect to the canonical map ρ, with the induced order from Aff(T(A)) (i.e., an element g ∈ G is positive if and only if g is positive in Aff(T(A))), is isomorphic to Z and so the positive cone of G is finitely generated.) The following lemma was stated and proved in [9] for the K0-group of an Elliott-Thomsen algebra. The argument in fact shows the following (for the reader's convenience, we include the proof). (In fact the ordered groups arising are the same.) Lemma 2.8 (Theorem 3.14 of [9]). Consider (Zl, (Zl)+) with the standard (direct sum) order. Let G be a subgroup of Zm, and put G+ = G ∩ (Zm)+. Then the positive cone G+ is finitely generated (as a semigroup). / / / / / / O O o o o o / / / / / / / THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 6 Proof. Let us first show that G+ \ {0} has only finitely many minimal elements. Suppose, otherwise, that {qn} is an infinite set of minimal elements of G+ \ {0}. Write qn = (m(1, n), m(2, n), ..., m(j, n)) ∈ Zm + , where m(i, n) are positive integers (including zero), i = 1, 2, ..., m and n = 1, 2, .... If there is an integer M ≥ 1 such that m(i, n) ≤ M for all i and n, then {qn} is a finite set. So we may assume that {m(i, n)} is unbounded for some 1 ≤ i ≤ m. Passing to a subsequence of {nk} such that limk→∞ m(i, nk) = +∞, we may assume that limn→∞ m(i, n) = +∞. We may assume that, for some j, {m(j, n)} is bounded. Otherwise, by passing to a subsequence, we may assume that limn→∞ m(i, n) = +∞ for all i ∈ {1, 2, ..., m}. Therefore limn→∞ m(i, n) − m(i, 1) = +∞. It follows that, for some n ≥ 1, m(i, n) > m(i, 1) for all i ∈ {1, 2, ..., m}. Therefore, qn > q1, which contradicts the fact that qn is minimal. By passing to a subsequence, we may write {1, 2, ..., m} = N ⊔ B such that limn→∞ m(i, n) = +∞ if i ∈ N and {m(i, n)} is bounded if i ∈ B. Therefore, {m(j, n)} has only finitely many different values if j ∈ B. Thus, by passing to a subsequence again, we may assume that m(j, n) = m(j, 1) if j ∈ B. Therefore, for some n > 1, m(i, n) > m(i, 1) for all n if i ∈ N, and m(j, n) = m(j, 1) for all n if j ∈ B. It follows that qn ≥ q1. This is impossible, since qn is minimal. This shows that G+ has only finitely many minimal elements. To show that G+ is generated by these minimal elements, fix an element q ∈ G+ \ {0}. It suffices to show that q is a finite sum of minimal elements in G+. If q is not minimal, consider the set of all elements in G+ \ {0} which are (strictly) smaller than q. This set is finite. Choose one which is minimal among them, say p1. Then p1 is minimal element in G+ \ {0}, as otherwise there is one smaller than p1. Since q is not minimal, q 6= p1. Consider q − p1 ∈ G+ \ {0}. If q − p1 is minimal, then q = p1 + (q − p1). Otherwise, we repeat the same argument to obtain a minimal element p2 ≤ q − p1. If q − p1 − p2 is minimal, then q = p1 + p2 + (q − p1 − p2). Otherwise we repeat the same argument. This process is finite. Therefore q is a finite sum of minimal elements in G+ \ {0}. (cid:3) With Lemma 2.8, one has Lemma 2.9. Let A be an NCCC. Then the ordered group (ρA(K0(A)), ρA(K0(A)) ∩ Aff +(T(A))) is finitely generated (as an ordered group). (In other words, the positive cone is finitely generated as a semigroup.) Proof. With the fixed NC cell complex decomposition of A, consider the standard embedding Π : A → mMi=1 Mdi(C(Xi)). Define ρ : K0(A) ∋ [p] 7→ (rank(Π1(p)), ..., rank(Πm(p))) ∈ Zm. Clearly, the map ρ is positive. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 7 Define G = ρ(K0(A)) and G+ = ρ(K0(A)) ∩ (Zm)+. It follows from Lemma 2.8 that the cone (G, G+) is a finitely generated ordered group. In order to prove the lemma, one only has to show that there is an isomorphism (G, G+) ∼= (ρA(K0(A)), ρA(K0(A)) ∩ Aff +(T(A))). Define θ : ρ(K0(A)) ∋ (x1, x2, ..., xm) 7→ (τ 7→ mXi=1 xi di αi(τ )) ∈ Aff(T(A)). Then θ is injective. Note that for any τ ∈ T(A) with the decomposition one has and therefore τ = α1µ0 + α2µ2 + · · · + αmµm, ρA(p)(τ ) = = = αiZXi αiZXi tri(Πi(p)(x))dµi(x) rank(Πi(p)) di dµi,j αi rank(Πi(p)) di = θ(ρ(p)), mXi=1 mXi=1 mXi=1 θ(ρ(K0(A))) = θ(G) = ρA(K0(A)). then the affine map τ 7→ Pm It is clear that θ(G+) ⊆ ρA(K0(A)) ∩ Aff +(T(A)). Moreover, if θ(x1, ..., xm) ∈ Aff +(T(A)), αi(τ ) is positive, and hence each xi must be positive; that is, θ induces an order isomorphism between (G, G+) and ρA(K0(A)), ρ(K0(A)) ∩ Aff +(T(A)), as desired. (cid:3) xi di i=1 Lemma 2.10. Let A be an NCCC. Then, for any finite set F ⊆ Aff(T(A)) and any ε > 0, there are positive continuous affine maps θ1 : Aff(T(A)) → Rs and θ2 : Rs → Aff(T(A)) for some s ∈ N such that kθ2 ◦ θ1(f ) − f k∞ < ε, f ∈ F . Proof. The statement clearly holds for A a finite dimensional C-algebra. Assume that A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)), and suppose that the statement holds for B. Let (F , ε) be given. Note that Aff(T(A)) is the pullback of Aff(T(B)) and CR(Dn) in the same manner as A. For each f ∈ F , write it as f = (fB, fD), where fB ∈ Aff(T(B)) and fD ∈ CR(Dn). Denote by FB the set of fB's. Since B satisfies the statement, there are continuous positive affine maps θB,1 : Aff(T(B)) → RsB and θB,2 : RSB → Aff(T(B)) such that (2.3) kθB,2 ◦ θB,1(fB) − fBk∞ < ε/3, f ∈ F . THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 8 Choose γ ∈ (0, 1) such that provided dist(x, y) < γ. kfD(x) − fD(y)k < ε/3, f ∈ F , Define gγ : [0, 1] ∋ x 7→  x ∈ [0, 1 − γ], 1, linear, x ∈ [1 − γ, 1 − γ/2], 0, x ∈ [1 − γ/2, 1]; and consider the linear map χγ : C(Sn−1) ∋ f 7→ (x 7→ (1 − gγ(kxk)) · f ( x kxk )) ∈ C(Dn). Then χγ is a positive affine map from Aff(T(Mk(C(Sn−1)))) to Aff(T(Mk(C(Dn)))). Note that (2.4) kfDgγ − (fD − χγ(fDSn−1))k∞ = sup{(1 − gγ(kxk))(fD(x) − fD( x kxk )) : x ∈ Dn} < ε/3 for any f ∈ F . Pick an open cover U of Dn such that the variation of the function fDgγ on any open subset in U is at most ε/3 for any f ∈ F . Moreover, one requires that the diameter of each open set in U should be at most γ/4. Choose {φU : U ∈ U} to be a partition of unity subordinated to U, and fix xU ∈ U for each U ∈ U. Put Uγ = {Ui : U ∩ Sn−1 = Ø}, and write Uγ = {U1, U2, ..., UUγ }. Since the diameter of each Ui ∈ U is at most γ/4, one has that if U /∈ Uγ, then gγ(xU ) = 0, and hence (2.5) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) fDgγ − XU ∈Uγ Define and (fDgγ)(xU )φU(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) fDgγ − XU ∈Uγ fDgγ −XU ∈U (fDgγ)(xU )φU − XU /∈Uγ (fDgγ)(xU )φU(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ ε/3. (fDgγ)(xU )φU(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) θ1 : Aff(T(A)) ∋ (f, g) 7→ θB,1(f ) ⊕ ((ggγ)(xUi)) ∈ RsB ⊕ RUγ , θ2 : RsB ⊕ RUγ ∋ ((ξ1, ..., ξsB), (η1, ..., ηUγ )) 7→ (θB,2((ξ1, ..., ξsB)), χγ(ϕ(θB,2((ξ1, ..., ξsB)))) + Uγ Xi=1 ηiφUi) ∈ Aff(T(A)). Then, a straightforward calculation shows that θ2 ◦ θ1((fB, fD)) = (θB,2(θB,1(fB)), χγ(ϕ(θB,2(θB,1(fB)))) + Uγ Xi=1 (fDgγ)(xUi)φUi) THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 9 By the inductive assumption, one has that for any f ∈ F , kfB − θB,2(θB,1(fB))k < ε/3 < ε, and Therefore as desired. fD − (χγ(fDSn−1) + fD − (χγ(ϕ(θB,2(θB,1(fB)))) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Uγ Xi=1 (fDgγ)(xUi)φUi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ < (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) UγXi=1 (fDgγ)(xUi)φUi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Uγ Xi=1 ≤ (cid:13)(cid:13)(cid:13)gγfD −X(fDgγ)(xp)φp(cid:13)(cid:13)(cid:13)∞ (fD − χγ(fDSn−1) − (fDgγ)(xUi)φUi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ + ε/3 (by (2.3)) + ε/3 ≤ ε (by (2.5)). + 2ε/3 (by (2.4)) kθ2 ◦ θ1(f ) − f k < ε, f ∈ F , (cid:3) Remark 2.11. In fact, as shown in [4] (Lemma 2.6), the statement of Lemma 2.10 always holds if T(A) is replaced by an arbitrary compact metrizable Choquet simplex. 3. An existence theorem Let A be an NCCC. The main result in this section (Theorem 3.7) is that any almost compatible pair (κ, γ) from an NCCC to Q can be lifted to an algebra homomorphism, where κ : K0(A) → Q and γ : Aff(T(A)) → Aff(T(Q)) ∼= R. Lemma 3.1. Let P be an m × n matrix with integer entries, and let ξ ∈ Rm with each entry a positive number (including zero). Assume that each entry of P ξ is rational. Then, for any ε > 0, there is ζ ∈ Qm with positive entries (including zero) such that (1) P ξ = P ζ, and (2) kξ − ζk∞ < ε. Proof. By deleting the columns of P corresponding to the 0's of ξ, one may assume that each entry of ξ is strictly positive. Let us show that ker P = ker P ∩ Qm. It is clear that ker P ⊇ Qm ∩ ker P . Note that P is rational, so that one can choose a basis of ker P (as a real vector space) consisting of rational vectors, from which it follows that dimR(ker P ) ≤ dimQ(ker P ∩ Qm), THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 10 and hence dimR(ker P ) ≤ dimR(ker P ∩ Qm). This forces ker P = Qm ∩ ker P . Again by the rationality of P , there is a vector η ∈ Qm such that and hence P η = P ξ, P −1({P ξ}) = P −1({P η}) = η + ker P = η + Qm ∩ ker P = η + Qm ∩ ker P . It is clear that ξ ∈ η + Qm ∩ ker P . On noting that any entry of ξ is strictly positive, and all vectors in η + Qm ∩ ker P are rational, for any ε > 0, it follows that there is ζ ∈ η + Qm ∩ ker P such that kζ − ξk∞ < ε and each entry of ζ is positive. (cid:3) Lemma 3.2. Let A be a unital subhomogeneous C-algebra such that Primd(A) has finitely many connected components for each d. Let (F , ε) be given. Then, for any compatible pair (κ, γ) satisfying, where κ : K0(A) → Q = K0(Q) and γ : Aff(T(A)) → R = Aff(T(A)), there is a homomorphism φ : A → Q such that (1) [φ]0 = κ, and (2) γ( f )(tr) − tr(φ(f )) < ε, f ∈ F , where tr is the canonical trace of Q. Moreover, φ can be chosen to have finite dimensional range. Proof. Without loss of generality, one may assume that F is inside the unit ball. Since Q has unique trace, it is enough to show that for any κ : K0(A) → Q and τ ∈ T(A) with there is a homomorphism φ : A → Q such that [φ]0 = κ and τ (p) = κ(p), p ∈ K0(A), τ (f ) − tr(φ(f )) < ε, f ∈ F . By Corollary 2.5, there is a probability measure µi on each Xi, and αi ∈ [0, 1] with αi = 1 mXi=1 αiZXi such that τ (a) = mXi=1 where tri is the canonical trace of Mki(C). tri(Πi(a))dµi, a ∈ A, THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 11 Note that for each projection p ∈ A ⊗ K, one has (3.1) τ (p) = = mXi=1 mXi=1 αiZXi rank(Πi(p)) di dµi αi rank(Πi(p)) di . Since K0(A) is finitely generated, there are p1, p2, ..., pr ∈ K0(A) which generate K0(A). Con- sider the matrix P =   rank(Π1(p1)) rank(Π1(p2)) ... rank(Π2(p1)) rank(Π2(p2)) ... rank(Π1(pr)) rank(Π2(pr)) d1 d2 · · · · · · ... · · · · · · rank(Πm(p1)) rank(Πm(p2)) ... rank(Πm(pr)) dm   , and the vector By (3.1), one has ξ = ( α1 d1 , α2 d2 , ..., αm dm )T. P ξ = (τ (p1), τ (p2), ..., τ (pr), 1)T = (κ(p1), κ(p2), ..., κ(pr), 1)T ∈ Qr+1. Then, by Lemma 3.1, there is a positive rational vector ξ ∈ Qm such that (3.2) Write P ξ = P ζ and kξ − ηk∞ < ε/2l(k1 + · · · + kl). ζ = ( β1 d1 , β2 d2 , ..., βm dm )T. By (3.2), one has that (3.3) αi − βi < ε/2m, 1 ≤ i ≤ m. Since P ξ = P ζ, one has mXi=1 βi rank(πx(pj)) di = mXi=1 αi rank(πx(pj)) di , j = 1, ..., r. Since p1, p2, ..., pl generate K0(A), one has (3.4) mXi=1 βi rank(πx(p)) di = mXi=1 αi rank(πx(p)) di , p ∈ K0(A). It also follows from P ξ = P ζ that β1 + · · · + βm = 1. Consider τ ′ ∈ T(A) defined by τ ′(a) = βiZXi mXi=1 tri(Πi(a))dµi, a ∈ A. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 12 Then, by (3.4) and (3.3), one has and τ ′(p) = τ (p), p ∈ K0(A), τ (f ) − τ ′(f ) < ε/2, f ∈ F . Consider each probability measure µi, and choose a discrete measure µi such that Write tri(Πi(f ))dµi −ZXi (cid:12)(cid:12)(cid:12)(cid:12) ZXi tri(Πi(f ))dµi(cid:12)(cid:12)(cid:12)(cid:12) < ε/2, liXk=1 δxik, 1 li µi = f ∈ F . for some xik ∈ Xi, where δxik is the Dirac measure concentrated at xik. Without loss of generality, one may assume that all li are the same and equal to l for some l. Define One then has and τ (a) = βiZXi mXi=1 tri(Πi(a))dµi, a ∈ A. τ (p) = τ (p), p ∈ K0(A), Write βi = qi/q for natural numbers qi ≤ q. Since P βi = 1, one has that τ (f ) − τ (f ) < ε, f ∈ F . X qi = q. Define a unital homomorphism by φ : A ∋ a 7→ mMi=1 lMk=1 where N = d1 · · · dmlq. Then (πxik (a) ⊕ · · · ⊕ πxik(a) ) ∈ MN (C), {z d1···di−1di+1···dmqi } i=1Pl tr(φ(a)) = Pm mXi=1 mXi=1 qi q 1 l = = lXk=1 k=1(d1 · · · di−1di+1 · · · dmqi)Tr(πxik(a)) d1 · · · dnql d1 · · · di−1di+1 · · · dmdi d1 · · · dnq qi 1 l lXk=1 tri(πxik(a)) ( tri(πxik(a))) = τ (a). Let ι : MN (C) → Q be an unital embedding. Then the homomorphism ι◦φ satisfies the condition of the lemma. (cid:3) THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 13 Lemma 3.3. Let A be a NCCC. For any finite set F ⊆ A and any ε > 0, there is a finite set H ⊆ A+ such that for any σ > 0, there is δ > 0 such that if τ ∈ T(A) satisfies τ (h) > σ, h ∈ H, there exists τ ′ ∈ T(A) such that (1) τ ′(p) = τ (p), p ∈ K0(A), (2) τ ′(f ) − τ (f ) < ε, f ∈ F , and (3) αi(τ ′) > δ, all i. Moreover, one may require that δ depends only on σ. Proof. If A = Mn1(C) ⊕ · · · ⊕ Mnk(C). Then H = {1Mni (C) : 1 ≤ i ≤ k} satisfies the lemma with δ = σ for any given σ (in fact, one has that τ ′ = τ in this case). Let A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)), where B is NCCC, and assume that the lemma holds for B. Let F ⊆ A, ε > 0 be given. For each f ∈ A, write f = (fB, fD) for fB ∈ B and fD ∈ Mk(C(Dn)). Since B is assumed to satisfy the lemma, there is HB such that for any σ, there is δB(σ) > 0 such that if τB ∈ T(B) satisfies there is τ ′ B ∈ T(A) such that and Choose γ > 0 such that τB(h) > σ, h ∈ HB, τ ′ B(p) = τB(p), p ∈ K0(B), τ ′ B(fB) − τB(fB) < ε/2, f ∈ F , αi(τ ′ B) > δB(σ), all i. kfD(x) − fD(y)k < ε/2, f ∈ F , if dist(x, y) < γ. Define gγ : [0, 1] ∋ x 7→  x ∈ [1 − γ/2, 1], 1, linear, x ∈ [1 − γ, 1 − γ/2], 0, x ∈ [0, 1 − γ/2]; similarly, define gγ/2 and g2γ. For each h ∈ HB define hD ∈ Mk(C(Dn)) by hD(x) =  0, 4kxk+2γ−4 γ kxk ∈ [0, 1 − γ/2], ϕ(h)( x kxk), kxk ∈ [1 − γ/2, 1 − γ/4], ϕ(h)( x kxk), kxk ∈ [1 − γ/4, 1], where ϕ : B → Mk(C(Sn−1)) is the gluing map. For each h ∈ HB, define h = (h, hD) ∈ A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 14 Also define g2γ = (0, g2γ1Mk(C)), gγ/2 = (0, gγ/21Mk(C)) ∈ A. Then H = {h, g2γ, 1 − gγ/2 : h ∈ HB} satisfies the condition of the lemma. Indeed, for any σ > 0, set Let τ ∈ T(A) be such that δ = σδB(σ). τ (h) > σ, h ∈ H. Without loss of generality, one may assume that τ ((f, g)) = ατB(f ) + βZDn\Sn−1 gdµ, where µ is a discrete probability measure of Dn \ Sn−1. Since τ (gγ) > σ, one has β > σ. Write µ =Pi βiδxi. For each xi with kxik ≥ γ Then define the trace τB,xi(f ) = 2 , define τB,xi ∈ T(B) by 1 k Tr(ϕ(f )( xi kxik )). τ ((f, g)) = ατB(f ) + β( Xkxik≥ γ = α′τB(f ) + β′Z(Dn)◦ 2 gdµ′, βiτB,xi(f )) + β( Xkxik< γ 2 βi)Z(Dn)◦ gd( 1 Pkxik< γ 2 βi Xkxik< γ 2 βiδxi) τB(f ) = 2 ατB(f ) + β(Pkxik≥ γ α +Pkxik≥ γ βi Xkxik< γ Pkxik< γ 1 2 2 2 µ′ = βiτB,xi(f )) ββi βiδxi, ββi, , α′ = α + Xkxik≥ γ β′ = β( Xkxik< γ βi). 2 2 where and Note that (3.5) (3.6) Noting that one has By the choice of γ, one has τ (p) = τ (p), p ∈ K0(A). τ (f ) − τ (f ) < ε/2, f ∈ F . τ (g2γ) = τ (g2γ) and τ (1 − gγ/2) ≥ τ (1 − gγ/2), τ (g2γ) > σ and τ (1 − gγ/2) > σ. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 15 A straightforward calculation shows that β′ ≥ τ (g2γ) and α′ = τ (1 − gγ/2), and therefore (3.7) β′ > σ and α′ > σ. Also noting that for any h ∈ HB, one has and Therefore, τ (h) ≥ τ (h) > σ, τ (h) = α′τB(h). By the inductive assumption, there is τ ′ B ∈ T(B) such that τB(h) > σ/α′ > σ, h ∈ HB. (3.8) (3.9) and (3.10) Then the trace τ ′ B(p) = τB(p), p ∈ K0(B), τ ′ B(fB) − τB(fB) < ε/2, f ∈ F , αi(τ ′ B) > δB(σ), all i. τ ′(f, g) = α′τ ′ B(f ) + β′Z(Dn)◦ gdµ′ satisfies the condition of the lemma. By (3.5) and (3.8), one has τ ′(p) = τ (p), p ∈ K0(A), Indeed, by (3.6) and (3.9), one has (3.11) τ ′(f ) − τ (f ) < ε, f ∈ F . By (3.7) and (3.10), one has as desired. αi(τ ′) > σδB(σ), all i, (cid:3) Lemma 3.4. Let P be a m × n matrix. Assume that m ≥ n and rank(P ) = n. Let σ > 0 be given. Then, for any ε > 0, there is δ > 0 such that for any vectors ξ ∈ Rm and κ ∈ Rn with (1) ξi > σ, i = 1, ..., m, and (2) kκ − P ξk∞ < δ, there is ζ ∈ Rm with ζi ≥ 0, i = 1, ..., m such that (3) P ζ = κ, and (4) kξ − ζk∞ < ε. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 16 Proof. Regard P as a linear map from Rm → Rn. Since rank(P ) = n, one has that P is surjective. Therefore, for the given ǫ < σ, there is δ > 0 such that (3.12) Bk·k∞(0, δ) ⊆ P (Bk·k∞(0, ε)). Then the δ satisfies the condition of the lemma. Indeed, let ξ ∈ Rm and κ ∈ Rn be given, and satisfy Then kκ − P ξk∞ < δ. κ − P (ξ) ∈ Bk·k∞(0, δ) ⊆ Rn. By (3.12), there is θ ∈ Bk·k∞(0, ε) ⊆ Rm such that and hence P (θ) = κ − P (ξ), P (ξ + θ) = κ. Since each entries of ξ is at least σ > ε, one has that and therefore is the desired vector. ξ + θ ∈ (R+)m, ζ = ξ + θ (cid:3) Remark 3.5. Let A be an NCCC, and let κ : K0(A) → Q be a positive map. Since A is of type I, it is exact; therefore, κ (regarded as a state of K0(A)) is induced by a trace. That is, there is τ ∈ T(A) such that In particular, this implies that κ factors through ρA(K0(A)). κ(p) = τ (p), p ∈ K0(A). Lemma 3.6. Let A be an NCCC. Let (F , ε) be given. Let p1, p2, ..., pn ∈ K0(A) be such that {p1, ..., pn} is a set of generators for ρA(K0(A)) (as an abelian group) (still use the same notation for the image of pi). Then, there is a finite set H ⊆ A+ such that for any σ > 0, there is δ > 0 such that if κ : K0(A) → Q and τ ∈ T(A) satisfy (1) τ (h) > σ, h ∈ H, and (2) τ (pi) − κ(pi) < δ, i = 1, ..., n, then there is τ ′ ∈ T(A) such that (3) τ ′(f ) − τ (f ) < ε, f ∈ F , and (4) τ ′(pi) = κ(pi), i = 1, ..., n. Proof. Without loss of generality, one may assume that F is contained in the unit ball of A. One may also assume that p1, p1, ..., pn ∈ ρA(K0(A)) are Q-independent. Let H ⊆ A+ be the subset of Lemma 3.3 with respect to A, F , and ε/2. Then H satisfies the lemma. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 17 In fact, for any σ > 0, let δ′ be the constant of Lemma 3.3 with respect to σ. Consider the m × n matrix P =   rank(Π1(p1)) rank(Π2(p1)) d1 rank(Π1(p2)) d2 rank(Π2(p2)) d1 ... d2 ... rank(Π1(pn)) rank(Π2(pn)) d1 d2 · · · · · · ... · · · rank(Πm(p1)) dm rank(Πm(p2)) dm ... rank(Πm(pn)) dm   . Since p1, p1, ..., pn ∈ ρA(K0(A)) are Q-independent, the matrix P has rank n and it satisfies the assumptions of Lemma 3.4. Applying Lemma 3.4 to δ′ (in the place of σ) and ε/2m, one obtains δ. Let (κ, τ ) be a pair satisfies and (3.13) τ (pi) − κ(pi) < δ, i = 1, ..., n, τ (h) > σ, h ∈ H. By (3.13) and Lemma 3.3, there is τ ′ ∈ T(A) such that (3.14) and Set τ ′(pi) = τ (pi), τ ′(f ) − τ (f ) < ε/2, i = 1, ..., n, f ∈ F , αi(τ ′) > δ′, ∀i. ξ = (α1(τ ′), α2(τ ′), ..., αm(τ ′))T and κ = (κ(p1), κ(p2), ..., κ(pn))T, and then one has kP ξ − κk∞ = max{τ ′(pi) − κ(pi) : i = 1, ..., n} = max{τ (pi) − κ(pi) : i = 1, ..., n} (by (3.14)) < δ By Lemma 3.4, there is a positive vector ζ = (β1, ..., βm) such that (3.15) and Consider the trace It is clear that By (3.15), one has P ζ = (κ(p1), κ(p2), ..., κ(pn))T αi − βi < ε/2m, ∀i. τ ′′ =X βiµi. τ ′′(f ) − τ (f ) < ε, f ∈ F . τ ′′(p) = κ(p), p ∈ K0(A). Moreover, the trace τ ′′ is indeed a state since τ ′′(1A) = κ([1A]) = 1, as desired. (cid:3) THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 18 Theorem 3.7. Let A be a unital NCCC, and let σ > 0. Let (F , ε) be given. Let p1, p2, ..., pn ∈ K0(A) be such that the set {p1, ..., pn} generates the group ρA(K0(A)) (let us still use the same notation for the image of pi). Then, there is a finite set H ⊆ A+ such that for any σ > 0, there is δ > 0 such that if κ : K0(A) → Q and τ ∈ T(A) satisfy (1) τ (pi) − κ[pi] < δ, i = 1, ..., n, and (2) τ (h) > σ, h ∈ H, then there is a homomorphism φ : A → Q such that (3) [φ]0 = κ, and (4) τ (f ) − tr(φ(f )) < ε, f ∈ F , where tr is the canonical trace of Q. Moreover, φ can be chosen to have finite dimensional range. Proof. Let H ⊆ A+ denote the finite set of Lemma 3.6 with respect to the data A, F , ε/2. Then H satisfies the theorem. Indeed, given σ > 0, consider the constant δ of Lemma 3.6 with respect to σ. Let (κ, τ ) be a pair as above -- δ-compatible on pi, i = 1, ..., n, and such that By Lemma 3.6, there is τ ′ ∈ T(A) such that the pair (κ, τ ′) is exactly compatible on the pi and τ (h) > σ, h ∈ H. Then, by Lemma 3.2, there is a homomorphism φ : A → Q such that τ ′(f ) − τ (f ) < ε/2, f ∈ F . and [φ]0 = κ tr(φ(f )) − τ ′(f ) < ε/2, f ∈ F , and therefore satisfying the condition of the theorem. (cid:3) 4. A decomposition theorem and a uniqueness theorem Recall (see, for instance, [10]) Lemma 4.1 (The Marriage Lemma). Let A and B be two finite subsets of a metric space. Suppose that for any set X ⊆ A, one has #{y ∈ B : dist(y, X) < ε} ≥ #X, then there is a set B′ ⊆ B and a one-to-one pairing between A and B′ such that the distance between the points in each pair is at most ε. The following statement is a generalization of the Marriage Lemma due to Gong in a unpub- lished manuscript: THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 19 Lemma 4.2. Let A, B be two finite subsets of a metric space with subsets A1 ⊂ A and B1 ⊂ B. Suppose that for each X ⊂ A1, one has and for each Y ⊂ B1, one has #{y ∈ B : dist(y, X) < ε} ≥ #X #{x ∈ A : dist(x, Y ) < ε} ≥ #Y. Then there are sets A2 ⊆ A and B2 ⊆ B such that A1 ⊆ A2 and B1 ⊆ B2, and the elements of A2 and B2 can be (bijectively) paired to within ε. Proof. The proof is similar to that of [10]. We spell out the details for reader's convenience. In the case #(B1) = 0, i.e., B1 is empty, then this is the case of the classical Marriage Lemma (Lemma 4.1) in the case of A1 and B. (For this case, X0 can be chosen to be A1.) We define partial order for (m, n) ∈ Z+ × Z+ (where Z+ = {0, 1, 2, ...} is the set of positive integers) by (m, n) ≤ (m1, n1) if m ≤ m1 and n ≤ n1. We denote (m, n) < (m1, n1) if (m, n) ≤ (m1, n1) and (m, n) 6= (m1, n1). We will prove the lemma by induction on #(A1) and #(B1). That is, we assume that if the result is true for the case (cid:0)#(A1), #(B1)(cid:1) < (m, n) and prove the lemma to be true for the case #(A1) = m and #(B1) = n. The rest of the proof divides into two cases. Case 1. For any nonempty set X ⊆ A1, #{y ∈ B; dist(y, X) < ε} ≥ #(X) + 1 and for any nonempty set Y ⊆ B1, #{x ∈ A; dist(x, Y ) < ε} ≥ #(Y ) + 1. Choose any a ∈ A1, there is a b ∈ B such that dist(a, b) < ε. We will pair a ∈ A1 with b ∈ B. Then let A = A\{a} with A1 = A1\{a}, and B = B \{b} with B1 = B1 if b 6∈ B1 or B1 = B1\{b} if b ∈ B1. It is easy to verify that A ⊇ A1 and B ⊇ B1 satisfy the condition of the lemma with (cid:0)#( A1), #( B1)(cid:1) = either (m − 1, n) or (m − 1, n − 1). That is, (cid:0)#(A1), #(B1)(cid:1) < (m, n). By the induction assumption, there is a subset A2 ⊇ A1 and B2 ⊇ B1 such that X0 can be paired one by one within ε. Then the sets A2 = A2 ∪ {a} and B2 = B2 ∪ {b} satisfy the lemma. Case 2. The conditions of Case 1 do not hold. Then either there is X ⊆ A1 or Y ⊆ B1 does not satisfy the condition in Case 1. That is, either there is X ⊆ A1 such that #{y ∈ B; dist(y, X) < ε} = #(X), or there is Y ⊆ B1 such that #{x ∈ A; dist(x, Y ) < ε} = #(Y ). Without loss of generality, let us assume there is Y1 ⊆ B1 such that #{x ∈ A; dist(x, Y1) < ε} = #(Y1). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 20 Let X1 = {x ∈ A; dist(x, Y1) < ε}. Then for any subset Z ⊆ Y1, {x ∈ A; dist(x, Z) < ε} ⊆ X1 and therefore {x ∈ A; dist(x, Z) < ε} = {x ∈ X1; dist(x, Z) < ε} which has at least #(Z) elements. That is, Y1 and X1 satisfy the condition for the classical Pairing Lemma with Y1 = A, X1 = B as in 2.10. Hence Y1 and X1 can be paired one by one to within ε. Let B = B \ Y1 with subset B1 = B1 \ Y1, and let A = A \ X1 with subset A1 = A1 \ (X1 ∩ A1). Let us verify that B ⊇ B1 and A ⊇ A1 satisfy the condition of the lemma. Let Y ⊆ B1. Then {x ∈ A; dist(x, Y ∪ Y1) < ε} = {x ∈ A1; dist(x, Y ) < ε} ∪ X1. Since #(X1) = #(Y1) and {x ∈ A; dist(x, Y ∪ Y1) < ε} ≥ #(Y ) + #(Y1) = #(Y ) + #(X1), we have, {x ∈ A1; dist(x, Y ) < ε} ≥ #(Y ). Let X ⊂ A1. Since for point y ∈ Y1, dist(y, X) ≥ ε, we have #{y ∈ B; dist(y, X) < ε} = #{y ∈ B; dist(y, X) < ε} ≥ #(X). So the conditions of our lemma hold for A ⊇ A1 and B ⊇ B1 with #( B1) = #(B1) − #(Y1) < n and #( A1) ≤ #(A1) ≤ m. So by the inductive assumption, there exist A2 ⊇ A1 and B2 ⊇ B1 such that A2 and B2 can be paired element by element to within ε. Then the sets A2 = A2 ∪ X1 and B2 = B2 ∪ Y1 satisfy the lemma. (cid:3) Definition 4.3 (See 2.2 of [9]). Let A be a unital C-algebra with T(A) 6= Ø. Recall that each self-adjoint a ∈ A induces a ∈ Aff(T(A)) by a(τ ) = τ (a), τ ∈ T(A). Denote this map by ρA. 1 the set of positive elements with norm at most 1, and then denote by A+ Denote by A+ 1,q the image of A+ 1 in Aff(T(A)) under the canonical map ρA. As a consequence of this generalized version of the Marriage Lemma, one has Lemma 4.4. Let A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)) be an NCCC with n ≥ 1. Let ∆ : A+ 1,q \ {0} → (0, 1) be an order preserving map. Let F ⊆ A be a finite set, and let ε > 0, 1 > γ > 0, and M ∈ N. There are finite sets H1, H2 ⊆ A+ and δ > 0 such that for any unital homomorphisms φ, ψ : A → Mm(C) such that (1) τ (φ(h)), τ (ψ(h)) > ∆(h), h ∈ H1, and (2) τ (φ(h)) − τ (ψ(h)) < δ, h ∈ H2, there are unital homomorphisms φ′, ψ′ : A → Mm(C) such that (3) kφ′(f ) − φ(f )k < ε, kψ′(f ) − ψ(f )k < ε, f ∈ F , (4) SP(φ′) ∩B(1 −γ) = SP(ψ′) ∩B(1 −γ), and each point in SP(φ′) ∩B(1 −γ) has multiplicity at least M, where B(1 − γ) ⊆ Dn is the closed ball with radius 1 − γ. Proof. Let (F , ε) be given. For each f ∈ F , write f = (fB, fD) where fB ∈ B and fD ∈ Mk(C(Dn)). Set FB = {fB : f ∈ F } and FD = {fD : f ∈ F }. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 21 Choose η > 0 such that for any x, y ∈ Dn with dist(x, y) < 4η, one has (4.1) kfD(x) − fD(y)k < ε, f ∈ F . Without loss of generality, one may assume that 5η < γ. Fix an η-dense finite subset {x1, x2, ..., xl} ⊆ B(1 − γ). For each subset define {y1, y2, ..., yt} ⊆ {x1, x2, ..., xl}, h{y1,...,yt} = max{1 − 1 η dist(x, {y1, ..., yt}η), 0}, where Yη denotes the η-neighborhood of Y if Y is a subset of a metric space. Using {y1, y2, ..., yt}2η 6= {y1, y2, ..., yt}3η, choose a positive function g{y1,...,yt} ∈ C(Dn) such that 0 < g{y1,...,yt} ≤ 1 and For functions hy1,...,yt and gy1,...,yt, regard them as the elements supp(g{y1,...,yt}) ⊆ {y1, ..., yt}3η \ {y1, ..., yt}2η. (0, hy1,...,yt1Mk(C(Dn))), (0, gy1,...,yt1Mk(C(Dn))) ∈ A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)), and still denote them by hy1,...,yt and gy1,...,yt, respectively. Put and (4.2) H1 = {g{y1,...,yt} : {y1, ..., yt} ⊆ {x1, x2, ..., xl}}, H2 = {h{y1,...,yt} : {y1, ..., yt} ⊆ {x1, x2, ..., xl}}, δ = min{∆(g{y1,...,yt}) : {y1, ..., yt} ⊆ {x1, ..., xl}}}. Also pick a finite open cover U of B(1 − γ) such that each U ∈ U has diameter at most η, [U ∈U U ⊆ B(1 − γ/2) and U \ [V 6=U V 6= Ø. Then choose continuous functions sU,1, ..., sU,M ∈ C(Dn) such that supp(sU,i) ⊆ [V 6=U V and supp(sU,i) ∩ supp(sU,j) = Ø, i 6= j. Regard each sU,i as an element of A, and put S = {sU,i : U ∈ U, i = 1, ..., M}. Then H1 := H1 ∪ S, H2 and δ have the property stated in the conclusion of the lemma. Indeed, let φ, ψ : A → Mm(C) be unital homomorphisms satisfying (4.3) and (4.4) τ (φ(h)), τ (ψ(h)) > ∆(h), h ∈ H1, τ (φ(h)) − τ (ψ(h)) < δ, h ∈ H2. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 22 Let Y ⊆ SP(φ) ∩ B(1 − γ) be an arbitrary subset. Let denote the subset of the points y satisfying dist(y, Y ) < η. Then {y1, y2, ..., yt} ⊆ {x1, x2, ..., xl} #Y ≤ m · tr(φ(h{y1,...,yt})) ≤ m · tr(ψ(h{y1,...,yt})) + m · δ ≤ m · tr(ψ(h{y1,...,yt})) + m · ∆(gy1,...,yt) ≤ #(SP(ψ) ∩ {y1, ..., yt}2η) + m · ∆(gy1,...,yt) ≤ #(SP(ψ) ∩ {y1, ..., yt}3η) (by (4.3)) (by (4.4)) (by (4.2)) ≤ #(SP(ψ) ∩ Y4η). The same argument shows that #X ≤ #(SP(φ) ∩ X4η), X ⊆ SP(ψ) ∩ B(1 − γ). Then, applying Lemma 4.2 with and A = SP(φ) ∩ B(1 − η), A1 = SP(φ) ∩ B(1 − γ) B = SP(ψ) ∩ B(1 − η), B1 = SP(ψ) ∩ B(1 − γ), one obtains A2 and B2 such that (4.5) and (4.6) SP(φ) ∩ B(1 − γ) ⊆ A2 ⊂ SP(φ) ∩ B(1 − η) SP(ψ) ∩ B(1 − γ) ⊆ B2 ⊂ SP(ψ) ∩ B(1 − η), and A2 and B2 can be paired up to 4η. Write A2 = {zφ,1, zφ,2, ..., zφ,s} and B2 = {zψ,1, zψ,2, ..., zψ,s} for some s, where (4.7) dist(zφ,i, zψ,i) < 4η, i = 1, ..., s. Then, up to unitary equivalence, there are decompositions φ = φ ⊕ sMi=1 πzφ,i and ψ = ψ ⊕ sMi=1 πzψ,i, where SP φ ∩ B(1 − γ) = SP ψ ∩ B(1 − γ) = Ø, and the homomorphisms φ′ = φ = φ ⊕ sMi=1 πzφ,i and ψ′ = ψ ⊕ sMi=1 πzφ,i have the required properties except the requirement on multiplicity. Indeed, by (4.7) and (4.1), one has kφ(f ) − φ′(f )k = 0 < ε and kψ(f ) − ψ′(f )k < ε, f ∈ F . THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 23 By (4.5) and (4.6), one has that SP(φ′) ∩ B(1 − γ) = {zφ,1, ..., zφ,s} ∩ B(1 − γ) = SP(ψ′) ∩ B(1 − γ). To satisfy the multiplicity condition, one needs to perturb φ′ and ψ′ further. Since φ(sU,i) > ∆(sU,i) > 0, U ∈ U, i = 1, ..., M, one has that for any U ∈ U, at least M of {zφ1, ..., zφs} are inside U \ SV 6=U V (counting multiplicity). Then there is a grouping of grouping {zφ1, ..., zφs} such that each group is insider at most one open set U and has elements at least M if it is covered by an open set U. Therefore, after a perturbation, one may assume that each point in {zφ1, ..., zφs} ∩ B(1 − γ) has multiplicity at least M, and hence φ′ and ψ′ satisfy the desired multiplicity condition. (cid:3) Theorem 4.5. Let A = B ⊕Mk(C(Sn−1)) Mk(C(Dn)) be an NCCC, and let ∆ : A+ 1,q \{0} → (0, 1) be an order preserving map. Let F ⊆ A be a finite set. Let ε > 0, η > 0 and K ∈ N \ {0}. There are finite sets H1, H2 ⊆ A+ and δ > 0 such that for any unital homomorphisms φ, ψ : A → Mm(C) such that (1) τ (φ(h)), τ (φ(h)) > ∆(h), (2) τ (φ(h)) − τ (ψ(h)) < δ, h ∈ H1, and h ∈ H2, there exist unital homomorphisms φ, ψ : A → Mm(C) such that (3) (cid:13)(cid:13)(cid:13) φ(f ) − φ(f )(cid:13)(cid:13)(cid:13) < ε, (cid:13)(cid:13)(cid:13) ψ(f ) − ψ(f )(cid:13)(cid:13)(cid:13) < ε, f ∈ F , (4) φ and ψ have decompositions φ = φ0 ⊕ φ1 ⊕ · · · ⊕ φ1 and ψ = ψ0 ⊕ ψ1 ⊕ · · · ⊕ ψ1 such that φ1 and ψ1 are unitarily equivalent, and {z K } K {z } tr( φ0(a)) ≤ η · tr( φ(a)) and tr( ψ0(a)) ≤ η · tr( ψ(a)), a ∈ F . Proof. The statement holds if A is finite dimensional. Assume that the statement holds for B, and let us show that the statement holds for A. Let (F , ε) be given. For each f ∈ F , write f = (fB, fD) where fB ∈ B and fD ∈ Mk(C(Dn)). Set FB = {fB : f ∈ F } and FD = {fD : f ∈ F }. For each r > 0, define gr : [0, 1] ∋ x 7→  x ∈ [0, 1 − r], 1, linear, x ∈ [1 − r, 1 − r/2], 0, x ∈ [1 − r/2, 1]; Also define gr = (0, gr(kxk)1Mk(C(Dn))) ∈ A. Pick γ > 0 such that (4.8) kfD(x) − fD(y)k < ε/8, f ∈ F , dist(x, y) < γ, and consider gγ. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 24 Consider the linear map χγ : Mk(C(Sn−1)) ∋ f 7→ (x 7→ (1 − gγ(kxk)) · f ( x kxk )) ∈ Mk(C(Dn)) Clearly, χγ is a positive linear map, and then the map Γγ : B ∋ f 7→ (f, χγ(ϕ(f ))) ∈ A is positive and injective, where ϕ : B → Mk(C(Sn−1)) is the gluing map. If f ∈ B satisfies that τ (f ) = 0, τ ∈ T(B), then τ (χ(f )) = 0, τ ∈ T(Mk(C(Dn))), and therefore Hence the map τ (Γγ(f )) = 0, τ ∈ T(A). ∆B : B1 is well defined and is order preserving. q ∋ f 7→ ∆(Γγ(f )) ∈ (0, 1) Applying the inductive hypothesis to B with ∆B, FB, ε/2, K and η, one obtains HB,1, HB,2, and δB. Define Put HB,1 = {Γγ(h) : h ∈ HB,1} and HB,2 = {Γγ(h) : h ∈ HB,2}. Let γ be a positive number such that σB = min{∆(h) : h ∈ HB,1}. (4.9) kχγ(h)(x) − χγ(h)(y)k < min{ ∆( \1 − gγ/2)δB 4 , σB 4 , ε 8 }, h ∈ HB,1 ∪ HB,2 ∪ FD for any x, y ∈ Dn satisfying dist(x, y) < γ. Let HD,1 ⊆ A, HD,2 ⊆ A and δD denote the finite sets and constant of Lemma 4.4 with respect to F ∪ HB,1 ∪ HB,2 (in the place of F ), min{ε/8, ∆( \1 − gγ/2)δB/8, σB/8} (in the place of ε), γ, M = ⌊2K/η⌋ + 1, and ∆. Then and H1 = HB,1 ∪ HD,1 ∪ {1 − gγ/2, gγ}, H2 = HB,2 ∪ HD,2, δ = min{∆( \1 − gγ/2)δB/4, σB/4, δD} satisfy the statement. In fact, let φ, ψ : A → Mm(C) be unital homomorphisms satisfying (4.10) and (4.11) τ (φ(h)), τ (φ(h)) > ∆(h), h ∈ H1 τ (φ(h)) − τ (ψ(h)) < δ, h ∈ H2. Since δ ≤ δD, by Lemma 4.4, there are homomorphisms φ′, ψ′ : A → Mm(C) THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 25 such that (4.12) (4.13) and kφ′(f ) − φ(f )k < min{ε/8, ∆( \1 − gγ/2)δB/8, σB/8}, f ∈ F ∪ HB,1 ∪ HB,2, kψ′(f ) − ψ(f )k < min{ε/8, ∆( \1 − gγ/2)δB/8, σB/8}, f ∈ F ∪ HB,1 ∪ HB,2, Sp(φ′) ∩ B(1 − γ) = Sp(ψ′) ∩ B(1 − γ) = {x1, x2, ..., xl}, for some x1, x2, ..., xl ∈ B(1 − γ) with multiplicity at least M. Therefore, up to unitary equiva- lence, there are decompositions (4.14) (4.15) φ′ = φ′ B ⊕ ( πxφ,i) ⊕ ( mφMi=1 mψMi=1 lMi=1 lMi=1 πxi), πxi). ψ′ = ψ′ B ⊕ ( πxψ,i) ⊕ ( for some mφ, mψ ∈ N and some xφ,1, ..., xφ,mφ, xψ,1, ..., xψ,mψ ∈ Dn with 1 > kxφ,ik ≥ γ and 1 > kxψ,jk ≥ γ. By (4.10), (4.12) and (4.13), one has (4.16) tr(φ′(h)), tr(ψ′(h)) > ∆(h), h ∈ HB,1. 7 8 It also follows from (4.11) (4.12) and (4.13) that for any h ∈ HB,1 ∪ HB,2, (4.17) tr(φ′(h)) − tr(ψ′(h)) < tr(φ′(h)) − tr(ψ′(h)) + ∆( \1 − gγ/2)δB/4 < δ + ∆( \1 − gγ/2)δB/4 ≤ ∆( \1 − gγ/2)δB/2. Therefore, by the decompositions (4.14) and (4.15), one has (4.18) tr(φ′ B(h) ⊕ ( mφMi=1 h(xφ,i))) − tr(ψ′ B(h) ⊕ ( mψMi=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) h(xψ,i)))(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) < ∆( \1 − gγ/2)δB/2 For each point xφ,i (or xψ,j), replace the homomorphism πxφ,i (or πxψ,i) by the homomorphism ψ,i = xψ,i/ kxψ,ik ∈ Sn−1). By the choice of φ,i = xφ,i/ kxφ,ik ∈ Sn−1 (or x′ ), where x′ (or πx′ πx′ γ (see (4.9)), one has ψ,i φ,i and (4.19) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mφMi=1 (4.20) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mψMi=1 Put πx′ φ,i (h) − πx′ ψ,i (h) − mφMi=1 mψMi=1 πxφ,i(h)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) πxψ,i(h)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mφMi=1 B ⊕ ( < min{ < min{ ∆( \1 − gγ/2)δB 4 , σB 4 , ε 8 }, h ∈ HB1,1 ∪ HB1,2 ∪ FD, ∆( \1 − gγ/2)δB 4 , σB 4 , ε 8 }, h ∈ HB1,1 ∪ HB1,2 ∪ FD. φ′′ B = φ′ πx′ φ,i ) and ψ′′ B = ψ′ B ⊕ ( mψMi=1 πx′ ψ,i ). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 26 Since x′ φ,i, x′ ψ,j ∈ Sn−1, the homomorphisms φ′′ B and ψ′′ B factor through B. Define (4.21) φ′′ = φ′′ B ⊕ ( lMi=1 πxi) and ψ′′ = ψ′′ B ⊕ ( lMi=1 πxi). By (4.19), (4.20), one has (4.22) kφ′(f ) − φ′′(f )k < ε/8 and kψ′′(f ) − ψ′′′(f )k < ε/8, f ∈ F . By (4.19), (4.20), and (4.16), one has (4.23) tr(φ′′(h)), tr(ψ′′(h)) > ∆(h), h ∈ HB,1. 3 4 By (4.19), (4.20) and (4.18), one has (4.24) tr(φ′′(h)) − tr(ψ′′(h)) < 3 4 ∆( \1 − gγ/2)δB, h ∈ HB,2. For each point xi with kxik > 1 − γ, replace the homomorphism πxi by πx′ i = factors through B. Also note that for each such , where x′ i xi/ kxik ∈ Sn−1. Clearly, the homomorphism πx′ xi, one has i (4.25) tr(πx′ i (h)) > tr(πxi(h)), h ∈ HB,1. By the choice of γ (see (4.8)), one also has (fD) − πxi(fD)(cid:13)(cid:13) < ε/8, f ∈ F . i (cid:13)(cid:13)πx′ B ⊕ ( Mkxik≥1−γ φ′′′ B = φ′′ πx′ i ) and ψ′′′ B = ψ′′ B ⊕ ( Mkx′ ik≥1−γ πx′ i ), (4.26) Set and consider (4.27) φ′′′ = φ′′′ B ⊕ ( Mkxik<1−γ πxi) and ψ′′ = ψ′′′ B ⊕ ( lMkxik<1−γ πxi). By (4.26), one has (4.28) kφ′′(f ) − φ′′′(f )k < ε/8 and kψ′′(f ) − ψ′′′(f )k < ε/8, f ∈ F . By (4.25) and (4.23), one has (4.29) tr(φ′′′(h)), tr(ψ′′′(h)) > ∆(h), h ∈ HB,1. 3 4 By (4.24), one has (4.30) tr(φ′′′(h)) − tr(ψ′′′(h)) < 3 4 ∆( \1 − gγ/2)δB, h ∈ HB,2. Denote by Then, by (4.29), N = rank(φ′′′ B(1)) = rank(ψ′′′ B (1)). N m ≥ tr(φ′′′(1 − gγ/2)) > 3 4 ∆( \1 − gγ/2), THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 27 and Therefore (4.31) m − N m ≥ tr(φ′′′(gγ)) > ∆(d(gγ)). < 1 − ∆(d(gγ)). N m 3 4 ∆( \1 − gγ/2) < Then, by considering the unital homomorphisms φ′′′ B, ψ′′′ B : B → MN (C), it follows from (4.30) and (4.31) that (4.32) 1 N (Tr(φ′′′ B(h)) − Tr(φ′′′ B(h))) = m N (tr(φ′′′(Γγ(h))) − tr(φ′′′(Γγ(h)))) < δB, h ∈ HB,2. It also follows from (4.29) that and 1 N 1 N Tr(φ′′′ B(h)) ≥ tr(φ′′′ B(h)) = tr(φ′′′(Γγ(h))) > ∆(h), Tr(ψ′′′ B (h)) ≥ tr(ψ′′′ B (h)) = tr(ψ′′′(Γγ(h))) > ∆(h), h ∈ HB,1, h ∈ HB,1. Then, it follows from the inductive hypothesis that there are homomorphisms φB, ψB : B → MN (C) such that (4.33) (cid:13)(cid:13)(cid:13) φB(fB) − φ′′′ and there are decompositions (cid:13)(cid:13)(cid:13) ψB(fB) − ψ′′′ B(fB)(cid:13)(cid:13)(cid:13) < ε/2 and } {z K B (fB)(cid:13)(cid:13)(cid:13) < ε/2, {z K f ∈ F , } φB = φB,0 ⊕ φB,1 ⊕ · · · ⊕ φB,1 and ψB = ψB,0 ⊕ ψB,1 ⊕ · · · ⊕ ψB,1 with φB,1 and ψB,1 unitarily equivalent and (4.34) (4.35) 1 N 1 N rank( φB,0(a)) ≤ η · rank( ψB,0(a)) ≤ η · 1 N 1 N rank( φB(a)), a ∈ FB, rank( ψB(a)), a ∈ FB. By (4.31) and (4.34), one has that for any a ∈ FB, one has (4.36) tr( φB,0(a)) = 1 m rank( φB,0(a)) = N m 1 N rank( φB,0(a)) ≤ η · 1 m rank( φB(a)) = η · tr( φB(a)), and for the same reason, (4.37) tr( ψB,0(a)) ≤ η · tr( ψB(a)), a ∈ FB. Consider the map Lkxik<1−γ πxi, there is a decomposition Mkxik<1−γ πxi = Mkxik<1−γ πx′ i πx′′ i ) ⊕ · · · ⊕ ( Mkxik<1−γ ⊕ ( Mkxik<1−γ {z K πx′′ i ) , } THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 28 where each x′ K. Since each xi has multiplicity at least M = ⌊2K/η⌋ + 1, one has that i are same as xi but with different multiplies, and x′ i and x′′ i has multiplicity at most (4.38) Then set and i πx′ tr( Mkxik<1−γ φ0 = φB,0 ⊕ Mkxik<1−γ φ1 = φB,1 ⊕ ( Mkxik<1−γ (a)) ≤ η · tr( Mkxik<1−γ πxi(a)), a ∈ A. πx′ i and πx′′ i ) and πx′ i , ψ0 = ψB,0 ⊕ Mkxik<1−γ ψ1 = ψB,1 ⊕ ( Mkxik<1−γ πx′′ i ), where φB,0, φB,1, ψB,0, ψB,1 are regarded as maps on A. The homomorphisms and these decompositions have the properties required in the statement of the theorem. Indeed, by (4.12), (4.13), (4.22), (4.28), and (4.33), one has φ = φ0 ⊕ φB,1 ⊕ · · · ⊕ φB,1 and ψ = ψ0 ⊕ ψB,1 ⊕ · · · ⊕ ψB,1 K } {z {z } (cid:13)(cid:13)(cid:13)φ(f ) − φ(f )(cid:13)(cid:13)(cid:13) < kφ(f ) − φ′′′(f )k +(cid:13)(cid:13)(cid:13)φ′′′(f ) − φ(f )(cid:13)(cid:13)(cid:13) < ε/2 + ε/2 = ε, (cid:13)(cid:13)(cid:13)ψ(f ) − ψ(f )(cid:13)(cid:13)(cid:13) < kψ(f ) − ψ′′′(f )k +(cid:13)(cid:13)(cid:13)ψ′′′(f ) − ψ(f )(cid:13)(cid:13)(cid:13) < ε/2 + ε/2 = ε, and f ∈ F , f ∈ F . It is also clear that φ1 and ψ1 are unitarily equivalent, and it follows from (4.36), (4.38) and (4.38) that the maps φ0 and ψ0 satisfy the desired trace condition. (cid:3) Recall (from [9]) Lemma 4.6 (Lemma 4.13 of [9]). Let A be a unital separable nuclear residually finite dimensional C-algebra satisfying the UCT, and let ∆ : A+ 1,q \ {0} → (0, 1) be an order preserving map. For any finite set F ⊆ A and any ε > 0, there exist δ > 0, a finite set G ⊆ A, a finite set P ⊆ K(A), a finite set H ⊆ A1 + \ {0}, and an integer K ≥ 1 satisfying the following condition: For any two unital G-δ-multiplicative linear maps φ1, φ2 : A → Mn(C) (for some integer n) and any unital homomorphism ψ : A → Mm(C) with m ≥ n such that (1) τ ◦ ψ(g) ≥ ∆(g), g ∈ H, (2) [φ1]P = [φ2]P, there exists a unitary U ∈ MKm+n(C) such that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) φ1(f ) ⊕ ψ(f ) ⊕ · · · ⊕ ψ(f ) −u∗(φ2(f ) ⊕ ψ(f ) ⊕ · · · ⊕ ψ(f ) < ε, f ∈ F . {z K } {z K )u(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) } THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 29 Theorem 4.7. Let A be an NCCC. Let ∆ : A+ 1,q \ {0} → (0, 1) be an order preserving map. Let (F , ε) be given. Then there are finite sets G ⊆ A, H1 ⊆ A+, H2 ⊆ A+ and P ⊆ K(A), and positive numbers δ, σ > 0 satisfying the following condition: If φ, ψ : A → Mm(C) are unital G-δ-multiplicative maps such that (1) [φ]P = [ψ]P, (2) tr(φ(h)) > ∆(h) and tr(φ(h)) > ∆(h), h ∈ H1, and (3) tr(φ)(h) − tr(ψ)(h) < σ, h ∈ H2, then there is a unitary u ∈ Mm(C) such that kφ(f ) − u∗ψ(f )uk < ε, f ∈ F . Proof. Applying Lemma 4.6 to the data A, ∆/4, and (F , ε/3), one obtains (G′, δ′, P, H, K) satisfying the conditions of Lemma 4.6. Applying Theorem 4.5 to the data A, ∆/2, F ∪H, min{ε/6, ∆(H)/4} (in place of ε), η = 1/2K, and K, one obtains H1, H2, and 2σ (in place of δ) satisfying the conclusion of Theorem 4.5. Without loss of generality, one may assume that H1 and H2 are in the unital ball of A. Applying Corollary 5.5 of [9] to F ∪ H1 ∪ H2, min{ε/6, ∆(H1)/2, ∆(H2)/2, σ/4}, min{1/2K, ∆(H1)/2, ∆(H2)/2, σ/4(1 + σ)}, one obtains (G, δ). Then G, δ, σ, H1, H2, and P satisfy the conclusion of the theorem. Indeed, let φ, ψ : A → Mm(C) be G-δ-multiplicative maps such that (1) [φ]P = [ψ]P, (2) tr(φ(h)) > ∆(h) and tr(φ(h)) > ∆(h), h ∈ H1, and (3) tr(φ)(h) − tr(ψ)(h) < σ, h ∈ H2. By Theorem 5.5 of [9], there are φ0, φ1, ψ0, ψ1 : A → Mm(C) such that φ0, ψ0 are G′-δ′- multiplicative, φ1, ψ1 are homomorphisms, such that (4.39) kφ(a) − φ0(a) ⊕ φ1(a)k < min{ε/6, ∆(H1)/2, ∆(H2)/2, σ/4}, a ∈ F ∪ H1 ∪ H2, (4.40) kψ(a) − ψ0(a) ⊕ ψ1(a)k < min{ε/6, ∆(H1)/2, ∆(H2)/2, σ/4}, a ∈ F ∪ H1 ∪ H2, and tr(φ0(1)) = tr(ψ0(1)) < min{1/2K, ∆(H1)/2, ∆(H2)/2, σ/4(1 + σ)}. Consider the unital homomorphisms φ1, ψ1 : A → pMm(C)p, where p = φ1(1) = ψ1(1). One has that (1) τ (φ1(h)), τ (φ1(h)) > ∆(h)/2, (2) τ (φ1(h)) − τ (ψ1(h)) < 2σ, h ∈ H1, and h ∈ H2, where τ ∈ T(pMm(C)p). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 30 By Theorem 4.5, up to unitary equivalence, there are homomorphisms φ′ 1, ψ′ 1, µ : A → pMm(C)p such that < min{ε/6, ∆(H)/4}, a ∈ F ∪ H, < min{ε/6, ∆(H)/4}, a ∈ F ∪ H, (4.41) (4.42) and (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) φ1(a) − φ′ 1(a) ⊕ µ(a) ⊕ · · · ⊕ µ(a) ψ1(a) − ψ′ 1(a) ⊕ µ(a) ⊕ · · · ⊕ µ(a) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) } } {z K {z K τ (φ′ 1(a)) ≤ 1 2K · τ (φ1(a)) and τ (ψ′ 1(a)) ≤ 1 2K Therefore, one has that · τ (ψ1(a)), a ∈ F ∪ H, τ ∈ T(pMm(C)p). τ (µ(h)) > ∆(h), 1 4 h ∈ H, τ ∈ T(qMm(C)q), where q = µ(1). Consider the map and note that (φ0 ⊕ φ′ 1) ⊕ (µ ⊕ · · · ⊕ µ ) and (ψ0 ⊕ ψ′ 1) ⊕ (µ ⊕ · · · ⊕ µ ), {z K } {z K } tr(φ0(1) ⊕ φ′ 1(1)) = tr(ψ0(1) ⊕ ψ′ 1(1)) < tr(µ(1)). It then follows from Lemma 4.6 that there is a unitary u ∈ Mm(C) such that for any a ∈ F , (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (φ0(a) ⊕ φ′ 1(a)) ⊕ (µ(a) ⊕ · · · ⊕ µ(a) ) − u∗((ψ0(a) ⊕ ψ′ 1(a)) ⊕ (µ(a) ⊕ · · · ⊕ µ(a) {z K } {z K It then follows from (4.39), (4.40), (4.41), and (4.42) that as desired. kφ(a) − u∗ψ(a)uk < ε, a ∈ F , ))u(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) } < ε 3 . (cid:3) Note that KL(A, Q) ∼= Hom(K0(A), K0(Q)), a straightforward consequence is Corollary 4.8. Let A be an NCCC. Let ∆ : A+ 1,q \ {0} → (0, 1) be an order preserving map. Let (F , ε) be given. Then there are finite sets H1 ⊆ A+ and H2 ⊆ A+ and a positive number δ > 0 satisfying the following condition: If φ, ψ : A → Q are unital homomorphisms such that (1) [φ]0 = [ψ]0, (2) tr(φ(h)) > ∆(h) and tr(φ(h)) > ∆(h), h ∈ H1, and (3) tr(φ)(h) − tr(ψ)(h) < δ, h ∈ H2, then there is a unitary u ∈ Q such that kφ(f ) − u∗ψ(f )uk < ε, f ∈ F . THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 31 It is also worth pointing out the following corollary: Corollary 4.9. Let A be a subhomogeneous C-algebra. Let ∆ : A+ 1,q \ {0} → (0, 1) be an order preserving map. Let (F , ε) be given. Then there are finite sets G ⊆ A, H1 ⊆ A+, H2 ⊆ A+ and P ⊆ K(A), and positive numbers δ, σ > 0 satisfying the following condition: If φ, ψ : A → Mm(C) are unital G-δ-multiplicative maps such that (1) [φ]P = [ψ]P, (2) tr(φ(h)) > ∆(h) and tr(φ(h)) > ∆(h), h ∈ H1, and (3) tr(φ)(h) − tr(ψ)(h) < σ, h ∈ H2, then there is a unitary u ∈ Mm(C) such that kφ(f ) − u∗ψ(f )uk < ε, f ∈ F . Proof. This follows the fact that any subhomogeneous C-algebra can be locally approximated by NCCCs (Theorem 2.15 of [5]). (cid:3) 5. Tracial factorization and tracial approximation Recall that Definition 5.1 ([13], [2]). Let S be a class of unital C-algebras. A C-algebra A is said to be tracially approximated by the C-algebras in S, and one writes A ∈ TAS, if the following condition holds: For any finite set F ⊆ A, any ε > 0, and any nonzero a ∈ A+, there is a nonzero sub-C-algebra S ⊆ A such that S ∈ S, and if p = 1S, then (1) kpf − f pk < ε, f ∈ F , (2) pf p ∈ε S, f ∈ F , and (3) 1 − p is Murray-von Neumann equivalent to a subprojection of aAa. One particularly important class S of building blocks is the class of Elliott-Thomsen algebras. Definition 5.2. ([6], [1]) A C-algebra C is said to be an Elliott-Thomsen algebra if C = {(a, f ) ∈ E ⊕ (F ⊗ C([0, 1])) : f (0) = 0(a), f (1) = 1(a)} for some finite dimensional C-algebras E, F , where 0, 1 : E → F are unital homomorphisms. Denote by π∞ the standard quotient map π∞ : A ∋ (a, f ) 7→ a ∈ E. Let us denote the class of unital Elliott-Thomsen algebras by C, and denote the class of unital Elliott-Thomsen algebras with trivial K1-group by C0. Remark 5.3. In fact, by Corollary 29.3 of [9], one has TAC = TAC0. Remark 5.4. In fact, the class of unital Elliott-Thomsen algebras is exactly the class of NCCCs with dimensions of cells at most one; see [5]. For TAC0 algebras, one has the following classification theorem. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 32 Theorem 5.5 (Corollary 28.7 of [9]). Let A, B be unital separable amenable simple C-algebras satisfying the UCT. Assume that A, B are Jiang-Su stable, and assume that A ⊗ Q ∈ TAC0 and B ⊗ Q ∈ TAC0. Then A ∼= B if and only if Ell(A) ∼= Ell(B). In this section, let us show that for any separable simple unital locally ASH algebra A, one has that A ⊗ Q ∈ TAC0. Theorem 5.6. Let A be a unital simple separable locally approximately subhomogeneous (ASH) C-algebra satisfying A ∼= A ⊗ Q. Then, for any finite set F ⊆ A and any ε > 0, there exist an Elliott-Thomsen algebra C with K1(C) = {0}, a unital completely positive linear map Φ : A → C, and a unital embedding Ψ : C → A such that (1) Φ is F -ε-multiplicative, and (2) τ (f ) − τ (Ψ(Φ(f ))) < ε, f ∈ F , τ ∈ T(A). Proof. Without loss of generality, one may assume that 1 ∈ F and each element of F is self- adjoint and has norm at most 1. Let A be a unital separable simple locally ASH algebra satisfying A ∼= A ⊗ Q. By Theorem 2.15 of [5], the C-algebra A can be locally approximated by unital NCCCs. Therefore, without loss of generality, one may assume that there is a sub-C-algebra A1 ⊆ A such that A1 is a NCCC and F ⊆ A1. Put GA1 = ρA1(K0(A1)), G+ A1 = ρA1(K0(A1)) ∩ Aff +(T(A1)), uA1 = ρA1([1]), and fix a set {p1, ..., pn} which generates the group GA1. Note that by Lemma 2.9, the positive cone G+ A1 is finitely generated. For each h ∈ A+, define ∆(h) = inf{τ (ι(h)) : τ ∈ T(A)}, where ι : A1 → A is the embedding map. Since A is simple, the map ∆ induces a order preserving map from A1,q + to (0, 1). Let us still denote this by ∆. Let H1 ⊆ (A1)+, H2 ⊆ (A1)+ and δ0 > 0 be the finite sets and constant of Corollary 4.8 with respect to F · F , ε/4, and ∆/4. Let H be the subset of Theorem 3.7 with respect to A1, F ∪ H1 ∪ H2 (in the place of F ), min{ε/4, δ0/4, ∆(h)/4 : h ∈ H1} (in the place of ε). Put σ = 1 2 min{∆(h) : h ∈ H}, and denote by δ1 the constant of Theorem 3.7 with respect to σ. Put G = ρA(K0(A)), G+ = ρA(K0(A)) ∩ Aff +(T(A)), u = ρA([1]). Then (G, G+, u) is a unperforated order-unit group, and there is a natural pairing between (G, G+, u) and T(A) induced by ρA. Still denote this pairing map by ρA. Note that one has the THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 33 following commutative diagram: [ι]0 K0(A1) K0(A) ρA1 ρA GA1 (ι)∗ / G. lim −→ Consider ((G, G+, u), T(A), ρA). By Theorem 5.2.2 of [1], there is an inductive system C ′ = (C ′ i, η′ i) with C ′ i ∈ C0 and ηi injective such that there is an isomorphism Ξ : ((G, G+, u), T(A), ρA) → ((K0(C ′), K0(C ′), [1]), T(C ′), ρ′ C). Consider C = C ′ ⊗ Q = lim −→ (Ci, ηi), where Ci = C ′ i ⊗ Q and ηi = η′ i ⊗ idQ. One has ((G, G+, u), T(A), ρA) ∼= ((K0(C), K0(C), [1]), T(C), ρC), and let us still denote by Ξ the isomorphism. In the remaining part of the paper, let us also use Ξ to denote its restriction to G or to Aff(T(A)), depending on the context. Consider the following diagram: GA1 [ι]0 / G Ξ K0(C1) / K0(C2) / · · · [η2]0 [η1]0 / K0(C). Since the positive cone of GA1 is finitely generated (Lemma 2.9), the positive map [ι]0 can be lifted to a positive homomorphism GA1 → K0(Cn) for sufficiently large n. Without loss of generality, one may assume that [ι]0 has a lifting κ : GA1 → K0(C1), making the diagram commutative: [ι]0 GA1 κ G Ξ K0(C1) / K0(C2) / · · · [η2]0 [η1]0 / K0(C). By Lemma 2.10, after a telescoping of the inductive system (Ci, ηi), there is also an approximate lifting, making the diagram of affine function spaces, Aff(T(A1)) γ (ι)∗ Aff(T(A))) Ξ Aff(T(C1))) / Aff(T(C2))) (η1)∗ (η2)∗ / · · · / Aff(T(C))), approximately commutative, and such that (5.1) (5.2) τ (κ([pi])) − γ([pi])(τ ) < δ1, τ ∈ T(C1), 1 ≤ i ≤ n, γ(h)(τ ) > 1 2 ∆(h) > σ, h ∈ H ∪ H1, τ ∈ T(C1), / / / / / / / / / / / / / / / / / and (5.3) Write where THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 34 (cid:12)(cid:12)(cid:12)(Ξ−1 ◦ (η1,∞)∗ ◦ γ)( f )(τ ) − (ι)∗( f )(τ )(cid:12)(cid:12)(cid:12) < ε/8, τ ∈ T(A), f ∈ F . C1 = C ′ 1 ⊗ Q = {(a, f ) ∈ E ⊕ (F ⊗ C([0, 1])) : f (0) = 0(a), f (1) = 1(a)}, E = pMi=1 Q, F = lMj=1 Q, for natural numbers p, l, and 0, 1 : E → F are unital homomorphisms. On each interval [0, 1]j, choose a partition 0 = t0 < t1 < t2 < · · · < tk−1 < tk = 1 such that (5.4) (cid:12)(cid:12)(cid:12)γ( f )(τs) − γ( f )(τti)(cid:12)(cid:12)(cid:12) < min{ε/4, δ0/2}, where τs = tr ◦ πs and tr is the canonical trace of Q. One may assume that k is sufficiently large that s ∈ [ti, ti+1], f ∈ F ∪ H2, For each 0 < i < k, define 2π/(k − 1) < ε/8. and κi = [πti]0 ◦ κ : K0(A1) → K0(Q) ∼= Q, γi = (πti)∗ ◦ γ : Aff(T(A1)) → Aff(T(Q)) ∼= R. By (5.1), each pair (κi, γi) is δ1-compatible on [pi], 1 ≤ i ≤ n. By (5.2), one has that γi(h)(tr) > σ, h ∈ H. Therefore, by Theorem 3.7, there is a homomorphism φi : A1 → Q such that [φi]0 = κi and (5.5) (cid:12)(cid:12)(cid:12)γi(h)(tr) − tr(φi(h))(cid:12)(cid:12)(cid:12) < min{ε/4, δ0/4, ∆(h)/4 : h ∈ H1}, Together with (5.2), it then follows that h ∈ F ∪ H1 ∪ H2. (5.6) tr(φi(h)) > γi(h)(tr) − ∆(h) > 1 4 1 4 ∆(h), h ∈ H1. (5.7) It also follows from (5.5) and (5.4) that for any h ∈ H2 and any 1 ≤ i ≤ k − 2, tr(φi(h)) − tr(φi+1(h)) ≤ (cid:12)(cid:12)(cid:12)tr(φi(h)) − γi(h)(tr)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)γi(h)(tr) − γi+1(h)(tr)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)tr(φi+1(h)) − γi+1(h)(tr)(cid:12)(cid:12)(cid:12) < δ0/4 + δ0/2 + δ0/4 = δ0. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 35 Since πi is homotopic to πi+1, it is clear that [κi] = [κi+1], 1 ≤ i ≤ k − 2, and therefore (5.8) [φ]0 = [φi+1]0, 1 ≤ i ≤ k − 2. Denote by π∞ : C1 → E the standard quotient map, and consider κ∞ = [π∞]0 ◦ κ : K0(A1) → K0(E) and γ∞ = [π∞] ◦ γ : Aff(T(A1)) → Aff(T(E)). The same argument as above shows that there is a homomorphism φ∞ : A1 → E such that [φ∞]0 = κ∞, (5.9) and (5.10) (cid:12)(cid:12)(cid:12)γ∞(h)(τ ) − τ (φ∞(h))(cid:12)(cid:12)(cid:12) < min{ε/4, δ0/4, ∆(h)/4 : h ∈ H1}, Define h ∈ H1 ∪ H2, τ ∈ T(E). φ0 = 0 ◦ φ∞ and φk = 1 ◦ φ∞ and consider the restrictions of these maps to the j-th direct summand; still denote them by φ0 and φk respectively. It then follows from (5.9) that (5.11) [φ0]0 = [φk] = [φi]0, 1 ≤ i ≤ k − 1, and it follows from (5.10) and (5.2) that (5.12) tr(φ0(h)), tr(φk(h)) > ∆(h), h ∈ H1. 1 4 Moreover, with (5.10) and (5.4), the same argument as that of (5.7) shows that (5.13) tr(φ0(h)) − tr(φ1(h)) < δ0 and tr(φk−1(h)) − tr(φk(h)) < δ0, h ∈ H2. Thus, with (5.11), (5.6), (5.12), (5.7) and (5.13), Corollary 4.8 implies that there are unitaries u1, u2, ..., uk−1 ∈ Q such that (cid:13)(cid:13)φi(f ) − u∗ i+1φi+1(f )ui+1(cid:13)(cid:13) < ε/4, Define v0 = 1, and f ∈ F · F , 0 ≤ i ≤ k − 2. vi = uiui−1 · · · u1, i = 1, ..., k − 1. Then, for any 0 ≤ i ≤ k − 2 and any f ∈ F · F , one has kAd(vi) ◦ φi(f ) − Ad(vi+1) ◦ φi+1(f )k = k(ui · · · u1)∗φi(f )(ui · · · u1) − (ui+1 · · · u1)∗φi+1(f )(ui+1 · · · u1)k = (cid:13)(cid:13)φi(f ) − u∗ i+1φi+1(f )ui+1(cid:13)(cid:13) < ε/4. Replacing each homomorphism φi by Ad(vi) ◦ φi for i = 1, ..., k − 1, and still denoting it by φi, one has kφi(f ) − φi+1(f )k < ε/4, f ∈ F · F , 0 ≤ i ≤ k − 2. Note that the replacement of φi does not change the induced map on the invariant, and hence one still has [φk−1]0 = [φk]0, THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 36 and tr(φ0(h)), tr(φk(h)) > ∆(h), h ∈ H1, 1 4 tr(φk−1(h)) − tr(ψk(h)) < δ0, h ∈ H2. Applying Corollary 4.8 again, we obtain a unitary w ∈ Q such that kw∗φk−1(f )w − φk(f )k < ε/4, f ∈ F · F . Since Q is AF, and any unitary in a finite dimensional C-algebra can be connected to the identity along a path with length at most π (i.e., it has exponential length at most π), there are unitaries 1 = w0, w1, ..., wk−2, wk−1 = w ∈ Q kwi − wi−1k < 2π/(k − 1) < ε/8. (cid:13)(cid:13)w∗ i φi(f )wi − w∗ i+1φi+1(f )wi+1(cid:13)(cid:13) < 3ε/8, f ∈ F · F , 0 ≤ i ≤ k − 2. Replace each homomorphism φi again by Ad(wi) ◦ φi, 1 ≤ i ≤ k − 1, and still denote it by φi. One then has (5.14) kφi(f ) − φi+1(f )k < 3ε/8, f ∈ F · F , 0 ≤ i ≤ k − 1. Define the positive linear map φ : A1 → C([0, 1]j, Q) by Φj(f )(t) = ti+1 − t ti+1 − ti φi(f ) + t − ti ti+1 − ti φi+1(f ), if t ∈ [ti, ti+1]. By (5.14), the map Φj is F -ε-multiplicative. It also follows from (5.4) that if t ∈ [ti, ti+1], then, for any f ∈ F , such that Hence, (5.15) (cid:12)(cid:12)(cid:12)γ( f )(τt) − τt(Φj(f ))(cid:12)(cid:12)(cid:12) ≤ (cid:12)(cid:12)(cid:12)γ( f )(τti) − τt(Φj(f ))(cid:12)(cid:12)(cid:12) + ε/4 (by (5.4)) = (cid:12)(cid:12)(cid:12)(cid:12)γi( f )(tr) − ( ≤ (cid:12)(cid:12)(cid:12)γi( f )(tr) − tr(φi(f ))(cid:12)(cid:12)(cid:12) + 5ε/8 (by (5.14)) < ε/4 + 5ε/8 = 7ε/8 (by (5.5)). ti+1 − t ti+1 − ti tr(φi(f )) + t − ti ti+1 − ti tr(φi+1(f )))(cid:12)(cid:12)(cid:12)(cid:12) + ε/4 Repeat this construction for all j = 1, ..., l, and note that the maps Φ1, Φ2, ..., Φl induce a map Φ : A1 → C1. Since each Φj is F -ε multiplicative, the map Φ is F -ε-multiplicative. By (5.15), one has (5.16) f ∈ F , τ ∈ T(C1). (cid:12)(cid:12)(cid:12)γ( f )(τ ) − τ (Φ(f ))(cid:12)(cid:12)(cid:12) < 7ε/8, Now, let us construct an embedding Ψ : C1 → A such that τ (f ) − τ (Ψ(Φ(f ))) < ε, f ∈ H, τ ∈ T(A). THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 37 Since A ∼= A ⊗ Q, the subgroup H := ker ρA ⊆ K0(A) is divisible, and therefore the exact sequence splits. Pick a decomposition 0 / H / K0(A) ρA / G / 0 K0(A) = G ⊕ H. Since K0(A) is weakly unperforated, the order on (G ⊕ H) is completely determined by that on G. Define Then κ′ is a positive homomorphism, and the pair κ′ : G ∋ g 7→ (g, 0) ∈ G ⊕ H = K0(A). (κ′ ◦ (Ξ−1K0(C)), ΞAff(T(C))) is compatible. It induces a positive homomorphism θ : Cu∼(C) → Cu∼(A). By Theorem 1 of [26], there is a homomorphism ψ : C → A such that the Cuntz map induced by ψ is θ. In particular, one has that (5.17) Then the map (ψ)∗ = Ξ−1Aff(T(C)). Ψ = ψ ◦ η1,∞ satisfies the conclusion of the theorem (together with Φ). Indeed, since C is simple, the map ψ is an embedding, and therefore Ψ is an embedding. Moreover, for any f ∈ F , one has τ (ι(f )) − τ (Ψ ◦ Φ(f )) = (cid:12)(cid:12)(cid:12)(ι)∗( f )(τ ) − (Ψ)∗((Φ)∗( f ))(τ )(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(ι)∗( f )(τ ) − (ψ)∗((ηi,∞ ◦ Φ)∗( f ))(τ )(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(ι)∗( f )(τ ) − (Ξ−1 ◦ (ηi,∞)∗ ◦ (Φ)∗)( f )(τ )(cid:12)(cid:12)(cid:12) < (cid:12)(cid:12)(cid:12)(ι)∗( f )(τ ) − (Ξ−1 ◦ (ηi,∞)∗ ◦ γ)( f )(τ )(cid:12)(cid:12)(cid:12) + 7ε/8 (by (5.16)) ≤ ε/8 + 7ε/8 = ε (by (5.17)) (by (5.3)), as desired. (cid:3) Remark 5.7. With a slight modification (a perturbation of the linear map γ), the same argu- ment as Theorem 5.6 shows that the same statement holds for C-algebras which are tracially approximated by subhomogeneous C-algebras. The passage from Theorem 5.6 to the actual tracial approximation is an application of the following very important theorem due to Winter: / / / / THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 38 Theorem 5.8 (Theorem 2.2 of [32]). Let S be a class of separable unital C-algebras which have a finite presentation with weakly stable relations. Suppose further that S is closed under taking direct sums and under taking tensor products with finite dimensional C-algebras, and that S contains all finite dimensional C-algebras. Let A be a separable, simple, unital C-algebra with dimnucA < ∞ and T(A) 6= Ø, and let be a system of maps with the following properties: ( A σi / Bi i / A )i∈N (1) Bi ∈ S, i ∈ N, (2) i is an embedding for each i ∈ N, (3) σi is a completely positive contraction for each i ∈ N, (4) ¯σ : A →Qi∈N Bi/Li∈N Bi induced by σi is a unital homomorphism (5) sup{τ (iσi(a) − a) : τ ∈ T(A)} → 0, as i → ∞ for each a ∈ A. Then A ⊗ Q ∈ TAS. With this and Theorem 5.6, one has Theorem 5.9. Let A be a unital simple separable locally ASH C-algebra. Then A ⊗ Q ∈ TAC0, where C0 is the class of unital Elliott-Thomsen algebras with trivial K1-group. In particular, if A ∼= A ⊗ Z, where Z is the Jiang-Su algebra, then A is classifiable (by means of the naive Elliott invariant). (The converse is also true.) Proof. By Theorem 3.1 of [5], one has dr(A⊗Q) ≤ 2, and in particular, dimnuc(A⊗Q) ≤ 2 < +∞. It then follows from Theorems 5.6 and 5.8 that A⊗Q ∈ TAC0. By the classification theorem of [9] (based in particular on the deformation technique of [33] and [17] -- see also [19]), the C-algebra A is classifiable. (cid:3) Corollary 5.10. Let A be a simple separable unital locally ASH (respectively, locally AH) algebra. Then A ⊗ Z is an ASH (respectively, AH) algebra. Proof. By Theorem 5.9, the C-algebra A ⊗ Z is classifiable by means of the Elliott invariant. By [1] and [8], the Elliott invariant for separable, Jiang-Su stable, simple, unital, finite C- algebras (in particular, locally ASH algebras) is exhausted by ASH algebras (by Theorem 3 of [8], finiteness implies stable finiteness in this setting). Furthermore, by [30], the Elliott invariant for separable, Jiang-Su stable, simple, unital, locally AH algebras is exhausted by AH algebras. (In both settings, the models have no dimension growth.) (cid:3) The classification of locally ASH algebras (Theorem 5.9) in fact allow us to recover the recent classification result for the C-algebra of a minimal homeomorphism -- assumed to have mean dimension zero but not to be uniquely ergodic ([27], [18] -- the uniquely ergodic case was dealt with in [3], or in [29] on the ease the space is finite dimensional): Corollary 5.11 (Corollary 5.3 of [18]). Let X be a compact metrizable space, and let σ : X → X be a minimal homeomorphism. Then the C-algebra (C(X)⋊σ Z)⊗Z is classifiable. In particular, if (X, σ) has mean dimension zero, the C-algebra C(X) ⋊σ Z is classifiable. / / THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 39 Proof. By Theorem 4.1 of [5], the C-algebra (C(X) ⋊σ Z) ⊗ Q is locally ASH. By Theorem 5.9, the C-algebra (C(X) ⋊σ Z) ⊗ Q belongs to the class TAC0, and hence the C-algebra (C(X) ⋊σ Z) ⊗ Z is classifiable. If (X, σ) has mean dimension zero, then it follows from [3] that C(X) ⋊σ Z ∼= (C(X) ⋊σ Z) ⊗ Z, and so the C-algebra C(X) ⋊σ Z is classifiable. (cid:3) References [1] G. A. Elliott. An invariant for simple C-algebras. Canadian Mathematical Society. 1945 -- 1995, 3:61 -- 90, 1996. [2] G. A. Elliott and Z. Niu. On tracial approximation. J. Funct. Anal., 254(2):396 -- 440, 2008. URL: http://dx.doi.org/10.1016/j.jfa.2007.08.005, doi:10.1016/j.jfa.2007.08.005. [3] G. A. Elliott and Z. Niu. The C-algebra of a minimal homeomorphism of zero mean dimension. 06 2014. URL: http://arxiv.org/abs/1406.2382, arXiv:1406.2382. [4] G. A. Elliott and Z. Niu. On the classification of simple unital C-algebras with finite decomposition rank. Preprint, 2015. [5] G. A. Elliott, Z. Niu, L. Santiago, and A. Tikuisis. Decomposition rank of approximately subhomogeneous C-algebras. 05 2015. URL: http://arxiv.org/abs/1505.06100, arXiv:1505.06100. [6] G. A. Elliott and K. Thomsen. The state space of the K0-group of a simple separable C-algebra. Geom. Funct. Anal., 4(5):522 -- 538, 1994. URL: http://dx.doi.org/10.1007/BF01896406, doi:10.1007/BF01896406. [7] G. A. Elliott and A. S. Toms. Regularity properties in the classification program for sep- 2008. URL: amenable C-algebras. Bull. Amer. Math. arable http://dx.doi.org/10.1090/S0273-0979-08-01199-3, doi:10.1090/S0273-0979-08-01199-3. 45(2):229 -- 245, (N.S.), Soc. [8] G. Gong, X. Jiang, Canad. Math. Bull., doi:10.4153/CMB-2000-050-1. and H. Su. Obstructions 43(4):418 -- 426, 2000. URL: to Z-stability for unital simple C-algebras. http://dx.doi.org/10.4153/CMB-2000-050-1, [9] G. Gong, H. Lin, and Z. Niu. Classification of finite simple amenable Z-stable C-algebras. 01 2015. URL: http://arxiv.org/abs/1501.00135, arXiv:1501.00135. [10] P. R. Halmos and H. E. Vaughan. The marriage problem. Amer. J. Math., 72:214 -- 215, 1950. [11] E. Kirchberg. The classification of purely infinite C-algebras using Kasparov's theory. Preprint, 1994. [12] E. Kirchberg algebra O2. J. Reine Angew. Math., 525:17 -- 53, 2000. URL: http://dx.doi.org/10.1515/crll.2000.065, doi:10.1515/crll.2000.065. and N. C. Phillips. Embedding exact C-algebras in the Cuntz of [13] H. Lin. Tracially AF C-algebras. Trans. Amer. Math. Soc., 353(2):693 -- 722, 2001. [14] H. Lin. Simple nuclear C-algebras of tracial topological rank one. J. Funct. Anal., 251(2):601 -- 679, 2007. URL: http://dx.doi.org/10.1016/j.jfa.2007.06.016, doi:10.1016/j.jfa.2007.06.016. [15] H. Lin. Asymptotically unitary equivalence and asymptotically inner automorphisms. Amer. J. Math., 131(6):1589 -- 1677, 2009. URL: http://dx.doi.org/10.1353/ajm.0.0086, doi:10.1353/ajm.0.0086. simple amenable C-algebras. [16] H. Lin. Asymptotic unitary equivalence and classification of In- http://dx.doi.org/10.1007/s00222-010-0280-9, vent. Math., doi:10.1007/s00222-010-0280-9. 183(2):385 -- 450, 2011. URL: [17] H. Lin. Localizing the Elliott conjecture at strongly self-absorbing C-algebras, II. J. Reine Angew. Math., 692:233 -- 243, 2014. [18] H. Lin. Crossed products and minimal dynamical systems. 02 2015. URL: http://arxiv.org/abs/1502.06658, arXiv:1502.06658. THE CLASSIFICATION OF SIMPLE SEPARABLE UNITAL LOCALLY ASH ALGEBRAS 40 [19] H. Lin and Z. Niu. Lifting KK-elements, asymptotic unitary equivalence and classification of simple C- algebras. Adv. Math., 219(5):1729 -- 1769, 2008. URL: http://dx.doi.org/10.1016/j.aim.2008.07.011, doi:10.1016/j.aim.2008.07.011. [20] H. Matui and Y. Sato. Strict comparison and Z-absorption of nuclear C-algebras. Acta Math., 209(1):179 -- 196, 2012. URL: http://dx.doi.org/10.1007/s11511-012-0084-4, doi:10.1007/s11511-012-0084-4. [21] H. Matui and Y. Sato. Decomposition rank of UHF-absorbing C-algebras. Duke Math. J., 163(14):2687 -- 2708, 2014. URL: http://dx.doi.org/10.1215/00127094-2826908, doi:10.1215/00127094-2826908. [22] Z. Niu. A classification of tracially approximate splitting interval algebras. I. The building blocks and the limit algebras. C. R. Math. Acad. Sci. Soc. R. Can., 35(1):1 -- 34, 2014. [23] Z. Niu. A classification of tracially approximate splitting interval algebras. II. Existence theorem. C. R. Math. Acad. Sci. Soc. R. Can., 37(1):1 -- 32, 2015. [24] Z. Niu. A classification of tracially approximate splitting interval algebras. III. Uniqueness theorem and isomorphism theorem. C. R. Math. Acad. Sci. Soc. R. Can., 37(2):41 -- 75, 2015. [25] N. C. Phillips. A classification theorem for nuclear purely infinite simple C-algebras. Doc. Math., 5:49 -- 114 (electronic), 2000. [26] L. Robert. Classification of inductive limits of 1-dimensional NCCW complexes. Adv. Math., 231(5):2802 -- 2836, 2012. URL: http://dx.doi.org/10.1016/j.aim.2012.07.010, doi:10.1016/j.aim.2012.07.010. [27] K. R. Strung. C-algebras of minimal dynamical systems of the product of a cantor set and an odd dimensional sphere. 03 2014. URL: http://arxiv.org/abs/1403.3136, arXiv:1403.3136. [28] A. S. Toms. K-theoretic rigidity and slow dimension growth. Invent. Math., 183(2):225 -- 244, 2011. URL: http://dx.doi.org/10.1007/s00222-010-0273-8, doi:10.1007/s00222-010-0273-8. [29] A. S. Toms and W. Winter. Minimal dynamics and the classification of C-algebras. Proc. Natl. Acad. Sci. USA, 106(40):16942 -- 16943, 2009. URL: http://dx.doi.org/10.1073/pnas.0903629106, doi:10.1073/pnas.0903629106. [30] J. Villadsen. The range of the Elliott invariant of the simple AH-algebras with slow dimension growth. K-Theory, 15(1):1 -- 12, 1998. [31] W. Winter. Nuclear dimension and Z-stability of pure C-algebras. Invent. Math., 187(2):259 -- 342, 2012. URL: http://dx.doi.org/10.1007/s00222-011-0334-7, doi:10.1007/s00222-011-0334-7. [32] W. Winter. Classifying crossed product C-algebras. 08 2013. URL: http://arxiv.org/abs/1308.5084, arXiv:1308.5084. [33] W. Winter. Localizing the Elliott conjecture at strongly self-absorbing C-algebras. J. Reine Angew. Math., 692:193 -- 231, 2014. [34] W. Winter and J. Zacharias. The nuclear dimension of C*-algebras. Adv. Math., 224(2):461 -- 498, 2010. URL: http://dx.doi.org/10.1016/j.aim.2009.12.005, doi:10.1016/j.aim.2009.12.005. Department of Mathematics, University of Toronto, Toronto, Ontario, Canada M5S 2E4. E-mail address: [email protected] Department of Mathematics, University of Puerto Rico. E-mail address: [email protected] University of Oregon and East China Normal University. E-mail address: [email protected] Department of Mathematics, University of Wyoming, Laramie, WY, 82071, USA. E-mail address: [email protected]
1106.2419	3	1106	2012-04-14T20:28:35	On a quantitative operator K-theory	[ "math.OA", "math.KT", "math.MG" ]	In this paper, we develop a quantitative K-theory for filtered C-algebras. Particularly interesting examples of filtered C-algebras include group C-algebras, crossed product C-algebras and Roe algebras. We prove a quantitative version of the six term exact sequence and a quantitative Bott periodicity. We apply the quantitative K-theory to formulate a quantitative version of the Baum-Connes conjecture and prove that the quantitative Baum-Connes conjecture holds for a large class of groups.	math.OA	math	ON QUANTITATIVE OPERATOR K-THEORY HERV´E OYONO-OYONO AND GUOLIANG YU Abstract. In this paper, we develop a quantitative K-theory for filtered C ∗-algebras. Particularly interesting examples of filtered C ∗-algebras include group C ∗-algebras, crossed product C ∗-algebras and Roe algebras. We prove a quantitative version of the six term exact sequence and a quantitative Bott periodicity. We apply the quantitative K-theory to formulate a quantitative version of the Baum-Connes conjecture and prove that the quantitative Baum- Connes conjecture holds for a large class of groups. 2 1 0 2 r p A 4 1 ] . A O h t a m [ 3 v 9 1 4 2 . 6 0 1 1 : v i X r a Contents Introduction 0. 1. Quantitative K-theory 1.1. Filtered C∗-algebras 1.2. Almost projections/unitaries 1.3. Definition of quantitative K-theory 1.4. Elementary properties of quantitative K-theory 1.5. Morita equivalence 1.6. Lipschitz homotopies 2. Controlled morphisms 2.1. Definition and main properties 2.2. Controlled exact sequences 3. Extensions of filtered C∗-algebras 3.1. Semi-split filtered extensions 3.2. Controlled boundary maps 3.3. Long exact sequence 3.4. The mapping cones 4. Controlled Bott periodicity 4.1. Tensorization in KK-theory 4.2. The controlled Bott isomorphism 4.3. The six term (λ, h)-exact sequence 5. Quantitative K-theory for crossed product C∗-algebras 5.1. Lengths and propagation 5.2. Kasparov transformation 5.3. Application to K-amenability 6. The quantitative Baum-Connes conjecture 6.1. The Rips complex 2 3 3 4 7 10 12 15 16 17 19 20 20 22 26 29 32 32 38 39 40 40 41 44 46 46 2000 Mathematics Subject Classification. 19K35,46L80,58J22. Key words and phrases. Baum-Connes Conjecture, Coarse Geometry, Group and Crossed product C ∗-algebras Novikov Conjecture, Operator Algebra K-theory, Roe Algebras. Oyono-Oyono is partially supported by the ANR "Kind" and Yu is partially supported by a grant from the U.S. National Science Foundation. 1 2 H. OYONO-OYONO AND G. YU 6.2. Quantitative statements 7. Further comments References 48 52 52 0. Introduction The receptacles of higher indices of elliptic differential operators are K-theory of C∗-algebras which encode the (large scale) geometry of the underlying spaces. The following examples are important for purpose of applications to geometry and topology. equivariant elliptic differential operators on covering space [1, 2, 5, 11]; • K-theory of group C∗-algebras is a receptacle for higher index theory of • K-theory of crossed product C∗-algebras and more generally groupoid C∗- algebras for foliations serve as receptacles for longitudinally elliptic oper- ators [3, 4]; • the higher indices of elliptic operators on noncompact Riemannian mani- folds live in K-theory of Roe algebras [15]. The local nature of differential operators implies that these higher indices can be defined in term of idempotents and invertible elements with finite propagation. Using homotopy invariance of the K-theory for C∗-algebras, these higher indices give rise to topological invariants. In the context of Roe algebras, a quantitative operator K-theory was introduced to compute the higher indices of elliptic operators for noncompact spaces with fi- nite asymptotic dimension [19]. The aim of this paper is to develop a quantitative K-theory for general C∗-algebras equipped with a filtration. The filtration struc- ture allows us to define the concept of propagation. Examples of C∗-algebras with filtrations include group C∗-algebras, crossed product C∗-algebras and Roe alge- bras. The quantitative K-theory for C∗-algebras with filtrations is then defined in terms of homotopy of quasi-projections and quasi-unitaries with propagation and norm controls. We introduce controlled morphisms to study quantitative operator K-theory. In particular, we derive a quantitative version of the six term exact sequence. In the case of crossed product algebras, we also define a quantitative ver- sion of the Kasparov transformation compatible with Kasparov product. We end this paper by using the quantitative K-theory to formulate a quantitative version of the Baum-Connes conjecture and prove it for a large class of groups. This paper is organized as follows: In section 1, we collect a few notations and definitions including the concept of filtered C∗-algebras. We use the concepts of almost unitary and almost projection to define a quantitative K-theory for filtered C∗-algebras and we study its elementary properties. In section 2, we introduce the notion of controlled morphism in quantitative K-theory. Section 3 is devoted to extensions of filtered C∗-algebras and to a controlled exact sequence for quantitative K-theory. In section 4, we prove a controlled version of the Bott periodicity and as a consequence, we obtain a controlled version of the six-term exact sequence in K-theory. In section 5, we apply KK-theory to study the quantitative K-theory of crossed product C∗-algebras and discuss its application to K-amenability. Finally ON A QUANTITATIVE OPERATOR K-THEORY 3 in section 8, we formulate a quantitative Baum-Connes conjecture and prove the quantitative Baum-Connes conjecture for a large class of groups. 1. Quantitative K-theory In this section, we introduce a notion of quantitative K-theory for C∗-algebras with a filtration. Let us fix first some notations about C∗-algebras we shall use throughout this paper. b1 . . . bk   of Mn1+···+nk (B). • If B is a C∗-algebra and if b1, . . . , bk are respectively elements of Mn1(B), . . . , Mnk (B), we denote by diag(b1, . . . , bk) the block diagonal matrix   • If X is a locally compact space and B is a C∗-algebra, we denote by C0(X, B) the C∗-algebra of B-valued continuous functions on X vanishing at infinity. The special cases of X = (0, 1], X = [0, 1), X = (0, 1) and X = [0, 1], will be respectively denoted by CB, B[0, 1), SB and B[0, 1]. • For a separable Hilbert space H, we denote by K(H) the C∗-algebra of • If A and B are C∗-algebras, we will denote by A ⊗ B their spatial tensor compact operators on H. product. 1.1. Filtered C∗-algebras. Definition 1.1. A filtered C∗-algebra A is a C∗-algebra equipped with a family (Ar)r>0 of linear subspaces indexed by positive numbers such that: • Ar ⊂ Ar′ if r 6 r′; • Ar is stable by involution; • Ar · Ar′ ⊂ Ar+r′ ; • the subalgebra [r>0 Ar is dense in A. If A is unital, we also require that the identity 1 is an element of Ar for every positive number r. The elements of Ar are said to have propagation r. • Let A and A′ be respectively C∗-algebras filtered by (Ar)r>0 and (A′r)r>0. A homomorphism of C∗ -algebras φ : A−→A′ is a filtered homomorphism (or a homomorphism of filtered C∗-algebras) if φ(Ar) ⊂ A′r for any positive number r. • If A is a filtered C∗-algebra and X is a locally compact space, then In particular the C0(X, A) is a C∗-algebra filtered by (C0(X, Ar))r>0. algebras CA, A[0, 1], A[0, 1) and SA are filtered C∗-algebras. • If A is a non unital filtered C∗-algebra, then its unitarization eA is filtered by (Ar + C)r>0. We define for A non-unital the homomorphism for a ∈ A and z ∈ C. ρA : eA → C; a + z 7→ z Prominent examples of filtered C∗-algebra are provided by Roe algebras asso- ciated to proper metric spaces, i.e. metric spaces such that closed balls of given radius are compact. Recall that for such a metric space (X, d), a X-module is a Hilbert space HX together with a ∗-representation ρX of C0(X) in HX (we shall 4 H. OYONO-OYONO AND G. YU write f instead of ρX (f )). If the representation is non-degenerate, the X-module is said to be non-degenerate. A X-module is called standard if no non-zero function of C0(X) acts as a compact operator on HX . The following concepts were introduced by Roe in his work on index theory of elliptic operators on noncompact spaces [15]. Definition 1.2. Let HX be a standard non-degenerate X-module and let T be a bounded operator on HX . (i) The support of T is the complement of the open subset of X × X {(x, y) ∈ X × X s.t. there exist f and g in C0(X) satisfying f (x) 6= 0, g(y) 6= 0 and f · T · g = 0}. (ii) The operator T is said to have finite propagation (in this case propagation less than r) if there exists a real r such that for any x and y in X with d(x, y) > r, then (x, y) is not in the support of T . (iii) The operator T is said to be locally compact if f · T and T · f are compact for any f in C0(X). We then define C[X] as the set of locally compact and finite propagation bounded operators of HX , and for every r > 0, we define C[X]r as the set of element of C[X] with propagation less than r. We clearly have C[X]r · C[X]r′ ⊂ C[X]r+r′. We can check that up to (non- canonical) isomorphism, C[X] does not depend on the choice of HX . Definition 1.3. The Roe algebra C∗(X) is the norm closure of C[X] in the algebra L(HX ) of bounded operators on HX . The Roe algebra in then filtered by (C[X]r)r>0. Although C∗(X) is not canonically defined, it was proved in [9] that up to canonical isomorphisms, its K-theory does not depend on the choice of a non- degenerate standard X-module. Furthermore, K∗(C∗(X)) is the natural receptacle for higher indices of elliptic operators with support on X [15]. If X has bounded geometry, then the Roe algebra admits a maximal version [7] filtered by (C[X]r)r>0. Other important examples are reduced and maximal crossed product of a C∗-algebra by an action of a discrete group by automorphisms. These examples will be studied in detail in section 5. 1.2. Almost projections/unitaries. Let A be a unital filtered C∗-algebra. For any positive numbers r and ε, we call • an element u in A a ε-r-unitary if u belongs to Ar, ku∗ · u − 1k < ε and ku · u∗ − 1k < ε. The set of ε-r-unitaries on A will be denoted by Uε,r(A). • an element p in A a ε-r-projection if p belongs to Ar, p = p∗ and kp2−pk < ε. The set of ε-r-projections on A will be denoted by Pε,r(A). For n integer, we set Uε,r n (A) = Uε,r(Mn(A)) and Pε,r n (A) = Pε,r(Mn(A)). For any unital filtered C∗-algebra A, any positive numbers ε and r and any positive integer n, we consider inclusions and n (A) ֒→ Pε,r Pε,r n (A) ֒→ Uε,r Uε,r n+1(A); p 7→(cid:18)p 0 0(cid:19) n+1(A); u 7→(cid:18)u 0 1(cid:19) . 0 0 ON A QUANTITATIVE OPERATOR K-THEORY 5 This allows us to define and Uε,r ∞ (A) = [n∈N ∞ (A) = [n∈N Pε,r Uε,r n (A) Pε,r n (A). Remark 1.4. Let r and ε be positive numbers with ε < 1/4; (i) If p is an ε-r-projection in A, then the spectrum of p is included in 2 2 , 1+√1+4ε (cid:17) ∪(cid:16) 1+√1−4ε 1 − ε < kuk < 1 + ε/2, 1 − ε/2 < ku−1k < 1 + ε, ku∗ − u−1k < (1 + ε)ε. (cid:16) 1−√1+4ε 2 , 1−√1−4ε 2 (ii) If u is an ε-r-unitary in A, then (cid:17) and thus kpk < 1 + ε. and κ0,ε(t) = 1 if t > 1+√1−4ε (iii) Let κ0,ε : R → R be a continuous function such that κ0,ε(t) = 0 if t 6 1−√1−4ε . If p is an ε-r-projection in A, then κ0,ε(p) is a projection such that kp − κ0,ε(p)k < 2ε which moreover does not depends on the choice of κ0,ε. From now on, we shall denote this projection by κ0(p). 2 2 (iv) If u is an ε-r-unitary in A, set κ1(u) = u(u∗u)−1/2. Then κ1(u) is a unitary such that ku − κ1(u)k < ε. 1 − 2ε, then κ0(p) and q are homotopic projections [18, Chapter 5]. (v) If p is an ε-r-projection in A and q is a projection in A such that kp− qk < (vi) If u and v are ε-r-unitaries in A, then uv is an ε(2 + ε)-2r-unitary in A. Definition 1.5. Let A be a C∗-algebra filtered by (Ar)r>0. • Let p0 and p1 be ε-r-projections. We say that p0 and p1 are homotopic ε- r-projections if there exists a ε-r-projection q in A[0, 1] such that q(0) = p0 and q(1) = p1. In this case, q is called a homotopy of ε-r-projections in A and will be denoted by (qt)t∈[0,1]. • If A is unital, let u0 and u1 be ε-r-unitaries. We say that u0 and u1 are homotopic ε-r-unitaries if there exists an ε-r-unitary v in A[0, 1] such that v(0) = u0 and v(1) = u1. In this case, v is called a homotopy of ε-r-unitaries in A and will be denoted by (vt)t∈[0,1]. Example 1.6. Let p be a ε-projection in a filtered unital C∗-algebra A. Set ct = cos πt/2 and st = sin πt/2 for t ∈ [0, 1] and let us considerer the homotopy of projec- tions (ht)t∈[0,1] with ht =(cid:18) c2 t (cid:19) in M2(C) between diag(1, 0) and diag(0, 1). Set (qt)t∈[0,1] = (diag(p, 0)+(1−p)⊗ht)t∈[0,1]. Since q2 t (p2−p)⊗I2, we see that (qt)t∈[0,1] is a homotopy of ε-r-projections between diag(1, 0) and diag(p, 1− p) in M2(A). t −qt = s2 ctst s2 t ctst Next result will be frequently used throughout the paper and is quite easy to prove. Lemma 1.7. Let A be a C∗-algebra filtered by (Ar)r>0. 6 H. OYONO-OYONO AND G. YU 4 (i) If p is an ε-r-projection in A and q is a self-adjoint element of Ar such that kp − qk < ε−kp2−pk , then q is ε-r-projection . In particular, if p is an ε-r-projection in A and if q is a self-adjoint element in Ar such that kp− qk < ε, then q is a 5ε-r-projection in A and p and q are connected by a homotopy of 5ε-r-projections. (ii) If A is unital and if u is an ε-r-unitary and v is an element of Ar such that ku − vk < ε−ku∗u−1k , then v is an ε-r-unitary . In particular, if u is an ε-r-unitary and v is an element of Ar such that ku − vk < ε, then v is an 4ε-r-unitary in A and u and v are connected by a homotopy of 4ε-r-unitaries. 3 Lemma 1.8. There exists a real λ > 4 such that for any positive number ε with ε < 1/λ, any positive real r, any ε-r-projection p and ε-r-unitary W in a filtered unital C∗-algebra A, the following assertions hold: (i) W pW ∗ is a λε-3r-projection of A; (ii) diag(W pW ∗, 1) and diag(p, 1) are homotopic λε-3r-projections. Proof. The first point is straightforward to check from remark 1.4. For the second point, with notations of example 1.6, use the homotopy of ε-r-unitaries (cid:16) W c2 t +s2 t (W−1)stct W s2 (W−1)stct t +c2 t (cid:17)t∈[0,1] =(cid:0)(cid:0) ct −st st ct (cid:1) · diag(W, 1) · ( ct st −st ct )(cid:1)t∈[0,1] to connect by conjugation diag(W pW ∗, 1) to diag(p, W W ∗) and then connect to diag(p, 1) by a ray. (cid:3) Recall that if two projections in a unital C∗-algebra are close enough in norm, then there are conjugated by a canonical unitary. To state a similar result in term of ε-r-projections and ε-r-unitaries, we will need the definition of a control pair. Definition 1.9. A control pair is a pair (λ, h), where • λ > 1; • h : (0, 1 4λ ) → (0, +∞); ε 7→ hε is a map such that there exists a non- increasing map g : (0, 1 4λ ) → (0, +∞), with h 6 g. Lemma 1.10. There exists a control pair (λ, h) such that the following holds: for every positive number r, any ε in (0, 1 4λ ) and any ε-r-projections p and q of a filtered unital C∗-algebra A satisfying kp−qk < 1/16, there exists an λε-hεr-unitary W in A such that kW pW ∗ − qk 6 λε. Proof. We follow the proof of [18, Proposition 5.2.6]. If we set • then z = (2κ0(p) − 1)(2κ0(q) − 1) + 1, kz − 2k 6 2kκ0(p) − κ0(q)k 6 8ε + 2kp − qk and hence z is invertible for ε < 1/16. • Moreover, if we set U = zz−1 and since zκ0(q) = κ0(p)z, then we have κ0(q) = U κ0(p)U∗. ON A QUANTITATIVE OPERATOR K-THEORY 7 Let us define z′ = (2p − 1)(2q − 1) + 1. Then we have kz − z′k 6 9ε and kz′k 6 3. If ε is small enough, then kz′∗z′ − 4k 6 2 and hence the spectrum of z′∗z′ is in [2, 6]. Let us consider the expansion in power serie Pk∈N aktk of t 7→ (1 + t)−1/2 on (0, 1) and let nε be the smallest integer such that Pnε6k ak/2k 6 ε. Let us set then W = z′/2Pnε )k. Then for a suitable λ (not depending on A, p, q or ε), we get that W is a λε-(4nε + 2)r-unitary which satisfies the required condition. (cid:3) k=0 ak( z′∗z′−4 4 Remark 1.11. The order of h when ε goes to zero in lemma 1.10 is Cε−3/2 for some constant C. 1.3. Definition of quantitative K-theory. For a unital filtered C∗-algebra A, we define the following equivalence relations on Pε,r • if p and q are elements of Pε,r ∞ (A) × N and on Uε,r (q, l′) if there exists a positive integer k and an element h of Pε,r such that h(0) = diag(p, Ik+l′ ) and h(1) = diag(q, Ik+l). ∞ (A), l and l′ are positive integers, (p, l) ∼ ∞ (A[0, 1]) ∞ (A), u ∼ v if there exists an element h of • if u and v are elements of Uε,r Uε,r ∞ (A[0, 1]) such that h(0) = u and h(1) = v. If p is an element of Pε,r class of (p, l) modulo ∼ and if u is an element of Uε,r equivalence class modulo ∼. Definition 1.12. Let r and ε be positive numbers with ε < 1/4. We define: ∞ (A) and l is an integer, we denote by [p, l]ε,r the equivalence ∞ (A) we denote by [u]ε,r its ∞ (A): ∞ (A) × N/ ∼ for A unital and (i) K ε,r 0 (A) = Pε,r K ε,r 0 (A) = {[p, l]ε,r ∈ Pε,r( A) × N/ ∼ such that dim κ0(ρA(p)) = l} for A non unital. 1 (A) = Uε,r ∞ ( A)/ ∼ (with A = A if A is already unital). (ii) K ε,r Remark 1.13. We shall see in lemma 1.24 that as it is the case for K-theory, K ε,r ∗ (•) can indeed be defined in a uniform way for unital and non-unital filtered C∗-algebras. It is straightforward to check that for any unital filtered C∗-algebra A, if p is an ε-r-projection in A and u is an ε-r-unitary in A, then diag(p, 0) and diag(0, p) are homotopic ε-r-projections in M2(A) and diag(u, 1) and diag(1, u) are homotopic ε-r-unitaries in M2(A). Thus we obtain the following: Lemma 1.14. Let A be a filtered C∗-algebra. Then K ε,r equipped with a structure of abelian semi-group such that 0 (A) and K ε,r 1 (A) are and [p, l]ε,r + [p′, l′]ε,r = [diag(p, p′), l + l′]ε,r [u]ε,r + [u′]ε,r = [diag(u, v)]ε,r, for any [p, l]ε,r and [p′, l′]ε,r in K ε,r 0 (A) and any [u]ε,r and [u′]ε,r in K ε,r 1 (A). According to example 1.6, for every unital filtered C∗-algebra A, any ε-r-projection p in Mn(A) and any integer l with n > l, we see that [In − p, n − l]ε,r is an inverse for [p, l]ε,r. Hence we obtain: Lemma 1.15. If A is a filtered C∗-algebra, then K ε,r 0 (A) is an abelian group. Although K ε,r 1 (A) is not a group, it is very close to be one. 8 H. OYONO-OYONO AND G. YU Lemma 1.16. Let A be a filtered C∗-algebra. Then for any ε-r-unitary u in Mn( A) (with A = A if A is already unital), we have [u]3ε,2r + [u∗]3ε,2r = 0 in K 3ε,2r (A). 1 Proof. If u is an ε-r-unitary in a unital filtered C∗-algebra A, then according to −st ct )(cid:1)t∈[0,1] point (vi) of remark 1.4, we see that(cid:0)diag(1, u)(cid:0) ct −st is a homotopy of 3ε-2r-unitaries between diag(u, u∗) and diag(uu∗, 1). Since kuu∗− 1k < ε, we deduce from lemma 1.7 that uu∗ and 1 are homotopic 3ε-2r-unitaries and hence we get the lemma. st ct (cid:1) · diag(1, u∗) · ( ct st (cid:3) Remark 1.17. According to lemma 1.16, if we define the equivalence relation on Uε,r ∞ (A) to be homotopy within U3ε,2r 1 (A) can be endowed with an abelian group structure. ∞ (A), then K ε,r We have for any filtered C∗-algebra A and any positive numbers r, r′, ε and ε′ with ε 6 ε′ < 1/4 and r 6 r′ natural semi-group homomorphisms 0 1 • ιε,r : K ε,r • ιε,r : K ε,r • ιε,r ∗ = ιε,r • ιε,ε′,r,r′ • ιε,ε′,r,r′ • ιε,ε′,r,r′ 0 (A)−→K0(A); [p, l]ε,r 7→ [κ0(p)] − [Il]; 1 (A)−→K1(A); [u]ε,r 7→ [u]; 0 ⊕ ιε,r 1 ; 0 (A)−→K ε′,r′ : K ε,r 1 (A)−→K ε′,r′ : K ε,r ⊕ ιε,ε′,r,r′ = ιε,ε′,r,r′ (A); [p, l]ε,r 7→ [p, l]ε′,r′; (A); [u]ε,r 7→ [u]ε′,r′. ∗ 1 1 1 0 1 0 If some of the indices r, r′ or ε, ε′ are equal, we shall not repeat it in ιε,ε′,r,r′ Remark 1.18. Let p0 and p1 be two ε-r-projections in a filtered C∗-algebra such that κ0(p0) and κ0(p1) are homotopic projections. Then for any ε in (0, 1/4), this homotopy can be approximated for some r′ by a ε-r′-projection. Hence, using point (iii) of remark 1.4, there exists a homotopy (qt)t∈[0,1] of ε-r′ projections in A such that kp0 − q0k < 3ε and kp1 − q1k < 3ε. We can indeed assume that r′ > r and thus by lemma 1.7, we get that p0 and p1 are homotopic as 15ε-r′-projections. Proceeding in the same way for the odd case we eventually obtain: ∗ . there exists λ > 1 such that for any filtered C∗-algebra A, any ε ∈ (0, 1 any positive number r, the following holds: 4λ ) and ∗ (A) such that ιε,r Let x and x′ be elements in K ε,r ∗ (x) = ιε,r there exists a positive number r′ with r′ > r such that ιε,λε,r,r′ K λε,r′ ∗ Lemma 1.19. Let p be a matrix in Mn(C) such that p = p∗ and kp2 − pk < ε for some ε in (0, 1/4). Then there is a continuous path (pt)t∈[0,1] in Mn(C) such that ∗ (x′) in K∗(A), then (x) = ιε,λε,r,r′ (x′) in (A). ∗ ∗ • p0 = p; • p1 = Ik with k = dim κ0(p); • p∗t = pt and kp2 t − ptk < ε for every t in [0, 1]. Proof. The selfadjoint matrix p satisfies kp2 − pk < ε if and only if the eigenvalues of p satisfy the inequality i.e. −ε < λ2 − λ < ε, λ ∈(cid:18) 1 − √1 + 4ε 2 , 1 − √1 − 4ε 2 (cid:19)[(cid:18)√1 − 4ε + 1 2 , √1 + 4ε + 1 2 (cid:19) . ON A QUANTITATIVE OPERATOR K-THEORY 9 Let λ1, . . . , λk be the eigenvalues of p lying in (cid:16) 1−√1+4ε λk+1, . . . , λn be the eigenvalues of p lying in (cid:16) √1−4ε+1 t ∈ [0, 1] 2 2 , √1+4ε+1 2 , 1−√1−4ε 2 (cid:17) and let (cid:17). We set for • λi,t = tλi for i = 1, . . . , k; • λi,t = tλi + 1 − t for i = k + 1, . . . , n. Since λ 7→ λ2 − λ is decreasing on (cid:16) 1−√1+4ε (cid:16)√1−4ε+1 (cid:17) then, √1+4ε+1 2 2 2 , −ε < λ2 i,t − λi,t < ε , 1−√1−4ε 2 (cid:17) and increasing on for all t in [0, 1] and i = 1, . . . , n. If we set pt = u · diag(λ1,t, . . . , λn,t) · u∗ where u is a unitary matrix of Mn(C) such that p = u · diag(λ1, . . . , λn) · u∗, then • p0 = p; • p1 = κ0(p); • p∗t = pt and kp2 t − ptk < ε for every t in [0, 1]. Since there is a homotopy of projections in Mn(C) between κ0(p) and Ik with k = dim κ0(p), we get the result. (cid:3) As a consequence we obtain: Corollary 1.20. For any positive numbers r and ε with ε < 1/4, then K ε,r 0 (C) → Z; [p, l]ε,r 7→ dim κ0(p) − l is an isomorphism. Lemma 1.21. Let u be a matrix in Mn(C) such that ku∗u − Ink < ε and kuu∗ − Ink < ε for ε in (0, 1/4). Then there is a continuous path (ut)t∈[0,1] in Mn(C) such that • u0 = u; • u1 = In; • ku∗t ut − Ink < ε and kutu∗t − Ink < ε for every t in [0, 1]. Proof. Since u is invertible, u∗u and uu∗ have the same eigenvalues λ1, . . . , λn, and thus ku∗t ut − Ink < ε and kutu∗t − Ink < ε if and only if λi ∈ (1 − ε, 1 + ε) for i = 1, . . . , n. Let us set 1 n , . . . , λ−t/2 ) · w∗ where w is a unitary matrix of Mn(C) • ht = w · diag(λ−t/2 such that u∗u = w · diag(λ1, . . . , λn) · w∗; • vt = u · ht for all t ∈ [0, 1]. Then v∗t vt = w · diag(λ1−t Since λ1−t i −1 < ε for all all t ∈ [0, 1], we get that kv∗t vt−Ink < ε and kvtv∗t −Ink < ε for every t in [0, 1]. The matrix v1 is unitary and the result then follows from path-connectness of Un(C). (cid:3) n ) · w∗. , . . . , λ1−t 1 As a consequence we obtain: Corollary 1.22. For any positive numbers r and ε with ε < 1/4, then we have K ε,r 1 (C) = {0}. 10 H. OYONO-OYONO AND G. YU 1.4. Elementary properties of quantitative K-theory. Let A1 and A2 be two unital C∗-algebras respectively filtered by (A1,r)r>0 and (A2,r)r>0 and consider ∞ (A1 ⊕ A2) ∼= A1 ⊕ A2 filtered by (A1,r ⊕ A2,r)r>0. Since we have identifications Pε,r Pε,r ∞ (A1) × Pε,r ∞ (A2) induced by the inclusions A1 ֒→ A1 ⊕ A2 and A2 ֒→ A1 ⊕ A2, we see that we have isomorphisms K ε,r 0 (A1)⊕K ε,r 1 (A1⊕A2). Lemma 1.23. Let A be a filtered non unital C∗-algebra and let ε and r be positive numbers with ε < 1/4. We have a natural splitting ∞ (A1 ⊕ A2) ∼= Uε,r 0 (A1⊕A2) and K ε,r ∞ (A2) and Uε,r 0 (A2) ∼−→ K ε,r ∞ (A1) × Uε,r 1 (A1)⊕K ε,r 1 (A2) ∼−→ K ε,r Proof. Viewing A as a subalgebra of A, the group homomorphisms 0 ( A) ∼=−→ K ε,r K ε,r 0 (A) ⊕ Z. K ε,r 0 ( A) −→ K ε,r [p, l]ε,r 0 (A) ⊕ Z 7→ ([p, dim κ0(ρA(p))]ε,r, dim κ0(ρA(p)) − l) and 0 (A) ⊕ Z −→ K ε,r K ε,r 7→ (cid:20)(cid:18)p ([p, l]ε,r, k − k′) 0 ( A) 0 0 Ik(cid:19) , l + k′(cid:21)ε,r are inverse one of the other. (cid:3) Let us set A+ = A ⊕ C equipped with the multiplication (a, x) · (b, y) = (ab + xb + ya, xy) for a and b in A and x and y in C. Notice that • A+ is isomorphic to A ⊕ C with the algebra structure provided by the direct sum if A is unital; • A+ = A if A is not unital. Let us define also ρA in the unital case by ρA : A+ → C; (a, x) 7→ x. We know that in usual K-theory, we can equivalently define for A unital the Z2-graded group K∗(A) as A+ by K0(A) = ker ρA,∗ : K0(A+) → K0(C) ∼= Z and K1(A) = K1(A+). ε,r(A). If Let us check that this is also the case for our Z2-graded semi-groups K∗ the C∗-algebra A is filtered by (Ar)r>0, then A+ is filtered by (Ar + C)r>0. Let us define for a unital filtered algebra A K′0 ε,r(A) = {[p, l]ε,r ∈ Pε,r(A+) × N/ ∼ such that dim κ0(ρA(p)) = l} and K′1 ε,r(A) = Uε,r(A+)/ ∼ . Proceeding as we did in the proof of lemma 1.23, we obtain a natural splitting 0 (A+) ∼=−→ K′0 K ε,r ε,r(A) ⊕ Z. But then, using the identification A+ ∼= A ⊕ C and in view of lemmas 1.19 and 1.21, we get ON A QUANTITATIVE OPERATOR K-THEORY 11 Lemma 1.24. The Z2-graded semi-groups K ε,r isomorphic. ∗ (A) and K′ ∗ ε,r(A) are naturally This allows us to state functoriallity properties for quantitative K-theory. If φ : A → B is a homomorphism of unital filtered C∗-algebras, then since φ preserve ε-r-projections and ε-r-unitaries, it obviously induces for any positive number r and any ε ∈ (0, 1/4) a semi-group homomorphism ∗ (A) −→ K ε,r ∗ (B). : K ε,r φε,r ∗ In the non unital case, we can extend any homomorphism φ : A → B to a homomor- phism φ+ : A+ → B+ of unital filtered C∗-algebras and then we use lemmas 1.23 and 1.24 to define φε,r ∗ (B). Hence, for any positive number r and ∗ any ε ∈ (0, 1/4), we get that K ε,r 1 (•)) is a covariant additive functor from the category of filtered C∗-algebras (together with filtered homomorphism) to the category of abelian groups (resp. semi-groups). 0 (•) (resp. K ε,r ∗ (A) −→ K ε,r : K ε,r Definition 1.25. (i) Let A and B be filtered C∗-algebras. Then two homomorphisms of filtered C∗-algebras ψ0 : A → B and ψ1 : A → B are homotopic if there exists a path of homomorphisms of filtered C∗-algebras ψt : A → B for 0 6 t 6 1 between ψ0 and ψ1 and such that t 7→ ψt is continuous for the pointwise norm convergence. (ii) A filtered C∗-algebra A is said to be contractible if the identity map and the zero map of A are homotopic. Example 1.26. If A is a filtered C∗-algebra A, then the cone of A is a contractible filtered C∗-algebra. CA = {f ∈ C([0, 1], A) such that f (0) = 0} We have then the following obvious result: Lemma 1.27. If φ : A → B and φ′ : A → B are two homotopic homomorphisms of filtered C∗-algebras, then φε,r for every positive numbers ε and r with ε < 1/4. In particular, if A is a contractible filtered C∗-algebra, then K ε,r ∗ (A) = 0 for every positive numbers ε and r with ε < 1/4. ∗ = φ′ε,r ∗ Let A be a C∗-algebra filtered by (Ar)r>0 and let (Bk)k∈N be an increasing sequence of C∗-subalgebras of A such that [k∈N Bk is dense in A. Assume that Sr>0 Bk ∩ Ar is dense in Bk for every integer k. Then for every integer k, the C∗-algebra Bk is filtered by (Bk ∩ Ar)r>0. If A is unital, then Bk is unital for some k, and thus we will assume without loss of generality that Bk is unital for every integer k. Proposition 1.28. Let A be a unital C∗-algebra filtered by (Ar)r>0 and let (Bk)k∈N be an increasing sequence of C∗-subalgebra of A such that (Bk ∩ Ar) is dense in Bk for every integer k, (Bk ∩ Ar) is dense in Ar for every positive number r. • [r>0 • [k∈N Then the Z2-graded semi-groups K ε,r ∗ (A) and lim k K ε,r ∗ (Bk) are isomorphic. 12 H. OYONO-OYONO AND G. YU Proof. In particular, we see that [k∈N Bk is dense in A. Let us denote by Υ∗,ε,r : lim k K ε,r ∗ (Bk) → K ε,r ∗ (A) 12 . Since [k∈N the homomorphism of semi-group induced by the family of inclusions Bk ֒→ A where k runs through integers. We give the proof in the even case, the odd case being analogous. Let p be an element of Pε,r n (A) and let δ = kp2 − pk > 0 and choose α < ε−δ (Bk ∩ Ar) is dense in Ar, there is an integer k and a selfadjoint element q of Mn(Bk ∩ Ar) such that kp − qk < α. According to lemma 1.19, q is a ε-r projection. Let q′ be another selfadjoint element of Mn(Bk ∩ Ar) such that kp − q′k < α. Then kq − q′k < 2α and if we set qt = (1 − t)q + tq′ for t ∈ [0, 1], then kq2 t − qtqk + kqtq − q2k + kq2 − qk + kq − qtk t − qtk 6 kq2 6 kqt − qk(kqtk + kqk + 1) + 4α + δ 6 12α + δ < ε, and thus q and q′ are homotopic in Pε,r some Mn(Bk ∩ Ar) satisfying kq − pk < kp2−pk image of [q, l]ε,r in lim k inverse for Υ0,ε,r. We proceed similarly in the odd case. n (A) and q in , we define Υ′0,ε,r([p, l]ε,r) to be the K ε,r ∗ (Bk). Then Υ′0,ε,r is a group homomorphism and is an n (Bk). Therefore, for p ∈ Pε,r (cid:3) 12 1.5. Morita equivalence. For any unital filtered algebra A, we get an identifi- cation between Pε,r ∞ (Mk(A)) and Pε,r ∞ (A). This identification gives rise to a natural group isomorphism between K ε,r 0 (A) and K ε,r 0 (Mk(A)), and this isomorphism is induced by the inclusion of C∗-algebras nk (A) and therefore between Pε,r n (Mk(A)) and Pε,r ιA : A ֒→ Mk(A); a 7→ diag(a, 0). Namely, if we set e1,1 = diag(1, 0, . . . , 0) ∈ Mk(C), definition of the functoriality yields ιε,r A,∗ for any p in Pε,r [p, l]ε,r = [p ⊗ e1,1 + Il ⊗ (Ik − e1,1), l]ε,r ∈ K ε,r n (A) and any integer l with l 6 n. We can verify that 0 (Mk(A)) (ιε,r A,∗ )−1[q, l]ε,r = [q, kl]ε,r for any q in Pε,r n (Mk(A)) and any integer l with l 6 n, where on the right hand side of the equality, the matrix q of Mn(Mk(A)) is viewed as a matrix of Mnk(A). ∞ (Mk(A)) 1 (A) and 1 (Mk(A)). This isomorphism is also induced by the inclusion ιA and we have In a similar way, we obtain in the odd case an identification between Uε,r ∞ (A) providing a natural semi-group isomorphism between K ε,r and Uε,r K ε,r for any x in Uε,r ιA,∗[x]ε,r = [x ⊗ e1,1 + In ⊗ (Ik − e1,1)]ε,r ∈ K ε,r n (A). 1 (Mk(A)) Let us deal now with the non-unital case. For usual K-theory, Morita equivalence for non-unital C∗-algebra can be deduced from the unital case by using the six- term exact sequence associated to the split extension 0 → A → A → C → 0. But for quantitative K-theory this splitting only gives rise (in term of section 2.1) to a controlled isomorphism (see corollary 4.9). In order to really have a genuine ON A QUANTITATIVE OPERATOR K-THEORY 13 isomorphism, we have to go through the tedious following computation. If B is a non-unital C∗-algebra, let us identify Mk( B) with Mk(B) ⊕ Mk(C) equipped with the product (b, λ) · (b′, λ′) = (bb′ + λb′ + bλ′, λλ′) for b and b′ in Mk(B) and λ and λ′ in Mk(C). Under this identification, if A is not unital, let us check that the semi-group homomorphism Φ1 : K ε,r 1 ( A) → K ε,r 1 ( ^Mk(A)); [(x, λ)]ε,r 7→ [(x ⊗ e1,1, λ]ε,r induced by the inclusion ιA is invertible with inverse given by the composition Ψ1 : K ε,r 1 ( ^Mk(A)) → K ε,r 1 (Mk( A)) ∼=→ K ε,r 1 ( A), where the first homomorphism of the composition is induced by the inclusion ^Mk(A) → Mk( A); (a, z) 7→ (a, zIk). Let (x, λ) be an element of Uε,r n ( A), with x ∈ Mn(A) and λ ∈ Mn(C). Then Ψ1 ◦ Φ1[(x, λ)]ε,r = [(x ⊗ e1,1, λ ⊗ Ik)]ε,r, where we use the identification Mnk(C) ∼= Mn(C)⊗ Mk(C) to see x⊗ e1,1 and λ⊗ Ik respectively as matrices in Mnk(A) and Mnk(C). According to lemma 1.21, as a ε-r-unitary of Mn(C), λ is homotopic to In. Hence [(x ⊗ e1,1, λ ⊗ Ik)]ε,r = [(x ⊗ e1,1, λ ⊗ e1,1 + In⊗Ik−1)] and from this we get that Ψ1 ◦ Φ1 is induced in K-theory by the inclusion map A ֒→ Mk( A); a 7→ diag(a, 0) which is the identity homomorphism (according to the unital case). Conversely, let (y, λ) be an element in Uε,r n ( ^Mk(A)) with and λ ∈ Mn(C). Then y ∈ Mn(Mk(A)) ∼= Mn(A) ⊗ Mk(C) Φ1 ◦ Ψ1[(y, λ)]ε,r = [(y ⊗ e1,1, λ ⊗ Ik)]ε,r, where Let • y ⊗ e1,1 belongs to Mn(Mk(A))⊗ Mk(C) ∼= Mn(A)⊗ Mk(C)⊗ Mk(C) (the first two factors provide the copy of Mn(Mk(A)) where y lies in and e1,1 lies in the last factor). • λ⊗Ik belongs to the algebra Mn(Mk(C)) ∼= Mn(C)⊗Mk(C) that multiplies Mn(A) ⊗ Mk(C) ⊗ Mk(C) on the first two factors. σ : Mn(A) ⊗ Mk(C) ⊗ Mk(C) → Mn(A) ⊗ Mk(C) ⊗ Mk(C) be the C∗-algebra homomorphism induced by the flip of Mk(C)⊗ Mk(C). This flip can be realized by conjugation of a unitary U in Mk(C) ⊗ Mk(C) ∼= Mk2(C). Let (Ut)t∈[0,1] be a homotopy in Uk2 (C) between U and Ik2 . Let us define A = {(x, z⊗Ik); x ∈ Mn(A)⊗Mk(C)⊗Mk(C), z ∈ Mn(C)} ⊂ Mn( ^Mk(A))⊗Mk(C), where z ⊗ Ik is viewed as z ⊗ Ik ⊗ Ik in Mn( ^Mk(A)) ⊗ Mk(C) ∼= Mn(C) ⊗ ^Mk(A) ⊗ Mk(C). Then for any t ∈ [0, 1], A → A; (x, z ⊗ Ik) 7→ ((In ⊗ Ut) · x · (In ⊗ Ut)−1, z ⊗ Ik) 14 H. OYONO-OYONO AND G. YU is an automorphism of C∗-algebra. Hence, (cid:0)(In ⊗ Ut) · (y ⊗ e1,1) · (In ⊗ U−1 nk ( ^Mk(A)) between (y ⊗ e1,1, λ ⊗ Ik) and (σ(y ⊗ e1,1), λ ⊗ Ik). The is a path in Uε,r range of σ(y ⊗ e1,1) being in the range of the projection In ⊗ e1,1 ⊗ Ik, we have an orthogonal sum decomposition ), λ ⊗ Ik(cid:1)t∈[0,1] t (σ(y ⊗ e1,1), λ ⊗ Ik) = (σ(y ⊗ e1,1), λ ⊗ e1,1) + (0, λ ⊗ (Ik − e1,1)) (recall that λ ⊗ e1,1 and λ ⊗ (Ik − e1,1) multiply Mn(A) ⊗ Mk(C) ⊗ Mk(C) on the first two factors). By lemma 1.21, λ is homotopic to In in Uε,r n (C) and thus (σ(y ⊗ e1,1), λ ⊗ Ik) is homotopic to (σ(y ⊗ e1,1), λ ⊗ e1,1) + (0, In ⊗ (Ik − e1,1)) in nk ( ^Mk(A))) which can be viewed as Uε,r diag((y, λ), (0, Ik(k−1)) in Mk(Mn( ^Mk(A)). From this we deduce that [(y, λ)]ε,r = [(y ⊗ e1,1, λ ⊗ Ik)]ε,r in 1 ( ^Mk(A)). K ε,r For the even case, by an analogous computation, we can check that the group homomorphisms K ε,r 0 ( A) → K ε,r 0 ( ^Mk(A)); [(p, q), l)]ε,r 7→ [(p ⊗ e1,1), q, l]ε,r and 0 ( A); [(p, q), l)]ε,r 7→ [(p, q ⊗ Ik), kl]ε,r, 0 ( ^Mk(A)) → K ε,r K ε,r 0 (Mk(A)) → K ε,r respectively induce by restriction homomorphisms Φ0 : K ε,r 0 (Mk(A)) and Ψ0 : K ε,r 0 (A) which are inverse of each other, where in the right hand side of the last formula, we have viewed p ∈ Mn(Mk(A)) as a matrix in Mnk(A) and q ⊗ Ik ∈ Mn(C) ⊗ Mk(C) as a matrix in Mnk(C). Since Φ0 is induced by ιA, we get from lemma 1.23 that ιε,r 0 (Mk(A)) is an A,∗ isomorphism. 0 (A) → K ε,r 0 (A) → K ε,r : K ε,r Let A be a C∗-algebra filtered by (Ar)r>0. Then K(H)⊗ A is filtered by (K(H)⊗ Ar)r>0 and applying proposition 1.28 to the increasing family (Mk(A)+)k∈N of C∗- subalgebras of ^K(H) ⊗ A, lemmas 1.23 and 1.24, and the discussion above, we deduce the Morita equivalence for K ε,r ∗ (•). Proposition 1.29. If A is a filtered algebra and H is a separable Hilbert space, then the homomorphism A → K(H) ⊗ A; a 7→ a   0   . . . induces a (Z2-graded) semi-group isomorphism (the Morita equivalence) ∗ (A) → K ε,r for any positive number r and any ε ∈ (0, 1/4). A : K ε,r Mε,r ∗ (K(H) ⊗ A) ON A QUANTITATIVE OPERATOR K-THEORY 15 1.6. Lipschitz homotopies. Definition 1.30. If A is a C∗-algebra and C a positive integer, then a map h = [0, 1] → A is called C-Lipschitz if for every t and s in [0, 1], then kh(t) − h(s)k 6 Ct − s. Proposition 1.31. There exists a number C such that for any unital filtered C∗- algebra A and any positive number r and ε < 1/4 then : (i) if p0 and p1 are homotopic in Pε,r l and a C-Lipschitz homotopy in Pε,r diag(p1, Ik, 0l). (ii) if u0 and u1 are homotopic in Uε,r n (A), then there exist integers k and n+k+l(A) between diag(p0, Ik, 0l) and C-Lipschitz homotopy in U3ε,2r n (A) then there exist an integer k and a n+k (A) between diag(u0, Ik) and diag(u1, Ik). Proof. (i) Notice first that if p is an ε-r-projection in A, then the homotopy of ε-r- projections of M2(A) between (cid:18)1 0 Let (pt)t∈[0,1] be a homotopy between p0 and p1 in Pε,r 0 0(cid:19) and (cid:18)p 0 1 − p(cid:19) in example 1.6 is 2-Lipschitz. 0 ε−kp2 t−ptk 4 n (A). Set α = inf t∈[0,1] ant let t0 = 0 < t1 < . . . < tk = 1 be a partition of [0, 1] such that kpti − pti−1k < α for i ∈ {1, . . . , k}. We construct a homotopy of ε-r-projections with the required property between diag(p0, In(k−1), 0) and diag(p1, In(k−1), 0) in Mn(2k−1)(A) as the composition of the following homotopies. • We can connect diag(pt0 , In(p−1), 0) and diag(pt0 , In, 0, . . . , In, 0) within Pε,r n(2k−1)(A) by a 2-Lipschitz homotopy. • As we noticed at the beginning of the proof, we can connect diag(pt0 , In, 0, . . . , In, 0) and diag(pt0 , In − pt1 , pt1, . . . , In − ptk , ptk ) within Pε,r • The ε-r-projections diag(pt0 , In − pt1, pt1, . . . , In − ptk , ptk ) and diag(pt0 , In − pt0, . . . , ptk−1, In − ptk−1 , ptk ) satisfy the norm estimate of the assumption of lemma 1.7(i) and hence then can be connected within Pε,r n(2k−1)(A) by a ray which is clearly a 1-Lipschitz homotopy. • Using once again the homotopy of example 1.6, we see that diag(pt0 , In− pt0, . . . , ptk−1, In − ptk−1, ptk ) and diag(0, In, . . . , 0, In, ptk ) are con- nected within Pε,r • Eventually, diag(0, In, . . . , 0, In, ptk ) and diag(ptk , In(k−1), 0) are con- n(2k−1)(A) by a 2-Lipschitz homotopy. n(2k−1)(A) by a 2-Lipschitz homotopy. n(2k−1)(A) by a 2-Lipschitz homotopy. (ii) Let (ut)t∈[0,1] be a homotopy between u0 and u1 in Uε,r n (A). Set α = and let t0 = 0 < t1 < . . . < tp = 1 be a partition of inf t∈[0,1] [0, 1] such that kuti − uti−1k < α for i ∈ {1, . . . , p}. We construct a homo- topy with the required property between diag(u0, I2np) and diag(u1, I2np) within U3ε,2r n(2p+1)(A) as the composition of the following homotopies. nected within Pε,r t ut−Ink 3 ε−ku∗ • Since Inp and diag(u∗t1ut1, . . . , u∗tputp ) satisfy the norm estimate of the assumption of lemma 1.7(ii), then diag(ut0, Inp) is a 3ε-2r-unitary that can be connected to diag(ut0, u∗t1ut1, . . . , u∗tp utp) in U3ε,2r n(p+1)(A) by a 1-Lipschitz homotopy. 16 H. OYONO-OYONO AND G. YU • Proceeding as in the first point of lemma 1.8, we see that diag(In, u∗t1, . . . , u∗tp, Inp) and diag(u∗t1 , . . . , u∗tp , In(p+1)) can be con- nected within Uε,r n(2p+1)(A) by a 2-Lipschitz homotopy and thus, in view of remark 1.4, diag(ut0 , u∗t1ut1, . . . , u∗tp utp , Inp) = diag(In, u∗t1, . . . , u∗tp, Inp) · diag(ut0, ut1, . . . , utp, Inp) and diag(u∗t1, . . . , u∗tp, In(p+1)) · diag(ut0 , ut1, . . . , utp, Inp) = diag(u∗t1 ut0, . . . , u∗tp utp−1, utp, Inp) can be connected within U3ε,2r n(2p+1)(A) by a 4-Lipschitz homotopy. • Since ku∗tiuti−1 − Ink < ε, we get by using once again lemma 1.7(ii) that diag(u∗t1 ut0, . . . , u∗tp utp−1, utp, Inp) and diag(Inp, utp, Inp) can be connected within U3ε,2r n(2p+1)(A) by a 1-Lipschitz homotopy. • Eventually, diag(Inp, utp , Inp) can be connected to diag(utp , I2np) within U3ε,2r (2p+1)n(A) by a 2-Lipschitz homotopy. (cid:3) Corollary 1.32. There exists a control pair (αh, kh) such that the following holds: For any unital filtered C∗-algebra A, any positive numbers ε and r with ε < 1 4αh n (A), then there is for some and any homotopic ε-r-projections q0 and q1 in Pε,r integers k and l an αhε-kh,εr-unitary W in Uαhε,kh,εr (A) such that n+k+l k diag(q0, Ik, 0l) − W diag(q1, Ik, 0l)W ∗k < αhε. Proof. According to proposition 1.31, we can assume that q0 and q1 are connected by a C-Lipschitz homotopy (qt)t∈[0,1], for some universal constant C. Let t0 = 0 < t1 < ··· < tp = 1 be a partition of [0, 1] such that 1/32C < ti − ti−1 < 1/16C. With notation of lemma 1.10, pick for every integer i in {1, . . . , p} a λε-lε-unitary Wi in A such that kWiqti−1 W ∗i − qtik < λε. If we set W = Wp ··· W1, then W is a 3pλε-plεr-unitary such that kW q0W ∗ − q1k < 2pλε. Since p < 2C, we get the result. (cid:3) 2. Controlled morphisms As we shall see in section 3, usual maps in K-theory such as boundary maps factorize through semi-group homomorphism of quantitative K-theory groups with expansion of norm control and propagation controlled by a control pair. This motivates the notion of controlled morphisms for quantitative K-theory in this section. Recall that a controlled pair is a pair (λ, h), where • λ > 1; • h : (0, 1 4λ ) → (0, +∞); ε 7→ hε is a map such that there exists a non- increasing map g : (0, 1 4λ ) → (0, +∞), with h 6 g. The set of control pairs is equipped with a partial order: (λ, h) 6 (λ′, h′) if λ 6 λ′ and hε 6 h′ε for all ε ∈ (0, 1 4λ′ ) ON A QUANTITATIVE OPERATOR K-THEORY 17 2.1. Definition and main properties. For any filtered C∗-algebra A, let us de- fine the families K0(A) = (K ε,r 1 (A))0<ε<1/4,r>0 and K∗(A) = (K ε,r ∗ (A))0<ε<1/4,r>0. Definition 2.1. Let (λ, h) be a controlled pair, let A and B be filtered C∗-algebras, and let i, j be elements of {0, 1,∗}. A (λ, h)-controlled morphism 0 (A))0<ε<1/4,r>0, K1(A) = (K ε,r is a family F = (F ε,r)0<ε< 1 F : Ki(A) → Kj(B) 4λ ,r>0 of semigroups homomorphisms F ε,r : K ε,r (B) (A) → K λε,hεr j i such that for any positive numbers ε, ε′, r and r′ with 0 < ε 6 ε′ < 1 hεr 6 hε′ r′, we have 4λ and F ε′,r′ ◦ ιε,ε′,r,r′ i = ιλε,λε′,hεr,hε′ r′ j ◦ F ε,r. controlled morphism. If it is not necessary to specify the control pair, we will just say that F is a Let A and B be filtered algebras. Then it is straightforward to check that if F : Ki(A) → Kj (B) is a (λ, h)-controlled morphism, then there is group homo- morphism F : Ki(A) → Kj(B) uniquely defined by F ◦ ιε,r ◦ F ε,r. The homomorphism F will be called the (λ, h)-controlled homomorphism induced by F . A homomorphism F : Ki(A) → Kj(B) is called (λ, h)-controlled if it is induced by a (λ, h)-controlled morphism. If we don't need to specify the control pair (λ, h), we will just say that F is a controlled homomorphism. i = ιλε,hεr j Example 2.2. (i) Let A = (Ar)r>0 and B = (Br)r>0 be two filtered C∗-algebras and let f : A → B be a homomorphism. Assume that there exists d > 0 such that f (Ar) ⊂ Bdr for all positive r. Then f gives rise to a bunch of semi- group homomorphisms (cid:16)f ε,r and hence to a (1, d)-controlled morphism f∗ : K∗(A) → K∗(B). (B)(cid:17)0<ε< 1 ∗ (A) → K ε,dr : K ε,r 4 ,r>0 ∗ ∗ (ii) The bunch of semi-group isomorphisms 4 ,r>0 of proposition 1.29 defines a (1, 1)-controlled morphism ∗ (K(H)⊗A))0<ε< 1 ∗ (A) → K ε,r A : K ε,r (Mε,r and MA : K∗(A) → K∗(K(H)⊗A) M−1 A : K∗(K(H)⊗A) → K∗(A) inducing the Morita equivalence in K-theory. If (λ, h) and (λ′, h′) are two control pairs, define h ∗ h′ : (0, 1 4λλ′ ) → (0, +∞); ε 7→ hλ′εh′ε. Then (λλ′, h ∗ h′) is a control pair. Let A, B1 and B2 be filtered C∗-algebras, let i, j and l be in {0, 1,∗} and let F = (F ε,r)0<ε< 1 ,r>0 : Ki(A) → Kj (B1) be a (αF , kF )-controlled morphism, let G = (Gε,r)0<ε< 1 ,r>0 : Kj (B1) → Kl(B2) be a (αG, kG)-controlled morphism. Then G◦F : Ki(A) → Kl(B2) is the (αGαF , kG∗kF )- controlled morphism defined by the family (GαF ε,kF,εr ◦ F ε,r)0<ε< 1 ,r>0. 4αF αG 4αF 4αG 18 H. OYONO-OYONO AND G. YU Remark 2.3. The Morita equivalence for quantitative K-theory is natural, i.e MB ◦ f = (IdK(H))⊗f ) ◦ MA for any homomorphism f : A → B of filtered C∗-algebras. Notation 2.4. Let A and B be filtered C∗-algebras, let (λ, h) be a control pair, and let F = (F ε,r)0<ε< 1 ,r>0) be a (αF , kF )-controlled morphism (resp. a (αG, kG)-controlled morphism). Then we write F ,r>0 : Ki(A) → Kj(B) (resp. G = (Gε,r)0<ε< 1 ∼ G if (λ,h) 4αF 4αG • (αF , kF ) 6 (λ, h) and (αG, kG) 6 (λ, h). • for every ε in (0, 1 4λ ) and r > 0, then ιαF ε,λε,kF,εr,hεr j ◦ F ε,r = ιαG ε,λε,kG,εr,hεr j ◦ Gε,r. (λ,h) If F and G are controlled morphisms such that F ∼ G for a control pair (λ, h), then F and G induce the same morphism in K-theory. Remark 2.5. Let F : Ki(A2) → Kj(B1) (resp. F′ : Ki(A2) → Kj(B1)) be a (αF , kF )-controlled (resp. a (αF ′, kF ′)-controlled) morphisms and let G : Ki′ (A1) → Ki(A2) (resp. G′ : Kj(B1) → Kl(B2)) be a (αG, kG)-controlled (resp. a (αG′ , kG′ )- controlled) morphism. Assume that F ∼ F′ for a control pair (λ, h), then (λ,h) • G′ ◦ F • F ◦ G (αG′ λ,kG′∗h) (αG λ,h∗kG ) ∼ ∼ G′ ◦ F′; F′ ◦ G. the controlled morphism induced by IdA. If i is an element in {0, 1,∗} and A a filtered C∗-algebra, we denote by IdKi(A) Let F : Ki(A1) → Ki′ (B1), F′ : Kj(A2) → Kl(B2), G : Ki(A1) → Kj (A2) and G′ : Ki′ (B1) → Kl(B2) be controlled morphisms and let (λ, h) be a control pair. Then the diagram Ki′ (B1) −−−−→ Kl(B2) G′ Ki(A1) G−−−−→ Kj(A2) xF ′ Fx is called (λ, h)-commutative (or (λ, h)-commutes) if G′ ◦ F Definition 2.6. Let (λ, h) be a control pair, and let F : Ki(A) → Kj(B) be a (αF , kF )-controlled morphism with (αF , kF ) 6 (λ, h). ∼ F′ ◦ G. • F is called left (λ, h)-invertible if there exists a controlled morphism (λ,h) G : Kj (B) → Ki(A) (λ,h) such that G ◦ F a left (λ, h)-inverse for F . Notice that definition of (αF αG, kF ∗ kG) 6 (λ, h). ∼ IdKi(A). The controlled morphism G is then called ∼ implies that • F is called right (λ, h)-invertible if there exists a controlled morphism (λ,h) G : Kj (B) → Ki(A) (λ,h) such that F ◦ G right (λ, h)-inverse for F . ∼ IdKi(B). The controlled morphism G is then called a ON A QUANTITATIVE OPERATOR K-THEORY 19 • F is called (λ, h)-invertible or a (λ, h)-isomorphism if there exists a con- trolled morphism G : Kj (B) → Ki(A) which is a left (λ, h)-inverse and a right (λ, h)-inverse for F . The con- trolled morphism G is then called a (λ, h)-inverse for F (notice that we have in this case necessarily (αG, kG) 6 (λ, h)). We can check easily that indeed, if F is left (λ, h)-invertible and right (λ, h)- invertible, then there exists a control pair (λ′, h′) with (λ, h) 6 (λ′, h′), depending only on (λ, h) such that F is (λ′, h′)-invertible. Definition 2.7. Let (λ, h) be a control pair and let F : Ki(A) → Kj(B) be a (αF , kF )-controlled morphism. • F is called (λ, h)-injective if (αF , kF ) 6 (λ, h) and for any 0 < ε < 1 (A), then F ε,r(x) = 0 in K αF ε,kF,εr 4λ , any (B) implies r > 0 and any x in K ε,r that ιε,λε,r,hεr (x) = 0 in K λε,hεr i i i (A); • F is called (λ, h)-surjective, if for any 0 < ε < 1 4λαF , any r > 0 and (B), there exists an element x in K λε,hεr (A) such that j i any y in K ε,r F λε,hλεr(x) = ιε,αF λε,r,kF,λεhεr j j (y) in K αF λε,kF,λεhεr j (B). Remark 2.8. (i) If F : K1(A) → Ki(B) is a (λ, h)-injective controlled morphism. Then according to lemma 1.16, there exists a control pair (λ′, h′) with (λ, h) 6 (λ′, h′) depending only on (λ, h) such that for any 0 < ε < 1 4λ′ , any r > 0 and any x and x′ in K ε,r (B) implies that ιε,λ′ε,r,h′ εr 1 (A) , then F ε,r(x) = F ε,r(x′) in K αF ε,kF,εr (x) = ιε,λ′ε,r,h′ εr (x′) in K λ′ε,h′ εr (A); i 1 1 1 (ii) It is straightforward to check that if F is left (λ, h)-invertible, then F is (λ, h)-injective and that if F is right (λ, h)-invertible, then there exists a control pair (λ′, h′) with (λ, h) 6 (λ′, h′), depending only on (λ, h) such that F is (λ′, h′)-surjective. (iii) On the other hand, if F is (λ, h)-injective and (λ, h)-surjective, then there exists a control pair (λ′, h′) with (λ, h) 6 (λ′, h′), depending only on (λ, h) such that F is a (λ′, h′)-isomorphism. 2.2. Controlled exact sequences. Definition 2.9. Let (λ, h) be a control pair, 4αG 4αF • Let F = (F ε,r)0<ε< 1 ,r>0 : Ki(A) → Kj (B1) be a (αF , kF )-controlled morphism, and let G = (Gε,r)0<ε< 1 ,r>0 : Kj (B1) → Kl(B2) be a (αG, kG)- controlled morphism, where i, j and l are in {0, 1,∗} and A, B1 and B2 are filtered C∗-algebras. Then the composition Ki(A) F→ Kj(B1) G→ Kl(B2) , any r > 0 and any y in K ε,r is said to be (λ, h)-exact at Kj(B1) if G ◦ F = 0 and if for any 0 < ε < (B1) such that Gε,r(y) = 0 in 4 max{λαF ,αG} K αG ε,kG,εr (B2), there exists an element x in K λε,hεr (A) such that 1 j j i F λε,hλεr(x) = ιε,αF λε,r,kF,λεhεr j (y) 20 H. OYONO-OYONO AND G. YU in K αF λε,kF,λεhεr j (B1). • A sequence of controlled morphisms ···Kik−1 (Ak−1) Fk−1→ Kik (Ak) Fk→ Kik+1 (Ak+1) Fk+1→ Kik+2 (Ak+2)··· is called (λ, h)-exact if for every k, the composition Kik−1 (Ak−1) Fk−1→ Kik (Ak) Fk→ Kik+1 (Ak+1) is (λ, h)-exact at Kik (Ak). Remark 2.10. If the composition Ki(A) F→ K1(B1) G→ Kl(B2) is (λ, h)-exact, then according to lemma 1.16, there exists a control pair (λ′, h′) with (λ, h) 6 (λ′, h′) depending only on (λ, h), such that for any 0 < ε < , any r > 0 and any y and y′ in K ε,r there exists an element x in K λ′ε,h′ εr 4 max{λ′αF ,αG} 1 (B1) , then Gε,r(y) = Gε,r(y′) in K αF ε,kF,εr (B) implies that (A) such that 1 j i ε,αF λ′ε,r,kF,λ′ εh′ ι j εr ε,αF λ′ε,r,kF,λ′ εh′ (y′) = ι j εr (y) + F λ′ε,h′ εr(x) in K αF λ′ε,kF,λεh′ εr 1 (B1). 3. Extensions of filtered C∗-algebras The aim of this section is to establish a controlled exact sequence for quantitative K-theory with respect to extension of filtered C∗-algebras admitting a completely positive cross section that preserves the filtration. We also prove that for these extensions, the boundary maps are induced by controlled morphisms. As in K- theory, one is a map of exponential type and the other is an index type map, and the later in turn fits in a long (λ, h)-controlled exact sequence for some universal control pair (λ, h). 3.1. Semi-split filtered extensions. Let A be a C∗-algebra filtered by (Ar)r>0 and let J be an ideal of A. Then A/J is filtered by ((A/J)r)r>0, where (A/J)r is the image of Ar under the projection A → A/J. Assume that the C∗-algebra extension 0 → J → A → A/J q → 0 admits a contractive filtered cross-section s : A/J → A, i.e such that s((A/J)r)) ⊂ Ar for any positive number. For any x ∈ J and any number ε > 0 there exists a positive number r and an element a of Ar such that kx − ak < ε. Let us set y = a − s ◦ q(a). Then y belongs to Ar ∩ J and moreover ky − xk = ka − x + s ◦ q(x − a)k 6 ka − xk + ks ◦ q(a − x)k 6 ka − xk + kq(x − a)k 6 2ε. Hence, [r>0 (Ar ∩ J) is dense in J and therefore J is filtered by (Ar ∩ J)r>0. Definition 3.1. Let A be a C∗-algebra filtered by (Ar)r>0 and let J be an ideal of A. The extension of C∗-algebras 0 → J → A → A/J → 0 ON A QUANTITATIVE OPERATOR K-THEORY 21 is said to be filtered and semi-split (or a semi-split extension of filtered C∗-algebras) if there exists a completely positive cross-section such that s : A/J → A s((A/J)r)) ⊂ Ar for any number r > 0. Such a cross-section is said to be semi-split and filtered. We have the following analogous of the lifting property for unitaries of the neutral component. Lemma 3.2. There exists a control pair (αe, ke) such that for any semi-split ex- tension of filtered C∗-algebras 0 −→ J −→ A q −→ A/J −→ 0 n+j with A unital, the following holds: for every positive numbers r and ε with ε < 1 4αe and any ε-r-unitary V homotopic to In in Uε,r n (A/J), then for some integer j, there exists a αeε-ke,εr-unitary W homotopic to In+j in Uαeε,ke,εr (A) and such that kq(W ) − diag(V, Ij )k < αeε. Proof. According to proposition 1.31, we can assume that V and In are connected by a C-Lipschitz homotopy (Vt)t∈[0,1], for some universal constant C. Let t0 = 0 < t1 < ··· < tp = 1 be a partition of [0, 1] such that 1/16C < ti − ti−1 < 1/8C. Then we get that kVi−1 − Vik < 1/8 and hence kVi−1V ∗i − Ink < 1/2. Let lε be the smallest integer such that Pk>lε+1 2−k/k < ε and Pk>lε+1 logk 2/k! < ε and let us consider the polynomial functions Pε(x) = Plε k=0 xk/k! and Qε(x) = −Plε k=1 xk/k. We get then kVi−1V ∗i − Pε ◦ Qε(1 − Vi−1V ∗i )k 6 3ε. Choose a completely positive section s(A/J) → A such that s(1) = 1 and let us set W t i = Pε(s(tQε(In − Vi−1V ∗i ))) for t in [0, 1] and i in {1, . . . , p}. Since Vi−1V ∗i is closed to the unitary Vi−1V ∗i (ViV ∗i−1Vi−1V ∗i )−1/2, then W t i is uniformly (in t and i) closed to exp ts(log(Vi−1V ∗i (ViV ∗i−1Vi−1V ∗i )−1/2)) which is unitary (the logarithm is well defined since Vi−1V ∗i (ViV ∗i−1Vi−1V ∗i )−1/2 is closed to In) and hence W t i is a αε-2lεr- unitary for some universal α. Hence W 1 n (A) homotopic i i ) − Vi−1V ∗i k < 3αε. If we set now W = W 1 to In and such that kq(W 1 p and since p 6 16C, then W satisfies the required property. (cid:3) is a αε-2lεr-unitary in Uε,r 1 ··· W 1 Lemma 3.3. There exists a control pair (α, k) such that for any semi-split exten- sion of filtered C∗-algebras 0 → J → A → A/J → 0 with A unital, any semi-split filtered cross section s : A/J → A with s(1) = 1 and any ε-r-projection p in A/J with 0 < ε < 1 4λ , there exists an element yp in Jkεr such that k1+yp−e2ıπs(k0(p))k < αε/3. In particular 1 + yp is a αε-kεr-unitary of J +; Proof. Let lε be the smallest integer such that +∞Xl=lε+1 10l/l! < ε. 22 H. OYONO-OYONO AND G. YU (2ıπs(p))l l! Let us define zp = lεXl=0 (cid:13)(cid:13)(cid:13)zp − e2ıπs(κ0(p))(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lεXl=0 6 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lεXl=0 6 ks(p) − s(κ0(p))ke10 + ε 6 (2e10 + 1)ε. . Then zp belongs to Mn(Alεr) and we have (2ıπs(p))l l! +∞Xl=0 − (2ıπs(κ0(p))l l! (2ıπs(p))l − (2ıπs(κ0(p)l) l! (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +∞Xl=lε+1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (2ıπs(κ0(p))l l! (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) If we set yp = zp − s ◦ q(zp), then yp ∈ Mn(J ∩ Alεr) and kzp − (1 + yp)k = ks ◦ q(zp) − 1k 6 (cid:13)(cid:13)(cid:13)q(cid:16)zp − e2ıπs(κ0(p))(cid:17)(cid:13)(cid:13)(cid:13) < λε, with λ = (2e10 + 1). Therefore we have k1 + yp − e2ıπs(κ0(p))k < 2λε. The end of the statement is then a consequence of lemma 1.7. (cid:3) 3.2. Controlled boundary maps. For any extension 0 → J → A → A/J → 0 of C∗-algebras we denote by ∂J,A : K∗(A/J) → K∗(J) the associated (odd degree) boundary map. Proposition 3.4. There exists a control pair (αD, kD) such that for any semi-split extension of filtered C∗-algebras −→ A/J −→ 0, there exists a (αD, kD)-controlled morphism of odd degree 0 −→ J −→ A q DJ,A = (∂ε,r J,A)0<ε 1 4αD ,r : K∗(A/J) → K∗(J) which induces in K-theory ∂J,A : K∗(A/J) → K∗(J). Proof. Let s : A/J → A be a semi-split filtered cross-section. Let us first prove the result when A is unital. (i) Let p be an element of Pε,r n (A/J). Then ∂J,A([κ0(p)]) is the class of e2ıπs(κ0(p)) in K1(J). Fix a control pair (α, k) as in lemma 3.3 and pick any yp in Mn(Jkεr) such that k1 + yp− e2ıπs(k0(p))k < αε/3. Then 1 + yp is an αε-kεr-unitary of Mn(J +), and according to lemma 1.7, any two such αε-kεr-unitaries are homotopic in U 3αε,kεr (J +). Applying lemma 3.3 to A/J[0, 1], we see that the map n Pε,r n (A/J) −→ U3αε,kεr n (J +); p 7→ 1 + yp preserves homotopies and hence gives rise to a bunch of well defined semi- group homomorphism ∂ε,r J,A : K ε,r 0 (A/J) −→ K 3αε,kεr 1 (J); [p, l]ε,r 7→ [1 + yp]3αε,kεr which in the even case satisfies the required properties for a controlled homomorphism. ON A QUANTITATIVE OPERATOR K-THEORY 23 n (A/J), pick any element v in some Uε,r (ii) In the odd case, we follow the route of [18, Chapter 8]. For any element u of Uε,r j (A/J) such that diag(u, v) is homotopic to In+j in U3ε,2r n+j (A/J) (we can choose in view of lemma 1.16 v = u∗). According to lemma 3.2, and up to replace v by diag(v, Ik) for some integer k, there exists an element w in U3αeε,2ke,3εr (A) such that kq(w) − diag(u, v)k 6 3αeε. Let us set x = w diag(In, 0)w∗. Then x is an element in P6αeε,4ke,3εr Let us set now h = x − diag(In, 0) − s ◦ q(x − diag(In, 0)). Then h is a (A) such that kq(x) − diag(In, 0)k 6 9αeε. n+j n+j self-adjoint element of M2n(A4ke,3εr ∩ J) such that kx − diag(In, 0) − hk 6 9αeε, and therefore h + diag(In, 0) belongs to P45αeε,4ke,3εr n+j (J). Define then ∂ε,r J,A([u]ε,r) = [h + diag(In, 0), n]450αeε,4ke,3εr . It is straightforward to check that (compare with [18, Chapter 8]). • two choice of elements satisfying the conclusion of lemma 3.2 relatively (J) (this to diag(u, v) give rise to homotopic elements P450αeε,4ke,3εr is a consequence of lemma 1.7). n+j • Replacing u by diag(u, Im) and v by diag(v, Ik) gives also rise to the same element of K 450αeε,4ke,3εr (J). 0 Applying now lemma 3.2 to the exact sequence 0 → J[0, 1] → A[0, 1] → A/J[0, 1] → 0, J,A([u]ε,r) we get that ∂ε,r • only depends on the class of u in K ε,r • does not depend on the choice of v such that diag(u, v) is connected 1 (A/J); to In+j in Uε,r n+j(A/J). • If A is not unital, use the exact sequence 0 → J → eA → gA/J → 0 J,A as the composition to define ∂ε,r 0 (A/J) ֒→ K ε,r K ε,r 0 (gA/J) ∂ε,r J, eA−→ K 450αeε,4ks,3εr 1 (J), where the inclusion in the composition is induced by the inclusion A/J ֒→ gA/J ∼= eA/J. • Since the set of filtered semi-split cross-section s : A/J → A such that s((A/J)r) ⊂ Ar is convex, the definition of ∂ε,r J,A actually does not depend on the choice of such a section. • Using lemma 1.7, (αD, kD), then DJ,A = (∂ε,r inducing the (odd degree) boundary map ∂J,A : K∗(A/J) → K∗(J). it is plain to check that for a suitable control pair ,r is a (αD, kD)-controlled morphism J,A)0<ε 1 4αD (cid:3) For a semi-split extension of filtered C∗-algebras 0 −→ J −→ A q −→ A/J −→ 0, 24 H. OYONO-OYONO AND G. YU J,A : K0(A/J) → K1(J), for the restriction of DJ,A to K0(A/J) and we set D0 D1 J,A : K1(A/J) → K0(J), for the restriction of DJ,A to K1(A/J). Remark 3.5. (i) Let A and B be two filtered C∗-algebras and let φ : A → B be a filtered homomorphism. Let I and J be respectively ideals in A and B and assume that • φ(I) ⊂ J; • there exists semi-split filtered cross-sections s : A/I → A and s′ : B/J → J such that s′ ◦ φ = φ ◦ s, where φ : A/I → B/J is the homomorphism induced by φ, (ii) Let 0 −→ J −→ A then DJ,B ◦ φ∗ = φ∗ ◦ DI,A. −→ A/J −→ 0 be a split extension of filtered C∗- algebras, i.e there exists a homomorphism of filtered C∗-algebras s : A/J → A such that q ◦ s = IdA/J . Then we have DJ,A = 0. q For a filtered C∗-algebra A, we have defined the suspension and the cone re- spectively as SA = C0((0, 1), A) and CA = C0((0, 1], A). Then SA and CA are filtered C∗-algebras and evaluation at the value 1 gives rise to a semi-split filtered extension of C∗-algebras (1) and in the even case, the corresponding boundary ∂SA,CA : K0(A) → K1(SA) implements the suspension isomorphism and has the following easy description when A is unital: if p is a projection, then ∂SA,CA[p] is the class in K1(SA) of the path of unitaries 0 → SA → CA → A → 0 [0, 1] → Un(A); t 7→ pe2ıπt + 1 − p. Let us show that we have an analogous description in term of almost projection. Notice that if q is an ε-r-projection in A, then zq : [0, 1] → A; t 7→ qe2ıπt + 1 − q is a 5ε-r-unitary in fSA. Using this, we can define a (5, 1)-controlled morphism ZA = (Z ε,r A )0<ε<1/20,r>0 : K0(A) → K1(SA) in the following way: n (A) and any integer k let us set • for any q in Pε,r Vq,k : [0, 1] → U5ε,r • define then Z ε,r n (fSA) : t 7→ diag(e−2kıπt, 1, . . . , 1) · (1 − q + qe2ıπt); A ([q, k]ε,r) = [Vq,k]5ε,r. (λ,h) CA,SA ∼ ZA. Proposition 3.6. There exists a control pair (λ, h) such that for any filtered C∗- algebra A, then D0 Proof. Let [q, k]ε,r be an element of K ε,r n (A) and k integer. We can assume without loss of generality that n > k. Namely, up to replace n by 2n and using a homotopy between diag(q, 0) and diag(0, q) in Pε,r 2n (A), we can indeed assume that q and diag(Ik, 0) commute. As in the proof of proposition 3.4, define lε as the smallest integer such that P∞l=lε+1 10l/l! < ε. Let us consider the following 0 (A), with q in Pε;r paths in Mn(A) z : [0, 1]−→Mn(A); t 7→ lεXl=0 (2ıπ(tq + (1 − t) diag(Ik, 0)))l/l! ON A QUANTITATIVE OPERATOR K-THEORY 25 and z′ : [0, 1]−→Mn(A); t 7→ exp(2ıπ diag(−tIk, 0))(1 − q + e2ıπtq). Since q and Ik commutes, then exp(2ıπ(diag(−tIk, 0) + tq)) = exp(2ıπ diag(−tIk, 0)) · exp(2ıπtq) and hence z(t) = exp(2ıπ diag(−tIk, 0)) exp(2ıπtq) − ∞Xl=lε+1 (2ıπ(tq + (1 − t) diag(Ik, 0)))l/l!. We get therefore kz(t) − z′(t)k 6 ε + kqe2ıπt + (1 − q) − exp 2ıπtqk 6 ε + 2kκ0(q) − qk + k exp 2ıπtκ0(q) − exp 2ıπtqk 6 ε(5 + 4e4π). Let us set y : [0, 1] : −→Mn(A); t 7→ z(t) − 1 − (1 − t) diag(Ik, 0) For some αs > α∂ , we get then that 1 + y and z′ are homotopic elements in U αsε,k∂,εr for the extension of equation (1), we get in view of the proof of proposition 3.4, (fSA). Using the semi-split filtered cross-section A → CA; a 7→ [t 7→ ta] (2ıπ)l/l! − t (2ıπq)l/l!. n lεXl=1 lεXl=1 ια∂ ε,αsε,k∂,εr 1 ◦ ∂ε,r SA,CA([q, k]ε,r) = [1 + y]αsε,k∂,εr, and thus we deduce ια∂ ε,αsε,k∂,εr 1 ◦ ∂ε,r SA,CA([q, k]ε,r) = [z′]αsε,k∂,εr. We get the result by using a homotopy of unitaries in Mn(fSA) between and t 7→ exp(2ıπ diag(−tIk, In−k)). t 7→ diag(e−2kπt, 1, . . . , 1) (cid:3) The inverse of the suspension isomorphism is provided, up to Morita equivalence let us consider the unilateral shift S on ℓ2(N), i.e the by the Toeplitz extension: operator defined on the canonical basis (en)n∈N of ℓ2(N) by S(en) = en+1 for all integer n. Then the Toeplitz algebra T is the C∗-subalgebra of L(ℓ2(N)) generated by S. The algebra of compact operators K(ℓ2(N)) is an ideal of T and we get an extension of C∗-algebras 0 → K(ℓ2(N)) → T ρ → C(S1) → 0, called the Toeplitz extension, where S1 denote the unit circle. Let us define T0 = ρ−1(C0((0, 1)), where C0(0, 1) is viewed as a subalgebra of C(S1). We obtain then an extension of C∗-algebras For any C∗-algebra A, we can tensorize this exact sequence to obtain an extension 0 → K(ℓ2(N)) → T0 → C0(0, 1) → 0. ρ 0 → K(ℓ2(N)) ⊗ A → T0 ⊗ A→SA → 0 26 H. OYONO-OYONO AND G. YU which is filtered and semi-split when A is a filtered C∗-algebra. Proposition 3.7. There exists a control pair (λ, h) such that for any unital filtered C∗-algebra A. D1 K(ℓ2(N))⊗A,T0⊗A ◦ ZA (λ,h) ∼ MA Proof. Let q be a ε-r-projection in Mn(A). We can assume indeed without loss of generality that n = 1.The Toeplitz extension is semi-split by the section induced by the completely positive map s : C(S1) −→ T ; f 7→ Mf , where if π0 stands for the projection L2(S1) ∼= ℓ2(Z) → l2(N), then Mf is the composition (f· being the pointwise multiplication by f ). Notice first that (cid:0) S 1−SS∗ (cid:1) is a unitary lift of S1 → M2(C); z 7→ diag(z, ¯z) in M2(T ) under the homomorphism induced by ρ : T → C(S1). Under the section induced by s, we see that zq lifts to 1 ⊗ (1 − q) + S ⊗ q, and hence l2(N) ֒→ ℓ2(Z) ∼= L2(S1) f·→ L2(S1) π0→ l2(N), S∗ 0 W =(cid:18)S 1 − SS∗ S∗ (cid:19) ⊗ q + I2 ⊗ (1 − q) 0 is a lift in U5ε,r W ∗ diag(1, 0)W is closed to 2 (T0 ⊗ A) of diag(zq, z∗q ). Since kq(1 − q)k < ε, we see that (cid:18) S∗ 1 − SS∗ S(cid:19)(cid:18)1 0 0 0 0(cid:19)(cid:18)S 1 − SS∗ S∗ (cid:19) ⊗ q2 +(cid:18)1 0 0 0(cid:19) ⊗ (1 − q)2. 0 Hence, W ∗ diag(1, 0)W is an element of P10ε,2r SS∗) ⊗ q). Since 2 (T0⊗A) which is closed to diag(1, (1− MA([q, 0]ε,r) = [diag(0, (1 − SS∗) ⊗ q)]ε,r, we get the existence of a positive real αt such that the proposition holds. (cid:3) 3.3. Long exact sequence. We follow the route of [18, Sections 6.3, 7.1 and 8.2] to state for semi-split extensions of filtered C∗-algebras (λ, h)-exact long exact sequences in quantitative K-theory, for some universal control pair (λ, h). Proposition 3.8. There exists a control pair (λ, h) such that for any semi-split extension of filtered C∗-algebras 0 −→ J −→ A/J −→ 0, −→ A q  the composition K∗(J) j∗→ K∗(A) q∗→ K∗(A/J) is (λ, h)-exact at K∗(A). Proof. We can assume without loss of generality that A is unital. In the even 0 (A) such that q∗(y) = 0 in K ε,r case, let y be an element of K ε,r 0 (A/J), let e be an ε-r-projection in Mn(A) and let l be a positive integer such that y = [e, k]ε,r. Up to stabilization, we can assume that k 6 n and that q(e) is homotopic to pk = diag(Ik, 0) as an ε-r-projection in Mn(A/J). According to corollary 1.32, there exists up to stabilization a αhε-kh,εr-unitary W of Mn(A/J) such that kW q(e)W ∗ − pkk 6 αhε. ON A QUANTITATIVE OPERATOR K-THEORY 27 The 3αsε-2kh,εr-unitary diag(W, W ∗) of M2n(A/J) is homotopic to I2n. Let choose as in lemma 3.2, a control pair (α, l), an integer j and a αε-lεr-unitary V of M2n+j(A) such that kq(V ) − diag(W, W ∗, Ik+j )k 6 αε. 0 If we set e′ = V diag(e, 0)V ∗, then e′ is a 4αε-2lεr-projection in M2n+j(A). If s : A/J → A is a semi-split filtered cross-section such that s(1) = 1, define f = e′ − s ◦ q(e′ − diag(In, 0)). We see that f belongs to M2n+j(J +) and moreover, since kf − e′k 6 (4α + αh)ε, then according to lemma 1.7, f is for a suitable λ a λε-2lεr-projection of M2n+k(J +) homotopic to e′. Then x = [f, k]λε,2lεr defines a class in K λε,2lεr (J). As in the proof of (ii) of lemma 1.8 we can choose λ big enough so that diag(e′, I2n+j ) and diag(e, 0, I2n+j) are homotopic λε-2kh,εr-projections of M2n(A) and hence we get the result in the even case. For the odd case, let y be an element in K ε,r 1 (A/J) and let us choose an ε-r-unitary V in some Mn(A) such that y = [V ]ε,r. In view of lemma 3.2 and up to enlarge the size of the matrix V , we can assume that kq(V ) − q(W )k 6 αeε with W a αeε-ke,εr-unitary of Mn(A) homotopic to In. Hence W ∗V and V are homotopic 3αeε-(ke,ε + 1)r-unitary of Mn(A). If we set 1 (A) such that q∗(y) = 0 in K ε,r U = W ∗V + s ◦ q(In − W ∗V ), then the coefficients of the matrix U − In lie in J. Moreover, since kU − W ∗V k 6 (2αe + 1)ε, we obtain that U is a λε-(lε + 1)r-unitary for some λ > 1. Hence, x = [U ]λε,(ke,ε+1)r defines a class in K λε,(ke,ε+1)r (cid:3) (J) with the required property. 1 Proposition 3.9. There exists a control pair (λ, h) such that for any semi-split extension of filtered C∗-algebras 0 −→ J −→ A/J −→ 0, −→ A q  the composition K1(A) q∗→ K1(A/J) D1 J,A→ K0(J) is (λ, h)-exact at K1(A/J). Proof. We can assume without loss of generality that A is unital. Let y be an element of K ε,r (A/J) and let U be an ε-r-unitary of Mn(A/J) such that y = [U ]ε,r. With notation of lemma 3.2, let j be an integer and W be a 3αeε-2ke,3εr-unitary in M2n+j(A) such that J,A(y) = 0 in K α∂ ε,k∂,εr 1 (A/J) such that ∂ε,r 0 kq(W ) − diag(U, U∗, Ij )k 6 αε. Set x = W diag(In, 0)W ∗ and h = x − diag(In, 0) − s ◦ q(x − diag(In, 0) as in the proof of proposition 3.4. Since ∂ε,r J,A(y) = 0, we can up to take a larger n assume that h + diag(In, 0) is homotopic to diag(In, 0) as an αDε-kD,εr-projection of M2n+j( J). Since x is close to h+ diag(In, 0), we get from corollary 1.32 that up to take a larger j, there exists for a control pair (α, l), depending only on the control pairs (αh, kh) and (αD, kD) of corollary 1.32 and lemma 3.3, an αε-lεr-unitary V ′ in M2n+j(eJ) such that W diag(In, 0)W ∗ − V ′ diag(In, 0)V ′∗k 6 αε. 28 H. OYONO-OYONO AND G. YU Then V = ρJ (V ′)V ′−1W ∗ is a 10(α + αe)ε-(lε + ke,ε)r-unitary in M2n+j(A) such that Since for a suitable constant α′ depending only on α we have kq(V ) − diag(U, U∗, Ij)k 6 αε. kρJ (V ′) diag(In, 0)ρJ (V ′∗) − diag(In, 0)k 6 α′ε, kV diag(In, 0)V ∗ − diag(In, 0)k 6 α′′ε we obtain that and kV ∗ diag(In, 0)V − diag(In, 0)k 6 α′′ε for some constant α′′ depending only on α′ that we can choose indeed larger than (10α + αe). Hence the n × n-left upper corner X of V is a α′′ε-(lε + l′ε)r-unitary in Mn(A) such that kq(X) − Uk 6 α′′ε. Hence we get the result. Proposition 3.10. There exists a control pair (λ, h) such that for any semi-split extension of filtered C∗-algebras 0 −→ J −→ A/J −→ 0, −→ A (cid:3)  q the composition K1(A/J) D1 J,A→ K0(J) ∗→ K0(A) ∗ (y) = 0 in K ε,r is (λ, h)-exact in K0(J). Proof. It is enough to prove the result for A unital. Let y be an element of K ε,r 0 (J) such that ε,r 0 (A), let e be an ε-r-projection in Mn(J +) and k be a positive integer such that y = [e, k]ε,r. If we set pk = diag(Ik, 0), we can indeed assume without loss of generality that kq(e) − pkk 6 2ε (where J + is viewed as a subalgebra of A). Up to stabilization, we can also assume that e is homotopic to pk as an ε-r-projection in Mn(A). According to corollary 1.32, there exists up to stabilization a αhε-kh,εr-unitary W of Mn(A) such that Up to replace n by 2n, W by diag(W, W ∗) and e by diag(e, 0), we can assume that W is a 3αhε-2kh,εr-unitary homotopic to In. Since ke − W pkW ∗k 6 αhε. kq(W )pkq(W ∗) − pkk 6 kq(W )pkq(W ∗) − q(e)k + kq(e) − pkk < (2 + αh)ε, then kq(W ∗)pkq(W ) − pkk < (2 + 4αh)ε. Hence for an α′ > 1 depending only on αh, the left-up n × n corner V1 and the right bottom corner V2 of q(W ) are α′ε-2ke,εr-unitaries of Mn(A/J) such that kq(W )q(W ∗) − diag(V1, V2) diag(V1, V2)∗k < (αh + α′)ε and kq(W ∗)q(W ) − diag(V1, V2)∗ diag(V1, V2)k < (αh + α′)ε. Hence q(W ) is close to diag(V1, V2) and hence there is a λ > 1 depending only on αe such that as a λε-2kh,εr-unitary of Mn(A/J), then diag(V1, V2) is homotopic ON A QUANTITATIVE OPERATOR K-THEORY 29 to q(W ) and hence to In. We can indeed choose λ big enough such that if we set x = [V1]λε,2ke,ε r, then ∂λε,2ke,εr J,A (x) = [e, k]λα∂ ε,k∂,αε2ke,εr = ιε,r,λε,2ke,εr ∗ (y). (cid:3) From propositions 3.8, 3.9 and 3.10 we can derive the analogue of the long exact sequence in K-theory. Theorem 3.11. There exists a control pair (λ, h) such that for any semi-split extension of filtered C∗-algebras 0 −→ J −→ A/J −→ 0, −→ A q  the sequence K1(J) is (λ, h)-exact. ∗−→ K1(A) q∗−→ K1(A/J) DJ,A−→ K0(J) ∗−→ K0(A) q∗−→ K0(A/J) As a consequence, using the exact sequence (2) 0 → SA → CA → A → 0, and in view of lemma 1.27 and point (iii) of remark 2.8, we deduce in the setting of the semigroup K ε,r ∗ (•) the analogue of the suspension isomorphism in K-theory. Corollary 3.12. Let D1 SA,CA : K1(A) → K0(SA) be the controlled boundary morphism associated to the semi-split and filtered extension of equation (2) for a filtered C∗-algebra A. A = D1 A is (λ, h)-invertible. • There exists a control pair (λ, h) such that for any filtered C∗-algebra A, then D1 • Moreover, we can choose a (λ, h)-inverse which is natural: there exists a control pair (αβ, kβ) and for any filtered C∗-algebra A a (λ, h)-controlled morphism B0 ,r>0 : K0(SA) → K1(A) which is an (λ, h)- inverse for D1 B ◦ fS = f ◦ B0 A for any homomorphism f : A → B of filtered C∗-algebras, where fS : SA → SB is the suspension of the homomorphism f . A = (βε,r A and such that B0 A )0<ε< 1 4αβ 3.4. The mapping cones. We end this section by proving that the mapping cones construction can be performed in the framework of quantitative K-theory. Let be a filtered semi-split extension of C∗-algebras. Let us set A/J[0, 1) = C0([0, 1), A/J) and define the mapping cone of q: 0 → J → A → A/J→0 q Cq = {(x, f ) ∈ A ⊕ A/J[0, 1); such that f (0) = q(x)}. Using a semi-split filtered cross-section for q, we see that Cq is filtered by Let us set (Cq ∩ (Ar ⊕ A/J[0, 1))r)r>0 . eq : J → Cq; x 7→ (x, 0) 30 and H. OYONO-OYONO AND G. YU We have then a semi-split extension of filtered C∗-algebras φq : SA/J → Cq; f 7→ (0, f ). 0 → J ej→ Cq π2→ A/J[0, 1) → 0, where π2 is the projection on the second factor of A ⊕ A/J[0, 1). Lemma 3.13. There exists a control pair (λ, h) such that eq,∗ is (λ, h)-invertible for any semi-split extension of filtered C∗-algebras 0 → J → A Proof. The even case is a consequence of theorem 3.11. We deduce the odd case from the even one using corollary 3.12. (cid:3) → A/J → 0. q It is a standard fact in K-theory that the boundary of an extension of C∗-algebras 0 → J → A → A/J → 0 can be obtain using the equality q eq,∗ ◦ ∂J,A = φq,∗ ◦ ∂A/J , where ∂A/J = ∂SA/J,CA/J stands for the boundary map of the extension 0 → SA/J → CA/J → A/J → 0 (corresponding to the evaluation at 1). We have a similar result in quantitative K-theory: Lemma 3.14. With above notations, we have eq,∗ ◦ DJ,A = φq,∗ ◦ DA/J , where DA/J stands for DSA/J,CA/J . Proof. We can assume without loss of generality that A is unital. Let us fix a semi-split filtered cross-section s : A/J → A such that s(1) = 1. Let p be an ε-r projection in A/J. Using the notations of the proof of proposition 3.3, define for t in [0, 1] • xt = lεXl=1 (2ıπts(p))l − t(2ıπ)ls(pl) l! in A; • ft : [0, 1] → A/J : σ 7→ l! Then, (1 + (yt, ft))t∈[0,1] is a path of αε-kεr unitary in C+ Moreover, lεXl=1 ((2ıπ(1 − σ)t + σ)p)l − ((1 − σ)t + σ)(2ıπp)l . q with x0 = 0 and f1 = 0. q • x1 belongs to J and satisfies the conclusion of lemma 3.3 starting from the ε-r-projection p and with respect to the semi-split extension of filtered C∗-algebras 0 → J → A → A/J → 0 and to the semi-split filtered cross- section s; • f0 belongs to SA/J and satisfies the conclusion of lemma 3.3 starting from the ε-r-projection p and with respect to the semi-split extension of filtered C∗-algebras 0 → SA/J → CA/J→A/J → 0 corresponding to evaluation at 1 and to the semi-split filtered cross-section A/J 7→ CA/J; a 7→ [t 7→ ta]. Hence, following the construction of proposition 3.4 in the even case, we obtain that eq,∗ ◦ DJ,A and φq,∗ ◦ DA/J coincide on K0(A/J). Let us check now the odd case. Let u be an ε-r-unitary in Mn(A/J). Pick any ε-r-unitary in some Mj(A/J) such that diag(u, v) is homotopic to In+j in U3ε,2r n+j (A/J). According to lemma 3.2, and up to replace v by diag(v, Ik) for some ON A QUANTITATIVE OPERATOR K-THEORY 31 integer k, there exists an element w in U3αeε,2ke,3εr (A) homotopic to In+j as a 3αeε- 2ke,3εr-unitary and such that kq(w) − diag(u, v)k 6 3αeε. Let (wt)t∈[0,1] be a path in U3αeε,2ke,3εr (A) with w0 = In+j and w1 = w and set yt = q(wt) diag(In, 0)q(w∗t ). As in the proof of proposition 3.4, we see that yt is an element in P12αeε,4ke,3εr (A/J) such that ky1 − diag(In, 0)k 6 9αeε. Define n+j n+j n+j g : [0, 1] → Mn+j(A/J); t 7→ yt − diag(In, 0) − t(y1 − diag(In, 0)). n+j Then g + diag(In, 0) is the element of P12αeε,4ke,3εr (S+A/J) that we get from u and v when we perform the construction of proposition 3.4 in the odd case with respect to the extension 0 → SA/J → CA/J → A/J → 0. Let us set now xt = wt diag(In, 0)w∗t and ht = xt − diag(In, 0) − ts ◦ q(x1 − diag(In, 0)) for t in [0, 1]. Then diag(In, 0) + ht belongs to P12αeε,4ke,3εr (A) and diag(In, 0) + h1 is the element of P12αeε,4ke,3εr (J) that we get from u and v when we perform the construction of proposition 3.4 in the odd case with respect to the extension 0 → J → A → A/J → 0. Eventually, if we define n+j n+j q Ht : [0, 1] → Mn+j(A/J); σ 7→ g(1−σ)t+σ, then ((ht, Ht) + diag(In, 0))t∈[0,1] is a homotopy in P12αeε,4ke,3εr and ((h1, 0) + diag(In, 0)). Thus we obtain the result in the odd case. (C+ n+j q ) between ((0, g) + diag(In, 0)) As a consequence, we get that the controlled suspension morphism is compatible with the controlled boundary maps. Proposition 3.15. There exists a control pair (λ, h) such that for any semi-split extension of filtered C∗-algebras 0 → J → A → A/J → 0, the following diagrams are (λ, h)-commutative: (cid:3) K0(A/J) DA/J−−−−→ K1(SA/J) yDSJ,SA DJ,Ay DJ−−−−→ K0(SJ) K1(J) and K1(A/J) DA/J−−−−→ K0(SA/J) yDSJ,SA DJ,Ay DJ−−−−→ K1(SJ) K0(J) , where DJ and DA/J stands respectively for the controlled suspension morphisms DSJ,CJ and DSA/J,CA/J . Proof. Let qs : SA → SA/J the suspension of the homomorphism q : A → A/J. Applying lemma 3.14 to the extensions 0 → J → A → A/J → 0 and 0 → SJ → SA → SA/J → 0 and using the naturality of controlled boundary maps mentioned 32 H. OYONO-OYONO AND G. YU in remark 3.5, we get eqs,∗ ◦ DSJ,SA ◦ DA/J = φqs,∗ ◦ DSA/J ◦ DA/J = DSCq ◦ φq,∗ ◦ DA/J = DSCq ◦ eq,∗ ◦ DJ,A = eqs,∗ ◦ DJ ◦ DJ,A The proposition is then a consequence of lemma 3.13. (cid:3) 4. Controlled Bott periodicity The aim of this section is to prove that there exists a control pair (λ, h) such that given a filtered C∗-algebra A, then Bott periodicity K0(A) ∼=→ K0(S2A) is induced in K-theory by a (λ, h)-isomorphism K0(A) → K0(S2A). As an application, we use the controlled boundary morphism of proposition 3.4 to close the controlled exact sequence of 3.11 into a six-term (λ, h)-exact sequence for some universal control pair (λ, h). This will be achieved by using the full power of KK-theory. 4.1. Tensorization in KK-theory. Let A be a C∗-algebra and let B be a C∗- algebra filtered by (Br)r>0. Within all this section, we will assume for sake of simplicity that Br is closed for every positive number r (which is the case for Roe algebras and crossed product algebras). Let us define A⊗Br as the closure in the spatial tensor product A⊗B of the algebraic tensor product of A and Br. Then the C∗-algebra A⊗B is filtered by (A⊗Br)r>0. Moreover, if J is a semi-split ideal of A, i.e 0 → J → A → A/J → 0 is a semi-split extension of C∗algebras, then 0 → J⊗B → A⊗B → A/J⊗B → 0 is a semi-split extension of filtered C∗-algebras. Recall from [11] that for C∗- algebras A1, A2 and D, G. Kasparov defined a tensorization map τD : KK∗(A1, A2) → KK∗(A1⊗D, A2⊗D) in the following way: let z be an element in KK∗(A1, A2) represented by a K-cycle (π, T,E), where • E is a right A2-Hilbert module; • π is a representation of A1 into the algebra L(E) of adjointable operators • T is a self-adjoint operator on E satisfying the K-cycle conditions, i.e. of E; [T, π(a)], π(a)(T 2 − IdE ) are compact operators on E for any a in A1. Then τD(z) ∈ KK∗(A1⊗D, A2⊗D) is represented by the K-cycle (π⊗IdD, T⊗IdD,E⊗D). In what follows, we show that if A1 and A2 are C∗-algebras, if B is a fil- tered C∗-algebra and if z is an element in KK∗(A1, A2), then the homomorphism K∗(A1⊗B) → K∗(A2⊗B) provided by left multiplication by τB(z) is induced by a controlled morphism. Moreover, we have some compatibility results with respect to Kasparov product. As an outcome, we obtain a controlled version of the Bott periodicity that induces in K-theory the Bott periodicity. Proposition 4.1. Let A1 and A2 be C∗-algebras, let B be a filtered C∗-algebra and let z be an element in KK1(A1, A2). Then there exists an (αD, kD)-controlled ON A QUANTITATIVE OPERATOR K-THEORY 33 morphism TB(z) = (τ ε,r B (z))0<ε< 1 4αD ,r>0 : K∗(A1⊗B) → K∗(A2⊗B) of degree 1 inducing in K-theory the right multiplication by τB(z). Proof. Recall that z can be indeed represented by a odd A1-A2-K-cycle (π, T,H⊗A2), where H is a separable Hilbert space, π is a representation of A1 in the algebra L(H⊗A2) of adjointable operators of H⊗A2 and T is a self-adjoint operator in L(H⊗A2) satisfying the K-cycle conditions. Let us set PB = IdH⊗A2⊗B +T⊗IdB , πB = π⊗IdB and define the C∗-algebra E(π,T ) = {(x, y) ∈ A1⊗BML(H⊗A2⊗B) such that PB·πB(x)·PB−y ∈ K(H)⊗A2⊗B}. 2 Since PB has no propagation, the C∗-algebra E(π,T ) is filtered by (E(π,T ) r )r>0 with E(π,T ) r = {(x, PB · πB(x) · PB + y); x ∈ A1⊗Br and y ∈ K(H) ⊗ A2⊗Br}. The extension of filtered C∗-algebras (3) 0 −→ K(H) ⊗ A2⊗B −→ E(π,T ) −→ A1 ⊗ B −→ 0 is semi-split by the cross-section s : A1⊗B → E(π,T ); x 7→ (x, PB · πB(x) · PB). Let us show that the associated controlled boundary (degree one) map DK(H)⊗A2⊗B,E(π,T ) : K∗(A1⊗B) → K∗(K(H)⊗A2⊗B) only depends on the class z of (π, T,H ⊗ A2) in KK1(A1, A2). Assume that (π, T,H ⊗ A2[0, 1]) is a A1-A2[0, 1]-K -cycle providing a homotopy between two A1-A2-K-cycles (π0, T0,H ⊗ A2) and (π1, T1,H ⊗ A2). For t ∈ [0, 1] we denote by • et : A2[0, 1] → A2 the evaluation at t; • Ft ∈ L(H ⊗ A2) the fiber at t of an operator F ∈ L(H ⊗ A2[0, 1]); • πt : A1 → L(H ⊗ A2) the representation induced by π at the fiber t; (with P = T +1 • st : A2⊗B → E(πt,Tt); x 7→ (x, Pt,B · πt,B(x) · Pt,B) 2 ); Then the homomorphism E(π,T ) → E(πt,Tt); (x, y) 7→ (x, yt) satisfies the con- ditions of remark 3.5 (with s : A2⊗B → E(π,T ); x 7→ (x, PB · πB(x) · PB) and st : A2⊗B → E(π,Tt)) and thus we get that (IdK(H) ⊗ et⊗IdB)∗ ◦ DK(H)⊗A1⊗B[0,1],E(π,T ) = DK(H)⊗A1⊗B,E(πt ,Tt), and according to lemma 1.27, we deduce that DK(H)⊗A1⊗B2,E(π0,T0) = DK(H)⊗A1⊗B,E(π1,T1) . This shows that for a A1-A2-K-cycle (π, T,H ⊗ A2), then DK(H)⊗A1⊗B,E(π,T ) de- pends only on the class z of (π, T,H ⊗ A2) in KK1(A1, A2). Finally we define TB(z) = (τ ε,r B (z))0<ε< 1 4αD def== M−1 A2⊗B ◦ DK(H)⊗A1⊗B,E(π,T ) , where • (π, T,H⊗A2) is any A1-A2-K-cycles representing z; • MA2⊗B is the Morita equivalence (see example 2.2). 34 H. OYONO-OYONO AND G. YU The result then follows from the observation that up to the Morita equivalence K∗(K(H) ⊗ A2⊗B) ∼=→ K∗(A2⊗B), the boundary ∂K(H)⊗A1⊗B,E(π,T ) corresponding to the exact sequence (3) is induced by right multiplication by τB(z). (cid:3) Remark 4.2. Let B be a filtered C∗-algebra. (i) For any C∗-algebras A1 and A2 and any elements z and z′ in KK1(A1, A2) then TB(z + z′) = TB(z) + TB(z′). (ii) Let 0 → J → A → A/J → 0 be a semi-split extension of filtered C∗- algebras and let [∂J,A] be the element of KK1(A/J, J) that implements the boundary map ∂J,A. Then we have TB([∂J,A]) = DJ⊗B,A⊗B. (iii) For any C∗-algebras A1, A2 and D and any K-cycle (π, T,H⊗A2) for KK1(A1, A2), we have a natural identification between E(πD ,TD) and E(π,T )⊗D. Hence, for any element z in KK1(A1, A2) then TB(τD(z)) = TB⊗D(z). For a a filtered C∗-algebra B and a homomorphism f : A1 → A2 of C∗-algebras, we set fB : A1⊗B → A2⊗B for the filtered homomorphism induced by f . Proposition 4.3. Let B be a filtered C∗-algebra and let A1 and A2 be two C∗- algebras. (i) For any C∗-algebra A′1, any homomorphism of C∗-algebras f : A1 → A′1 (ii) For any C∗-algebra A′2, any homomorphism of C∗-algebras g : A2 → A′2 and any z in KK1(A′1, A2), we have TB(f∗(z)) = TB(z) ◦ fB,∗; and any z in KK1(A1, A2), we have TB(g∗(z)) = gB,∗ ◦ TB(z). Proof. (i) Let A′1 be a filtered C∗-algebra, let f : A1 → A′1 be a homomorphism of C∗-algebras and let (π, T, H ⊗ A2) be an odd A′1-A2-K-cycle. With the notations of the proof of proposition 4.1, the homomorphism f E : Ef ∗(π,T ) → E(π,T ); (x, y) 7→ (fB(x), y) fits in the commutative diagram 0 −−−−→ K(H) ⊗ A2⊗B −−−−→ Ef ∗(π,T ) −−−−→ A1⊗B −−−−→ 0 . f Ey yfB =y 0 −−−−→ K(H) ⊗ A2⊗B −−−−→ E(π,T ) −−−−→ A′1⊗B −−−−→ 0 Moreover fB and f E intertwines the semi-split and filtered cross-sections A1⊗B → Ef ∗(π,T ); x 7→ (x, PB · πB ◦ fB(x) · PB) and A′1⊗B → E(π,T ); x 7→ (x, PB · πB(x) · PB) and thus, we get by remark 3.5 that for all z in KK1(A′1, A2). TB(f∗(z)) = TB(z) ◦ f∗ ON A QUANTITATIVE OPERATOR K-THEORY 35 (ii) Let A′2 be a C∗-algebra and let g : A2 → A′2 be a homomorphism of C∗-algebras. For any element F in L(H ⊗ A2), let us denote by F = F⊗A2IdA′ 2 ∈ L(H ⊗ A2⊗A2 A′2). Notice that H⊗ A2⊗A2A′2 can be viewed as a right A′2-Hilbert-submodule of H⊗A′2 and under this identification, for any F in K(H) ⊗ A2, then F is the restriction to H⊗A2⊗A2A′2 of the homomorphism (IdK(H)⊗g)(F ). Let z be an element of KK1(A1, A2) represented by a K-cycle (π, T,H⊗A2). Consider the A1-A2-K-cycle (π′, T ′,H′⊗A2) with H′ = H1 ⊕ H2 ⊕ H3, where H1, H2 and H3 are three copies of H, π′ = 0 ⊕ 0 ⊕ π and T ′ = IdH1⊗A2 ⊕ IdH2⊗A2 ⊕ T . Then (π′, T ′,H′⊗A2) is again a K-cycle repre- senting z and g∗(z) is represented by the K-cycle (π′′, T ′′,E), where • E = H1 ⊗ A′2LH2 ⊗ A′2LH3 ⊗ A2⊗A2A′2; • π′′ = 0 ⊕ 0 ⊕ π; • T ′′ = IdH1⊗A′ Using Kasparov stabilization theorem, we get that H2⊗A′2L H3⊗A2⊗A2A′2 is isomorphic as a right-A′2-Hilbert module to H⊗ A′2 and hence, using this identification, we can represent g∗(z) using a standard right-A′2-Hillbert module, as in the proof of proposition 4.1. Then, under the above identi- fication H2 ⊗ A′2LH3 ⊗ A2⊗A2A′2 ∼= H ⊗ A′2, gE : E(π,T ) → Eg∗(π,T ) 2 ⊕ IdH2⊗A′ 2 ⊕ T . (x, y) 7→ (x, P ′′Bπ′′(x)P ′′B + (IdK(H′)⊗B⊗g)(y − P ′Bπ′(x)P ′B)) restricts to a homomorphism K(H1⊕H2⊕H3)⊗A2⊗B → K(H1⊕H)⊗A′2⊗B. We get now a commutative diagram 0 −−−−→ K(H1 ⊕ H2 ⊕ H3) ⊗ A2⊗B −−−−→ E(π′,T ′) −−−−→ A1⊗B −−−−→ 0 . −−−−→ E(π′′,T ′′) −−−−→ A1⊗B −−−−→ 0 0 −−−−→ K(H1 ⊕ H) ⊗ A′2⊗B gEy gEy y= Hence, we get by remark 3.5 that DK(H)⊗A′ 2⊗B,E(π′′ ,T ′′ ) = gE,∗ ◦ DK(H)⊗A2⊗B,E(π′,T ′ ). But the restriction of gE to the corner K(H1)⊗A2⊗B of the C∗-algebra K(H1⊕H2⊕H3)⊗A2⊗B is IdK(H1)⊗g⊗IdB. Since the Morita equivalence MA′ 2⊗B : K∗(A′2⊗B) ∼=→ K∗(K(H1 ⊕ H)⊗A′2⊗B) can be implemented by an inclusion of A′2⊗B in a corner of K(H1)⊗A′2⊗B, and similarly for the Morita equivalence MA2⊗B : K∗(A2⊗B) ∼=→ K∗(K(H1 ⊕ H2 ⊕ H3)⊗A2⊗B), we deduce that the two following compositions coincide: K∗(A2⊗B)) and gB,∗−→ K∗(A′2⊗B) 2 ⊗B MA′ −→ K∗(K(H1 ⊕ H)⊗(A′2⊗B)) K∗(A2⊗B) MA2⊗B−→ K∗(K(H1 ⊕ H2 ⊕ H3)⊗A2⊗B) gE,∗−→ K∗(K(H1 ⊕ H)⊗A′2⊗B). 36 H. OYONO-OYONO AND G. YU Hence we get for any z in KK1(A1, A2). TB(g∗(z)) = g∗ ◦ TB(z) (cid:3) Let us now extend the definition of TB to the even case. Consider for a suit- able control pair (αB, kB) and any filtered C∗-algebra A the (αB, kB)-controlled morphism of odd degree BA : K∗(SA) → K∗(A) defined • by B0 • by M−1 A on K0(SA) as in corollary 3.12; A ◦ DK(ℓ2(N))⊗A,T0⊗A on K1(SA) using the Toeplitz extension 0 → K(ℓ2(N)) ⊗ A → T0 ⊗ A→SA → 0 (see the discussion at the end of section 3.2). Then, according to corollary 3.12 and proposition 3.7, there exists a control pair (λ, h) such that BA is a right (λ, h)-inverse for DSA,CA for any filtered C∗-algebra A. Let us set αT = λαB and kT = h ∗ kB. Now, let B be a filtered C∗-algebra, let A1 and A2 be C∗-algebras, then define for any z in KK0(A1, A2) the (αT , kT )-controlled morphism TB(z) = (τ ε,r B )0<ε< 1 4αT ,r>0 : K∗(A1⊗B) → K∗(A2⊗B) by where TB(z) def==BA2⊗B ◦ TB(z ⊗A2 [∂A2 ]) • [∂A2 ] = [∂SA2,CA2] ∈ KK1(A2, SA2) corresponds to the boundary of the • ⊗A2 stands for Kasparov product. exact sequence 0 → SA2 → CA2 → A → 0; Up to compose on the left with ιαD ε,αT ε,kD r,kT r TB(•) also as an (αT , kT )-controlled morphism. Theorem 4.4. Let B be a filtered C∗-algebra, let A1 and A2 be C∗-algebras , we can in the odd case define ∗ (i) For any element z in KK∗(A1, A2), then TB(z) : K∗(A1⊗B) → K∗(A2⊗B) is a (αT , kT )-controlled morphism with same degree as z that induces in K-theory right multiplication by τB(z). (ii) For any elements z and z′ in KK∗(A1, A2) then TB(z + z′) = TB(z) + TB(z′). (iii) Let A′1 be a filtered C∗-algebras and let f : A1 → A′1 be a homomorphism of C∗-algebras, then TB(f∗(z)) = TB(z) ◦ fB,∗ for all z in KK∗(A′1, A2). (iv) Let A′2 be a C∗-algebra and let g : A′2 → A2 be a homomorphism of C∗- algebras then TB(g∗(z)) = gB,∗ ◦ TB(z) for any z in KK∗(A1, A′2). (αT ,kT ) (v) TB([IdA1 ]) (vi) For any C∗-algebra D and any element z in KK∗(A1, A2), we have TB(τD(z)) = TB⊗D(z). ∼ IdK∗(A1⊗B). Proof. Since BA2⊗B is a right (λ, h)-inverse for DSA2⊗B,CA2⊗B, it induces in K- theory a right inverse (indeed an inverse) for the (degree 1) boundary map ∂SA2⊗B,CA2⊗B : K∗(A2⊗B) → K∗(SA2⊗B). ON A QUANTITATIVE OPERATOR K-THEORY 37 But since TB(z⊗A2[∂SA2⊗B,CA2⊗B]) induces in K-theory right multiplication by z⊗A2 [∂SA2⊗B,CA2⊗B], we eventually get that TB(z⊗A2 [∂SA2⊗B,CA2⊗B]) induced in K-theory the composition K∗(A1⊗B) ⊗A1⊗B τB (z) −→ K∗(A2⊗B) and hence we get the first point. ∂SA2 ⊗B,CA2 ⊗B −→ K∗(SA2⊗B) Point (ii) is a consequence of remark 4.2. Point (iii) is a consequence of proposition 4.3. Point (iv) is a consequence of proposition 4.3 and of the naturality of B• (see remark 3.5 and corollary 3.12), point (v) holds by definition of B•. Point (vi) is a consequence of point (iii) of remark 4.2. (cid:3) We end this section by proving the compatibility of TB with Kasparov product. Theorem 4.5. There exists a control pair (λ, h) such that the following holds : let A1, A2 and A3 be C∗-algebras and let B be a filtered C∗-algebra. Then for any z in KK∗(A1, A2) and any z′ in KK∗(A2, A3), we have TB(z⊗A2z′) (λ,h) ∼ TB(z′) ◦ TB(z). Proof. We first deal with the case z even. According to [12, Lemma 1.6.9], there exists a C∗-algebra A4 and homomorphisms θ : A4 → A1 and η : A4 → A2 such that • the element [θ] of KK∗(A4, A1) induced by θ is invertible. • z = η∗([θ]−1). Since θ∗([θ]−1) = [IdA1 ] in KK∗(A1, A1), we get in view of remark 2.5 and of points (iii), (iv) and (v) of theorem 4.4 that (λ,h) ∼ TB(θ∗(z⊗A2z′)) ◦ TB([θ]−1), TB(z⊗A2 z′) with (λ, h) = (α2 , kT ∗ kT ). But by bi-functoriality of KK-theory, we have T θ∗(z⊗A2 z′) = η∗(z′) and then the result is a consequence of points (iii) and (iv) of theorem 4.4. We can proceed similarly when z′ is even. Let us prove now the result when z and z′ are odd. Then [∂A2 ] = [∂SA2,CA2] is an invertible element in KK1(A2, SA2) and z⊗A2z′ = z⊗A2 [∂A2 ]⊗SA2 [∂A2]−1⊗A2z′ and hence using the even case, we get that (4) But TB(z⊗A2z′) (λ,h) ∼ TB([∂A2 ]−1⊗A2 z′) ◦ TB(z⊗A2 [∂A2]). TB([∂A2 ]−1⊗A2 z′) (λ′,h′) = ∼ BA3⊗B ◦ TB([∂A2 ]−1⊗A2z′⊗A3 [∂A3]) BA3⊗B ◦ TB(z′⊗A3 [∂A3]) ◦ TB([∂A2]−1) (5) for some control pair (λ′, h′), depending only on (λ, h) and (αT , kT ), where equation (5) holds by the even case applied to z′⊗A3[∂A3 ] and [∂A2]−1. Hence, for a control pair (λ′′, h′′)-depending only on (λ, h), we get applying the even case to [∂A2]−1 and z⊗A2[∂A2 ] that (6) ∼ BA3⊗B ◦ TB(z′⊗A3 [∂A3]) ◦ TB(z). TB(z⊗A2 z′) (λ′′,h′′) In view of this equation, we deduce the odd case from the controlled Bott period- if we set [∂] = [∂C0(0,1),C0(0,1]] ∈ icity, which will be proved in the next lemma: KK1(C, C0(0, 1)), then there exists a controlled (α, k) such that TA([∂]−1) is an 38 H. OYONO-OYONO AND G. YU (α, k)-inverse for DA for any filtered C∗-algebra A. Indeed, from this claim and since for some control pair (α′, k′), the (αB, kB)-controlled morphism BA is for every filtered C∗-algebra A a right (α′, k′)-inverse for TA([∂]), we get that TA([∂]−1) (α′′,k′′) ∼ BA for some controlled pair (α′′, k′′) depending only on (α′, k′) and (αT , kT ). Noticing by using point (vi) of theorem 4.4, that TA3⊗B([∂]−1) = TB([∂A3 ]−1), the proof of the theorem in the odd case is then by equation (6) a consequence of the even case applied to [∂A3]−1 and z′⊗A3[∂A3 ] (cid:3) 4.2. The controlled Bott isomorphism. We prove in this subsection a con- trolled version of Bott periodicity. The proof use the even case of theorem 4.5 and is needed for the proof of the odd case. Let A = (Ar)r>0 be a filtered C∗-algebra and let us assume that Ar is closed for every positive number r. Let us denote for short as before DSA,CA by DA and [∂SA,CA] by [∂A] for any filtered C∗-algebra A and let us set [∂] = [∂C]. (λ,h) Theorem 4.6. There exists a control pair (α, k) such that for every filtered C∗- algebra A, then TA([∂]−1) is an (α, k)-inverse for DA. Proof. Consider the even element z = [∂]⊗S[∂S] of KK∗(C, S2), where S = C0(0, 1) and S2 = SS. The lemma is a consequence of the following claim: there exists a ∼ TA(z) for any C∗-algebra A. Before control pair (λ, h) such that DSA ◦ DA proving the claim, let us see how it implies the lemma. Notice first that by point (ii) of remark 4.2, we have DA = TA([∂]). Since by associativity of Kasparov product [∂]−1⊗Cz = [∂S], we get from theorem 4.5 applied to the even case, that there exists a control pair (λ′, h′) such that for any filtered C∗-algebra A, then TA(z) ◦ TA([∂]−1) ◦ DA ∼ DSA ◦ DA. Using the claim and since z is an invertible element of KK∗(C, S2), we obtain from theorem 4.5 applied to the even case that there exists a control pair (α, k) such that TA([∂]−1) is a left (α, k)-inverse for DA. Using associativity of the Kasparov product, we see that [∂] = z⊗S2 [∂S]−1. Then applying twice theorem 4.5, on one hand to [∂]−1 and z⊗S2[∂S]−1 and on the other hand to [∂]−1⊗z and [∂S]−1, we get that there exists a control pair (α′, k′) such that TA([∂]) ◦ TA([∂]−1) ∼ TSA([∂]−1) ◦ TSA([∂]). But according to what we have seen before, TSA([∂]−1) ◦ TSA([∂]) Let us now prove the claim. It is known that up to Morita equivalence, [∂A]−1 is the element of KK1(SA, A) corresponding to the boundary element of the Toeplitz extension ∼ IdK∗(SA). (α′,k′) (α,k) (λ,h) 0 → K(ℓ2(N)) ⊗ A → T0 ⊗ A→SA → 0. A : K0(A) → K1(SA) and D1 Let us respectively denote by D0 A : K1(A) → K0(SA) the restriction of DA to K0(A) and K1(A). According to proposition 3.7, there exists a control pair (λ′, h′) such that, on even elements (λ′,h′) TA([∂]−1) ◦ D0 (7) Since [∂S] = [∂]−1⊗z, we get by left composition by TA(z) in equation (7) and by us- ing theorem 4.5 in the even case that there exists a control pair (λ, h) depending only on (λ′, h′) and such that that D1 A (z) : K0(A) → K0(S2A) A (z) (here T 0 ∼ IdK0(A). SA◦D0 ∼ T 0 (λ,h) A A ON A QUANTITATIVE OPERATOR K-THEORY 39 stands for the restriction of TA(z) to K0(A)). For the odd case, we know from corollary 3.12 that there exists a control pair (λ′′, h′′) such that D1 S2A : K1(S2A) → K0(S3A) is (λ′′, h′′)-invertible. Using the previous case, and since by associativity of the Kasparov product, we have [∂A]⊗SAτSA(z) = τA(z)⊗[∂S2A], we get by ap- plying twice theorem 4.5 in the even case that there exists a control pair (λ′′′, h′′′) such that D1 A (z) : K1(A) → K1(S2A) A(z), where T 1 S2A ◦T 1 S2A : K1(S2A) → K0(S3A) is (λ′′, h′′)- is the restriction of TA(z) to K1(A). Since D1 invertible, we get the result by remark 2.5. S2A ◦D0 ∼ D1 SA ◦D1 (λ′′′,h′′′) (cid:3) A 4.3. The six term (λ, h)-exact sequence. Recall from proposition 3.15 that there exists a control pair (λ, h) such that for any semi-split extension of filtered C∗- algebras 0 → J → A → A/J → 0, the following diagrams are (λ, h)-commutative: and K0(A/J) DA/J−−−−→ K1(SA/J) DJ,Ay yDSJ,SA DJ−−−−→ K0(SJ) K1(J) K1(A/J) DA/J−−−−→ K0(SA/J) yDSJ,SA DJ,Ay DJ−−−−→ K1(SJ) K0(J) As a consequence, by using theorem 4.6 and proposition 3.11, we get Theorem 4.7. There exists a control pair (λ, h) such that for any semi-split ex- tension of filtered C∗-algebras 0 −→ J  −→ A q −→ A/J −→ 0, with Ar closed for every positive number r, then the following six-term sequence is (λ, h)-exact K0(J) ∗−−−−→ K0(A) q∗−−−−→ K0(A/J) DJ,Ax K1(A/J) q∗←−−−− K1(A) ∗←−−−− K1(J) DJ,Ay Remark 4.8. Let us consider with notations of section 3.4 the semi-split extension of filtered C∗-algebras φq→ Cq 0 → SA/J (8) where π1 : Cq → A is the projection on the first factor of Cq. Since we have a π2→ A/J[0, 1) → 0, and since semi-split extension of filtered algebras 0 → J A/J[0, 1) is a contractible filtered C∗-algebra, we see in view of theorem 4.7 that ej,∗ : K∗(J) → K∗(Cq) is a controlled isomorphism. It is then plain to check that up to the controlled isomorphism ej,∗ and DA/J : K∗(SA/J) → K∗(A/J), we get from the semi-split extension of filtered C∗-algebras of equation (8) (for a possibly different control pair) the controlled six-term exact sequence of theorem 4.7. π1→ A → 0, ej→ Cq If we apply theorem 4.7 to a filtered and split extension, we get: 40 H. OYONO-OYONO AND G. YU Corollary 4.9. There exists a control pair (λ, h) such that for every split extension of filtered C∗-algebra 0 → J → A → A/J → 0, with Ar closed for every positive number r and any filtered split cross-section s : A/J → A, then K∗(J) ⊕ K∗(A/J) −→ K∗(A); (x, y) 7→ ∗(x) + s∗(y) is (λ, h)-invertible. 5. Quantitative K-theory for crossed product C∗-algebras In this section, we study quantitative K-theory for crossed product C∗-algebras and discuss its applications to K-amenability. Let Γ be a finitely generated group. A Γ-C∗-algebra is a separable C∗-algebra equipped with an action of Γ by automorphisms. Recall that the convolution alge- bra Cc(Γ, A) of finitely supported A-valued functions on Γ admits two canonical C∗- completions, the reduced crossed product A⋊redΓ and the maximal crossed product A⋊maxΓ. Moreover, there is a canonical epimorphism λΓ,A : A⋊maxΓ → A⋊redΓ which is the identity on Cc(Γ, A). 5.1. Lengths and propagation. Recall that a length on Γ is a map ℓ : Γ → R+ such that • ℓ(γ) = 0 if and only if γ is the identity element e of Γ; • ℓ(γγ′) 6 ℓ(γ) + ℓ(γ′) for all element γ and γ′ of Γ. • ℓ(γ) = ℓ(γ−1). In what follows, we will assume that ℓ is a word length arising from a finite generat- ing symmetric set S, i.e ℓ(γ) = inf{d such that γ = γ1 ··· γd with γ1, . . . , γd in S}. Let us denote by B(e, r) the ball centered at the neutral element of Γ with radius r, i.e B(e, r) = {γ ∈ Γ such that ℓ(γ) 6 r}. For any positive number r, we set (A⋊redΓ)r def=={f ∈ Cc(Γ, A) with support in B(e, r)}. Then the C∗-algebra A⋊redΓ is filtered by ((A⋊redΓ)r)r>0. def=={f ∈ Cc(Γ, A) with support in B(e, r)}, then the C∗-algebra setting (A⋊maxΓ)r A⋊maxΓ is filtered by ((A⋊maxΓ)r)r>0 (notice that as sets, (A⋊redΓ)r = (A⋊maxΓ)r). It is straightforward to check that two word lengths give rise for A⋊redΓ (resp. for A⋊maxΓ) to quantitative K-theories related by a (1, c)-controlled isomorphism for a constant c. In the same way, For a homomorphism f : A → B of Γ-C∗-algebras, we denote respectively by fΓ,red : A⋊redΓ → B⋊redΓ and fΓ,max : A⋊maxΓ → B⋊maxΓ the homomorphisms respectively induced by f on the reduced and on the maximal crossed product. −→ A/J −→ 0, we For any semi-split extension of Γ-C∗-algebras 0 −→ J have semi-split extensions of filtered C∗-algebras q  −→ A 0 −→ J ⋊redΓ Γ,red−→ A⋊redΓ qΓ,red−→ A/J ⋊redΓ −→ 0 and 0 −→ J ⋊maxΓ Γ,max−→ A⋊maxΓ qΓ,max−→ A/J ⋊maxΓ −→ 0 and hence, by theorem 4.7, we get: ON A QUANTITATIVE OPERATOR K-THEORY 41 Proposition 5.1. There exists a control pair (λ, h) such that for any semi-split extension of Γ-C∗-algebras 0 −→ J  −→ A q −→ A/J −→ 0, the following six-term sequences are (λ, h)-exact −−−−−→ K0(A⋊redΓ) K0(J ⋊redΓ) Γ,red,∗ DJ⋊red Γ,A⋊red Γx and DJ⋊red Γ,A⋊max Γx K1(A/J ⋊redΓ) qΓ,red,∗←−−−−− K1(A⋊redΓ) K0(J ⋊maxΓ) Γ,max,∗ −−−−−→ K0(A⋊maxΓ) K1(A/J ⋊maxΓ) qΓ,max,∗←−−−−− K1(A⋊maxΓ) qΓ,red,∗ −−−−−→ K0(A/J ⋊redΓ) DJ⋊red Γ,A⋊red Γy Γ,red,∗←−−−−− K1(J ⋊redΓ) qΓ,max,∗ −−−−−→ K0(A/J ⋊maxΓ) DJ⋊max Γ,A⋊max Γy Γ,max,∗←−−−−− K1(J ⋊maxΓ) 5.2. Kasparov transformation. In this subsection we see how a slight modifica- tion of the argument used in section 4.1 allowed to define a controlled version of the Kasparov transformation compatible with Kasparov product. Notice first that every element z of KK Γ ∗ (A, B) can be represented by a K-cycle, (π, T,H ⊗ B), where • H is a separable Hilbert space; • the right Hilbert B-module H ⊗ B is acted upon by Γ; • π is an equivariant representation of A in the algebra L(H⊗B) of ad- jointable operators on H ⊗ B; • T is a self-adjoint operator on H⊗ B satisfying the K-cycle conditions, i.e. [T, π(a)], π(a)(T 2 −IdH⊗B) and π(a)(γ(T )− T ) belongs to K(H)⊗ B, for every a in A and γ ∈ Γ. Let TΓ = T ⊗B IdB⋊redΓ be the adjointable element of (H ⊗ B) ⊗B B⋊redΓ ∼= H ⊗ B⋊redΓ induced by T and let πΓ be the representation of A⋊redΓ in the algebra L(H⊗B⋊redΓ) of adjointable operators of H⊗B⋊redΓ induced by π. Then (πΓ, TΓ,H ⊗ B⋊redΓ) is a A⋊redΓ-B⋊redΓ-K-cycle and the Kasparov transform Γ (z) of this K-cycle in KK∗(A⋊redΓ, B⋊redΓ). In the odd [11] of z is the class J red case, let us set P = IdH⊗B+T . Then P induces an adjointable operator PΓ = P ⊗B IdB⋊redΓ of (H ⊗ B) ⊗B B⋊redΓ ∼= H ⊗ B⋊redΓ. Let us define E(π,T ) = {(x, y) ∈ A⋊redΓ⊕L(H⊗B⋊redΓ) such that PΓ·πΓ(x)·PΓ−y ∈ K(H)⊗B⋊redΓ}. Since PΓ has no propagation, the C∗-algebra E(π,T ) is filtered by (E(π,T ) E(π,T ) )r>0 with = {(x, PΓ · πΓ(x) · PΓ + y); x ∈ (A⋊redΓ)r and y ∈ K(H) ⊗ (B⋊redΓ)r}. 2 r r The extension of C∗-algebras 0 −→ K(H) ⊗ B⋊redΓ −→ E(π,T ) −→ A⋊redΓ −→ 0 is filtered semi-split by the cross-section s : A⋊redΓ → E(π,T ); x 7→ (x, PΓ · πΓ(x) · PΓ). Let us show that DK(H)⊗B⋊redΓ,E(π,T ) only depends on the class of (π, T,H⊗ B) in KK Γ 1 (A, B). Assume that (π, T,H ⊗ B[0, 1]) is a Γ-equivariant A-B[0, 1]-K-cycle 42 H. OYONO-OYONO AND G. YU providing a homotopy between two Γ-equivariant A-B-K-cycles (π0, T0,H⊗ B) and (π1, T1,H ⊗ B). For t ∈ [0, 1] we denote by • et : B[0, 1]⋊redΓ → B⋊redΓ the evaluation at t; • Ft ∈ L(H⊗B⋊redΓ) the fiber at t of an operator F ∈ L(H⊗B[0, 1]⋊redΓ); • πΓ,t the representation of A ⋊red Γ induced by πΓ at the fiber t; • st : A ⋊red Γ → E(πt,Tt); x 7→ (x, PΓ,t · πΓ,t · PΓ,t) (with P = T +1 2 ); Then the homomorphism E(π,T ) → E(πt,Tt); (x, y) 7→ (x, yt) satisfies the con- ditions of remark 3.5 (with s : A⋊redΓ → E(π,T ); x 7→ (x, PΓ · πΓ(x) · PΓ) and st : A ⋊red Γ → E(πt,Tt)) and thus we get that (IdK(H) ⊗ et)∗ ◦ DK(H)⊗B[0,1]⋊redΓ,E(π,T ) = DK(H)⊗B⋊redΓ,E(πt ,Tt), and according to lemma 1.27, we deduce that DK(H)⊗B⋊redΓ,E(π0,T0) = DK(H)⊗B⋊redΓ,E(π1,T1) . This shows that for a Γ-equivariant A-B-K-cycles (π, T,H⊗B), then DK(H)⊗B⋊redΓ,E(π,T ) depends only on the class z of (π, T,H⊗ B) in KK Γ 1 (A, B). Eventually, if we define Γ (z) = M−1 J red B⋊redΓ ◦ DK(H)⊗B⋊redΓ,E(π,T ), where • (π, T,H ⊗ B) is any Γ-equivariant A-B-K-cycles representing z; • MB⋊redΓ is the Morita equivalence (see example 2.2). we get as in section 4.1 Proposition 5.2. Let A and B be Γ-C∗-algebras. Then for any element z of KK Γ 1 (A, B), there is a odd degree (αD, kD)-controlled morphism Γ (z) = (J red,ε,r J red Γ (z))0<ε< 1 4αD ,r>0 : K∗(A ⋊red Γ) → K∗(B⋊redΓ) such that (i) J red (ii) J red Γ (x) induces in K-theory the right multiplication by J red Γ Γ (z); is additive, i.e J red Γ (z + z′) = J red Γ (z) + J red Γ (z′). (iii) Let A′ be a Γ-C∗-algebra and let f : A → A′ be a homomorphism Γ-C∗- algebras, then for any z in KK Γ Γ (f∗(z)) = J red J red 1 (A′, B). Γ (z) ◦ fΓ,red,∗ (iv) Let B′ be a Γ-C∗-algebra and let g : B → B′ be a homomorphism of Γ-C∗-algebras, then J red Γ (g∗(z)) = gΓ,red,∗ ◦ J red 1 (A, B). for any z in KK Γ Γ (z) (v) If 0 → J → A → A/J → 0 is a semi-split exact sequence of Γ-C∗-algebras, let [∂J,A] be the element of KK Γ 1 (A/J, J) that implements the boundary map ∂J,A. Then we have J red Γ ([∂J,A]) = DJ ⋊redΓ,A⋊redΓ. ON A QUANTITATIVE OPERATOR K-THEORY 43 We can now define J red and kJ = kT ∗kD. If A and B are Γ-C∗-algebra and if z is an element in KK Γ then we set with notation of section 4.1 for even element in the following way. Set αJ = αT αD 0 (A, B), Γ Γ (z) = (J red,ε,r J red Γ (z))0<ε< 1 4αT ,r def==TB⋊redΓ([∂]−1) ◦ J red Γ (z ⊗B [∂SB]). According to theorem 4.6, there exists a control pair (λ, h) such that for any Γ-C∗- algebra A, then J red we can assume indeed that J red morphism. As for theorem 4.4, we get. ∼ IdK∗(A⋊redΓ). Up to compose with ιαD ε,αJ ε,kD,εr,kJ ,εr Γ (•) is also, in the odd case a (αJ , kJ )-controlled Γ ([IdA]) (λ,h) ∗ , Theorem 5.3. Let A and B be Γ-C∗-algebras. ∗ (A, B), then (i) For any element z of KK Γ J red Γ (z) : K∗(A ⋊red Γ) → K∗(B⋊redΓ) is a (αJ , kJ )-controlled morphism of same degree as z that induces in K-theory right multiplication by J red ∗ (A, B), then (ii) For any z and z′ in KK Γ Γ (z). J red Γ (z + z′) = J red Γ (z) + J red Γ (z′). and any z in KK Γ (iii) For any Γ-C∗-algebra A′, any homomorphism f : A → A′ of Γ-C∗-algebras (iv) For any Γ-C∗-algebra B′, any homomorphism g : B → B′ of Γ-C∗-algebras Γ (f∗(z)) = J red Γ (g∗(z)) = gΓ,∗ ◦ J red Γ (z) ◦ fΓ,∗. Γ (z). ∗ (A′, B), then J red ∗ (A, B), then J red and any z in KK Γ Using the same argument as in the proof of theorem 4.5, we see that J red Γ is compatible with Kasparov products. Theorem 5.4. There exists a control pair (λ, h) such that the following holds: for every Γ-C∗-algebras A, B and D, any elements z in KK Γ ∗ (A, B) and z′ in KK Γ ∗ (B, D), then J red Γ (z ⊗B z′) (λ,h) ∼ J red Γ (z′) ◦ J red Γ (z). We can perform a similar construction for maximal cross products. Theorem 5.5. Let A and B be Γ-C∗-algebras. (i) For any element z of KK Γ ∗ (A, B), there exists a (αJ , kJ )-controlled mor- Γ (z) = (J max,ε,r phism J max with same degree as z that induces in K-theory right multiplication by J max Γ ,r : K∗(A⋊maxΓ) → K∗(B⋊maxΓ) (z))0<ε< 1 4αJ Γ (ii) For any z and z′ in KK Γ (z) and such that λΓ,B,∗ ◦ J max ∗ (A, B), then (z + z′) = J max (z) = J red (z) + J max J max Γ Γ Γ Γ (z′). Γ (z) ◦ λΓ,A,∗. and any z in KK Γ (iii) For any Γ-C∗-algebra A′, any homomorphism f : A → A′ of Γ-C∗-algebras (iv) For any Γ-C∗-algebra B′, any homomorphism g : B → B′ of Γ-C∗-algebras (f∗(z)) = J max (g∗(z)) = gΓ,max,∗ ◦ J max (z) ◦ fΓ,max,∗. (z). ∗ (A′, B), then J max ∗ (A, B), then J max and any z in KK Γ Γ Γ Γ Γ Moreover, there exists a controlled pair (λ, h) such that, • for any Γ algebra A, then J max Γ ([IdA]) (λ,h) ∼ IdK∗(A⋊maxΓ); 44 H. OYONO-OYONO AND G. YU • For any semi-split extension of Γ algebras 0 → J → A → A/J → 0, then J max Γ ([∂J,A]) (λ,h) ∼ DJ,A. Theorem 5.6. There exists a control pair (λ, h) such that the following holds: for every Γ-C∗-algebras A, B and D, any elements z in KK Γ ∗ (A, B) and z′ in KK Γ ∗ (B, D), then J max Γ (z ⊗B z′) (λ,h) ∼ J max Γ (z′) ◦ J max Γ (z). 5.3. Application to K-amenability. The original definition of K-amenability is due to J. Cuntz [6]. For our purpose, it is more convenient to use the equivalent definition given by P. Julg and A. Valette in [10]. If Γ is a discrete group, let us 0 (C, C) of the K-cycle (IdC, 0, C), where C is provided denote by 1Γ the class in KK Γ with the trivial action on Γ. Definition 5.7. Let Γ be a discrete group. Then Γ is K-amenable if 1Γ can be represented by a K-cycle such that the action of Γ on the underlying Hilbert space is weakly contained in the regular representation. (The previous definition indeed also makes sense for locally compact groups.) Example 5.8. Amenable groups are obviously K-amenable. Typical example on non-amenable K-amenable groups are free groups [6]. More generally, J. L. Tu proved in [17] that group which satisfies the strong Baum-Connes conjecture (i.e with γ = 1) are K-amenable. Examples of such group are groups with the Haagerup property [8] and fundamental groups of compact and oriented 3-manifolds [13]. For a Γ-C∗-algebra B and an element T of L(H⊗B), where H is a separable Hilbert space, let us set TΓ,max = T⊗BIdB⋊maxΓ and TΓ,red = T⊗BIdB⋊redΓ. If A is a Γ-C∗-algebra and π : A → L(H⊗B) is a Γ-equivariant representation, let πΓ,red : A⋊redΓ → L(H⊗B⋊redΓ) and πΓ,max : A⋊maxΓ → L(H⊗B⋊maxΓ) be respectively the reduced and the maximal representation induced by π. Then, we have the following (compare with the proof of [10, proposition 3.4]). Proposition 5.9. Let Γ be a K-amenable discrete group and let A and B be Γ- C∗-algebras. Then any elements of KK Γ ∗ (A, B) can be represented by a K-cycle (π, T,H⊗B) such that the homomorphism πΓ,max : A⋊maxΓ → L(H⊗B⋊maxΓ) factorises through the homomorphism λΓ,A : A⋊maxΓ → A⋊redΓ, i.e there exists a homomorphism such that πΓ,red,max : A⋊redΓ → L(H⊗B⋊maxΓ) As a consequence, for any Γ-C∗-algebra A, then πΓ,max = πΓ,red,max ◦ λΓ,A. is an isomorphism [6]. λΓ,A,∗ : K∗(A⋊maxΓ) → K∗(A⋊redΓ) We have the following analogous result for quantitative K-theory. Theorem 5.10. There exists a control pair (λ, h) such that λΓ,A,∗ : K∗(A⋊maxΓ) → K∗(A⋊redΓ) is a (λ, h)-isomorphism for every Γ-C∗-algebra A. ON A QUANTITATIVE OPERATOR K-THEORY 45 Proof. Let (π, T,H⊗SA) be a Γ-equivariant K-cycle as in proposition 5.9 repre- senting the element [∂A] of KK Γ 1 (A, SA) corresponding to the extension 0 → SA → CA → A → 0. Let then choose πΓ,A,red,max : A⋊redΓ → L(H⊗B⋊maxΓ) such that πΓ,max = πΓ,red,max ◦ λΓ,A. Let us set P = T +IdH⊗SA and then define 2 E(π,T ) red = {(x, y) ∈ A⋊redΓ ⊕ L(H ⊗ SA⋊redΓ) such that PΓ,red · πΓ,red(x) · PΓ,red − y ∈ K(H) ⊗ SA⋊redΓ}, E(π,T ) max = {(x, y) ∈ A⋊maxΓ ⊕ L(H ⊗ SA⋊maxΓ) such that PΓ,max · πΓ,max(x) · PΓ,max − y ∈ K(H) ⊗ SA⋊maxΓ} and E(π,T ) red,max = {(x, y) ∈A⋊redΓ ⊕ L(H ⊗ SA⋊maxΓ) such that PΓ,max · πΓ,red,max(x) · PΓ,max − y ∈ K(H) ⊗ A⋊maxΓ} Then E(π,T ) red , E(π,T ) max and E(π,T ) red,max are respectively filtered by {(x, PΓ,red · πΓ,red(x) · PΓ,red + y); x ∈ A⋊redΓr and y ∈ K(H) ⊗ SA⋊redΓr}, {(x, PΓ,max · πΓ,max(x) · PΓ,max + y); x ∈ SA⋊maxΓr and y ∈ K(H) ⊗ SA⋊maxΓr} and {(x, PΓ,max·πΓ,red,max(x)·PΓ,max + y); x ∈ A⋊redΓr and y ∈ K(H)⊗ SA⋊maxΓr}. Moreover, the extension of C∗-algebras 0 −→ K(H) ⊗ SA⋊redΓ −→ E(π,T ) 0 −→ K(H) ⊗ SA⋊maxΓ −→ E(π,T ) red −→ A⋊redΓ −→ 0, max −→ A⋊maxΓ −→ 0 and 0 −→ K(H) ⊗ SA⋊maxΓ −→ E(π,T ) red,max −→ A⋊redΓ −→ 0 provided by the projection on the first factor are respectively semi-split by the filtered cross-sections sred : A⋊redΓ → E(π,T ) smax : A⋊maxΓ → E(π,T ) red ; x 7→ (x, PΓ,red · πΓ,red(x) · PΓ,red), max ; x 7→ (x, PΓ,max · πΓ,max(x) · PΓ,max) and sred,max : A⋊redΓ → E(π,T ) max ; x 7→ (x, PΓ,max · πΓ,red,max(x) · PΓ,max). Let us set and f1 : E(π,T ) max → E(π,T ) red,max : (x, y) 7→ (λΓ,A,∗(x), y) f2 : E(π,T ) red,max → E(π,T ) red : (x, y) 7→ (x, y⊗A⋊maxΓIdA⋊redΓ). =y λΓ,K(H)⊗SAy f1y f2y yλΓ,A y= 46 H. OYONO-OYONO AND G. YU The the three above extensions fit in a commutative diagram 0 −−−−→ K(H) ⊗ SA⋊maxΓ −−−−→ E(π,T ) max −−−−→ A⋊maxΓ −−−−→ 0 0 −−−−→ K(H) ⊗ SA⋊maxΓ −−−−→ E(π,T ) red,max −−−−→ A⋊redΓ −−−−→ 0 0 −−−−→ K(H) ⊗ SA⋊redΓ −−−−→ E(π,T ) red −−−−→ A⋊redΓ −−−−→ 0 which satisfy the conditions of remark 3.5 relatively to sred, smax and sred,max, and hence we deduce (9) and (10) DK(H)⊗SA⋊maxΓ,E(π,T ) red,max ◦ λA,Γ,∗ = DK(H)⊗SA⋊maxΓ,E(π,T ) max λK(H)⊗SA,Γ,∗ ◦ DK(H)⊗SA⋊maxΓ,E(π,T ) red,max = DK(H)⊗SA⋊redΓ,E(π,T ) red Let us set then D′A = M−1 SA⋊maxΓ ◦ DSA⋊maxΓ,E(π,T ) red,max : K∗(A⋊redΓ) → K∗(SA⋊maxΓ). Since we have by definition of the quantitative Kasparov transformation the equal- ities JΓ,red([∂A]) = M−1 SA⋊redΓ ◦ DSA⋊redΓ,E(π,T ) red and JΓ,max([∂A]) = M−1 SA⋊maxΓ ◦ DSA⋊maxΓ,E(π,T ) max , we deduce by using equations (9) and (10), theorems 5.3, 5.5, 5.4 and 5.6 and naturality of Morita equivalence, that there exists a control pair (λ, h) such that JΓ,max([∂A]−1) ◦ D′A is a (α, h)-inverse for λΓ,A,∗. (cid:3) 6. The quantitative Baum-Connes conjecture In this section, we formulate a quantitative version for the Baum-Connes con- jecture and we prove it for a large class of groups. 6.1. The Rips complex. Let Γ be a finitely generated group equipped with a lenght ℓ arising from a finite and symmetric generating set. Recall that for any positive number d, then the d-Rips complex Pd(Γ) is the set of finitely supported probability measures on Γ with support of diameter less than d for the distance induced by ℓ. We equip Pd(Γ) with the distance induced by the norm khk = sup{kh(γ)k; γ ∈ Γ} for h ∈ C0(Γ, C). Since ℓ is a proper function, i.e. B(e, r) is finite for every positive number r, we see that Pd(Γ) is a finite dimension and locally finite simplicial complexe and the action of Γ by left translations is simplicial, proper and cocompact. Let us denote by • Vd(Γ) the closed subset of elements of Pd(Γ) with support in B(e, d). • Wd(Γ) the closed subset of elements of Pd(Γ) with support in B(e, 2d); Then Vd(Γ) is a compact subset of Wd(Γ) and contains a fundamental domain for the action of Γ on Pd(Γ). Lemma 6.1. The compact Vd(Γ) is contained in the interior of Wd(Γ). ON A QUANTITATIVE OPERATOR K-THEORY 47 Proof. Let h be an element in Vd(Γ) and choose an element γ in B(e, d) such that h(γ) > 0. Then if g is an element of Pd(Γ) such that kg − hk < h(γ), we get that g(γ) 6= 0 and thus every element γ′ of the support of g satisfies ℓ(γ−1γ′) < d. Hence g belongs to Wd(Γ). (cid:3) Lemma 6.2. There is a continuous function φ : Pd(Γ) → [0, 1] compactly supported in Wd(Γ) such that γ(φ) = 1. Xγ∈Γ Proof. Let ψ : Pd(Γ) → [0, 1] a continuous function compactly supported in the interior of Wd(Γ) and such that ψ(x) = 1 if x belongs to Vd(Γ). Since Vd(Γ) contains a fundamental domain for the action of Γ on Pd(Γ), we get that Pγ∈Γ ψ(γx) > 0 for all x in Pd(Γ) (notice that the sum Pγ∈Γ ψ(γx) is locally finite). We define ψ(x) then φ(x) = Pγ∈Γ ψ(γx) for any x in Pd(Γ). (cid:3) Let us define sΓ,d as the cardinality of the finite set Then for any function φ as in lemma 6.2, the function {γ ∈ Γ such that γWd(Γ) ∩ Wd(Γ) 6= ∅}. eφ : Γ → C0(Pd(Γ)); γ 7→Xγ∈Γ φ1/2γ(φ1/2) is a projection of C0(Pd(Γ))⋊redΓ with propagation less than sΓ,d. Moreover, since the set of function satisfying the condition of lemma 6.2 is an affine space, we get that for any positive number ε and r with ε < 1/4 and r > sΓ,d, the class [eφ, 0]ε,r ∈ K ε,r 0 (C0(Pd(Γ))⋊redΓ) does not depend on the chosen function φ. Let us set then rΓ,d,ε = kJ ,ε/αJ sΓ,d. Recall that kJ can be chosen non increasing and in this case, rΓ,d,ε is non decreasing in d and non increasing in ε. Definition 6.3. For any Γ-C∗-algebra A and any positive numbers ε, r and d with ε < 1/4 and r > rΓ,d,ε, we define the quantitative assembly map µε,r,d Γ,A,∗ : KK Γ ∗ (C0(Pd(Γ)), A) → K ε,r 7→ (cid:0)J z Γ ∗ (A ⋊red Γ) red, ε αJ , r kJ,ε/αJ (z)(cid:1)(cid:18)[eφ, 0] ε αJ kJ,ε/αJ (cid:19) . r , Then according to theorem 5.3, the map µε,r,d Γ,A is a homomorphism of groups (resp. semi-groups) in even (resp. odd) degree. For any positive numbers d and d′ such that d 6 d′, we denote by qd,d′ : C0(Pd′(Γ)) → C0(Pd(Γ)) the homomorphism induced by the restriction from Pd′ (Γ) to Pd(Γ). It is straightforward to check that if d, d′ and r are positive numbers such that d 6 d′ and r > rΓ,d′,ε, then Γ,A = µε,r,d′ µε,r,d Γ,A ◦ qd,d′,∗. Moreover, for every positive numbers ε, ε′, d, r and r′ such that ε 6 ε′ 6 1/4, rΓ,d,ε 6 r, rΓ,d,ε′ 6 r′, and r < r′, we get by definition of a controlled morphism that (11) ιε,ε′,r,r′ ∗ ◦ µε,r,d Γ,A,∗ = µε′,r′,d Γ,A,∗ . 48 H. OYONO-OYONO AND G. YU Furthermore, the quantitative assembly maps are natural in the Γ-C∗-algebra, i.e. if A and B are Γ-C∗-algebras and if φ : A → B is a Γ-equivariant homomorphism, then φΓ,red,∗,ε,r ◦ µε,r,d Γ,A,∗ = µε,r,d Γ,B,∗ ◦ φ∗ for every positive numbers r and ε with r > rΓ,d,ε and ε < 1/4. These quantitative assembly maps are related to the usual assembly maps in the following way: recall from [2] that there is a bunch of assembly maps with coefficients in a Γ-C∗-algebra A defined by µd Γ,A,∗ : KK Γ ∗ (C0(Pd(Γ)), A) → K∗(A ⋊red Γ) z 7→ [eφ] ⊗C0(Pd(Γ))⋊Γ JΓ(z). For every positive numbers r and ε with r > rΓ,d,ε and ε < 1/4, we have (12) ∗ ◦ µε,r,d ιε,r Γ,A,∗ Γ,A,∗ ◦ qd,d′,∗ = µd Γ,A,∗ for all positive numbers d and d′ with d 6 d′, the family of assembly maps (µd Γ,A)d>0 gives rise to a homomorphism Recall that since µd′ Γ,A,∗. = µd µΓ,A,∗ : lim d>0 KK Γ ∗ (C0(Pd(Γ)), A) −→ K∗(A ⋊red Γ) called the Baum-Connes assembly map. 6.2. Quantitative statements. Let us consider for a Γ-C∗-algebra A and positive numbers d, d′, r, r′, ε and ε′ with d 6 d′, ε′ 6 ε < 1/4, rΓ,d,ε 6 r and r′ 6 r the following statements: QIΓ,A,∗(d, d′, r, ε): for any element x in KK Γ 0 in K ε,r ∗ (A ⋊red Γ) implies that q∗d,d′(x) = 0 in KK Γ ∗ (C0(Pd(Γ)), A), then µε,r,d Γ,A,∗ ∗ (C0(Pd′ (Γ)), A). (x) = (A ⋊red Γ), there exists an element QSΓ,A,∗(d, r, r′, ε, ε′): for every y in K ε′,r′ ∗ x in KK Γ ∗ (C0(Pd(Γ)), A) such that ∗ Using equation (12) and remark 1.18 we get µε,r,d Γ,A,∗ (x) = ιε′,ε,r′,r (y). Proposition 6.4. Assume that for all positive number d there exists a positive number ε with ε < 1/4 for which the following holds: for any positive number r with r > rΓ,d,ε, there exists a positive number d′ with d′ > d such that QIΓ,A(d, d′, r, ε) is satisfied. Then µΓ,A,∗ is one-to-one. We can also easily prove the following: Proposition 6.5. Assume that there exists a positive number ε′ with ε′ < 1/4 such that the following holds: for any positive number r′ , there exist positive numbers ε, d and r with ε′ 6 ε < 1/4, rΓ,d,ε 6 r and r′ 6 r such that QSΓ,A(d, r, r′, ε, ε′) is true. Then µΓ,A,∗ is onto. The following results provide numerous examples of finitely generated groups that satisfy the quantitative statements. ON A QUANTITATIVE OPERATOR K-THEORY 49 Theorem 6.6. Let A be a Γ-C∗-algebra. Then the following assertions are equiv- alent: (i) µΓ,ℓ∞(N,K(H)⊗A),∗ is one-to-one, (ii) For any positive numbers d, ε and r > rΓ,d,ε with ε < 1/4 and r > rΓ,d, there exists a positive number d′ with d′ > d for which QIΓ,A(d, d′, r, ε) is satisfied. Proof. Assume that condition (ii) holds. Let x be an element in some KK Γ ∗ (C0(Pd(Γ)), ℓ∞(N,K(H) ⊗ A)) such that µd Γ,ℓ∞(N,K(H)⊗A),∗(x) = 0. Using equation (12), we get that ιε′,r′ (x)) = 0 for any ε′ in (0, 1/4) and r′ > rΓ,d,ε′ and hence, by remark 1.18, we can find ε and r > rΓ,d,ε such that µε,r,d (x) = 0. Recall from [14, Proposition 3.4] that we have an iso- Γ,ℓ∞(N,K(H)⊗A),∗ morphism (µε′,r′,d Γ,A,∗ ∗ (13) KK Γ 0 (C0(Pd(Γ)), ℓ∞(N,K(H) ⊗ A)) ∼=−→ KK Γ 0 (C0(Pd(Γ)), A)N induced on the j th factor and up to the Morita equivalence KK Γ 0 (C0(Pd(Γ)), A) ∼= KK Γ 0 (C0(Pd(Γ)),K(H) ⊗ A) by the j th projection ℓ∞(N,K(H) ⊗ A) → K(H) ⊗ A. Let (xi)i∈N be the ele- ment of KK Γ 0 (C0(Pd(Γ)), A)N corresponding to x under this identification and let d′ > d be a number such that QIΓ,A(d, d′, r, ε) holds. Naturality of the quanti- tative assembly maps implies that µε,r,d (xi) = 0 and hence that qd,d′,∗(xi) = 0 Γ,A,∗ in KK Γ ∗ (C0(Pd′ (Γ)), A) for every integer i. Using once again the isomorphism of ∗ (C0(Pd′ (Γ)), ℓ∞(N,K(H)⊗ A) and equation (13), we get that qd,d′,∗(x) = 0 in KK Γ hence µΓ,ℓ∞(N,K(H)⊗A),∗ is one-to-one. Let us prove the converse in the even case, the odd case being similar. As- sume that there exists positive numbers d, ε and r with ε < 1/4 and r > rΓ,d,ε and such that for all d′ > d, the condition QIΓ,A(d, d′, r, ε) does not hold. Let us prove that µΓ,ℓ∞(N,K(H)⊗A),∗ is not one-to-one. Let (di)i∈N be an increasing and unbounded sequence of positive numbers such that di > d for all integer i. For all integer i, let xi be an element in KK Γ (xi) = 0 in K0(A ⋊red Γ) and qd,di,∗(xi) 6= 0 in KK Γ 0 (C0(Pdi (Γ)), A). Let x be the element of 0 (C0(Pd(Γ)), ℓ∞(N,K(H) ⊗ A)) corresponding to (xi)i∈N under the identifica- KK Γ tion of equation (13). Let (pi)i∈N be a family of ε-r-projections, with pi in some Mli( ^A ⋊red Γ) and n an integer such that 0 (C0(Pd(Γ)), A) such that µε,r,d Γ,A,∗ µε,r,d Γ,ℓ∞(N,K(H)⊗A),∗ (x) = [(pi)i∈N, n]ε,r 0 (ℓ∞(N,K(H) ⊗ A)⋊redΓ). By naturality of µε,r,d in K ε,r , we get that [pi, n]ε,r = 0 Γ,•,∗ in K ε,r 0 (A ⋊red Γ) for all integer i. We see by using proposition 1.31 that then ιε,r ∗ ([(pi)i∈N, n]) = 0 in K0(ℓ∞(N,K(H) ⊗ A)⋊redΓ). We eventually obtain that Γ,A(x) = ιε,r µd Γ,A (x) = 0. Since qd,di,∗(x) 6= 0 for every integer i, we get that µΓ,ℓ∞(N,K(H)⊗A),∗ is not one-to-one. Theorem 6.7. There exists λ > 1 such that for any Γ-C∗-algebra, the following assertions are equivalent: ∗ ◦ µε,r,d (cid:3) (i) µΓ,ℓ∞(N,K(H)⊗A),∗ is onto; 50 H. OYONO-OYONO AND G. YU (ii) For any positive numbers ε and r′ with ε < 1 4λ , there exist positive num- bers d and r with rΓ,d,ε 6 r and r′ 6 r for which QSΓ,A(d, r, r′, λε, ε) is satisfied. ∗ (y) = z, with 0 < ε < 1 Proof. Choose λ as in remark 1.18. Assume that condition (ii) holds. Let z be an el- (ℓ∞(N,K(H)⊗ 4λ and r′ > 0. Let yi be the image ement in K∗(ℓ∞(N,K(H)⊗A)⋊redΓ) and let y be an element in K ε,r′ ∗ A)⋊redΓ) such that ιε,r′ of y under the composition (14) K ε,r′ (A ⋊red Γ), where the first map is induced by the evaluation ℓ∞(N,K(H) ⊗ A) −→ K(H) ⊗ A at i and the second map is the Morita equivalence of proposition 1.29. Let d and r be numbers with r > r′ and r > rΓ,d,ε and such that QSΓ,A(d, r, r′, λε, ε) holds. Then for any integer i, there exists a xi in KK Γ ∗ (C0(Pd(Γ)), A) such that µλε,r,d Γ,A,∗ (ℓ∞(N,K(H)⊗A)⋊redΓ) → K ε,r′ ∗ (K(H)⊗A ⋊red Γ) ∼=→ K ε,r′ ∗ (xi) = ιε,λε,r′,r ∗ ∗ (yi) in K ε,r x ∈ KK Γ ∗ (A⋊redΓ). Let ∗ (C0(Pd(Γ)), ℓ∞(N,K(H) ⊗ A)) be the element corresponding to (xi)i∈N under the identification of equation (13). By naturality of the quantitative assembly maps, we get according to proposition 1.31 and up to replace λ by 3λ (for the odd case) that (x) = ιε,λε,r′,r (y) in K ε,r µλε,r,d Γ,ℓ∞(N,K(H)⊗A)),∗ ∗ ∗ (ℓ∞(N,K(H) ⊗ A)⋊redΓ). We have hence Γ,ℓ∞(N,K(H)⊗A)),∗(x) = ιε,r′ µd ∗ (y) = z, and therefore µΓ,ℓ∞(N,K(H)⊗A),∗ is onto. Let us prove the converse in the even case, the odd case being similar. Assume that there exist positive numbers ε and r′ with ε < 1 4λ such that for all positive numbers r and d with r > r′ and r > rΓ,d,ε, then QSΓ,A(d, r, r′, λε, ε) does not hold. Let us prove then that µΓ,ℓ∞(N,K(H)⊗A),∗ is not onto. Let (di)i∈N and (ri)i∈N be increasing and unbounded sequences of positive numbers such that ri > rΓ,di,λε and ri > r′. Let yi be an element in K ε,r′ (yi) is not in the range of µλε,ri,di (ℓ∞(N,K(H)⊗ A)⋊redΓ) such that for every integer i, the image of y under the composition of equation (14) 0 (C0(Pd′(Γ)), ℓ∞(N,K(H)⊗ A)) is yi. Assume that for some d′, there is an x in KK Γ such that ιε,r′ (x). Using remark 1.18, we see that there exists a positive number r with r′ 6 r and rΓ,d′,λε 6 r and such that . There exists an element y in K ε,r′ (A ⋊red Γ) such that ιε,λε,r′,ri Γ,ℓ∞(N,K(H)⊗A),∗ (y) = µd′ Γ,A,∗ ∗ ∗ 0 0 ιε,λε,r′,r ∗ ◦ µε,r′,d′ Γ,ℓ∞(N,K(H)⊗A),∗ (x) = ιε,λε,r′,r ∗ (y). But then, if we choose i such that ri > r and di > d′ we get by using naturality of the assembly map and equation (11) that ιε,λε,r′,ri (yi) belongs to the image of µλε,ri,di Γ,A,∗ , which contradicts our assumption. ∗ (cid:3) Replacing in the proof of (ii) implies (i) of theorems 6.6 and 6.7 the algebra ℓ∞(N,K(H) ⊗ A) by Qi∈N(K(H) ⊗ Ai) for a family (Ai)i∈N of Γ-C∗-algebras, we can prove the following result. ON A QUANTITATIVE OPERATOR K-THEORY 51 Theorem 6.8. Let Γ be a discrete group. (i) Assume that for any Γ-C∗-algebra A, the assembly map µΓ,A,∗ is one-to- one. Then for any positive numbers d, ε and r > rΓ,d,ε with ε < 1/4 and r > rΓ,d, there exists a positive number d′ with d′ > d such that QIΓ,A(d, d′, r, ε) is satisfied for every Γ-C∗-algebra A; (ii) Assume that for any Γ-C∗-algebra A, the assembly map µΓ,A,∗ is onto. Then for some λ > 1 and for any positive numbers ε and r′ with ε < 1 4λ , there exist positive numbers d and r with rΓ,d,ε 6 r and r′ 6 r such that QSΓ,A(d, r, r′, λε, ε) is satisfied for every Γ-C∗-algebra A. In particular, if Γ satisfies the Baum-Connes conjecture with coefficients, then Γ satisfies points (i) and (ii) above. Recall from [16, 20] that if Γ coarsely embeds in a Hilbert space, then µΓ,A,∗ is one-to-one for every Γ-C∗-algebra A. Hence we get: Corollary 6.9. If Γ coarsely embeds in a Hilbert space, then for any positive num- bers d, ε and r > rΓ,d,ε with ε < 1/4 and r > rΓ,d, there exists a positive number d′ with d′ > d such that QIΓ,A(d, d′, r, ε) is satisfied for every Γ-C∗-algebra A; The quantitative assembly maps admit maximal versions defined with notations of definition 6.3 for any Γ-C∗-algebra A and any positive number ε, r and d with ε < 1/4 and r > rΓ,d,ε, as µε,r,d Γ,A,max,∗ ∗ (A⋊maxΓ) : KK Γ ∗ (C0(Pd(Γ)), A) → K ε,r 7→ (cid:0)J z Γ max, ε αJ , r kJ,ε/αJ (z)(cid:1)(cid:18)[eφ, 0] ε αJ kJ,ε/αJ (cid:19) . r , As in the reduced case, we have using the same notations • for any positive number d and d′ such that d 6 d′, then µε,r,d Γ,A,max,∗ = µε,r,d′ Γ,A,max,∗ ◦ qd,d′,∗. • for every positive numbers ε, ε′, d, r and r′ such that ε 6 ε′ 6 1/4, rΓ,d,ε 6 r, rΓ,d,ε′ 6 r′, and r < r′, then ◦ µε,r,d ιε,ε′,r,r′ ∗ Γ,A,max,∗ = µε′,r′,d Γ,A,max,∗ . • the maximal quantitative assembly maps are natural in the Γ-C∗-algebras. Moreover, by theorem 5.5(i), the maximal quantitative assembly maps are compat- ible with the reduced ones, i.e µε,r,d . The surjectivity of the Γ,A,max,∗ Γ,A,∗ Baum-Connes assembly map µΓ,A,∗ implies that the map Γ,A,∗ ◦ µε,r,d = λε,r is onto. We have a similar statement in the setting of quantitative K-theory. λΓ,A,∗ : K∗(A⋊maxΓ) → K∗(A⋊redΓ) Theorem 6.10. There exists λ > 1 such the following holds : let Γ be a discrete group and assume that for any Γ-C∗-algebra A, the assembly map µΓ,A,∗ is onto. Then for any positive numbers ε and r, with ε < 1 4λ , there exists a positive number r′ with r′ > r such that • for any Γ-C∗-algebra A; • for any x in K ε,r there exists y in K λε,r′ ∗ ∗ (A⋊redΓ), (A⋊maxΓ) such that λλε,r′ Γ,A,∗ (y) = ιε,λε,r,r′ ∗ (x). 52 H. OYONO-OYONO AND G. YU 7. Further comments The definition of quantitative K-theory can be extended to the framework of filtered Banach algebras, i.e. Banach algebra A equipped with a family (Ar)r>0 of linear subspaces indexed by positive numbers such that: • Ar ⊂ Ar′ if r 6 r′; • Ar · Ar′ ⊂ Ar+r′ ; • the subalgebra [r>0 Ar is dense in A. Since we no more have an involution, we need to introduce instead a norm control for almost idempotents. Let ε be in (0, 1/4) and let r and N be positive numbers. An element e of A is an ε-r-N -idempotent if • e is in Ar; • ke2 − ek < ε; • kek < N ; Similarly, if A is a unital, an element x in A is called ε-r-N -invertible if • x is in Ar; • kxk < N ; • there exists an element y in Ar such that kyk < N , kxy − 1k < ε and kyx − 1k < ε. Quantitative K-theory can then be defined in the setting of ε-r-N -idempotents and of ε-r-N -invertibles. We obtain in this way a bunch of abelian semi-groups (K ε,r,N ∗ (A))ε∈(0,1/4),r>,N >1. Let us set for a fixed N > 1 (A))ε∈(0,1/4),r>0. KN ∗ (A) = (K ε,r,N ∗ If A is a filtered C∗-algebra and e an ε-r-N -idempotent in A, then there is an obvious 0 (A). Approximating ((2e∗ − 1)(2e − 1) + (1, 1)-controlled morphism K0(A) → KN 1)1/2e((2e∗ − 1)(e − 1) + 1)−1/2 by using a power serie (compare with the proof of lemma 1.10), we get that for every N > 1, there exists a control pair (λN , hN ) such that K0(A) → KN 0 (A) is a (λN , hN )-controlled isomorphism. Using the polar decomposition, we have a similar statement in the odd case. References [1] Michael Atiyah. Elliptic operators, discrete groups and von Neumann algebras. In Col- loque "Analyse et Topologie" en l'Honneur de Henri Cartan (Orsay, 1974), pages 43 -- 72. Ast´erisque, No. 32 -- 33. Soc. Math. France, Paris, 1976. [2] Paul Baum, Alain Connes, and Nigel Higson. Classifying space for proper actions and K- theory of group C ∗-algebras. In C ∗-algebras: 1943 -- 1993 (San Antonio, TX, 1993), volume 167 of Contemp. Math., pages 240 -- 291. Amer. Math. Soc., Providence, RI, 1994. [3] Alain Connes. A survey of foliations and operator algebras. In Operator algebras and ap- plications, Part I (Kingston, Ont., 1980), volume 38 of Proc. Sympos. Pure Math., pages 521 -- 628. Amer. Math. Soc., Providence, R.I., 1982. [4] Alain Connes and Georges Skandalis. The longitudinal index theorem for foliations. Publ. Res. Inst. Math. Sci., 20(6):1139 -- 1183, 1984. [5] Alain Connes and Henri Moscovici. Cyclic cohomology, the Novikov conjecture and hyperbolic groups. Topology, 29(3):345 -- 388, 1990. [6] Joachim Cuntz. K-theoretic amenability for discrete groups. J. Reine Angew. Math., 344:180 -- 195, 1983. [7] Guihua Gong, Qin Wang, and Guoliang Yu. Geometrization of the strong Novikov conjecture for residually finite groups. J. Reine Angew. Math., 621:159 -- 189, 2008. ON A QUANTITATIVE OPERATOR K-THEORY 53 [8] Nigel Higson and Gennadi Kasparov. E-theory and KK-theory for groups which act properly and isometrically on Hilbert space. Invent. Math., 144(1):23 -- 74, 2001. [9] Nigel Higson, John Roe, and Guoliang Yu. A coarse Mayer-Vietoris principle. Math. Proc. Cambridge Philos. Soc., 114(1):85 -- 97, 1993. [10] Pierre Julg and Alain Valette. K-theoretic amenability for SL2(Qp), and the action on the associated tree. J. Funct. Anal., 58(2):194 -- 215, 1984. [11] Gennadi Kasparov. Equivariant KK-theory and the Novikov conjecture. Invent. Math., 91(1):147 -- 201, 1988. [12] Vincent Lafforgue. K-th´eorie bivariante pour les alg`ebres de Banach, groupoıdes et conjecture de Baum-Connes. Avec un appendice d'Herv´e Oyono-Oyono. J. Inst. Math. Jussieu, 6(3):415 -- 451, 2007. [13] Michel Matthey, Herv´e Oyono-Oyono, and Wolfgang Pitsch. Homotopy invariance of higher signatures and 3-manifold groups. Bull. Soc. Math. France, 136(1):1 -- 25, 2008. [14] Herv´e Oyono-Oyono and Guoliang Yu. K-theory for the maximal Roe algebra of certain expanders. J. Funct. Anal., 257(10):3239 -- 3292, 2009. [15] John Roe. Coarse cohomology and index theory on complete Riemannian manifolds. Mem. Amer. Math. Soc., 104(497), 1993. [16] Georges Skandalis, Jean-Louis. Tu, and Guoliang Yu. The coarse Baum-Connes conjecture and groupoids. Topology, 41(4):807 -- 834, 2002. [17] Jean-Louis Tu. La conjecture de Baum-Connes pour les feuilletages moyennables. K-Theory, 17(3):215 -- 264, 1999. [18] N. E. Wegge-Olsen. K-theory and C ∗-algebras. Oxford Science Publications. The Clarendon Press Oxford University Press, New York, 1993. A friendly approach. [19] Guoliang Yu. The Novikov conjecture for groups with finite asymptotic dimension. Ann. of Math. (2), 147(2):325 -- 355, 1998. [20] Guoliang Yu. The coarse Baum-Connes conjecture for spaces which admit a uniform embed- ding into Hilbert space. Invent. Math., 139:201 -- 240, 2000. Universit´e de Lorraine, Metz , France E-mail address: [email protected] Vanderbilt University, Nashville, USA E-mail address: [email protected]
1409.8523	2	1409	2015-07-08T15:20:10	Unbounded Operators on Hilbert $C^*$-Modules	[ "math.OA", "math.FA" ]	Let $E$ and $F$ be Hilbert $C^$-modules over a $C^$-algebra $\CAlg{A}$. New classes of (possibly unbounded) operators $t:E\to F$ are introduced and investigated. Instead of the density of the domain $\Def(t)$ we only assume that $t$ is essentially defined, that is, $\Def(t)^\bot=\{0\}$. Then $t$ has a well-defined adjoint. We call an essentially defined operator $t$ graph regular if its graph $\Graph(t)$ is orthogonally complemented in $E\oplus F$ and orthogonally closed if $\Graph(t)^{\bot\bot}=\Graph(t)$. A theory of these operators is developed. Various characterizations of graph regular operators are given. A number of examples of graph regular operators are presented ($E=C_0(X)$, a fraction algebra related to the Weyl algebra, Toeplitz algebra, Heisenberg group). A new characterization of affiliated operators with a $C^*$-algebra in terms of resolvents is given.	math.OA	math	UNBOUNDED OPERATORS ON HILBERT C∗-MODULES REN´E GEBHARDT AND KONRAD SCHM UDGEN Abstract. Let E and F be Hilbert C ∗-modules over a C ∗-algebra A. New classes of (possibly unbounded) operators t : E → F are introduced and in- vestigated. Instead of the density of the domain D(t) we only assume that t is essentially defined, that is, D(t)⊥ = {0}. Then t has a well-defined adjoint. We call an essentially defined operator t graph regular if its graph G(t) is orthog- onally complemented in E ⊕ F and orthogonally closed if G(t)⊥⊥ = G(t). A theory of these operators and related concepts is developed. Various character- izations of graph regular operators are given. A number of examples of graph regular operators are presented (E = C0(X), a fraction algebra related to the Weyl algebra, Toeplitz algebra, Heisenberg group). A new characterization of affiliated operators with a C ∗-algebra in terms of resolvents is given. Contents Introduction 1. 2. Orthogonally closed operators on Hilbert C∗-modules 3. Operators on the commutative C∗-algebra C0(X) 4. Graph regular operators 4.1. Definition and basics on graph regular operators 4.2. The (a, a∗, b)-transform 4.3. Quotients of adjointable operators 4.4. Absolute value 4.5. Bounded transform 4.6. Polar decomposition 4.7. Graph regular operators on C0(X) 4.8. Functional calculus of graph regular normal operators 5. Associated operators and affiliated operators 6. Examples 6.1. Matrices of commutative C∗-algebras and its multipliers 6.2. A fraction algebra related to the Weyl algebra 6.3. Unbounded Toeplitz operators 6.4. Heisenberg group References 1 3 9 14 14 15 18 20 20 23 24 26 26 32 32 33 35 37 38 1. Introduction Hilbert C∗-modules are a well established tool in the theory of C∗-algebras and their applications. They have been invented by I. Kaplansky [K53] for commutative C∗-algebras and by W. Paschke [P73] and M. Rieffel [R74] in the general case. Standard textbooks are [L95] and [MT05]. Unbounded operators on Hilbert C∗- modules play an important role for the study of noncompact quantum groups [W91], in KK-theory [BJ83, K97] and in noncommutative geometry [GVF]. 2010 Mathematics Subject Classification. Primary 46 L 08; Secondary 47 D 40, 47 L 05. Key words and phrases. Hilbert C ∗-modules, unbounded operators, affiliated operators. 1 2 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Let E and F be Hilbert C∗-modules over a C∗-algebra A. By an operator from E into F we mean an A-linear and C-linear mapping t of an A-submodule D(t) of E into F . Such an operator t is called regular if t is closed, D(t) is dense in E, D(t∗) is dense in F , and I + t∗t is a bijection of E. Regular operators form the most important class of unbounded operators on Hilbert C∗-modules. A nice presentation of their theory can be found in Chapters 9 and 10 of Lance' book [L95]. Regular operators were invented by S. Baaj [B81, BJ83] and extensively studied by S.L. Woronowicz in two seminal papers [W91, WN92]. Woronowicz considered the case when E is the C∗-algebra A itself and called the corresponding operators affiliated with A. That is, the affiliated operators are precisely the regular operators on the Hilbert C∗-modules E = A. Unbounded operators on Hilbert C∗- modules are studied in [H89], [K97], [K02], [FS10], [KL12], [Pal99], [Pie06]. A generalization of regular operators are the semiregular operators introduced by A. Pal [Pal99]. Note that semiregular operators are always densely defined. The aim of the present paper is to introduce several new classes of unbounded operators on C∗-modules and to develop the basics of their theory. Let us briefly explain the main new concepts by avoiding technical subtleties. Precise definitions and further explanations will be given in the corresponding sections of the text. Suppose that t is an operator from E into F such that its domain D(t) is essential, that is, D(t)⊥ = {0}. Here D(t)⊥ denotes the orthogonal complement of D(t) with respect to the A-valued scalar product of E. Then the operator t has a well-defined adjoint operator t∗. An operator t is called orthogonally closed if its graph G(t) = {(x, tx); t ∈ D(t)} is orthogonally closed, that is, if G(t)⊥⊥ = G(t) in E ⊕ F . An orthogonally closed operator t is called graph regular if its graph G(t) is orthogonally complemented in E ⊕ F . Graph regularity is the most important new concept of operators appearing in this paper. This notion is also of interest in the case when the operator t is bounded and only essentially defined. It should be emphasized that all these operators are not necessarily densely defined! Instead we assume only that their domains are essential, that is, they have trivial orthogonal complements, and we replace the closure of the graph G(t) in E ⊕ F by its double orthogonal complement G(t)⊥⊥. The difference between graph regularity and regularity can be nicely illustrated by the Hilbert C∗-module C0(X) for the C∗-algebra C0(X), where X is a locally compact Hausdorff space. Then both classes of operators are given by multiplica- tion operators: the regular operators by functions of C(X), but the graph regular operators by functions which are only continuous up to a no-where dense set and for which the modulus goes to infinity in neighbourhoods of discontinuities (see Theorem 15). This example shows another essential difference: While a regular operator can be transported into a densely defined closed operator by each induced representation, for a graph regular operator this is only possible for certain rep- resentations. To include such phenomena was one motivation for studying graph regular operators. We will elaborate this elsewhere more in detail. This paper is organized as follows. Section 2 contains basic definitions and facts on essentially defined operators, their adjoints and orthogonally closed operators. Graph regular operators are introduced and studied in Section 4. Basic proper- ties of these operators are derived (Theorems 1 and 2) and two characterizations of graph regular operators in terms of bounded operators are obtained. The first one, the (a, a∗, b)-transform, gives a purely algebraic description of graph regular opera- tors by a triple of adjointable operators (Theorem 7). The second characterization (Theorems 11 and 12) concerns the bounded transform. This transform played already a crucial role in Woronowicz' approach to affiliated operators. To each graph regular operator t : E → F we associate a regular operator t0 : E0 → F0 UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 3 acting between essential submodules E0 and F0 of E and F , respectively. As a byproduct of these considerations we use the notion of essentially defined partial isometries to prove a general result on the polar decomposition of adjointable op- erators (Theorem 13) and of graph regular operators (Theorem 14). Further, we show that quotients of adjointable operators provide a large source of examples of graph regular operators (Theorems 8 and 9). The functional calculus of normal regular operators is extended to normal graph regular operators (Theorem 16). In Section 5 we specialize to the case when E is the C∗-algebra A itself and A is faithfully realized on a Hilbert space H. Then the regular operators on the Hilbert C∗-module E = A are precisely Woronowicz' affiliated operators. Suppose that t is a densely defined closed operator on H. We shall say that t is associated with A if the operators at = (I + t∗t)−1, at∗ = (I + tt∗)−1, and bt = t(I + t∗t)−1 are in the multiplier algebra M(A). A number of results as well as examples and counter- examples on associated and affiliated operators are derived. One of the main new results (Theorem 17) provides a characterization of affiliated operators in terms of resolvents: If λ ∈ ρ(t), then t is affiliated with A if and only if (t−λI)−1 ∈ M(A) and (t − λI)−1A and (t∗ − λ I)−1A are dense in A. This seems to be a useful criterion for proving that operators are affiliated, since the resolvent is better understood in operator theory than the bounded transform. In Sections 3 and 6 we develop various classes of examples. A particular em- phasis is on graph regular operators that are not regular on the corresponding C∗-modules. Section 3 and Subsection 4.7 contain a careful treatment of the com- mutative case E = C0(X), where X is a locally compact Hausdorff space. Among others, essentially defined orthogonally closed operators and graph regular oper- ators are characterized in this case. In Subsection 6.1 we consider some simple examples for matrices over commutative C∗-algebras. In Subsection 6.2 the posi- tion operator Q = x and the momentum operator P = −i d dx become graph regular operators on a C∗-algebra obtained from a fraction algebra. In Subsection 6.3 some unbounded Toeplitz operators are described as graph regular operators on the Toeplitz C∗-algebra, while in Subsection 6.4 a graph regular operator on the C∗-algebra of the Heisenberg group is constructed. Finally, let us fix some notation that will be used in this paper. Throughout, A denotes a C∗-algebra, Ah := {a ∈ A : a∗ = a} is its hermitian part, and A+ := {a ∈ Aa ≥ 0} is its cone of positive elements. If H is a Hilbert space, we denote by B(H) the bounded operators on H, by K(H) the compact operators on H, and by C(H) the set of densely defined closed operators on H. For an operator t ∈ C(H), let D(t) be its domain on H, N (t) its null space, R(t) its range and ρ(t) its resolvent set. 2. Orthogonally closed operators on Hilbert C∗-modules First we recall a standard definition. Definition 1. A (right) pre-Hilbert C∗-module E over A is a right A-module E equipped with a map h., .i : E × E → A satisfying the following conditions: λ(xa) = (λx)a = x(λa), λ ∈ C, x ∈ E, a ∈ A, hαx + βy, zi = α hx, zi + β hy, zi , α, β ∈ C, x, y, z ∈ E, a ∈ A, x, y ∈ E, x, y ∈ E, hx, yai = hx, yi a, hx, yi = hy, xi∗ , hx, xi ≥ 0, x ∈ E, hx, xi = 0 =⇒ x = 0, x ∈ E. 4 REN ´E GEBHARDT AND KONRAD SCHM UDGEN A pre-Hilbert C∗-module E over A is called a Hilbert C∗-module over A, briefly a Hilbert A-module, if (E,k.kE) is complete, where k.kE is the norm on E given by kxkE := k hx, xi k1/2 A , x ∈ E. Example 1. The C∗-algebra A itself is a Hilbert A-module over A by taking the multiplication as right action and the A-valued scalar product ha, bi := a∗b, a, b ∈ E. In this case kakE = k ha, aik1/2 A = kakA for a ∈ E. Suppose that E is a Hilbert A-module. If x, y ∈ E and hx, yi = 0, we write x⊥y. For a subset F of E, the set F ⊥ := {x ∈ E∀y ∈ F : hx, yi = 0} is called the orthogonal complement of F . Clearly, F ⊥ is a submodule of E and x⊥y is equivalent to y⊥x, since hx, yi = hy, xi∗. If F, G are subsets of E, then F ⊆ F ⊥⊥, F ⊆ G =⇒ G⊥ ⊆ F ⊥, F ⊥ = F ⊥⊥⊥. We only verify the last equality F ⊥ = F ⊥⊥⊥. The first inclusion yields F ⊥ ⊆ (F ⊥)⊥⊥. On the other hand, the relation F ⊆ F ⊥⊥ implies that (F ⊥⊥)⊥ ⊆ F ⊥. Therefore, F ⊥ = F ⊥⊥⊥. Let F and G be subsets of E. We set F + G := {f + gf ∈ F, g ∈ G} and write F ⊕ G := F + G if F ⊆ G⊥. Since F ⊆ G⊥ implies G ⊆ G⊥⊥ ⊆ F ⊥, we always have F ⊕ G = G ⊕ F . Since F ⊆ (F ⊥)⊥, it is justified to write F ⊕ F ⊥. Definition 2. A subset F of E is said to be essential if F ⊥ = {0}. A submodule F of E is called orthogonally closed if F = F ⊥⊥ and orthogonally complemented if F ⊕ F ⊥ = E. Since F ⊥ is always closed in E, each orthogonally closed submodule is closed. Example 2. For the C∗-algebra C0(X) of continuous functions on a locally com- pact Hausdorff X space vanishing at infinity each closed ideal is of the form IO := {f ∈ C0(X)∀x ∈ X \ O : f (x) = 0} for some open subset O ⊆ X. The mapping O → IO is a bijection of the open subsets of X onto the closed ideals of C0(X). For x ∈ X we have x ∈ O if and only if there exists f ∈ IO with f (x) 6= 0. The following facts are easily verified: = I(X\O)◦. • I⊥ O • IO is essential in C0(X) if and only if O is dense in X. • IO is orthogonally closed if and only if O coincides with the interior of its • IO is orthogonally complemented if O is closed. closure. The following simple facts will be often used: If F, G are submodules of E, then (F ∩ G)⊥ ⊇ (F ⊥ + G⊥)⊥⊥ and (F + G)⊥ = F ⊥ ∩ G⊥. Further, if F, G are orthogonally closed, then (F ∩ G)⊥ = (F ⊥ + G⊥)⊥⊥. Lemma 1. For any subset F of E, the set F ⊕ F ⊥ is essential. Proof. If x ∈ (F ⊕ F ⊥)⊥ = F ⊥ ∩ F ⊥⊥, then hx, xi = 0 and hence x = 0. (cid:3) Lemma 2. If F, G are submodules of E with F ⊆ G and F ⊥ ∩ G = {0}, then F ⊥ ∩ (G ⊕ G⊥) = G⊥. Proof. Clearly, F ⊥ ⊇ G⊥, so F ⊥ ∩ (G ⊕ G⊥) ⊇ F ⊥ ∩ G⊥ = G⊥. Now assume that x = g + g⊥ ∈ F ⊥ with g ∈ G, g⊥ ∈ G⊥. But then x − g⊥ = g ∈ G ∩ F ⊥, so g = 0 and x = g⊥ ∈ G⊥. (cid:3) UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 5 The direct sum E ⊕ F of two Hilbert A-modules E and F is a right A-module. If we define a mapping h., .iE⊕F : (E ⊕ F ) × (E ⊕ F ) → A by h(e1, f1), (e2, f2)iE⊕F := he1, e2iE + hf1, f2iF , then this module becomes also a Hilbert A-module. Now we turn to operators on Hilbert A-modules. By an operator t from E into F we mean a C-linear and A-linear mapping defined on a right A-submodule D(t) of E, called the domain of t. The symbol t : E → F always denotes an operator from E into F . The C-linearity and A-linearity of t mean that e1, e2 ∈ E, f1, f2 ∈ F, t(λx) = λt(x) and t(xa) = t(x)a for λ ∈ C, x ∈ D(t), a ∈ A. For an operator t : E → F , its null space N (t) := {x ∈ Etx = 0} is a right A-submodule of E, its range R(t) := {txx ∈ D(t)} is a right A-submodule of F and its graph G(t) := {(x, tx)x ∈ D(t)} is a right A-submodule of E ⊕ F . As in the case of ordinary Hilbert space operators we say t is closed if G(t) is closed in E ⊕ F and t is closable if there exists an operator s which is a closed extension of t. In this case there exists a unique closed operator, denoted by t and called the closure of t, such that G(t) = G(t). An operator t : E → E is called positive if htx, xi ≥ 0 for all x ∈ D(t). Definition 3. An operator t : E → F is called essentially defined if D(t) is an essential submodule of E. Suppose that t : E → F is an essentially defined operator. Set D(t∗) := {y ∈ F∃z ∈ E : ∀x ∈ D(t) : htx, yiF = hx, ziE}. Since D(t) is an essential submodule, the element z ∈ E is uniquely determined by y. We define t∗y := z. Then t∗ : F → E is an operator, called the adjoint of t, and htx, yi = hx, t∗yi for x ∈ D(t), y ∈ D(t∗). The operators of the set L(E, F ) := {t : E → FD(t) = E,D(t∗) = F} are called adjointable. Note that L(E) := L(E, E) is a unital C∗-algebra. Definition 4. An essentially defined operator t : E → E is called symmetric if t ⊆ t∗, and self-adjoint if t = t∗. Let v : E ⊕ F → F ⊕ E denote the unitary operator (x, y) 7→ (−y, x). Proposition 1. Suppose that t : E → F is essentially defined. Then: (1) G(t∗) = vG(t)⊥. (2) N (t∗) = R(t)⊥. (3) If t is injective and R(t)⊥ = {0}, then t∗ is injective and (t∗)−1 = (t−1)∗. (4) G(t) ⊕ vG(t∗) is essential. The proof of these statements is similar to the Hilbert space case; we omit the details. Lemma 3. Let r : E → F be essentially defined. Suppose that D(r∗) = F and R(r) is essential. If D ⊆ D(r) is essential, then R(r ↾D) is essential. Proof. Let x ∈ F be such that hx, ryi = 0 for all y ∈ D. Then hr∗x, yi = 0 for all y ∈ D. Since D⊥ = {0} by assumption, we conclude that r∗x = 0. That is, x ∈ N (r∗) = R(r)⊥ = {0}. (cid:3) Proposition 2. Let t, t1, t2 be essentially defined operators from E into F and let s be an essentially defined operator from F into G. Then: 6 REN ´E GEBHARDT AND KONRAD SCHM UDGEN (1) t1 ⊆ t2 implies t∗1 ⊇ t∗2. (2) If t1 + t2 is essentially defined, then (t1 + t2)∗ ⊇ t∗1 + t∗2. If D(t2) ⊆ D(t1) and D(t∗1) = E, then t1 + t2 is essentially defined and (t1 + t2)∗ = t∗1 + t∗2. (3) If st is essentially defined, then (st)∗ ⊇ t∗s∗. If R(t) ⊆ D(s) and D(s∗) = G, then st is essentially defined and (st)∗ = t∗s∗. (4) If t is injective, D(s) ⊆ R(t) and D((t−1)∗) = F , then st is essentially defined and (st)∗ = t∗s∗. Proof. Assertions (1) -- (3) are shown by simple computations. We prove (4). By Lemma 3, the domain D(st) = t−1D(s) is essential. For x ∈ D((st)∗) and y ∈ D(s) we derive hsy, xi =(cid:10)(st)t−1y, x(cid:11) =(cid:10)t−1y, (st)∗x(cid:11) =(cid:10)y, (t−1)∗(st)∗x(cid:11) . Therefore, x ∈ D(s∗) and s∗x = (t−1)∗(st)∗x = (t∗)−1(st)∗x ∈ D(t∗). That is, D((st)∗) ⊆ D(t∗s∗). Now (3) completes the proof. (cid:3) Definition 5. An essentially defined operator p is a projection if p = p2 = p∗. The next proposition characterizes projections and shows that they are in one- to-one correspondence to orthogonally closed submodules. Proposition 3. For a submodule G of E we define an operator pG : E → E by D(pG) := G ⊕ G⊥, pG(x + y) := x, x ∈ G, y ∈ G⊥. Then pG is essentially defined and pG = p2 (pG)∗ is a projection if and only if G is orthogonally closed. G ⊆ p∗G = pG⊥⊥. In particular, pG = Conversely, if p is a projection, then p = pG for some orthogonally closed sub- module G of E. Proof. By Lemma 1, p := pG is essentially defined. Obviously, p2 = p ⊆ p∗. Let z ∈ D(p∗). Then there exists w ∈ E such that hx, zi = hp(x + y), zi = hx + y, wi , x ∈ G, y ∈ G⊥. Equivalently, hx, z − wi = hy, wi for all x ∈ G, y ∈ G⊥, so both sides are zero. The latter is equivalent to z− w ∈ G⊥ and w ∈ G⊥⊥, that is, z ∈ G⊥⊕ G⊥⊥. Therefore, since w = p∗z, the operator p∗ is given by D(p∗) = G⊥⊥ ⊕ G⊥, p∗(x + y) = x, x ∈ G⊥⊥, y ∈ G⊥. This completes the proof of the first half. Now assume that p = p2 = p∗ and let G := R(p) ⊆ D(p). For x = py ∈ G we have px = p2y = py = x. For x ∈ G⊥, we obtain hx, pyi = 0 = h0, yi for y ∈ D(p). Therefore, x ∈ D(p∗) = D(p) and px = p∗x = 0. Thus, we have shown that pG ⊆ p. Further, the assumption p = p2 implies that R(p) = N (1 − p). Hence G = R(p) = N (1 − p) = N (1 − p∗) = R(1 − p)⊥. Therefore, G = G⊥⊥ and pG is self-adjoint. Since p is also self-adjoint and pG ⊆ p, (cid:3) we conclude that p = pG. Definition 6. Let t : E → F be an operator. A submodule D of D(t) is called an essential core for t if G(t ↾D)⊥ = G(t)⊥. By Lemma 1 this is equivalent to the relation (t ↾D)∗ = t∗ if D is essential. Clearly, D is an essential core for t if and only if G(t ↾D)⊥ ⊆ G(t)⊥, or equivalently, if the sum G(t ↾D)⊥ + G(t) is an orthogonal sum G(t ↾D)⊥ ⊕ G(t). Example 3. Let b ∈ L(E, F ) and let D be an essential submodule of E. Then D is an essential core for b, since (b ↾D)∗ ⊇ b∗ and b∗ is everywhere defined. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 7 Definition 7. A (not necessarily essentially defined) operator t : E → F is or- thogonally closed if G(t) is orthogonally closed, that is, if G(t)⊥⊥ = G(t), and orthogonally closable if G(t)⊥⊥ = G(s) for some (orthogonally closed) operator s. By Proposition 1 the adjoint is always orthogonally closed. Theorem 1. Let t : E → F be essentially defined. Then D(t∗) is an essential submodule of F if and only if t is orthogonally closable. Proof. Suppose that D(t∗) is an essential submodule. Then t∗∗ exists and it follows easily that t ⊆ t∗∗. Applying Lemma 1 twice, first to t and then to t∗, we obtain G(t∗∗) = G(t)⊥⊥. Hence t∗∗ is orthogonally closed and t is orthogonally closable. Now suppose that G(t)⊥⊥ = G(s) for some (essentially defined) operator s. Let z ∈ D(t∗)⊥. Then h(−z, 0), (y, t∗y)i = − hz, yi + h0, t∗yi = 0 for all y ∈ D(t∗), so (z, 0) ∈ vG(t∗)⊥ = G(t)⊥⊥ = G(s) and hence z = 0, since s is an operator. (cid:3) Let us define the sets C′o(E, F ) := {t : E → FD(t)⊥ = {0},D(t∗)⊥ = {0}}, Co(E, F ) := {t ∈ C′o(E, F )t is orthogonally closed}, C′o(E) := C′o(E, E), Co(E) := Co(E, E). Then, by Theorem 1, C′o(E, F ) is the set of essentially defined operators t : E → F that are orthogonally closable. Theorem 2. (1) Suppose that t ∈ C′o(E, F ). Then we have: (a) t ⊆ t∗∗, t∗ = t∗∗∗, and G(t)⊥⊥ = G(t∗∗). (b) t = t∗∗ if and only if t ∈ Co(E, F ). (c) N (t∗∗) is orthogonally closed. (d) If R(t) and R(t∗) are essential, then t and t∗∗ are injective, t−1 is essentially defined and orthogonally closable, and (t∗∗)−1 = (t−1)∗∗. (2) Suppose that t ∈ Co(E, F ). Then: (a) N (t) is orthogonally closed. (b) If R(t) and R(t∗) are essential, then t is injective and t−1 is essentially defined and orthogonally closed. Proof. (1a): By Theorem 1, t∗∗ exists. Applying Proposition 1 several times we obtain G(t∗∗) = vG(t∗)⊥ = G(t)⊥⊥, so that t ⊆ t∗∗. This implies that t∗∗∗ exists. Using Proposition 1 once again we get G(t∗∗∗) = vG(t)⊥⊥⊥ = vG(t)⊥ = G(t∗), hence t∗∗∗ = t∗. (1b) follows at once from the equality G(t)⊥⊥ = G(t∗∗). (1c): We derive N (t∗∗)⊥⊥ = R(t∗)⊥⊥⊥ = R(t∗)⊥ = N (t∗∗). (1d): By Proposition 1, t∗ is injective and (t∗)−1 = (t−1)∗. By assumption, R(t∗) is essential. Therefore, using Proposition 1 again we conclude that t∗∗ is injective and (t∗∗)−1 = ((t∗)−1)∗ = (t−1)∗∗. (cid:3) (2a) and (2b) follow from (1c) and (1d), respectively. The operator t∗∗ is called the orthogonal closure of the orthogonally closable operator t. Lemma 4. Let t ∈ Co(F, G) and b ∈ L(E, F ). Suppose that the operator tb is essentially defined. Then tb ∈ Co(E, G) and (tb)∗ = (b∗t∗)∗∗. Proof. From Proposition 2 it follows that (tb)∗ ⊇ b∗t∗ and the latter is essentially defined. Therefore, tb ⊆ (tb)∗∗ ⊆ (b∗t∗)∗ = t∗∗b∗∗ = tb. This implies that (tb)∗ = (b∗t∗)∗∗. (cid:3) Definition 8. An operator t ∈ C′o(E) is called essentially self-adjoint if t∗ = t∗∗. An operator t ∈ Co(E, F ) is called normal if t∗t = tt∗. 8 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Proposition 4. Let t ∈ C′o(E, F ) and let D be a submodule of D(t). If D is an essential core for t, then D is an essential submodule of E and (t ↾D)∗ = t∗, t ⊆ (t ↾D)∗∗ = t∗∗. In particular, D is an essential core for t∗∗ and if t is orthogonally closed, then t = (t ↾D)∗∗. Proof. Define an operator pE : E ⊕ F → E by pE(x, y) := x for x ∈ E, y ∈ F . It is straightforward to check that the adjoint operator p∗E : E → E ⊕ F acts by p∗E(x) = (x, 0) for x ∈ E. Thus, pE ∈ L(E ⊕ F, E). Now, using Proposition 2,(3) for the second equality, Proposition 3 for the third equality, and the relations G(t ↾D)⊥⊥ = G(t)⊥⊥ = G(t∗∗) for the fourth equality, we derive (1) G(t↾D )p∗E) = N (pG(t↾D )⊥⊥p∗E) = N (pG(t∗∗)p∗E). D⊥ = R(pEpG(t↾D ))⊥ = N (p∗ Let x ∈ N (pG(t∗∗)p∗E). Then x ∈ D(pG(t∗∗)p∗E) and (x, 0) ∈ G(t∗∗) ⊕ vG(t∗). Hence there exist elements y ∈ D(t∗∗) and z ∈ D(t∗) such that x = y+t∗z and 0 = t∗∗y−z. Since (0, 0) = pG(t∗∗)p∗Ex = pG(t∗∗)(x, 0) = (y, t∗∗y), we get y = 0, so z = 0 and x = 0. Thus, N (pG(t∗∗)p∗E) = {0}. Hence D is essential by (1). The other (cid:3) statements are easily verified. Some algebraic properties on orthogonally closable operators are collected in the next proposition. The proofs are straightforward and we omit the details. Proposition 5. Let t ∈ C′o(E, F ). (1) If s ∈ L(E, F ), then t + s ∈ C′o(E, F ). (2) If s ∈ L(F, G) is injective with s−1 ∈ L(G, F ), then st ∈ C′o(E, G). (3) If s ∈ L(G, E) is injective with s−1 ∈ L(E, G), then ts ∈ C′o(G, F ). All these statements remain valid if C′o is replaced by Co. Remark 1. In this Remark we want to emphasize that the theory developed so far is valid in a much more general setting. For this we introduce some definitions. Let A be a complex ∗-algebra. A ∗-bimodule for A is a bimodule X for A equipped with an involution (that is, an antilinear mapping x → x∗ of X such that (x∗)∗ = x for x ∈ X) satisfying (ax)∗ = x∗a∗ and (xa)∗ = a∗x∗ for a ∈ A, x ∈ X. Definition 9. A quadratic ∗-space over A is a triple (E, X,h., .i) of a right A- module E, a ∗-bimodule X for A and a map h., .i : E × E → X such that for α, β ∈ C, x, y, z, x1, . . . , xn ∈ E, a ∈ A: hαx + βy, zi = αhx, zi + β hy, zi , hx, yai = hx, yi a, hx, yi = hy, xi∗ , j=1 hxj , xji = 0 implies x1 = ··· = xn = 0. Xn All preceding notions and results on essentially defined operators and their ad- joints, on orthogonally closed operators and on graph regular operators remain valid for quadratic ∗-spaces rather than Hilbert A-modules. Indeed, an inspection of the definitions and proofs shows that only the axioms from Definition 9 are needed. The same is true for Theorem 6, Theorem 7 and Proposition 7. However, from now on C∗-algebra properties are essentially used. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 9 3. Operators on the commutative C∗-algebra C0(X) Throughout this section we suppose that X is a locally compact Hausdorff space. We consider the C∗-algebra C0(X) of continuous functions on X vanishing at infin- ity as a Hilbert C∗-module E (see Example 1) and study operators t : E → E. The main aim of this section is to describe orthogonally closed operators on E and to show that Co(E) consists of multiplication operators. For this reason we investigate multiplication operators in detail. To state the results we introduce some notation that is inspired by [KL12, Section 6]. For a function m : X → C we set reg(m) :={x ∈ Xm is continuous in a neighborhood of x}, regb(m) :=nx ∈ ∂reg(m)∃U ⊆ reg(m) open, m : U → C continuous, reg∞(m) :=nx ∈ ∂reg(m)∃U ⊆ reg(m) open, m : U → C continuous, with x ∈ U, m ≡ m on U ∩ reg(m)} , with x ∈ U, m ≡ m on U ∩ reg(m), m(x) = ∞} , sing-suppr(m) :=∂reg(m) \ (regb(m) ∪ reg∞(m)). Clearly, reg(m) is the largest open set on which m is continuous. Further, reg(m) ∪ regb(m) is the largest open set contained in reg(m) on which m re- stricted to reg(m) has a (indeed unique) continuous C-valued extension. Finally, reg(m)∪ regb(m)∪ reg∞ (m) is the largest open set contained in reg(m) on which m restricted to reg(m) has a (unique) continuous C-valued extension. In partic- (m) is contained in (reg(m))◦. The space X is a ular, reg(m) ∪ regb(m) ∪ reg∞ disjoint union X = reg(m) ∪ regb(m) ∪ reg∞ (m) ∪ sing-suppr(m) ∂reg(m) {z ∪ (X \ reg(m)) (X\reg(m))◦ . } {z } By F0(X) we denote the set of functions on X vanishing at infinity. We note E = C0(X) = {f ∈ F0(X)reg(f ) = X}. Let us briefly summarize the results that will be obtained in this section. For an arbitrary function m on X a multiplication operator tm is defined and we introduce an equivalence relation to characterize those functions giving the same operator (Lemma 7). Then we show that tm is essentially defined if and only if reg(m) is dense in X and that in this case tm is already orthogonally closed and its adjoint is tm (Theorem 3). Further, we prove that t ∈ Co(E) if and only if t = tm for some function m : X → C (Theorem 4). In Example 4 we give a normal operator t for which the domains D(t) and D(t∗) are different. For m, m : X → C we write m ≃ m ⇔ reg(m) = reg( m) and m ≡ m on reg(m) ∩ reg( m). In Lemma 5(2) it will be shown that ≃ is an equivalence relation. To simplify the notation we associate to each function m : X → C a function m on X defined by m(x) =(m(x) m(x) , x ∈ X \ regb(m) , x ∈ regb(m) , where m is one of the functions appearing in the definition of regb(m). In fact, any two of those functions have the same values at x, since x is in the boundary of reg(m) and the two continuous functions coincide on this set. Hence m is well- defined. Lemma 5(3) shows that m ≃ m. 10 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Lemma 5. (1) Let m1, m2 : X → C. If m1 ≃ m2, then reg∞(m1) = reg∞(m2) and sing-suppr(m1) = sing-suppr(m2). (2) ≃ is an equivalence relation on the set of functions from X to C. (3) If m : X → C, then m ≃ m, reg( m) = reg(m) ∪ regb(m), regb( m) = ∅. Proof. Note that the intersection of two open and dense sets is again open and dense. (m2). This proves (1). (1): From R := reg(m1) = reg(m2) we conclude that reg(m1) ∩ reg(m2) is dense in R. Since m1 and m2 are equal and continuous on this set, any x in R is in reg(m1) ∪ regb(m1) if and only if it is in reg(m2) ∪ regb(m2). By the same argument any x in R is in reg(m1) ∪ regb(m1) ∪ reg∞ (m1) if and only if it is in reg(m2) ∪ regb(m2) ∪ reg∞ (2): Obviously ≃ is reflexive and symmetric, so it remains to show transitivity. Let m1 ≃ m2 and m2 ≃ m3. Clearly, reg(m1) = reg(m2) = reg(m3). Arguing as above, reg(m1) ∩ reg(m2) ∩ reg(m3) is open and dense in the latter set. Again by continuity of m1 and m3 on reg(m1) ∩ reg(m3), these functions coincides on this set, since they do on reg(m1) ∩ reg(m2) ∩ reg(m3). That is, m1 ≃ m3. (3): Clearly, m is continuous on reg(m) ∪ regb(m), so the latter is contained in reg( m). Since m and m coincide on the open set X \ reg(m), we even have reg( m) ⊆ reg(m). In particular, reg(m) = reg( m) and since m and m are equal on reg(m)∩reg( m) = reg(m), it follows that m ≃ m. By (1) the proof is complete, since reg(m) ∪ regb(m) = reg( m) ∪ regb( m) implies that regb( m) = ∅. (cid:3) (m) ∪ sing-suppr(m) does not change any of In particular, changing m on reg∞ Now we define the multiplication operators tm. the sets reg(m), regb(m), reg∞ (m), sing-suppr(m), and X \ reg(m). Definition 10. For a function m : X → C let f ∈ D(tm). It is straightforward to check that tm is indeed an operator on E. D(tm) := {f ∈ Edmf ∈ E}, tmf :=dmf , Lemma 6. Let m, f : X → C and x ∈ X. If f is continuous at x and f (x) 6= 0, then x ∈ reg(m) if and only if x ∈ reg(mf ). Proof. Since f (x) 6= 0, there is an open set Uf containing x such that f (x′) 6= 0 for all x′ ∈ Uf . By definition, x /∈ reg(m) if for any neighbourhood U of x, there is an x′ ∈ U such that m is discontinuous at x′. This holds if and only if for any neighbourhood U of x, there exists x′ ∈ U such that mf is discontinuous at x′. This means that x /∈ reg(mf ). (cid:3) We now show that the operator tm depends only on the equivalence class of m. Lemma 7. Let m : X → C be given. Then D(tm) = {f ∈ Ef ≡ 0 on X \ reg( m), ∂reg( m) ⊆ reg(dmf ),bmf ∈ F0(X)}, f ∈ D(tm), x ∈ X \ reg∞ (tmf )(x) = m(x)f (x), (m). For x ∈ X there is an f ∈ D(tm) such that f (x) 6= 0 if and only if x ∈ reg( m). If m, m : X → C, then tm = t m if and only if m ≃ m. In particular, tm = t m. Proof. By Definition 10 D(tm) consists of those f ∈ E for which reg(dmf ) = X and dmf ∈ F0(X). Since reg( m) ⊆ reg(dmf ) for f ∈ E, D(tm) is the set of all f ∈ E such that X \ reg( m) ⊆ reg(dmf ) anddmf ∈ F0(X). Let f ∈ E with bmf ∈ F0(X). Suppose that ∂reg( m) ⊆ reg(dmf ) and f ≡ 0 X\reg( m) = X\reg( m)∪∂reg( m) is contained in reg(dmf ). To showdmf ∈ F0(X), on X \ reg( m). Then mf ≡ 0 on X \ reg( m) and so on X \ reg( m). Hence UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 11 For f ∈ D(tm) it is obvious that tmf ≡ mf on reg( m). Since f ≡ 0 on Now we suppose that f ∈ D(tm). In particular, reg(mf ) is dense in X, hence continuity of f there exists an open set U ⊆ X \ reg( m) such that f (y) 6= 0 for y ∈ U . From Lemma 6 it follows that U is even contained in X \ reg( mf ). But this contradicts the density of reg(mf ) ⊆ reg( mf ) in X, hence f (x) = 0. In particular let ǫ > 0. Since bmf ∈ F0(X), there exists an compact set K ⊆ X such that bmf ≤ ǫ on X\K. By continuity ofdmf on X the same is true for this function, since bmf and dmf coincide on the dense set X \ ∂reg(m). That isdmf ∈ F0(X), finally f ∈ D(tm). ∂reg( m) ⊆ X = reg(dmf ). Assume that x ∈ X \ reg( m) and f (x) 6= 0. By the dmf coincides with bmf on reg( m) and bmf ≡ 0 on X \ reg( m). So dmf ∈ F0(X) implies bmf ∈ F0(X) and the description of D(tm) is proven. X \ reg( m), it remains to show dmf ≡ 0 on (X \ reg( m)) ∪ sing-suppr(m). For this it suffices to prove the following: If x ∈ X with f (x) = 0 and dmf (x) 6= 0, (m). Indeed, dmf does not vanish on a neighbourhood U of x. Let m : U → C denote the functiondmf /f on U , where α/0 := ∞ for α ∈ C. Then m is continuous, sincedmf does not vanish on U , and m coincides with m on U ∩ reg(m). Further, since f (x) = 0 anddmf (x) 6= 0, we have m(x) = ∞. Hence x ∈ reg∞(m). Let x ∈ X. By the preceding, if x /∈ reg( m), then f (x) = 0 for all f ∈ D(tm). If x ∈ reg( m), there exists a function f : X → C with compact support contained in reg( m) and f (x) 6= 0, since X is locally compact. Then f ∈ D(tm), since mf is continuous everywhere and has compact support. Next we prove that tm = t m if and only if m ≃ m. Assume first that tm = t m. then x ∈ reg∞ Then, by the preceding, x ∈ reg( m) ⇔ ∃f ∈ D(tm) : f (x) 6= 0 ⇔ ∃f ∈ D(t m) : f (x) 6= 0 ⇔ x ∈ reg( m). Thus reg( m) = reg( m) which implies that reg(m) = reg( m) = reg( m) = reg( m). For x ∈ reg(m) ∩ reg( m) we choose f ∈ D(tm) such that f (x) 6= 0. Then reg(m) ∩ reg( m) and the latter set is dense in reg( m), since m ≃ m. Further, m(x)f (x) =dmf (x) =dmf (x) = m(x)f (x), so that m(x) = m(x). Hence m ≃ m. Conversely, assume that m ≃ m. Let f ∈ D(tm). Then dmf ≡ mf ≡ mf on f ≡ 0 on X \ reg( m) by f ∈ D(tm) and hence dmf ≡ 0 ≡ mf on X \ reg( m). Thus,dmf ≡ mf on a dense set. Therefore, sincedmf is continuous on X, we have dmf = dmf . Thus, f ∈ D(t m) and t mf = tmf . This proves that tm ⊆ t m. By Note that the sets reg(m), regb(m), reg∞ (m) and sing-suppr(m) remain un- changed if m is replaced by its complex conjugate function m. Moreover, m = m. Theorem 3. Let m : X → C. The operator tm is essentially defined if and only if reg(m) is dense in X. In this case we have t∗m = tm and tm ∈ Co(E). Proof. By Lemma 7 we can assume without loss of generality that m = m. Assume that reg(m) is dense in X. Let g⊥D(tm). For x ∈ reg(m) there exists f ∈ D(tm) such that f (x) 6= 0 by Theorem 7. From g(x)f (x) = hg, fi (x) = 0 we conclude that g(x) = 0. Hence, by density of reg(m), g = 0. That is, D(tm)⊥ = {0}. Conversely, suppose that reg(m) is not dense. Since X is a locally compact Hausdorff space, there is a function g ∈ C0(X), g 6= 0, with support contained in X \ reg(m). Then hf, gi = 0 for f ∈ D(tm) by Theorem 7 and so D(tm)⊥ 6= {0}. Now assume that reg(m) is dense. We show that t∗m = tm. Let f ∈ D(tm) and g ∈ D(tm). For x ∈ reg(m) = reg(m) we get symmetry, tm = t m. (cid:3) htmf, gi (x) = m(x)f (x)g(x) = hf, tmgi (x). 12 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Since reg(m) is dense, we conclude that htmf, gi = hf, tmgi. Thus, tm ⊆ t∗m. Let f ∈ D(t∗m). There exists h ∈ E such that cmg(x)f (x) = htmg, fi (x) = hg, hi (x) = g(x)h(x), reg(m). Since reg(m) is dense, we getdmf = h ∈ E. Thus f ∈ D(tm). For x ∈ reg(m) we choose g ∈ D(tm) such that g(x) 6= 0. Hence h ≡ mf on Finally, from reg(m) = reg(m) we easily derive t∗∗m = (tm)∗ = tm ∈ Co(E). (cid:3) The domain D(tm) can be trivial if m is continuous on an dense set, since reg(m) g ∈ D(tm), x ∈ X. is empty if there is no open set of continuous points of m. Lemma 8. Let m, n : X → C be two functions. Then: (1) tm + tn ⊆ tm+n and tmtn ⊆ tmn. (2) If reg(m) and reg(n) are dense in X, then tm + tn, tmtn ∈ C′o(E) and (tm + tn)∗∗ = tm+n, (tmtn)∗∗ = tmn. Proof. We verify the statements about the product; the proof for the sum is similar. (1): Let f ∈ D(tmtn). Then reg(nf ) and reg(mcnf ) are dense and open in X, so is their intersection which is contained in reg(mnf ). On this set, mnf ≡ mcnf . Hence \(mn)f = [mcnf ]∧ ∈ E, that is, f ∈ D(tmn) and tmtnf = tmnf . (2): First we prove that D(tmtn)⊥ = {0}. Let g⊥D(tmtn) and x ∈ reg(n) ∩ reg(m). Since the latter set is open and X is a locally compact Hausdorff space, there exists f ∈ E such that its support is contained in this set and f (x) 6= 0. Clearly, nf ∈ E and mnf ∈ E, hence f ∈ D(tmtn). But f (x)g(x) = hf, gi (x) = 0, so g(x) = 0. Since reg(n) ∩ reg(m) is dense, g = 0. Thus, tmtn is essentially defined. By (1), (tmtn)∗ ⊇ t∗mn = tmn, and tmn is essentially defined by Theorem 3, since reg(m) ∩ reg(n) ⊆ reg(mn) is dense. Now we show that D((tmtn)∗) ⊆ D(tmn) which in turn implies the last assertion (tmtn)∗∗ = t∗mn = tmn. Let f ∈ D((tmtn)∗). Then there exists g ∈ E such that hf, tmtnhi = hg, hi for all h ∈ D(tmtn). Arguing as above, for x ∈ reg(n) ∩ reg(m) there exists h ∈ D(tmtn) with h(x) 6= 0. Thereore, f mn ≡ g on the dense set reg(m) ∩ reg(n). Hence [mnf = g ∈ E, that is, f ∈ D(tmn). (cid:3) Remark 2. Set m(x) := ei/x on X = [0, 1]. Then t∗mtm ( tm2. Theorem 4. Let m : X → C. Suppose that reg(m) is dense in X. Then: (1) tm is normal and t∗mtm is essentially self-adjoint. (2) D(t∗mtm) is an essential core for tm. (3) R(1 + t∗mtm) = {g ∈ E∀x ∈ sing-suppr(m) : g(x) = 0}. In particular, R(1 + t∗mtm) is essential. Proof. By Lemma 7 we can assume that m = m. Then tm ∈ Co(E) and t∗m = tm by Theorem 3. (1): Using Lemma 8(2) we get (t∗mtm)∗∗ = (tmtm)∗∗ = tm2. Since the latter is Now we prove that tmtm = tmtm . Since both operators are restrictions of tm2, it suffices to show that their domains coincide. By symmetry it is enough to prove self-adjoint, t∗mtm is essentially self-adjoint. that D(tmtm) ⊆ D(tmtm). Let f ∈ D(tmtm), that is, dmf ∈ E and [mdmf ]∧ ∈ E. We have to show that dmf ∈ E and [mdmf ]∧ ∈ E. Clearly, [mdmf ]∧ = [m2f ]∧ = [mdmf ]∧ ∈ E, since these functions coincide on reg(m). Further, dmf2 = htmf, tmfi = htmtmf, fi = [m2f ]∧f. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 13 Hence [m2f ]∧(x) = 0 implies thatdmf (x) = 0 for x ∈ X. Using this fact it follows that the function 2 h(x) :=dmf (x)(cid:16)\m2f (x)(cid:17)2 0 /(cid:12)(cid:12)(cid:12)\m2f (x)(cid:12)(cid:12)(cid:12) , if \m2f (x) 6= 0, , if \m2f (x) = 0 (2): We show that G(tm ↾D(t∗ belongs to E and coincides with mf on reg(m). Thus [mf ]∧ ∈ E. mtm))⊥ ⊆ G(tm)⊥. Assume that h(g, h), (f, tmf )i = 0 for all f ∈ D(tmtm). Let x ∈ reg(m). Since X is a locally compact Hausdorff space, there exists fx ∈ C0(X) with support contained in reg(m) and fx(x) 6= 0. Clearly, fx ∈ D(tmtm). Then g(x)fx(x) = −h(x)tmfx(x) = −h(x)m(x)fx(x), hence g ≡ −hm on reg(m). Now for all f ∈ D(tm) we have h(g, h), (f, tmf )i ≡ gf + hmf ≡ 0 on reg(m), that is, (g, h)⊥G(tm), since reg(m) is dense. (3): Let x ∈ sing-suppr(m) = sing-suppr(m) and f ∈ D(tmtm). By Theorem 7, f (x) = 0, (tmf )(x) = 0 and (tmtmf )(x) = 0. Thus ((1 + t∗mtm)f )(x) = 0 which proves one inclusion. Conversely, let g ∈ E with g(x) = 0 for all x ∈ sing-suppr(m). The functions [1/(1 + m2)]∧, [m/(1 + m2)]∧ and [m2/(1 + m2)]∧ are bounded and continuous on reg(m)∪ reg∞ ∞(m). Hence, setting f := [g/(1 +m2)]∧, we have gm f ∈ E, dmf =(cid:20) 1 + m2(cid:21)∧ ∈ E, 1 + m2(cid:21)∧ [mdmf ]∧ =(cid:20) gm2 ∈ E, by using that g ≡ 0 on sing-suppr(m). That is, f ∈ D(t∗mtm) and (1 + t∗mtm)f ≡ g on reg(m), so that g ∈ R(1 + t∗mtm), since reg(m) is dense in X. (cid:3) The next theorem is one of our main results in this Section. It says that all es- sentially defined orthogonally closable operators on E are multiplication operators. Theorem 5. If t ∈ C′o(E), then there is a function m : X → C such that t = tm. Proof. Let t ∈ C′o(E). We abbreviate D := D(t) and D∗ := D(t∗). We set O := ∪f∈DOf , O∗ := ∪f∈D∗Of with Of := {x ∈ Xf (x) 6= 0}. Further, since D and D∗ are essential, O and O∗ are dense in X. Hence O′ := O∩O∗ is also dense. For x ∈ X, we have g(x)(tf )(x) = hg, tfi (x) = ht∗g, fi (x) = (t∗g)(x)f (x) f ∈ D, g ∈ D∗. If x ∈ O′, there are f ∈ D and g ∈ D∗ such that f (x) 6= 0 and g(x) 6= 0. Then for m(x) := (tf )(x)/f (x) = (t∗g)(x)/g(x). In particular, this shows that m(x) is independent of the chosen functions f and g and that m is continuous on O′. Now let f ∈ D and x ∈ O′. Then there is a g ∈ D∗ such that g(x) 6= 0, so (tf )(x) = m(x)f (x). Similarly, (t∗g)(x) = m(x)g(x) for g ∈ D∗ and x ∈ O′. We now extend m arbitrarily to a function defined on the whole set X. It follows that dmf ∈ E for f ∈ D and cmg ∈ E for g ∈ D∗. Then we have f ∈ D(tm) and tmf = tf for f ∈ D. Likewise, g ∈ D(tm) and tmg = t∗g for g ∈ D∗. Thus, t ⊆ tm and t∗ ⊆ tm. Therefore, tm = (tm)∗ ⊆ t∗∗ = t ⊆ tm, that is, tm = t. (cid:3) The next example gives a normal operator tm such that D(tm) 6= D(tm). Example 4. Set X := [0, 1] and define m, f : X → C by m(x) :=(ei/x/x , x 6= 0 , x = 0 0 , f (x) :=(e−i/xx , x 6= 0 , x = 0 0 . 14 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Then reg(m) = (0, 1] and f ∈ C0(X). Further, dmf = 1, so f ∈ D(tm). On the other side, (mf )(x) = e−2i/x for x ∈ (0, 1], so reg(dmf ) = (0, 1] and f /∈ D(tm). This proves that D(tm) 6= D(tm). By Theorem 4, the operator tm is normal. 4. Graph regular operators 4.1. Definition and basics on graph regular operators. Graph regular oper- ators are the most important new class of operators introduced in this paper. Definition 11. An operator t : E → F is called • graph regular if t is essentially defined and orthogonally closed and its graph G(t) is orthogonally complemented in E ⊕ F , • regular if t is closed, D(t) is dense in E, D(t∗) is dense in F , and R(1 + t∗t) is dense in E. The preceding is the definition of a regular operator given in [L95, p. 96]. Each regular operator is graph regular by [L95, Theorem 9.3]. By an equivalent definition, an operator t : E → F is graph regular if t is closed, D(t) and D(t∗) are essential in E and F , respectively, and R(1 + t∗t) and R(1 + tt∗) are dense in E and F , respectively. (The equivalence to Definition 11 is easily verified by using some arguments from the proof of Theorem 6(1) below.) We denote by Rgr(E, F ) the set of all essentially defined graph regular operators and by R(E, F ) the set of regular operators from E into F . Let us abbreviate Rgr(E) := Rgr(E, E) and R(E) := R(E, E). A number of basic properties of graph regular operators are collected in the following theorem. Theorem 6. (1) For t ∈ C′o(E, F ) the following statements are equivalent: (a) t ∈ Rgr(E, F ). (b) G(t) ⊕ vG(t∗) = E ⊕ F . (c) R(1 + t∗t) = E and R(1 + tt∗) = F . (2) If t ∈ Co(E, F ) and t∗ ∈ Rgr(F, E), then t ∈ Rgr(E, F ). (3) If t ∈ Rgr(E, F ), then (a) D(t∗t) is an essential core for t. (b) t∗ ∈ Rgr(F, E). Proof. (1a) ⇒ (1b) follows from the relation vG(t∗) = G(t)⊥ by Proposition 1. Hence t is orthogonally closed and graph regular. (1b) ⇒ (1a): Since G(t) ⊆ G(t)⊥⊥, (1b) clearly implies that G(t) = G(t)⊥⊥. (1b) ⇒ (1c): If x ∈ E, then (x, 0) ∈ G(t)⊕vG(t∗), so there are y ∈ D(t), z ∈ D(t∗) such that x = y + t∗z, 0 = ty − z. Then y ∈ D(t∗t) and x = (1 + t∗t)y ∈ R(1 + t∗t). In the same way one shows that F = R(1 + tt∗). (1c) ⇒ (1b): Since R(1+t∗t) = E, we have E⊕0 = R(1+t∗t)⊕0 ⊆ G(t)⊕vG(t∗). For x ∈ D(t∗t), we set y := tx ∈ D(t∗). Then ((1 + t∗t)x, 0) = (x + t∗y, tx − y) ∈ G(t) ⊕ vG(t∗). Similarly, 0 ⊕ F = 0 ⊕ R(1 + tt∗) ⊆ G(t) ⊕ vG(t∗). Thus E ⊕ F ⊆ G(t) ⊕ vG(t∗) which yields G(t) ⊕ vG(t∗) = E ⊕ F . (2) is obtained from t = t∗∗ and (1). (3a): We have hx, yi + htx, tyi = hx, (1 + t∗t)yi for x ∈ D(t) and y ∈ D(t∗t), so that G(t ↾D(t∗t))⊥ ∩ G(t) = {0}. With Lemma 2 it follows G(t)⊥ = G(t ↾D(t∗t))⊥ ∩ (G(t) ⊕ vG(t∗)) = G(t ↾D(t∗t))⊥ ∩ E = G(t ↾D(t∗t))⊥, so D(t∗t) is an essential core for t. (3b): This follows from (2), since t and t∗ are orthogonally closed. (cid:3) UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 15 A special situation is treated in the following example. Example 5. Let H be a separable Hilbert space, A a C∗-algebra of compact oper- ators acting on H, and E a Hilbert A-module. Then we have Co(E) = Rgr(E) = R(E). Indeed, since R(E) ⊆ Rgr(E) ⊆ Co(E) by definition, it suffices to prove that Co(E) ⊆ R(E). Let t ∈ Co(E). Since all closed submodules of any Hilbert C∗- module over a C∗-algebra of compact operators are orthogonally complemented [Mg97], all essential submodules are dense. Hence t and t∗ are densely defined on E. Further, since t is closed, t is semiregular in the sense of [Pal99]. As shown in [Pal99, Proposition 5.1], for any Hilbert A-module E of a C∗-algebra of compact operators semiregular operators are always regular. Thus, t ∈ R(E). In the very special case E = A = K(H) we have R(E) = C(H), since then the regular operators on E are the affiliated operators with A and C(H) is the set of affiliated operators with K(H) as noted in [W91]. Proposition 6. Let t ∈ Rgr(E, F ) and q ∈ L(E, F ). Suppose that r ∈ L(G, E) and s ∈ L(F, G) are invertible with r−1 ∈ L(E, G) and s−1 ∈ L(G, F ). Then the operators t + q, tr and st are essentially defined and graph regular. Proof. Let pE and pF denote the projections from E ⊕ F onto E and F , respec- tively. Clearly, t + q, tr, and st are essentially defined and orthogonally closed by Proposition 2. In particular, their graphs are closed. Since t is graph regular, G(t) is orthogonally complemented, so there is a projection p ∈ L(E ⊕ F ) with R(p) = G(t). We now obtain G(t + q) = {(x, tx + qx)x ∈ D(t)} = {(pEpv, pF pv + qpEpv)v ∈ E ⊕ F} = R((pE, pF + qpE)p), = R((r−1pE, pF )p), = R((pE, spF )p). G(tr) = {(r−1x, tx)x ∈ D(t)} = {(r−1pEpv, pF pv)v ∈ E ⊕ F} G(st) = {(x, stx)x ∈ D(t)} = {(pEv, spF pv)v ∈ E ⊕ F} Thus the closed subspaces G(t + q), G(tr), and G(st) are ranges of adjointable operators, hence they are orthogonally complemented by [L95, Theorem 3.2]. (cid:3) The next lemma describes a cases where graph regularity implies regularity. Lemma 9. If t ∈ Rgr(E, F ), R(t) ⊆ D(t∗) and R(t∗) ⊆ D(t) , then t ∈ R(E, F ). Proof. We have to prove that D(t) and D(t∗) are dense in E and F , respectively. For D(t) this follows from the relations E = R(1 + t∗t) ⊆ D(t) + R(t∗) ⊆ D(t). Since t∗ ∈ Rgr(F, E) by Theorem 6, we can replace t by t∗ in the preceding and obtain the density of D(t∗). (cid:3) 4.2. The (a, a∗, b)-transform. In this section we establish a one-to-one correspon- dence between graph regular operators and certain triples of adjointable operators. As noted in Remark 1 this works in a purely algebraic setting and it neither requires the C∗-condition nor even a norm. Definition 12. For Hilbert A-modules E and F , let AB(E, F ) denote the set of all triples (a, a∗, b) of operators a ∈ L(E), a∗ ∈ L(F ), b ∈ L(E, F ) such that a and a∗ are self-adjoint, N (a) = {0}, N (a∗) = {0}, and bb∗ = a∗ − a2 ∗, b∗b = a − a2, ab∗ = b∗a∗. 16 REN ´E GEBHARDT AND KONRAD SCHM UDGEN In particular 0 ≤ a ≤ I and 0 ≤ a∗ ≤ I in this case; further kbk ≤ 1. We call the map t → (at, at∗ , bt) described in Theorem 7 the (a, a∗, b)-transform. Theorem 7. If t ∈ Rgr(E, F ), then (at, at∗ , bt) ∈ AB(E, F ), where at := (1 + t∗t)−1, at∗ := (1 + tt∗)−1, bt := t(1 + t∗t)−1. Further, N (bt) = N (t), bt∗ = b∗t , and the projection onto the graph of t is given by p =(cid:18)at bt b∗t 1 − at∗(cid:19) ∈ L(E ⊕ F, E ⊕ F ). If (a, a∗, b) ∈ AB(E, F ), then ta,a∗,b ∈ Rgr(E, F ), where ∗ )∗, ta,a∗,b := (ba−1)∗∗ = (b∗a−1 and we have t∗a,a∗,b = ta∗,a,b∗ . The map t 7→ (at, at∗ , bt) is a bijection from Rgr(E, F ) onto AB(E, F ) with inverse (a, a∗, b) 7→ ta,a∗,b. Proof. First we suppose that t ∈ Rgr(E, F ). Then R(1 + t∗t) = E, so at is defined on the whole module E. It is straightforward to verify that 1 + t∗t is positive and injective for each essentially defined operator t. Therefore at is positive and has a trivial kernel. Analogous statements hold for at∗ . Further, bt is defined on E, since R(at) ⊆ D(t). Similarly, bt∗ is defined on F . For x := (1 + t∗t)x′ ∈ E and y := (1 + tt∗)y′ ∈ F , where x′ ∈ E, y′ ∈ F , we compute hbtx, yi = htx′, (1 + tt∗)y′i = htx′, y′i + htx′, tt∗y′i = htx′, y′i + ht∗tx′, t∗y′i = h(1 + t∗t)x′, t∗y′i = hx, bt∗ yi . Hence bt = (bt∗ )∗ ∈ L(E, F ). From bt∗ = (bt)∗ = (tat)∗ we obtain bt∗bt ⊇ att∗tat = at(1 − at). t . Further, bt∗ at∗ = t∗a2 Since at(1− at) is defined on the whole E, the latter yields bt∗ bt = at− a2 (1 + t∗t)t∗ = t∗(1 + tt∗) and R(a2 t∗ = 1 ↾D(t∗t) t∗a2 t∗ ) = D(tt∗tt∗) ⊆ D(t∗tt∗) imply that t∗ = at(1 + t∗t)t∗a2 t∗ = att∗(1 + tt∗)a2 The preceding proves that (at, at∗ , bt) ∈ AB(E, F ). Clearly, R(bt∗ ) ⊆ R(t∗), so N (t) ⊆ N (bt). Suppose that btx = 0 for some x ∈ E. Then (at − a2 t )x = b∗t btx = 0, so x = atx ∈ D(t∗t) ⊆ D(t). Further, (1 + t∗t)x = x, so t∗tx = 0 and from htx, txi = ht∗tx, xi = 0 it follows that x ∈ N (t). Thus, N (t) = N (bt). The statement concerning the projection is easily verified. Conversely, we now assume that (a, a∗, b) ∈ AB(E, F ). We define t := ba−1 and ∗ . Since D(t)⊥ = R(a)⊥ = N (a) = {0}, t is essentially defined. Similarly, s := b∗a−1 s is essentially defined. For x ∈ E, y ∈ F we have t∗ = atbt∗ . ht(ax), a∗yi = hbx, a∗yi = ha∗bx, yi = hbax, yi = hax, b∗yi = hax, s(a∗y)i , Our next aim is to prove that R(a) is an essential core for s∗. Since s∗ = a−1 so t ⊆ s∗ and s ⊆ t∗. In particular, t ∈ C′o(E, F ). Proposition 2, it suffices to show that G(ba−1)⊥ ⊆ G(a−1 Then h(r, s), (ax, bx)i = 0 for all x ∈ E, so ar + b∗s = 0. Further, we have ∗ b by ∗ b)⊥. Let (r, s) ∈ G(ba−1)⊥. a∗(br + (1 − a∗)s) = a∗br + (a∗ − a2 ∗)s = bar + bb∗s = b(ar + b∗s) = 0. Since a∗ is injective, this yields s = a∗s − br. Using the assumption b∗a∗ = ab∗ we obtain Let x ∈ D(a−1 ∗ b). Then there exists a (unique) element z ∈ F with bx = a∗z. b∗z = a−1b∗a∗z = a−1b∗bx = a−1(a − a2)x = (1 − a)x. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 17 Now we compute (cid:10)(r, s), (x, a−1 ∗ bx)(cid:11) = hr, xi +(cid:10)s, a−1 ∗ bx(cid:11) = hr, xi + hs, zi = hr, xi + ha∗s − br, zi = hr, xi + hs, a∗zi − hr, b∗zi = hr, xi + hs, bxi − hr, (1 − a)xi = hb∗s, xi + har, xi = 0. ∗ b). This proves that R(a) is an essential core for s∗. Therefore, (r, s)⊥G(a−1 so that t∗∗ = s∗, that is, (ba−1)∗∗ = (b∗a−1 Since t ⊆ s∗ and D(t) = R(a) is an essential core for s∗, we have G(t)⊥⊥ = G(s∗), ∗ )∗. Finally, we derive 1 + t∗t∗∗ ⊇ 1 + t∗t = 1 + a−1b∗ba−1 = 1 + a−1(a − a2)a−1 = a−1, 1 + t∗∗t∗ ⊇ 1 + s∗s = 1 + a−1 ∗ = a−1 ∗ . ∗ = 1 + a−1 ∗ (a∗ − a2 ∗ bb∗a−1 ∗)a−1 Hence at∗∗ = a ∈ L(E) and at∗ = a∗ ∈ L(F ), so that t ∈ Rgr(E, F ). Further, we have bt∗∗ = t∗∗at∗∗ = t∗∗a ⊇ ta = b ∈ L(E, F ) and so bt∗∗ = b. (cid:3) t∗ bt = btan From at ∈ L(E) it follows at once that the operator 1 + t∗t is self-adjoint for any t ∈ Rgr(E, F ) by Proposition 1. Lemma 10. If t ∈ Rgr(E, F ), then f (at∗ )bt = btf (at) for all f ∈ C([0, 1]). Proof. The operators at, at∗ , bt are adjointable and at∗ bt = btat by Theorem 7. Hence an t for all n ∈ N, so f (at∗)bt = btf (at) for all polynomials f . Since (cid:3) the polynomials are uniformly dense in C([0, 1]), the assertion follows. Lemma 11. Let t ∈ Co(E, F ) and suppose that t and t∗ are bounded. Then we have t ∈ Rgr(E, F ) if and only if t ∈ L(E, F ). Proof. The if direction is trivial. To prove the only if part assume that t is graph regular. Since t is orthogonally closed, t is closed. Because t is closed and bounded, the domain D(t) is closed in E. By Theorem 6(3b), t∗ is also graph regular. There- fore, replacing t by t∗, it follows that D(t∗) is also closed. Hence D(t∗t) is closed. Since t is graph regular, we have at ∈ L(E). Therefore, by [L95, Theorem 3.2], R(at) = D(t∗t) is orthogonally complemented. But R(at)⊥ = N (at) = {0}, so R(at) = E. In particular, E = D(t∗t) ⊆ D(t). Hence D(t) = E. By a similar reasoning we obtain D(t∗) = F . Therefore, t ∈ L(E, F ). (cid:3) Corollary 1. Let t ∈ Rgr(E). Then t is normal if and only if at = at∗ . In this case bt is normal and the operators at and bt commute. Proof. Since t ∈ Rgr(E), we have (at, at∗ , bt) ∈ AB(E) by Theorem 7, so b∗t bt = at − a2 t , btb∗t = at∗ − a2 t∗ , btat = at∗ bt. The first statement is clear and in this case is b∗t bt = btb∗t and btat = atbt. (cid:3) In the next proposition E and F are Hilbert C∗-modules of (possibly different !) C∗-algebras. Proposition 7. Suppose that t ∈ Rgr(E) and φ ∈ Hom(L(E),L(F )). Then there exists an orthogonally closed operator φ(t) : F → F such that φ(at)F is an essential core for φ(t) and φ(t)(φ(at)x) = φ(bt)x, x ∈ F. Moreover, hφ(t)x, yi = hx, φ(t∗)yi for x ∈ D(φ(t)), y ∈ D(φ(t∗)). If N (φ(at)) and N (φ(at∗ )) are trivial, then φ(t) ∈ Rgr(F ), φ(t)∗ = φ(t∗), and aφ(t) = φ(at), aφ(t)∗ = φ(at∗ ), bφ(t) = φ(bt). 18 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Proof. Clearly, 0 ≤ φ(at) ≤ I, 0 ≤ φ(at∗ ) ≤ I, and φ(bt)∗φ(bt) = φ(at) − φ(at)2, φ(bt)φ(bt)∗ = φ(at∗ ) − φ(at∗ )2, φ(at)φ(bt)∗ = φ(bt)∗φ(at∗ ), since φ is a ∗-homomorphism. If x ∈ N (φ(at)), then hφ(bt)x, φ(bt)xi = hx, φ(b∗t bt)xi =(cid:10)x, φ(at)x − φ(at)2x(cid:11) = 0, so x ∈ N (φ(bt)). Therefore, the map φ(t)0 : φ(at)x 7→ φ(bt)x (x ∈ F ) is well- defined. Similarly, the kernel of φ(at∗ ) is contained in the kernel of φ(b∗t ). Further, it is easy to see that G(φ(t∗)0) ⊆ vG(φ(t)0)⊥ = {(x, y) ∈ F ⊕ Fφ(b∗t )x = φ(at)y, y ∈ N (φ(at))⊥} and the latter is the graph of an operator. Hence φ(t)0 and φ(t∗)0 are orthogonally closable. If we denote the corresponding orthogonal closures by φ(t) and φ(t∗), the first half of the proposition is shown. If the kernels of φ(at) and φ(at∗ ) are trivial, then (φ(at), φ(at∗ ), φ(bt)) ∈ AB(F ) and all statements follow from Theorem 7. (cid:3) If the kernel of φ(at) is not trivial, it can happen that the domain of the operator φ(t) is only {0}, see Example 7 below. Corollary 2. Let A be a (non-degenerated) concrete C∗-algebra on H. Let φ be the embedding of L(A) = M(A) into B(H) = L(H). such that φ(t) = T . (1) For any T ∈ C(H) with aT , aT ∗ , bT ∈ M(A) there exists a unique t ∈ Rgr(A) (2) If A contains the compact operators, then we have T := φ(t) ∈ C(H) and aT , aT ∗ , bT ∈ M(A) for t ∈ Rgr(A). In particular, Rgr(A) can be identified with those T ∈ C(H) for which aT , aT ∗ , bT ∈ M(A). Proof. (1): Since C(H) = Rgr(H), we have (aT , aT ∗ , bT ) ∈ AB(H) by Theorem 7. By assumption aT , aT ∗ , and bT are elements of M(A). To show that (aT , aT ∗ , bT ) ∈ AB(A) it suffices to prove that aT and aT ∗ are injective as operators on A. Clearly, they are injective as operators on H. Assume that aT a = 0 for some a ∈ A. Then, aT aξ = 0 for ξ ∈ H. Hence aξ = 0 for all ξ ∈ H, so that a = 0. Thus, aT is injective on A. Similarly, aT ∗ is injective on A. Using once more Theorem 7 it follows that there exists an operator t ∈ Rgr(A) such that at = aT , at∗ = aT ∗ , bt = bT . Further, φ(t) = T , since T (φ(at)ξ) = T (atξ) = T aT ξ = bT ξ = btξ = φ(bt)ξ, ξ ∈ H, and R(aT ) = D(T ∗T ) is a core for T . (2): If t ∈ Rgr(A), then (at, at∗ , bt) ∈ AB(A). We show that the kernels of φ(at) and φ(at∗ ) are trivial. Assume that φ(at)ξ = 0 for some nonzero vector ξ ∈ H. Since A contains all compact operators, the rank one projection pξ onto C·ξ is in A. Therefore, atpξ = φ(at)pξ = 0 which contradicts the injectivity of at as operator on A. Hence at, similarly at∗, is injective on H. Therefore, T ∈ Rgr(H) = C(H) (cid:3) by Proposition 7. Example 6. From Corollary 2 it follows immediately that Rgr(K(H)) = C(H) and Rgr(B(H)) = C(H), since M(K(H)) = M(B(H)) = B(H). 4.3. Quotients of adjointable operators. A large class of examples of un- bounded graph regular operators can be obtained as quotients of adjointable oper- ators. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 19 x ∈ G, D(t) = R(a), Theorem 8. Let a ∈ L(G, E) and b ∈ L(G, F ). Suppose that N (a) ⊆ N (b) and N (a∗) = {0}. If the operator t : E → F defined by t(ax) := bx, is closed, then t ∈ Rgr(E, F ) and t∗ = (a∗)−1b∗. Proof. Since N (a) ⊆ N (b) and R(a)⊥ = N (a∗) = {0}, t is well-defined and essen- tially defined. Since the graph of t is the set {(ax, bx)x ∈ G}, it is the range of the adjointable operator q : G → E ⊕ F defined by q(x) := (ax, bx). Since this range is closed by assumption this range, [L95, Theorem 3.2] applies and shows that the range is orthogonally complemented. Hence t is graph regular. The adjoint of t is (cid:3) then easily computed; we omit the details. Corollary 3. Let x ∈ L(F, E) and assume that N (x) = N (x∗) = {0}. Then x−1 ∈ Rgr(F, E) and (x−1)∗ = (x∗)−1. Proof. Since x−1 is closed, the assertion follows from Theorem 8 by letting b the (cid:3) identity on F . Corollary 4. Let a ∈ L(E) and let p, q ∈ C[X] be relatively prime. Assume that R(q(a)) is essential and N (q(a)) ⊆ N (p(a)). Let t : E → E be the operator defined by D(t) := R(q(a)), t(q(a)x) := p(a)x, x ∈ E. Then t is graph regular. Proof. In order to apply Theorem 8 we only have to prove that t is closed. Let (xn)n∈N be a sequence in E such that p(a)xn → xp ∈ E and q(a)xn → xq ∈ E. Since p and q are relative prime, there are polynomials p, q ∈ C[X] such that pp + qq = 1. Then xn = (p(a)p(a) + q(a)q(a))xn → p(a)xp + q(a)xq =: xr ∈ E, so xp = p(a)xr and xq = q(a)xr. That is, t is closed. Theorem 9. Let a ∈ L(G, E), b ∈ L(G, F ), a∗ ∈ L(H, F ), and b∗ ∈ L(H, E) be such that b∗a∗ = a∗b∗. Assume that N (a∗) = N (a∗ ∗) = {0}. Then N (a) ⊆ N (b) and N (a∗) ⊆ N (b∗). The operators t and t′ defined by (cid:3) D(t) := R(a), D(t′) := R(a∗), t(ax) := bx, t′(a∗y) := b∗y, x ∈ G, y ∈ H, are essentially defined, orthogonally closable and they satisfy (t′)∗∗ ⊆ t∗, t∗∗ ⊆ (t′)∗. If in addition t and t′ are closed and ab∗ = b∗a∗ ∗, then t∗ = t′ and t ∈ Rgr(E, F ). Proof. Suppose that a∗x = 0 for some x ∈ H. Then 0 = b∗a∗a = a∗b∗x and hence b∗x = 0, since N (a∗) = {0}. This shows that N (a∗) ⊆ N (b∗). In a similar manner, the assumption N (a∗ ∗) = {0} implies that N (a) ⊆ N (b). Hence the operators t and t′ are well-defined. It is obvious that t and t′ are essentially defined. From the relations (a∗)−1b∗a∗ = b∗ and t∗y = (a∗)−1b∗y, y ∈ G, we get t′ ⊆ t∗. Since t′ ⊆ t∗ and t′ is essentially defined, so is t∗. Since t is also essentially defined, Interchanging the role of t and t′ we conclude that t is orthogonally closable. t ⊆ (t′)∗ and t′ is orthogonally closable. Applying the involution to the relations t ⊆ (t′)∗ and t′ ⊆ t∗ we obtain (t′)∗∗ ⊆ t∗ and t∗∗ ⊆ (t′)∗ which proves the first half of the proposition. ∗. Since t and t′ are closed, Now suppose that t and t′ are closed and ab∗ = b∗a∗ G := G(t) ⊕ vG(t′) = {(ax + b∗y, bx − a∗y)x ∈ E, y ∈ F} 20 REN ´E GEBHARDT AND KONRAD SCHM UDGEN is a closed submodule of E ⊕ F . We define q(x, y) := (ax + b∗y, bx − a∗y) for (a, b) ∈ E ⊕ F . Then q ∈ L(E ⊕ F ) and R(q) = G. By [L95, Theorem 3.2], G is orthogonally complemented. It is easily calculated that q∗(x′, y′) = (a∗x′ + b∗y′, b∗ ∗x′ − a∗ ∗y′) for (x′, y′) ∈ E ⊕ F. We show that N (q∗) = {0}. Suppose that q∗(x′, y′) = 0. Then a∗x′ + b∗y′ = 0 and a∗ ∗y′ − b∗ ∗x′ = 0, so we obtain 0 = aa∗x′ + ab∗y′ = aa∗x′ + b∗a∗ ∗x′, b∗ ∗y′ = aa∗x′ + b∗b∗ ∗x′i = 0. Thus, a∗x′ = 0 and b∗ The latter implies that ha∗x′, a∗x′i + hb∗ ∗x′ = ∗y′ = 0. Therefore, x′ = 0 and y′ = 0 by the assumption N (a∗) = N (a∗ a∗ ∗) = {0}. That is, N (q∗) = {0}. Hence, G⊥ = R(q)⊥ = N (q∗) = {0}. Therefore, since G is orthogonally complemented, we have G = E ⊕ F . This proves that t′ = t∗. (cid:3) 4.4. Absolute value. The next theorem is concerned with the absolute value of graph regular operators. Theorem 10. Suppose that t ∈ Rgr(E) and define ∗x′. D(t) := R(a1/2 t ), t(a1/2 t x) := (1 − at)1/2x, x ∈ E. Then t ∈ Rgr(E) is self-adjoint and positive. Further, we have t2 = t∗t, at = at, and bt = bt. Proof. Clearly, t is essentially defined. We prove that t is closed. Let (xn) be a sequence in E such that a1/2 t xn → x ∈ E and (1 − at)1/2xn → y ∈ E. Then xn = (1− at)xn + atxn → a1/2 t x + (1− at)1/2y =: x′, so that x = a1/2 t x′. This shows that t is closed. Therefore, t is graph regular by Proposition 8. and b = b∗ = (1 − at)1/2 we conclude that t = t∗, since the relations b∗a∗ = a∗b∗ and ab∗ = b∗ ∗a∗ are fulfilled. Further, D(t2) ⊇ R(at) = D(t∗t). It is easily checked that (1 + t2)at = 1 = (1 + t∗t)at, so t2 ⊇ t∗t. Since t∗t is self-adjoint and t2 = t∗t is symmetric, we obtain t2 = t∗t. We derive Applying Theorem 9 with a = a∗ = a1/2 t t xE =D(1 − at)1/2, a1/2 t x), a1/2 Dt(a1/2 bt = tat = ta1/2 t xE =D(at − a2 t = (at − a2 t )1/2x, xE ≥ 0 for x ∈ E, t )1/2 = (b∗t bt)1/2 = bt. so t is positive. Clearly, we have at = at. Finally, we compute t a1/2 t = (1 − at)1/2a1/2 (cid:3) Remark 3. In contrast to the Hilbert space case the domains D(t) and D(t) do not coincide in general, even more, neither D(t) ⊆ D(t) nor D(t) ⊇ D(t) holds. Indeed, let E = A := C([0, 1]) and set m(x) := x−1ei/x for x ∈ (0, 1]. Then, by Theorem 15 below, the operator tm is graph regular, since reg(m) = (0, 1] and sing-suppr(m) = ∅. It is easily verified that tm = tm. Define f (x) := xe−i/x for x ∈ (0, 1] and f (0) = 0. Then f ∈ D(tm), but f /∈ D(tm). The function g(x) = x is in D(tm), but not in D(tm). 4.5. Bounded transform. The main results of this section give a characterization of graph regular operators in terms of the bounded transform. Let us begin with some important notation. Let Z(E, F ) denote the set of all z ∈ L(E, F ) such that kzk ≤ 1 and N (I − z∗z) = {0} and let Z d(E, F ) be the set of those z ∈ Z(E, F ) for which R(I − z∗z) is dense in E. In [L95, Lemma 10.3] it was shown that z ∈ Z d(E, F ) implies that z∗ ∈ Z d(F, E). The following lemma contains the analogues statement for Z(E, F ). UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 21 Lemma 12. If z ∈ L(E, F ), then N (I−z∗z) = {0} if and only if N (I−zz∗) = {0}. In particular, z ∈ Z(E, F ) if and only if z∗ ∈ Z(F, E). Proof. It suffices to show one direction, since z can be replaced by z∗. Assume that x ∈ N (I − zz∗) \ {0}. Then kz∗xk2 = hx, zz∗xi = hx, xi = kxk2. Hence z∗x 6= 0. But (I − z∗z)z∗x = z∗(I − zz∗)x = 0, so z∗x ∈ N (I − z∗z) 6= {0}. (cid:3) For z ∈ Z(E, F ) we define an operator tz : E → F by D(tz) := (I − z∗z)1/2E, tz(I − z∗z)1/2x := zx, x ∈ E. Since R((I − z∗z)1/2)⊥ = N ((I − z∗z)1/2) = N (I − z∗z) = {0}, tz is essentially defined. Clearly, R(z) = R(tz). Theorem 11. If z ∈ Z(E, F ), then tz ∈ Rgr(E, F ). Further, the mapping z → tz is injective from Z(E, F ) into Rgr(E, F ). We have t∗z = tz∗ and atz = I − z∗z, z = tza1/2 tz . Proof. First we prove that the operator tz is closed. Let (xn) be a sequence of D(z) such that ((I − z∗z)∗)1/2xn → x and zxn → y. Then xn = (I − z∗z)∗xn + z∗zxn → ((I − z∗z)∗)1/2x + z∗y, so x ∈ D(tz) and tzx = y. Thus, tz is closed and tz ∈ Rgr(E, F ) by Proposition 8. Set a = (1 − z∗z)1/2, b = z, a∗ = (1 − zz∗)1/2, b∗ = z∗. Since z∗(1 − zz∗)1/2 = (1− z∗z)1/2z∗, we then have b∗a∗ = a∗b∗ and ab∗ = b∗a∗ ∗. Hence Theorem 9 applies and yields z∗t = zt∗. Finally, by Proposition 2(4), t∗z = ((I − z∗z)1/2)−1z∗, so t∗ztz = ((I − z∗z)1/2)−1z∗z((I − z∗z)1/2)−1 = ((I − z∗z)1/2)−1(I − z∗z)((I − z∗z)1/2)−1 − (((I − z∗z)1/2)−1)2 = I − ((I − z∗z)1/2)−1((I − z∗z)1/2)−1. Therefore, (I + t∗ztz)−1 = I − z∗z. In particular, D(t∗ztz)⊥ = R(I − z∗z)⊥ = {0} and ((I + t∗ztz)−1)∗ = (I + t∗ztz)−1. This also implies that t∗ztz is self-adjoint and tz((I + t∗ztz)−1)1/2 = z. (cid:3) According to [L95, Theorem 10.4], the mapping z 7→ tz is a bijection from the set Z d(E, F ) onto the set R(E, F ) of regular operators. For the extended mapping acting on Z(E, F ) the situation is more subtle. It is still an injective mapping into the set Rgr(E, F ) of graph regular operators, but it is not sujective as shown by Example 7 below. Lemma 13. If z ∈ Z(E, F ), then z ∈ Z(E) and tz = tz. Further, we have N (z) = N (tz) = N (z) = N (tz). Proof. Since 1 − z∗z = 1 − z2 and kzk = kzk ≤ 1, the first statement is clear. Further, atz = (1 + t∗ztz)−1 = 1 − z∗z, so D(tz) = R(a1/2 tz ) = R((1 − z∗z)1/2) = R(1 − z2) = D(tz), tz(1 − z∗z)x = (1 − atz )1/2x = (z∗z)1/2x = zx = tz(1 − z2)x, x ∈ E, that is, tz = tz. Since kernels of orthogonally closed operators are orthogonally closed, we obtain N (z) = R(z∗)⊥ = R(tz∗ )⊥ = R(t∗z)⊥ = N (tz). Because N (z) = N (z), this completes the proof. (cid:3) The following lemma restates [L95, Proposition 3.7]. It will be used several times in the proof of Theorem 12 below. Lemma 14. If a ∈ L(E)+, then R(a) = R(aε) for any ε > 0. 22 REN ´E GEBHARDT AND KONRAD SCHM UDGEN t ) is contained in its domain. Its adjoint is an extension of a1/2 Now suppose that t ∈ Rgr(E, F ). Recall that at ∈ L(E)+, at∗ ∈ L(F )+, and is essentially defined, since t∗, hence it . Hence bt = b∗t∗ ∈ L(E, F ) by Theorem 7. The operator ta1/2 R(a1/2 is also essentially defined. Moreover, by Proposition 2(3), (a1/2 ta1/2 Definition 13. The bounded transform zt of t ∈ Rgr(E, F ) is defined by is orthogonally closed; in particular, ta1/2 t∗)∗ = ta1/2 is closed. t t t t t t zt := ta1/2 t ↾ D(t∗t) . Since t ∈ Rgr(E, F ), D(t∗t) is essential in E and D(tt∗) is essential in F . Theorem 12. Suppose t ∈ Rgr(E, F ). Set E0 := D(t∗t) and F0 := D(tt∗). Then zt ∈ Z d(E0, F0) and z∗t = zt∗, where the adjoint z∗t is taken in L(E0, F0). Further, t∗ t ↾ zt = a1/2 at = (1 − zt∗ zt)∗, D(t∗t), N (zt) = N (t), R(zt) ⊆ R(t), bt = zta1/2 at∗ = (1 − ztzt∗ )∗, t , and treg := tzt ↾E0 is a regular operator from E0 to F0 satisfying treg ⊆ t = (tzt )∗∗, where t∗∗zt is the biadjoint of the operator tzt : E → F . Proof. From Theorem 7 we already know that t = a−1 t∗ bt. Using this we derive (2) Here the second equality follows from Lemma 10 applied with f (x) = √x. t∗ bt ⊆ a−1 t∗ bt = a−1 t∗ bta1/2 t = ta1/2 a1/2 t∗ t = a1/2 t∗ a−1 t∗ a1/2 . t Now we prove that a1/2 t∗ t is bounded with norm not exceeding 1. Let x ∈ D(t). Using that t∗at∗ t = bt∗t = b∗t t ⊆ (t∗bt)∗ = (t∗tat)∗ = (1 − at)∗ = 1 − at, we derive t∗ txk2 = kDa1/2 ka1/2 t∗ tx, a1/2 t∗ txEk = k ht∗at∗ tx, xi k = k h(1 − at)x, txi k ≤ k1 − atkkxk2 ≤ kxk2. t∗ t and ta1/2 t concides on D(t) and so on its subspace is closed, we conclude By (2) the operators a1/2 D(t∗t). Since both operators are bounded on D(t∗t) and ta1/2 that zt ≡ ta1/2 D(t∗t) = a1/2 D(t∗t) . t∗ t ↾ ↾ t t The latter equality implies that R(zt) is contained in R(a1/2 t∗ ). Lemma 14 yields R(a1/2 t∗ ) = R(at∗ ) = D(tt∗) = F0, so that R(zt) ⊆ F0. Hence zt becomes a bounded operator from E0 into F0. Analogously, zt∗ is a bounded operator from F0 into E0. We show that z∗t = zt∗. Let z∗E denote the adjoint of zt considered as an t operator from E into F . Clearly, z∗E t∗ ⊇ zt∗. This implies that z∗t = z∗E t t ⊇ (a1/2 ↾F0= zt∗ . Therefore, zt ∈ L(E0, F0). t∗ t)∗ ⊇ t∗a1/2 Next we verify the formulas for at, at∗ , and bt. First we note that t∗ ↾F0 a1/2 zt∗ zt ↾D(t∗t) = t∗a1/2 (3) Hence at = (1 − zt∗ zt)∗∗ by Example 3, so at = a∗t = (1 − zt∗zt)∗∗∗ = (1 − zt∗zt)∗. The formula for at∗ is proven in a similar manner. Since R(a1/2 ) ⊆ E0, we obtain zta1/2 t∗ t ↾D(t∗t)= t∗at∗ t ↾D(t∗t)= 1 − at ↾D(t∗t) . t t ↾E0 a1/2 t = tat = bt. t = ta1/2 Using relation (3) and again Lemma 14 we get D(t∗t)) = R(at ↾ R(1 − zt∗ zt) = R(at ↾ = R(a2 t ) = R(at) = D(t∗t) = E0. R(at)) ⊇ R(at ↾R(at)) Hence zt ∈ Z d(E0, F0). Therefore, treg = tzt is a regular operator from E0 to F0. Clearly, treg ⊆ t. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 23 Now we prove that t = (tzt )∗∗. Since tzt ⊆ t and D(tzt ) = R((1 − z∗t zt)1/2) = R(at ↾ D(t∗t)) = R(at ↾ R(at)) ⊇ R(a2 t ), it suffices to show that R(a2 for some x ∈ D(t). Then, for all y ∈ E, t ) is an essential core for t. Assume (x, tx)⊥G(t ↾R(a2 t )) 0 =(cid:10)(x, tx), (a2 t y, ta2 t y)(cid:11) =(cid:10)x, a2 t y(cid:11) = hatx, atyi + htx, btatyi = hatx, atyi + hb∗t tx, atyi = hatx + b∗t tx, atyi t y(cid:11) +(cid:10)tx, ta2 Therefore, since R(at) is essential, 0 = (at + b∗t t)x = (at + 1 − at)x = x. That is, t ))⊥ = {0}. Since t is graph regular and hence G(t) ⊕ G(t)⊥ = E, it G(t) ∩ G(t ↾R(a2 follows from Lemma 2 that G(t ↾R(a2 t ))⊥ = G(t ↾R(a2 t ))⊥ ∩ (G(t) ⊕ G(t)⊥) = G(t)⊥. t ) is an essential core for t which completes the proof of the equality Thus R(a2 t = (tzt)∗∗. Clearly, R(zt) ⊆ R(t). Finally, we show that N (zt) = N (t). Let x ∈ N (t) ⊆ D(t∗t). Since zt ⊇ a1/2 t∗ t ↾D(t∗t), we have x ∈ D(zt) and ztx = 0, so N (t) ⊆ N (zt). Conversely, let x ∈ N (zt). Then (1 − at)x = zt∗ ztx = 0 and x = atx ∈ D(t∗t) ⊆ D(t). We obtain (1 + t∗t)x = x and t∗tx = 0. From htx, txi = ht∗tx, xi = 0 we get tx = 0. Hence N (zt) ⊆ N (t). (cid:3) Definition 14. The operator treg := tzt ∈ R(E0, F0) from Theorem 12 is called the regular operator associated with the graph regular operator t ∈ Rgr(E, F ). Remark 4. There are two other possibilities to define the bounded transform for a graph regular operator t and both of them are natural in some sense. Define z′t := ta1/2 t and z′′t := (bt(a1/2 t )−1)∗∗. t )−1 is the restriction of z′t = ta1/2 Note that z′t ∈ Co(E, F ) and zt is the restriction of z′t to D(t∗t). It is easily seen that bt(a1/2 )−1 is essentially defined and orthogonally closable and its orthogonal closure may also be taken as a bounded transform z′′t of t. Then z′t = z′′t ) is an essential core for z′t. We were not able to prove or disprove this. But one can show that if and only if R(a1/2 ). Hence bt(a1/2 to R(a1/2 t t t t zt ⊆ z′′t ⊆ z′t, (z′′t )∗ = z′t∗ ≡ t∗a1/2 t∗ , (z′t)∗ = z′′t∗ ≡ (bt∗ (a1/2 t∗ )−1)∗∗. Note that the operator t can be recovered from both transforms z′t and z′′t as well. We dont know wether or not the equality (ta1/2 Remark 5. The relationship between t ∈ Rgr(E, F ), treg ∈ R(E0, F0), and their bounded tranforms should be studied be further in detail. Despite of this it seems that graph regular operators form an important notion in its own, because they act on the given Hilbert C∗-module E. Though each symmetric operator on a Hilbert space is a restriction of a self-adjoint operator in possibly larger space, symmetric operators are a basic concept. )∗ = t∗a1/2 t∗ holds. t 4.6. Polar decomposition. Definition 15. Let E′ ⊆ E and F ′ ⊆ F be orthogonally closed. An operator v ∈ Co(E, F ) is called a partial isometry with initial space E′ and final space F ′ if v∗v is the projection pE′ and vv∗ is the projection pF ′. In this case v∗ is also a partial isometry with initial space F ′ and final space E′. Using this general notion of essentially defined partial isometries there is the following theorem on a polar decomposition of adjointable operators. 24 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Theorem 13. Let t ∈ L(E, F ). Then there is a partial isometry v ∈ Co(E, F ) with initial space R(t∗) and final space R(t) such that t = vt, t = v∗t and R(v) = R(t), R(v∗) = R(t∗), N (v) = N (t), and N (v∗) = N (t∗) if and only if R(t) and R(t∗) are orthogonally closed. Proof. The only if direction follows easily from the definition of a partial isometry. To prove the if part we assume that R(t) and R(t∗) are orthogonally closed. Define a map w : R(t) → R(t) by w(tx) := tx for x ∈ E. Then w is well-defined and isometric, since htx, txi = htx,txi for x ∈ E. The continuous extension of w to a map from R(t) onto R(t) is also an isometry which is denoted again by w. Now we define v : E → F by v(x + y) := wx, x ∈ R(t), y ∈ N (t). Clearly, D(v)⊥ = (N (t) ⊕ R(t))⊥ = N (t)⊥ ∩ R(t)⊥ = N (t)⊥ ∩ N (t) = {0}, so v is essentially defined. Further, t = vt, N (v) = N (t) and R(v) = R(t). Let v′(x + y) := w−1x, x ∈ R(t), y ∈ N (t∗). (cid:10)v(x + x⊥), y(cid:11) =(cid:10)x + x⊥, z(cid:11) for all x ∈ R(t) and x⊥ ∈ N (t). Choosing x = 0, we As above, v′ is essentially defined, t = v′t, N (v′) = N (t∗) and R(v′) = R(t). It is easily seen that v′ ⊆ v∗ and v ⊆ (v′)∗. Therefore, since v′ is essentially defined, so is v∗. Hence v is orthogonally closable by Theorem 1. We show that v∗ = v′. Let y ∈ D(v∗). Then there is an element z ∈ E such that conclude that z ∈ N (t)⊥ = R(t∗)⊥⊥ = R(t)⊥⊥ = R(t). Thus, R(v∗) ⊆ R(v′). Putting now x⊥ = 0, we get htx′, yi = hvtx′, yi = htx′, zi for all x′ ∈ E. Hence t∗y = tz = tv∗y and N (v∗) ⊆ N (t∗) = N (v). Now v′ ⊆ v∗, R(v∗) ⊆ R(v′) and N (v∗) ⊆ N (v′) clearly imply that v′ = v∗. In a similar manner it is shown that v′∗ = v. Obviously, v∗v is the projection onto the orthogonally closed submodule R(t∗) and vv∗ the projection onto R(t). (cid:3) Theorem 14. Let z ∈ Z(E, F ). There exists a partial isometry v ∈ Co(E, F ) with initial space R(tz) and final space R(tz) such that tz = v∗tz, tz = vtz, R(v) = R(tz), R(v∗) = R(t∗z), N (v) = N (tz), and N (v∗) = N (t∗z) if and only if R(z) and R(z∗) are orthogonally closed. In this case, z = vz. Proof. Since R(z) = R(tz), R(z∗) = R(t∗z), N (z) = N (tz), and N (z∗) = N (t∗z), one direction follows at once from the definition of a partial isometry. Conversely, assume that R(z∗) = R(z) and R(z) are orthogonally closed. By Proposition 13, there is a partial isometry v ∈ Co(E, F ) with z = vz, z = v∗z and R(v) = R(z), R(v∗) = R(z∗), N (v) = N (z) and N (v∗) = N (z∗). Finally, vtz(1 − z2)1/2x = vzx = zx = tz(1 − z2)1/2x for x ∈ E, so that vtz = tz. Similarly, v∗tz = tz. (cid:3) Note that the partial isometry belongs to L(E, F ) if and only if R(z) and R(z∗) are orthogonal complements: D(v) = R(z∗) ⊕ N (z) and D(v∗) = R(z) ⊕ N (z∗). 4.7. Graph regular operators on C0(X). Before we turn to the functional cal- culus of graph regular normals we reconsider the commutative case from Section 3. By Theorem 5, Co(C0(X)) consists of multiplication operators. Theorem 15 charac- terizes the graph regular operators among them as those for which sing-suppr(m) is empty. Lemma 15. For any function m : X → C the following are equivalent: UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 25 (1) tm is injective if and only if {x ∈ reg(m)m(x) 6= 0} is dense in reg(m). (2) If tm is injective and m does not vanish on X, then t−1 m = t1/m. Proof. (1): Set N := {x ∈ reg(m)m(x) = 0}. Assume that N contains a nonempty open set U . Since X is locally compact and Hausdorff, there is a non-zero function f ∈ E with support contained in U . Then mf = 0, f ∈ D(tm) and tmf = 0. Hence tm is not injective. On the other hand, assume that f ∈ D(tm) and tmf = 0. Then m(x)f (x) = (tmf )(x) = 0 for x ∈ reg(m) . Thus f ≡ 0 on reg(m) \ N . If the latter is dense in reg(m), then f ≡ 0 on reg(m) by the continuity of f . Futher, we have f ≡ 0 on X \ reg(m) by Lemma 7, since f ∈ D(tm). That is, f = 0. (2): We have reg(1/m) = reg(m). In particular, t1/m is injective. From Lemma 8(1) it follows that tmt1/m and t1/mtm are restrictions of the identity. Therefore, t1/m ⊆ t−1 m ⊆ t1/m, so equality is (cid:3) proven. Theorem 15. If m : X → C, then tm ∈ Rgr(C0(X)) if and only if reg(m) is dense in X and sing-suppr(m) is empty. In this case we have t∗mtm = tm2 and 1/m. The last inclusion gives t−1 m and tm ⊆ t−1 Proof. The first assertion is clearly a corollary of Theorem 4. atm = t , 1 1+m2 btm = t m 1+m2 . Suppose that tm ∈ Rgr(E). Then atm is self-adjoint, hence t∗mtm is self-adjoint. Further, by Lemma 8, t∗mtm is contained in the self-adjoint operator tm2. Hence t∗mtm = tm2. Finally, using Lemma 8(2) and Lemma 15(2), we compute , atm = (1 + t∗mtm)−1 = (1 + tm2)−1 = (t1+m2)−1 = t btm = b∗∗tm = (tmatm)∗∗ = (tmt )∗∗ = t m 1+m2 1 1 1+m2 . 1+m2 (cid:3) Corollary 5. Suppose that m : X → C is bounded and reg(m) dense in X. In particular, we have Then tm is graph regular if and only if reg( m) = X. tm ∈ L(C0(X)) = Cb(X) in this case. Proof. Since m is bounded, reg∞ from Theorem 15. (m) is empty. Hence the statement follows directly (cid:3) The next example shows that not all operators t ∈ Rgr(C0(R)) are of the form tz for some z ∈ Z(C0(R)). Moreover there is a representation π of L(C0(R)) such that the domain of π(t) consists only of the zero element. Example 7. Consider the operator t := tm on C0(R), where m(x) = 1/x for x 6= 0. Then t is self-adjoint and graph regular by Theorem 15. We show that there is no z ∈ Z(C0(R)) such that t = tz. Assume to the by Theorem 11. Further, t = t1/(√1+m2 ). Choose g ∈ C0(R) x xg(x) √1+x2 is t g /∈ D(t). Let π be the ∗-representation of L(C0(R)) = {tnn ∈ Cb(R)} given by π(tn) = contrary that t = tz for z ∈ Z(C0(R)). Then z = ta1/2 we have at = t1/(1+m2) and therefore a1/2 t g(x) = xg(x)/(√1 + x2). Since the function 1 with g(0) 6= 0. Then a1/2 continuous on R\{0} and has no continuous extension to R, we have a1/2 Hence ta1/2 is not everwhere defined which contradicts the equality z = ta1/2 n(0). Then π(at) = 0, so the domain of π(t) is {0} (compare Proposition 7). Example 8 (Continuing Example 4). Recall that the operator tm from Example 4 is normal, but D(tm) 6= D(t∗m). Since reg(m) = (0, 1] and 0 ∈ reg∞ (m), tm is graph regular by Theorem 15. That is, even for graph regular operators t the statements . t t t 26 REN ´E GEBHARDT AND KONRAD SCHM UDGEN (1) t∗t = tt∗ (that is, t is normal), (2) D(t) = D(t∗) and ktfk = kt∗fk for all f ∈ D(t), are not equivalent! 4.8. Functional calculus of graph regular normal operators. Let A and B be C∗-algebras, where A is non-unital and B is unital. Clearly, each ∗-homomorphism φ : A → B extends uniquely to a ∗-homomorphism of the unitization A∼ := A ⊕ C by φ(a + α) := φ(a) + α1 for a ∈ A, α ∈ C. Let ζ denote the identity map of C. Considered as an operator on C0(C), ζ is a regular operator, but on the unitization C0(C)∼ the operator ζ is no longer regular. On the other hand, since aζ = aζ = (1 + ζ2)−1 ∈ C0(C) the operator ζ is graph regular according to Theorem 7. Further, bζ = ζ(1 + ζ2)−1 ∈ C0(C). Theorem 16. Let E be a Hilbert A-module and let t ∈ Rgr(E) be normal. Then there exists a unique φt ∈ Hom(C0(C)∼,L(E)) with N (φt(aζ)) = {0} and φt(ζ) = t. Proof. Set D := {z ∈ C : z ≤ 1/2}, F := {(z1, z2) ∈ [0, 1] × D : z22 = z1 − z2 1} ⊆ [0, 1] × D. Then ∂F = {(0, 0)} and F = F \∂F . By Corollary 1, aζ is self-adjoint, bζ is normal, and aζ and bζ commute. Their joint spectrum σ(aζ, bζ) is contained in F . Similar statements hold for at and bt. Uniqueness: Let φ ∈ Hom(C0(C)∼,L(E)) be such that N (φt(aζ)) = {0} and φ(ζ) = t. Then, by Proposition 7, φ(aζ ) = aφ(ζ) = at and φ(bζ) = bφ(ζ) = bt. For f ∈ C(F ) the functional calculus of commuting bounded normal operators yields (4) Each function g + β ∈ C0(C)∼ is of the form f (aζ, bζ) for some function f ∈ C(F ) with f ↾∂F≡ β. Indeed, we have φ(f (aζ, bζ)) = f (φ(aζ ), φ(bζ )) = f (at, bt). g(z) = g(aζ(z)−1bζ(z)) = f (aζ, bζ)(z), where f (z1, z2) :=(g(z2/z1) + β , (z1, z2) ∈ F β , (z1, z2) ∈ ∂F . To show that f is continuous at ∂F , assume (z1, z2) → (0, 0). From z22 = z1 − z2 g vanishes at infinity. This proves the uniqueness assertion. it follows z2/z1) =p1/z1 − 1 → ∞ since z1 → 0. Therefore g(z2/z1) → 0, since Existence: Equation (4) defines a ∗-homomorphism from C0(C)∼ into L(E). Inserting f (z1, z2) := z1 into (4) it follows that N (φ(aζ )) = N (at). Note the latter is trivial. Similarly, φ(bζ) = bt. Frm Proposition 7 we get aφ(ζ) = φ(aζ) = at and (cid:3) bφ(ζ) = φ(bζ ) = bt. From Theorem 7 we finally conclude that φ(ζ) = t. 1 5. Associated operators and affiliated operators Throughout this section we assume that the Hilbert A-module E is the C∗- algebra A itself equipped with the A-valued scalar product ha, bi := a∗b, a, b ∈ A, and that A is realized as a nondegenerate C∗-algebra on a Hilbert space H. Then L(E) is the multiplier algebra M(A) = {x ∈ B(H) : xA ⊆ A,Ax ⊆ A}. and R(E) is the set Aη of affiliated operators in the sense of Woronowicz [W91]. Recall that Aη is the set of operators t ∈ C(H) for which at = (I + t∗t)−1 ∈ M(A), bt = t(I + t∗t)−1 ∈ M(A), and atA is dense in A. We write tηA if t ∈ Aη. Note that t = bta−1 Definition 16. We say that an operator t ∈ C(H) is associated with A and write tµA if t ∈ Rgr(E). The set of associated operators with A is denoted by Aµ. for tηA. t UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 27 That is, by Theorem 7, t ∈ C(H) is in Aµ if and only if at, at∗ , bt ∈ M(A). Note that the density of the set atA in A is not required for tµA. Obviously, Aµ ⊆ Aη. Further, t ∈ Aµ is in Aη if and only if atA is dense in A. Lemma 16. M(A) = {tµAt ∈ B(H)}. Proof. If t ∈ M(A), then I + t∗t ∈ M(A) and so at ∈ M(A) and bt = tat ∈ M(A), hence tµA. Conversely, suppose that t is bounded. Then I + t∗t is bounded and at ∈ M(A), so that a−1 (cid:3) t = I + t∗t ∈ M(A). Therefore, t = bta−1 t ∈ M(A). For tµA all three operators at, at∗ , bt have to be in the multiplier algebra M(A), while for tηA it is only required that at, bt ∈ M(A) (and the density of atA). From tηA it follows that at∗ ∈ M(A). Therefore, it is natural to ask whether or not at ∈ M(A) and bt ∈ M(A) already imply that tµA, that is, at∗ ∈ M(A). This is true if t ∈ C(H) is normal, since then at = at∗ . Proposition 8 below contains an number of other sufficient conditions. In Example 11 we will show that this is not true in general. The following simple relations appeared already in Definition 12. Lemma 17. Let t ∈ C(H). Then: t∗ = btbt∗ and at∗ bt = btat. t∗ = btan−1 Proof. (1): We have at∗ − a2 by a similar reasoning starting with the operator att∗at. bt∗ for n ∈ N. t∗ = tt∗a2 t∗ = ta2 t t t∗ = btbt∗. The second equality follows (1) at∗ − a2 (2) b∗t = bt∗ . t∗ − an+1 (3) an (2): Let x, y ∈ H. Then at∗ y ∈ D(tt∗) ⊆ D(t∗) and using (1) we obtain hbtx, at∗ yi = hatx, t∗at∗yi = hx, atbt∗yi = hx, bt∗ at∗ yi . Therefore b∗t = bt∗, since R(at∗ ) is dense and bt and bt∗ are bounded. (3) is easily derived from (1). (cid:3) Proposition 8. Suppose that at, bt ∈ M(A). Each of the following conditions imply that at∗ ∈ M(A) and so tµA. (1) 0 ∈ ρ(t). (2) kat∗k < 1, or equivalently, tt∗ ≥ ε for some ǫ > 0. (3) M(A)sa is closed under strong convergence of monotone sequences. (4) tt∗ = qt∗t for some q > 0. Proof. Clearly, from (1) it follows that 0 ∈ ρ(t∗) which in turn implies (2). (2), (3): By Lemma 17 and the assumptions at, bt ∈ M(A) we have at∗ − an+1 t∗ = bt(I + . . . + an−1 If (2) is fulfilled, then an+1 t∗ → 0 in M(A), hence at∗ ∈ M(A). On the other side, an+1 t∗ ∈ M(A)sa is monotone decreasing and strongly converging. Hence again by assumption (3) it follows at∗ ∈ M(A). (4) finally follows from the relations at∗ = (I + tt∗)−1 = (I + qt∗t)−1 = q−1(I + (q−1 − 1)at)−1at ∈ M(A). )b∗t ∈ M(A)sa. t (cid:3) Proposition 9. Suppose that t ∈ C(H) and 0 ∈ ρ(t). Then tµA if and only if t−1 ∈ M(A). Proof. Since 0 ∈ ρ(t), (t∗)−1 = (t−1)∗ ∈ B(H). Simple computations show that I − at∗ = (I + (t−1)∗t−1)−1. I − at = (I + t−1(t−1)∗)−1, From these identities we conclude that t−1 ∈ M(A), so (t−1)∗ ∈ M(A), implies that at, bt, at∗ ∈ M(A), that is, tµA. bt = (t−1)∗(I − at), 28 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Conversely, suppose that tµA. Then, by Lemma 17,(1), we have bt∗ = (bt)∗ ∈ (cid:3) M(A) and at∗ ∈ M(A). Therefore, t−1 = bt∗ (I − at∗ )−1 ∈ M(A). Corollary 6. If t ∈ C(H) and tµA, then (I + t∗t)µA. Proof. Since tµA, we have (I + t∗t)−1 = at ∈ M(A). Since 0 ∈ ρ(1 + t∗t) and (1 + t∗t)−1 = at ∈ M(A), we obtain (1 + t∗t)µA by Proposition 9. (cid:3) Corollary 7. Suppose that tµA and sµA. (1) If 0 ∈ ρ(t) and λ ∈ ρ(t) with 0 < λ < 1/kt−1k, then (t − λ)µA. (2) If 0 ∈ ρ(t) ∩ ρ(s), then tsµA. Proof. Both assertions follow immediately from Proposition 9. For (1) we use the equality (t − λI)−1 = λ−1t−1(λ−1 − t−1)−1 ∈ M(A), while for (2) we note that 0 ∈ ρ(ts) and (ts)−1 = s−1t−1 ∈ M(A). (cid:3) Next we investigate affiliated operators and their resolvents. Before we turn to the main result we prove two simple lemmas. Lemma 18. Let A be a C∗-algebra acting on a Hilbert space H. Let s ∈ B(H) and x, y ∈ M(A). Suppose that xA and yA are dense in A. If sx ∈ M(A) and s∗y ∈ M(A), then s ∈ M(A). Proof. Let a ∈ A. Since xA is dense in A, there are elements an ∈ A, n ∈ N , such that xan → a in A. Hence sxan → sa in A. Since sx ∈ M(A) by assumption, sxan ∈ A and so sa ∈ A. Replacing x by y and s by s∗ it follows that s∗a ∈ A. Therefore, s ∈ M(A). (cid:3) Lemma 19. Let A be a C∗-algebra and x, y ∈ M(A). Suppose that λy ≥ xx∗ for some λ > 0. If xA is dense in A, so is yA. In particular, xA is dense in A if and only if xx∗A is. Proof. Assume to the contrary that yA 6= xA = A. Then the closure of (yA)∗ is a proper left ideal. Hence there exists a state ω of A that annihilates (yA)∗ (see e.g. [Dix77, Lemma 2.9.4]). Let πω be the GNS representation of A associated with the state ω and let ϕω be the corresponding cyclic vector ϕω. We denote the extension of πω to the multiplier algebra M(A) also by the symbol πω. Then (5) 0 = ω((ya)∗) = hπω(a∗y)ϕω, ϕωi = hπω(y)ϕω, πω(a)ϕωi for all a ∈ A, so that πω(y)ϕω = 0. Therefore, ω(xa)2 = hπω(a)ϕω, πω(x∗)ϕωi2 = kπω(a)ϕωk2kπω(x∗)ϕωk2 = kπω(a)ϕωk2hπω(xx∗)ϕω, ϕωi ≤ kπω(a)ϕωk2λhπω(y)ϕω, ϕωi = 0 for a ∈ A, that is, ω annihilates xA. Hence xA is not dense in A which is a contradiction, since we assumed that xA = A. Applying this to the case y = xx∗ we conclude that xx∗A is dense provided that xA is dense. Since the converse implication is trivial, it follows that xx∗A is dense if and only if xA is dense. (cid:3) The following theorem appeared in [Sch05]. Theorem 17. Suppose A is a C∗-algebra acting on a Hilbert space H and t is a densely defined closed operator on H with non-empty resolvent set. Let λ ∈ ρ(t). Then tηA if and only if (t − λI)−1 ∈ M(A) and (t − λI)−1A and (t∗ − λI)−1A are dense in A. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 29 Proof. Since tηA is equivalent to (t − λI)ηA (see [W91], p. 412, Example 1), we can assume without restriction of generality that λ = 0. Then t−1 and (t∗)−1 are in B(H). First we suppose that tηA. Set x := (I + (tt∗)−1)−1 and s := t−1. Since tηA implies t∗ηA, it follows that zt∗ = t∗(I + tt∗)−1/2 = (zt)∗ ∈ M(A). Therefore, we obtain (I +tt∗)−1 = I−ztz∗t ∈ M(A) and hence (I +tt∗)−1/2 ∈ M(A). These relations imply that sx ≡ t−1(I + (tt∗)−1)−1 = t∗(tt∗)−1(I + (tt∗)−1)−1 = t∗(I + tt∗)−1 = zt∗ (I + tt∗)−1/2 ∈ M(A). (6) Since x := (I +(tt∗)−1)−1 = I−(I +tt∗)−1 ∈ M(A) and x−1 is also bounded, we have x−1 ∈ M(A) and hence xA = A. Recall that sx ∈ M(A) by (6). Now we interchange the roles of t and t∗ and set y := (I + (t∗t)−1)−1. By a similar reasoning as in (6) we derive s∗y ∈ M(A). Further, y ∈ M(A) and yA = A. Hence the assumptions of Lemma 18 are satisfied, so we obtain t−1 = s ∈ M(A). Recall that (I + t∗t)−1A is dense in A, because tηA. Therefore, since (I + t∗t)−1A = (t∗t)−1(I + (t∗t)−1)−1A ⊆ (t∗t)−1A = t−1(t∗)−1A ⊆ t−1A, t−1A is dense in A. Replacing t by t∗, it follows that (t∗)−1A is dense in A. This completes the proof of the only if part. Conversely, let us assume that t−1 ∈ M(A) and that t−1A and (t∗)−1A are dense in A. Then I − z∗t zt = (I + t∗t)−1 = t−1(t−1)∗(I + t−1(t−1)∗)−1 ∈ M(A) and zt(I − z∗t zt)1/2 = t(I + t∗t)−1 = (t−1)∗(I + t−1(t−1)∗)−1 ∈ M(A). Therefore, setting x := (I − z∗t zt)1/2 and s := zt, we have x ∈ M(A) and sx ∈ M(A). Since t has a bounded inverse, there exists ǫ ∈ (0, 1/4) such that t∗t ≥ 2ǫI. Then I + t∗t ≤ 1 ǫ t∗t and hence (I + t∗t)−1 ≥ ǫt−1(t−1)∗. Therefore, since t−1A is dense in A by assumption, (I + t∗t)−1A = (I − z∗t zt)A = x2A is dense in A by Lemma 19. Since x ≥ 0, xA dense in A again by Lemma 19. By the assumptions we can interchange the roles of t and t∗. Then we obtain y := (I − ztz∗t )1/2 ∈ M(A) and s∗y = z∗t y ∈ M(A). Further, (I + tt∗)−1A = (I − ztz∗t )A = y2A in A and hence yA are dense in A. Thus, zt ∈ M(A) by Lemma 18 and hence tηA. (cid:3) 2ǫ t∗t + t∗t ≤ 1 The preceding theorem has a number of interesting corollaries. For these results we assume that A is a C∗-algebra acting on a Hilbert space H and t and s are densely defined closed operators on H. Corollary 8. Suppose that t, sηA, λ ∈ ρ(t) and µ ∈ ρ(s). Then we have −λµ ∈ ρ(ts − λs − µt) and (ts − λs − µt)ηA. Proof. By some straightforward arguments one verifies that (7) (ts − λs − µt + λµI)−1 = (s − µI)−1(t − λI)−1, ((ts − λs − µt)∗ + λµI)−1 = (t∗ − λI)−1(s∗ − λI)−1. (8) Hence −λµ ∈ ρ(ts − λs − µt). From the only if part of Theorem 17 it follows that the operators in (7) and in (8) belong to M(A) and that they maps A densely into A. Therefore, by the if part of Theorem 17, (ts − λs − µt)ηA. (cid:3) Proposition 10. Suppose that λ ∈ ρ(t), s(t− λI)−1 ∈ M(A) and ks(t− λI)−1k < 1. Then (t + s)ηA. Proof. By Theorem 17, (t − λI)−1 ∈ M(A) and (t − λI)−1A and (t∗ − λI)−1A are dense in A. By the assumption we have r := s(t − λI)−1 ∈ M(A) and krk < 1. Therefore (I + r)−1 is bounded and an element of M(A), since I + r ∈ M(A). Further, since t − λI and I + r are bijective and (t + s − λI)ϕ = (I + r)(t − λI)ϕ for ϕ ∈ D(t) ⊆ D(s), the map t + s− λI : D(t) → H is bijective. Hence λ ∈ ρ(t + s) 30 REN ´E GEBHARDT AND KONRAD SCHM UDGEN and (t + s − λI)−1 = (t − λI)−1(I + r)−1 ∈ M(A). Because I + r is an invertible element of M(A), the density of (t − λI)−1A implies the density of (t + s − λI)−1A in A. Finally, (t∗ + s∗ − λI)−1 = (I + r∗)−1(t∗ − λI)−1 maps A densely into A, since kr∗k < 1. Now applying again Theorem 17 we obtain (t + s)ηA. (cid:3) Corollary 9. Let A ⊆ B(H) a C∗-algebra and suppose that t, sηA. If λ ∈ ρ(t), 0 ∈ ρ(s), and kλ(t − λI)−1k < 1, then tsηA. Proof. By Corollary 8 we have 0 ∈ ρ(ts − λs) and (ts − λs)ηA. Since tηA and λ ∈ ρ(t), it follows from Theorem 17 that (t − λ)−1 ∈ M(A). Therefore, λs(ts − λs)−1 = λs((t − λI)s)−1 = λ(t − λI)−1 ∈ M(A). Hence, since kλ(t − λ)−1k < 1 by assumption, Proposition 10 applies to the opera- tors t := ts − λs and s := λs and implies that t + s = tsηA. (cid:3) Corollary 10. Let A ⊆ B(H) a C∗-algebra. Suppose that tηA and λ, µ ∈ ρ(t). For an operator s on H we have s(t − λI)−1 ∈ M(A) if and only if s(t − µI)−1 ∈ M(A). Proof. Since (t− λI)−1 ∈ M(A) and (t− µI)−1 ∈ M(A) by Theorem 17, the assertion follows from the identity s(t−λI)−1−s(t−µI)−1 = (λ−µ)s(t−µI)−1(t−λI)−1 = (λ−µ)s(t−λI)−1(t−µI)−1. (cid:3) Corollary 11. Let A ⊆ B(H) a C∗-algebra. Suppose that tηA is a self-adjoint operator and s is a symmetric t-bounded operator on H with t-bound less than 1. If s(t − λI)−1 ∈ M(A) for some λ ∈ ρ(t), then (t + s)ηA and t + s is self-adjoint. Proof. The proof can be given by repeating the standard proof of the Kato-Rellich (cid:3) theorem and using Lemma 10 and Corollary 10. It is natural to ask whether or not the second density assumption in Theorem 17 can be omitted, that is, when does the density of (t− λI)−1A for (t− λI)−1 ∈ M(A) imply the density of (t∗ − λI)−1A in A? A counterexample is provided by Example 9 below. The next proposition shows that the answer is affirmative if the distance of (t − λI)−1 to the set A−1 of invertible elements of A is zero. Proposition 11. Let t be a densely defined closed operator and A a C∗-algebra acting on a Hilbert space H. Suppose that t has a bounded inverse t−1 contained in M(A). Assume that dist(t−1,A−1) = 0. If t−1A dense in A, so is (t∗)−1A. Proof. Set x := t−1. Then (t∗)−1 = (t−1)∗ = x∗. Assume to the contrary that x∗A = (t∗)−1A is not dense in A. Then, as in the proof of Lemma 19, there is a state ω on A that annihilates (x∗A)∗. Arguing as in line (5) it follows that πω(x)ϕω = 0. Let x = vx be the polar decomposition of the operator x. Since x ∈ M(A), we have x = (x∗x)1/2 ∈ M(A). For ε > 0 let fε denote a continuous function on R such that fε(τ ) = 0 on [0, ε 2 , ε] and fε(τ ) = u on [ε, +∞). By [Ped98] Theorem 6.1, applied to the multiplier algebra M(A), there exists a unitary operator uε ∈ M(A) such that 2 ], fε(τ ) ≤ ε on [ ε vfε(x) = uεfε(x) ∈ M(A). Clearly, x = limε→+0 fε(x) in M(A). Therefore, 0 = πω(x)ϕω = πω(vx)ϕω = lim ε→+0 πω(vfε(x))ϕω = lim ε→+0 πω(uε)πω(fε(x))ϕω, UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 31 so that 0 = limε→+0 πω(fε(x))ϕω = πω(x)ϕω. For a ∈ A we now obtain 0 = hπω(x)2ϕω, πω(a)ϕωi = hπω(x∗x)ϕω πω(a)ϕωi = hϕω, πω(x∗xa)ϕωi = ω((x∗xa)∗). This implies that x∗xA is not dense in A. Hence xA is not dense by Lemma 19 (cid:3) which is the desired contradiction. Example 9. Let H be the Hilbert space l2(N2) and let A be the C∗-algebra The multiplier algebra of A is A =(cid:18)K(H) K(H) K(H) B(H)(cid:19) . M(A) =(cid:18)B(H) K(H) K(H) B(H)(cid:19) . (9) (10) (11) Let {ekl}k,l∈N0 be the standard orthonormal basis of H. Let s ∈ B(H) be the shift operator given by sekl = ek+1,l and let P0 be the orthogonal projection onto N (s∗). Clearly, {e0,l}l∈N0 is an orthonormal basis of P0H. Further, let {λkl}k,l∈N0 be a double sequence of positive numbers such that limk,l→∞ λkl = 0. Define a self- adjoint compact operator on H by rekl := λklekl, k, l ∈ N0. Let x ∈ B(H ⊕ H) defined by the operator matrix x :=(cid:18)s 0 s∗(cid:19) . r Since λkl > 0 for all k, l, the compression P0r ↾ P0H of r to P0H has trivial kernel and dense range. Using this fact it is easily seen that N (x) = {0} and R(x) is dense in H ⊕ H. Hence t := x−1 is a densely defined closed operator on the Hilbert space H ⊕ H. By (10), we have t−1 = x ∈ M(A). Statement: t−1A = xA is dense in A, while (t∗)−1A = x∗A is not dense in A. Proof. Let y be an element of A. Then y given by an operator matrix a, b, c ∈ K(H) and d ∈ B(H), and have (12) y :=(cid:18)a b c d(cid:19) , xy =(cid:18)sa + rc s∗c sb + rd s∗d (cid:19) . Since K(H) = s∗sK(H) ⊆ s∗K(H) ⊆ K(H), we have s∗K(H) = K(H). Similarly, s∗B(H) = B(H). Since the range of r contains all rank one operators ekl ⊗ enm, sK(H) + rK(H) is dense in K(H). Therefore, by (9) and (12), xA is dense in A. Next we prove the second assertion. First we note that P0(rK(H) + sB(H)) = P0(rK(H)) ⊆ P0K(H). This implies that rK(H) + sB(H) is not dense in B(H). Therefore, since x∗y =(cid:18) s∗a ra + sc s∗b rb + sd(cid:19) , it follows from (9) that the set x∗A is not dense in A. (cid:3) 32 REN ´E GEBHARDT AND KONRAD SCHM UDGEN 6. Examples 6.1. Matrices of commutative C∗-algebras and its multipliers. In this sub- section we use matrices over commutative C∗-algebras to construct simple examples of operators that help to delimit the general theory. Let A be a C∗-algebra. If Aij ⊆ A for i, j ∈ {1, 2} set Let X be a locally compact non-compact Hausdorff space and set A21 A22 (cid:19) :=(cid:26)(cid:18) a11 a12 (cid:18) A11 A22 A0 :=(cid:18) C0(X) C0(X) a21 a22 (cid:19)aij ∈ Aij for i, j ∈ {1, 2}(cid:27) . C0(X) C0(X)∼ (cid:19) , C0(X) C0(X) (cid:19) , A :=(cid:18) C0(X) C0(X) where C0(X)∼ := C0(X) + C · 1. A straightforward computation shows that the left multiplier algebra LM(A) and the multiplier algebra M(A) are given by From this we can read off that all elements of the form LM(A) =(cid:18) Cb(X) C0(X) M(A) =(cid:18) Cb(X) C0(X) C0(X) C0(X)∼ (cid:19) . Cb(X) C0(X)∼ (cid:19) , (cid:18) ∗ ∗ (cid:19) ∈ LM(A) with f ∈ Cb(X) \ C0(X) ∗ f act as operators t on A defined on the whole space such that the adjoints are not defined on the whole space. From now on let X = R. 1 0 Example 10. Let f = 1 and set the other matrix entries zero. Then t acts as D(t) = A, t =(cid:18) 0 Hence t∗ is essentially defined. Further, it is easily checked that t∗∗ = t and 0 (cid:19) . 0 (cid:19) , D(1 + t∗t) = A, 1 + t∗t =(cid:18) 2 0 0 1 (cid:19) , D(1 + tt∗) = A0, 1 + tt∗ =(cid:18) 1 0 0 2 (cid:19) . and D(t∗) = A0, t∗ =(cid:18) 0 From these formuals we read off that R(1 + t∗t) = A and R(1 + tt∗) = A0 ( A. Hence at is adjointable, while at∗ is not. In fact, bt is also not adjointable, since 1 0 D(bt) = A, bt =(cid:18) 0 1/2 0 (cid:19) /∈ M(A). 0 In the following slighty more sophisticated example at and bt are both ad- jointable, while at∗ is not adjointable. c For the adjoint t∗ we obtain Example 11. Let f, g ∈ C(R) be functions given by f (x) := xq1 + sin2(x) and g(x) := xp1 + cos2(x). Then f (x)2 + g(x)2 = 3x2. Define t : A → A by t =(cid:18) 0 f D(t) :=(cid:26)(cid:18) a b g (cid:19) . t =(cid:18) 0 g (cid:19) . D(t∗) :=(cid:26)(cid:18) a b 1+f2+g2 ! ∈ A. at =(cid:18) 1 d (cid:19) ∈ A f c, f d, gc ∈ C0(R), gd ∈ C0(R)∼(cid:27), d (cid:19) ∈ A f a + gc ∈ C0(R), f b + gd ∈ C0(R)∼(cid:27), 1+f2+g2 (cid:19) ∈ A, It is now easily verified that 1 + t∗t is surjective and bt = 0 1+f2+g2 0 1 0 0 0 0 f f g c UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 33 The operator at∗ is computed as D(at∗ ) = A0, at∗ = 1 1 + f2 + g2(cid:18) 1 + g2 −gf −f g 1 + f2 (cid:19) . That is, at ∈ A and bt ∈ A are adjointable, but at∗ /∈ M(A) is not adjointable. Example 12. Now we consider the operator t given by D(t) = A, t =(cid:18) 0 1 0 1 (cid:19) , and D(t∗) = A0, t∗ =(cid:18) 0 0 1 1 (cid:19) . Then one easily proves that R(t) * D(t∗) and D(t∗t) = A0 ( A = D(t). particular, D(t∗t) is not dense in the domain D(t)! 6.2. A fraction algebra related to the Weyl algebra. Let P = −i d dt and Q = t be the momentum and position operators acting as self-adjoint operators on the Hilbert space L2(R). Fix α, β ∈ R\{0} and define bounded operators x and y by In x := (Q − αiI)−1 and y := (P − βiI)−1. It is not difficult to verify that these operators satisfy the commutation relations (13) (14) x − x∗ = 2αix∗x = 2αixx∗, y − y∗ = 2βiy∗y = 2βiyy∗, xy − yx = −ixy2x = −iyx2y, xy∗ − y∗x = −ix(y∗)2x = −iy∗x2y∗. J0 := yxX = xyX = X yx = X xy Let X be the unital ∗-subalgebra of B(L2(R)) generated by x and y. Since the operators x, x∗, y, y∗ are bijections of the Schwartz space S(R), so are their inverses. Hence τx(·) := x·x−1 and τy(·) := y·y−1 are automorphisms of the algebra L(S(R)) of linear operators on the Schwartz space S(R). From the relations (13) and (14) we conclude that τx and τy leave the algebra X invariant, so they are algebra automorphisms of X . Hence (15) and J0 is a two-sided ∗-ideal of the ∗-algebra X . Let Fx be the unital ∗-subalgebra of X generated by x, that is, Fx is the com- mutative ∗-algebra of polynomials f (x, x∗) in x and x∗ with complex coefficients. Likewise, Fy denotes the unital ∗-subalgebra of X generated by y. Note that Fx ∩ Fy = C · 1. From the relations it follows easily that X is the direct sum of vector spaces Fx + Fy and J0. Hence each element a ∈ X can be written as (16) where f1, g1, f2, g2 are polynomials with g1(0, 0) = f2(0, 0) = 0 and b1, b2 ∈ X . Moreover, these triples {f1, g1, b1} and {f2, g2, b2} are uniquely determined by a. Let ε > 0 and λ ∈ R. We denote by χε the characteristic function of the interval [0, ε]. Put ωε,λ(t) := 1√ε χε(t − λ). Lemma 20. For polynomials f ∈ C[x, x∗] and g ∈ C[y, y∗], where g(0, 0) = 0, and b ∈ X we have (17) a = f1(x, x∗) + g1(y, y∗) + yxb1 = f2(x, x∗) + g2(y, y∗) + xyb2, lim (18) (19) lim lim ε→+0 hf (x, x∗) ωε,λ, ωε,λi = f ((λ − αi)−1, (λ + αi)−1), ε→+0 hg(y, y∗) ωε,λ, ωε,λi = 0, ε→+0 hyxb ωε,λ, ωε,λi = 0. hϕωε,λ, ωε,λi = ϕ(λ) +Z λ+ε ε−1(ϕ(t) − ϕ(λ))dt → ϕ(λ) λ Proof. Suppose that ϕ ∈ C1(R). Then we have which in turn implies (17). 34 REN ´E GEBHARDT AND KONRAD SCHM UDGEN Next we prove (18). The crucial step for this is to show that (20) lim ε→+0 yωε,λ = lim ε→+0 y∗ωε,λ = 0 in L2(R). Without loss of generality we can assume that β < 0. Since β < 0, for the resolvent of P = −i d dt we obtain (yϕ)(t) = ((P − βiI)−1ϕ)(t) = −iZ ∞ t Hence we compute (yωε,λ)(t) = 0 for t ≥ λ + ε, eβ(s−t)ϕ(s)ds, ϕ ∈ L2(R). (yωε,λ)(t) = −i √ε eβs dt = −i β√ε (eβ(λ+ε−t) − 1) for λ ≤ t ≤ λ + ε, and (yωε,λ)(t) = −i √ε t e−βtZ λ+ε e−βtZ λ+ε λ eβs dt = −i β√ε e−βt(eβ(λ+ε) − eβ(λ−ε)) for t ≤ λ. From these formulas we easily derive that limε→+0 yωε,λ = 0. Replacing β by −β a similar reasoning yields limε→+0 y∗ωε,λ = 0. Since the operator y is bounded and g(0, 0) = 0, (20) implies (18). Since kωε,λk = 1, it follows from (20) that hyxb ωε,λ, ωε,λi = hxb ωε,λ, y∗ωε,λi ≤ kxbk ky∗ωε,λk → 0 which proves (19). (cid:3) Next we define two circles Kα and Kβ intersecting in the origin by Kα := {(z, 0) ∈ C2 : z − z = 2αiz2}, Kβ := {(0, w) ∈ C2 : w − w = 2βiw2}. Let R = R ∪ {∞} be the one point compactification of the real line. The maps λ → (zλ, 0) := ((λ − αi)−1, 0) and λ → (0, wλ) := (0, (λ − βi)−1), where z∞ := 0 and w∞ := 0, are bijection of R onto Kα and Kβ, respectively. Let z ∈ Kα and w ∈ Kβ. For a as in (16) we define πx,z(a) = f1(z, z), πy,w(a) = g2(w, w). Lemma 21. For all z ∈ Kα and w ∈ Kβ, πx,z and πy,w are one dimensional ∗-representations of the ∗-algebra X such that (21) Moreover, for f ∈ Fx and g ∈ Fy we have (22) and πy,w(a) ≤ kak πx,z(a) ≤ kak for a ∈ X . kfk = sup z∈Kα f (z, z) and kgk = sup w∈Kβ g(w, w). Proof. A simple computation based on the relations (13) and (14) shows that πx,z and πy,w are well-defined ∗-homomorphisms of X . Let λ ∈ R. From the formulas (17), (18) and (19) we infer that πx,zλ(a) = f1((λ − αi)−1, (λ + αi)−1) = lim ε→+0 haωε,λ, ωε,λi. Therefore, since kωε,λk = 1, we obtain πx,zλ(a) ≤ kak. Passing to the limit λ → ∞, we get πx,z0(a) ≤ kak. This proves the first inequality of (21) for all z ∈ Kα. The inequality πy,w(a) ≤ kak follows in a similar manner by interchanging the role of x and y and using the counterparts of formulas (17), (18) and (19). Finally, we prove (22). One inequality of the equality (22) was shown by (21). The reversed inequality follows at once from the fact that the spectra of the self- (cid:3) adjoint operators P and Q are equal to R. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 35 It was noted in [Sch10] (and is easily verified by using (13)), the ∗-algebra X is algebraically bounded, that is, given a ∈ X there exist γa > 0 and finitely many elements ai ∈ X such that γa = a∗a +Pi a∗i ai. Hence kakun := sup π kπ(a)k, a ∈ X , a defines a C∗-norm on X , where the supremum is taken over all ∗-representation of X , and the completion Xun of (X ;k · kun) is called the universal C∗-algebra of X . (The supremum is finite, since kπ(a)k ≤ γ1/2 for all a ∈ X .) By decomposition theory it suffices to take all irreducible ∗-representations π. As proved in [Sch10], the irreducible ∗-representation of X are the one-dimensional representations πx,z and πy,w,where z ∈ Xx and w ∈ Kβ, and the identity representation π0 acting on the Hilbert space L2(R). From (21) it follows that the universal C∗-norm k · kun coincides with the operator norm k · k on L2(R). Hence the universal C∗-algebra Xun is just the closure X of X in B(L2(R)). By (22) the closures of Fx and Fy in B(L2(R)) are the commutative C∗-subalgebras C(Kα) and C(Kβ), respectively, and the closure J of J0 in B(L2(R)) is a two-sided ∗-ideal of the C∗-algebra Xun = X . Lemma 22. J0 is a two-sided essential ideal of Xun. Proof. Let a ∈ Xun be such that aJ0 = {0}. Then axy = 0 in L2(R). Since x and (cid:3) y are bijection, this implies a = 0. The operator x∗y∗yx is an integral operator on L2(R) with kernel K(t, s) : (2β)−1(t + αi)−1(s − αi)−1e−βt−s. Since K ∈ L2(R2), the operator x∗y∗yx = yx2 is compact, so are yx and hence yx. Hence J0 ⊆ K(L2(R)) by (15) and therefore J = K(L2(R)). Now we define two operators, denoted by Q and P , on the C∗-algebra Xun by Q := αiI + x−1, D(Q) := xXun and P := βi + y−1, D(P ) := yXun, that is, Q(xa) = αixa + a and P (ya) = βiya + a, where a ∈ Xun. Theorem 18. Q and P are graph regular self-adjoint operators on the C∗-algebra Xun = X . Proof. We carry out the proof for Q; a similar reasoning yields the assertions for P . Since xXun and x∗Xun contain the essential ideal J0 (by Lemma 22), xXun and x∗Xun are essential in the C∗-algebra Xun. Therefore, by Theorem 3, x−1 and (x∗)−1 are graph regular operators on Xun, so Q and P are graph regular by Proposition Corollary 6. Further, (x−1)∗ = (x∗)−1 and hence Q∗ = −αiI + (x∗)−1 by Theorem 3 and Proposition 2 (2). From the first relation of (13) it follows that −αiI + (x∗)−1 = αiI + x−1. Hence Q = Q∗, that is, Q is self-adjoint. (cid:3) The operators Q and P are not regular on Xun, since neither xXun nor yXun is dense in Xun. Note that the corresponding restrictions of Q and P are affiliated with the essential ideal J = K(L2(R)) of the C∗-algebra Xun. 6.3. Unbounded Toeplitz operators. Let L2(T) be the Hilbert space of square integrable functions on the unit circle T with scalar product hf, gi :=Z 1 0 f (e2πit)g(e2πit) dt f, g ∈ L2(T), and let P denote the projection of L2(T) on the closed subspace H 2(T) generated by {zn := e2πitnn ∈ N0}. For φ ∈ L∞(T) the Toeplitz operator Tφ is the bounded operator on the Hilbert space H 2(T) is defined by Tφf := P φf , f ∈ H 2(T). 36 REN ´E GEBHARDT AND KONRAD SCHM UDGEN The C∗-algebra generated by the unilateral shift S := Tz is the Toeplitz algebra T := {Tφφ ∈ C(T)} ∔ K(H 2(T)). Our aim is to construct a class of examples of graph regular (unbounded) Toeplitz operators on the C∗-algebra T . Let p, q ∈ C[z] be relatively prime polynomials such that q has no zeros in the open unit disc D. Then the Toeplitz operator with rational symbol p/q is defined by D(Tp/q) := {f ∈ H2(T) p q f ∈ H2(T)}, Tp/qf := p q f (f ∈ D(Tp/q)), Since Tp/q is a multiplication operator, Tp/q is a closed densely defined operator on the Hilbert space H2(T). Theorem 19. Suppose that p, q are relatively prime polynomials such that q has no zero in the open unit disc. Then the Toeplitz operator Tp/q is associated with the Toeplitz algebra T . Further, Tp/q is affiliated with the Toeplitz algebra if and only if in addition q has no zero on the unit circle. Proof. Since q has no zero in D, q is an outer function (see e.g. [RR85]). Now we use an argument from [Sar08, Section 3]. Since p and q are relatively prime, we have p2 + q2 > 0 on the closed unit disc D. Therefore, by the Riezs- Fej´er Theorem [RR85], there exists a polynomial r ∈ C[z] such that r has no zero in D and p2 + q2 = r2 on T. Let f := q/r and g := p/r. Then f and g are continuous and in the unit ball of H∞(T), f is outer, f2 + g2 = 1 on T. Upon multiplying r by some constant of modulus one we can assume that f (0) > 0. From [Sar94, Proposition 5.3] it follows that D(Tp/q) = f H 2(T) and Tp/q = TgT −1 f . Moreover, T ∗p/q = (T −1 f )∗T ∗g = T −1 f Tg. Using these facts we compute TgTgTf = T −1 1 + T ∗p/qTp/q = 1 + T −1 f T −1 f = (Tf Tf )−1, = T −1 f f (Tf Tf + TgTg)T −1 f = T −1 f (Tf2 + Tg2)T −1 f 1 + Tp/qT ∗p/q = 1 + TgT −1 f T −1 f Tg = 1 + Tg(Tf Tf )−1Tg = 1 + Tg(1 − TgTg)−1Tg = 1 + (1 − TgTg)−1TgTg = (1 − TgTg)−1. Hence aTp/q = Tf Tf and aT ∗ p/q = I − TgTg are in T . Further, f Tf Tf = TgTf ∈ T . bTp/q = Tp/qATp/q = TgT −1 , bTp/q ∈ T , Tp/q is associated with the C∗-algebra T . Since aTp/q , aT ∗ p/q Suppose now q has a zero at some λ ∈ T. Then a has a zero at λ as well. For z ∈ T let ωz be the character on T given by (23) If Tφ + K ∈ T , then Tf Tf (Tφ + K) = Tf2φ + K for some K ∈ K(H2(T)). Hence (φ ∈ C(T), K ∈ K(H 2(T)). ωz(Tφ + K) = φ(z) ωλ(aTp/q (Tφ + K)) = ωλ(Tf2φ + K) = f (λ)2φ(λ) = 0. Therefore, aTp/qT is not dense in T and hence Tp/q is not affiliated with T . so in particular, Tp/q is affiliated with T . On the other hand, if q has no zero on T, then p/q ∈ C(T) and hence Tp/q ∈ T , (cid:3) The simplest interesting example is the following. Example 13. Set p(z) = 1 and q(z) = 1 − z, so that p/q = 1/(1 − z). Then, by Theorem 19, T1/(1−z) associated with T , but T1/(1−z) is not affiliated with T . In fact, T1/(1−z) = (I − S)−1. UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 37 6.4. Heisenberg group. Let H be the 3-dimensional Heisenberg group, that is, H is the Lie group whose differential manifold is the vector space R3 and whose multiplication is given by (x1, x2, x3)(x′1, x′2, x′3) := (x1 + x′1, x2 + x′2, x3 + x′3 + 1 2 (x1x′2 − x′1x2)). The C∗-algebra C∗(H) of the Lie group H was described in [LT11]. We briefly repeat this result. First we recall that C∗(H) is defined as the completion of L1(H) with respect to the norm kfk = sup{kπU (f )k : U unitary representation of H}. where πU is the ∗-representation of L1(H) associated with U , that is, U (x1, x2, x3)f (x1, x2, x3) dx1dx2dx3, f ∈ L1(H). πU (f ) :=ZR3 The irreducible unitary representations of H consist of a series Uλ, λ ∈ R×, of infinite dimensional representations acting on L2(R) and of a series Ua, a ∈ R2, of one dimensional representations. For (x1, x2, x3) ∈ H, these representations act as (Uλ(x1, x2, x3)ξ)(s) = e−2πiλ(x3+ 1 2 x1x2+sx2)ξ(s − x1), ξ ∈ L2(R), λ ∈ R×, Ua(x1, x2, x3) = e−2πi(a1x1+a2x2), a = (a1, a2) ∈ R2. The Lie algebra of H has a basis {X, Y, Z} with commutation relations [X, Y ] = Z, [X, Z] = [Y, Z] = 0 and we have dUλ(iZ) = 2πλI and dUa(iZ) = 0. Now let F be the C∗-algebra of all operator fields F = (F (λ); λ ∈ R) satifying the following conditions: (i) F (λ) is a compact operator on L2(R) for each λ ∈ R×, (ii) F (0) ∈ C0(R2), (iii) R× ∋ λ → F (λ) ∈ B(L2(R) is norm continuous, (iv) limλ→∞ kF (λ)k = 0. Let η be a fixed function of the Schwartz space S(R) of norm one in L2(R). For ξ ∈ L2(R), let Pξ denote the projection on the one dimensional subspace C · ξ. Then for h ∈ C0(R2) and λ ∈ R× := R\{0}, the operator νλ(h) is defined by (24) h(x1, x2)Pη(λ;x1,x2)λ−1dx1dx2, νλ(h) :=ZR2 where h denotes the Fourier transform of h and By Proposition 2.14 in [LT11], we have η(λ; x1, x2)(s) := λ1/4e2πix1s η(cid:0)λ1/2(s + x2λ−1)(cid:1), λ→0 kνλ(h)k = hk∞ for h ∈ C0(R2). lim (25) x1, x2, s ∈ R. Then, according to Theorem 2.16 in [LT11], the C∗-algebra C∗(H) is the C∗- subalgebra of C∗(H) formed by all operator fields F ∈ F such that (26) lim λ→0 kF (λ) − νλ(F (0))k = 0, where νλ : C0(R2) → F is defined by (24), and for c ∈ C∗(H), we have F (c)(λ) = πUλ , λ ∈ R×, and F (c)(0)(a) = πUa (c), a ∈ R2. On the other hand, it was proved in [WN92] that the Lie algebra generators X, Y, Z act as skew-adjoint regular operators on the C∗-algebra C∗(H). We show that the range R(iZ) is essential in C∗(H). Assume that G(λ) ∈ C∗(H) and G(λ) ∈ R(iZ)⊥. Since dUλ(iZ) = 2πλI for λ ∈ R×, R(iZ) contains all vector fields F (λ) ∈ F of compact support contained in R×. This implies that G(λ) = 0 38 REN ´E GEBHARDT AND KONRAD SCHM UDGEN on R×. Therefore, limλ→0 νλ(G(0)) = 0 by (26) and hence [G(0) = 0 by (25), so G(0) ∈ C0(R2) is zero. Thus G = 0 in C∗(H) which proves that R(iZ) is essential. Further, iZ is self-adjoint, so (iZ)−1 is by Proposition 1. Since iZ is graph regular, so is (iZ)−1 by Proposition 3. Note that (iZ)−1 is not regular, because dUa(iZ) = 0 for a ∈ R2 and hence (iZ)−1 is not densely defined. Theorem 20. (iZ)−1 is a graph regular self-adjoint operator on the C∗-algebra C∗(H). References [B81] S. Baaj, Multiplicateurs non bornes, Thesis, Universite Pierre et Marie Curie 37(1981), 1 -- 44. [BJ83] S. Baaj, P. Julg, Theorie bivariante de Kasparov et operateurs non bornes dans les C- modules hilbertiens, C. R. Acad. Sci. Paris Sr. I Math. 296 (1983), no. 21, 875-878. [Dix77] J. Dixmier, C-algebras, North-Holland, Amsterdam, 1977. [FS10] M. Frank, K. Sharifi, Generalized inverses and polar decomposition of unbounded regular operators on Hilbert C ∗-modules, J. Operator Theory 64(2010), 377-386. [GVF] J.M. Gracia-Bondia, J. Varilly, H. Figueroa, Elements of Noncommutative Geometry, Birkhauser, Boston, 2001. [H89] M. Hilsum, Fonctorialite en K-Theorie bevariante pour les varietes lipschitziennes, K- Theory 3(1987), 401-440. [KL12] J. Kaad, M. Lesch, A local global principle for regular operators in Hilbert C ∗-modules, J. Funct. Anal. 262(2012), 4540-4569. [K53] I. Kaplansky, Moduls over operator algebras, Amer. J. Math. 75(1953), 839-858. [Kas80] G.G. Kasparov, Hilbert C ∗-modules: Theorems of Stinespring and Voiculescu, J. Oper- ator Theory 4(1980), 133 -- 150. [K97] D. Kucerovsky, The KK-product of unbounded modules, K-Theory 11(1997), 17 -- 34. [K02] D. Kucerovsky, Functional calculus and representations of C0(X) on a Hilbert module, Quart. J. Math. 53(2002), 467 -- 477. [L95] E. C. Lance, Hilbert C ∗-modules - A toolkit for operator algebraists, Cambridge Univ. Press, 1995. [LT11] J. Ludwig, L. Turowska, The C ∗-algebras of the Heisenberg group and of thread-like Lie groups, Math. Z. 268 (2011), 897-930. [Mg97] B. Maganja, Hilbert C ∗-modules in which all closed submodules are complemented, Proc. Math. Soc. 125(1997), 849 -- 852. [MT05] V. M. Manuilov, E. V. Troitsky, Hilbert C ∗-Modules, Amer. Math. Soc, RI, 2005. [Pal99] A. Pal, Regular operators on Hilbert C ∗-modules, J. Operator Theory 42 (1999), 331-350. [P73] W. L. Paschke, Inner product moduls over B∗-algebras, Trans. Amer. Math. Soc. 182(1973), 443-468. [Ped98] G. K. Pedersen, Factorization in C*-algebras, Expo. Math. 16(1998), 145-156. [Pie06] F. Pierrot, Op´erateurs r´eguliers dans les C ∗-modules et structure des C ∗-alg`ebres de groups de Lie semisimples complexes simplement connexes, J. Lie Theory 16 (2006), 651- 689. [R74] M. A. Rieffel, Induced representations of C ∗-algebras, Adv. Math. 13(1974), 176-257. [RR85] M. Rosenblum, J. Rovnyak, Hardy classes and operator theory, Oxford Univ. Press, 1985. [Sar94] D. Sarason, Sub-Hardy Hilbert Spaces in the Unit Disc, John Wiley Sons, Inc., New York, 1994. [Sar08] D. Sarason, Unbounded Toeplitz Operators, Integr. Equ. Oper. Theory 61 (2008), 281- 298. [Sch05] K. Schmudgen, Unbounded operators affiliated with C ∗-algebras, Preprint, Leipzig, 2005. [Sch10] K. Schmudgen, Algebras of fractions and strict Positivstellensatze for ∗-algebras, J. reine angew. Math. 647(2010), 57-88. [W91] S. L. Woronowicz, Unbounded elements affiliated with C ∗-algebras and non-compact quan- tum groups, Commun. Math. Phys. 136(1991), 399-432. [WN92] S. L. Woronowicz, K. Napi´orkowski, Operator theory in the C ∗-algebra framework, Re- ports Math. Phys. 31(1992), 353-371. K. Schmudgen; Universitat Leipzig, Mathematisches Institut, Augustusplatz 10/11, D-04109 Leipzig, Germany; E-Mail: [email protected] UNBOUNDED OPERATORS ON HILBERT C ∗-MODULES 39 R. Gebhardt; Max Planck Institute for Mathematics in the Sciences, Inselstrasse 22, D-04103 Leipzig, Germany; E-Mail: [email protected]
1605.04900	2	1605	2016-06-06T18:08:06	C*-algebras have a quantitative version of Pelczynski's property (V)	[ "math.OA", "math.FA" ]	A Banach space X has Pelczynski's property (V) if for every Banach space Y every unconditionally converging operator T: X -> Y is weakly compact. H. Pfitzner proved that C*-algebras have Pelczynski's property (V). In the preprint "H. Krulisova: Quantification of Pelczynski's property (V)" the author explores possible quantifications of the property (V) and shows that C(K) spaces for a compact Hausdorff space K enjoy a quantitative version of the property (V). In this paper we generalize this result by quantifying Pfitzner's theorem. Moreover, we prove that in dual Banach spaces a quantitative version of the property (V) implies a quantitative version of the Grothendieck property.	math.OA	math	C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) HANA KRULISOV ´A Abstract. A Banach space X has Pe lczy´nski's property (V) if for every Ba- nach space Y every unconditionally converging operator T : X → Y is weakly compact. H. Pfitzner proved that C ∗-algebras have Pe lczy´nski's property (V). In the preprint [8] the author explores possible quantifications of the prop- erty (V) and shows that C(K) spaces for a compact Hausdorff space K enjoy a quantitative version of the property (V). In this paper we generalize this result by quantifying Pfitzner's theorem. Moreover, we prove that in dual Ba- nach spaces a quantitative version of the property (V) implies a quantitative version of the Grothendieck property. 1. Introduction In 1994, H. Pfitzner proved that C∗-algebras have Pe lczy´nski's property (V) (see [10]). The aim of this paper is to prove a quantitative version of Pfitzner's result. In this way we continue the study of quantitative versions of Pe lczy´nski's property (V) presented in the preprint [8]. Section 2 summarizes all essential definitions and basic facts contained mostly in the preprint [8]. In Section 3 we slightly improve Behrends's quantitative version of Rosenthal's ℓ1 -- theorem [2, Section 3], which we use to prove the main theorem in Section 4. Section 5 is devoted to the relationship of quantitative versions of Pe lczy´nski's property (V) and the Grothendieck property in dual Banach spaces. 2. Preliminaries We follow the notation of [8] with one exception. Because we deal also with C∗-algebras, we write X ′ (instead of X ∗) for a dual to a Banach space X, since the ∗ in C∗-algebras is already reserved for the involution. All Banach spaces are considered either real or complex, unless stated otherwise. The closed unit ball of a Banach space X is denoted by BX . 2.1. Pe lczy´nski's property (V) and its quantification. Let us recall some essential definitions and facts (explained in more detail in [8] with many comments). 2010 Mathematics Subject Classification. 46B04,46L05,47B10. Key words and phrases. Pe lczy´nski's property (V); C ∗-algebra; Grothendieck property. The research was supported by the Grant No. 142213/B-MAT/MFF of the Grant Agency of the Charles University in Prague and by the Research grant GA CR P201/12/0290. 1 A series P∞ n=1 xn in a Banach space X is said to be is a bounded sequence of scalars, • unconditionally convergent if the seriesP∞ • weakly unconditionally Cauchy (wuC ) if P∞ n=1 x′(xn) converges. n=1 tnxn converges whenever (tn) for all x′ ∈ X ′ the series 2 HANA KRULISOV ´A A bounded linear operator T : X → Y between Banach spaces X and Y is called unconditionally converging if Pn T xn is an unconditionally convergent series in Y whenever Pn xn is a weakly unconditionally Cauchy series in X. It is not difficult to show that T is unconditionally converging if and only if for every series Pn xn in X with ∞ sup x′∈BX ′ Xn=1 x′(xn) < ∞ the series Pn T xn converges. We say that a Banach space X has Pe lczy´nski's property (V) if for every Banach space Y every unconditionally converging operator T : X → Y is weakly compact. To quantify the property (V) means to replace the implication (1) T is unconditionally converging ⇒ T is weakly compact by an inequality measure of weak non-compactness of T ≤ C · measure of T not being unconditionally converging, where C is some positive constant depending only on X, and the two measures are positive numbers for each operator T and are equal to zero if and only if T is weakly compact or unconditionally converging, respectively. This inequality is a strengthening of the original implication (1). For this purpose we use the following quantities. For a bounded sequence (xn) in a Banach space X we define ca(cid:0)(xn)(cid:1) = inf n∈N sup{kxk − xlk : k, l ∈ N, k, l ≥ n}. It is a measure of non-Cauchyness of a sequence (xn), hence in Banach spaces it measures non-convergence. Let T : X → Y be a bounded linear operator between Banach spaces X and Y . We set uc(T ) = sup(ca (cid:18) n Xi=1 T xi(cid:19)n! : (xn) ⊂ X, sup x′∈BX ′ x′(xn) ≤ 1) . ∞ Xn=1 Then uc(T ) measures how far is the operator T from being unconditionally con- verging. Let A be a bounded subset of a Banach space X. The de Blasi measure of weak non-compactness of the set A is defined by ω(A) = inf{d(A, K) : ∅ 6= K ⊂ X is weakly compact}, where d(A, K) = sup{dist(a, K) : a ∈ A}. De Blasi has proved that ω(A) = 0 if and only if A is relatively weakly compact (see [4]). Other quantities which measure relative weak non-compactness are for example γ(A) = sup{ lim n lim m x′ m(xn) − lim m lim n x′ m(xn) : (xn) is a sequence in A, (x′ m) is a sequence in BX ′ , and the limits exist} or wckX (A) = sup{dist(clust(X ′′,w∗)(xn), X) : (xn) is a sequence in A}, C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) 3 where clust(X ′′,w∗)(xn) stands for the set of all w∗-cluster points of the sequence (xn) in X ′′. The quantities γ(A) and wckX (A) are equivalent by [1, Theorem 2.3] in the following sense: (2) wckX (A) ≤ γ(A) ≤ 2 wckX (A). However, the quantity ω(A) is not equivalent to the other two (see [1, Corollary 3.4]). We have only (3) by [1, Theorem 2.3]. wckX (A) ≤ ω(A) For measuring weak non-compactness of a bounded linear operator T : X → Y between Banach spaces X and Y we use the quantities ω(T (BX)), γ(T (BX)), and wckY (T (BX )), which we denote simply by ω(T ), γ(T ), and wkY (T ). We say that a Banach space X has a quantitative version of Pe lczy´nski's property (V) -- we denote it by (Vq) -- if there is a constant C > 0 such that for every Banach space Y and every operator T : X → Y (4) γ(T ) ≤ C · uc(T ). If it is possible to replace γ(T ) in (4) with ω(T ), we say that X has the property (Vq)ω. If γ(T ) in (4) is replaced by ω(T ′), where T ′ : Y ′ → X ′ denotes the dual operator to T , we say that X has the property (Vq)∗ ω. In [8, Proposition 3.2] it is proved that a Banach space X has the property (Vq) if and only if there exists a constant C > 0 such that for each bounded subset K of the dual space X ′ where γ(K) ≤ C · η(K), ∞ η(K) = supn lim sup n x′∈K x′(xn) : (xn) ⊂ X, sup sup x′∈BX ′ Xn=1 x′(xn) ≤ 1o. Using the above-described characterization we will prove in Section 4 that C∗-algebras have the property (Vq). Note that the quantity η is translation-invariant, that is, η(K) = η(K + z′), K ⊂ X ′, z′ ∈ X ′. (5) This follows from the fact that (xn) weakly null whenever P xn is a wuC series in X. 2.2. Measures of weak and weak∗ non-Cauchyness of sequences in Banach spaces. In sections 4 and 5 we will use the following standard quantities, analogous to the quantity ca, which measure how far is a sequence in a (dual) Banach space from being weakly (weak∗) Cauchy. Let X be a Banach space and let (xn) be a bounded sequence X. We set δ(xn) = sup x′∈BX ′ lim n→∞ k,l≥nx′(x′ sup k) − x′(x′ l). This quantity is a measure of weak non-Cauchyness of the sequence (xn). Further- more, let us set δ(xn) = inf {δ(xnk ) : (xnk ) is a subsequence of (xn)} . It measures how close can subsequences of (xn) be to be weakly Cauchy. 4 HANA KRULISOV ´A If (x′ n) is a bounded sequence in X ′, we set k,l≥n x′ sup n) = sup x∈BX δw∗(x′ lim n→∞ k(x) − x′ l(x). The last quantity is a measure of weak∗ non-Cauchyness of the sequence (x′ n). The quantity δ(xn) equals 0 if and only if the sequence (xn) is weakly Cauchy. n) is weak∗ Cauchy. If δ(xn) = 0, it it Analogously, δw∗ (x′ not clear whether (xn) admits a weakly Cauchy subsequence. n) = 0 if and only if (x′ 2.3. Selfadjoint elements and selfadjoint functionals. Let A be a C∗-algebra. Let us denote by Asa the selfadjoint elements of A, that is Asa = {a ∈ A : a = a∗}. If f is a bounded linear Then Asa is a real Banach space and A = Asa + iAsa. functional on A, f ∗ is the functional defined by f ∗(x) = f (x∗), x ∈ A. Let (A′)sa denote the set {f ∈ A′ : f = f ∗} of selfadjoint functionals on A. Then (A′)sa is a real Banach space, and is isometrically isomorphic to (Asa)′. We will write A′ sa for both these spaces. Every functional x′ ∈ A′ can be decomposed as x′ = f + ig where f, g ∈ A′ sa. It suffices to set f = (x′ + (x′)∗)/2, g = (x′ − (x′)∗)/(2i). 3. A quantitative version of Rosenthal's ℓ1 -- theorem For proving the main result we need the quantitative version of Rosenthal's ℓ1 -- theorem proved by E. Behrends in [2, Section 3]. In this section we revise his theorem, because it turns out that one of the estimates there can be easily improved. We will then use this improved version. Let us remind Behrends's definition [2, 3.1]. Definition. Let (xn) be a bounded sequence in a Banach space X and ε > 0. We say that (xn) admits ε -- ℓ1 -- blocks if for every infinite M ⊂ N there are scalars a1, . . . , ar with Par = 1 and i1, . . . , ir in M such that (cid:13)(cid:13)P aρxiρ(cid:13)(cid:13) ≤ ε. The revised version of the quantitative Rosenthal's ℓ1 -- theorem for complex Ba- nach spaces is the following. Theorem 3.1. Let X be a complex Banach space X and ε > 0. Let (xn) be a sequence in X which admits ε -- ℓ1 -- blocks. Then there is a subsequence (xnk ) of (xn) such that for every x′ ∈ X ′ with kx′k = 1 the diameter of the set of cluster points of the sequence (x′(xnk ))k is at most πε. Remark. In the original Behrends' theorem [2, Theorem 3.3] there is a larger constant 8/√2 in place of π. A similar result with the better constant π has been obtained (in a different way) by I. Gasparis [5]. Sketch of the proof of Theorem 3.1. The proof is essentially the same as the original one. Suppose that the conclusion were not true. We can find δ > 0 such that the number sup x′∈SX ′ {diameter of the set of accumulation points of (x′(xnk ))k} is greater than πε + δ for any subsequence (xnk ) of (xn). Fix τ ∈ (0, 1) such that (2 + supn kxnk)τ < δ π . Similarly to the one in the proof of [2, Theorem 3.3 (or 3.2)] we can prove the following lemma. C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) 5 Lemma. The sequence (xn) admits a subsequence (without loss of generality still denoted by (xn)) which satisfies the following conditions: (i) Whenever C and D are disjoint finite subsets of N, there are z0, w0 ∈ C with w0 ≥ πε + δ and x′ ∈ X ′ with kx′k = 1 such that x′(xn)− z0 ≤ τ for n ∈ C and x′(xn) − (z0 + w0) ≤ τ for n ∈ D. (ii) There are i1 < ··· < ir in N and a1, . . . , ar ∈ C which satisfy Finally, the time has come for the modification. By [11, Lemma 6.3] we find Set C = {1, . . . , r} \ D. For these sets C and D we find z0, w0, and x′ from (i) of the lemma. It follows that r Xρ=1 D ⊂ {1, . . . , r} such that aρ = 1, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xρ∈D ε ≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) aρxiρ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≥(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρx′(xiρ )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xρ=1 Xρ=1 ≥(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρ(z0 + w0)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xρ∈C aρz0 + Xρ∈D ≥ w0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) − z0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xρ=1 Xρ∈D δ π − (1 + z0)τ ≥ ε + = ε + r r r r Xρ=1 aρ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ ε. r Xρ=1 aρxiρ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ τ, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xρ=1 r 1 π ≥ aρ = 1 π . =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) r aρx′(xiρ ) + Xρ∈D Xρ∈C Xρ=1 − τ ≥ w0 π − z0τ − τ ≥ aρ =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xρ∈D − τ r aρx′(xiρ )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aρz0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) πε + δ − τ Xρ=1 π − (1 + z0)τ aρw0 + δ π − (2 + sup n kxnk)τ > ε, which is a contradiction. (cid:3) 4. Main theorem This section is devoted to our main result -- a quantitative version of Pfitzner's theorem (Theorem 4.1 below). We also prove a "real version" of this theorem (Theorem 4.2). Theorem 4.1. Let A be a C∗-algebra. Then for every bounded K ⊂ A′ (6) wckA′(K) ≤ π · η(K). Therefore A has the property (Vq). Proof. The quantities γ(K) and wckA′ (K) are equivalent by [1, Theorem 2.3], more specifically, the inequality (6) implies γ(K) ≤ 2π · η(K). If this holds for each bounded K ⊂ A′, Proposition [8, 3.2] mentioned also in Section 2 gives that A has the property (Vq). Let us show the inequality (6). Let K ⊂ A′ be bounded. The case wckA′ (K) = 0 is trivial. Suppose that wckA′(K) > 0 and fix an arbitrary λ ∈ (0, wckA′(K)). By the definition of the quantity wckA′(K) we find a sequence (x′ n) in K such that dist(cid:0)clust(A′′′,w∗)(x′ n), A′(cid:1) > λ. 6 HANA KRULISOV ´A Since every dual of a C∗-algebra is a predual of a von Neumann algebra, we deduce from [13, Theorem III.2.14] (see also [6, Example IV.1.1(b)]) that A′ is L-embedded -- it means that A′ is complemented in A′′′ by a projection P satisfying kx′′′k = kP x′′′k + kx′′′ − P x′′′k, x′′′ ∈ A′′′. Consequently, from [7, Theorem 1] we have δ(x′ nk ) : (x′ n) = inf{δ(x′ ≥ 2 dist(cid:0)clust(A′′′,w∗)(x′ n), A′(cid:1) > 2λ. nk ) is a subsequence of (x′ k)} Fix an arbitrary ε > 0. We now prove the following claim. Claim. There is a sequence of self-adjoint elements (xk) in BA satisfying xixj = 0, i, j ∈ N, i 6= j, and a subsequence (x′ nk ) of the sequence (x′ n) such that nk (xk)(cid:12)(cid:12) > (1 − ε)2 λ (cid:12)(cid:12)x′ π , k ∈ N. n is canonically decomposed in the following way: x′ Proof. Each x′ where fn, gn ∈ A′ are selfadjoint functionals. It suffices to find (xk) and (x′ that n = fn + ign, nk ) such fnk (xk) > (1 − ε)2 λ π or gnk (xk) > (1 − ε)2 λ π . Indeed, since selfadjoint functionals attain real values on selfadjoint elements of A, we have nk (xk)(cid:12)(cid:12) = fnk (xk) + ignk (xk) ≥((cid:12)(cid:12) Re(fnk (xk) + ignk(xk))(cid:12)(cid:12) = fnk (xk) (cid:12)(cid:12)x′ (cid:12)(cid:12) Im(fnk (xk) + ignk (xk))(cid:12)(cid:12) = gnk (xk) We begin by proving that there is a strictly increasing sequence of indices (nk) such that δ(fnk ) > λ or δ(gnk ) > λ. If δ(fn) > λ, the proof is over, so suppose that n) > 2λ + 2τ . By the definition of δ(fn) δ(fn) ≤ λ. Let us find τ > 0 satisfying δ(x′ there is a subsequence (fnk ) of the sequence (fn) with δ(fnk ) < λ + τ . We claim that the corresponding subsequence (gnk ) of (gn) satisfies δ(gnk ) > λ. To obtain a contradiction, suppose that δ(gnk ) ≤ λ. Using the definition of δ(gnk ) we find a strictly increasing sequence of indices (kl) such that δ(gnkl ) < λ + τ . Then . δ(x′ nkl ) = δ(fnkl = sup + ignkl ) x′′∈BA′′ x′′∈BA′′ ≤ sup ≤ sup x′′∈BA′′ lim l→∞ lim l→∞ lim l→∞ + sup x′′∈BA′′ sup sup p,q≥l(cid:12)(cid:12)x′′(fnkp + ignkp ) − x′′(fnkq + ignkq )(cid:12)(cid:12) p,q≥l(cid:16)(cid:12)(cid:12)x′′(fnkp ) − x′′(fnkq )(cid:12)(cid:12) +(cid:12)(cid:12)x′′(gnkp ) − x′′(gnkq )(cid:12)(cid:12)(cid:17) p,q≥l(cid:12)(cid:12)x′′(fnkp ) − x′′(fnkq )(cid:12)(cid:12) p,q≥l(cid:12)(cid:12)x′′(gnkp ) − x′′(gnkq )(cid:12)(cid:12) ) < λ + τ + λ + τ = 2λ + 2τ, sup sup lim l→∞ = δ(fnkl ) + δ(gnkl which contradicts the fact that δ(x′ n) > 2λ + 2τ . Without loss of generality we may assume that we have found a subsequence (fnk ) of the sequence (fn) with δ(fnk ) > λ and such that (fnk ) = (fn). By passing C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) 7 to a further subsequence we can also ensure that inf n∈N kfnk supn∈N kfnk > 1 − ε. Indeed, the sequence (fn) is bounded, hence we can find its subsequence (fnk ) such that the limk→∞ kfnkk exists. This limit is nonzero, because otherwise we would have δ(fn) = 0. We thus obtain the desired subsequence by omitting finitely many members of (fnk ). The inequality δ(fn) > λ says that for every subsequence (fnk ) of (fn) there is some x′′ ∈ A′′ with kx′′k = 1 such that the diameter of the set of accumulation points of the sequence (x′′(fnk ))k is greater than λ. By Theorem 3.1 the sequence (fn) does not admit λ π -- ℓ1 -- blocks, i.e. there is an infinite M ⊂ N such that whenever i=1 ai = 1, and n1 < ··· < nr are indices in M , we have π . Hence there is a subsequence (fnk ) of (fn) such that for each a1, . . . , ar ∈ C satisfy Pr (cid:13)(cid:13)Pr i=1 aifni(cid:13)(cid:13) > λ nonzero (αk) ∈ ℓ1 and N ∈ N large enough N ∞ > λ π (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 By letting N → ∞ we obtain λ π fnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk PN k=1 αk αk ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αkfnk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 Xk=1 Therefore we have for each (αk) ∈ ℓ1 kfnkk ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) λ Xk=1 Xk=1 π π supk∈N kfnkk αk ≤ Xk=1 Let us set αk λ ∞ ∞ ∞ ∞ . . ≤ ∞ Xk=1 αk. αk fnk kfnkk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) r = λ π supk∈N kfnkk and θ = (1 − ε) r inf k∈Nkfnkk. Then λ π λ π θ = (1 − ε) inf k∈N kfnkk supk∈N kfnkk ≥ (1 − ε) inf n∈N kfnk supn∈N kfnk ≥ (1 − ε)2 λ Without loss of generality we can assume that (fnk ) = (fn). Then(cid:0) fn (ak) ∈ ℓ1, sequence consisting of selfadjoint elements which satisfies αk, ≤ αk π ∞ ∞ ∞ r . kfnk(cid:1)n is a basic Xk=1 αk ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 fk kfkk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 that is (36) of [10] (where a′ k = fk). By Pfitzner's proof of [10, Theorem 1] we obtain a sequence (xk) in A and a subsequence (fnk ) of (fn) for which (35) of [10] is valid (where a′ nk = fnk ), i.e. xk are selfadjoint elements in BA such that xixj = 0, i, j ∈ N, i 6= j, and fnk (xk) > θ ≥ (1 − ε)2 λ π , k ∈ N. This completes the proof of the claim. 8 HANA KRULISOV ´A Let (xk) and (x′ nk ) be sequences obtained by the claim. Since x′ nk (xk) > (1 − ε)2 λ π , k ∈ N, we have lim sup k→∞ sup x′∈K x′(xk) ≥ (1 − ε)2 λ π . N N N N Xk=1 µ(xk) = But P xk is a wuC series in A satisfying supx′∈BA′ P x′(xk) ≤ 1. Indeed, all xk are contained in a commutative subalgebra B of A, which can be identified with the space C0(Ω) for some Ω by the Gelfand representatiton. Then xk, k ∈ N, are real continuous functions on Ω with kxkk = supξ∈Ω xk(ξ) ≤ 1 and {xi 6= 0} ∩ {xj 6= 0} = ∅, i 6= j. Let x′ ∈ A′, and let us set µ = x′ ↾B∈ B′ = C0(Ω)′ = M(Ω). For each N ∈ N we get ZΩ Xk=1 x′(xk) = ≤ZΩ 1 dµ = kµk ≤ kx′k. Therefore supx′∈BA′ P∞ k=1 x′(xk) ≤ 1. We thus obtain η(K) ≥ (1 − ε)2 λ arbitrarily, it follows that η(K) ≥ 1 Remark. It is not clear whether C∗-algebras have also the property (Vq)∗ ω. From [8, Theorem 4.1] it follows that the answer is affirmative for commutative C∗-algebras. In fact we do not know any example of a Banach space with the prop- erty (Vq) but not (Vq)∗ ω. Regarding the property (Vq)ω, we know from [8, Propo- sition 4.3] that some (commutative) C∗-algebras enjoy this property and some do not. π . Since ε > 0 and λ < wckA′ (K) were chosen π wckA′(K), which completes the proof. (cid:3) Xk=1Z{xk6=0} xk dµ xk dµ(cid:12)(cid:12)(cid:12)(cid:12) Xk=1(cid:12)(cid:12)(cid:12)(cid:12) ≤ The following theorem is a kind of "real version" of Theorem 4.1. Theorem 4.2. Let A be a C∗-algebra. Then the space Asa has the property (Vq), more precisely, for every bounded K ⊂ A′ (7) sa wckA′(K) ≤ η(K). (0, wckA′ sa (K)). We find (fn) in K such that Proof. The proof is analogous to the previous one, it suffices to use real versions of the theorems that have allowed us to prove Theorem 4.1. Let us sketch it briefly. sa(K) > 0 and an arbitrary λ ∈ sa(cid:1) > λ. Consider a bounded set K ⊂ A′ dist(cid:0)clust((A′ Since A′ is L-embedded, the real version of A′ (let us denote it by (A′)R) is also L-embedded. But (A′)sa is a 1-complemented subspace of (A′)R and is therefore L-embedded by [6, Proposition IV.1.5]. We thus get sa)′′,w∗)(fn), A′ sa with wckA′ from [7, Theorem 1]. Let us fix ε > 0. By passing to a subsequence we arrange that δ(fn) > 2λ inf n∈N kfnk supn∈N kfnk > 1 − ε. C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) 9 By the real version of the quantitative Rosenthal's ℓ1 -- theorem [2, Theorem 3.2] the sequence (fn) admits λ -- ℓ1 -- blocks, which yields a subsequence (fnk ) of the sequence (fn) that for every (αn) ∈ ℓ1 satisfies λ supk∈N kfnkk ∞ Xk=1 ∞ αk ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 αk fnk kfnkk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ ∞ Xk=1 αk. Then we proceed exactly as in the proof of Theorem 4.1 to obtain the desired conclusion. (cid:3) 5. Relation to the Grothendieck property Let us remind that a Banach space X has the Grothendieck property if every weak∗ convergent sequence in its dual is weakly convergent. It is well known that for dual Banach spaces the property (V) implies the Grothendieck property. In this section we prove that this implication holds even for suitable quantitative versions of these properties. One possible quantification of the Grothendieck property has already been stud- ied in [3] and [9]. Let us remind the definition: Let c > 0. A Banach space X is c-Grothendieck if (8) whenever (x′ n) is a bounded sequence in X ′. δ(x′ n) ≤ c · δw∗(x′ n) A Banach space X has the Grothendieck property if and only if for every sequence n) in X ′ the following implication holds: (x′ n) is weak∗ Cauchy ⇒ (x′ (x′ n) is weakly Cauchy. The inequality (8) quantifies this implication. But we can look at the Grothendieck property also in another way: X has the Grothendieck property if and only if every sequence (x′ n) in X ′ satisfies the implication If we replace this implication by an inequality n) is weak∗ Cauchy ⇒ {x′ (x′ wckX ′(cid:0){x′ n : n ∈ N} is a relatively weakly compact set. n : n ∈ N}(cid:1) ≤ c · δw∗(x′ n) where c > 0 is some constant not depending on (x′ n), we obtain another quantitative version of the Grothendieck property. We will prove that all dual Banach spaces with the property (Vq) have this kind of quantitative Grothendieck property (see Corollary 5.2). We do not know whether the latter quantitative Grothendieck property implies the former one (with a larger constant). Theorem 5.1. Let X be a Banach space. Then for every bounded sequence (x′′ n) in X ′′ 2 δw∗ (x′′ n). Proof. Let (x′′ n : n ∈ N}(cid:1) ≤ 1 η(cid:0){x′′ n) be a bounded sequence in X ′′. The case η(cid:0){x′′ n : n ∈ N}(cid:1) = 0 is n : n ∈ N}(cid:1) > 0 and fix δ ∈(cid:0)0, η(cid:0){x′′ n : n ∈ N}(cid:1)(cid:1). Let us n : n ∈ N}(cid:1) > δ + ε. By the definition of the quantity η we k in X ′ with supx′′∈BX ′′ P∞ k=1 x′ k=1 x′′(x′ k) ≤ 1 such that k) > δ + ε. Since (x′ k) is a weakly null sequence, there are n) and (y′ n) > n) is a bounded sequence in X ′′, k) which for all n ∈ N satisfy y′′ trivial. Suppose that η(cid:0){x′′ find ε > 0 such that η(cid:0){x′′ can find a wuC series P∞ lim supk→∞ supn∈N x′′ subsequences of (y′′ δ + ε. The sequence (y′ n) is weakly null in X ′ and (y′′ n(x′ n) of (x′′ k) of (x′ n(y′ 10 HANA KRULISOV ´A hence by Simons' extraction lemma [12, Theorem 1] there is a strictly increasing sequence of indices (nk) such that for all k ∈ N nk (y′ y′′ nm ) < ε. Xm∈N m6=k Let us define αk =((−1)k sgn−1(cid:0)y′′ 0, nk (y′ nk )(cid:1), nk (y′ y′′ nk (y′ y′′ nk ) 6= 0, nk ) = 0, k ∈ N, where sgn denotes the complex signum function, i.e. sgn(z) = z z , z ∈ C \ {0}. Set x′ = w∗- lim N→∞ N Xk=1 αky′ nk ∈ X ′. ∞ Then x′ ∈ BX ′ because for all x ∈ BX αkz′ ∞ Xk=1 x′(x) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) For each k ∈ N even nk (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∞ ≤ Re y′′ nk (x′) = αky′′ nk (y′ z′ nk (x) ≤ Xk=1 Xn=1 nk ) + Re Xm∈N m6=k/2 x′ n(x) ≤ sup x′′∈BX ′′ ∞ Xn=1 x′′(x′ n) ≤ 1. y′′ nk (α2my′ n2m)! y′′ nk (α2m−1y′ nk (y′ y′′ nk (y′ ≥ y′′ − Re Xm∈N nk ) − Xm∈N nk ) − Xm∈N y′′ nk (y′ > (δ + ε) − ε = δ. = y′′ nk (y′ m6=k/2 m6=k n2m−1 )! n2m ) − Xm∈N nm) nk (y′ y′′ n2m−1 ) Analogously, for each k ∈ N odd nk (x′) = αky′′ Re y′′ nk (y′ nk ) + Re Xm∈N y′′ nk (α2my′ y′′ nk (α2m−1y′ n2m)! n2m−1)! m6=(k+1)/2 − Re Xm∈N nk ) + Xm∈N < −(δ + ε) + ε = −δ. ≤ −y′′ nk (y′ m6=k y′′ nk (y′ nm ) Therefore inf n∈N k,l≥ny′′ sup nk (x′) − y′′ nl(x′) ≥ inf n∈N sup k,l≥n(cid:12)(cid:12) Re(cid:0)y′′ nk (x′) − y′′ nl(x′)(cid:1)(cid:12)(cid:12) ≥ 2δ. C-ALGEBRAS HAVE A QUANTITATIVE VERSION OF PE LCZY ´NSKI'S PROPERTY (V) 11 It follows that δw∗ (x′′ arbitrarily, we obtain the desired inequality. n) ≥ δw∗ (y′′ nk ) ≥ 2δ. Since δ < η(cid:0){x′′ n : n ∈ N}(cid:1) was chosen (cid:3) Corollary 5.2. Let X be a Banach space and C > 0. Suppose that each bounded K ⊂ X ′′ satisfy (9) (i.e. X ′ enjoys the property (Vq)). Then for every bounded sequence (x′′ n) in X ′′ wckX ′′ (K) ≤ C · η(K) n : n ∈ N}(cid:1) ≤ 1 wckX ′′(cid:0){x′′ 2 C · δw∗(x′′ n). n : n ∈ N}. Proof. It suffices to combine the previous theorem with the inequality (9) applied to K = {x′′ (cid:3) Corollary 5.3. Let A be a von Neumann algebra. Then A has a quantitative version of the Grothendieck property -- more precisely, for every bounded sequence (x′ n) in A′ Proof. Since every von Neumann algebra is a C∗-algebra and a dual Banach space, the assertion follows from Theorem 4.1 and Corollary 5.2. (cid:3) wckA′(cid:0){x′ n : n ∈ N}(cid:1) ≤ 1 2 π δw∗ (x′ n). Acknowledgement The author wishes to express her gratitude to the referee, who suggested how to improve Theorem 5.1 and shorten its proof. She also acknowledges many sugges- tions and comments of her supervisor Ondrej Kalenda during the preparation of the paper. References [1] C. Angosto and B. Cascales. Measures of weak noncompactness in Banach spaces. Topology Appl., 156(7):1412 -- 1421, 2009. [2] Ehrhard Behrends. New proofs of Rosenthal's l1-theorem and the Josefson-Nissenzweig the- orem. Bull. Polish Acad. Sci. Math., 43(4):283 -- 295 (1996), 1995. [3] Hana Bendov´a. Quantitative Grothendieck property. J. Math. Anal. Appl., 412(2):1097 -- 1104, 2014. [4] Francesco S. De Blasi. On a property of the unit sphere in a Banach space. Bull. Math. Soc. Sci. Math. R. S. Roumanie (N.S.), 21(69)(3-4):259 -- 262, 1977. [5] I. Gasparis. ǫ-weak Cauchy sequences and a quantitative version of Rosenthal's ℓ1-theorem. J. Math. Anal. Appl., 434(2):1160 -- 1165, 2016. [6] P. Harmand, D. Werner, and W. Werner. M -ideals in Banach spaces and Banach algebras, volume 1547 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1993. [7] O. F. K. Kalenda, H. Pfitzner, and J. Spurn´y. On quantification of weak sequential complete- ness. J. Funct. Anal., 260(10):2986 -- 2996, 2011. [8] Hana Krulisov´a. Quantification of Pe lczy´nski's property (V). Preprint is available at http://arxiv.org/abs/1509.06610. [9] Jindrich Lechner. 1-Grothendieck C(K) spaces. Preprint is available at http://arxiv.org/abs/1511.02202. [10] H. Pfitzner. Weak compactness in the dual of a C ∗-algebra is determined commutatively. Math. Ann., 298(2):349 -- 371, 1994. [11] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third edition, 1987. [12] S. Simons. On the Dunford-Pettis property and Banach spaces that contain c0. Math. Ann., 216(3):225 -- 231, 1975. [13] M. Takesaki. Theory of operator algebras. I, volume 124 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin, 2002. Reprint of the first (1979) edition, Operator Algebras and Non-commutative Geometry, 5. 12 HANA KRULISOV ´A Department of Mathematical Analysis, Faculty of Mathematics and Physics, Charles University in Prague, Sokolovsk´a 83, 186 75, Praha 8, Czech Republic E-mail address: [email protected]
1108.2598	1	1108	2011-08-12T10:14:23	Traces on symmetrically normed operator ideals	[ "math.OA" ]	For every symmetrically normed ideal $\mathcal{E}$ of compact operators, we give a criterion for the existence of a continuous singular trace on $\mathcal{E}$. We also give a criterion for the existence of a continuous singular trace on $\mathcal{E}$ which respects Hardy-Littlewood majorization. We prove that the class of all continuous singular traces on $\mathcal{E}$ is strictly wider than the class of continuous singular traces which respect Hardy-Littlewood majorization. We establish a canonical bijection between the set of all traces on $\mathcal{E}$ and the set of all symmetric functionals on the corresponding sequence ideal. Similar results are also proved in the setting of semifinite von Neumann algebras.	math.OA	math	TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS F. SUKOCHEV AND D. ZANIN We dedicate this paper to the memory of Nigel Kalton whose influence on us both has been profound. Without his collaboration this paper would never have been written. Abstract. For every symmetrically normed ideal E of compact operators, we give a criterion for the existence of a continuous singular trace on E. We also give a criterion for the existence of a continuous singular trace on E which respects Hardy-Littlewood majorization. We prove that the class of all contin- uous singular traces on E is strictly wider than the class of continuous singular traces which respect Hardy-Littlewood majorization. We establish a canonical bijection between the set of all traces on E and the set of all symmetric func- tionals on the corresponding sequence ideal. Similar results are also proved in the setting of semifinite von Neumann algebras. 1. Introduction In his groundbreaking paper [6], J. Dixmier proved the existence of positive singular traces (that is, linear positive unitarily invariant functionals which vanish on all finite dimensional operators) on the algebra B(H) of all bounded linear operators acting on infinite-dimensional separable Hilbert space H. Namely, if ψ : R+ → R+ is a concave increasing function such that (1) lim t→∞ ψ(2t) ψ(t) = 1, then there is a singular trace τω, defined for every positive compact operator A ∈ B(H) by setting (2) τω(A) = ω( 1 ψ(n) n Xk=1 sk(A)). Here, {sk(A)}k∈N is the sequence of singular values of the compact operator A ∈ B(H) taken in the descending order and ω is an arbitrary dilation invariant gen- eralised limit on the algebra l∞ of all bounded sequences. This trace is finite on 0 ≤ A ∈ B(H) if and only if A belongs to the Marcinkiewicz ideal (see e.g. [14],[15],[27]) Mψ := {A ∈ B(H) : sup n∈N 1 ψ(n) n Xk=1 sk(A) < ∞}. 2000 Mathematics Subject Classification. 47L20, 47B10, 46L52. Key words and phrases. Symmetric functionals, singular traces. 1 2 F. SUKOCHEV AND D. ZANIN In [18], Dixmier's result was extended to an arbitrary Marcinkiewicz ideal Mψ with the following condition on ψ (3) lim inf t→∞ ψ(2t) ψ(t) = 1. All the traces defined above by formula (2) vanish on the ideal L1 consisting of all compact operators A ∈ B(H) such that P∞ An ideal E of algebra B(H) is said to be symmetrically normed if {sk(B)}k∈N ≤ {sk(A)}k∈N and A ∈ E implies that kBkE ≤ kAkE (see [14], [15], [29]1, [28], [20]). Since the ideal Mψ is just a special example of symmetrically normed operator ideal, the following question (suggested in [18], [16], [17], [7]) arises naturally. k=1 sk(A) < ∞. Question 1. Which symmetrically normed operator ideals admit a nontrivial sin- gular trace2? In analyzing Dixmier's proof of the linearity of τω given by (1), it was observed in [18] (see also [3]) that τω possesses the following fundamental property, namely if 0 ≤ A, B ∈ Mψ are such that (4) sk(B) ≤ sk(A), ∀n ∈ N, n Xk=1 n Xk=1 then τω(B) ≤ τω(A). Such a class of traces was termed "fully symmetric"in [20], [30] (see also earlier papers [8],[25], where the term "symmetric"was used). It is natural to consider such traces only on fully symmetrically normed operator ideals E (that is, on symmetrically normed operator ideals E satisfying the condition: if A, B satisfy (4) and A ∈ E, then B ∈ E and kBkE ≤ kAkE). In fact, it was established in [8] that every Marcinkiewicz ideal Mψ with ψ satisfying the condition (3) possesses fully symmetric traces. Furthermore, in the recent paper [18], the following unexpected result was established. If ψ satisfies the condition (3), then every fully symmetric trace on Mψ is a Dixmier trace τω for some ω. The following question ( also suggested in [18], [7], [16], [17]) arises naturally. Question 2. Which fully symmetrically normed operator ideals admit a nontrivial singular trace which is fully symmetric? In papers [16],[17] the following two problems (closely related to Question 1 and Question 2) were also suggested. Question 3. Which fully symmetrically normed operator ideals admit a trace which is not fully symmetric? Let us fix an orthonormal basis {en}n∈N in H. An operator A ∈ B(H) is called diagonal if (Aen, em) = 0 for every n 6= m. Question 4. Let the mapping ϕ : E → C be unitarily invariant. Suppose that ϕ is linear on the subset of all diagonal operators from E. Does it imply that ϕ is a trace on E? 1 We have to caution the reader that in Theorem 1.16 of [29] the assertion (b) does not hold for the norm of an arbitrary symmetrically normed ideal E (see e.g. corresponding counterexamples in [19, p. 83]). 2 In this paper, we exclusively deal with positive traces TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 3 In some very special cases (for principal ideals contained in L1, which are, strictly speaking, not symmetrically normed ideals), Question 3 was answered in the affir- mative3 in [33]. In [20], question 3 was answered in the affirmative for the special case of Marcinkiewicz ideals under the assumption (1). It should be pointed out that the method used in [20] cannot be extended to an arbitrary Marcinkiewicz ideal Mψ and, furthermore, cannot be extended to a general symmetrically normed op- erator ideal. Question 4 was answered in [20] in full generality using deep results from [11, 10] (see also [9]). The following theorem is the main result of this paper. It yields answers to Questions 1 -- 3. In the course of the proof of Theorem 5, we also present a new (and very simple) proof answering Question 4. Prior to stating Theorem 5, we make a few preliminary observations, for which we are grateful to the referee. Any trace ϕ : E → C obeys the condition 1 m ϕ(A⊕m) = ϕ(A), A ∈ E, m ≥ 1. Here, the direct sum A⊕m is formed with respect to some arbitrary Hilbert space isomorphism H ⊕m ≃ H. Thus, traces are closely related to the following convex (see Lemma 11 below) functional on E. 1 m kA⊕mkE, A ∈ E. π : A → lim m→∞ The non-triviality of the functional π : E → R is an obvious necessary condition for the existence of a trace. Theorem 5. Let E be a symmetrically normed operator ideal. Consider the fol- lowing conditions. (1) There exist nontrivial singular traces on E. (2) There exist nontrivial singular traces on E, which are fully symmetric. (3) There exist nontrivial singular traces on E, which are not fully symmetric. (4) E 6= L1 and there exist an operator A ∈ E such that (5) lim m→∞ 1 m kA⊕mkE > 0. (i) The conditions (1) and (4) are equivalent for every symmetrically normed operator ideal E. (ii) The conditions (1), (2) and (4) are equivalent for every fully symmetrically normed operator ideal E. (iii) The conditions (1) − (4) are equivalent for every fully symmetrically normed operator ideal E equipped with a Fatou norm. Recall that the norm on a symmetrically normed operator ideal E is called a Fatou norm if the unit ball of E is closed with respect to strong (or, equivalently, weak) operator convergence. Observe that classical ideals (such as Schatten-von Neumann ideals Lp, Marcinkiewicz, Orlicz and Lorentz ideals [14], [15], [29]) have a Fatou norm. In fact, in some standard references on the subject (e.g. Simon's book [29]), the requirement that symmetrically normed operator ideal has a Fatou norm appears to be a part of the definition. Similarly, in the book [24], devoted to the study of symmetric4 function spaces (which are a commutative counterpart 3We are grateful to the referee for this remark. 4termed there "rearrangement invariant". 4 F. SUKOCHEV AND D. ZANIN of symmetrically normed operator ideals), an assumption that the norm is a Fatou norm is incorporated into the definition [24, p. 118]. The proof of Theorem 5 is given in Section 7. In fact, in this paper we will prove a more general result for symmetric spaces associated with semifinite von Neumann algebras. The precise statements are given in Section 4 (see Theorems 23, 28, 29), Section 5 (see Theorems 33, 35, 36) and Section 6 (see Theorems 47, 48). The appendix contains the proof of important technical results for which we were unable to find a suitable reference. We also present a new and short proof of the Figiel-Kalton theorem from [13]. Finally, we say a few words about our proof and its relation to the previous results in the literature. Our strategy is based on the approach from recent papers [30] and [21], where condition (5) was connected to the geometry of E (see also [2]). The condition (5) is easy to verify in concrete situations. For example, the following corollary of Theorem 5 strengthens the main result of [20] and complements earlier results of J. Varga [32]. Corollary 6. Every Marcinkiewicz ideal Mψ with ψ satisfying the condition (3) admits a trace which is not fully symmetric. Indeed, it is proved in [1, Proposition 2.3] that the condition (4) of Theorem 5 is equivalent to the condition (3) for the Marcinkiewicz ideal Mψ. Some examples of symmetrically normed operator ideals, which are not Marcinkiewicz ideals, pos- sessing symmetric traces were presented in [7]. These results are also an immediate corollary of Theorem 5. For completeness, we note that the assertion (ii) in Theorem 5 holds for a wider class of relatively fully symmetrically normed operator ideals. The latter class is defined as follows: if A, B ∈ E are such that (4) holds, then kBkE ≤ kAkE . It coincides with the class of all symmetrically normed subspaces of a fully symmetric operator ideal (see [19]) 2. Definitions and preliminaries The theory of singular traces on symmetric operator ideals rests on some classical analysis which we now review for completeness. As usual, L∞(0, ∞) is the set of all bounded Lebesgue measurable functions on the semi-axis equipped with the uniform norm. Given a function x ∈ L∞(0, ∞), one defines its decreasing rearrangement t → µ(t, x) by the formula (see e.g. [22]) µ(t, x) = inf{s ≥ 0 : m({x > s}) ≤ t}. Let H be a Hilbert space and let B(H) be the algebra of all bounded operators on H equipped with the uniform norm. Let M ⊂ B(H) be a semi-finite von Neumann algebra equipped with a fixed faithful and normal semi-finite trace τ. M is said to be atomic (see [31, Definition 5.9]) if every nonzero projection in M contains a nonzero minimal projection. M is said to be atomless if there is no minimal projections in M. For every A ∈ M, the generalised singular value function t → µ(t, A) is defined by the formula (see e.g. [12]) µ(t, A) = inf{kApk : τ (1 − p) ≤ t}. If, in particular, M = B(H), then µ(A) is a step function and, therefore, can be identified with the sequence of singular numbers of the operators A (the singular TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 5 values are the eigenvalues of the operator A = (A∗A)1/2 arranged with multiplicity in decreasing order). Equivalently, µ(A) can be defined in terms of the distribution function dA of A. That is, setting we obtain dA(s) = τ (EA(s, ∞)), s ≥ 0, µ(t, A) = inf{s ≥ 0 : dA(s) ≤ t}, t > 0. Here, EA denotes the spectral measure of the operator A. Using the Jordan decomposition, every operator A ∈ B(H) can be uniquely written as A = (ℜA)+ − (ℜA)− + i(ℑA)+ − i(ℑA)−. Here, ℜA := 1/2(A + A∗) (respectively, ℑA := 1/2i(A − A∗)) for any operator A ∈ B(H) and B+ = BEB(0, ∞) ( respectively, B = −BEB(−∞, 0)) for any self-adjoint operator B ∈ B(H). Recall that ℜA, ℑA ∈ M for every A ∈ M and B+, B− ∈ M for every self-adjoint B ∈ M. Further, we need to recall the important notion of Hardy -- Littlewood majoriza- tion. Let A, B ∈ (L1 + L∞)(M). The operator B is said to be majorized by A and written B ≺≺ A if and only if Z t 0 µ(s, B)ds ≤Z t 0 µ(s, A)ds, t ≥ 0. We have (see [12]) A + B ≺≺ µ(A) + µ(B) ≺≺ 2σ1/2µ(A + B) for every positive operators A, B ∈ (L1 + L∞)(M). If s > 0, the dilation operator σs is defined by setting (σs(x))(t) = x( t s ), t > 0 in the case of the semi-axis. In the case of the interval (0, 1), the operator σs is defined by (σsx)(t) =(x(t/s), 0, t ≤ min{1, s} s < t ≤ 1. Similarly, in the sequence case, we define an operator σn by setting and an operator σ1/2 by setting n times n times σn(a1, a2, · · · ) = (a1, · · · , a1 , a2, · · · , a2 , · · · ) {z } {z } σ1/2 : (a1, a2, a3, a4, · · · ) → ( a1 + a2 2 , a3 + a4 2 , · · · ). Definition 7. The Banach space E(M, τ ) ⊂ (L1 + L∞)(M) is said to be a sym- metric operator space if the following conditions hold. (1) Given A ∈ E(M, τ ) and B ∈ (L1 + L∞)(M) with µ(B) = µ(A), we have B ∈ E(M, τ ) and kBkE = kAkE. (2) Given 0 ≤ A ∈ E(M, τ ) and 0 ≤ B ∈ (L1 + L∞)(M) with B ≤ A, we have B ∈ E(M, τ ) and kBkE ≤ kAkE. 6 F. SUKOCHEV AND D. ZANIN The space E(M, τ ) is called fully symmetric if for every A ∈ E(M, τ ) and every B ∈ (L1 + L∞)(M) with B ≺≺ A, we have B ∈ E(M, τ ) and kBkE ≤ kAkE. The norm on a symmetric space E(M, τ ) is a Fatou norm if the unit ball of E(M, τ ) is closed with respect to strong (or, equivalently, weak) operator conver- gence. Every symmetric space equipped with a Fatou norm is necessarily fully symmetric. A linear functional ϕ : E(M, τ ) → C is said to be symmetric if ϕ(B) = ϕ(A) for every positive A, B ∈ E(M, τ ) such that µ(B) = µ(A). A linear functional ϕ : E(M, τ ) → C is said to be fully symmetric if ϕ(B) ≤ ϕ(A) for every positive A, B ∈ E(M, τ ) such that B ≺≺ A. Every fully symmetric functional is symmetric and bounded. The converse fails [20]. A functional ϕ : E(M, τ ) → C is called singular if ϕ = 0 on (L1 ∩ L∞)(M). If E(M, τ ) 6⊂ L1(M), then every symmetric functional is singular. If E = E(0, ∞) and if ϕ : E → R is a symmetric functional, then sϕ(x) = ϕ(σsx) for every x ∈ E. If E = E(0, 1) and if ϕ : E → R is a singular symmetric functional, then sϕ(x) = ϕ(σsx) for every x = µ(x) ∈ E. Let E be a fully symmetric Banach space either on the interval (0, 1) or on the semi-axis. We need the notion of an expectation operator (see [2]). Let A = {Ak} be a (finite or infinite) sequence of disjoint sets of finite measure and denote by A the collection of all such sequences. Denote by A∞ the complement of ∪kAk. The expectation operator E(·A) : L1 + L∞ → L1 + L∞ is defined by setting E(xA) =Xk 1 m(Ak) (ZAk x(s)ds)χAk . Note that we do not require A∞ to have finite measure. Every expectation operator is a contraction both in L1 and L∞. Therefore, E(xA) ≺≺ x, x ∈ L1 + L∞. It follows that E(·A) is also contraction in E. It will be convenient to introduce the following notation. If A is a discrete subset of the semi-axis (i.e. a subset without limit points inside (0, ∞)), then the elements of A ∪ {0} partition the semi-axis. This partition consists of a (finite or infinite) sequence of sets of finite measure. We identify this partition with the set A. Elements of A will be called nodes of the partition A. The corresponding averaging operator will be denoted by E(·A). Let E be a symmetric Banach space either on the interval (0, 1) or on the semi- axis. Define the sets DE = Lin({x ∈ E : x = µ(x)}) = {µ(a) − µ(b), a, b ∈ E}, ZE = Lin({x1 − x2 : 0 ≤ x1, x2 ∈ E, µ(x1) = µ(x2)}). Let C be a Hardy operator defined by setting (Cx)(t) = 1 t Z t 0 x(s)ds. The following theorem was proved in [13]. For convenience of the reader, we give a new and simple proof in the appendix. TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 7 Theorem 8. Let E be a symmetric space on the semi-axis and let x ∈ DE. We have x ∈ ZE if and only if Cx ∈ E. A similar assertion is also valid for the interval (0, 1) provided that R 1 0 x(s)ds = 0. The following uniform submajorization was introduced by Kalton and Sukochev in [19]. Let x, y ∈ L1(0, 1) (or x, y ∈ (L1 + L∞)(0, ∞)). We say that y ⊳ x if there exists m ∈ N such that (6) Z b ma µ(s, y)ds ≤Z b a µ(s, x)ds, ∀ma ≤ b. Let x, y ∈ l∞. We say that y ⊳ x if there exists m ∈ N such that (7) b Xk=ma+1 µ(k, y) ≤ b Xk=a+1 µ(k, x) ∀ma + 1 ≤ b. The following important theorem was proved in [19] (see Theorem 5.4 and The- orem 6.3 there). Theorem 9. Let x, y ∈ L1(0, 1) or x, y ∈ (L1 + L∞)(0, ∞) or x, y ∈ l∞ be such that y ⊳ x. For every ε > 0, the function (1 − ε)y belongs to a convex hull of the set {z : µ(z) ≤ µ(x)}. This theorem led to the following fundamental result (see [19]). Theorem 10. Let E = E(0, 1) (or E = E(0, ∞) or E = E(N)) be a symmetric Banach space either on the interval (0, 1) or on the semi-axis or on N. It follows that the corresponding set E(M, τ ) is a symmetric Banach space. Also, the uniform submajorization permits us to prove the convexity of the functional π : E → R defined in Section 1. Lemma 11. The functional π : E → R is convex on every symmetrically normed operator ideal E. Proof. Let E be the corresponding symmetrically normed ideal of l∞. For every A, B ∈ E, it follows from Proposition 8.6 of [19] that µ(A + B) ⊳ µ(A) + µ(B). Hence, σmµ(A + B) ⊳ σm(µ(A) + µ(B)). By Theorem 9, we have kσmµ(A + B)kE ≤ kσm(µ(A) + µ(B))kE ≤ kσmµ(A)kE + kσmµ(B)kE. Note that kA⊕mkE = kσmµ(A)kE. Dividing by m and letting m → ∞, we obtain π(A + B) ≤ π(A) + π(B). (cid:3) 3. Lifting of symmetric functionals In this section, we explain a canonical bijection between symmetric functionals and traces. In what follows, we require that a semifinite von Neumann algebra M be either atomless or atomic with traces of all atoms being 1. For an atomless von Neumann algebra M, we have (see e.g. [12]) Z t 0 µ(s, A)ds = sup{τ (pA) : p ∈ P (M), τ (p) = t}, A ∈ M. 8 F. SUKOCHEV AND D. ZANIN For a atomic von Neumann algebra M, we have (see e.g. [12]) µ(k, A) = sup{τ (pA) : p ∈ P (M), τ (p) = m}, A ∈ M. m Xk=1 In either case, this implies a remarkable inequality (see e.g. [12]) (8) µ(A + B) ≺≺ µ(A) + µ(B) ≺≺ 2σ1/2µ(A + B), 0 ≤ A, B ∈ (L1 + L∞)(M). Lemma 12. Let E = E(0, 1) (or E = E(0, ∞) or E = E(N)) be a symmetric Banach space either on the interval (0, 1) or on the semi-axis or on N. If x, y ∈ E+ are such that y ⊳ x, then ϕ(y) ≤ ϕ(x) for every positive symmetric functional ϕ on E. Proof. Fix ε > 0. By Theorem 9, there exist zk ∈ E, 1 ≤ k ≤ n, and positive numbers λk, 1 ≤ k ≤ n, such that µ(zk) ≤ µ(x) for every 1 ≤ k ≤ n and (1 − ε)y = λkzk, n Xk=1 λk = 1. n Xk=1 Since ϕ is positive and symmetric, it follows that ϕ(zk) ≤ ϕ(zk) = ϕ(µ(zk)) ≤ ϕ(µ(x)) = ϕ(x). Therefore, (1 − ε)ϕ(y) ≤ ϕ(x). Since ε > 0 is arbitrarily small, the assertion follows. (cid:3) The following assertion is essentially known. However, we provide the full proof for readers convenience. Lemma 13. Let M be a semifinite atomless von Neumann algebra and let A, B ∈ (L1 + L∞)(M, τ ) be positive operators. 2a µ(s, A + B)ds ≤Z b Z b (µ(s, A) + µ(s, B))ds ≤Z 2b Z b 2a 2a a (µ(s, A) + µ(s, B))ds, ∀2a ≤ b, µ(s, A + B)ds, ∀2a ≤ b. Similar assertion is valid for atomic von Neumann algebra M. Proof. Applying inequality (8) to the operators A, B, we obtain that and Subtracting this inequalities, we obtain 0 Z b Z 2a Z b 0 0 µ(s, A + B)ds ≤Z b µ(s, A + B)ds ≥Z a µ(s, A + B)ds ≤Z b 0 (µ(s, A) + µ(s, B))ds (µ(s, A) + µ(s, B))ds. (µ(s, A) + µ(s, B))ds. a Proof of the second inequality is identical. 2a (cid:3) The following theorem answers Question 4 in the affirmative, as also does [20, Theorem 5.2]. The proof below is very simple and based on a completely different approach. TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 9 Theorem 14. Let E = E(0, 1) (or E = E(0, ∞) or E = E(N)) be a symmetric Banach space either on the interval (0, 1) or on the semi-axis or on N and let E(M, τ ) be the corresponding symmetric Banach operator space. (1) If ϕ is a positive symmetric functional on E, then there exists a positive symmetric functional L(ϕ) on E(M, τ ) such that ϕ(x) = L(ϕ)(A) for all positive x ∈ E and A ∈ E(M, τ ) such that µ(A) = µ(x). (2) If ϕ is a positive symmetric functional on E(M, τ ), then there exists a positive symmetric functional L−1(ϕ) on E such that ϕ(A) = L−1(ϕ)(x) for all positive x ∈ E and A ∈ E(M, τ ) such that µ(A) = µ(x). Proof. We will only prove (1). Proof of (2) is identical. Let A, B ∈ E+(M, τ ). It follows from Lemma 13 that µ(A + B) ⊳ µ(A) + µ(B) ⊳ 2σ1/2µ(A + B). It follows from Lemma 12 that ϕ(µ(A + B)) ≤ ϕ(µ(A) + µ(B)) ≤ ϕ(2σ1/2µ(A + B)) = ϕ(µ(A + B)). It follows that L(ϕ) is additive on E+(M, τ ). We than extend it to E(M, τ ) by linearity. (cid:3) Theorem 14 provides a very natural bijection between the set of all symmetric functionals on E and that on E(M, τ ), observed first for the case of fully symmetric functionals in [8]. Next corollary follows immediately. Corollary 15. Let E and E(M, τ ) be as in Theorem 14. The functional ϕ is fully symmetric on E if and only if L(ϕ) is a fully symmetric functional on E(M, τ ). We also need a lifting between sequence and function spaces. The following space was introduced in [21]. Let A = {[n − 1, n]}n∈N be a partition of the semi-axis. Clearly, E(·A) maps L1 + L∞ into the set of step functions which can be identified with sequences. Proposition 16. Let E be a symmetric Banach sequence space and let F be the linear space of all such functions x ∈ L∞ for which E(µ(x)A) ∈ E. The space F equipped with the norm is a symmetric Banach function space. kxkF = kxk∞ + kE(µ(x)A)kE The fact that the space F is a Banach space is non-trivial. Proof of this fact was missing in both [19] and [21]. We include it in the appendix. Below, we assume that E is embedded into F. Theorem 17. Let E = E(N) be a symmetric Banach sequence space and let F be the corresponding function space. (1) If ϕ is a positive symmetric functional on E, then there exists a positive symmetric functional L(ϕ) on F such that ϕ(E(µ(x)A)) = L(ϕ)(x) for all positive x ∈ F. (2) If ϕ is a positive symmetric functional on F, then its restriction on E is a positive symmetric functional. This restriction is an inverse operation for the L in (1). 10 F. SUKOCHEV AND D. ZANIN Proof. Let us prove (1) ϕ(σ1/2a) = 1/2ϕ(a1, a3, · · · ) + 1/2ϕ(a2, a4, · · · ) = = 1/2ϕ(a1, 0, a2, 0, · · · ) + 1/2ϕ(0, a2, 0, a4, · · · ) = 1/2ϕ(a) for every a ∈ E. Let x, y ∈ F be positive. It follows from Lemma 50 that E(µ(x + y)A) ⊳ E(µ(x) + µ(y)A) ⊳ 2σ1/2E(µ(x + y)A). It follows from Lemma 12 that ϕ(E(µ(x + y)A)) = ϕ(E(µ(x) + µ(y)A)) and (1) follows. The first assertion of (2) is trivial. Clearly, µ(x) − E(µ(x)A) ∈ (L1 ∩ L∞)(0, ∞). If E 6= l1, then ϕ(y) = 0 for every y ∈ (L1 ∩ L∞)(0, ∞) and every symmetric functional ϕ on F. If E = l1, then F = (L1 ∩ L∞)(0, ∞) and the only symmetric functional on both spaces is an integral. The second assertion of (2) follows. (cid:3) 4. Existence of symmetric functionals In this section, we present results concerning existence of symmetric functionals on symmetric function spaces. The main results of this section are Theorem 23, Theorem 28 and Theorem 29. We need the following variation of the Hahn-Banach theorem. Lemma 18. Let E be a partially ordered linear space and let p : E → R be convex and monotone functional. For every x0 ∈ E, there exists a positive linear functional ϕ : E → R such that ϕ ≤ p and ϕ(x0) = p(x0). Proof. The existence of ϕ follows from the Hahn-Banach theorem. We only have to prove that ϕ ≥ 0. If z ≥ 0, then ϕ(x0 − z) ≤ p(x0 − z). Therefore, ϕ(z) ≥ ϕ(x0) − p(x0 − z) = p(x0) − p(x0 − z) ≥ 0 due to the fact that z ≥ 0 and p is monotone. (cid:3) Define operators Mm : (L1 +L∞)(0, ∞) → (L1+L∞)(0, ∞) (or, Mm : L1(0, 1) → L1(0, 1)) by setting (Mmx)(t) = 1 t log(m)Z t t/m x(s), m ≥ 2. Lemma 19. If 0 ≤ x ∈ L1 + L∞ (or, 0 ≤ x ∈ L1(0, 1)), then Z b/m a x(s)ds ≤Z b a (Mmx)(s)ds ≤Z b a/m x(s)ds provided that ma ≤ b. In particular, m−1σmx ⊳ Mmx ⊳ x provided that x = µ(x). Proof. Clearly, (Mmx)(s)ds = a Z b a/mZ min{ms,b} max{a,s} = 1 log(m)Z b x(s) log( min{ms, b} max{a, s} )ds. x(s)ds dt t = 1 t/m log(m)Z b a Z t log(m)Z b 1 a/m dt t x(s)ds = TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 11 The integrand does not exceed x(s) log(m) and the second inequality follows im- mediately. The integrand is positive and is equal to x(s) log(m) for s ∈ (a, b/m). The first inequality follows. (cid:3) Corollary 20. If E is a symmetric Banach function space either on the interval (0, 1) or on the semi-axis, then Mm : E → E is a contraction for m ∈ N. Proof. Let x = µ(x) ∈ E. It follows from Lemma 19 that Mmx ⊳ x. It follows from theorem 9 that, for every ε > 0, the function (1 − ε)Mmx belongs to a convex hull of the set {z : µ(z) ≤ µ(x)}. Therefore, Mmx ∈ E and (1 − ε)kMmxkE ≤ kxkE. Since ε is arbitrarily small, the assertion follows. (cid:3) Lemma 21. Let E be a symmetric Banach space either on the interval (0, 1) or on the semi-axis. Let p : DE → R be convex and monotone functional. If p = 0 on ZE ∩ DE, then p extends to a convex monotone functional p : E → R by setting Also, p(x) = 0 for every x ∈ ZE. p(x) = p(µ(x+) − µ(x−)). Proof. If x ∈ DE, then x − µ(x+) + µ(x−) ∈ ZE ∩ DE. Therefore, p(x − µ(x+) + µ(x−)) = 0 and, due to the convexity of p, p(x) = p(µ(x+) − µ(x−)). This proves the correctness of the definition. For x, y ∈ E, we have µ((x + y)+) − µ((x + y)−) − µ(x+) + µ(x−) − µ(y+) + µ(y−) ∈ ZE ∩ DE. It follows that p(µ((x + y)+) − µ((x + y)−) − µ(x+) + µ(x−) − µ(y+) + µ(y−)) = 0 and p(x + y) = p(µ((x + y)+) − µ((x + y)−)) = = p(µ(x+) − µ(x−) + µ(y+) − µ(y−)) ≤ p(x) + p(y). Since p is monotone on DE, then p(y) ≤ 0 for every 0 ≥ y ∈ DE. It follows that p(y) = p(−µ(y)) ≤ 0 for 0 ≥ y ∈ E. Therefore, p(x + y) ≤ p(x) + p(y) ≤ p(x) for every 0 ≥ y ∈ E. (cid:3) Lemma 22. Let E be a symmetric Banach space either on the interval (0, 1) or on the semi-axis. The functional p : x → lim sup m→∞ k(Mmx)+kE, x ∈ DE satisfies the assumptions of Lemma 21. Also, for every x ∈ DE, we have p(x) ≤ kxkE. Proof. It follows from Corollary 20 that k(Mmx)+kE ≤ kMmxkE ≤ kxkE, x ∈ E. It follows that p(x) = lim sup m→∞ k(Mmx)+kE ≤ kxkE, x ∈ DE. Clearly, the mappings x → (Mmx)+ are convex and monotone. So are the mappings x → k(Mmx)+kE. Therefore, p : DE → R is a convex and monotone functional. 12 F. SUKOCHEV AND D. ZANIN If x ∈ ZE ∩ DE, then by Theorem 8 Cx ∈ E. Therefore, (Mmx)(t) ≤ 1 log(m) ( ≤ 1 log(m) ( 1 m 1 t Z t/m 0 x(s)ds + 1 t Z t 0 x(s)ds) ≤ σmCx + Cx)(t). Since kσmkE→E ≤ m (see [22, Theorem II.4.5]), it follows that k(Mmx)+kE ≤ 2 log(m) kCxkE and p(x) = 0. (cid:3) Theorem 23. Let E = E(0, ∞) be a symmetric Banach space on the semi-axis. For a given 0 ≤ x ∈ E, there exists a symmetric linear functional ϕ : E → R such that ϕ(x) = lim m→∞ kσm(µ(x))kE. 1 m Proof. Without loss of generality, x = µ(x). Let p be the convex monotone func- tional constructed in Lemma 22. It follows from Lemma 18 that there exist a positive linear functional ϕ on E such that ϕ ≤ p and ϕ(x) = p(x). Since p(z) = 0 for every z ∈ ZE, it follows that ϕ(z) = 0 for every z ∈ ZE. Therefore, ϕ is a symmetric functional. Since ϕ(z) ≤ p(z) ≤ kzkE for every z = µ(z) ∈ E, it follows that kϕkE ∗ ≤ 1. Therefore, ϕ(x) = ϕ( 1 m σmx) ≤ 1 m kσmxkE. Passing m → ∞, we obtain ϕ(x) ≤ lim m→∞ 1 m kσmµ(x)kE. On the other hand, It follows from Lemma 19 that m−1σmx ⊳ Mmx. Therefore, p(x) = lim sup m→∞ kMmxkE ≥ lim m→∞ 1 m kσmµ(x)kE. The assertion follows immediately. (cid:3) Consider the functional π : E → E (identical to the one defined in Section 1). (9) π(x) = lim m→∞ 1 m kx⊕mkE, x ∈ E. Note that π(−x) = π(x) for every x ∈ E. If p is a functionals defined in Lemma 22, then p(−x) = 0 for positive x ∈ E. Therefore, p 6= π. However, the assertion below follows from Theorem 23. Lemma 24. Let E = E(0, ∞) be a symmetric Banach space on the semi-axis. Let p and π be the convex functionals on E defined in Lemma 22 and (9), respectively. For every positive x ∈ E, we have p(x) = π(x). Proof. For every x ∈ E, consider the functional ϕ constructed in Theorem 23. By construction, we have ϕ(x) = p(x) = π(x). (cid:3) If E 6⊂ L1(0, ∞), then the functional ϕ constructed in Theorem 23 is necessarily singular. The case E ⊂ L1 requires more detailed treatment. TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 13 Lemma 25. Let E be a symmetric (respectively, fully symmetric) Banach function space either on the interval (0, 1) or on the semi-axis. Let {ϕi}i∈I ∈ E∗ be a net and let ϕ ∈ E∗ be such that ϕi → ϕ ∗−weakly. (1) If every ϕi is symmetric, then ϕ is symmetric. (2) If every ϕi is fully symmetric, then ϕ is fully symmetric. Proof. Let each ϕi be symmetric. If 0 ≤ x1, x2 ∈ E are such that µ(x1) = µ(x2), then ϕ(x1) = lim i∈I ϕi(x1) = lim i∈I ϕi(x2) = ϕ(x2). Hence, ϕ is symmetric. Let each ϕi be fully symmetric. Thus, ϕi(x) ≤ 0 for every x ∈ DE such that Cx ≤ 0. Therefore, ϕ(x) = limi∈I ϕi(x) ≤ 0 for every x ∈ DE such that Cx ≤ 0. Let x1, x2 ∈ E be positive elements such that x1 ≺≺ x2. Therefore, z = µ(x1) − µ(x2) ∈ DE and Cz ≤ 0. It follows from above that ϕ(z) ≤ 0. Hence, ϕ is a fully symmetric functional. (cid:3) Lemma 26. Let E be a symmetric (respectively, fully symmetric) Banach function space either on the interval (0, 1) or on the semi-axis and let ϕ be a symmetric (respectively, fully symmetric) functional on E. The formula ϕsing(x) = lim n→∞ ϕ(µ(x)χ(0,1/n)), 0 ≤ x ∈ E. defines a singular symmetric (respectivley, fully symmetric) linear functional on E. Proof. If x, y ∈ E are positive functions, then µ(x + y)χ(0,1/n) ⊳ (µ(x) + µ(y))χ(0,1/n) ⊳ 2σ1/2µ(x + y)χ(0,1/n). Taking the limit as n → ∞, we derive from Lemma 12 that ϕsing(µ(x + y)) = ϕsing(µ(x) + µ(y)). Since ϕ is symmetric, it follows that ϕsing(x + y) = ϕsing(µ(x + y)) = ϕsing(µ(x) + µ(y)) = ϕsing(x) + ϕsing(y). Hence, ϕsing is an additive functional on E+. Therefore, it extends to a linear functional on E. Clearly, ϕsing is symmetric. Second assertion is trivial. (cid:3) In fact, the construction in Lemma 26 gives a singular part of the functional ϕ as defined by Yosida-Hewitt theorem. Lemma 27. Let E = E(0, ∞) ⊂ L1(0, ∞) be a symmetric Banach function space on the semi-axis and let ϕ be a symmetric functional on E. If ϕsing is a functional constructed in Lemma 26, then ϕ−ϕsing is a normal functional (that is, an integral). Proof. It is clear that 0 ≤ ϕsing(z) ≤ kzk∞ lim n→∞ ϕ(χ(0,1/n)) = 0 for every positive z ∈ (L1 ∩ L∞)(0, ∞). It follows that ϕsing(µ(x)χ(1/n,∞)) = 0 for every x ∈ E. Therefore, (10) (ϕ − ϕsing)(x) = lim n→∞ ϕ(µ(x)χ(1/n,∞)) = lim n→∞ (ϕ − ϕsing)(µ(x)χ(1/n,∞)). 14 F. SUKOCHEV AND D. ZANIN On the other hand, for every positive z ∈ (L1 ∩ L∞)(0, ∞) with kzk∞ = 1, we have z ≺ χ(0,kzk1). It is proved in [30, Theorem 23] that z belongs to the closure (in the topology of L1 ∩ L∞) of the set {u ≥ 0 : µ(u) = χ(0,kzk1)}. Thus, (ϕ − ϕsing)(z) = (ϕ − ϕsing)(χ(0,kzk1)) = kzk1(ϕ − ϕsing)(χ(0,1)). By linearity, (11) (ϕ − ϕsing)(z) = (ϕ − ϕsing)(χ(0,1)) ·Z ∞ 0 It follows that from (10) and (11) that z(s)ds, ∀z ∈ (L1 ∩ L∞)(0, ∞). (ϕ−ϕsing)(x) = lim n→∞Z ∞ 1/n µ(s, x)ds·(ϕ−ϕsing )(0, 1) =Z 1 0 x(s)ds·(ϕ−ϕsing )(χ(0,1)) for every positive function x ∈ E. The assertion follows immediately. (cid:3) Theorem 28. Let E ⊂ L1(0, ∞) be a symmetric Banach space on the semi-axis. For a given 0 ≤ x ∈ E, there exists a singular symmetric linear functional ϕsing such that ϕsing(x) = lim m→∞ 1 m kσm(µ(x))χ(0,1)kE. Proof. Apply Theorem 23 to the function µ(x)χ(0,1/n). It follows that there exists a symmetric linear functional ϕn such that kϕnkE ∗ ≤ 1 and ϕn(µ(x)χ(0,1/n)) = lim m→∞ 1 m kσm(µ(x)χ(0,1/n))kE ≥ lim m→∞ 1 m kσm(µ(x))χ(0,1)kE. Since the unit ball in E∗ is ∗−weakly compact (Banach-Alaoglu theorem), there exists a convergent subnet ψi = ϕF (i), i ∈ I, of the sequence ϕn, n ∈ N. Let ψi → ϕ. It follows from Lemma 25 that ϕ is a symmetric functional. By the definition of a subnet (see [26, Section IV.2]), for every fixed n ∈ N, there exists in ∈ I such that F (i) > n for every i > in. Thus, for every i > in, we have ψi(µ(x)χ(0,1/n)) ≥ ϕF (i)(µ(x)χ(0,1/F (i))) ≥ lim m→∞ 1 m kσm(µ(x))χ(0,1)kE. The subnet ψi, in < i ∈ I converges to the same limit ϕ. Therefore, ϕ(µ(x)χ(0,1/n)) ≥ lim m→∞ 1 m kσm(µ(x))χ(0,1)kE. Now, taking the limit as n → ∞, we obtain the inequality ϕsing(x) ≥ lim m→∞ 1 m kσm(µ(x))χ(0,1)kE, where ϕsing is a singular symmetric functional defined in Lemma 26. The opposite inequality is trivial. (cid:3) Theorem 29. Let E be a symmetric Banach space on the interval (0, 1). For a given 0 ≤ x ∈ E, there exists a singular symmetric linear functional ϕsing such that ϕsing(x) = lim m→∞ 1 m kσm(µ(x))kE. TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 15 Proof. Let F be a symmetric Banach space on the semi-axis with a norm given by the formula kxkF = kµ(x)χ(0,1)kE + kxk1, ∀x ∈ F. Clearly, F ⊂ L1(0, ∞). Applying Theorem 28, we obtain a symmetric singular functional ϕ on F such that ϕ(x) = lim m→∞ 1 m kσm(µ(x))χ(0,1)kF = lim m→∞ 1 m kσm(µ(x))kE. (cid:3) 5. Existence of fully symmetric functionals In this section, we present results concerning existence of fully symmetric func- tionals on fully symmetric function spaces. The main results of this section are Theorem 33, Theorem 35 and Theorem 36. Lemma 30. Let E be a symmetric Banach function space either on the interval (0, 1) or on the semi-axis. If x, z ∈ DE are such that Cx ≤ Cz, then CMmx ≤ CMmz. Proof. Let x = µ(a) − µ(b) and z = µ(c) − µ(d) with a, b, c, d ∈ E. It follows from assumption Cx ≤ Cz that C(µ(a) + µ(d)) ≤ C(µ(b) + µ(c)) or, equivalently, µ(a) + µ(d) ≺≺ µ(b) + µ(c). Arguing as in Lemma 19, we have with It is now clear that 0 0 Z t (Mmz)(s)ds =Z t h(s, t) =  log(t/s) log(m) 1, , z(s)h(s, t)ds 0 ≤ s ≤ t/m t/m ≤ s ≤ t 0 Z t Z t 0 Mm(µ(a) + µ(d))(s)ds =Z t Mm(µ(b) + µ(c))(s)ds =Z t 0 0 (µ(s, a) + µ(s, d))h(s, t)ds, (µ(s, b) + µ(s, c))h(s, t)ds. Clearly, h is positive and decreasing with respect to s. It follows from [22, Equality 2.36] that Mm(µ(a) + µ(d)) ≺≺ Mm(µ(b) + µ(c)) and the assertion follows. (cid:3) Lemma 31. Let E be a fully symmetric Banach function space either on the in- terval (0, 1) or on the semi-axis and let x = µ(x) ∈ E. If z ∈ DE is such that Cx ≤ Cz, then p(x) ≤ p(z). Proof. Since Mmx is decreasing, it follows from Lemma 30 that Z t 0 µ(s, Mmx)ds =Z t 0 (Mmx)(s)ds ≤Z t 0 (Mmz)+(s)ds ≤Z t 0 µ(s, (Mmz)+)ds. Therefore, (Mmx)+ = Mmx ≺≺ (Mmz)+. The assertion follows now from the definition of the functional p. (cid:3) 16 F. SUKOCHEV AND D. ZANIN Lemma 32. Let E be a fully symmetric Banach function space either on the in- terval (0, 1) or on the semi-axis. Let p be the functional constructed in Lemma 22. The functional q(x) = inf{p(z) : z ∈ DE, Cx ≤ Cz}, x ∈ DE satisfies the assumptions of Lemma 21. Proof. It is clear from the definition of q that q ≤ p and that q is a positive functional. We claim that q is convex on DE. Let x1, x2 ∈ DE. Fix ε > 0 and select z1, z2 ∈ DE such that Cxi ≤ Czi and p(zi) ≤ q(xi) + ε for i = 1, 2. Thus, C(x1 + x2) ≤ C(z1 + z2) and q(x1 + x2) ≤ p(z1 + z2) ≤ p(z1) + p(z2) ≤ q(x1) + q(x2) + 2ε. Since ε is arbitrarily small, the claim follows. We claim that q is monotone on DE. Let x1, x2 ∈ DE be such that x1 ≤ x2. Fix ε > 0 and select z ∈ DE such that Cx2 ≤ Cz and p(z) ≤ q(x2) + ε. Thus, Cx1 ≤ Cx2 ≤ Cz and q(x1) ≤ p(z) ≤ q(x2) + ε. Since ε is arbitrarily small, the claim follows. For x ∈ ZE ∩ DE, we have 0 ≤ q(x) ≤ p(x) = 0 and, therefore, q(x) = 0. s (cid:3) The following theorem is the first main result of this section. Theorem 33. Let E = E(0, ∞) be a fully symmetric Banach space on the semi- axis. For a given 0 ≤ x ∈ E, there exists a fully symmetric linear functional ϕ : E → R such that ϕ(x) = lim m→∞ kσm(µ(x))kE. 1 m Proof. Without loss of generality, x = µ(x). Let q be the convex monotone func- tional constructed in Lemma 32. It follows from Lemma 18 that there exist a positive linear functional ϕ on E such that ϕ ≤ q and ϕ(x) = q(x). It is clear that ϕ ≤ q ≤ p. Since p(z) = 0 for every z ∈ ZE, it follows that ϕ(z) = 0 for every z ∈ ZE. Therefore, ϕ is a symmetric functional. For every z ∈ DE with Cz ≤ 0, we have ϕ(z) ≤ q(z) ≤ p(0) = 0. Let x1, x2 ∈ E be positive elements such that x1 ≺≺ x2. Therefore, z = µ(x1) − µ(x2) ∈ DE and Cz ≤ 0. It follows from above that ϕ(z) ≤ 0. Hence, ϕ is a fully symmetric functional. Since ϕ(z) ≤ q(z) ≤ p(z) ≤ kzkE for every z = µ(z) ∈ E, it follows that kϕkE ∗ ≤ 1. Therefore, ϕ(x) = ϕ( 1 m σmx) ≤ 1 m kσmxkE. Passing m → ∞, we obtain ϕ(x) ≤ lim m→∞ 1 m kσmµ(x)kE. On the other hand, q(x) = p(x) by Lemma 31. By Lemma 19, we have m−1σmx ⊳ Mmx. Therefore, ϕ(x) = q(x) = p(x) = lim sup m→∞ kMmxkE ≥ lim m→∞ 1 m kσmµ(x)kE. The assertion follows immediately. (cid:3) TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 17 If π : E → E is a convex functional defined in (9), then π(−x) = π(x) for every x ∈ E. If q is a functional defined in Lemma 32, then q(−x) = 0 for positive x ∈ E. Therefore, q 6= π. However, the assertion below follows from Theorem 33. Lemma 34. Let E = E(0, ∞) be a fully symmetric Banach space on the semi- axis. Let q and π be the convex functionals on E defined in Lemma 32 and (9), respectively. For every positive x ∈ E, we have q(x) = π(x). Proof. For every x ∈ E, consider the functional ϕ constructed in Theorem 33. By construction, we have ϕ(x) = q(x) = p(x) = π(x). (cid:3) The proofs of the two following theorems are very similar to that of Theorem 28 (respectively, Theorem 29) and are, therefore, omitted. The only difference is that the reference to Theorem 23 (respectively, Theorem 28) has to be replaced with the reference to Theorem 33 (respectively, Theorem 35). Theorem 35. Let E ⊂ L1(0, ∞) be a fully symmetric Banach space on the semi- axis. For a given 0 ≤ x ∈ E, there exists a singular fully symmetric linear functional ϕsing such that ϕsing(x) = lim m→∞ 1 m kσm(µ(x))χ(0,1)kE. Theorem 36. Let E ⊂ L1(0, 1) be a fully symmetric Banach space on the inter- val (0, 1). For a given 0 ≤ x ∈ E, there exists a singular fully symmetric linear functional ϕsing such that ϕsing(x) = lim m→∞ 1 m kσm(µ(x))kE. 6. The sets of symmetric and fully symmetric functionals are different In this section, we demonstrate that the sets of symmetric and fully symmetric functionals on a given fully symmetric space E are distinct (provided that one of these sets is non-empty). The main results are Theorem 47 and Theorem 48. Let x = µ(x) ∈ (L1 + L∞)(0, ∞) (or x = µ(x) ∈ L1(0, 1)) and let X(t) = R t 0 x(s)ds. For every θ > 0, let an(θ) be such that X(an(θ)) = (3/2)nθ for every n ∈ Z such that an(θ) does exist. Given a sequence κ = {κn}n∈Z ∈ (N ∪ {∞})Z, let Bκ,θ = {κna3n(θ), where n ∈ Z is such that κ2 na3n(θ) < a3n+1(θ)}. If κn = m for all n ∈ N, we write Bm,θ instead of Bκ,θ. Also, set Am = {man(1) : m2an(1) < an+1(1), n ∈ Z}. Lemma 37. If x = µ(x) ∈ L1 + L∞ and if Ci, 1 ≤ i ≤ k, are discrete sets, then E(x ∪k i=1 Ci) ≺≺ E(xCi). k Xi=1 18 F. SUKOCHEV AND D. ZANIN Proof. It is sufficient to verify Z t 0 E(x ∪k i=1 Ci)(s)ds ≤ only at the nodes of E(x ∪k However, if t ∈ Ci for some i, then k Xi=1Z t 0 E(xCi)(s)ds i=1 Ci), that is at the nodes of E(xCi) for every i. Z t 0 and we are done. E(x ∪k i=1 Ci)(s)ds = X(t) =Z t 0 E(xCi)(s)ds (cid:3) We will need the following lemma. Lemma 38. If x = µ(x) ∈ L1 + L∞ and if κ ≥ κ′ (that is κn ≥ κ′ then n for every n), (12) E(xBκ,θ) ≺≺ 3 2 E(xBκ′,θ). Proof. Let n ∈ Z be such that κ2 a3n+1(θ). Therefore, na3n(θ) < a3n+1(θ). It follows that κ′2 n a3n(θ) < Z κna3n(θ) 0 0 na3n(θ) E(xBκ,θ)(s)ds ≤Z a3n+1(θ) x(s)ds = 3/2Z κ′ E(xBκ,θ)(s)ds ≤ 3/2Z t ≤ 3/2Z κ′ Z t 0 0 0 0 Hence, we have (13) x(s)ds = 3/2Z a3n(θ) 0 na3n(θ) x(s)ds ≤ E(xBκ′,θ)(s)ds. E(xBκ′,θ)(s)ds for every t being a node of the partition Bκ,θ. Thus, (13) holds for every t > 0. (cid:3) Remark 39. The inequality (12) holds if κn ≥ κ′ the inequality κ2 na3n(θ) < a3n+1(θ). n only for such n ∈ Z that satisfy Lemma 40. Let E be a fully symmetric Banach function space either on the in- terval (0, 1) or on the semi-axis. Let x = µ(x) ∈ E and y = µ(y) ∈ E be such that ϕ(y) ≤ ϕ(x) for every positive symmetric functional ϕ ∈ E∗. There exists 0 ≤ um ∈ E such that um → 0 in E and Z b ma y(s)ds ≤Z mb a (x + um)(s)ds, ∀ma ≤ b. Proof. Let p be a convex positive functional considered in Lemma 22. By Lemma 18, there exists a positive functional ϕ ∈ E∗ such that ϕ ≤ p and ϕ(y−x) = p(y−x). We have p(z) = 0 for every z ∈ ZE and, therefore, ϕ(z) = 0 for every z ∈ ZE. Therefore, ϕ is a positive symmetric linear functional on E. By the assumption, ϕ(y − x) ≤ 0 and, therefore, p(y − x) = 0. Hence, by the definition of p, we have um = (Mm(y − x))+ → 0 in E. Clearly, Mmy ≤ Mmx + um. It follows from Lemma 19 that Z b ma y(s)ds ≤Z mb ma (Mmy)(s)ds ≤Z mb ma (Mmx + um)(s)ds ≤Z mb a (x + um)(s)ds. (cid:3) TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 19 For each sequence κ and λ > 0, we define the sequence κλ by setting κλ n =(κn, ∞, κn ≥ λ κn < λ. Lemma 41. If m ∈ N, x = µ(x) ∈ L1 + L∞ and 0 ≤ u ∈ L1 + L∞ are such that Z b ma E(xBκ,θ)(s)ds ≤Z mb a (x + u)(s)ds, ∀ma ≤ b ∈ R, m−1σmE(xBκ100m,θ) ≺≺ 30µ(u). then (14) Proof. If κ100m trivial. n = ∞ for every n ∈ Z, then E(xBκ100m,θ) = 0 and the assertion is Let n ∈ Z be such that κ2 na3n(θ) < a3n+1(θ) and κn ≥ 100m. It follows that (15) Z mκna3n(θ) 0 u(s)ds ≥Z mκna3n(θ) a3n(θ) (x + u)(s)ds −Z mκna3n(θ) a3n(θ) x(s)ds. By the assumption, we have (16) Z mκna3n(θ) a3n(θ) (x + u)(s)ds ≥Z κna3n(θ) ma3n(θ) E(xBκ,θ)(s)ds. Note that mκna3n(θ) < a3n+1(θ). It follows from (15) and (16) that (17) Z mκna3n(θ) 0 u(s)ds ≥Z κna3n(θ) ma3n(θ) E(xBκ,θ)(s)ds −Z a3n+1(θ) a3n(θ) x(s)ds. Let n′ be the maximal integer number such that n′ < n and κ2 It is clear that n′ a3n′(θ) < a3n′+1(θ). κ2 n′a3n′ (θ) < a3n′+1(θ) ≤ a3n−2(θ) < ma3n(θ) and (18) E(xBκ,θ) = X(κna3n(θ)) − X(κn′a3n′ (θ)) κna3n(θ) − κn′ a3n′(θ) ≥ X(a3n(θ)) − X(a3n−2(θ)) κna3n(θ) on the interval (ma3n(θ), κna3n(θ)). If κ2 n′a3n′ (θ) ≥ a3n+1(θ) for every n′ < n, then X(κna3n(θ)) (19) E(xBκ,θ) = κna3n(θ) ≥ X(a3n(θ)) κna3n(θ) on the interval (ma3n(θ), κna3n(θ)). It follows from (17) and (18) (or (19)) that Z mκna3n(θ) 0 u(s)ds ≥ κn − m κn · (1 − 4 9 )X(a3n(θ)) − 1 2 X(a3n(θ)). Since κn ≥ 100m, it follows that Z mκna3n(θ) 0 u(s)ds ≥ ((1 − 1 100 )(1 − 4 9 ) − 1 2 )X(a3n(θ)) = 1 20 X(a3n(θ)) = = 1 30 X(a3n+1(θ)) ≥ 1 30 X(κna3n(θ)) = 1 30Z κna3n(θ) 0 E(xBκ100m,θ)(s)ds. for every t being a node of the partition Bκ100m,θ. Therefore, u(s)ds ≤ 30Z mt 0 µ(s, u)ds 20 F. SUKOCHEV AND D. ZANIN It follows immediately that (20) Z t 0 or, equivalently, 0 E(xBκ100m,θ)(s)ds ≤ 30Z mt Z t Z t/m E(xBκ100m,θ)(s)ds ≤ 30Z mt E(xBκ100m,θ)(s)ds ≤ 30Z t 0 0 0 0 µ(s, u)ds, t > 0 µ(s, u)ds, t > 0. (cid:3) The assertion follows immediately. Lemma 42. Let E be a fully symmetric Banach function space either on the in- terval (0, 1) or on the semi-axis. If x = µ(x) ∈ E is such that ϕ(y) ≤ ϕ(x) for every positive symmetric functional ϕ on E and every 0 ≤ y ≺≺ x, then λ−1σλE(xBκλ,θ) → 0 as λ → ∞. Proof. Since E(xBκ,θ) ≺≺ x, it follows from the assumption and Lemma 40 that there exists 0 ≤ um → 0 such that Z b ma E(xBκ,θ)(s)ds ≤Z mb a (x + um)(s)ds, ∀ma ≤ b ∈ R. For every λ ≥ 100m, we have κ100m ≤ κλ. It follows from Lemma 41 that 1 λ σλE(xBκλ,θ) ≺≺ 1 m σmE(xBκλ,θ) Lemma38 ≺≺ 3 2m σmE(xBκ100m,θ) (14) ≺≺ 45µ(um). The assertion now follows immediately. (cid:3) Proposition 43. Let E be a fully symmetric Banach function space either on the interval (0, 1) or on the semi-axis equipped with a Fatou norm. If x = µ(x) ∈ E is such that ϕ(y) ≤ ϕ(x) for every positive symmetric functional ϕ on E and every 0 ≤ y ≺≺ x, then m−1σmE(xBm,θ) → 0 as m → ∞. Proof. For every m, r ∈ N, set κm,r n =(m 0 ≤ n < r ∞ r ≤ n and κm,r = {κm,r r → ∞. It follows from the definition of Fatou norm that n }n∈Z. Clearly, E(xBκm,r,θ) → E(xBm,θ) almost everywhere when lim r→∞ kσmE(xBκm,r,θ)kE = kσmE(xBm,θ)kE. Select rm so large that (21) 1 m kσmE(xBκm,rm ,θ)kE > 1 2m kσmE(xBm,θ)kE. Now define the sequence κ = {κn}n∈Z by setting κn = inf m≥1 κm,rm n = inf rm>n m, n ∈ Z. TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 21 Clearly, rκn ≥ n and, therefore, κn → ∞ as n → ∞. In particular, the set {n : κn < λ} is finite for every λ ∈ N. Set M (λ) = max{λ, max ( κn<λ a3n+1(θ) a3n(θ) )1/2}. If m > M (λ), then m2a3n(θ) ≥ a3n+1(θ) whenever κn < λ. Thus, κn ≥ λ when- ever m2a3n(θ) < a3n+1(θ). Hence, κλ )2a3n(θ) < a3n+1(θ). )2a3n(θ) < a3n+1(θ). Therefore, κλ According to Remark 39, it follows that n = κn whenever (κm,rm for every n ∈ Z such that (κm,rm n ≤ κm,rm n n n E(xBκm,rm ,θ) ≺≺ 3 2 E(xBκλ,θ). Since m ≥ λ, it follows that (22) 1 m σmE(xBκm,rm ,θ) ≺≺ 3 2λ σλE(xBκλ,θ). By Lemma 42, for every ε > 0, there exists λ such that (23) It follows that 1 λ kσλE(xBκλ,θ)kE < 1 3 ε. 1 m kσmE(xBm,θ)kE (21) ≤ 2 m kσmE(xBκm,rm ,θ)kE (22) ≤ 3 λ kσλE(xBκλ,θ)kE (23) < ε for every m > M (λ). Since ε > 0 is arbitrarily small, the assertion follows. (cid:3) Lemma 44. Let E be a fully symmetric Banach space either on the interval (0, 1) or on the semi-axis equipped with a Fatou norm. If x = µ(x) ∈ E is such that ϕ(y) ≤ ϕ(x) for every positive symmetric functional ϕ on E and every 0 ≤ y ≺≺ x, then m−1σmE(xAm) → 0 as m → ∞. Proof. It is clear that ak(3/2) = ak+1(1) and ak((3/2)2) = ak+2(1) for every k ∈ N. It follows that Therefore, by Lemma 37, we have Bm,1 ∪ Bm,3/2 ∪ Bm,(3/2)2 = Am. (24) E(xAm) ≺≺ E(xBm,1) + E(xBm,3/2) + E(xBm,(3/2)2). The assertion follows now from Proposition 43. (cid:3) Lemma 45. Let x = µ(x) ∈ L1 + L∞(0, ∞) be a function on the semi-axis. If x /∈ L1(0, ∞), then, for every t > 0 and every m ∈ N, we have (25) X(t) ≤ 2 3 X(m4t) + 3 2Z m4t 0 E(xAm)(s)ds. Proof. For a given t > 0, there exists n ∈ Z such that t ∈ [an(1), an+1(1)]. If an+1(1) > m2an(1), then If an+1(1) ≤ m2an(1) and an+2(1) > m2an+1(1), then 0 Z m4t Z m4t E(xAm)(s)ds ≥Z man(1) E(xAm)(s)ds ≥Z man+1(1) 0 0 0 E(xAm)(s)ds = X(man(1)) ≥ 2 3 X(t). E(xAm)(s)ds = X(man+1(1)) ≥ X(t). 22 F. SUKOCHEV AND D. ZANIN If an+2(1) ≤ m2an+1(1) and an+1(1) ≤ m2an(1), then X(m4t) ≥ X(an+2(1)) = 3 2 X(an+1(1)) ≥ 3 2 X(t) and the assertion follows. (cid:3) The situation in the case that x ∈ L1 is slightly more complicated. Lemma 46. If x = µ(x) ∈ L1(0, 1) or x ∈ L1(0, ∞), then there exists constant C such that for every t > 0 (26) X(t) ≤ 2 3 X(m4t) + 3 2Z m4t 0 E(xAm)(s)ds + CZ m4t 0 χ[0,1](s)ds. Proof. Consider first the case of the semi-axis. Fix n0 such that X(an0) ≤ 4/9X(∞). For a givne t ∈ [a, an0], there exists n ∈ Z such that n < n0 and t ∈ [an, an+1]. Then, the argument in Lemma 45 applies mutatis mutandi. For every t ≥ an0 we have X(t) ≤ X(∞) min{an0, 1} min{m4t, 1} = X(∞) min{an0, 1}Z m4t 0 χ[0,1](s)ds. Setting C = X(∞)/ min{an0, 1}, we obtain the assertion. The same argument applies in the case of the interval (0, 1) by replacing X(∞) (cid:3) by X(1). The following two theorems are crucial for the proof of the implication (3) ⇔ (4) in Theorem 5. Theorem 47. Let E be a fully symmetric Banach space either on the interval (0, 1) or on the semi-axis and let x ∈ E. Suppose that the norm on E is a Fatou norm. If ϕ(y) ≤ ϕ(x) for every positive symmetric functional on E and every 0 ≤ y ≺≺ x, then (27) lim m→∞ 1 m kσm(µ(x))kE = 0 provided that one of the following conditions is satisfied (1) E = E(0, 1) is a space on the interval (0, 1). (2) E = E(0, ∞) is a space on the semi-axis and E(0, ∞) 6⊂ L1(0, ∞). Proof. Without loss of generality, x = µ(x). If x /∈ L1, then by Lemma 45, Z t/m4 0 or, equivalently, 3 2Z t 0 E(xAm)(s)ds, ∀t > 0 x(s)ds ≤ x(s)ds + 0 2 3Z t 1 m4 σm4 x ≺≺ 3 2 Applying m−1σm to the both parts, we obtain 3 2 1 m5 σm5 x ≺≺ σmx + 1 m x + 2 3 2 3 E(xAm). 1 m σmE(xAm). Take norms and let m → ∞. It follows from Lemma 44 that lim m→∞ 1 m kσmxkE ≤ 2 3 lim m→∞ 1 m kσmxkE. This proves (27). TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 23 If x ∈ L1 and C are as in Lemma 46, then it follows from Lemma 46 that Z t/m4 0 x(s)ds ≤ or, equivalently, 2 3Z t 0 x(s)ds + 3 2Z t 0 E(xAm)(s)ds + CZ t 0 χ[0,1](s)ds, ∀t > 0 1 m4 σm4 x ≺≺ 2 3 x + 3 2 E(xAm) + Cχ(0,1). Applying m−1σm to the both parts, we obtain 1 m5 σm5 x ≺≺ 2 3 1 m σmx + 3 2 1 m σmE(xAm) + C 1 m σmχ(0,1). Take norms and let m → ∞. For every symmetric space E on the interval (0, 1) and for every symmetric space E on the semi-axis such that E 6⊂ L1(0, ∞) we have m−1kσmχ(0,1)kE → 0. It follows from Lemma 44 that lim m→∞ 1 m kσmxkE ≤ 2 3 lim m→∞ 1 m kσmxkE and again (27) follows. (cid:3) Theorem 48. Let E = E(0, ∞) be a fully symmetric Banach space on the semi- axis equipped with a Fatou norm such that E(0, ∞) ⊂ L1(0, ∞). If ϕ(y) ≤ ϕ(x) for every positive symmetric functional on E and every 0 ≤ y ≺≺ x, then (28) lim m→∞ 1 m kσm(µ(x))χ(0,1)kE = 0. Proof. Fully symmetric Banach space F on the interval (0, 1) consists of those z ∈ E supported on the interval (0, 1). Let x1 = µ(x)χ(0,1) ∈ F. Suppose that It clearly follows that lim m→∞ 1 m kσm(µ(x))χ(0,1)kE > 0. lim m→∞ 1 m kσm(µ(x1))kF > 0. By Theorem 47, there exists 0 ≤ y1 ≺≺ x1 and a positive symmetric functional ϕ ∈ F ∗ such that ϕ(y1) > ϕ(x1). Let ϕsing be a singular part of the functional ϕ constructed in Lemma 26. It follows from Lemma 26 that ϕsing is symmetric. By Lemma 27, the difference ϕ − ϕsing is a symmetric normal functional on F (that is, an integral). Therefore, ϕsing(y1) > ϕsing(x1). Now we show that the functional ϕsing can be extended from F to E by setting ϕsing(z) = lim n→∞ ϕsing(µ(z)χ(0,1/n)), 0 ≤ z ∈ E. Repeating the argument in Lemma 26, we prove that the extension above is additive on E+. Thus, the functional ϕsing ∈ E∗ is positive and symmetric. Since y1 ≺≺ x and ϕsing(y1) > ϕsing(x1) = ϕsing(x), the assertion follows. (cid:3) 24 F. SUKOCHEV AND D. ZANIN 7. Proof of Theorem 5 In this section, we prove an assertion more general then that of Theorem 5. The assertion of Theorem 5 follows from that of Theorem 49 by setting M = B(H). In what follows, the semifinite von Neumann algebra M is either atomless or atomic so that the trace of every atom is 1. Theorem 49. Let E(M, τ ) be a symmetric operator space. Consider the following conditions. (1) There exist nontrivial positive singular symmetric functionals on E(M, τ ). (2) There exist nontrivial singular fully symmetric functionals on E(M, τ ). (3) There exist positive symmetric symmetric functional on E(M, τ ) which are not fully symmetric. (4) If E(M, τ ) 6⊂ L1(M, τ ), then there exists an operator A ∈ E(M, τ ) such that (29) (30) lim m→∞ 1 m kσmµ(A)kE > 0. If E(M, τ ) ⊂ L1(M, τ ), then there exists an operator A ∈ E(M, τ ) such that lim m→∞ 1 m k(σmµ(A))χ(0,1)kE > 0. (i) The conditions (1) and (4) are equivalent for every symmetric operator space E(M, τ ). (ii) The conditions (1), (2) and (4) are equivalent for every fully symmetric op- erator space E(M, τ ). (iii) The conditions (1)-(4) are equivalent for every fully symmetric operator space E(M, τ ) equipped with a Fatou norm. Proof. Implications (2) ⇒ (1) and (3) ⇒ (1) are trivial. (1) ⇒ (4) Let E(M, τ ) be a symmetric operator space with a singular symmetric functional ϕ. Let A ∈ E(M, τ ) be an operator such that ϕ(A) 6= 0. Without loss of generality, A ≥ 0. If E(M, τ ) 6⊂ L1(M, τ ), then ϕ(A) = 1 m ϕ(A ⊕ · · · ⊕ A ) ≤ kϕkE ∗(M,τ ) · 1 m kσmµ(A)kE. Passing m → ∞, we obtain the required inequality (29). Let now E(M, τ ) ⊂ L1(M, τ ). If M is atomic, then E(M, τ ) = L1(M, τ ) and the assertion is trivial. Let M be atomless. Since ϕ is a singular functional and A − AEA(µ( 1 m , A), ∞) ∈ (L1 ∩ L∞)(M, τ ), ∀m ∈ N, m times {z } we infer that = 1 m ≤ kϕkE ∗(M,τ )· 1 m ϕ(A) = ϕ(AEA(µ( 1 m , A), ∞)) = ϕ(AEA(µ( , A), ∞) ⊕ · · · ⊕ AEA(µ( , A), ∞) ) ≤ 1 m 1 m 1 m m times {z kσmµ(AEA(µ( , A), ∞))kE ≤ kϕkE ∗(M,τ )· } k(σmµ(A))χ(0,1)kE. 1 m TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 25 Passing m → ∞, we obtain the required inequality (30). (4) ⇒ (1) Firstly, we assume that the algebra M is finite. Without loss of generality, τ (1) = 1. Let E(M, τ ) be a symmetric operator space and let E(0, 1) be the corresponding symmetric function space. By the assumption, there exists an element x = µ(A) ∈ E(0, 1) such that m−1σmx 6→ 0 in E(0, 1). By Theorem 29, there exists a positive singular symmetric functional 0 6= ϕ ∈ E(0, 1)∗. Let L(ϕ) be a functional on E(M, τ ) defined in Theorem 14. Clearly, L(ϕ) is a nontrivial positive symmetric functional on E(M, τ ). The case when M is an infinite atomless von Neumann algebra can be treated in a similar manner. The only difference is that the reference to Theorem 29 has to be replaced with the reference to either Theorem 28 or Theorem 23. Let E(M, τ ) be a symmetric operator space on a atomic von Neumann algebra M and let E(N) be the corresponding symmetric sequence space. It follows from the assumption that E(M, τ ) 6= L1(M, τ ) or, equivalently, E(N) 6= l1. By the assumption, there exists an element x = µ(A) ∈ E such that m−1σmx 6→ 0 in E. Let F (0, ∞) be a symmetric function space constructed in Proposition 16. Since E(N) 6= l1, it follows that F (0, ∞) 6⊂ L1(0, ∞). Recall that the space E(N) is naturally embedded into the space F (0, ∞) and that the norms k · kE and k · kF are equivalent on E(N). We have x ∈ F and m−1σmx 6→ 0 in F (0, ∞). By Theorem 23, there exists a positive symmetric functional 0 ≤ ϕ ∈ F (0, ∞)∗. The restriction of the functional ϕ to E(N) is a nontrivial positive symmetric functional on E(N). Let L(ϕ) be a functional on E(M, τ ) defined in Theorem 14. Clearly, L(ϕ) is a nontrivial positive symmetric functional on E(M, τ ). (4) ⇒ (2) The proof is very similar to that of the implication (4) ⇒ (1) and is, therefore, omitted. The only difference is that references to Theorem 29, Theorem 28 or Theorem 23 have to be replaced with references to Theorem 36, Theorem 35 or Theorem 33, respectively. (4) ⇒ (3) Firstly, we assume that the algebra M is finite. Without loss of generality, τ (1) = 1. Let E(M, τ ) be a symmetric operator space and let E(0, 1) be the corresponding symmetric function space. By the assumption, there exists an element x = µ(A) ∈ E(0, 1) such that m−1σmx 6→ 0 in E(0, 1). By Theorem 47, there exists a positive symmetric but not fully symmetric functional ϕ ∈ E(0, 1)∗. Let L(ϕ) be a functional on E(M, τ ) defined in Theorem 14. Clearly, L(ϕ) is a symmetric but not fully symmetric functional on E(M, τ ). The case when M is an infinite atomless von Neumann algebra can be treated in a similar manner. The only difference is that the reference to Theorem 47 has to be replaced with the reference to either Theorem 47 or Theorem 48. Let E(M, τ ) be a symmetric operator space on a atomic von Neumann algebra M and let E(N) be the corresponding symmetric sequence space. It follows from the assumption that E(M, τ ) 6= L1(M, τ ) or, equivalently, E(N) 6= l1. By the assumption, there exists an element x = µ(A) ∈ E such that m−1σmx 6→ 0 in E. Let F (0, ∞) be a symmetric function space constructed in Proposition 16. Since E(N) 6= l1, it follows that F (0, ∞) 6⊂ L1(0, ∞). Recall that the space E(N) is naturally embedded into the space F (0, ∞) and that the norms k · kE and k · kF are equivalent on E(N). We have x ∈ F and m−1σmx 6→ 0 in F (0, ∞). By Theorem 47, there exists a positive symmetric functional ϕ ∈ F (0, ∞)∗ and a function 0 ≤ y ≺≺ x such that ϕ(y) > ϕ(x). Set z = E(µ(y){(n − 1, n)}n∈N). Clearly, z ∈ E(N) and ϕ(z) = ϕ(y) > ϕ(x). Hence, the restriction of the functional ϕ to E(N) is a positive 26 F. SUKOCHEV AND D. ZANIN symmetric but not fully symmetric functional on E(N). Let L(ϕ) be a functional on E(M, τ ) defined in Theorem 14. Clearly, L(ϕ) is a positive symmetric but not fully symmetric functional on E(M, τ ). (cid:3) In this appendix, we set A = {(n − 1, n)}n∈N. 8. Appendix Lemma 50. If x, y ∈ (L1 + L∞)(0, ∞) are positive functions, then E(µ(x + y)A) ⊳ E(µ(x)A) + E(µ(y)A) ⊳ 2σ1/2E(µ(x + y)A). Proof. Recall that µ(x + y) ≺≺ µ(x) + µ(y) ≺≺ 2σ1/2µ(x + y). It follows that Let now a, b be positive integers. Subtracting the above inequalities, we obtain 0 0 0 0 Z b µ(s, x + y)ds ≤Z b µ(s, x + y)ds ≥Z a Z 2a E(µ(x + y)A)(s)ds =Z b Z b (µ(s, x) + µ(s, y))ds =Z b ≤Z b E(µ(x) + µ(y)A)(s)ds ≤Z 2b Z b 2a 2a 2a a (µ(s, x) + µ(s, y))ds, (µ(s, x) + µ(s, y))ds. µ(s, x + y)ds ≤ 2a E(µ(x) + µ(y)A)(s)ds. E(µ(x + y)A)(s)ds. a Similarly, we have (cid:3) (cid:3) Corollary 51. The quasi-norm in Construction 16 is a norm. Proof. It follows from Lemma 50 that provided that x, y are positive functions. By Theorem 9, E(µ(x + y)A) ⊳ E(µ(x) + µ(y)A) kE(µ(x + y)A)kE ≤ kE(µ(x)A)kE + kE(µ(y)A)kE. Lemma 52. Let y = µ(y) ∈ (L1 + L∞)(0, ∞). It follows that Z b 2−kλa λ λ − 1Z b a y(s)ds ≤ (σ2k y)(s)ds provided that b ≥ λa. Proof. Let α be the average value of y on the interval [2−kλa, 2−kb]. Clearly, y ≤ α on the interval [2−kλa, b] and y ≥ α on the interval [2−ka, 2−kb]. Thus, σ2k y ≥ α on the interval [a, b]. Therefore, Z b 2−kλa y(s)ds ≤ (b − 2−kλa)α ≤ λ λ − 1 (b − a)α ≤ λ λ − 1Z b a (σ2k y)(s)ds. (cid:3) TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 27 Theorem 53. If {xn}n∈N be a Cauchy sequence in F, then there exists x ∈ F such that xn → x in F. Proof. For every k > 0, there exists mk such that kxm − xmk kF ≤ 4−k for m ≥ mk. Set yk = xmk+1 − xmk . Clearly, kykkF ≤ 4−k for every k ∈ N. In particular, the seriesP∞ Set zn =P∞ k=1 yk converges in L∞(0, ∞). k=n σ2k µ(yk). We claim that zn ∈ F and zn → 0 in F. Indeed, µ(yk) ≤ kykk∞χ(0,1) + T E(µ(yk)A). Here, T is a shift to the right. It follows that E(µ(zn)A) ≤ ∞ Xk=n σ2k (kykk∞χ(0,1) + T E(µ(yk)A)). Therefore, kznkF ≤ kznk∞ + ∞ Xk=n 2kkkykk∞χ(0,1) + T E(µ(yk)A)kE ≤ ≤ kznk∞ + 2k+1kykkF ≤ 1 3 ∞ Xk=n · 41−n + 22−n = o(1). It follows from Lemma 8.5 of [19] that Z b λa µ(s, ∞ Xk=n yk)ds ≤ ∞ Xk=nZ b 2−kλa µ(s, yk)ds. It follows from Lemma 52 that Z b 2−kλa µ(s, yk)ds ≤ λ λ − 1Z b a (σ2k µ(yk))(s)ds. Therefore, Hence, Z b λa µ(s, ∞ Xk=n yk)ds ≤ λ λ − 1Z b a zn(s)ds. yk ⊳ λ λ − 1 zn. ∞ Xk=n Since λ > 1 is arbitrarily large, it follows from Theorem 9 that k ∞ Xk=n ykkF ≤ kznkF → 0. Thus, the seriesP∞ k=1 yk does converge in F. The assertion follows immediately. (cid:3) 28 F. SUKOCHEV AND D. ZANIN 9. Proof of Figiel-Kalton theorem The proof of Theorem 8 follows from the combinations of Lemmas below. Lemma 54. Let E be a symmetric Banach space either on the interval (0, 1) or on the semi-axis. If x ∈ ZE, then C(µ(x+) − µ(x−)) ∈ E. k=1(xk − yk) with xk, yk ∈ E+ and µ(xk) = µ(yk), 1 ≤ k ≤ n. Proof. Let x = Pn Set z = x+ + yk = x− + n Xk=1 xk. n Xk=1 It follows from the definition of C and (8) that Cµ(z) ≤ C(x+) + Cµ(yk) = C(µ(x+) − µ(x−)) + Cµ(x−) + Using the second inequality in (8), we obtain Cµ(xk). n Xk=1 n Xk=1 µ(s, xk))ds ≤Z (n+1)t 0 Z t 0 (µ(s, x−) + Therefore, n Xk=1 µ(s, z)ds ≤Z t 0 µ(s, z)ds + ntµ(t, z). Cµ(z) ≤ Cµ(z) + C(µ(x+) − µ(x−)) + nµ(z). It follows that C(µ(x−) − µ(x+)) ≤ nµ(z). Similarly, C(µ(x+) − µ(x−)) ≤ nµ(z) and the assertion follows. (cid:3) Lemma 55. Let E be a symmetric Banach space either on the interval (0, 1) or on the semi-axis. If x ∈ DE, then C(µ(x+) − µ(x−)) ∈ Cx + E. Proof. Since x ∈ DE, it follows that x = µ(a) − µ(b) with a, b ∈ E. Set u = µ(a) − x+ ≥ 0. Clearly, µ(a) = u + x+ and µ(b) = u + x−. It follows from the definition of C and (8) that Cµ(a) ≤ Cµ(u) + Cµ(x+) = C(µ(x+) − µ(x−)) + Cµ(u) + Cµ(x−). Using the second inequality in (8), we obtain It follows that Similarly, Cµ(x−) + Cµ(u) ≤ Cµ(b) + µ(b). Cx ≤ C(µ(x+) − µ(x−)) + µ(b). Cx ≥ C(µ(x+) − µ(x−)) − µ(a) and the assertion follows. (cid:3) Lemma 56. Let E = E(0, ∞) be a symmetric space on the semi-axis. If x ∈ DE is such that Cx ∈ E, then x ∈ ZE. Proof. Define a partition A = {(2n, 2n+1)}n∈Z and set x1 = E(xA). If x = µ(a) − µ(b) with a, b ∈ E, then x1 = E(µ(a)A) − E(µ(b)A). Clearly, E(µ(a)A) ≤ σ2µ(a) ∈ E, E(µ(b)A) ≤ σ2µ(b) ∈ E are decreasing functions. It follows that x1 ∈ DE. It is easy to see that Cx1 − Cx ≤ 2σ2(µ(a) + µ(b)). TRACES ON SYMMETRICALLY NORMED OPERATOR IDEALS 29 Therefore, Cx1 ∈ E. Define a function z ∈ E by setting z(t) = (Cx1)(2n+1), t ∈ (2n, 2n+1). Clearly, x1 = 2z − σ2z ∈ ZE. Consider the function x − x1 on the interval (2n, 2n+1). By Kwapien theorem [23], there exist positive equimeasurable functions y1n, y2n supported on (2n, 2n+1) such that µ(y1n) = µ(y2n), ky1nk∞, ky2nk∞ ≤ 6k(x − x1)χ(2n,2n+1)k∞. Set y1 =Pn∈N y1n and y2n =Pn∈N y2n. It follows that y1, y2 ∈ E+. Since x − x1 = y1 − y2 and µ(y1) = µ(y2), it follows that x − x1 ∈ ZE. The assertion follows immediately. (cid:3) References [1] S. Astashkin, F. Sukochev, Banach-Saks property in Marcinkiewicz spaces. J. Math. Anal. Appl. 336 (2007) 1231 -- 1258. [2] M. Braverman, A. Mekler, The Hardy-Littlewood property for symmetric spaces, Siberian Math. J. 18 (1977), 371 -- 385. [3] A. Carey, F. Sukochev, Dixmier traces and some applications to noncommutative geometry. (Russian) Uspekhi Mat. Nauk 61 (2006), no. 6 (372), 45 -- 110; translation in Russian Math. Surveys 61 (2006), no. 6, 1039 -- 1099. [4] A. Connes, Noncommutative Geometry Academic Press, San Diego, 1994. [5] A. Connes, The action functional in noncommutative geometry. Comm. Math. Phys. 117 (1988), no. 4, 673 -- 683. [6] J. Dixmier, Existence de traces non normales (French) C. R. Acad. Sci. Paris Ser. A-B 262 1966 A1107 -- A1108. [7] P. Dodds, B. de Pagter, A. Sedaev, E. Semenov, F. Sukochev, Singular symmetric func- tionals and Banach limits with additional invariance properties. (Russian) Izv. Ross. Akad. Nauk Ser. Mat. 67 (2003), no. 6, 111 -- 136; translation in Izv. Math. 67 (2003), no. 6, 1187 -- 1212. [8] P. Dodds, B. de Pagter, E. Semenov, F. Sukochev, Symmetric functionals and singular traces Positivity 2 (1998), no. 1, 47 -- 75. [9] K. Dykema, T. Figiel, G. Weiss, M. Wodzicki, Commutator structure of operator ideals. IMADA preprint PP-1997 -- 22, May 13, 1997, Odense. [10] K. Dykema, T. Figiel, G. Weiss, M. Wodzicki, Commutator structure of operator ideals. Adv. Math. 185 (2004), no. 1, 1 -- 79. [11] K. Dykema, N. Kalton, Sums of commutators in ideals and modules of type II factors. Ann. Inst. Fourier (Grenoble) 55 (2005), no. 3, 931 -- 971. [12] T. Fack, H. Kosaki, Generalized s-numbers of τ -measurable operators Pacific J. Math. 123 (1986), no. 2, 269 -- 300. [13] T. Figiel, N. Kalton, Symmetric linear functionals on function spaces Function Spaces, Interpolation Theory and Related Topics. Lund, 2000 [14] I. Gohberg, M. Krein, Introduction to the theory of linear nonselfadjoint operators. Trans- lations of Mathematical Monographs, Vol. 18 American Mathematical Society, Providence, R.I. 1969. [15] I. Gohberg, M. Krein, Theory and applications of Volterra operators in Hilbert space. Trans- lations of Mathematical Monographs, Vol. 24 American Mathematical Society, Providence, R.I. 1970. [16] D. Guido, T. Isola, Singular traces on semifinite von Neumann algebras. J. Funct. Anal. 134 (1995), no. 2, 451 -- 485. [17] D. Guido, T. Isola, Dimensions and singular traces for spectral triples, with applications to fractals. J. Funct. Anal. 203 (2003), no. 2, 362 -- 400. [18] N. Kalton, A. Sedaev, F. Sukochev, Fully symmetric functionals on a Marcinkiewicz space are Dixmier traces Adv. Math. 226 (2011) 3540 -- 3549. [19] N. Kalton, F. Sukochev, Symmetric norms and spaces of operators J.reine angew. Math. (2008), 1 -- 41 30 F. SUKOCHEV AND D. ZANIN [20] N. Kalton, F. Sukochev, Rearrangement-invariant functionals with applications to traces on symmetrically normed ideals Canad. Math. Bull. 51 (2008), 67 -- 80. [21] N. Kalton, F. Sukochev, D. Zanin, Orbits in symmetric spaces. II Studia Math. 197 (2010), no. 3, 257 -- 274. [22] S. Krein, Ju. Petunin, E. Semenov, Interpolation of linear operators, Nauka, Moscow, 1978 (in Russian); English translation in Translations of Math. Monographs, Vol. 54, Amer. Math. Soc., Providence, RI, 1982. [23] S. Kwapien, Linear functionals invariant under measure preserving transformations Math. Nachr. 119 (1984), 175 -- 179. [24] J. Lindenstrauss, L. Tzafriri, Classical Banach Spaces II: Function Spaces, Springer, 1996. [25] S. Lord, A. Sedaev, F. Sukochev, Dixmier traces as singular symmetric functionals and applications to measurable operators. J. Funct. Anal. 224 (2005), no. 1, 72 -- 106. [26] M. Reed, B. Simon, Methods of modern mathematical physics. I. Functional analysis. Second edition. Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York, 1980. [27] A. Pietsch, About the Banach envelope of l1 , ∞. Rev. Mat. Complut. 22 (2009), no. 1, 209 -- 226. [28] R. Schatten, Norm ideals of completely continuous operators. Second printing. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 27 Springer-Verlag, Berlin-New York 1970. [29] B. Simon, Trace ideals and their applications. Second edition. Mathematical Surveys and Monographs, 120. American Mathematical Society, Providence, RI, 2005. [30] F. Sukochev, D. Zanin, Orbits in symmetric spaces J.Funct.Anal. 257 (2009), no.1, 194-218. [31] M. Takesaki, Theory of operator algebras. I. Reprint of the first (1979) edition. Encyclopae- dia of Mathematical Sciences, 124. Operator Algebras and Non-commutative Geometry, 5. Springer-Verlag, Berlin, 2002. [32] J. Varga, Traces on irregular ideals. Proc. Amer. Math. Soc. 107 (1989), no. 3, 715 -- 723. [33] M. Wodzicki, Vestigia investiganda. Mosc. Math. J. 2 (2002), no. 4, 769 -- 798, 806. School of Mathematics and Statistics, University of New South Wales, Sydney, 2052, Australia. E-mail address: [email protected] School of Mathematics and Statistics, University of New South Wales, Sydney, 2052, Australia. E-mail address: [email protected]
1811.00100	4	1811	2019-06-02T17:40:58	Pullbacks of graph C*-algebras from admissible pushouts of graphs	[ "math.OA", "math.QA" ]	We define an admissible decomposition of a graph $E$ into subgraphs $F_1$ and $F_2$, and consider the intersection graph $F_1\cap F_2$ as a subgraph of both $F_1$ and $F_2$. We prove that, if the graph $E$ is row finite and its decomposition into the subgraphs $F_1$ and $F_2$ is admissible, then the graph C-algebra $C^(E)$ of $E$ is the pullback C-algebra of the canonical surjections from $C^(F_1)$ and $C^(F_2)$ onto $C^(F_1\cap F_2)$.	math.OA	math	PULLBACKS OF GRAPH C-ALGEBRAS FROM ADMISSIBLE PUSHOUTS OF GRAPHS PIOTR M. HAJAC Instytut Matematyczny, Polska Akademia Nauk ´Sniadeckich 8, 00-656 Warszawa, Poland E-mail: [email protected] and Department of Mathematics University of Colorado Boulder 2300 Colorado Avenue Boulder, CO 80309-0395 USA SARAH REZNIKOFF Mathematics Department, Kansas State University 138 Cardwell Hall, Manhattan, KS 66502 US E-mail: [email protected] MARIUSZ TOBOLSKI Instytut Matematyczny, Polska Akademia Nauk ´Sniadeckich 8, 00-656 Warszawa, Poland E-mail: [email protected] Abstract. We define an admissible decomposition of a graph E into subgraphs F1 and F2, and consider the intersection graph F1 ∩ F2 as a subgraph of both F1 and F2. We prove that, if the graph E is row finite and its decomposition into the subgraphs F1 and F2 is admissible, then the graph C-algebra C∗(E) of E is the pullback C-algebra of the canonical surjections from C∗(F1) and C∗(F2) onto C∗(F1 ∩ F2). 1. Introduction and preliminaries Pushouts of graphs have proven to be very useful in the theory of free groups [11]. We hope that our approach to pullbacks of graph algebras through pushouts of underlying graphs will also turn out to be beneficial. A graph C-algebra is the universal C-algebra associated to a directed graph. If one considers a specific class of morphisms of directed graphs (e.g., see [1, Definition 1.6.2]), then the graph C-algebra construction yields a covariant functor from the category of di- rected graphs to the category of C-algebras. On the other hand, Hong and Szyma´nski [7] showed that a pushout diagram in the category of directed graphs can lead to a pullback of C-algebras. The purpose of this paper is to find conditions on the pushout diagram of 2010 Mathematics Subject Classification. 46L45, 46L55, 46L85. Key words and phrases. graph algebra, pushout and pullback, gauge action, quantum spaces: spheres, balls, lens spaces, weighted projective spaces. 1 graphs that give rise to the pullback diagram of the associated graph C-algebras. This leads to a notion of an admissible decomposition of a directed graph, which we present in Section 2. The main result is contained in Section 3 and examples are in Section 4. Our result is closely related to [8, Corollary 3.4], where it is proven, in an appro- priate form, for k-graphs without sinks but not necessarily row finite. Herein, we focus our attention on row-finite 1-graphs but possibly with sinks. Thus our results are com- plementary and lead to the following question: Is it possible to get rid of both of these assumtions ("row finite" and "no sinks") at the same time to prove a more general pushout-to-pullback theorem? In this paper, by a graph E we will always mean a directed graph, i.e. a quadruple (E0, E1, sE, rE), where E0 is the set of vertices, E1 is the set of edges, sE : E1 → E0 is the source map and rE : E1 → E0 is the range map. A graph E is called row finite if each vertex emits only a finite number of edges. Next, E is called finite if both E0 and E1 are finite. A vertex is called a sink if it does not emit any edge. By a path µ in E of length µ = k > 0 we mean a sequence of composable edges µ = e1e2 . . . ek. We treat vertices as paths of length zero. The set of all finite paths for a graph E is denoted by Path(E). One extends the source and the range maps to Path(E) in a natural way. We denote the extended source and range maps by sPE and rPE, respectively. Definition 1.1. The graph C-algebra C∗(E) of a row-finite graph E is the universal C-algebra generated by mutually orthogonal projections P :=(cid:8)Pv v ∈ E0(cid:9) and partial isometries S :=(cid:8)Se e ∈ E1(cid:9) satisfying the Cuntz -- Krieger relations [4]: (cid:88) S∗ e Se = PrE (e) SeS∗ e = Pv for all e ∈ E1, and for all v ∈ E0 that are not sinks. (CK1) (CK2) e∈s−1 E (v) The datum {S, P} is called a Cuntz -- Krieger E-family. One can show that the above relations imply the standard path-algebraic relations: (1.1) S∗ f Se = 0 for e (cid:54)= f, PsE (e)Se = Se = SePrE (e) . Any graph C-algebra C∗(E) can be endowed with a natural circle action (1.2) α : U (1) −→ Aut(C∗(E)) defined by its values on the generators: (1.3) αλ(Pv) = Pv , αλ(Se) = λSe , where λ ∈ U (1), v ∈ E0, e ∈ E1. The thus defined circle action is called the gauge action. A subset H of E0 is called hereditary iff, for any v ∈ H such that there is a path starting at v and ending at w ∈ E0, we have w ∈ H. (Note that we can equivalently define the property of being hereditary by replacing "path" with "edge".) A subset H of E0 is called saturated iff there does not exist a vertex v /∈ H such that (1.4) 0 < s−1 E (v) < ∞ and 2 rE(s−1 E (v)) ⊆ H. Saturated hereditary subsets play a fundamental role in the theory of gauge-invariant It follows from [2, Lemma 4.3] that, for any hereditary ideals of graph C-algebras. subset H, the algebraic ideal generated by {Pv v ∈ H} is of the form IE(H) = span(cid:8)SxS∗ y x, y ∈ Path(E) , rPE(x) = rPE(y) ∈ H(cid:9) . (1.5) Here, for any path µ = e1 . . . ek, we adopt the notation Sµ := Se1 . . . Sek. Furthermore, if µ is a vertex, then Sµ := Pµ. By [2, Theorem 4.1 (b)], quotients by closed ideals generated by saturated hereditary subsets can also be realised as graph C-algebras by constructing a quotient graph. Given a hereditary subset H of E0, the quotient graph E/H is given by (E/H)0 := E0 \ H (E/H)1 := E1 \ r−1 (1.6) Note that the restriction-corestriction of the range map rE to (E/H)1 → (E/H)0 makes sense for any H, but the same restriction-corestriction of the source map sE exists because H is hereditary. E (H). and Moreover, if H is also saturated, we obtain the -isomorphism (1.7) C∗(E)/IE(H) ∼= C∗(E/H), where IE(H) is the norm closure of IE(H). 2. Admissible decompositions of graphs Given two graphs E = (E0, E1, sE, rE) and G = (G0, G1, sG, rG), one can define a graph morphism f : E → G as a pair of mappings f 0 : E0 → G0 and f 1 : E1 → G1 satisfying (2.1) sG ◦ f 1 = f 0 ◦ sE and rG ◦ f 1 = f 0 ◦ rE . We call the thus obtained category the category of directed graphs. A subgraph of a graph E = (E0, E1, sE, rE) is a graph F = (F 0, F 1, sF , rF ) such that (2.2) F 0 ⊆ E0, F 1 ⊆ E1, ∀ e ∈ F 1 : sF (e) = sE(e) and rF (e) = rE(e). Next, let F1 and F2 be two subgraphs of a graph E. We define their intersection and union as follows: F1 ∩ F2 := (F 0 1 ∩ F 0 2 , F 1 1 ∩ F 1 2 , s∩, r∩), ∀ e ∈ F 1 1 ∩ F 1 2 : s∩(e) := sE(e), r∩(e) := rE(e), F1 ∪ F2 := (F 0 1 ∪ F 0 2 , F 1 1 ∪ F 1 2 , s∪, r∪), ∀ e ∈ F 1 1 ∪ F 1 2 : s∪(e) := sE(e), r∪(e) := rE(e). (2.3) To consider pushout diagrams in the category of directed graphs, we follow the con- vention used in [5]. If a graph E has two subgraphs F1 and F2 such that E = F1 ∪ F2, 3 then the following diagram E (2.4) F1 F2 F1 ∩ F2 is automatically a pushout diagram. Let us illustrate the concept of a pushout diagram of graphs with the following example: (2.5) We are now ready to define an admissible decomposition of a row-finite graph: Definition 2.1. An unordered pair {F1, F2} of subgraphs of a row-finite graph E is called an admissible decomposition of E iff the following conditions are satisfied: . (1) E = F1 ∪ F2 , (2) if v is a sink in F1 ∩ F2 , then v is a sink in Fi , i = 1, 2, 2 = r−1 (3) F 1 2 ), i = 1, 2. 1 ∩ F 0 1 ∩ F 1 (F 0 Fi Note that, by (1) in Definition 2.1, E is a pushout of F1 and F2 over their intersection. Observe also that Diagram (2.5) gives an example of an admissible decomposition of a graph. Definition 2.1 prompts the following two lemmas. Lemma 2.2. Let {F1, F2} be an admissible decomposition of a row-finite graph E. Then F1 ∩ F2 = Fi/(F 0 i ), for i = 1, 2. 1 ∩ F 0 i \ (F 0 2 )) and Fi = E/(E0 \ F 0 i \ (F 0 2 implies e /∈ F 1 1 ∩ F 0 1 ∩F 1 Proof. First, note that F 0 sFi(e) /∈ F 0 1 ∩F 0 2 ) is hereditary in Fi. Indeed, take e ∈ F 1 2 . Hence, by Definition 2.1(3), we have rFi(e) /∈ F 0 i . Then 1 ∩F 0 2 . 4 ; ; c c c c ; ; = = a a : : d d Therefore, we can define Fi/(F 0 nition 2.1(3). i \ (F 0 1 ∩ F 0 2 )), which coincides with F1 ∩ F2 due to Defi- Next, note that i = (F 0 E0 \ F 0 i = F 0 (2.6) where j (cid:54)= i and j = 1, 2, so we already know that E0 \ F 0 that it is hereditary in E, we only need to exclude edges starting in E0 \ F 0 in E0 \ F 0 j . They do not exist because E1 = F 1 j , so E0 \ F 0 i ∩ F 0 j ), is hereditary in Fj. To see i and ending is hereditary in E. j ) \ F 0 j \ (F 0 i ∪ F 0 i ∪ F 1 j \ F 0 i = F 0 i i E (F 0 i ). To this end, taking advantage of the admis- It remains to verify that F 1 i = r−1 sibility of (Fi ∩ Fj) ⊆ Fi, we compute i ) \ F 1 (2.7) r−1 E (F 0 (F 0 i = r−1 (F 0 i ) \ F 1 i ⊆ r−1 i = r−1 E (F 0 Fj i ), we conclude that F 1 Fj (cid:4) Therefore, as F 1 Lemma 2.3. Let {F1, F2} be an admissible decomposition of a row-finite graph E. Then i \ (F 0 F 0 1 ∩ F 0 i ∩ F 0 i = r−1 j ) \ F 1 E (F 0 i = (F 1 i ∩ F 1 i ), as desired. j ) \ F 1 i = ∅. 2 ) is saturated in Fi and in E for i = 1, 2. i \ (F 0 1 ∩ F 0 2 ) were not saturated in Fi, then there would exist a vertex v in i \ (F 0 F 0 i \ (F 0 1 ∩ F 0 2 )) = F 0 1 ∩ F 0 2 Proof. If F 0 (2.8) such that s−1 (v) (cid:54)= ∅ and rFi(s−1 (v)) ⊆ F 0 i \ (F 0 1 ∩ F 0 2 ). (2.9) Thus we would have a vertex in F1 ∩ F2 that is a sink in F1 ∩ F2 but not in Fi, which contradicts Definition 2.1(2). Fi Fi Much in the same way, if F 0 i \ (F 0 1 ∩ F 0 2 ) were not saturated in E, then there would exist a vertex w ∈ E0 \ (F 0 i \ (F1 ∩ F 0 (2.10) where j (cid:54)= i and j = 1, 2, such that E (w) (cid:54)= ∅ and rE(s−1 s−1 (2.11) Hence, there is e ∈ s−1 E (w) such that rE(e) /∈ F 0 e ∈ F 1 again contradicts Definition 2.1(2). i , so w = sE(e) ∈ F 0 E (w)) ⊆ F 0 i \ (F 0 1 ∩ F 0 2 ). i ∪ F 1 j . As E1 = F 1 j , it follows that i . Consequently, w is a sink in F1 ∩ F2 but not in Fi, which (cid:4) 2 )) = F 0 j , 3. Pullbacks of graph C-algebras Let {F1, F2} be an admissible decomposition of a row-finite graph E. Then, by Lemma 2.2 and Lemma 2.3, we can take an advantage of the formula (1.7) to define the canonical quotient maps: (3.1) (3.2) (3.3) π1 : C∗(E) −→ C∗(E)/IE(F 0 π2 : C∗(E) −→ C∗(E)/IE(F 0 1 \ (F 0 χ1 : C∗(F1) −→ C∗(F1)/IF1(F 0 2 \ F 0 1 \ F 0 1 ∩ F 0 1 ) ∼= C∗(F1), 2 ) ∼= C∗(F2), 2 )) ∼= C∗(F1 ∩ F2), 5 (3.4) χ2 : C∗(F2) −→ C∗(F2)/IF2(F 0 2 \ (F 0 1 ∩ F 0 2 )) ∼= C∗(F1 ∩ F2). Note that quotient maps are automatically U (1)-equivariant for the gauge action. This brings us to the main theorem: Theorem 3.1. Let {F1, F2} be an admissible decomposition of a row-finite graph E. Then there exist canonical quotient gauge-equivariant ∗-homomorphisms rendering the following diagram (3.5) C∗(F1) π1 χ1 C∗(E) C∗(F1 ∩ F2) π2 χ2 C∗(F2) commutative. Moreover, this is a pullback diagram of U (1)-C-algebras. Proof. Note first that all the canonical surjections in the diagram are well defined due to the admissibility conditions of the decomposition of the graph E (see the discussion at the beginning of this section). The commutativity of the diagram is obvious as all maps are canonical surjections. Finally, using [9, Proposition 3.1] and the surjectivity of χ1 and χ2, to prove that (3.5) is a pullback diagram, it suffices to show that ker π1 ∩ ker π2 = {0} and that π2(ker π1) ⊆ ker χ2. Since ker π1 and ker π2 are closed ideals in a C-algebra, we know that (3.6) ker π1 ∩ ker π2 = ker π1 ker π2. 1 \ F 0 Next, as F 0 that 2 and F 0 2 \ F 0 1 are saturated hereditary subsets of E0, it follows from (1.7) (3.7) ker π1 = IE(F 0 2 \ F 0 1 ) and ker π2 = IE(F 0 1 \ F 0 2 ). Furthermore, using the characterization (1.5) of ideals generated by hereditary subsets, we know that an arbitrary element of ker π1 ker π2 is in the closed linear span of elements of the form SαS∗ βSγS∗ δ , where α, β ∈ Path(E) with rPE(α) = rPE(β) ∈ F 0 2 \ F 0 1 , (3.8) and γ, δ ∈ Path(E) with (3.9) The conclusion ker π1 ∩ ker π2 = {0} follows from the analysis of all possible paths satis- fying the above conditions. rPE(γ) = rPE(δ) ∈ F 0 1 \ F 0 2 . Indeed, it follows from (1.1) that S∗ 2 , rPE(β) ∈ F 0 2 \ F 0 βSγ (cid:54)= 0 is possible only if sPE(β) = sPE(γ). As 2 , if β = e1 . . . em and γ = f1 . . . fn, 1 \ F 0 1 and rPE(γ) ∈ F 0 1 ∪ F 1 E1 = F 1 we infer that (3.10) rE(em−1) = sE(em) ∈ F 0 2 and rE(fn−1) = sE(fn) ∈ F 0 1 . 6 x x & & & & x x 1 ∩ F 0 2 or rE(em−1) ∈ F 0 Hence rE(em−1) ∈ F 0 1 . Now, we continue by induction using Definition 2.1(3) for the intersection case of the alternative. This brings us to conclusion that sPE(β) ∈ F 0 1 . It follows that sPE(β) = sPE(γ) ∈ F 0 1 \ F 0 2 , we conclude that β (cid:54)= γ, so there exists the smallest index i such that ei (cid:54)= fi. Now, remembering the relations (CK1) and (1.1), we compute 2 . Much in the same way, we argue that sPE(γ) ∈ F 0 2 . Furthermore, as rPE(β) ∈ F 0 1 and rPE(γ) ∈ F 0 1 ∩ F 0 2 \ F 0 2 \ F 0 S∗ βSγ = S∗ = S∗ ei+1...emS∗ ei+1...emS∗ ei ei S∗ ei−1 . . . S∗ e1Se1 . . . Sei−1SfiSfi+1...fn SfiSfi+1...fn (3.11) = 0. Finally, if β or γ is a path of length zero, i.e. a vertex, then it is straightforward to conclude that S∗ βSγ = 0. Next, taking again an advantage of (1.5) and (1.7), we obtain ker χ2 = IF2(F 0 β α, β ∈ Path(F2), rPF2(α) = rPF2(β) ∈ F 0 2 \ F 0 1 ) = span{SαS∗ 2 \ F 0 Any element of IF2(F 0 1 ) is an element of IE(F 0 2 \ F 0 α ∈ Path(F2). Hence π2(IE(F 0 we conclude that π2(ker π1) ⊆ ker χ2. Remark 3.2. One can also prove Theorem 3.1 in the setting of Leavitt path algebras [1]. A proof of the Leavitt version of Theorem 3.1 is completely analogous due to [1, Corol- lary 2.5.11]. 2 \ F 0 1 }. 1 ), and π2(Sα) = Sα for all 1 ). Finally, from the continuity of π2, (cid:4) 1 )) ⊆ IF2(F 0 2 \ F 0 2 \ F 0 4. Examples We end the paper by providing motivating examples from noncommutative topology. 4.1. Even quantum spheres. Not only the graph at the top of the diagram (2.5) representing the generic Podle´s quantum sphere [10] admits a natural admissible decom- position, but also the finite graphs L2n [6, Section 5.1] representing, respectively, the C-algebras C(S2n q ) of all even quantum spheres enjoy natural admissible decomposi- tions {F 1 2n) = C∗(F 2 q ) of the Hong-Szyma´nski quantum 2n-ball [7, Section 3.1], and C∗(F 1 2n) coincides [6, Ap- pendix A] with the C-algebra C(S2n−1 ) of the boundary Vaksman-Soibelman quantum odd sphere [12]. Thus we recover in terms of graphs the classical fact that an even sphere is a gluing of even balls over the boundary odd sphere. 2n) coincides with the C-algebra C(B2n 2n}. Here C∗(F 1 2n ∩ F 2 2n, F 2 q As Theorem 3.1 applies, we infer that the diagram (4.1) C(B2n q ) π1 χ1 C(S2n q ) C(S2n−1 q ) 7 π2 χ2 C(B2n q ) y y % % % % y y is a pullback diagram. This fact was already proved in [7, Proposition 5.1] by direct considerations of generators and relations. The case n = 3 is illustrated by the diagram: (4.2) . 4.2. Quantum lens space L3 space L3 q(l; 1, l) can be viewed as the graph C-algebra (e.g., see [3]) of the graph L3 l : q(l; 1, l). The C-algebra C(L3 q(l; 1, l)) of the quantum lens v0 0 ··· v1 0 v1 1 v1 l−2 v1 l−1 . (4.3) The graph L3 yielding, by Theorem 3.1, the pullback diagram: l enjoys an admissible decomposition {L3 l−k}, where k ∈ {1, . . . l − 1}, k, L3 (4.4) C(L3 q(k; 1, k)) π1 χ1 C(L3 q(l; 1, l)) π2 C(L3 q(l − k; 1, l − k)) χ2 C(S1) . 8 : : d d 9 9 e e w w ) ) ( ( u u Recall that C∗(L3 1) ∼= C(S3 q ), so for l = 2 we obtain the following pullback diagram: (4.5) C(S3 q ) C(L3 q(2; 1, 2)) π1 χ1 C(S1) . π2 χ2 C(S3 q ) Since the above diagram is U (1)-equivariant, U (1)-fixed-point subalgebras: it induces a pullback diagram for (4.6) C(CP 1 q ) C(WP 1 q (1, 2)) π1 χ1 wC . π2 χ2 C(CP 1 q ) q ) and C(WP 1 Here C(CP 1 q (1, 2)) denote the quantum complex projective space (see [6, Section 2.3]) and the quantum weighted projective space (see [3, Section 3]), respectively. Interestingly, the C-algebras in the above diagram can be viewed as graph C-algebras, and an infinite graph representing C(WP 1 q (1, 2)) is a pushout of infinite graphs repre- q ) over the graph consisting of one vertex and no edges representing C senting C(CP 1 (see Diagram (4.7) below). However, this example is beyond the scope of Theorem 3.1, because the above diagram is no longer U (1)-equivariant and the infinite graphs are not row finite. (∞) (∞) (4.7) (∞) (∞) Here edges with (∞) denote countably infinitely many edges. 9 x x & & & & x x w w ' ' ' ' w 6 6 h h 5 5 i i Acknowledgements The work on this project was partially supported by NCN-grant 2015/19/B/ST1/03098 (Piotr M. Hajac, Mariusz Tobolski) and by a Simons Foundation Collaboration Grant (Sarah Reznikoff). Piotr M. Hajac is very grateful to Kansas State University for its hos- pitality and financial support provided by this Simons Foundation Collaboration Grant. It is a pleasure to thank Carla Farsi for drawing our attention to using pushouts of graphs in the theory of free groups, Tatiana Gateva -- Ivanova for a helpful discussion concerning admissible decompositions of graphs, and Aidan Sims for making us aware on how to prove our main result in a different context. References [1] G. Abrams, P. Ara, M. Siles Molina, Leavitt path algebras. Lecture Notes in Mathematics, 2191. Springer, London, 2017. [2] T. Bates, D. Pask, I. Raeburn, W. Szyma´nski, The C-algebras of row-finite graphs, New York J. Math., 6 (2000), 307 -- 324. [3] T. Brzezi´nski, W. Szyma´nski, The C-algebras of quantum lens and weighted projective spaces, J. Noncomm. Geom., 12 (2018), 195 -- 215. [4] J. Cuntz, W. Krieger, A class of C∗-algebras and topological Markov chains, Invent. Math., 56 (1980), 251 -- 268. [5] H. Ehrig, H.-J. Kreowski, Pushout-properties: an analysis of gluing constructions for graphs. Math. Nachr., 91 (1979), 135 -- 149. [6] J. H. Hong, W. Szyma´nski, Quantum spheres and projective spaces as graph algebras, Comm. Math. Phys., 232 (2002), 157 -- 188. [7] J. H. Hong, W. Szyma´nski, Noncommutative balls and mirror quantum spheres, J. Lond. Math. Soc., 77 (2008), 607 -- 626. [8] A. Kumjian, D. Pask, A. Sims, M. F. Whittaker. Topological spaces associated to higher-rank graphs. J. Combin. Theory Ser., A 143 (2016), 19 -- 41. [9] G. K. Pedersen, Pullback and Pushout Constructions in C-Algebra Theory, J. Funct. Anal., 167 (1999), 243 -- 344. [10] P. Podle´s, Quantum spheres, Lett. Math. Phys. 14 (1987), no. 3, 193 -- 202. [11] J. R. Stallings. Topology of finite graphs. Invent. Math., 71 (1983), no. 3, 551 -- 565. [12] L. L. Vaksman, Ya. S. Soibelman, Algebra of functions on the quantum group SU(n + 1) and odd-dimensional quantum spheres, Algebra i Analiz, 2 (1990), 101 -- 120. 10
1209.4094	3	1209	2016-06-09T20:17:12	Construction of enveloping actions	[ "math.OA", "math.DS" ]	We study the problem of constructing a globalization for partial actions on -algebras, C-algebras and Hilbert modules. For the first ones we give a necessary condition for the existence of a globalization and we prove this conditions is necessary and sufficient for C*-algebras. Using the linking algebra of a Hilbert module we translate this condition to the realm of partial action on Hilbert modules.	math.OA	math	CONSTRUCTION OF ENVELOPING ACTIONS DAMI ´AN FERRARO Abstract. We study the problem of constructing a globalization for partial actions on -algebras, C-algebras and Hilbert modules. For the first ones we give a necessary condition for the existence of a globalization and we prove this conditions is necessary and sufficient for C-algebras. Using the linking algebra of a Hilbert module we translate this condition to the realm of partial action on Hilbert modules. 1. Introduction Among the simplest examples of partial actions on C-algebras [5, 6, 7] we find restrictions of actions (also called global actions) to non invariant C-ideals [1]. Many aspects of these examples, as the representations and crossed products, can be studied using techniques developed for actions on C-algebras. Then it is interesting to know which partial actions can be globalized, that is: described as the restriction of a global action (the globalization). This problem was stated in [1], where it was solved in case the C-algebra is commutative. Partial actions can be defined in other categories, as topological spaces, rings and Hilbert modules [1, 2, 3, 4]. Here we will work with -algebras, C-algebras and Hilbert modules, mainly because C-algebras are, at once, -algebras and Hilbert modules. This work is organized as follows. In the first section we study the problem of globalizing partial actions on -algebras (-partial actions). Our intention is to study continuous partial action on C-algebras (C-partial actions) from a - algebraic point of view, which amounts to dump all the topological structure keep- ing the -algebraic properties we need. Hence, when necessary, we make some assumptions on the -algebras which are known to hold for C-algebras. Under these assumptions we give a necessary and sufficient condition for the existence of a globalization. Then we turn to consider all the structure of C-partial actions. At this point we show a C-partial action can be globalized to a C-partial action if and only if it can be globalized to a -partial action, independently of the topological structure. Finally, in the last section, we give a necessary and sufficient condition for the existence of a globalization of a partial action on a Hilbert bimodule (Hb- partial action). We specifically show that a Hb-partial action has a globalization if and only if it's linking partial action [1] has a globalization. Date: September 3, 2018. 2010 Mathematics Subject Classification. Primary 46L55. Secondary 46L40, 46L05. Key words and phrases. Partial actions, Enveloping actions. This work was partially supported by Mathamsud network U11-MATH05 (partially funded by ANII, Uruguay) and started during my visit to the Universidade Federal de Santa Catarina. I thank Professors Alcides Buss and Ruy Exel for receiving me there. 1 2 DAMI ´AN FERRARO 2. Partial action on -algebras Each one of the types of partial actions considered here has it's own notion of globalization. Besides, a partial action on a C-algebra is a partial action in a - algebra and in a Hilbert module. For this reason we introduce the terms "-partial action", "C-partial action" and "Hb-partial action". In case the partial action under consideration is global we substitute the term partial for global. We start by recalling some definitions and facts. 2.1. Algebras with involution. A -algebra is an algebra A over the complex field together with a conjugate linear function (the involution) A → A, a 7→ a∗, satisfying (ab)∗ = b∗a∗ and a∗∗ = a. Assume A is a -algebra. By a -ideal of A we mean a subspace I ⊂ A such that I ∗ = I and IA ⊂ A. A function between -algebras, φ : A → B, is a - homomorphism if it is linear, multiplicative and φ(a∗) = φ(a)∗. An automorphism of A is a bijective -homomorphism from A to A and the set of automorphism of A will be denoted Aut(A). An element a ∈ A is a right (left) annihilator if aA = {0} (Aa = {0}). Note a is a right annihilator if and only if a∗ is a left annihilator. We say A is non-degenerate if it does not have a right (or left) annihilator different from 0. Remark 2.1. If A is non-degenerate then for all a, b ∈ A the following conditions are equivalent: (1) a = b, (2) for all c ∈ A, ac = bc (3) for all c ∈ A, ca = cb and (4) for all c, d ∈ A, cad = cbd. A double centraliser of A is a pair µ = (L, R), where L, R : A → A are linear functions, L(ab) = L(a)b, R(ab) = aR(b) and aL(b) = R(a)b (for all a, b ∈ A). It is usual to write µa := L(a) and aµ := R(a). The -algebra structure of M (A) is given by point wise vector space operations, product (L, R)(M, S) := (L◦ M, S◦ R) and involution (L, R)∗ := (L′, R′), where L′(a) = R(a∗)∗ and R′(a) = L(a∗)∗. The function τ : A → M (A), where τ (a)b = ab and bτ (a) = ba, is -homomorphism which is injective if and only if A is non-degenerate. Remark 2.2. Double centralisers of -algebras can be treated as adjointable opera- tors of Hilbert modules. More precisely, if A is non-degenerate then (1) Given a function T : A → A there exists (L, R) ∈ M (A) such that T = L if and only if there exists a function T ∗ : A → A, with T (a)∗b = a∗T ∗(b) for all a, b ∈ A. (2) Given (L1, R1), (L2, R2) ∈ M (A) we have L1 = L2 if and only if R1 = R2. Take any group G and name AG the -algebra of functions from G to A (point wise operations). Given t ∈ G name θt the automorphism of AG defined by θt(f )r := frt. The canonical action of G on M (AG) is Θ : G → Aut(M (AG)) where Θt(L, R) = (θt ◦ L ◦ θt−1, θt ◦ R ◦ θt−1). 2.2. Partial actions and globalizations. Recall [7, 6] that a partial action of a group G on a set X is a pair σ := ({Xt}t∈G,{σt}t∈G) where (1) For all t ∈ G, Xt is a subset of X and σt : Xt−1 → Xt a function. (2) σe is the identity of X. (3) For all s, t ∈ G, if x ∈ Xt−1 and σt(x) ∈ Xs−1 , then x ∈ Xt−1s−1 and σs(σt(x)) = σst(x). CONSTRUCTION OF ENVELOPING ACTIONS 3 In case Xt = X for all t ∈ G, σ is said to be global and it is just a common action of G on X. Let us assume that σ and τ are partial actions of G on the sets X and Y, respectively. We say f : σ → τ is a morphism of partial actions on sets if f is a function from X to Y and for all t ∈ G : f (Xt) ⊂ Yt and f (σt(x)) = τt(f (x)) (for all x ∈ Xt−1). The identity morphism associated to σ is idσ := idX and the composition of morphism is just the composition of functions. With σ as before take a set U ⊂ X. Given t ∈ G define Ut := U ∩ σt(Xt−1 ∩ U ). It is obvious that Ut−1 ⊂ Xt−1 and σt(Ut−1) ⊂ Ut; hence it makes sense to define κt : Ut−1 → Ut as κt(x) = σt(x). Straightforward arguments imply the restriction of σ to U , defined as σU := ({Ut}t∈G,{κt}t∈G) , is partial action of G on U. From [1] we know every partial action on a set is isomorphic to the restriction of a global action on a set. A set U ⊂ X is said to be σ−invariant if for all t ∈ G, σt(Xt−1 ∩ U ) ⊂ U. Remark 2.3. If U ⊂ V ⊂ X then σU = σV U . Besides, if σ is global, σU is global if and only if U is σ−invariant. Definition 2.4. A -partial action of the (discrete) group G on the -algebra A is a set theoretic partial action of G on A, α = ({At}t∈G,{αt}t∈G) , such that At is a -ideal of A and αt a -homomorphism (for all t ∈ G). Example 2.5. Let α be a -global action of G on the -algebra A and let I be a -ideal of A. The restriction of α to I, αI := (cid:0){I ∩ αt(I)}t∈G,{αtI∩αt(I)}t∈G(cid:1) , is a -partial action of G on I. Example 2.6. Let A be a ∗−algebra and I a -ideal of it. The -partial action of Z2 = {0, 1} on A determined by I is αAI := ({A0, A1},{α0, α1}) where A0 = A, A1 = I, α0 = idA and α1 = idI . Morphisms of -partial actions are just morphisms of partial actions on set which are also -homomorphisms. The composition and identity are the natural ones. Definition 2.7 ([1, 3]). A -globalization the -partial action α (of G on A) is a 4-tuple Ξ = (B, β, I, ι) where: B is a -algebra, β is a -global action of G on B, I is a -ideal of B and ι : α → βI is an isomorphism. we say Ξ is non degenerate if B is non degenerate. We say Ξ is minimal if [I] := span{βt(I) : t ∈ G} equals B and, for convenience, Do not confuse our concept of minimality with topological minimality, here β may have many open invariant sets1. Note that Ξ is minimal iff the only β−invariant In case Ξ is not minimal, ([I], β[I], I, ι) is -ideal of B containing I is B itself. minimal because [I] is invariant and β global (Remark 2.3). Example 2.8. Given a complex vector space V and a conjugate linear bijection T : V → V with T 2 = idV , let V T be the -algebra obtained by considering on V the null product (uv = 0) and T as involution. Here we consider the entry wise conjugation T : Cn → Cn. Let α be the partial action of Z3 = {0, 1, 2} (with additive notation) on CT such that α0 = idCT and α1 = α2 = id{0}. There are two completely different globalizations for α, (C2T , β, I, ι) and (C3T , γ, J, κ), that we 1Note that there is no topology involved. 4 DAMI ´AN FERRARO now describe. Just set I = {(u, 0) : u ∈ C}, J = {(u, 0, 0) : u ∈ C}, ι(u) = (u, 0), κ(u) = (u, 0, 0), β1(u, v) = − 1 2 (u + √3v,− √3u + v) and γ1(u, v, w) = (w, u, v). Note β1 is the rotation by angle 2π/3 and a simple dimension argument implies that β is not isomorphic to γ. In fact it can be shown that these are the unique globalizations of α (up to isomorphism). Example 2.9. Consider the partial action of Example 2.6. Assume there exists a - ideal J of A such that A = I⊕J. Now form the (external) direct sum B := I⊕J ⊕J, considered as a -algebra with entry wise operations. Let β be the action of Z2 on B given by β1(a ⊕ b ⊕ c) = a ⊕ c ⊕ b and ι : A → B the unique linear map such that ι(a + b) = a ⊕ b, for all a ∈ I and b ∈ J. Then (B, β, I ⊕ J ⊕ 0, ι) is a minimal -globalization of αAI . The existence of the direct complement J is not necessary for the existence of a globalization, as we show with the next Example. Moreover, with it we also show that there are -partial actions on non degenerate -algebras with a -globalization but without a non degenerate -globalization. Example 2.10. Let U ∈ M4(C) be the matrix corresponding to the permutation (1 4)(2 3) (written as a product of cycles). Define an involution in M4(C) by the formula a∗ := uatu, where a 7→ a is the entry wise complex conjugation and a 7→ at is the usual matrix transposition2. Note that a∗ is obtained from a by performing a symmetry with respect to the anti-diagonal3. Name A the ∗−subalgebra of M4(C) formed by the matrices of the form   a11 a12 0 0 0 0 0 0 a13 a14 a23 a24 a34 0 0 a44   With I := {a ∈ A : a23 = 0} we have I = I ∗ and AA ⊂ I. Then I is a -ideal of A and it can be shown that A and I are non-degenerate. Now let αAI be the -partial action described in Example 2.6. Assume Ξ = (B, β, J, ι) is a -globalization of αAJ . For convenience we think J = A, ι = idA and α = βJ . Since B = A + β(A) and dim(A∩ β(A)) = dim(I) = 7, we know dim(B) = dim(A) + 1 = 9. Let u ∈ A be the matrix with 1 in the entry 2− 3 and 0 elsewhere. Then A = I ⊕ Cu and β1(u) /∈ A because u /∈ I = A∩ β1(A). Moreover, B = A⊕ Cβ1(u). As a vector space B is isomorphic to the external direct sum A⊕ C and, with this notation, the involution of B is (a⊕ λ)∗ = a∗⊕ λ (because u∗ = u). To describe the product of B we compute (a + λβ1(u))(b + µβ1(u)) = ab + µaβ1(u) + λβ1(u)b + λµβ1(u2). Firstly note that u2 = 0, so β1(u2) = 0. Secondly, aβ1(u) ∈ I and for all c ∈ I we have (aβ1(u) − au)c = aβ1(u)α1(c) − auc = aα1(uc) − auc = 0. Since I is non- degenerate and au ∈ I, aβ1(u) = au. In the same way we deduce that β1(u)b = ub. 2With this structure M4(C) is not a C-algebra because a∗a = 0 if ai,j = δ1(i)δ1(j). 3From lower left corner to upper right corner. CONSTRUCTION OF ENVELOPING ACTIONS 5 Now the product of B takes the form (a + λβ1(u))(b + µβ1(u)) = ab + µau + λub. Then the formula for the product of A ⊕ C should be (a ⊕ λ)(b ⊕ µ) := ab + µau + λub ⊕ 0. In fact A ⊕ C is a -algebra with this product and the involution described in the previous paragraph. Moreover, it is degenerate because u ⊕ −1 is a bilateral annihilator. The final step is to determine β, which amounts to give an expression for β1. Assume b = a+λu+µβ1(u), with a ∈ I and λ, µ ∈ C. Then β1(b) = β1(a)+λβ1(u)+ µu = a + µu + λβ1(u). In terms of A ⊕ C β1 should be given by β1((a + λu) ⊕ µ) = (a ⊕ µu) ⊕ λ. Now the reader can verify that β1 is in fact a -homomorphism of A ⊕ C with β2 1 = idA⊕C and β1A = α. If we set ι : A → A ⊕ C as the natural inclusion then (A ⊕ C, β, A ⊕ 0, ι) is a minimal -globalization of α. Furthermore, it is essentially the unique -globalization of α. The algebras from Example 2.8 are those in which all the elements are annihila- tors. Our source of inspiration are partial action on C-algebras and these algebras are non-degenerate because on such algebras the identity aa∗ = 0 implies a = 0. Then we will assume our -algebras are non degenerate when needed. Now our idea is to take a -partial action with a -globalization and try to construct another -globalization using just the -partial action. In this way we will be sure that the -partial action has a -globalization every time the construction can be performed. This new -globalization may be completely different from the initial one. Fix, for the rest of this section, a ∗−partial action α of G on A and a minimal -globalization Ξ = (B, β, I, ι) of α. The canonical morphism associated to Ξ, denoted π or πΞ in case is necessary to mention Ξ, is the -homomorphism π : B → M (AG), π(b)fr = ι−1(βr(b)ι(fr)). A simple computation shows that π : β → Θ is a morphism, where Θ is the canonical action of G on M (AG). Then π(B) is a Θ-invariant -subalgebra and π(A) is a -ideal of it. Moreover, Θπ(B) is a -global action and Θπ(B)π(I) = Θπ(I). The quadruple ΞΘ := (π(B), Θπ(B), π(I), π ◦ ι) is a minimal -globalization of α if and only if π ◦ ι is an injective function and π(It) = π(I) ∩ Θt(π(I)) (for all t ∈ G). Unfortunately these two conditions seem to be unrelated in general. Example 2.11. Let αAI be the -partial action from Example 2.10 and (A⊕C, β, A⊕ 0, ι) it's -globalization. Then AZ2 = A⊕A and π(a⊕λ)(b⊕c) = (a+λu)b⊕(a+λu)c. Since π(a⊕ λ) = π(a+ λu⊕ 0), we have π(A⊕ C) = π(A) and π is not injective. But π ◦ ι is injective because A is non degenerate. In this case the kernel of π, ker(π), is exactly the space generated by u ⊕ −1, which is the set of right annihilators of A ⊕ C, AnnR(A ⊕ C). This last condition is not accidental, as we now show. Lemma 2.12. Let Ξ = (B, β, I, ι) be a -globalization of the -partial action α (of G on A). If π is the canonical morphism associated to Ξ then ker(π) = AnnR(B). Moreover, if A is non-degenerate then π◦ ι is injective and π(B) is non degenerate. Proof. Assume b ∈ ker(π). Given a ∈ A and r ∈ G denote bδr the element of AG taking the value b at r and 0 elsewhere. Then 0 = ι(π(b)aδrr) = βr(b)ι(a). This implies bB = span bβr(ι(A)) = 0, so b ∈ AnnR(B). For the converse assume 6 DAMI ´AN FERRARO b ∈ AnnR(B) and take f ∈ AG. Then for all r ∈ G we have βr(b) ∈ AnnR(B), thus π(b)fr = ι−1(βr(b)ι(fr)) = 0 and b ∈ ker(π). In case A is non-degenerate and a ∈ ker(π ◦ ι), we have ι(a) ∈ AnnR(B). This implies a ∈ AnnR(A) and, so, a = 0. Finally assume π(b) ∈ AnnR(π(B)). Since π(B) is Θ−invariant, for all r ∈ G we have π(βr(b)) = Θr(π(b)) ∈ AnnR(π(B)). Then, for all r ∈ G and c, d ∈ A, 0 = π(βr(b))π(ι(c))dδee = ι−1(βr(b)ι(c)ι(d)). This implies 0 = βr(b)ι(c) for all r ∈ G and c ∈ A. Using the definition of π(b) we conclude that π(b) = 0. (cid:3) To give a sufficient condition for π ◦ ι to be an isomorphism we introduce as- similative ideals. A subset S of the -algebra C is assimilative (in C) if given c ∈ C such that cC ⊂ S, then c ∈ J. It is immediate that every subset of a unital -algebra is assimilative. Furthermore, a -ideal J of C is assimilative iff C/J is non-degenerate. Remark 2.13. Recall [4] that a ring R is s-unital if for all a ∈ R there exists b ∈ R such that ba = a. Evidently every s-unital -algebra is non-degenerate and any subset of a s-unital -algebra is assimilative in the algebra. Any sum of s-unital ideals is s-unital [4, Remark 2.5]. The next Proposition resumes all we have to say about uniqueness and non degeneracy of globalizations. Proposition 2.14. Let α be a -partial action of G on A with A non-degenerate and At assimilative in A, for all t ∈ G. Then (1) α has a -globalization if and only if α has a non-degenerate and minimal -globalization. (2) Given minimal -globalizations of α, Ξ = (B, β, I, ι) and Σ = (C, γ, J, κ), with Σ non-degenerate, there exists a unique morphism ΞρΣ : β → γ such that ΞρΣ ◦ ι = κ. Moreover, ΞρΣ is an isomorphism iff Ξ is non-degenerate iff πΞ is injective. (3) In case for all n ∈ N and t1, . . . , tn ∈ G the ideal At1 +··· + Atn is assimilative in A, every minimal -globalization of α is non-degenerate. Proof. Assume α has a ∗−globalization Ξ = (B, β, I, ι), which we can assume is minimal (see the comments after Definition 2.7). Let π be the canonical morphism associated to Ξ. We claim that π(It) = π(I) ∩ Θt(π(I)). Indeed, note that π(It) = π(I ∩ βt(I)) ⊂ π(I) ∩ Θt(π(I)). For the converse inclusion take a, b ∈ I and t ∈ G such that π(a) = Θt(π(b)). Given c ∈ A, aι(c) = ι(π(a)δc ee) = ι(Θt(π(a))δc ee) = βt(a)ι(c) ∈ It. As It is assimilative in I, a ∈ It and π(It) = π(I) ∩ Θt(π(I)). It is straightforward to show that π◦ι(A) = π(I) is -ideal of π(B), [π(I)] = π(B). From the last paragraph and Remark 2.1 we know π(B) is non-degenerate and π ◦ ι : α → Θπ(I) is an isomorphism. As Θπ(I) = Θπ(B)π(I), (π(B), Θπ(B), π(I), π◦ ι) is a non-degenerate and minimal -globalization of α. Thus we have proved (1). Lets prove the existence claim of statement (2). Let πΞ be the canonical mor- phism associated to Ξ. To show that the image of πΞ, Im(πΞ), is Im(πΣ) note that t (a, b) := ι−1(βt(ι(a))ι(b)) is the unique element given (t, a, b) ∈ G, the element uα of At such that αt(c)uα t (a, b) = αt(ca)b (for all c ∈ At−1 ). Indeed, as ι : α → βI is CONSTRUCTION OF ENVELOPING ACTIONS 7 an isomorphism and ι(At) = I ∩ βt(I) an ideal of B : αt(c)uα t (a, b) = ι−1(βt(ι(c))βt(ι(a))ι(b)) = ι−1(βt(ι(c))βt(ι(a))ι(b)) = ι−1(βt(ι(ca))ι(b)) = αt(ca)b. Uniqueness follows from the non degeneracy of At = αt(At−1 ). Then uniqueness implies ι−1(βt(ι(a))ι(b)) = uα t (a, b) = κ−1(γt(κ(a))κ(b)). This last equality implies that πΞ ◦ ι = πΣ ◦ κ. Then πΞ(I) = πΞ ◦ ι(A) = πΣ ◦ κ(A) = πΣ(J) and Im(πΞ) = [πΞ(I)] = [πΣ(J)] = Im(πΣ). From Lemma 2.12 we know ker(πΣ) = AnnR(C). Since C is non-degenerate this implies πΣ is injective. Now define ΣρΞ := π−1 Σ ◦ πΞ. In case Ξ is non degenerate, πΞ is injective and π−1 Ξ ◦ πΣ is the inverse of ΣρΞ. To close the circle note that in case ΣρΞ is an isomorphism, Ξ is non degenerate because Σ is non degenerate. Uniqueness of ΣρΞ follows from tree facts: it is a morphism of -global actions, ΣρΞI = κ ◦ ι−1 and span{βt(I) : t ∈ G} = B. To prove (3) assume Ξ is minimal, we will show that Ξ is non-degenerate. From the hypothesis we get that, for all n ∈ N and t1, . . . , tn ∈ G, the ideal It1 +···+Itn is assimilative in I. Assume b ∈ B satisfies Bb = {0}. Given that [I] = B, there exits n ∈ N, t1, . . . , tn ∈ G and b1, . . . , bn ∈ I such that b = Pn j=1 βtj (bj). We show b = 0 by induction in n. If n = 1 it follows that Ib1 = {0} and this implies b = 0 because I is non-degenerate. For n > 1 we have βr(b)c = βr(bβr−1(c)) = 0, for all c ∈ I and r ∈ G. Then Pn (bj)c ∈ Pn−1 (j = 1, . . . , n − 1) such that n tj bn = Pn−1 j=1 b′ (I) = βtj (I), so there j ∈ I such that βtn (bn) = Pn−1 j ). (cid:3) n we get bnc = −Pn−1 j ∈ It−1 ) ⊂ βtn ◦ βt−1 j ). This implies b = Pn−1 exists b′′ By induction it follows that b = 0. j=1 βrtj (bj)c = 0. With r = t−1 , for all c ∈ I. Thus there exists b′ j. Besides, βtn (b′ j) ∈ βtn (It−1 n tj j=1 βtj (b′′ j=1 βtj (bj + b′′ j=1 It−1 j=1 βt−1 n tj n tj n tj Before giving a sufficient condition for the existence of a -globalization we prove the following Lemma 2.15. Let A be a -algebra, I and J -ideals of A, a ∈ I and b ∈ J such that: I ∩ J and I + J are non-degenerate, Ia ⊂ J, Jb ⊂ I and ca = cb for all c ∈ I ∩ J. Then a = b. Proof. For all z ∈ I ∩ J and x ∈ I ∪ J we have xa, xb ∈ I ∩ J and zxa = zxb. Thus, for all x ∈ I + J, xa = xb and this implies a = b. (cid:3) The next result is a combination of [3, Theorem 4.5] and [4, Theorem 3.1]. Theorem 2.16. Let α = ({At}t∈G,{αt}t∈G) be a -partial action of G on A and consider the conditions (1) α has a -globalization. (2) For all (t, a, b) ∈ G × A × A there exists u ∈ At such that, for all c ∈ At−1 , αt(c)u = αt(ca)b. Then (1) implies (2). In case for all s, t ∈ G the -subalgebras At + As and At ∩ As are non-degenerate and At is assimilative in A, (2) implies (1). Moreover, in this last case α has a non-degenerate and minimal -globalization. Proof. The implication (1) ⇒ (2) is part of the proof of Proposition 2.14. For the converse note that the element u in (2) is uniquely determined by (t, a, b) because 8 DAMI ´AN FERRARO At is non-degenerate and αt(At−1 ) = At. The expression ut(a, b) will be used to denote the element u given for (t, a, b) in (2). Let AG be the -algebra of function from G to A with point wise operations and M (AG) it's multiplier algebra. Note A = Ae is non-degenerate, so AG is non-degenerate and we can appeal to Remark 2.2, what we do without explicit mention. We claim that there exists a unique injective homomorphism of -algebras ρ : A → M (AG) such that ρ(a)ft = ut(a, ft). Given a ∈ A, to show ρ(a) is a multiplier with adjoint ρ(a∗) it suffices to show that [ρ(a)f ]∗g = f ∗[ρ(a∗)g], for all f, g ∈ AG. The equality holds because for all f, g ∈ AG, t ∈ G and c, d ∈ At we have [ρ(a)f ]∗gt, f ∗[ρ(a∗)g]t ∈ At and d[ρ(a)f ]∗gtc = dut(a, ft)∗gtc = d[αt(αt−1 (c∗g∗t)a)ft]∗ = df ∗tαt(a∗αt−1 (gtc)) = αt(αt−1 (df ∗t)a∗αt−1(gtc)) = αt(αt−1 (df ∗t)a∗)gtc = df ∗[ρ(a∗)g]tc. The uniqueness of elements ut(a, b) implies a 7→ ut(a, b) is linear, so ρ is linear. To show ρ is multiplicative it suffices to prove, for all a, b, c ∈ A and t ∈ G, the identity ut(a, ut(b, c)) = ut(ab, c). Note ut(a, ut(b, c)), ut(ab, c) ∈ At and for all d ∈ At : dut(a, ut(b, c)) = αt(αt−1 (d)a)ut(b, c) = αt(αt−1 (d)ab)c = dut(ab, c). Thus ut(a, ut(b, c)) = ut(ab, c). Assume ρ(a) = 0. Given b ∈ A let δb on e and 0 elsewhere. Then 0 = ρ(a)δb 0 = cue(a, b) = αe(ca)b = cab. As A is non-degenerate, a = 0 and ρ is injective. e ∈ AG be the function taking the value b ee = ue(a, b). Given c ∈ A = Ae we have With Θ being the canonical action of G on M (AG) set B := span{Θt(ρ(A)) : t ∈ G}. So B is a Θ-invariant -sub algebra of M (AG) and the restriction of Θ to B, β, is a -global action. The proof will be completed once we show ρ(A) is an ideal of B, that ρ is an isomorphism between α and βρ(A) and that B is non-degenerate. βt(ρ(a))ρ(b) = ρ(ut(a, b)). From the definitions of β and ρ we obtain Lets show that βt(ρ(A))ρ(A) ⊂ ρ(At), for all t ∈ G. It suffices to show that βt(ρ(a))ρ(b)fr = urt(a, ur(b, fr)) and ρ(ut(a, b))fr = ur(ut(a, b), fr). Putting c := fr it suffices to show urt(a, ur(b, c)) = ur(ut(a, b), c). Note urt(a, ur(b, c)) ∈ Art, ur(ut(a, b), c) ∈ Ar, Arturt(a, ur(b, c) = αrt (At−1r−1a) ur(b, c) ∈ Art ∩ Ar and Arur(ut(a, b), c) = αr (Ar−1ut(a, b)) c ∈ αr(Ar−1 ∩ At) = Ar ∩ Art. Besides, for z ∈ Art ∩ Ar : zurt(a, ur(b, c)) = αrt(αt−1r−1(z)a)ur(b, c) = αr (αt(αt−1r−1(z)a)b) c = αr (αr−1(z)ut(a, b)) c = zur (ut(a, b), c) . Then Lemma 2.15 implies urt(a, ur(b, c)) = ur(ut(a, b), c). Now the inclusion βt(ρ(A))ρ(A) ⊂ ρ(At) (valid ∀ t ∈ G) implies ρ(A) is an ideal of B because Bρ(A) = X t∈G βt(ρ(A))ρ(A) ⊂ ρ(A) and ρ(A)B = (Bρ(A))∗ ⊂ ρ(A). CONSTRUCTION OF ENVELOPING ACTIONS 9 Take t ∈ G and a ∈ At−1 . To prove ρ(αt(a)) = βt(ρ(a)) is equivalent to prove ur(αt(a), b) = urt(a, b), for all b ∈ A and r ∈ G. As before note that ur(αt(a), b) ∈ Ar, urt(a, b) ∈ Art, Arur(αt(a), b) = αr (Ar−1αt(a)) b ∈ αr(Ar−1 ∩ At) = Ar ∩ Art and Arturt(a, b) = αrt (At−1r−1a) b ∈ αrt (At−1r−1At−1 ) = Art ∩ Ar. For z ∈ Ar ∩ Art : zur(αt(a), b) = αr(αr−1(z)αt(a))b = αr(αt(αt−1r−1(z)a))b = αrt(αt−1r−1(z)a)b = zurt(a, b). Then ur(αt(a), b) = urt(a, b) by Lemma 2.15. To finish the proof all we need to show is that ρ(At) = ρ(A) ∩ βt(ρ(A)). Note ρ(At) = ρ(αt(At−1 )) = βt(ρ(At−1 )) ⊂ ρ(A) ∩ βt(ρ(A)). Now take T ∈ ρ(A) ∩ βt(ρ(A)). Then there exist a, b ∈ A such that T = ρ(a) = βt(ρ(b)). Thus, for all f ∈ AG and r ∈ G : ur(a, fr) = ρ(a)fr = βt(ρ(a))fr = urt(a, fr). By replacing f with δb e and r with e we obtain ue(a, b) = ut(a, b), for all b ∈ A. But ue(a, b) = ab, then aA ⊂ At and this implies a ∈ At. Up to this point we have shown that Ξ := (B, ΘB, ρ(A), ρ) is a minimal - globalization of α. Note that the canonical morphism associated to Ξ, πΞ : B → M (AG), is just the natural inclusion B ⊂ M (AG) because πΞ ◦ ρ = ρ. Then Lemma 2.12 implies B is non-degenerate. (cid:3) With the previous Theorem we can extend Example 2.9 considerably. Let A be a -algebra and I an -ideal of it. Recall that an orthogonal complement for I is a -ideal J ⊂ A such that A = I ⊕ J. In this situation IJ = 0 because IJ ⊂ I ∩ J. Several remarks are in order. Firstly, if I is non degenerate the it has at most one orthogonal complement. Secondly, if J is an orthogonal complement for I, then A and I are non degenerate ⇔ I and J are non degenerate ⇔ A and J are non degenerate. Thirdly, I is assimilative in A every time A and I are non degenerate and I has an orthogonal complement. Finally, there are non degenerate assimilative ideals without an orthogonal complement. As an example take the complex sequences vanishing except on a finite set, considered as a -ideal of the complex sequences converging to zero. Corollary 2.17. Let α = ({At}t∈G,{αt}t∈G) be a -partial action of G on A. If for all s, t ∈ G the -subalgebras At + As and At ∩ As are non-degenerate and At has an orthogonal complement in A, then α has a non degenerate and minimal -globalization. Proof. From the hyphoteses it follows that A = Ae + Ae and At = At + At are non degenerate. Besides At has a unique orthogonal complement in A, so it is assim- ilative in A. Then it suffices to verify that α satisfies condition (2) from Theorem 2.16. Take (t, a, b) ∈ G× A× A. Since At−1 has an orthogonal complement in A, there exists a′ ∈ At−1 such that ca = ca′, for all c ∈ At−1. With u := αt(a′)b ∈ At we have, for all c ∈ At−1 : αt(c)u = αt(c)αt(a′)b = αt(ca′)b = αt(ca)b. (cid:3) 10 DAMI ´AN FERRARO It may happen that a -partial action has a minimal -globalization with many ideals At without an orthogonal complement. To exhibit an example consider let C0(R) be the -algebra of continuous function from R to C vanishing at ±∞. Let β be the action of R on C0(R) given by βt(f )(r) = f (r − t). As an ideal of C0(R) take A = C0(0, +∞) = {f ∈ C0(R) : fR\(0,+∞) ≡ 0} and let α be the restriction of β to A. Then we can construct a non degenerate and minimal -globalization of α using β. Every ideal of A is non degenerate because in A the identity aa∗ = 0 implies a = 0, but the only -ideals of A with an orthogonal complement are the trivial ones. For example, with t > 0 the ideal At = A ∩ βt(A) = C0(t, +∞) does not have an orthogonal complement in A. 2.2.1. About -partial actions on commutative algebras. We close this section with an adaptation of [1, Proposition 2.1] to -algebras. Proposition 2.18. Assume α is a -partial action of G on the commutative - algebra A and σ = (B, β, I, ι) a minimal -globalization. If At is non degenerate for all t ∈ G then B is commutative. Proof. Since B = span{βr(a) : r ∈ G, a ∈ I}, it suffices to prove that βr(a)βs(b) = βs(b)βr(a), for all a, b ∈ I and r, s ∈ G. Fix a, b ∈ A and r, s ∈ G. With t := r−1s we have βr(aβt(b)) = βr(a)βs(b) and βr(βt(b)a) = βs(b)βr(a). Then the proof will be completed once we show that aβt(b) = βt(b)a. Note that I is commutative, It := ι(At) non degenerate and aβt(b), βt(b)a ∈ I ∩ βt(I) = It. Then aβt(b) = βt(b)a if and only if caβt(b) = cβt(b)a, for all c ∈ It. Take c ∈ It, using that I is commutative and a, b, c, βt−1(ac), βt−1 (c), βt(b)a ∈ I, we obtain caβt(b) = βt(βt−1 (ac)b) = βt(bβt−1 (ac)) = βt(b)ac = cβt(b)a. (cid:3) 3. Partial actions on C-algebras Assume A is a C-algebra and G a topological group (we do not require it to be locally compact nor Hausdorff). Definition 3.1. The pair α = ({At}t∈G,{αt}t∈G) is a C-partial action (of G on A) if: • It is a -partial action of G on A. • For all t ∈ G, At is closed in A. • {At}t∈G is a continuous family, that is: for every open set U ⊂ A the set • The function {(t, a) ∈ G×A : a ∈ At−1} → A, (t, a) 7→ αt(a), is continuous. Morphisms between C-partial actions are just morphism between -partial ac- tions. Recall from [1, Example 2.1] that the restriction of a C-partial action to a C-ideal is a C-partial action. GU := {t ∈ G : At ∩ U 6= ∅} is open in G. Definition 3.2. A C-globalization of the C-partial action α, of G on A, is a 4-tuple Ξ = (B, β, I, ι) where: B is a C-algebra, β is a C-global action of G on B, I is a C-ideal of B and ι : α → βI is an isomorphism of -partial actions. Note we do not require ι to be a homeomorphism, this is automatic because every -homomorphism between C-algebras is contractive and has closed range. In case α is a C-partial action, it is also a -partial action and every C- globalization of α is a -globalization of α. Although not every -globalization of α is a C-globalization, the existence of a -globalization implies the existe CONSTRUCTION OF ENVELOPING ACTIONS 11 of a C-globalization. We will prove this claim in two steps: first we show the group's topology is irrelevant to globalize C-partial actions, then we construct a C-globalization from a -globalization. In the next statement Gdis denotes the group G with the discrete topology. Lemma 3.3. Let G be a topological group, β a C-global action of Gdis on B and I an ideal of B such that [I] := span{βt(I) : t ∈ G} is dense in B. Set α := βI , which is a C-partial action of Gdis on I. Then β is a C-partial action of G if and only if α is a C-partial action of G. Proof. The direct implication is [1, Example 2.1]. For the converse note that (as each βr is an isometry) G × B → B, (t, b) 7→ βt(b), is continuous if and only if for every b ∈ B the function evb : G → B, t 7→ βt(b), is continuous at e. Moreover, U := {b ∈ B : evb is continuous at e} is a closed β-invariant subspace of B, thus all we need to show is that I ⊂ U. Fix a ∈ I. From [1, Lemma 2.1] (using the family of ideals {βt(I)}t∈G) we get, for all t ∈ G, that kβt(a) − ak = sup{k(βt(a) − a)βr(b)k : r ∈ G, b ∈ I, kbk ≤ 1} = sup{k(βr−1t(a) − βr−1(a))bk : r ∈ G, b ∈ I, kbk ≤ 1} = sup{k(βrt(a) − βr(a))bk : r ∈ G, b ∈ I, kbk ≤ 1}. Fix r ∈ G and b ∈ I with kbk ≤ 1. Note that (βrt(a)−βr(a))b = βrt(a)b−βr(a)b ∈ βrt(I)I + βr(I)I = Irt + Ir. Thus, if {vλ}λ∈Λ is an approximate unit of Irt + Ir, we have k(βrt(a) − βr(a))bk = limλ k(βrt(a) − βr(a))bvλk. r such that vλ = c + d. This implies kck,kdk ≤ kvλk ≤ 1. One one hand, if {wµ}µ∈M is an approximate unit of It, we have Given λ ∈ Λ, vλ ∈ (Irt + Ir)+ and there exists c ∈ I + rt and d ∈ I + k(βrt(a) − βr(a))bck = kaβt−1r−1(bc) − βt−1(a)βt−1r−1(bc)k = lim µ kaβt−1r−1(bc) − βt−1(wµ)βt−1(a)βt−1r−1(bc)k ≤ lim sup µ ka − βt−1 (wµa)k = lim sup µ ka − αt−1 (wµa)k. To construct a particular approximate unit consider Mt := {c ∈ I + with it's natural order, then {µ}µ∈Mt is an approximate unit of It and t : kak < 1} k(βrt(a) − βr(a))bck ≤ lim With s = rt by symmetry we get µ∈Mt sup{ka − αt−1 (νa)k : ν ∈ Mt, ν ≥ µ} =: C(t). k(βrt(a) − βr(a))bdk = k(βst−1(a) − βs(a))bdk ≤ C(t−1). Putting all together we obtain kβt(a) − ak ≤ C(t) + C(t−1). All we need to show is that limt→e C(t) = 0. Take ε > 0. As α is a partial action of G on I there are neighbourhoods V ⊂ G and W ⊂ B of e and a, respectively, such that for s ∈ V and b ∈ Is−1 ∩ W we have kαs(b) − ak < ε/2. Take δ > 0 such that B(a, δ) ⊂ W. Then U := {r ∈ G : B(a, δ/2) ∩ Ir 6= ∅} is an open set containing e because {It}t∈G is a continuous family. For r ∈ U ∩ V −1 there exists b ∈ Ir ∩ B(a, δ/2), so limµ∈Mr ka − µak = dist(a, Ir) ≤ δ/2. Take µr ∈ Mr such that ka− νak < δ for all ν ∈ Mr with ν ≥ µr. 12 DAMI ´AN FERRARO Then for r ∈ U ∩ V −1 we have C(r) < ε because for all ν ∈ Mr with ν ≥ µr the inequality ka − αr−1(νa)k ≤ ε/2 holds and implies sup{ka − αr−1 (ν′a)k : ν′ ∈ Mr, ν′ ≥ ν} ≤ ε/2 < ε. (cid:3) Corollary 3.4. Let α = ({At}t∈G,{αt}t∈G) be a C-partial action and use the expression αdis to denote α as a C-partial action of Gdis. Then α has a C- globalization if and only if αdis has a C-globalization. Proof. As mentioned before the direct implication in immediate. For the converse assume αdis has a C-globalization Ξ = (B, β, I, ι). Without loss of generality we may assume Ξ is minimal (an enveloping action in the sense of F. Abadie [1]) this is [I] = B. Note βI is a C-partial action of G because it is isomorphic (as a -partial action) to α. Then the previous lemma implies β is a C-global action of G. Thus Ξ is a C-globalization of α. (cid:3) As promised before we now construct a C-globalization from a -globalization. Theorem 3.5. Let α = ({At}t∈G,{αt}t∈G) be a C-partial action. Then the following are equivalent: (1) α has a C-globalization. (2) α has a -globalization. (3) For all (t, a, b) ∈ G × A × A there exists u ∈ At such that, for all c ∈ At−1 , αt(c)u = αt(ca)b. Proof. We already know (1) implies (2). On C-algebras the condition a∗a = 0 implies a = 0, thus every C-ideal of a C-algebra is non-degenerate. The existence of approximate units implies every closed ideal is assimilative. Besides the sum4 of two C-ideals of a C-algebra is, again, a C-ideal. Then (2) and (3) are equivalent by Theorem 2.16. To show (2) implies (1), without loss of generality, we assume G is discrete (Corollary 3.4). From the previous paragraph and Proposition 2.14 we know α has a non degenerate and enveloping -globalization Ξ = (B, β, I, ι). Note I is a C-algebra because it is isomorphic (as a -algebra) to a C-algebra. Moreover, βI is a C-partial action isomorphic to α. Then it suffices to show that βI has a C-globalization and me may think I = A, ι is inclusion of A in B and α = βI . Let π be the canonical morphism associated to Ξ and Θ the canonical action of G on M (AG). From Proposition 2.14 we know π is injective and, as we think A ⊂ B, π(b)fr = βt(b)fr. structure inherited from AG and the supremum norm. Define b ) ⊂ AG b }. The set of bounded functions from G to A, AG C := {T ∈ M (AG) : T (AG b , is a C-algebra with the -algebra b ) ∪ T ∗(AG Note C is Θ invariant and that to show π(B) ⊂ C it suffices to prove that b and r ∈ G and π(A) ⊂ C. This last inclusion holds because given a ∈ A, f ∈ AG an approximate unit of Ar−1, {vλ}λ∈Λ, we have βr(a)fr ∈ Ar and kπ(a)frk = kβr(a)frk = lim λ kαr(vλ)βr(a)frk = lim λ kαr(vλa)frk ≤ kakkfk. 4Without closure. CONSTRUCTION OF ENVELOPING ACTIONS 13 b → AG Consider M (AG b ) as a C-algebra and let ρ : π(B) → M (AG b ) be defined as ρ(T )f = T f. To show ρ is injective it suffices to show that ρ ◦ π is injective. Take b ∈ B such that ρ ◦ π(b) = 0. There are a1, . . . , an ∈ A and t1, . . . , tn ∈ G such that b = Pn j=1 βtj (aj ). Given r ∈ G and c ∈ A we have bβr(c) = βr(βr−1(b)c) = βr(ρ ◦ π(b)δc r−1r−1 ) = 0. Thus bB = {0} and this implies b = 0. b ) → M (AG b as ψt(f )r = frt and Ψt : M (AG Given t ∈ G set ψt : AG b ) as Ψt(T ) = ψt ◦ T ◦ ψt−1 . Then Ψ is a C-global action of G on M (AG b ); ρ ◦ π : β → Ψ is a morphism of C-partial actions and the closure of ρ◦ π(B), D, is a Ψ-invariant C-subalgebra of M (AG b ). Name γ the restriction of Ψ to D. Note J := ρ(π(A)) is a C-ideal of D because ρ ◦ πA has closed range and J is an ideal of ρ ◦ π(B). From Remark 2.3 we get γJ = ΨDρ◦π(A) = Ψρ◦π(A) = Ψρ◦π(B)ρ◦π(A). Besides, ρ : Θπ(B) → Ψρ◦π(B) is an isomorphism. Then ρ◦ πA : α → ΨJ is an isomorphism and, consequently, (D, γ, J, ρ ◦ πA) is a C-globalization of α. (cid:3) Condition (3) asserts the existence of a certain element u for each (t, a, b) ∈ G × A × A. That element is the limit of the net given in (2) below. Proposition 3.6. Let α = ({At}t∈G,{αt}t∈G) a C-partial action. Then the fol- lowing are equivalent. (1) α has a C-globalization. (2) There exists U, V ⊂ A such that: (i) span AU = span V A = A and (ii) for all (t, a, b) ∈ G × U × V there exists an approximate unit of At−1 , {vλ}λ∈Λ, such that {αt(vλa)b}λ∈Λ is a Cauchy net (or converges). Proof. In case (1) holds set U = V = A and take (t, a, b) ∈ G × A × A. Let u be the element given in (2) of the previous theorem and {vλ}λ∈Λ and approximate unit of At−1 . Then {αt(vλ)}λ∈Λ is an approximate unit of At and {αt(vλ)u}λ∈Λ = {αt(vλa)b}λ∈Λ converges to u. Now assume (2) is true. Take (t, a, b) ∈ G × A × A for which there exists an approximate unit, {vλ}λ∈Λ, such that {αt(vλa)b}λ∈Λ converges to u. We claim that given any other approximate unit of At−1 , {wµ}µ∈M , the net {αt(wµa)b}µ∈M = {αt(wµ)u}µ∈M converges to u. Indeed lim λ αt(vλa)b = lim µ αt(wµ) lim λ αt(vλa)b = lim µ lim λ αt(wµvλa)b = lim µ αt(wµa)b. Then u is determined by (t, a, b), so we denote it ut(a, b). Moreover, if c ∈ At−1 then αt(c)ut(a, b) = lim λ αt(c)αt(vλa)b = lim λ αt(cvλa)b = αt(ca)b. Take a ∈ U, b ∈ V and c, d ∈ A and any approximate unit of At−1 , {vλ}λ∈Λ. Then {αt(vλca)bd}λ∈Λ converges to αt(cαt−1 (ut(a, b)d)) because, for all λ ∈ Λ, αt(vλca)bd = αt(vλc)ut(a, b)d = αt(vλcαt−1(ut(a, b)d)) and limλ αt(vλcαt−1 (ut(a, b)d)) = αt(cαt−1 (ut(a, b)d)). The conclusions of the previous paragraphs imply that (2) holds if we replace U for AU and V for V A. From now on we assume AU = U and V A = V. Given a, b ∈ U, c ∈ V and λ ∈ C note that ut(a, c) + λut(b, c) ∈ At and for all d ∈ At−1 : αt(d)(ut(a, b) + λut(b, c)) = αt(d(a + λb))c. In case a ∈ U and b, c ∈ V : αt(d)(ut(a, b) + λut(a, c)) = αt(da)(b + λc). Then, for all t ∈ G, there exists a unique bilinear function ut : span U × span V → At such that for all a ∈ span U, b ∈ span V and c ∈ At−1 : αt(c)ut(a, b) = αt(ca)b. Note that kut(a, b)k2 = 14 DAMI ´AN FERRARO kαt(αt−1 (ut(a, b)∗)a)bk ≤ kut(a, b)kkakkbk, then kut(a, b)k ≤ kakkbk. So there is a unique continuous bilinear function vt : A × A → At extending ut, for every t ∈ G. Given (t, a, b) ∈ G × A × A and c ∈ At−1 take sequences {an}n∈N ⊂ span U and {bn}n∈N ⊂ span V converging to a and b, respectively. Then αt(c)vt(a, b) = lim n αt(c)vt(an, bn) = lim n αt(can)bn = αt(ca)b. This shows (2) implies (3) of the previous theorem, so α admits a C-globalization. (cid:3) Now we use our criterion to give some conditions were a C-partial action can be globalized. On a unital C-algebra, A, for every unital ideal I there exists a unique ideal J such that A = I ⊕ J and IJ = {0}. In general, even for non unital A, we say a C-ideal I has an orthogonal complement if there exists a C-ideal J such that A = I ⊕ J and IJ = {0}. Corollary 3.7. Let α = ({At}t∈G,{αt}t∈G) be a C-partial action. t ∈ G, At has an orthogonal complement then α has a C-globalization. Proof. Given (t, a, b) ∈ G × A × A let J be the orthogonal complement of At−1 . Then there are a1 ∈ At−1 and a2 ∈ J such that a = a1 + a2. Note for all c ∈ At−1 , ca = ca1. Thus u := αt(a1)b ∈ At and αt(c)u = αt(ca1)b = αt(ca)b. If, for all (cid:3) The next corollary is a version for C-algebras of [3, Theorem 4.5]. The proof is omitted because it follows form the cited theorem and the previous corollary. Corollary 3.8. Let α = ({At}t∈G,{αt}t∈G) a C-partial action with A unital. Then α has a C-globalization if and only if At is a unital algebra, for all t ∈ G. There are certain -homomorphisms that force the existence of C-enveloping actions. For partial actions on locally compact and Hausdorff spaces the next result can be showed using [1, Proposition 2.1] instead of Theorem 3.5 (in that case the homomorphism φ defines a continuous function between the spectra of the algebras). Corollary 3.9. Let α = ({At}t∈G,{αt}t∈G) and β = ({Bt}t∈G,{βt}t∈G) be C- partial actions. If α has a C-globalization and there exists a -algebra's homomor- phism φ : A → M (B) such that: (1) for all t ∈ G, span{φ(a)b : a ∈ At, b ∈ B} = Bt and (2) for all t ∈ G, a ∈ At−1 and b ∈ Bt−1, φ(αt(a))βt(b) = βt(φ(a)b); then β has a C-globalization. Proof. By Cohen-Hewitt's Theorem Bt = {φ(a)b : a ∈ At, b ∈ B}, for all t ∈ G. The same theorem implies At = {ab : a, b ∈ At}, then Bt = {φ(a)b : a ∈ At, b ∈ Bt}. 1)c∗ 1 = φ(a∗ 1 and b2 = φ(a2)c2. Now let u be the element given for (t, a1, a2) by condition (3) of Theorem 3.5. Then c1φ(u)c2 ∈ Bt because u ∈ At. φ(e∗)d′∗. Then Given d ∈ Bt−1 choose d′ ∈ Bt−1 and e ∈ At−1 such that βt−1(c∗ Given (t, b1, b2) ∈ G×B×B take a1, a2 ∈ A and c1, c2 ∈ B such that b∗ 1βt(d∗)) = βt(d)c1φ(u)c2 = βt(d′)φ(αt(e)u)c2 = βt(d′)φ(αt(ea1)a2)c2 = βt(d′)φ(αt(ea1))φ(a2)c2 = βt(d′φ(e)φ(a1))b2 = βt(db1)b2. The rest follows directly from the previous theorem. (cid:3) CONSTRUCTION OF ENVELOPING ACTIONS 15 To close this section we give an example to show condition (1) from the previous statement can not we weakened to (1') Bt = span{φ(a)b : a ∈ At, b ∈ Bt}, for all t ∈ G. Let G = Z2 = {1,−1} (with multiplicative notation) A = B = C[0, 1]. As α consider the trivial global action of G on A. To give a partial action of G on B, β, we just need to specify β−1, which will be the identity on C0([0, 1)). For φ : C[0, 1] → M (C[0, 1]) = C[0, 1] just take the identity. Note (1') and (2) are satisfied, α has a C-globalization but β does not. 4. Partial action on equivalence bimodules In [1] F. Abadie defines Morita equivalence of C-partial actions using partial ac- tions on positive C-trings. Recall that a positive C-trings are exactly equivalence bimodules [1, 9, 8]. Here we adopt the terminology of equivalence bimodules. In this last section we give a necessary and sufficient condition for the existence of a globalization of a partial action on an equivalence bimodule and, as a consequence, we obtain the uniqueness of Morita enveloping actions (as was shown in [1]). We adopt the terminology of [8] and agree that "AXB is an equivalence bimodule" means "X is an A-B-equivalence bimodule". Take equivalence bimodules AXB and a CYD. A function φ : X → Y is an Hb-homomorphism if it is linear and for all x, y, z ∈ X , φ(xhy, ziB) = φ(x)hφ(y), φ(z)iD . Such functions are contractive and, with the previous notation, there exist unique -homomorphisms lφ : A → C and φr : B → D such that, for all x, y ∈ X , φl(Ahx, yi) = Chφ(x), φ(y)i and φr(hx, yiB) = hφ(x), φ(y)iD . The proof of these facts can be found in [1]. Given a C-ideal I of A, by Cohen-Hewitt IX := {ax : a ∈ I, x ∈ X} is a closed submodule of X . We denote I B the C-ideal of B induced by I (or IX ) through X , that is I B = span{hu, viB : u, v ∈ IX}. In a similar way we define, for a C-ideal J of B, X J and AJ . Recall that AI B = I and J = AJ B. For a closed subspace Z of X the following are equivalent: (i) there exists a C-ideal I of A such that Z = IX , (ii) there exists a C-ideal J of B such that Z = X J, (iii) XhZ,XiB ⊂ Z and (iv) AhX ,ZiX ⊂ Z. If these conditions are satisfied then AZ := span AhZ,Zi and ZB are C-ideals, Z = AZX = X Z B and we say Z is an ideal of X . Every ideal Z of X is an AZ − ZB-equivalence bimodule. Definition 4.1. The pair γ = ({γt}t∈G,{Xt}t∈G) is a Hb-partial action if: • X is an (A-B-)equivalence bimodule and G a topological group. • γ is a set theoretic partial action of G on X . • {Xt}t∈G is a continuous family of ideals of X . • For all t ∈ G, γt : Xt−1 → Xt is an Hb-homomorphism. • The function {(t, x) ∈ G×X : x ∈ Xt−1} → X , (t, x) 7→ γt(x), is continuous. Assume δ is a Hb-partial action G on CYD. We say φ : γ → δ is a Hb-morphism if it is a Hb-homomorphism from X to Y which is also a morphism of set theoretic partial actions. From [1] we know there are unique C-partial actions, α = ({αt}t∈G,{At}t∈G) and β = ({βt}t∈G,{Bt}t∈G) , such that • For all t ∈ G, At = AXt and Bt = XtB. • For all t ∈ G and x, y ∈ Xt−1 : αt(Ahx, yi) = Ahγt(x), γt(y)i and βt(hx, yiB) = hγt(x), γt(y)iB. 16 DAMI ´AN FERRARO We will call α the left side of γ and β the right side of γ and will denote them lγ and γr, respectively. Example 4.2. Every C-partial action, α, on a C-algebra A is a Hb-partial action on AAA. Besides α = lα = αr. Example 4.3. Given a Hb-global action of G on AXB, γ, and an ideal Y of X ; the restriction γY is a Hb-partial action on the AY -Y B-equivalence bimodule Y. In this case l(γY ) = (lγ)AY and (γY )r = γrY B. With restrictions of global actions, on one hand, and isomorphisms on the other we define globalizations. Definition 4.4. Let γ be an Hb-partial action of G on X . A globalization of γ is a 4−tuple Ξ = (Y, δ,Z, ι) such that: Y is an equivalence bimodule, δ is an Hb-global action of G on Y, Z is an ideal of Y and ι : γ → δZ is an isomorphism of Hb-partial actions. In case [Z] := span{δt(Z) : t ∈ G} is dense in Y we say Ξ is a minimal globalization. The nexus between Hb-partial actions and C-partial actions is the linking par- tial action [1]. To describe this action we start with an Hb-partial action of a group G on AXB, γ, and set α := lγ and β := γr. The linking algebra of X is the algebra of generalized compact operators of the A-Hilbert module X ⊕ A, L(X ) = K(X ⊕ A). In matrix representation L(X ) = (cid:0) A X eX B(cid:1) , with eX the B-A-equivalence bimodule adjoint to X . The linking partial action of γ, L(γ) = ({L(γ)t}t∈G,{L(X )t}t∈G) , is the unique C-partial action such that, for all t ∈, L(X )t = L(Xt) and ey b ) = (cid:16) αt(a) γt(x) ]γt(y) βt(b)(cid:17) ∀ x, y ∈ Xt−1 , a ∈ At−1 and b ∈ Bt−1. L(γ)t ( a x In case δ is an Hb-partial action of G on Y and π : γ → δ is an isomorphism, the morphism L(π) : L(γ) → L(δ) defined as L(π) ( a x ey b ) = (cid:16) lπ(a) π(x) gπ(y) πr(b)(cid:17) , is an isomorfphism with inverse L(π−1). Proposition 4.5. If Ξ = (CYD, δ,Z, ι) is an enveloping globalization of γ then • (L(Y), L(δ), L(Z), L(ι)) is a C-enveloping globalization of L(γ). • (C, lδ, CZ, lι) is a C-enveloping globalization of lγ. • (D, δr, Z D, ιr) is a C-enveloping globalization of γr. Proof. It is straightforward and is left to the reader. (cid:3) In the same way equivalence bimodules are constructed form C-algebras and projections, Hb-partial actions are constructed form C-partial actions and equi- variant projections. Theorem 4.6. Let α be a C-partial action of G on A and p ∈ M (A) a projection such that span ApA = A and αt(pa) = pαt(a), for all t ∈ G and a ∈ At−1 . If X := (1 − p)Ap, C := (1 − p)A(1 − p) and D := pAp then (1) X is a C-D-equivalence bimodule, (2) X is α-invariant, (3) the restriction of α to X , γ, is an Hb-partial action. CONSTRUCTION OF ENVELOPING ACTIONS 17 (4) C and D are α-invariant, lγ = αC and γr = αD. Moreover, α is isomorphic (as a C-partial action) to L(γ) and α has a C- globalization if and only if γ has an Hb-globalization. Proof. The proof of claims (1)-(4) are left to the reader. Besides, the usual iden- tification of A with L(X ) is an isomorphism of C-partial actions between α and L(γ). Then, in case γ has an Hb-globalization, L(γ) has a C-globalization and this implies α has a C-globalization. To prove the converse assume α has a C-globalization. To prove γ has a glob- alization we can assume, without loss of generality, that there exists a C-partial action, β of G on B, such that A is an ideal of B, α = βA and [A] = B. The key claim is that there exists a unique projection p ∈ M (B) such that βt(pb) = pβt(b) and pa = pa, for all t ∈ G, b ∈ B and a ∈ A. To prove the claim it suffices to show that for all t1, . . . , tn ∈ G and a1, . . . , an ∈ A nX j=1 k nX j=1 βtj (paj)k ≤ k βtj (aj)k. (1) From [1, Lemma 2.1] we conclude that it is enough to show that nX j=1 k βtj (paj)βr(b)k ≤ k nX j=1 βtj (aj)k, for all r ∈ G and b ∈ A with kbk < 1. Take r ∈ G and b ∈ B as before. Note pajβt−1 nX βtj (paj)βr(b)k = k βtj (pajβt−1 j r ∈ At−1 nX j r(b))k = k k j=1 j r so that βr−1tj (pajβt−1 j r(b))k j=1 j=1 j=1 nX nX nX nX j=1 j=1 = k ≤ k ≤ k αr−1tj (pajβt−1 pαr−1tj (aj βt−1 j r(b))k αr−1tj (aj βt−1 βr−1tj (ajβt−1 j r(b))k nX j r(b))k = k nX j r(b))k = k j=1 j=1 βtj (aj)k. Set Y := (1 − p)Ap ⊂ B, note that X ⊂ Y and define ι : X → Y as the canonical inclusion. Remark 2.3 implies βYX = βX = βAX = αX = γ, then (Y, βY ,X , ι) is a globalization of γ. (cid:3) Corollary 4.7. An Hb-partial action has a globalization if and only if it's linking partial action has a C-globalization. Proof. Let γ be a partial action of G on the A-B-equivalence bimodule X . The thesis follows easily from the previous theorem with α = L(γ) and p = ( 1 0 0 0 ) because L(γ)pL(X )(1−p) is isomorphic to γ. (cid:3) As a consequence of the Corollary we get that the group's topology does not affects the existence of Hb-globalizations because it does nor affects the existence of C-globalizations. This conclusion can be equally derived from our last result, 18 DAMI ´AN FERRARO which also implies that a C-partial action has a C-globalization if and only if it has a Hb-globalization. Theorem 4.8. An Hb-partial action has an Hb-globalization if and only if it's left and right sides have C-globalizations. Proof. The direct implication follows from Proposition 4.5. For the converse assume γ is an Hb-partial action of the topological group G on AXB. Assume α := lγ and β := γr have C-globalizations. Given t ∈ G set At := AXt and Bt := XtB. It suffices to show L(γ) has a C-globalization, for which we use Proposition 3.6 with U = V = {(cid:0) 0 x ey 0(cid:1) : x, y ∈ X}. Take (t, ξ, η) ∈ G × U × V. Let {vλ}λ∈Λ be an approximate unit of At−1 and {wµ}µ∈M one of Bt−1. Consider K = Λ × M with the ordered (λ, µ) ≤ (λ′, µ′) iff λ ≤ λ′ and µ ≤ µ′. Given κ = (λ, µ) ∈ K set dκ := (cid:16) vλ 0 For κ = (λ, µ), ξ = (cid:0) 0 x 0 wµ(cid:17) . Then {dκ}κ∈K is an approximate unit of L(X )t. ev 0 ) we have ey 0(cid:1) and η = ( 0 u L(γ)t(dκξ)η = (cid:16) hγt(vλx),vil 0 hγt(ywµ),uir(cid:17) . 0 By Cohen-Hewitt there exists b, c ∈ B and z, w ∈ X such that x = zb and v = wc. As β has a C-globalization {βt(wµb)c∗}µ∈M converges to an element p ∈ Bt−1. Then hγt(vλx), vil = hγt(vλazb), wcil = lim µ ν hγt(vλzwµfνb)c∗, wil lim = lim µ = lim ν hγt(vλzwµ)βt(fν b)c∗, wil = lim lim µ hγt(vλzwµ)p, wil µ hγt(vλzwµβt−1(p)), wil = hγt(vλzβt−1(p)), wil. Note zβt−1(p) ∈ Xt−1 , so that limλhγt(vλx), vil = hγt(zβt−1 (p)), wil. vergent. By symmetry {hγt(ywµ), uir}µ∈M is convergent. Then {L(γ)t(dκξ)η}κ∈K is con- (cid:3) References 1. Fernando Abadie, Enveloping actions and Takai duality for partial actions, J. Funct. Anal. 197 (2003), no. 1, 14 -- 67. 2. Wagner Cortes and Miguel Ferrero, Globalization of partial actions on semiprime rings, Con- temporary Mathematics 499 (2009), 27. 3. M. Dokuchaev and R. Exel, Associativity of crossed products by partial actions, enveloping actions and partial representations, Trans. Amer. Math. Soc. 357 (2005), no. 5, 1931 -- 1952. MR 2115083 (2005i:16066) 4. Michael Dokuchaev, ´Angel Del R´ıo, and Juan Sim´on, Globalizations of partial actions on nonunital rings, Proceedings of the American Mathematical Society 135 (2007), no. 2, 343 -- 352. 5. Ruy Exel, Circle actions on C-algebras, partial automorphisms, and a generalized Pimsner- Voiculescu exact sequence, J. Funct. Anal. 122 (1994), no. 2, 361 -- 401. MR 1276163 (95g:46122) , Twisted partial actions: a classification of regular C-algebraic bundles, Proc. London 6. Math. Soc. (3) 74 (1997), no. 2, 417 -- 443. 7. Kevin McClanahan, K-theory for partial crossed products by discrete groups, J. Funct. Anal. 130 (1995), no. 1, 77 -- 117. MR 1331978 (96i:46083) 8. Marc A. Rieffel, Morita equivalence for operator algebras, Operator algebras and applications, Part I (Kingston, Ont., 1980), Proc. Sympos. Pure Math., vol. 38, Amer. Math. Soc., Provi- dence, R.I., 1982, pp. 285 -- 298. 9. Heinrich Zettl, A characterization of ternary rings of operators, Advances in Mathematics 48 (1983), no. 2, 117 -- 143. CONSTRUCTION OF ENVELOPING ACTIONS 19 Departamento de Matem´atica y Estad´ıstica del Litoral, Universidad de la Rep´ublica, Gral. Rivera 1350. Salto. Uruguay. E-mail address: [email protected]
1102.5430	3	1102	2013-09-17T05:43:34	Cyclic Hilbert spaces and Connes' embedding problem	[ "math.OA", "math.FA" ]	Let $M$ be a $II_1$-factor with trace $\tau$, the linear subspaces of $L^2(M,\tau)$ are not just common Hilbert spaces, but they have additional structure. We introduce the notion of a cyclic linear space by taking those properties as axioms. In Sec.2 we formulate the following problem: "does every cyclic Hilbert space embed into $L^2(M,\tau)$, for some $M$?". An affirmative answer would imply the existence of an algorithm to check Connes' embedding Conjecture. In Sec.3 we make a first step towards the answer of the previous question.	math.OA	math	CYCLIC HILBERT SPACES AND CONNES' EMBEDDING PROBLEM Valerio CAPRARO and Florin R ADULESCU May 5, 2018 Abstract. Let M be a II1-factor with trace τ , the linear subspaces of L2(M, τ ) are not just common Hilbert spaces, but they have additional structure. We introduce the notion of a cyclic linear space by taking those properties as axioms. In Sec.2 we formulate the following problem: "does every cyclic Hilbert space embed into L2(M, τ ), for some M ?". An affirmative answer would imply the existence of an algorithm to check Connes' embedding Conjecture. In Sec.3 we make a first step towards the answer of the previous question. Contents 1 Cyclic Hilbert spaces 2 Relation with Connes' embedding conjecture 3 Extension of cyclic vector spaces 4 Problems we were not able to solve 5 Acknowledgement 1 Cyclic Hilbert spaces 1 3 4 10 11 Let M be a finite factor with unique normalized trace τ and let L2(M, τ ) be the Hilbert space obtained by taking the closure of the vector space M with respect to 1 the inner product (x, y) = τ (y∗x) 2 . Consider a finite-dimensional real Hilbert subspace H ⊆ Msa ⊆ L2(M, τ ) containing the identity. Observe that H is not just a common Hilbert space, but it has additional structure. 1 Proposition 1. The mapping ≪ a ⊗ b, c ⊗ d ≫ form on (H ⊗ H) ⊗ C and the following properties are satisfied . = τ (abdc) is a bilinear hermitian positive 1. the mappings v → v ⊗ 1 and v → 1 ⊗ v are isometric embeddings, i.e. (v, v) =≪ v ⊗ 1, v ⊗ 1 ≫=≪ 1 ⊗ v, 1 ⊗ v ≫ 2. ≪ ·,· ≫ is cyclic in the following sense ≪ a ⊗ b, c ⊗ d ≫=≪ c ⊗ a, d ⊗ b ≫ 3. ≪ ·,· ≫ is self-adjoint in the following sense ≪ a ⊗ b, c ⊗ d ≫= ≪ b ⊗ a, d ⊗ c ≫ 4. ≪ ·,· ≫ verifies the following property ≪ a ⊗ b, c ⊗ d ≫=≪ b ⊗ d, a ⊗ c ≫ 5. The mapping JH : (H ⊗ H)⊗ C → (H ⊗ H)⊗ C defined by setting JH (a⊗ b) = b⊗ a is an isometric involution, i.e. (a) JH (JH (a ⊗ b)) = a ⊗ b (b) ≪ JH(a ⊗ b), JH (a ⊗ b) ≫=≪ a ⊗ b, a ⊗ b ≫ In this article we want to consider Hilbert spaces which verify these five additional properties. Before giving the definition let us observe that properties 2. and 3. together imply properties 4. and 5. Indeed Lemma 2. Let (H, (·,·),≪ ·,· ≫) be a Hilbert space equipped with a bilinear positive hermitian form ≪ ·,· ≫ on (H ⊗ H) ⊗ C. If ≪ ·,· ≫ verifies 2. and 3. then it verifies also 4. and 5. 2 Proof. Suppose 2. and 3. are verified. Applying hermitianity, 2. and hermitianity again we get 4. Indeed ≪ a ⊗ b, c ⊗ d ≫= ≪ c ⊗ d, a ⊗ b ≫ = ≪ a ⊗ c, b ⊗ d ≫ = On the other hand, applying 3. and hermitianity, we get 5. Indeed =≪ b ⊗ d, a ⊗ c ≫ ≪ JH(a ⊗ b), JH (a ⊗ b) ≫=≪ b ⊗ a, b ⊗ a ≫= ≪ a ⊗ b, a ⊗ b ≫ = =≪ a ⊗ b, a ⊗ b ≫ Notice that we have not used the completeness with respect to (·,·). Thus we can give the following Definition 3. A cyclic pre-Hilbert space is a quadruple (V, (·,·), 1,≪ ·,· ≫), where (V, (·,·)) is a real pre-Hilbert space, 1 ∈ V is a pointed vector such that (1, 1) = 12 = 1 and ≪ ·,· ≫ is a bilinear complex-valued, hermitian positive form on (V ⊗V )⊗ C verifying properties 1.,2. and 3. (and, consequently, 4. and 5.). 2 Relation with Connes' embedding conjecture We have begun studying cyclic spaces motivated by Connes' embedding conjecture. Before explaining how they are related to each other, let us briefly recall Connes' embedding conjecture. Let R be the hyperfinite II1 factor (with unique trace denoted by τ ) and let ω ∈ β(N) \ N be a free ultrafilter on the natural number. One can construct the ultrapower Rω in the following way: first consider l∞(R) = {(xn)n ⊆ R : supnxn < ∞}; then consider its ideal Iω = {(xn)n ∈ l∞(Rω) : limn→ωτ (x∗ 2 = 0}; finally consider the quotient Rω = l∞(R)/Iω. It turns out to be a non weakly separable II1 factor with trace τRω (x + Iω) = limn→ωτ (xn), where (xn) is any representative sequence for x. Connes' embedding conjecture states that any II1-factor with separable predual embeds into Rω nxn) 1 ([Co]). This conjecture has become more and more interesting in recent years, since many authors have found lots of equivalent conditions showing that this conjecture is linked to 3 several branches of mathematics (like group theory and metric geometry), besides being transversal to most of the sub-specializations of Operator Algebras (see [Br], [Br2], [Ca-Pa], [Co-Dy], [El-Sz], [Ha-Wi2], [Ki], [Ne-Th], [Oz], [Pe], [Ra1], [Ra2], [Ra3], [Vo2], [Vo3] for some reference). Here is the problem we want to focus Problem 4. Does every separable cyclic space embed into some II1-factor with separable predual? We are interested in this problem because an affirmative answer would imply the existence of an algorithm to check Connes' embedding conjecture. Indeed 1. Take a II1-factor with separable predual M . If Prob.4 has affirmative answer, then we could theoretically enumerate all the inequalities verified by the moments of order 3 and 4 in M . They are positive definite polynomials of degree less than or equal to 4, that are quite easy to understand, being exactly the inequalities of a cyclic space. 2. Take these polynomials and calculate their own infimum on positive matrices of order n. Let εn ≥ 0 be such an infimum. Observe that Connes' embedding conjecture is Indeed Connes' true if and only if εn converges to 0, when n goes to infinity. embedding problem has an affirmative answer if and only if one can approximate the moments of order 3 and 4 (see [Ra3]). If the Connes embedding conjecture is true then the algorithm is infinite for at least one M . On the other hand, if it finite for some M , that is, it stops after a finite time, then the Connes embedding conjecture might be true or false. In this case, the algorithm could be used as a tool for constructing possible counter-example. 3 Extension of cyclic vector spaces The idea to answer Prob.4 is the following: suppose we have an orthonormal basis {xn} for L2(M, τ ), then we would have xixj = X αn ijxn n and thus the first requirement is that an element of V ⊗V should be actually an element of V ⊗ 1. It means that the first step is to extend the cyclic structure by adjoining elements. 4 More precisely we have to extend the cyclic structure on V to a cyclic structure on a space W of the shape V ⊕ RY , where Y is an indeterminate, in order to reconstruct step by step the product. We mean that, chosen arbitrarily y ∈ (V ⊗ V ) ⊗ C, y has to be represented as the indeterminate Y , i.e. y = 1 ⊗ Y = Y ⊗ 1. This is why extending the cyclic scalar product ≪ ·,· ≫ to one over ((V ⊕ RY )⊗ (V ⊕ RY ))⊗ C (where Y represents an arbitrary element in V ⊗ V ) means exactly that we are extending the scalar product on V to get the fixed product y of elements in V . Such a purpose forces some necessary assumptions on y: 1. y must be self-adjoint, in the sense that JV y = y. 2. y must have norm 1, i.e. ≪ y, y ≫= 1. 3. y is not an element of V ⊗ 1 or 1 ⊗ V (otherwise we would have trivial product). In this case we say that y is a non-trivial element in V ⊗ V . Unfortunately we are not able to extend the structure exactly, but just approximately. Definition 5. Let V be a finite dimensional cyclic vector space with orthonormal basis {x1, ...xn}. An ε-perturbation of the original scalar product ≪ ·,· ≫ is another scalar product ≪ ·,· ≫ε such that ≪ xi ⊗ xj, xk ⊗ xl ≫ − ≪ xi ⊗ xj, xk ⊗ xl ≫ε < ε ∀i, j, k, l ∈ {1, ...n} Proposition 6. Let V be a finite dimensional cyclic vector space and y a self-adjoint and non-trivial element in (V ⊗ V ) ⊗ C with norm 1. For every ε > 0, there exists an ε-perturbation of ≪ ·,· ≫ which extends to a cyclic structure on W := V ⊕ RY with the property y = 1 ⊗ Y = Y ⊗ 1. Proof of this proposition is quite technical, so we will divide it in several steps. Indeed, let 1, x2...xn an orthonormal basis of V (1 is the pointed vector on V ), we need to define the products ≪ Y ⊗ xi, xj ⊗ xl ≫ε ≪ Y ⊗ xi, Y ⊗ xj ≫ε ≪ Y ⊗ Y, xi ⊗ xj ≫ε ≪ Y ⊗ xi, xj ⊗ Y ≫ε ≪ Y ⊗ xi, Y ⊗ Y ≫ε ≪ Y ⊗ Y, Y ⊗ Y ≫ε 5 The remaining products ≪ xi ⊗ xj, xk ⊗ xl ≫ε will be defined in the course of the proof, when we find the suitable ε-perturbation of ≪ ·,· ≫. The most technical part of the proof is the definition of ≪ Y ⊗ xi, xj ⊗ xk ≫ε, which will be the first and second step. In the first step we follow a sequence of necessary conditions in order to construct a linear system whose solutions allow us to define such products; in the second step we solve this linear system. Before going into the first step, let us state some preliminary notions. By the fifth property in Prop.1, JV behaves on (V ⊗ V ) ⊗ C like an involution, so it is natural to fix the following terminology. Definition 7. An element x ∈ (V ⊗ V ) ⊗ C is called self-adjoint if JV x = x. For an element x ∈ (V ⊗ V ) ⊗ C which is not self-adjoint, its real part is Re(x) = x+JV x and the imaginary part is Im(x) = x−JV x . 2 2i Step 1 Let P be the projection of (V ⊗ V ) ⊗ C onto the first (V ⊗ C)⊥. Observe that ≪ Y ⊗ xi, xj ⊗ xk ≫ε=≪ Y ⊗ xi, P (xj ⊗ xk) + (1 − P )(xj ⊗ xk) ≫ε= =≪ Y ⊗ xi, P (xj ⊗ xk) ≫ε + ≪ Y ⊗ xi, (1 − P )(xj ⊗ xk) ≫ε Consider the second summand λk ij =≪ Y ⊗ xi, (1 − P )(xj ⊗ xk) ≫ε=≪ Y ⊗ xi, 1 ⊗ xk ≫ε=≪ 1 ⊗ Y, xk ⊗ xi ≫W =≪ y, xk ⊗ xi ≫ε=≪ y, xk ⊗ xi ≫ So we can think of the numbers λk ij as being pre-determined. Let us focus on the first summand: we are going to find a suitable perturbation in order to determine those numbers. Let ξi be the projection of the vector Y ⊗ xi on (V ⊗ V )⊗ C and ηi = P ξi. Then ≪ Y ⊗ xi, P (xj ⊗ xk) ≫ε=≪ ξi, P (xj ⊗ xk) ≫ε=≪ ηi, xj ⊗ xk ≫ε So we would solve our problem if we found suitable ηi's. Now observe that they should verify the following ≪ ηi, xj ⊗ xk ≫ε − ≪ Jεηj, xk ⊗ xi ≫ε= λji k 6 ij =: θk ij − λk This is just a linear system. Before attempting to solve it, let us write it separately for the real and the imaginary part. We get Re(θk ij) = Re ≪ ηi, xj ⊗ xk ≫ε −Re ≪ Jεηj, xk ⊗ xi ≫ε= 1 2 = (≪ ηi, xj ⊗ xk ≫ε +≪ ηi, xj ⊗ xk ≫ε− − ≪ Jεηj, xk ⊗ xi ≫ε −≪ Jεxj, xk ⊗ xi ≫ε) = 1 (≪ ηi, xj ⊗ xk ≫ε + ≪ Jεηi, xk ⊗ xj ≫ε − = 2 − ≪ Jεηj, xk ⊗ xi ≫ε − ≪ ηj, xi ⊗ xk ≫ε) = xj ⊗ xk + xk ⊗ xj ηj + Jεηj , ≫ε − ≪ 2 2 xk ⊗ xi + xi ⊗ xk 2 ηi + Jεηi 2 , ηi − Jεηi , xj ⊗ xk − xk ⊗ xj ηj − Jεηj , xk ⊗ xi − xi ⊗ xk 2i 2i =≪ Re(ηi), Re(xj ⊗ xk) ≫ε − ≪ Im(ηi), Im(xj ⊗ xk) ≫ε + − ≪ Re(ηj), Re(xk ⊗ xi) ≫ε − ≪ Im(ηj), Im(xk ⊗ xi ≫ε 2i 2i ≫ε − ≪ =≪ − ≪ ≫ε + ≫ε + By an analogous calculation we get Im(θk ij) =≪ Im(ηi), Re(xj ⊗ xk) ≫ε + ≪ Re(ηi), Im(xj ⊗ xk) ≫ε + + ≪ Im(ηj), Re(xk ⊗ xi) ≫ε − ≪ Re(ηj), Im(xk ⊗ xi ≫ε Thus we have to solve the equations < (...Re(ηi), Im(ηi), ..., Re(ηj ), Im(ηj )...), vk ij >= Re(θk ij) and being and < (...Re(ηi), Im(ηi), ..., Re(ηj ), Im(ηj )...), wk ij >= Im(θk ij) vk ij = (0, ...Re(xj ⊗ xk),−Im(xj ⊗ xk), ...0...,−Re(xk ⊗ xi),−Im(xk ⊗ xi), ...) wk ij = (0, ...Im(xj ⊗ xk), Re(xj ⊗ xk), ...0... − Im(xk ⊗ xi), Re(xk ⊗ xi)...) where the non-zero components are exactly the ones corresponding to i and j. Now observe that i, j are switchable everywhere and the case i = j is trivial. So we 7 can suppose i < j. Moreover, since the solvability of a linear system neither depend on permutations of the columns nor on multiplication by non-zero numbers, we can replace vk ij and wk ij by the following vk ij = (0, ...Re(xj ⊗ xk), Im(xj ⊗ xk), ...0... − Re(xk ⊗ xi), Im(xk ⊗ xi), ...) wk ij = (0, ... − Im(xj ⊗ xk), Re(xj ⊗ xk), ...0...Im(xk ⊗ xi), Re(xk ⊗ xi), ...) Such a re-writing concludes the first step. Step 2. The purpose of this step is to find a deformation of the xi's (namely: a perturbation of the scalar product) such that the new vk ij's become linearly independent so that we ij's, wk can solve the equations in Step 1. Suppose we have a linear combination which gives 0: X αk ijvk ij + X βk ijwk ij = 0 2≤i<j≤n,2≤k≤n 2≤i<j≤n,2≤k≤n Now fix i and look at this relation in the i-th component. We have of course the case i < j, but also a contribution that can be obtained from some j′ < i. So we can split the previous condition in the following ones: X (αk ijRe(xj ⊗ xk) − βk ijIm(xj ⊗ xk))+ 2≤i<j≤n,2≤k≤n and + X 2≤j ′<i≤n,2≤k′≤n X 2≤i<j≤n,2≤k≤n + X 2≤j ′<i≤n,2≤k′≤n (−αk′ j ′iRe(xk′ ⊗ xj ′) + βk′ j ′iIm(xk′ ⊗ xj ′)) (αk ijIm(xj ⊗ xk) + βk ijRe(xj ⊗ xk))+ (αk′ j ′iIm(xk′ ⊗ xj ′) + βk′ j ′iRe(xk′ ⊗ xj ′)) Now let si semicircular (see [Vo]) and ε′ > 0 small enough. Semicircularity guarantees √ε′si = xi are still an orthonormal basis, for any ε′ > 0. The choice of that √1 − ε′xi ⊕ ε′ small enough guarantees that the scalar product ≪ xi ⊗ xj, xk ⊗ xl ≫ε:=≪ xi ⊗ xj, xk ⊗ xl ≫ 8 is an ε-deformation. Moreover observe that in this deformation xi ⊗ xk are linearly In particular Re(xi ⊗ xj) and Im(xi ⊗ xj) independent and independent from xi ⊗ 1. are linearly independent over the real numbers. It follows that we can separate real and imaginary part in the previous conditions and get X 2≤i<j≤n,k=2,...n ijRe(xj ⊗ xk) − X αk 2≤j ′<i≤n,k′=2,...n αk′ j ′iRe(xk′ ⊗ xj ′) = 0 X 2≤i<j≤n,k=2,...n ijIm(xj ⊗ xk) + X αk 2≤j ′<i≤n,k′=2,...n αk′ j ′iIm(xk′ ⊗ xj ′) = 0 − X 2≤i<j≤n,k=2,...n ijIm(xj ⊗ xk) + X βk 2≤j ′<i≤n,k′=2,...n βk′ j ′iIm(xk′ ⊗ xj ′) = 0 X 2≤i<j≤n,k=2,...n ijRe(xj ⊗ xk) + X βk 2≤j ′<i≤n,k′=2,...n βk′ j ′iRe(xk′ ⊗ xj ′) = 0 Now, let us consider the first two conditions. If in the first sum i < k or in the second sum i > k′, the respective terms cannot cancel each other, so their coefficients must be zero. So one can have a term in the first sum equal to one in the second sum only in case i > k, i < k′, k corresponds to j′ in the second sum and k′ corresponds to j in the first sum. In this case one has αk j ′i = 0 from the second one. It follows that these coefficients must be zero. Similarly we obtain that ij − αk′ j ′i = 0 from the first condition and αk ij + αk′ the β's are equal to zero. Step 3. Here we want to define the scalar product ≪ ·,· ≫ε whenever Y appears twice. Recalling that the following properties have to be satisfied 1. ≪ Y ⊗ xi, Y ⊗ xj ≫ε= ≪ xi ⊗ Y, xj ⊗ Y ≫ε 2. ≪ Y ⊗ xi, Y ⊗ xj ≫ε= ≪ Y ⊗ Y, xi ⊗ xj ≫ε it follows that it will be enough to define the numbers ≪ Y ⊗ xi, Y ⊗ xj ≫ε and ≪ Y ⊗ xi, xj ⊗ Y ≫ε. So, we can define the matrix (≪ Y ⊗ xi, Y ⊗ xj ≫ε) as any positive matrix (Indeed the perturbation in the second step causes xi ⊗ xj to be linearly independent with respect to ≪ ·,· ≫ε and then the second of the previous conditions gives no further constrictions). Finally we can set ≪ Y ⊗ xi, xj ⊗ Y ≫ε= 0. 9 Forth Step: We can complete the proof very easily. Indeed we can set ≪ Y ⊗ xi, Y ⊗ Y ≫ε= 0, without contradictions. Finally, Bessel's inequality forces ≪ Y ⊗ Y, Y ⊗ Y ≫ε≥ X i,j ≪ Y ⊗ Y, xi ⊗ xj ≫ε 2 Also in this case there are no contradictions: it is enough to choose ≪ Y ⊗ Y, Y ⊗ Y ≫ large enough. (What is the smallest possible value?) Corollary 8. Let V be a cyclic finite dimensional space with orthonormal basis x1, ...xn. Then for every ε > 0 there exists a countably generated cyclic space Wε, with cyclic structure ≪ ·,· ≫ε, that verifies the following properties 1. W extends V as a vector space, i.e. {x1, ...xn, xn+1, ...} of W . the set {x1, ...xn} extends to a basis 2. ≪ ·,· ≫ε V is an ε-deformation of the cyclic structure on V . 3. dε(xi ⊗ xj, Wε ⊗ 1) = 0 for every i, j ∈ N, where dε(x, y) = √≪ x − y, x − y ≫ε Proof. It is enough to iterate the previous lemma, taking ε/2n at step n. 4 Problems we were not able to solve Let V be a separable cyclic vector space with orthonormal basis {xn}. We can think about xi as the operator on V defined by setting xi(xj) = xi ⊗ xj. Prop.6 guarantees that this operator is well defined by linearity, but the problem is that it could be unbounded. Indeed xi2 = Pn αn iixn2 could be infinite. So we have several open questions 1. Is the set of operators obtained in such a way a tracial algebra or at least as unbounded algebra of operators of type II in the sense of Inoue (see [In])? 2. Can we modify the proof of Prop.6 in such a way that we get bounded operators? 3. What is the relation between this construction and that of Netzer and Thom (see [Ne-Th]), who seem to obtain similar objects? 10 5 Acknowledgement The authors are grateful to Robin Hillier for reading the draft of the paper and for suggesting a correction. References [Br] N.P. Brown, Connes' embedding problem and Lance's WEP, Int.Math.Rev.Notices (2004) 10, 501-510. [Br2] , N.P.Brown, Topological Dynamical Systems Associated to II1-factors, preprint arXiv:1010.1214 [Ca-Pa] V. Capraro - L. Paunescu, Product between ultrafilter and applications to the Connes' embedding problem, preprint arXiv:math/0911.4978. Accepted by J. Oper. Theory. [Co] A.Connes, Classification of injective factors. Cases II1, II∞. III1, Ann. of Math. (2) 104 (1976) no. 1, 73-115. [Co-Dy] B. Collins and K. Dykema, Linearization of Connes' embedding problem, New York J. Math. 14 (2008) 617-641. [El-Sz] G. Elek - E. Szab´o, Hyperlinearity, essentially free actions and L2-invariants. The sofic property, Math. Ann. 332 (2005) n.2, 421-441. [Ha-Wi2] U. Haagerup and C. Winslow, The Effros-Marechal topology in the space of von Neumann algebras, II, J. Math. Anal. 171 (2000), 401-431. [Ki] E. Kirchberg, On semisplit extensions, tensor products and exactness of group C ∗- algebras, Inv. Math. 112 (1993), 449-489. [In] A. Inoue, Tomita-Takesaki theory in algebras of unbounded operators, Lecture notes in Mathematics 1699, Springer Verlag. [Ne-Th] T.Netzer and A. Thom, Tracial algebras and an embedding theorem, arxiv:math/1005.0823. 11 [Oz] N. Ozawa, About the QWEP conjecture, Intern. J. Math. 15 (2004), 501-530. [Pe] V. Pestov, Hyperlinear and sofic groups: a brief guide, The Bulletin of Symbolic Logic, vol. 14, number 4 (2008). [Ra1] F. Radulescu, The von Neumann algebras of the non-residually finite Baumslag group < a, bab3a−1 = b2 > embeds into Rω, Theta Ser. Adv. Math., vol. 9, Theta, Bucharest, 2008. [Ra2] F. Radulescu, A non-commutative, analytic version of Hilbert's 17-th problem in type II1 von Neumann algebras, To appear in Proceedings Theta Foundation. [Ra3] F. Radulescu, Convex sets associated with von Neumann algebras and Connes' approximate embedding problem, Math. Res. Lett. 6 (1999), no.2, 229-236. [Vo] D.Voiculescu, Circular and semicircular system and free product factor, Progress in Math., vol.92, Birkhauser, 1990. [Vo2] D. Voiculescu, The analogues of entropy and of Fischer's information measure in free probability theory, I, Commun. Math. Phys. 155 (1993), 71-92. [Vo3] D. Voiculescu, The analogues of entropy and of Fischer's information measure in free probability theory, II, Invent. Mat. 118 (1994), 411-440. Valerio CAPRARO - University of Rome "Tor Vergata" - [email protected] or [email protected] Florin R ADULESCU - University of Rome "Tor Vergata" - [email protected] 12
1811.05789	4	1811	2019-10-31T13:00:41	Dilations of markovian semigroups of Fourier multipliers on locally compact groups	[ "math.OA" ]	We prove that any weak* continuous semigroup $(T_t)_{t \geq 0}$ of Markov Fourier multipliers acting on a group von Neumann algebra $\mathrm{VN}(G)$ associated to a locally compact group $G$ can be dilated by a weak* continuous group of Markov $*$-automorphisms on a bigger von Neumann algebra. Our construction relies on probabilistic tools and is even new for the group $\mathbb{R}^n$. Our results imply the boundedness of the McIntosh's $\mathrm{H}^\infty$ functional calculus of the generators of these semigroups on the associated noncommutative $\mathrm{L}^p$-spaces.	math.OA	math	Dilations of markovian semigroups of Fourier multipliers on locally compact groups Cédric Arhancet Abstract We prove that any weak* continuous semigroup (Tt)t>0 of Markov Fourier multipliers acting on a group von Neumann algebra VN(G) associated to a locally compact group G can be dilated by a weak* continuous group of Markov ∗-automorphisms on a bigger von Neumann algebra. Our construction relies on probabilistic tools and is even new for the group Rn. Our results imply the boundedness of the McIntosh's H∞ functional calculus of the generators of these semigroups on the associated noncommutative Lp-spaces. 1 Introduction The study of dilations of operators is of central importance in operator theory and has a long tradition in functional analysis. Indeed, dilations are powerful tools which allow to reduce general studies of operators to more tractable ones. Suppose 1 < p < ∞. In the spirit of Sz.-Nagy's dilation theorem for contractions on Hilbert spaces, Fendler [Fen1] proved a dilation result for any strongly continuous semigroup (Tt)t>0 of positive contractions on an Lp-space Lp(Ω). More precisely, this theorem says that there exists a (bigger) measure space Ω′, two positive contractions J : Lp(Ω) → Lp(Ω′) and P : Lp(Ω′) → Lp(Ω) and a strongly continuous group of positive invertible isometries (Ut)t∈R on Lp(Ω′) such that (1.1) Tt = P UtJ for any t > 0, see also [Fen2]. Note that in this situation, J : Lp(Ω) → Lp(Ω′) is an isometric embedding whereas J P : Lp(Ω′) → Lp(Ω′) is a contractive projection. In the noncommutative setting, measure spaces and Lp-spaces are replaced by von Neumann algebras and noncommutative Lp-spaces and positive maps by completely positive maps. In their remarkable paper [JLM], Junge and Le Merdy essentially1 showed that there is no hope to have a "reasonable" analog of Fendler's result for semigroups of completely positive contractions acting on noncommutative Lp-spaces. It is a striking difference with the world of classical (=commutative) Lp-spaces of measure spaces. The semigroups of selfadjoint Markov Fourier multipliers plays a fundamental role in non- commutative harmonic analysis and operator algebras, see e.g. [Haa1]. In this paper, our main 2010 Mathematics subject classification: Primary 47A20, 47D03, 46L51 ; Secondary 47D07. Key words and phrases: semigroups, dilations, Markov operators, von Neumann algebras, noncommutative Lp-spaces, functional calculus, Fourier multipliers. 1. The authors prove that there exists no "reasonable" analog of a variant of Fendler's result for a discrete semigroup (T k)k>0 of completely positive contractions. 1 result (Theorem 3.1) gives a dilation of weak* continuous semigroups of Markov Fourier mul- tipliers acting on the group von Neumann algebra VN(G) of a locally compact group G in the spirit of (1.1) but at the level p = ∞. Our construction induces an isometric dilation similar to the one of Fendler's theorem for the strongly continuous semigroup induced by the semigroup (Tt)t>0 on the associated noncommutative Lp-space Lp(VN(G)) for any 1 6 p < ∞. Note that our paper [Arh4] gives a nonconstructive and complicated proof of such a dilation and effective only for discrete groups. Here our approach is very different, direct and short. In addition, it can also be used with non-discrete locally compact groups. We refer to [Arh1], [Arh2], [Arh4], [ALM], [AFM], [HaM] and [Ric] for strongly related things. One of the important consequences of Fendler's theorem is the boundedness, for the genera- tor of a strongly continuous semigroup (Tt)t>0 of positive contractions, of a bounded H∞ func- tional calculus which is a fundamental tool in various areas: harmonic analysis of semigroups, multiplier theory, Kato's square root problem, maximal regularity in parabolic equations, con- trol theory, etc. For detailed information, we refer the reader to [Haa], [JMX], [KW], to the survey [LeM1] and to the recent book [HvNVW2] and references therein. Our theorem also gives a similar result on H∞ functional calculus in the noncommutative context as explained in [JMX, Proposition 3.12] and [JMX, Proposition 5.8] in the case of semigroups of Fourier multipliers. The paper is organized as follows. The next Section 2 gives background. In particular, we give some information on crossed products since our construction relies on this notion. We also prove some elementary results which will be used in the sequel. Section 3 gives a proof of our main result of dilation of semigroups of Markov Fourier multipliers. In the last section 4, we describe some applications of our results to functional calculus. 2 Preliminaries Isonormal processes Let H be a real Hilbert space. An H-isonormal process on a probability space (Ω, µ) [Nua1, Definition 1.1.1] [Neer1, Definition 6.5] is a linear mapping W : H → L0(Ω) with the following properties: (2.1) for any h ∈ H the random variable W(h) is a centered real Gaussian, (2.2) (2.3) for any h1, h2 ∈ H we have E(cid:0)W(h1)W(h2)(cid:1) = hh1, h2iH . The linear span of the products W(h1)W(h2) · · · W(hm), with m > 0 and h1, . . . , hm in H, is dense in the real Hilbert space L2 R(Ω). Here L0(Ω) denote the space of measurable functions on Ω and we make the convention that the empty product, corresponding to m = 0 in (2.3), is the constant function 1. If (ei)i∈I is an orthonormal basis of H and if (γi)i∈I is a family of independent standard Gaussian random variables on a probability space Ω then for any h ∈ H, the family (γihh, eii)i∈I is summable in L2(Ω) and (2.4) W(h) def = Xi∈I γihh, eiiH , h ∈ H define an H-isonormal process. Recall that the span of elements eiW(h) is weak* dense in L∞(Ω) by [Jan1, Remark 2.15]. Using [HvNVW2, Proposition E.2.2], we see that (2.5) E(cid:0)eitW(h)(cid:1) = e− t 2khk2 H , t ∈ R, h ∈ H. 2 If u : H → H is a contraction, we denote by Γ∞(u) : L∞(Ω) → L∞(Ω) the (symmetric) second quantization of u acting on the complex Banach space L∞(Ω). Recall that the map Γ∞(u) : L∞(Ω) → L∞(Ω) preserves the integral2. If u is an isometry we have (2.6) h ∈ H and Γ∞(u) : L∞(Ω) → L∞(Ω) is a ∗-automorphism of the von Neumann algebra L∞(Ω). Fur- thermore, the second quantization functor Γ satisfies the following. In the part 1, we suppose that the construction3 is given by the concrete representation (2.4). We will only use this observation in the non-discrete case. Γ∞(u)(cid:0)eiW(h)(cid:1) = eiW(u(h)), Lemma 2.1 1. If L∞(Ω) is equipped with the weak* topology then the map H → L∞(Ω), h 7→ eiW(h) is continuous. 2. If π : G → B(H) is a strongly continuous orthogonal representation of a locally compact group, then G → B(L∞(Ω)), s 7→ Γ∞(πs) is a weak* continuous4 representation on the Banach space L∞(Ω). Proof : 1. Suppose that the sequence (hn) converges to h in H. Note that there exists an at most countable subset J of I such that i ∈ I − J implies hhn, eiiH = 0 and hh, eiiH = 0. By (2.4), note that W(hn) =Pi∈J γihhn, eiiH and W(h) =Pi∈J γihh, eiiH in L2(Ω), hence almost everywhere by [HvNVW2, Corollary 6.4.4]. For any i ∈ J, the sequence (hhn, eiiH ) converges to hh, eiiH . Since J is at most countable, W(hn) converges almost everywhere to W(h). It is not difficult to conclude using the dominated convergence theorem. 2. The second part can be proved with the arguments of the proof of [JMX, Lemma 9.3]. Von Neumann algebras Let M be a von Neumann algebra. We denote by U(M ) the group of all unitaries of M . Suppose that M is equipped with a semifinite normal faithful weight ψ. We denote by nψ the left ideal of all x ∈ M such that ψ(x∗x) < ∞. Let M and N be von Neumann algebras equipped with faithful normal semifinite weights ψM and ψN . We say that a positive linear map T : M → N is weight preserving if for any x ∈ m we have ψN (T (x)) = ψM (x). Note that a locally convex space X is said to be quasi-complete if every bounded Cauchy net of X converges [Osb1, Definition 4.23]. Recall that if H is a Hilbert space and if M is a von Neumann algebra acting on H then the σ-strong* topology on M is defined by the seminorms + ψM x 7→ (cid:0)P∞n=1(kx(ξn)k2 + kx∗(ξn)k2)(cid:1) 2 where ξn ∈ H and P∞n=1 kξnk2 < ∞. The following lemma is folklore. Unable to locate this statement in the literature, we give a short argument. 1 Lemma 2.2 Let M be a von Neumann algebra acting on a Hilbert space H. Then M equipped with the canonical locally convex structure which gives the σ-strong* topology is quasi-complete. Proof : First, we prove that M equipped with its canonical locally convex structure which gives the strong topology is quasi-complete. Consider a bounded Cauchy net (xi) of elements of M for this structure. Then, for any h ∈ H, (xi(h)) is a Cauchy net of elements of H, hence converges. So (xi) is a net which converges for the strong topology and which is norm-bounded 2. That means that for any f ∈ L∞(Ω) we haveRΩ Γ∞(u)f dµ =RΩ 3. The existence of a proof of Lemma (2.1) without (2.4) is unclear. 4. That means that B(L∞(Ω)) is equipped with the point weak* topology. f dµ. 3 by the uniform boundedness principle [KaR1, Theorem 1.8.9]. It is well-known that this implies that (xi) converges strongly to some bounded linear operator x. Since each xi belongs to M . We conclude that x belongs to M , we conclude that the strong structure is quasi-complete. Recall that a net (xi) is Cauchy in the strong* topology if and only if (xi) and (x∗i ) are Cauchy in the strong operator topology and that (xi) converges to x in the strong* topology if and only if (xi) and (x∗i ) converge to x and x∗ in the strong operator topology. Using this fact, it is not difficult to see that the canonical locally convex structure which gives the strong* topology is quasi-complete. Finally, recall that by [Tak1, Lemma 2.5] the σ-strong* topology coincide with the strong* topology on bounded subsets. So the proof is complete. Group von Neumann algebras Let G be a locally compact group equipped with a fixed left invariant Haar measure µG. Let VN(G) be the von Neumann algebra of G generated by the set (cid:8)λ(g) : g ∈ L1(G)(cid:9). It is called the group von Neumann algebra of G and is equal to the von Neumann algebra generated by the set {λs : s ∈ G} where λs : L2(G) → L2(G), f 7−→ (t 7→ f (s−1t)) is the left translation by s. Crossed products We refer to [Haa2], [Haa3], [Str1], [Sun] and [Tak2]. Let M be a von Neumann algebra acting on a Hilbert space H. Let G be a locally compact group equipped with some left Haar measure µG. Let α : G → M be a representation of G on M which is weak* continuous, i.e. for any x ∈ M and any y ∈ M∗, the map G → M , s 7→ hαs(x), yiM,M∗ is continuous. For any x ∈ M , we define the operators π(x) : L2(G, H) → L2(G, H) [Str1, (2) page 263] by (2.7) def = α−1 s (x)ξ(s), ξ ∈ L2(G, H), s ∈ G. (cid:0)π(x)ξ(cid:1)(s) These operators satisfy the following commutation relation [Str1, (2) page 292]: (2.8) (λs ⊗ IdH )π(x)(λs ⊗ IdH )∗ = π(αs(x)), x ∈ M, s ∈ G. Recall that the crossed product of M and G with respect to α is the von Neumann algebra M ⋊α G = (π(M ) ∪ {λs ⊗ IdH : s ∈ G})′′ on the Hilbert space L2(G, H) generated by the operators π(x) and λs ⊗ IdH where x ∈ M and s ∈ G. By [Str1, page 263] or [Dae1, Proposition 2.5], π is a normal injective ∗-homomorphism from M into M ⋊αG (hence σ-strong* continuous). We denote by K(G, M ) the space of σ-strong* continuous function f : G → M , s 7→ fs with compact support. If f ∈ K(G, M ) then f (G) is a σ-strong* compact subset of M , hence by [Osb1, Proposition 2.7 d)] a σ-strong* bounded subset of M . Hence it is a strong bounded subset and finally a norm-bounded subset of M by the principle of uniform boundedness [KaR1, Theorem 1.8.9]. Note that by [Str1, Proposition page 186] and [Str1, page 41], the bounded function G → M , s 7→ λs ⊗ IdH is σ-strong* continuous and the norm-bounded function s 7→ π(fs) is also σ-strong* continuous. Recall that the product of M is σ-strong* continuous on bounded subsets by [BrR1, Proposition 2.4.5]. We infer5 that the function G → M ⋊β G, s 7→ π(fs)(λs ⊗ IdH ) is σ-strong* continuous with compact support. So, by Lemma 2.2 and [Bou1, Corollary 2, III page 38] we can define the element RG fs ⋊ λs dµG(s) of the crossed product M ⋊α G by (2.9) fs ⋊ λs dµG(s) ZG def = ZG π(fs)(λs ⊗ IdH ) dµG(s). 5. In the book [Str1], the author considers weak* continuous functions, it is problematic since the product of M is not weak* continuous even on bounded sets by [KaR1, Exercise 5.7.9] (indeed this latter fact is equivalent to the weak continuity of the product on bounded sets by [Tak1, Lemma 2.5]). 4 If f, g ∈ K(G, M ) then essentially by the proof of [Haa2, Lemma 2.3] and [Str1, page 289] the function g ∗ f : G → M , t →RG g(s)f (s−1t) dµG(s) is σ-strong* continuous and we have (g ∗ f )(s) ⋊ λs dµG(s). (2.10) (cid:18)ZG gs ⋊ λs dµG(s)(cid:19)(cid:18)ZG fs ⋊ λs dµG(s)(cid:19) =ZG weak* dense subalgebra of M ⋊α G. It is very easy to see that the space of elements RG fs ⋊ λs dµG(s) for f ∈ K(G, M ) is a The following is a particular case6 of [Tak3, Proposition 3.5] and its proof, see also [Tak2, Theorem 1.7 (ii) p. 241]. Note that M is abelian in the statement. With [Bou1, Proposition 2, III page 35], the last part is an easy computation left to the reader. Proposition 2.3 Let M be an abelian von Neumann algebra acting on a Hilbert space H equipped with a continuous action α of a locally compact group G. Suppose that there exists a strongly continuous function u : G → U(M ) such that (2.11) u(sr) = u(s)αs(u(r)), s, r ∈ G. Then V : L2(G, H) → L2(G, H), ξ 7→ (s 7→ u(s−1)(ξ(s))) is a unitary and we have an ∗- isomorphism U : M ⋊α G → M ⋊α G, x 7→ V xV ∗ such that (2.12) U (λs ⊗ IdH ) = π(u(s)∗)(λs ⊗ IdH ) and U (π(x)) = π(x), s ∈ G, x ∈ M. Moreover, for any f ∈ K(G, M ), we have (2.13) U(cid:18)ZG fs ⋊ λs dµG(s)(cid:19) =ZG u(s)∗fs ⋊ λs dµG(s). Now, we suppose that M is finite and equipped with a normal finite faithful trace τ . By [Haa2, Lemma 3.3] [Str1, Theorem p. 301] [Tak2, Theorem 1.17], there exists a unique normal semifinite faithful weight ϕ⋊ on M ⋊β G which satisfies for any f, g ∈ K(G, M ) the fundamental "noncommutative Plancherel formula" (2.14) ϕ⋊(cid:18)(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)∗(cid:18)ZG gs ⋊ λs dµG(s)(cid:19)(cid:19) =ZG τ (f∗s gs) dµG(s) and the relations σϕ⋊ t (π(x)) = π(x) where x ∈ M and t ∈ R and σϕ⋊ t (λs ⊗ IdH ) = ∆it G(s)(λs ⊗ IdH )π([D(τ ◦ αs) : Dτ ]t), s ∈ G, t ∈ R. If M = C, we recover the Plancherel weight ϕG on VN(G). preserving, we obtain in particular If each αs : M → M is trace σϕ⋊ t (λs ⊗ IdH ) = ∆it G(s)(λs ⊗ IdH), s ∈ G, t ∈ R. Using [Bou1, Proposition 2, III page 35], we deduce that (2.15) σϕ⋊ t (cid:18)ZG fs ⋊ λs dµG(s)(cid:19) =ZG 6. The function u : G → U(M ) is a α-1-cocycle. ∆it G(s)fs ⋊ λs dµG(s), f ∈ K(G, M ), t ∈ R. 5 Fourier multipliers Let G be a locally compact group. We say that a weak* continuous operator T : VN(G)) → VN(G) is a Fourier multiplier if there exists a continuous function φ : G → C such that for any s ∈ G we have T (λs) = φ(s)λs. In this case φ is bounded and for any f ∈ L1(G) the element RG φ(s)f (s)λs dµG(s) belongs to VN(G) and (2.16) φ(s)f (s)λs dµG(s), i.e. T (λ(f )) = λ(φf ). T(cid:18)ZG f (s)λs dµG(s)(cid:19) =ZG In this case, we let T = Mφ and we say that φ is the symbol of T . We refer to the book [KaL1] and references therein for more information and to [ArK1] for Fourier multipliers on noncommutative Lp-spaces. Semigroups of Fourier multipliers Consider a locally compact group G with identity element e. Let (Tt)t>0 be a weak* continuous semigroup of selfadjoint unital completely positive Fourier multipliers. For any t > 0, let φt : G → C be the continuous symbol of Tt. Since each Tt is selfadjoint, φt is real-valued. By [DCH, Proposition 4.2], the function φt is of positive type. Moreover, for any t, t′ > 0 and any s ∈ G, the relation TtTt′(λs) = Tt+t′ (λs) gives φt(s)φt′ (s) = φt+t′ (s). Furthermore, for any s ∈ G, any x ∈ VN(G)∗ and any t > 0, we have φt(s)hλs, xiVN(G),VN(G)∗ = hTt(λs), xiVN(G),VN(G)∗ −−−→ t→0 hλs, xiVN(G),VN(G)∗ . We infer that limt→0 φt(s) = 1. Consequently, there exists a uniquely determined real number ψ(s) such that φt(s) = e−tψ(s), t > 0, s ∈ G. Since each Tt is unital, we have ψ(e) = 0. It is easy to check that ψ : G → R is continuous. By Schoenberg's theorem [BHV, Corollary C.4.19], we deduce that the function ψ is conditionally of negative type. By [BHV, Proposition 2.10.2], there exist a real Hilbert space H and an affine isometric action β : G → Isom(H), s 7→ βs such that the linear span of {βs(0) : s ∈ G} is dense in H and such that (2.17) ψ(s) = kβs(0)k2 H, s ∈ G. By [BHV, page 75], we can consider the associated orthogonal representation π : G → B(H), s 7→ πs of G on H. By [BHV, Lemma 2.2.1], there exists a 1-cocycle bψ : G → H with respect to π, i.e. a continuous function satisfying the cocycle law (2.18) πs(bψ(r)) = bψ(sr) − bψ(s), i.e. bψ(sr) = bψ(s) + πs(bψ(r)), s, r ∈ G, such that (2.19) βs(h) = πs(h) + bψ(s), h ∈ H. For any s ∈ G, we deduce that (2.20) ψ(s) (2.17) = kβs(0)k2 H 6 (2.19) = kbψ(s)k2 H . 3 Dilations of semigroups of Fourier multipliers The following theorem is our main result. Theorem 3.1 Let G be a locally compact group. Let (Tt)t>0 be a weak* continuous semigroup of selfadjoint unital completely positive Fourier multipliers on the group von Neumann algebra VN(G). Then, there exist a von Neumann algebra M equipped with a semifinite normal faithful weight ϕM , a weak* continuous group (Ut)t∈R of weight preserving ∗-automorphisms of M , a unital weight preserving injective normal ∗-homomorphism J : VN(G) → M such that (3.1) σϕM t ◦ J = J ◦ σϕG t for any t ∈ R and satisfying (3.2) Tt = EUtJ for any t > 0, where E : M → VN(G) is the canonical faithful normal weight preserving condi- tional expectation associated with J. Moreover, we have the following properties. 1. If G is discrete then ϕM is a normal finite faithful trace. 2. If G is unimodular then ϕM is a normal semifinite faithful trace. 3. If G is amenable then the von Neumann algebra M is injective. Proof : Here we use the continuous function ψ : C → R, the 1-cocycle bψ : G → H and the orthogonal representation π : G → B(H), s 7→ πs of G on H of Section 2. Let W : H → L0(Ω) be an H-isonormal process on a probability space (Ω, µ), see again Section 2. For any s ∈ G, we = Γ∞(πs) : L∞(Ω) → L∞(Ω) which is integral preserving. will use the second quantization αs In particular, if r, s ∈ G and if t ∈ R, we have def (3.3) αs(cid:0)e−√2itW(bψ(r))(cid:1) = Γ∞(πs)(cid:0)e−√2itW(bψ(r))(cid:1) (2.6) = e−√2itW(πs(bψ(r))). Since π is strongly continuous, by Lemma 2.1, we obtain a continuous action α : G → Aut(L∞(Ω)). def = L∞(Ω) ⋊α G equipped with its canonical normal So we can consider the crossed product M def = ϕ⋊. We denote by J : VN(G) → L∞(Ω) ⋊α G, λs 7→ 1 ⋊ λs semifinite faithful weight ϕM the canonical unital normal injective ∗-homomorphism. Using [Bou1, Proposition 2, III page 35], for any f ∈ K(G), we see that (3.4) J(cid:18)ZG f (s)λs dµG(s)(cid:19) =ZG f (s)1 ⋊ λs dµG(s). Since each αs : L∞(Ω) → L∞(Ω) is integral preserving, we have for any t ∈ R and any f ∈ K(G) σϕ⋊ t ◦ J(cid:18)ZG = J(cid:18)ZG f (s)λs dµG(s)(cid:19) (3.4) t (cid:18)ZG G(s)f (s)λs dµG(s)(cid:19) (2.15) = σϕ⋊ f (s)1 ⋊ λs dµG(s)(cid:19) (2.15) = ZG f (s)λs dµG(s)(cid:19). t (cid:18)ZG = J ◦ σϕG (3.4) ∆it ∆it G(s)f (s)1 ⋊ λs dµG(s) By density, we obtain (3.1). 7 Lemma 3.2 We have ϕ⋊ ◦ J = ϕG. : We will use [Str1, Theorem 6.2] with the weights ϕG and ϕ⋊ ◦ J. Note7 that λ(K(G)) Proof is a ∗-subalgebra which is σϕG -invariant, weak* dense in VN(G) and included in nϕG. For any f ∈ K(G), we have f (·)1 ∈ K(G, L∞(Ω)). So J(λ(f )) = RG f (s)1 ⋊ λs dµG(s) belongs to nϕ⋊. Hence ϕ⋊ ◦ J(cid:0)λ(f )∗λ(f )(cid:1) = ϕ⋊(cid:0)J(λ(f ))∗J(λ(f ))(cid:1) < ∞. Consequently, we have λ(f ) ∈ nϕ⋊◦J . We infer that nϕ⋊◦J is weak* dense in VN(G). By [Str1, page 19], we conclude that the weight ϕ⋊ ◦ J is semifinite. It is obvious that ϕ⋊ ◦ J is normal. For any t ∈ R, note that (3.4) ϕ⋊ ◦ J ◦ σϕG t (3.1) = ϕ⋊ ◦ σϕ⋊ t ◦ J = ϕ⋊ ◦ J. So the weights ϕ⋊ ◦ J and ϕG commutes by [Str1, pages 67-68]. Now, if f ∈ K(G), we have (3.4) ϕ⋊ ◦ J(cid:0)λ(f )∗λ(f )(cid:1) = ϕ⋊(cid:0)J(λ(f ))∗J(λ(f ))(cid:1) f (s)1 ⋊ λs dµG(s)(cid:19)∗ZG f (s)21 dµ(cid:19) dµG(s) =ZG = ϕ⋊(cid:18)(cid:18)ZG = ZG(cid:18)ZΩ (2.14) f (s)1 ⋊ λs dµG(s)(cid:19) f (s)2 dµG(s) (2.14) = ϕG(cid:0)λ(f )∗λ(f )(cid:1). We conclude with [Str1, Theorem 6.2] that ϕ⋊ ◦ J = ϕG. For any t ∈ R, we consider the function ut : G → U(L∞(Ω)), s 7→ e−√2itW(bψ (s)). The map bψ : G → H is continuous. By the point 1 of Lemma 2.1, the map H → L∞(Ω), h 7→ eiW(h) is continuous if L∞(Ω) is equipped with the weak* topology, hence with the weak operator topology when we consider that L∞(Ω) acts on L2(Ω). By composition, the function ut is continuous. Recall that by [KaR1, Exercice 5.7.5] or [Str1, page 41] the weak operator topology and the strong operator topology coincide on the unitary group U(L∞(Ω)). So the function ut is continuous if U(L∞(Ω)) is equipped with the strong operator topology. For any t ∈ R and any r, s ∈ G, note that ut(sr) = e−√2itW(bψ(sr)) (2.18) = e−√2itW(bψ(s))e−√2itW(πs(bψ(r))) (3.3) = ut(s)αs(cid:0)e−√2itW(bψ(r))(cid:1) = ut(s)αs(ut(r)). Hence (2.11) is satisfied. By Proposition 2.3, for any t ∈ R, we have a unitary Vt : L2(G, L2(Ω)) → L2(G, L2(Ω)), ξ 7→ (s 7→ ut(s−1)ξ(s)) and a ∗-isomorphism Ut : L∞(Ω) ⋊α G → L∞(Ω) ⋊α G, x 7→ VtxV ∗t such that (3.5) Ut(cid:18)ZG fs ⋊ λs dµG(s)(cid:19) (2.13) = ZG √2itW(bψ(s))fs ⋊ λs dµG(s) e for any f ∈ K(G, M ). For any t, t′ ∈ R any ξ ∈ L2(G, L2(Ω)), note that almost everywhere (cid:0)VtVt′ (ξ)(cid:1)(s) = ut(cid:0)s−1(cid:1)(ut′ (s−1)ξ(s) = e−√2itW(bψ(s−1))e−√2it′W(bψ(s−1))ξ(s) = e−√2i(t+t′)W(bψ(s−1))ξ(s) = ut+t′ (s−1)ξ(s) =(cid:0)Vt+t′ (ξ)(cid:1)(s). f (s)λs dµG(s). 7. Recall that λ(f ) =RG 8 We conclude that VtVt′ = Vt+t′ . Moreover, for any ξ, η ∈ L2(G, L2(Ω)) = L2(G × Ω), using dominated convergence theorem, we obtain (cid:10)Vt(ξ), η(cid:11)L2(G,L2(Ω)) =ZG×Ω Vt(ξ)(s)(ω)η(s, ω) dµG(s) dµ(ω) (ut(s−1)ξ(s))(ω)η(s, ω) dµG(s) dµ(ω) √2itW(bψ(s−1))(ω)ξ(s, ω)η(s, ω) dµG(s) dµ(ω). ξ(s, ω)η(s, ω) dµG(s) dµ(ω) = hξ, ηiL2(G,L2(Ω)). =ZG×Ω =ZG×Ω t→0 ZG×Ω −−−→ e So (Vt)t∈R is a weakly continuous group of unitaries hence a strongly continuous group by [Str1, Lemma 13.4] or [Tak2, page 239]. By [Tak2, page 238], we conclude that (Ut)t∈R is a weak* continuous group of ∗-automorphisms. Lemma 3.3 Each ∗-automorphism Ut preserves the weight ϕ⋊. Proof : We will use [Str1, Theorem 6.2] with the weights ϕ⋊ and ϕ⋊ ◦ Ut. Note that the space of elements RG fs ⋊ λs dµG(s) for f ∈ K(G, L∞(Ω)) is a ∗-subalgebra which is σϕ⋊ -invariant, weak* dense in L∞(Ω) ⋊α G and included in nϕ⋊. The formulas (2.15) and (3.5) show that each Ut and σϕ⋊ commute. So, we have t ϕ⋊ ◦ Ut ◦ σϕ⋊ t = ϕ⋊ ◦ σϕ⋊ t ◦ Ut = ϕ⋊ ◦ Ut. So the weights ϕ⋊ ◦ Ut and ϕ⋊ commutes by [Str1, pages 67-68]. It is (really) easy to check that the weight ϕ⋊ ◦ Ut is normal and semifinite. If f ∈ K(G, L∞(Ω)), we have ϕ⋊ ◦ Ut(cid:18)(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)∗(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)(cid:19) e (3.5) = ϕ⋊ (cid:18)Ut(cid:18)ZG = ϕ⋊ (cid:18)ZG = ZGZΩ = ϕ⋊(cid:18)(cid:18)ZG (2.14) (2.14) Ut(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)(cid:19)∗ √2itW(bψ(s))fs ⋊ λs dµG(s)(cid:19)∗(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)! e √2itW(bψ(s))fs ⋊ λs dµG(s)(cid:19)! √2itW(bψ(s))fs dµ dµG(s) =ZGZΩ f∗s fs dµ dµG(s) e−√2itW(bψ(s))f∗s e fs ⋊ λs dµG(s)(cid:19)∗(cid:18)ZG fs ⋊ λs dµG(s)(cid:19)(cid:19). We conclude with [Str1, Theorem 6.2] that ϕ⋊ ◦ Ut = ϕ⋊ for any t > 0. With [Str1, Theorem 10.1], we introduce the canonical faithful normal weight preserving conditional expectation E : L∞(Ω) ⋊α G → VN(G). Using [AcC, Theorem 7.5], it is entirely left to the reader to check that the conditional expectation is given by (3.6) E(cid:18)ZG gs ⋊ λs dµG(s)(cid:19) =ZG E(gs)λs dµG(s). 9 For any t > 0, we have EUtJ(cid:18)ZG f (s)λs dµG(s)(cid:19) (3.4) = EUt(cid:18)ZG f (s)1 ⋊ λ(s) dµG(s)(cid:19) (3.5) (3.6) (2.5) e = E(cid:18)ZG = ZG E(cid:0)e = ZG = ZG e− t √2itW(bψ(s))f (s) ⋊ λs dµG(s)(cid:19) √2itW(bψ(s))f (s)(cid:1)λs dµG(s) =ZG H f (s)λs dµG(s) =ZG = Tt(cid:18)ZG (2.16) 2 2kbψ (s)k2 e−tψ(s)f (s)λs dµG(s) (2.20) √2itW(bψ(s))(cid:1)f (s)λs dµG(s) E(cid:0)e e−tkbψ (s)k2 H f (s)λs dµG(s) f (s)λs dµG(s)(cid:19). Thus by density, for any t > 0, we obtain (3.2). Now, we prove the last assertions. Note each αs : L∞(Ω) → L∞(Ω) is integral-preserving. The first is well-known, e. g. [Ped1, Corollary 7.11.8]. The second is folklore. The third is [Ana1, Proposition page 301]. Remark 3.4 In the case of the Heat semigroup (Ht)t>0 on L∞(Rn), we can take H = Rn, bψ = IdRn and the action α is trivial. Here, we does not need a noncommutative von Neumann algebra for the dilation since M is a tensor product of commutative von Neumann algebras. This observation is a complement to [Arh4, Secton 4]. 4 Functional calculus We start with a little background on sectoriality and H∞ functional calculus. We refer to [Haa], [KW], [JMX], [HvNVW2] and [Arh2] for details and complements. Let X be a Banach space. A closed densely defined linear operator A : dom A ⊂ X → X is called sectorial of type ω if its def = {z ∈ C∗ : arg z < ω}, and for spectrum σ(A) is included in the closed sector Σω where Σω any angle ω < θ < π, there is a positive constant Kθ such that (cid:13)(cid:13)(λ − A)−1(cid:13)(cid:13)X→X 6 Kθ λ , λ ∈ C − Σθ. If −A is the negative generator of a bounded strongly continuous semigroup on a X then A is sectorial of type π 2 coincide with negative generators of bounded analytic semigroups. 2 . By [HvNVW2, Example 10.1.3], sectorial operators of type < π For any 0 < θ < π, let H∞(Σθ) be the algebra of all bounded analytic functions f : Σθ → C, equipped with the supremum norm kf kH∞(Σθ ) = sup(cid:8)f (z) : z ∈ Σθ(cid:9). Let H∞0 (Σθ) ⊂ H∞(Σθ) be the subalgebra of bounded analytic functions f : Σθ → C for which there exist s, c > 0 such that f (z) 6 czs (1+z)2s for any z ∈ Σθ. Given a sectorial operator A of type 0 < ω < π, a bigger angle ω < θ < π, and a function f ∈ H∞0 (Σθ), one may define a bounded operator f (A) by means of a Cauchy integral (see e.g. [Haa, Section 2.3] or [KW, Section 9]). The resulting mapping H∞0 (Σθ) → B(X) taking f to f (A) is an algebra homomorphism. By definition, A has a bounded H∞(Σθ) functional calculus provided that this homomorphism is bounded, that is if there exists a positive constant C such 6 Ckf kH∞(Σθ ) for any f ∈ H∞0 (Σθ). In the case when A has a dense range, the latter boundedness condition allows a natural extension of f 7→ f (A) to the full algebra H∞(Σθ). that (cid:13)(cid:13)f (A)(cid:13)(cid:13)X→X 10 Using the connection between the existence of dilations in UMD spaces and H∞ functional calculus together with the well-known angle reduction principle of Kalton-Weis relying on R- sectoriality, Theorem 3.1 allows us to obtain the following result (see [JMX, Proposition 3.12], [JMX, Proposition 5.8] and [KW, Corollary 10.9]). Here, we confine ourselves to unimodular groups by simplicity. Non-unimodular extensions and vector-valued versions of this result are left to the reader. Other applications will be given in subsequent papers, e.g. [ArK2]. Theorem 4.1 Let G be unimodular locally compact group. Let (Tt)t>0 be a weak* continuous semigroup of selfadjoint unital completely positive Fourier multipliers on VN(G). Suppose 1 < p < ∞. We let −Ap be the generator of the induced strongly continuous semigroup (Tt,p)t>0 on the Banach space Lp(VN(G)). Then for any θ > π 1 2 , the operator Ap has a completely bounded H∞(Σθ) functional calculus. p − 1 Acknowledgements. The author acknowledges support by the grant ANR-18-CE40-0021 (project HASCON) of the French National Research Agency ANR. Finally, I would like to thank the referee for useful remarks. References [AcC] L. Accardi and C. Cecchini. Conditional expectations in von Neumann algebras and a theorem of Takesaki. J. Funct. Anal. 45 (1982), 245 -- 273. [Ana1] C. Anantharaman-Delaroche. Action moyennable d'un groupe localement compact sur une algèbre de von Neumann. (French). Math. Scand. 45 (1979), no. 2, 289 -- 304. [Arh1] C. Arhancet. On Matsaev's conjecture for contractions on noncommutative Lp-spaces. Journal of Operator Theory 69 (2013), no. 2, 387 -- 421. [Arh2] C. Arhancet. Analytic semigroups on vector valued noncommutative Lp-spaces. Studia Math. 216 (2013), no. 3, 271 -- 290. [Arh4] C. Arhancet. Dilations of semigroups on von Neumann algebras and noncommutative Lp-spaces. J. Funct. Anal. 276 (2019), no. 7, 2279 -- 2314. [ALM] C. Arhancet and C. Le Merdy. Dilation of Ritt operators on Lp-spaces. Israel J. Math. 201 (2014), no. 1, 373 -- 414. [AFM] C. Arhancet, S. Fackler and C. Le Merdy. Isometric dilations and H ∞ calculus for bounded analytic semigroups and Ritt operators. Transactions of the American Mathematical Society 369 (2017), 6899 -- 6933. [ArK1] C. Arhancet and C. Kriegler. Projections, multipliers and decomposable maps on noncommu- tative Lp-spaces. Preprint, arXiv:1707.05591. [ArK2] C. Arhancet and C. Kriegler. Riesz transforms, Hodge-Dirac operators and functional calculus for Fourier multipliers. Preprint. [BHV] B. Bekka, P. de la Harpe, Pierre and A. Valette. Kazhdan's property (T). New Mathematical Monographs, 11. Cambridge University Press, Cambridge, 2008. [Bou1] N. Bourbaki. Integration. I. Chapters 1 -- 6. Translated from the 1959, 1965 and 1967 French originals by Sterling K. Berberian. Elements of Mathematics. Springer-Verlag, Berlin (2004). [BrR1] O. Bratteli and D. W. Robinson. Operator algebras and quantum statistical mechanics. 1. C ∗- and W ∗-algebras, symmetry groups, decomposition of states. Second edition. Texts and Monographs in Physics. Springer-Verlag, New York, 1987. [DCH] J. de Cannière and U. Haagerup. Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups. Amer. J. Math. 107 (1985), no. 2, 455 -- 500. 11 [Dae1] A. van Daele. Continuous crossed products and type III von Neumann algebras. London Mathematical Society Lecture Note Series, 31. Cambridge University Press, Cambridge-New York, 1978. [Fen1] G. Fendler. Dilations of one parameter semigroups of positive contractions on Lp spaces. Canad. J. Math. 49 (1997), no. 4, 736 -- 748. [Fen2] G. Fendler. On dilations and Transference for Continuous One-Parameter Semigroups of Positive Contractions on Lp-spaces. Ann. Univ. Sarav. Ser. Math. 9 (1998), no. 1. [Haa] M. Haase. The functional calculus for sectorial operators. Operator Theory: Advances and Applications, 169. Birkhäuser Verlag (2006). [Haa1] U. Haagerup. An Example of a nonnuclear C*-Algebra, which has the metric approximation property. Invent. Math. 50 (1978/79), no. 3, 279 -- 293. [Haa2] U. Haagerup. On the dual weights for crossed products of von Neumann algebras. I. Removing separability conditions. Math. Scand. 43 (1978/79), no. 1, 99 -- 118. [Haa3] U. Haagerup. On the dual weights for crossed products of von Neumann algebras. II. Application of operator-valued weights. Math. Scand. 43 (1978/79), no. 1, 119 -- 140. [HaM] U. Haagerup and M. Musat. Factorization and dilation problems for completely positive maps on von Neumann algebras. Comm. Math. Phys. 303 (2011), no. 2, 555 -- 594. [HvNVW2] T. Hytönen, J. van Neerven, M. Veraar and L. Weis. Analysis on Banach spaces, Volume II: Probabilistic Methods and Operator Theory. Springer, 2018. [Jan1] S. Janson. Gaussian Hilbert spaces. Cambridge Tracts in Mathematics, 129. Cambridge Uni- versity Press, Cambridge, 1997. [JLM] M. Junge, C. Le Merdy. Dilations and rigid factorisations on noncommutative Lp-spaces. J. Funct. Anal. 249 (2007), 220 -- 252. [JMX] M. Junge, C. Le Merdy and Q. Xu. H ∞ functional calculus and square functions on noncom- mutative Lp-spaces. Astérisque No. 305 (2006). [JMP2] M. Junge, T. Mei and J. Parcet. Noncommutative Riesz transforms-dimension free bounds and Fourier multipliers. J. Eur. Math. Soc. (JEMS) 20 (2018), no. 3, 529 -- 595. [KaL1] E. Kaniuth and A. T.-M. Lau. Fourier and Fourier-Stieltjes algebras on locally compact groups. Mathematical Surveys and Monographs, 231. American Mathematical Society, Providence, RI, 2018. [KaR1] R. V. Kadison and J. R. Ringrose. Fundamentals of the theory of operator algebras. Vol. I. Elementary theory. Reprint of the 1983 original. Graduate Studies in Mathematics, 15. American Mathematical Society, Providence, RI, 1997. [KW] P. C. Kunstmann and L. Weis. Maximal Lp-regularity for parabolic equations, Fourier multiplier theorems and H ∞-functional calculus. Functional analytic methods for evolution equations, Lect. Notes in Math. 1855, 65 -- 311, in Springer (2004). [LeM1] C. Le Merdy. Square functions, bounded analytic semigroups, and applications. Perspectives in operator theory, 191 -- 220, Banach Center Publ., 75, Polish Acad. Sci. Inst. Math., Warsaw, 2007. [Neer1] J. v. Neerven. Stochastic Evolution Equations. ISEM Lecture Notes 2007/08. [Nua1] D. Nualart. The Malliavin calculus and related topics. Second edition. Springer-Verlag, Berlin, 2006. [Osb1] M. S. Osborne. Locally convex spaces. Graduate Texts in Mathematics, 269. Springer, Cham, 2014. [Ped1] G. K. Pedersen. C ∗-algebras and their automorphism groups. London Mathematical Society Monographs, 14. Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], London-New York, 1979. 12 [PiX] G. Pisier and Q. Xu. Non-commutative Lp-spaces. 1459 -- 1517 in Handbook of the Geometry of Banach Spaces, Vol. II, edited by W.B. Johnson and J. Lindenstrauss, Elsevier (2003). [Ric] É. Ricard. A Markov dilation for self-adjoint Schur multipliers. Proc. Amer. Math. Soc. 136 (2008), no. 12, 4365 -- 4372. [Str1] S. Stratila. Modular theory in operator algebras. Translated from the Romanian by the author. Editura Academiei Republicii Socialiste România, Bucharest; Abacus Press, Tunbridge Wells, 1981. [Sun] V. Sunder. An invitation to von Neumann algebras. Springer-Verlag, New York, 1987. [Tak1] M. Takesaki. Theory of operator algebras. I. Reprint of the first (1979) edition. Encyclopaedia of Mathematical Sciences, 124. Operator Algebras and Non-commutative Geometry, 5. Springer- Verlag, Berlin, 2002. [Tak2] M. Takesaki. Theory of operator algebras. II. Encyclopaedia of Mathematical Sciences, 125. Operator Algebras and Non-commutative Geometry, 6. Springer-Verlag, Berlin, 2003. [Tak3] M. Takesaki. Duality for crossed products and the structure of von Neumann algebras of type III. Acta Math. 131 (1973), 249 -- 310. Cédric Arhancet 13 rue Didier Daurat, 81000 Albi, France URL: https://sites.google.com/site/cedricarhancet [email protected] 13
1104.1216	2	1104	2011-05-28T04:30:31	Residually finite actions and crossed products	[ "math.OA", "math.DS" ]	We study a notion of residual finiteness for continuous actions of discrete groups on compact Hausdorff spaces and how it relates to the existence of norm microstates for the reduced crossed product. Our main result asserts that an action of a free group on a zero-dimensional compact metrizable space is residually finite if and only if its reduced crossed product admits norm microstates, i.e., is an MF algebra.	math.OA	math	RESIDUALLY FINITE ACTIONS AND CROSSED PRODUCTS DAVID KERR AND PIOTR W. NOWAK Abstract. We study a notion of residual finiteness for continuous actions of discrete groups on compact Hausdorff spaces and how it relates to the existence of norm microstates for the reduced crossed product. Our main result asserts that an action of a free group on a zero-dimensional compact metrizable space is residually finite if and only if its reduced crossed product admits norm microstates, i.e., is an MF algebra. 1. Introduction Finite-dimensional approximation is a ubiquitous notion in the structure theory of C ∗-algebras, and it appears in a variety of different ways through properties like nuclearity, exactness, qua- sidiagonality, and the existence of norm microstates (i.e., being an MF algebra) [11]. Given that reduced crossed products by actions of countable discrete groups on compact metrizable spaces play an important role as examples and motivation in the study of C ∗-algebras, one would like to understand the extent to which various forms of finite-dimensional approximation in such crossed products are reflections of finite approximation properties at the level of the dynamics. Nuclearity and exactness are essentially measure-theoretic concepts and do not reflect any- thing intrinsically topological in the dynamics. Indeed nuclearity of the reduced crossed prod- uct is equivalent to the amenability of the action, which can be expressed in purely measure- dynamical terms [1], while exactness of the reduced crossed product is equivalent to the exactness of the acting group [11, Thm. 10.2.9], which can be characterized by the existence of an amenable action of the group on a compact metrizable space. On the other hand, quasidiagonality and the existence of norm microstates both involve the matricial approximation of multiplicative structure and are thus topological in nature, and their precise relation to the dynamics is for the most part poorly understood. Actually much of the difficulty stems from the group itself, as anything but the simplest geometry in the Cayley graph can cause severe complications for an operator analysis based on perturbations, and obtaining a topological understanding of the reduced crossed product amounts in part to knowning something about the reduced group C ∗-algebra, which sits inside the reduced crossed product in a canonical way. For free groups on two or more generators it is a deep theorem of Haagerup and Thorbjørnsen that the reduced group C ∗-algebra is MF [21]. For countable discrete groups G, if the reduced group C ∗-algebra C ∗ λ(G) is quasidiagonal then G is amenable [22], while C ∗ λ(G) is residually finite-dimensional if and only if G is residually finite and amenable [14, Cor. 4]. Not much else seems to be known however about quasidiagonality and the existence of norm microstates for reduced group C ∗-algebras. One important dynamical setting in which we do have a complete understanding of quasidi- agonality is that of integer actions. Pimsner showed in [31] that for a self-homeomorphism of Date: April 5, 2011. 1 2 DAVID KERR AND PIOTR W. NOWAK a compact metrizable space the following are equivalent: (i) the homeomorphism is pseudo- nonwandering (i.e., chain recurrent), (ii) the crossed product is embeddable into an AF algebra, (iii) the crossed product is quasidiagonal. Already though for Z2-actions it is not clear what dynamical condition should correspond to quasidiagonality. Lin succeeded however in showing that, for actions of Zd where d ≥ 1, embeddability of the crossed product into a simple AF algebra is equivalent to the existence of an invariant Borel probability measure of full support [26]. It is also known that almost periodic actions of certain amenable groups produce qua- sidiagonal crossed products [30]. Note that in all of these cases the reduced and full crossed products coincide and are nuclear due to the amenability of the acting group, and that a nuclear C ∗-algebra is quasidiagonal if and only if it admits norm microstates, i.e., is an MF algebra. The goal of the present paper is to initiate a dynamical study of the MF property in reduced crossed products that is primarily targeted at actions of free groups. Such a program owes its very possibility to the Haagerup-Thorbjørnsen result cited above which establishes the MF property of the reduced group C ∗-algebras of free groups on two or more generators, a conclusion whose validity is still unknown for the vast majority of nonamenable groups. For actions of a countable discrete group G on a compact metric space X we define a dynamical version of residual finiteness for groups, which we again call residual finiteness (Definition 2.1), and our main result asserts that, when G is free and X is zero-dimensional, the action is residually finite if and only if the reduced crossed product is an MF algebra (Theorem 3.10). For G = Z residual finiteness is the same as chain recurrence and the situation reduces to a special case of the equivalence of (i) and (iii) in the statement of Pimsner's theorem above. Pimnser showed in [31, Lemma 2] that, for a Z-action, chain recurrence is equivalent to every open set being topologically incompressible, and he establish the implication (iii)⇒(i) by using compressibility to construct a nonunitary isometry in the crossed product as an obstruction to quasidiagonality. For a free group Fr with r ≥ 2 we do not have an analogue of this non- compressibility characterization for residual finiteness, and so our proof that an Fr-action on a zero-dimensional space is residually finite if the reduced crossed product is MF must proceed by different means. The idea is to directly extract the finite approximations of the action from the existence of norm microstates. To carry out the required perturbation arguments we need to be able to work with projections, which explains our hypothesis that X is zero-dimensional. A residually finite action of G on X is, roughly speaking, one that approximately admits extensions consisting of actions of G on finite sets, i.e., there are actions of G on finite sets which map into X approximately equivariantly with approximately dense image (see Section 2). This can be viewed as a topological analogue of soficity for actions preserving a Borel probability measure, in which the approximations are measured in 2-norm and the maps into X must approximately push forward the uniform measure to the given measure on X. One can count the exponential growth of the number of such approximately equivariant maps up to some observational error to define a notion of entropy (see [8, 25] and especially Sections 2 and 3 of [24]) and this is one motivation for the study of finite dynamical approximations. It can be shown that all measure-preserving actions of Fr on a standard probability space are sofic (see [7]), and that the von Neumann algebra crossed product of such actions admit tracial microstates, i.e., embed into the ultrapower Rω of the hyperfinite II1 factor (use [10]). So residual finiteness in our topological context can be thought of as playing the role of measure-preservingness, while the MF property (existence of norm microstates) is the analogue of embeddability into Rω (existence of tracial microstates). Note however that in our setting one of the main points is to show that RESIDUALLY FINITE ACTIONS 3 the MF property implies residual finiteness, while in the von Neumann algebra case there is nothing to prove in this direction. After introducing residually finite actions in Section 2, we show in the first part of Section 3 that such actions give rise to an MF reduced crossed product whenever the reduced group C ∗-algebra of the acting group is MF. In fact we prove this more generally in Theorem 3.4 for actions on arbitrary separable C ∗-algebras which are quasidiagonal in the sense of Definition 3.2. Using this fact we then proceed in the second part of Section 3 to establish the equivalence of residual finiteness and the MFness of the reduced crossed product in the case of a free group acting on a zero-dimensional compact metrizable space (Theorem 3.10). Motivated by recent work of Rørdam and Sierakowski on paradoxical decompositions in the context of continuous actions on the Cantor set and purely infinite crossed products [34], we examine in Section 4 how paradoxical decomposability fits into our discussion of residual finiteness at the other extreme. Section 5 gives a list of characterizations of residual finiteness for minimal actions of free groups which incorporates paradoxical decomposability among other phenomena. In Section 6 we use spaces of probability measures to construct, for every nonamenable countable discrete group, an example of an action which is not residually finite although its restriction to any cyclic subgroup is residually finite. Finally, in Section 7 we revisit integer actions and observe in this case, combining arguments and results from [22, 3, 16], that the following conditions are equivalent: (i) the crossed product is a strong NF algebra, (ii) the OL∞ invariant of the crossed product is 1, and (iii) there is collection of transitive residually finite subsystems with dense union. We remark that Margulis and Vinberg defined in [29] a nontopological notion of residual finiteness for actions in which the maps in the finite modelling go in the other direction. We round out the introduction with a few words about notational convention. Throughout the paper G will always be a countable discrete group, with extra hypotheses added explicitly whenever appropriate. Actions of G on a compact Hausdorff space X will often be unnamed, in which case simply write G y X, and they will always be expressed using the concatenation (s, x) 7→ sx for x ∈ X and s ∈ G. When necessary we will refer to actions by means of a symbol such as α, in particular when we need to talk about the induced action on C(X), which will actually be expressed using this symbol, i.e., αs(f )(x) = f (s−1x) for f ∈ C(X), x ∈ X, and s ∈ G. We write λ for the left regular representation of G on ℓ2(G). The reduced crossed product of a continuous action G y X will be written C(X) ⋊λ G and the reduced group C ∗- algebra of G will be written C ∗ λ(G) associated to a group element s will invariably be denoted by us. For background on crossed products and group C ∗-algebras, especially from the kind of finite-dimensional approximation viewpoint of this paper, we refer the reader to [11]. λ(G). The canonical unitary in C(X) ⋊λ G or C ∗ For a compact Hausdorff space we write MX for the space of Borel probability measures on X equipped with the weak∗ topology, under which it is a compact Hausdorff space. Whenever convenient we will simultaneously regard elements of MX as states on C(X). The unitary group of a unital C ∗-algebra A will be written U(A). Acknowledgements. The first author was partially supported by NSF grant DMS-0900938 and the second author partially supported by NSF grant DMS-0900874. We thank Adam Sierakowski for comments and corrections, and Hanfeng Li for the argument in Example 2.5 and the simple proof of Lemma 5.1. 4 DAVID KERR AND PIOTR W. NOWAK 2. Residually finite actions Definition 2.1. A continuous action of G on a compact Hausdorff space X is said to be residually finite if for every finite set F ⊆ G and neighbourhood ε of the diagonal in X × X there are a finite set E, an action of G on E, and a map ζ : E → X such that ζ(E) is ε-dense in X and (ζ(sz), sζ(z)) ∈ ε for all z ∈ E and s ∈ F . Note that if S is a generating set for G then to verify residual finiteness it is sufficient for F in the definition to be quantified over all finite subsets of S by uniform continuity. We will mostly be interested in actions on compact metrizable spaces, in which case residual finiteness can be expressed in terms of a given compatible metric. Indeed if X is metrizable in the above definition and d is a compatible metric on X then residual finiteness is the same as saying that for every finite set F ⊆ G and ε > 0 there are a finite set E, an action of G on E, and a map ζ : E → X such that ζ(E) is ε-dense in X and d(ζ(sz), sζ(z)) < ε for all z ∈ E and s ∈ F . Note that if G admits a free residually finite continuous action on a compact metrizable space, then G must be a residually finite group, as is easy to see. Indeed this is the motivation for our terminology in the dynamical setting. In the case G = Z, residual finiteness is equivalent to chain recurrence, as explained at the beginning of Section 7. If X has no isolated points then a simple perturbation argument shows that in Definition 2.1 we can always take the set E to be a subset of X and the map ζ : E → X to be the inclusion. However, for general X this is not possible, as the following example demonstrates. For each i = 1, 2, 3 take Xi to be a copy of Z and consider the two-point compactification Xi ∪ {yi, zi} where yi and zi are the points at −∞ and +∞. Let Z act on Xi ∪ {yi, zi} so that it translates Xi via addition and fixes yi and zi. Now define X by taking the quotient disjoint union of the sets Xi ∪ {yi, zi} for i = 1, 2, 3 which collapses each of the sets {z1, y2, y3} and {y1, z2, z3} to single points. The following proposition shows that in the definition of residual finiteness it is enough to verify the ε-density condition in a local way. Proposition 2.2. A continuous action of G on a compact Hausdorff space X is residually finite if and only if for every x ∈ X, finite set F ⊆ G, and neighbourhood ε of the diagonal in X × X there are a finite set E, an action of G on E, and a map ζ : E → X such that x lies in the ε-neighbourhood of ζ(E) and (ζ(sz), sζ(z)) ∈ ε for all z ∈ E and s ∈ F . Proof. For the nontrivial direction, given a finite set F ⊆ G and neighbourhood ε of the diagonal in X × X take, for every x in some finite ε-dense set D ⊆ X, a finite set Ex and an action of G on Ex, and a map ζx : Ex → X such that x lies in the ε-neighbourhood of ζx(Ex) and (ζx(sz), sζx(z)) ∈ ε for all z ∈ Ex and s ∈ F . Now set E = `x∈D Ex and define ζ : E → X so that it restricts to ζx on Ex for each x ∈ D. Then ζ(E) is ε-dense in X and (ζ(sz), sζ(z)) ∈ ε for all z ∈ E and s ∈ F , verifying residual finiteness. (cid:3) Proposition 2.3. Let X be a compact Hausdorff space and G y X a residually finite continuous action. Then there is a G-invariant Borel probability measure on X. Proof. Let Λ be the net of pairs (F, ε) where F is a finite subset of G and ε is a neighbourhood of the diagonal in X × X and the order relation (F ′, ε′) ≻ (F, ε) means that F ′ ⊇ F and ε′ ⊆ ε. For RESIDUALLY FINITE ACTIONS 5 a given (F, ε) ∈ Λ there exist, by residual finiteness, a finite set D, an action of G on E, and a map ζ : E → X such that (ζ(sz), sζ(z)) ∈ ε for all z ∈ E and s ∈ F , and we define µF,ε to be the pullback under ζ of the uniform probability measure on E, i.e., µF,ε(f ) = E−1Pz∈E f (ζ(z)) for f ∈ C(X). Now take a weak∗ limit point of the net {µF,ε}(F,ε)∈Λ, which is easily verified to be a G-invariant Borel probability measure on X. (cid:3) Example 2.4. Suppose that G is residually finite. Let X be a compact Hausdorff space. Then the Bernoulli action G y X G given by s(xt)t∈G = (xs−1t)t∈G is easily seen to be residually finite. Example 2.5. Let f be an element in the group ring ZG. Then G acts on ZG/ZGf by left translation, and this gives rise to an action αf of G by automorphisms on the compact Abelian dual group Xf := \ZG/ZGf . When f is equal to d times the unit for some d ∈ N we obtain the Bernoulli action G y {1, . . . , d}G. Suppose now that G is residually finite and f is invertible as an element in the full group C ∗-algebra C ∗(G). Then the action αf is residually finite. Indeed in this case the points in Xf which are fixed by some finite-index normal subgroup of G are dense in Xf , which can be seen as follows. Let H be a finite-index normal subgroup of G. Then the canonical surjective ring homomorphism πH : ZG → Z(G/H) induces a surjective ring homomorphism ZG/ZGf → Z(G/H)/Z(G/H)πH (f ) which in turn induces an injective group homomorphism XπH (f ) ֒→ Xf . Note that XπH (f ), identified with its image in Xf , is equal to the set of points in Xf which are fixed by H. It thus suffices to show that SH∈H XπH (f ) happens precisely when the natural homomorphism ZG/ZGf →QH∈H is dense in Xf where H denotes the collection of finite-index normal subgroups of G. This Z(G/H)/Z(G/H)πH (f ) is injective. To verify this injectivity, let g be an element in the kernel. Then for every H ∈ H there is a wH ∈ Z(G/H) such that πH(g) = wH πH(f ). Since f is invertible in C ∗(G), wH is unique and its ℓ2-norm is bounded above by some constant not depending on H. Now if H1, H2 ∈ H and H1 ⊆ H2 then the image of wH1 in Z(G/H2) under the canonical map Z(G/H1) → Z(G/H2) is equal to wH2, and so we can define the projective limit of the wH as an element w in ZG. Then g = wf . Moreover, w lies in ℓ2(G) because the ℓ2-norms of the wH are uniformly bounded, and so w has finite support since it takes integer values. Therefore g ∈ ZGf and we obtain the desired injectivity. We remark that when f is invertible in ℓ1(G) and {Gi}∞ subgroups of G with T∞ i=1 is a sequence of finite-index normal i=j Gi = {e}, the measure entropy of αf with respect to the Haar measure and the sofic approximation sequence Σ arising from {Gi}∞ i=1 is equal to the exponential growth rate of the number of Gi-fixed points and to the logarithm of the Fuglede-Kadison determinant of f in the group von Neumann algebra of G [8]. This is also true for the topological entropy with respect to Σ more generally whenever f is invertible in the full group C ∗-algebra [25]. j=1S∞ 3. Residually finite actions and MF algebras A separable C ∗-algebra is said to be an MF algebra if it can be expressed as the inductive limit of a generalized inductive system of finite-dimensional C ∗-algebras [6, Defn. 3.2.1]. This is n=1 in N [6, Thm. 3.2.2], as well as to the existence of norm microstates (see Section 11.1 of [11]). The following characterizations are tailored for our purposes. equivalent to the embeddability of A into Q∞ n=1 Mkn for some sequence {kn}∞ n=1 Mkn/L∞ Proposition 3.1. Let A be a separable C ∗-algebra. The following are equivalent. 6 DAVID KERR AND PIOTR W. NOWAK (1) A is an MF algebra, (2) for every ε > 0 and finite set Ω ⊆ A there are a k ∈ N and a ∗-linear map ϕ : A → Mk (3) there is a dense ∗-subalgebra A0 of A such that for every ε > 0 and finite set Ω ⊆ A0 there are a k ∈ N and a ∗-linear map ϕ : A0 → Mk such that kϕ(ab) − ϕ(a)ϕ(b)k < ε such that kϕ(ab) − ϕ(a)ϕ(b)k < ε and (cid:12)(cid:12)kϕ(a)k − kak(cid:12)(cid:12) < ε for all a, b ∈ Ω, and (cid:12)(cid:12)kϕ(a)k − kak(cid:12)(cid:12) < ε for all a, b ∈ Ω. Moreover, if A is unital then the maps ϕ in (2) and (3) may be taken to be unital. n=1 Mkn for some n=1 in such a way that for each a ∈ A one has limn→∞ kank = kak for every Proof. (1)⇒(2). By Proposition 2.2.3(iii) of [6], A embeds into Q∞ (an) ∈ π−1({a}) where π :Qλ Mkn →Qλ Mkn/Lλ Mkn is the quotient map. Take a linear map ψ : A →Q∞ n=1 Mkn such that π ◦ ψ = Φ. We may assume ψ to be ∗-linear by replacing it with a 7→ (ψ(a) + ψ(a∗)∗))/2 if necessary. For each n let ψn : A → Mkn be the composition of ψ with the projection onto the nth factor. Then we have limn→∞ kψn(a)k = kak and limn→∞ kψn(ab) − ψn(a)ψn(b)k = 0 for all a, b ∈ A, from which (2) follows. n=1 Mkn/L∞ sequence {kn}∞ n=1 Ωn all a, b ∈ Ωn yielding (1). n. Now define a map θ : A →Q∞ (2)⇒(3). Trivial. (3)⇒(1). Since A is separable there is an increasing sequence Ω1 ⊆ Ω2 ⊆ · · · of finite subsets n is a dense ∗-subalgebra A1 of A. For each n ∈ N take a kn ∈ N and a of A0 such that S∞ ∗-linear map ϕn : A0 → Mkn such that kϕn(ab) − ϕn(a)ϕn(b)k < ε and (cid:12)(cid:12)kϕn(a)k − kak(cid:12)(cid:12) < ε for with the canonical projection mapQ∞ multiplicative, and hence extends to an injective ∗-homomorphism A →Q∞ to be unital, since we can lift the image of the unit under Φ to a projection (pn) in Qλ Mkn, n=1 pnMknpn/L∞ producing an injective unital ∗-homomorphism from A toQ∞ n=1 pnMknpn. This enables us to arrange ψ to be unital, so that each ψn is unital. It follows that the maps ϕ in conditions (2) and (3) may be taken to be unital. (cid:3) n=1 Mkn by composing a 7→ (ϕn(a))∞ n=1 n=1 Mkn. Then θ is isometric and k=1 Mkn, In the case that A is unital, note that in the proof of (1)⇒(2) the embedding Φ may be taken n=1 Mkn →Q∞ n=1 Mkn(cid:14)L∞ n=1 Mkn(cid:14)L∞ n=1 Mkn(cid:14)L∞ Note that the maps ϕ in Proposition 3.1 are typically not bounded. A C ∗-algebra is said to be quasidiagonal if it admits a faithful representation whose image is a quasidiagonal set of operators. Voiculescu showed that a separable C ∗-algebra A is quasidiagonal if and only if for every ε > 0 and finite set Ω ⊆ A there are a n ∈ N and a contractive completely positive map ϕ : A → Mn such that kϕ(ab) − ϕ(a)ϕ(b)k < ε and kϕ(a)k ≥ kak − ε for all a, b ∈ Ω [39]. Blackadar and Kirchberg showed that a separable nuclear C ∗-algebra is an MF algebra if and only if it is quasidiagonal [6, Thm. 5.2.2]. The reduced group C ∗-algebra of a free group on two or more generators is an example of an MF algebra which is not quasidiagonal [21]. Voiculescu's abstract characterization of quasdiagonality motivates the following definition, which extends the concept of residual finiteness to actions on noncommutative C ∗-algebras in view of Proposition 3.3. We will formulate and prove Theorems 3.4 and 3.5 within this general noncommutative framework. Definition 3.2. Let α be an action of G on a separable C ∗-algebra A. We say that α is quasidiagonal if for every finite set Ω ⊆ A, finite set F ⊆ G, and ε > 0 there are an d ∈ N, an action γ of G on Md, and a unital completely positive map ϕ : A → Md such that (1) kϕ(ab) − ϕ(a)ϕ(b)k < ε for all a, b ∈ Ω, RESIDUALLY FINITE ACTIONS 7 (2) kϕ(a)k ≥ kak − ε for all a ∈ Ω, and (3) kϕ(αs(a)) − γs(ϕ(a))k < ε for all a ∈ Ω and s ∈ F . Note that the existence of a quasidiagonal action on A in the sense of the above definition implies that A is quasidiagonal as a C ∗-algebra. Proposition 3.3. Let α be a residually finite continuous action of G on a compact metrizable space X. Then α is quasidiagonal as an action of G on C(X). Proof. Fix a compatible metric d on X. Let ε > 0 and let Ω be a finite subset of C(X). By uniform continuity there is a δ > 0 such that f (x) − f (y) < ε for all f ∈ Ω and all points x, y ∈ X satisfying d(x, y) < δ. By residual finiteness there are a finite set E, an action γ of G on E, and a map ζ : E → X such that ζ(E) is δ-dense in X and d(ζ(s−1z), s−1ζ(z)) < δ for all z ∈ E and s ∈ F . Define ϕ : C(X) → C(E) by ϕ(f )(z) = f (ζ(z)) for all f ∈ C(X) and z ∈ E. Then ϕ is a homomorphism, and for all f ∈ Ω and s ∈ F we have and kϕ(f )k = sup f (x) ≥ kf k − ε x∈ζ(E) showing that α is quasidiagonal. (cid:3) kϕ(αs(f )) − γs(ϕ(f ))k = sup z∈E(cid:12)(cid:12)f (s−1ζ(z)) − f (ζ(s−1z)(cid:12)(cid:12) < ε, The proof of the following theorem is reminiscent of part of the proof of Theorem 7 in [31]. Theorem 3.4. Suppose that C ∗ a separable C ∗-algebra A. Then A ⋊λ G is an MF algebra. λ(G) is an MF algebra. Let α be a quasidiagonal action of G on Proof. We view A as acting on a separable Hilbert space H via some faithful essential represen- tation and A ⋊λ G as acting on H ⊗ ℓ2(G) in the standard way, as determined by aus(ξ ⊗ δt) = α−1 st (a)ξ ⊗ δst for a ∈ A, s, t ∈ G, and ξ ∈ H, where {δs : s ∈ G} is the canonical basis of ℓ2(G). Since α is quasidiagonal there exist, for each n ∈ N, a positive integer kn, an action γn of G on Mkn, and a unital completely positive map ϕn : A → Mkn so that (1) limn→∞ kϕn(ab) − ϕn(a)ϕn(b)k = 0 for all a, b ∈ A, (2) limn→∞ kϕn(a)k = kak for all a ∈ A, and (3) limn→∞ kϕn(αs(a)) − γn,s(ϕn(a))k = 0 for all a ∈ A and s ∈ G. Note that if we view Mkn as acting on L2(Mkn, τ ) via the GNS construction with respect to the unique tracial state τ , then the action γ is implemented through conjugation by the unitaries ws for s ∈ G defined by wsxη = γs(x)η where η is the canonical cyclic vector. Thus by replacing Mkn with B(L2(Mkn, τ )) ∼= Mk2 and relabeling we may assume that there are unitary representations wn : G → U(ℓkn 2 ) such that n lim n→∞ kϕn(αs(a)) − wn,sϕn(a)w∗ n,sk = 0 for all a ∈ A and s ∈ G. Let Ω be a finite subset of the algebraic crossed product A ⋊alg G, which by definition consists of sums of the form Ps∈F asus where F is a finite subset of G and each as lies in A. Let ε > 0. We will show the existence of a d ∈ N and ∗-linear map β : A ⋊alg G → Md such that 8 DAVID KERR AND PIOTR W. NOWAK the proof by Proposition 3.1. Regard Mkn as acting on ℓkn 2 in the standard way. For N ∈ N we define the unital completely kβ(cc′) − β(c)β(c′)k < ε and (cid:12)(cid:12)kβ(c)k − kck(cid:12)(cid:12) < ε for all c, c′ ∈ Ω. This is sufficient to complete positive map ΦN : A →L∞ wn,s ⊗ λs for s ∈ G. Define a ∗-linear map ΘN : A ⋊alg G →L∞ 2 ⊗ ℓ2(G)) generated by Mkn ⊗ 1 and the operators ΘN (aus) = (ϕN (a)wN,s ⊗ λs, ϕN +1(a)wN +1,s ⊗ λs, . . . ) Write Dn for the C ∗-subalgebra of B(ℓkn ΦN (a) = (ϕN (a), ϕN +1(a), . . . ). n=N Dn by setting n=N Mkn by for a ∈ A and s ∈ G and extending linearly, which we can do since the subspaces Aus for s ∈ G are orthogonal with respect to the canonical conditional expectation from A ⋊λ G onto A. We aim to show the existence of an N ∈ N such that (1) kΘN (cc′) − ΘN (c)ΘN (c′)k < ε/2 for all c, c′ ∈ Ω, and (2) (cid:12)(cid:12)kΘN (c)k − kck(cid:12)(cid:12) < ε/2 for all c ∈ Ω. Since Ω is finite we can find an ε′ > 0 such that, for all n ∈ N and c ∈ Ω, if(cid:12)(cid:12)kΘn(c)k2 −kck2(cid:12)(cid:12) < ε′ then (cid:12)(cid:12)kΘn(c)k − kck(cid:12)(cid:12) < ε/2. Take a finite set F ⊆ G such that for every c ∈ Ω we can write c = Ps∈F ac,sus where ac,s ∈ A for each s ∈ F . Set M = maxc∈Ω,s∈F kac,sk. For each c ∈ Ω take a unit vector ηc in H ⊗ ℓ2(G) such that kck ≤ kcηck + ε′/2 and ηc =Pt∈K ξc,t ⊗ δt for some finite set K ⊆ G and vectors ξc,t ∈ H. For notational simplicity we may assume K to be the same for all c ∈ Ω. Take a δ > 0 such that (1 + 2F K)F 2δ2 ≤ ε′/2. Since the unital completely positive maps ϕn are asymptotically multiplicative and asymptotically isometric, by the version of Voiculescu's theorem which appears as Theorem 2.10 in [9] and stems from [35] we can find a N ∈ N and a n=N ℓkn 2 → H such that unitary U :L∞ for all c ∈ Ω, s ∈ F , and t ∈ K. For s ∈ G set ws = (wN,s, wN +1,s, . . . ) ∈ L∞ asymptotic equivariance of the maps ϕk, we may assume that N is large enough so that st (ac,s))U −1 − α−1 δ 2 n=N Mkn. By the (cid:13)(cid:13)U ΦN (α−1 (cid:13)(cid:13)ΦN (α−1 (cid:13)(cid:13)wstU −1α−1 st (ac,s)(cid:13)(cid:13) < stΦN (ac,s)wst(cid:13)(cid:13) < st (ac,s) − ΦN (ac,s)wstU −1(cid:13)(cid:13) < δ st (ac,s)) − w∗ δ 2 and hence for all c ∈ Ω, s ∈ F , and t ∈ K. We may moreover assume that N is large enough so that kΘN (cc′) − ΘN (c)ΘN (c′)k < ε/2 for all c, c′ ∈ Ω, since for a, b ∈ A, and s, t ∈ G we have kΘN (ausbut) − ΘN (aus)ΘN (but)k = sup n≥N ≤ sup n≥N kϕn(aαs(b))wn,st − ϕn(a)wn,sϕn(b)wn,tk k(ϕn(aαs(b)) − ϕn(a)ϕn(αs(b)))wn,stk + sup n≥N kϕn(a)(ϕ(αs(b))wn,s − wn,sϕn(b))wn,tk, where the last two suprema tend to zero as N → ∞, and the subspaces Aus for s ∈ G linearly span A ⋊alg G. RESIDUALLY FINITE ACTIONS 9 Let c ∈ Ω. Write U for the unitary operator from (L∞ t ζ ⊗ δt. Since ηc is a unit vector, for s ∈ F we have U (ζ ⊗ δt) = U w−1 n=N ℓkn 2 ) ⊗ ℓ2(G) to H ⊗ ℓ2(G) given by (cid:13)(cid:13)(cid:13)(cid:13)Xt∈K(cid:0)wstU −1α−1 2 st (ac,s) − ΦN (ac,s)wstU −1(cid:1)ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) = Xt∈K(cid:13)(cid:13)(cid:0)wstU −1α−1 ≤ Xt∈K(cid:13)(cid:13)wstU −1α−1 st (ac,s) − ΦN (ac,s)wstU −1(cid:1)ξc,t(cid:13)(cid:13) st (ac,s) − ΦN (ac,s)wstU −1(cid:13)(cid:13) < δ2. 2 2kξc,tk2 For any vectors x1, . . . , xn, y1, . . . , yn ∈ H we have (∗) we obtain and so applying this inequality and the crude bound 2 n n 2 n n n 2 n (cid:13)(cid:13)(cid:13)(cid:13) Xi=1 xi(cid:13)(cid:13)(cid:13)(cid:13) (cid:19)2 ≤(cid:18)(cid:13)(cid:13)(cid:13)(cid:13) yi(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13) (xi − yi)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 Xi=1 +(cid:18)1 + 2(cid:13)(cid:13)(cid:13)(cid:13) yi(cid:13)(cid:13)(cid:13)(cid:13) (cid:19)(cid:13)(cid:13)(cid:13)(cid:13) yi(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13) (xi − yi)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 Xi=1 Xi=1 ΦN (ac,s)wstU −1ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) kΘN (c) U −1ηck =(cid:13)(cid:13)(cid:13)(cid:13)Xs∈F Xt∈K ≤Xs∈F Xt∈K st (ac,s)ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) kcηck2 =(cid:13)(cid:13)(cid:13)(cid:13) U −1Xs∈F Xt∈K st (ac,s)ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)Xs∈F Xt∈K ≤ kΘN (c) U −1ηck2 +(cid:0)1 + 2kΘN (c) U −1ηck(cid:1) ×(cid:18)Xs∈F(cid:13)(cid:13)(cid:13)(cid:13)Xt∈K(cid:0)wstU −1α−1 ≤ kΘN (c)k2 + (1 + 2F K)F 2δ2 wstU −1α−1 α−1 2 2 kΦN (ac,s)k ≤ F K. (cid:19)2 st (ac,s) − ΦN (ac,s)wstU −1(cid:1)ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) = kΘN (c)k2 + ε′ 2 . 2 ≥ kcηck2 − ε′/2 ≥ kck2 − ε′. Consequently (cid:13)(cid:13)ΘN (c)(cid:13)(cid:13) Next let us show that kck2 ≥ kΘN (c)k2 − ε′. Take a unit vector η in H ⊗ ℓ2(G) such that kΘN (c)k ≤ kΘN (c) U −1ηk + ε′/2. The idea is to argue as in the above paragraph, reversing the roles of c and ΘN (c) and replacing ηc with η. Notice that the only way the particular choice of the vectors ηc entered into the above estimates, besides their being of norm one, was in obtaining the bound (∗). Here however we can simply take k U −1cηk ≤ kck as the replacement for this 10 DAVID KERR AND PIOTR W. NOWAK bound, in which case kΘN (c) U −1ηk2 =(cid:13)(cid:13)(cid:13)(cid:13)Xs∈F Xt∈K 2 ≤ k U −1cηk2 +(cid:0)1 + 2k U −1cηk(cid:1) ΦN (ac,s)wstU −1ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) ×(cid:18)Xs∈F(cid:13)(cid:13)(cid:13)(cid:13)Xt∈K(cid:0)ΦN (ac,s)wstU −1 − wstU −1α−1 ≤ kck2 + 3F 2δ2 (cid:19)2 st (ac,s)(cid:1)ξc,t ⊗ δst(cid:13)(cid:13)(cid:13)(cid:13) = kck2 + ε′ 2 and hence kck2 ≥ kΘN (c) U −1ηk2 − ε′/2 ≥ kΘN (c)k2 − ε′. By our choice of ε′ we conclude that for a ∈ A and s ∈ G and extending linearly. In view of the properties of ΘN we have kθ(cc′) − Now let M be an integer greater than or equal to N to be determined shortly. Set D = n=N Dn. Define the ∗-linear map θ : A ⋊alg G → D by declaring that θ(aus) = (ϕN (a)wn,s ⊗ λs, . . . , ϕM (a)wn,s ⊗ λs) (cid:12)(cid:12)kΘN (c)k − kck(cid:12)(cid:12) < ε/2 for all c ∈ Ω, so that the map ΘN satisfies the desired properties. LM θ(c)θ(c′)k < ε/3 for all c, c′ ∈ Ω and, by taking M large enough, (cid:12)(cid:12)kθ(c)k − kck(cid:12)(cid:12) < ε/2 for all Set k = PM represented on (LM LM λ(G) as canonically 2 ) ⊗ ℓ2(G) where the latter is identified in the standard way with 2 ) with Mk we view D as a λ(G), for λ(G) is MF then n=N kn. Note that D is a C ∗-subalgebra of B(LM 2 ⊗ ℓ2(G)). Via some fixed identification of B(LM λ(G). By hypothesis C ∗ n=1 Mkn is an embedding witnessing the fact that C ∗ λ(G) is MF, and hence so is Mk ⊗ C ∗ n=N (ℓkn n=N ℓkn 2 ) ⊗ C ∗ n=N ℓkn n=N ℓkn c ∈ Ω. C ∗-subalgebra of Mk ⊗ C ∗ if C ∗ we obtain an embedding λ(G) ֒→Q∞ n=1 Mkn(cid:14)L∞ λ(G) ֒→ Mk ⊗ ∞ Yn=1 Mk ⊗ C ∗ Mkn(cid:30) ∞ Mn=1 Mkn! ∼= ∞ Yn=1(cid:0)Mk ⊗ Mkn(cid:1)(cid:30) ∞ Mn=1(cid:0)Mk ⊗ Mkn(cid:1). Thus by Proposition 3.1 there is a d ∈ N and a ∗-linear map ϕ : Mk ⊗ C ∗ for all c, c′ ∈ Ω, λ(G) → Md such that, (1) kϕ(θ(c)θ(c′)) − ϕ(θ(c))ϕ(θ(c′))k < ε/3, (2) kϕ(θ(cc′) − θ(c)θ(c′))k < kθ(cc′) − θ(c)θ(c′)k + ε/3, and Set β = ϕ ◦ θ. Then β is ∗-linear, and for all c, c′ ∈ Ω we have (3) kϕ(θ(c))k − kθ(c)k(cid:12)(cid:12) < ε/2. kβ(cc′) − β(c)β(c′)k ≤ kϕ(θ(cc′) − θ(c)θ(c′))k + kϕ(θ(c)θ(c′)) − ϕ(θ(c))ϕ(θ(c′))k ≤ kθ(cc′) − θ(c)θ(c′)k + ε 3 + ε 3 < ε and completing the proof. (cid:12)(cid:12)kβ(c)k − kck(cid:12)(cid:12) ≤(cid:12)(cid:12)kϕ(θ(c))k − kθ(c)k(cid:12)(cid:12) +(cid:12)(cid:12)kθ(c)k − kck(cid:12)(cid:12) < ε 2 + ε 2 = ε, (cid:3) RESIDUALLY FINITE ACTIONS 11 λ(G) is quasidiagonal. Let α be a quasidiagonal action of G on Theorem 3.5. Suppose that C ∗ a separable nuclear C ∗-algebra A. Then A ⋊λ G is quasidiagonal. Proof. By a result of Rosenberg (see the appendix of [22]), the quasidiagonality of C ∗ λ(G) implies that G is amenable. Since A is nuclear, it follows that the crossed product A ⋊λ G is nuclear (see Section IV.3.5 of [2]). Since C ∗ λ(G) is quasidiagonal it is an MF algebra, and so A ⋊λ G is an MF algebra by Theorem 3.4. Since separable nuclear MF algebras are quasidiagonal [6, Thm. 5.2.2], we conclude that A ⋊λ G is quasidiagonal. (cid:3) An action of G on a compact Haudorff space X is said to be topologically free if the set of points in X with trivial isotropy group is dense. Corollary 3.6. Suppose that G is amenable. Let G y X be a topologically free residually finite action on a compact metrizable space. Then C(X) ⋊λ G is quasidiagonal. Proof. Because it admits a topologically free residually finite action, G must be a residually finite group, as is easy to verify. Since the full and reduced group C ∗-algebras of G coincide by amenability, it follows that C ∗ λ(G) is quasidiagonal [14, Cor. 4]. Since C(X) is nuclear we obtain the conclusion by Theorem 3.5. (cid:3) Now we fix an r ∈ {2, 3, . . . , ∞} and concentrate on actions of the free group Fr for the remain- der of the section. Our aim is to show that, for continuous actions of Fr on a zero-dimensional compact metrizable space, residual finiteness is equivalent to the reduced crossed product being an MF algebra. Note that such a crossed product cannot be quasidiagonal since it contains C ∗ λ(Fr), which, by a result of Rosenberg [22], is not quasidiagonal since Fr is nonamenable. The property of being an MF algebra is the appropriate substitute for quasidiagonality in the nona- menable case. Indeed quasidiagonality and the property of being an MF algebra are equivalent for separable nuclear C ∗-algebras [6, Thm. 5.2.2], and so quasidiagonality in the general group action setting can viewed as an artifact of nuclearity. The following two lemmas are standard types of perturbation results. Lemma 3.7. Let η > 0. Then there is a δ > 0 such that whenever d ∈ N and p and q are projections in Md satisfying kpqk < δ there exists a projection q′ ∈ (1 − p)Md(1 − p) such that kq′ − qk < η. Proof. Let δ be a strictly positive number less than η/6 to be further specified. Let p and q be projections in some matrix algebra Md such that kpqk < δ. Set a = (1 − p)q(1 − p). Then kq − ak = kpq + qp − pqpk ≤ 3kpqk < 3δ. Since ka2 − ak ≤ ka2 − q2k + kq − ak < k(a − q)ak + kq(a − q)k + 3δ < 9δ, if δ is small enough as a function of η there exists, by the functional calculus, a projection q′ ∈ (1−p)Md(1−p) such that kq′−ak < η/2. Since δ < η/6, we have kq′−qk ≤ kq′−ak+ka−qk < η/2 + η/2 = η, as desired. (cid:3) Lemma 3.8. Let n ∈ N and ε > 0. Then there exists a δ > 0 such that whenever d ∈ N and a1, . . . , an are n self-adjoint elements in Md satisfying (1) ka2 i − aik < δ for all i = 1, . . . , n, 12 DAVID KERR AND PIOTR W. NOWAK (2) kaiajk < δ for all distinct i, j = 1, . . . , n, and i=1 ai = 1 (3) Pn there exist pairwise orthogonal projections p1, . . . , pn ∈ Md such that Pn ε for all i = 1, . . . , n. i=1 pi = 1 and kpi −aik < Proof. We may assume that ε ≤ 1. Using Lemma 3.7 we successively choose numbers δn > δn−1 > · · · > δ1 > 0 such that δn = ε/2n and, for each k = 1, . . . , n − 1, whenever d ∈ N and p and q are projections in Md satisfying kpqk < (k + 1)δk there exists a projection q′ ∈ (1 − p)Md(1 − p) such that kq′ − qk < δk+1. Let δ be a strictly positive number less than δ1/4n to be further specified. Let a1, . . . , an be n self-adjoint elements in some matrix algebra Md such that ka2 i=1 ai = 1. By the functional calculus, we may take δ to be small enough as a function of ε in order to be able to find projections q1, . . . , qn ∈ Md such that kai − qik < ε/2n for all i = 1, . . . , n. Then for all distinct i, j = 1, . . . , n we have i − aik < δ for all i = 1, . . . , n, kaiajk < δ for all distinct i, j = 1, . . . , n, andPn kqiqjk ≤ k(qi − ai)qjk + kai(qj − aj)k + kaiajk ≤ kqi − aik + (kqik + kai − qik)kqj − ajk + δ < 4δ. Using our choice of the numbers δn, . . . , δ1 as given by Lemma 3.7, we successively construct pairwise orthogonal projections p1, . . . , pn ∈ Md such that kpi − qik < ε/2n for all i = 1, . . . , n. i=1 pi and q = qk+1 More precisely, at the kth stage we use the fact that the projections p = Pk satisfy k k kpqk ≤ Xi=1 kpiqk+1k ≤ Xi=1(cid:0)kpi − qik + kqiqk+1k(cid:1) ≤ δ1 + · · · + δk + 4kδ < (k + 1)δk to obtain a projection pk+1 with kpk+1 − qk+1k < δk+1. We then have kpi − aik < ε/n ≤ ε for all i = 1, . . . , n, and n n n 1 − =(cid:13)(cid:13)(cid:13)(cid:13) (ai − pi)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 i=1 pi must be zero, i.e., Pn which shows that the projection 1 −Pn pi(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 Xi=1 (cid:13)(cid:13)(cid:13)(cid:13) ≤ kai − pik < n · ε n ≤ 1, i=1 pi = 1. (cid:3) Lemma 3.9. Let r ∈ {2, 3, . . . , ∞}. Let X be a zero-dimensional compact metrizable space and Fr y X a continuous action. Suppose that C(X) ⋊λ Fr is an MF algebra. Then the action is residually finite. Proof. Since the MF property passes to C ∗-subalgebras and residual finiteness is a condition that is witnessed on finitely many group elements at a time, we may assume that r is finite. We may also assume that X is infinite, for otherwise the action is automatically residually finite. Then by metrizability we can take an enumeration q1, q2, . . . of the projections in C(X). On X we define the compatible metric d(x, y) = ∞ Xi=1 2−jqj(x) − qj(y). RESIDUALLY FINITE ACTIONS 13 Let ε > 0. Choose a n ∈ N such that 2−n+1 < ε. Take a partition of unity P = {p1, . . . , pm} ⊆ C(X) consisting of nonzero projections such that (i) each of the projections q1, . . . , qn is a sum of projections in P, and (ii) the clopen subsets of X of which the projections in P are characteristic functions all have diameter less than ε. Write u1, . . . , ur for the canonical unitaries in the crossed product C(X) ⋊λ Fr corresponding to the standard generators s1, . . . , sr of Fr. Write Ω for the set of all elements in C(X) ⋊λ Fr of the form up or pu∗ where p ∈ P ∪ {1} and u ∈ {1, u1, . . . , ur}. Write A for the unital commutative C ∗-subalgebra of C(X) ⋊λ Fr generated by elements of the form upu∗ where p ∈ P and u ∈ {1, u1, . . . , ur}. Let δ > 0 be such that ((1+ δ)2 + (1+ δ)+ 1)δ < 1/4. Let δ′ be a strictly positive number less than δ to be further specified. Since C(X) ⋊λ Fr is MF, by Proposition 3.1 there exist a d ∈ N and a unital ∗-linear map ϕ : A → Md such that (1) (cid:12)(cid:12)kϕ(a)k − kak(cid:12)(cid:12) < δkak for all a ∈ A, (2) kϕ(ukpiu∗ (3) kϕ(ukpiu∗ (4) kϕ(piu∗ (5) kϕ(uk)ϕ(uk)∗ − 1k < δ′ for all k = 1, . . . , r. k) − ϕ(pi)ϕ(u∗ l ) − ϕ(ukpiu∗ kulpju∗ k) − ϕ(uk)ϕ(piu∗ k)k < δ for all i = 1, . . . , m and k = 1, . . . , r, k)k < δ for all i = 1, . . . , m and k = 1, . . . , r, k)ϕ(ulpju∗ l )k < δ′ for all i, j = 1, . . . , m and k, l = 1, . . . , r, 1 · · · urpir u∗ Since every minimal projection in A has the form pi0u1pi1u∗ r for some i0, . . . , ir ∈ {1, . . . , m}, by repeated use of (1) and (2) and the triangle inequality we see that by taking δ′ small enough we can arrange for kϕ(p)ϕ(p′)k and kϕ(p)2 − ϕ(p)k to be as small as we wish for all distinct minimal projections p, p′ ∈ A. Thus assuming δ′ to be small enough we can construct a unital homomorphism Ψ : A → Md such that kΨ(a) − ϕ(a)k < δkak for all a ∈ A by using Lemma 3.8 to perturb the images of the minimal projections in A under ϕ to pairwise orthogonal projections in Md and defining Ψ(p) for a minimal projection p ∈ A to be the perturbation of ϕ(p). Take a maximal commutative subalgebra B of Md containing Ψ(A). By recoordinatizing via an automorphism of Md, we may assume B to be the diagonal subalgebra of Md. In view of condition (3) above, for each k = 1, . . . , r we can define the unitary vk = ϕ(uk)/pϕ(uk)ϕ(uk)∗ in Md and we may arrange that kϕ(uk) − vkk < δ by taking δ′ to be small enough. For i = 1, . . . , m and k = 1, . . . , r we have kϕ(uk)ϕ(pi)ϕ(uk)∗ − vkΨ(pi)v∗ kk ≤ k(ϕ(uk) − vk)ϕ(pi)ϕ(uk)∗k + kvk(ϕ(pi) − Ψ(pi))ϕ(uk)∗k + kvkΨ(pi)(ϕ(uk) − vk)∗k ≤ (1 + δ)2kϕ(uk) − vkk + (1 + δ)kϕ(pi) − Ψ(pi)k + kϕ(uk) − vkk ≤ ((1 + δ)2 + (1 + δ) + 1)δ < 1 4 in which case kΨ(ukpiu∗ k) − vkΨ(pi)v∗ kk ≤ kΨ(ukpiu∗ k) − ϕ(ukpiu∗ + kϕ(uk)kkϕ(piu∗ k)k + kϕ(ukpiu∗ k) − ϕ(pi)ϕ(u∗ k) − ϕ(uk)ϕ(piu∗ k)k k)k + kϕ(uk)ϕ(pi)ϕ(uk)∗ − vkΨ(pi)v∗ kk < δ + δ + (1 + δ)δ + 1 4 < 1. 14 DAVID KERR AND PIOTR W. NOWAK k) and vkΨ(pi)v∗ It follows that Ψ(ukpiu∗ trace. But the trace of vkΨ(pi)v∗ and k = 1, . . . , r the projections Ψ(ukpiu∗ each k = 1, . . . , r we can find a permutation matrix wk ∈ Mk such that wkΨ(pi)w∗ for all i = 1, . . . , m. Note that since wk is a permutation matrix we have wkBw∗ k are homotopic projections, and thus have the same k is the same as the trace of Ψ(pi). Thus for all i = 1, . . . , m k) and Ψ(pi) have the same trace. It follows that for k = Ψ(ukpiu∗ k) k = B. Write P (B) for the pure state space of B, which has d elements. Let ω ∈ P (B). Since A is a unital C ∗-subalgebra of C(X), by Gelfand theory there exists a point xω ∈ X which, when viewed as a pure state xω on C(X), restricts to ω ◦ Ψ on A. Set ζ(ω) = xω. This defines a map ζ : P (B) → X. Now for each i = 1, . . . , m the projection Φ(pi) is nonzero since kΦ(pi)k ≥ kϕ(pi)k − δ ≥ kpik − 2δ > 0, and so there is an ω ∈ P (B) such that ω(Φ(pi)) = 1 and hence pi(xω) = xω(pi) = 1. It follows that the image of ζ is ε-dense in X since the supports of the projections p1, . . . , pm all have diameter less than ε. Let Fr act on P (B) so that for each k = 1, . . . , r the generator sk acts as the bijection ω 7→ ω ◦ Ad w∗ k. Then for ω ∈ P (B), k = 1, . . . , r, and i = 1, . . . , m we have (sk xω)(pi) = xω(α−1 sk (pi)) = xω(u∗ kpiuk) = (ω ◦ Ψ)(u∗ = xskω(pi). kpiuk) = (ω ◦ Ad w∗ k)(Ψ(pi)) so that (sk xω)(qi) = xskω(qi) for i = 1, . . . , n and hence d(skζ(ω), ζ(skω)) = d(sk xω, xskω) = 2−i(sk xω)(qi) − xskω(qi) ≤ 1 2n−1 < ε. ∞ Xi=1 Since ε was an arbitrary positive number we conclude that the action is residually finite. (cid:3) Theorem 3.10. Let r ∈ {2, 3, . . . , ∞}. Let X be a zero-dimensional compact metrizable space and Fr y X a continuous action. Then the action is residually finite if and only if C(X) ⋊λ Fr is an MF algebra. Proof. Suppose that the action is residually finite. Then by Proposition 3.3 the action is qua- sidiagonal, and since C ∗ λ(Fr) is an MF algebra [21] we infer by Theorem 3.4 that C(X) ⋊λ Fr is an MF algebra. The other direction is Lemma 3.9. (cid:3) Note that for r = 1 the conclusion of Theorem 3.10 is valid without the zero-dimensionality hypothesis by Pimsner's result from [31] (see Section 7). 4. Paradoxical decompositions For a compact Hausdorff space X we write CX for the collection of clopen subsets of X and BX for the collection of Borel subsets of X. Definition 4.1. Suppose that G acts on a set X. Let S be a collection of subsets of X. Let k and l be integers with k > l > 0. We say that a set A ⊆ X is (G, S , k, l)-paradoxical (or simply (G, S )-paradoxical when k = 2 and l = 1) if there exist A1, . . . , An ∈ S and s1, . . . , sn ∈ G i=1 1siAi ≤ l · 1A. The set A is said to be completely (G, S )-nonparadoxical if it fails to be (G, S , k, l)-paradoxical for all integers k > l > 0. such that Pn i=1 1Ai ≥ k · 1A and Pn RESIDUALLY FINITE ACTIONS 15 Remark 4.2. Suppose that S is actually a subalgebra of the power set PX , which will always be the case in our applications. Then we can express the (G, S , k, l)-paradoxicality of a set A in S by partitioning copies of A instead of merely counting multiplicities. More precisely, A is (G, S , k, l)-paradoxical if and only if for each i = 1, . . . , k there exist an ni ∈ N and j=1 Ai,j = A for each i = 1, . . . , k and the sets si,jAi,j × {mi,j} ⊆ A × {1, . . . , l} for j = 1, . . . , ni and i = 1, . . . , k are pairwise disjoint. For the nontrivial direction, observe that if A1, . . . , An and s1, . . . , sn are as in the definition of (G, S , k, l)-paradoxicality then the sets of the form Ai,1, . . . , Ai,ni ∈ S , si,1, . . . , si,ni ∈ G, and mi,1, . . . , mi,ni ∈ {1, . . . , l} so that Sni A ∩(cid:18)(cid:18) \i∈P Ai(cid:19) \ [i∈{1,...,n}\P Ai(cid:19) ∩ s−1 j (cid:18)(cid:18) \i∈Q siAi(cid:19) \ [i∈{1,...,n}\Q siAi(cid:19), where P and Q are nonempty subsets of {1, . . . , n} with P ≤ k and Q ≤ l and j ∈ P , can be relabeled so as to produce the desired Ai,j. Suppose that G acts on a set X. Let S be a G-invariant subalgebra of the power set PX of X. The type semigroup S(X, G, S ) of the action with respect to S is the preordered semigroup (cid:26)[i∈I Ai × {i} : I is a finite subset of N and Ai ∈ S for each i ∈ I(cid:27). ∼ where ∼ is the equivalence relation under which P =Si∈I Ai×{i} is equivalent to Q =Si∈J Bi× {i} if there exist ni, mi ∈ N, Ci ∈ S , and si ∈ G for i = 1, . . . , k such that P =Fk and Q =Fk Bi × {i}(cid:19)(cid:21), (cid:20)[i∈I Ai × {i}(cid:19) ∪(cid:18) [i∈J+max I Bi × {i}(cid:21) =(cid:20)(cid:18)[i∈I Ai × {i}(cid:21) +(cid:20)[i∈J i=1 siCi × {mi}. Addition is defined by i=1 Ci × {ni} and for the preorder we declare that a ≤ b if b = a + c for some c. The following is a standard observation. Lemma 4.3. Let X be a compact metrizable space and G y X a continuous action. Let B be a nonempty Borel subset of X. Suppose that there is a G-invariant Borel probability measure µ on X with µ(B) > 0. Then B is completely (G, BX )-nonparadoxical. Proof. Let µ be a G-invariant Borel probability measure on X with µ(B) > 0. Suppose that B fails to be completely (G, BX )-nonparadoxical. Then there are k, l ∈ N with k > l for each i = 1, . . . , k there are an ni ∈ N and Bi,1, . . . , Bi,ni ∈ BX, si,1, . . . , si,ni ∈ G, and mi,1, . . . , mi,ni ∈ j= Bi,j = B for each i = 1, . . . , k and the sets si,jBi,j ×{mi,j} ⊆ B ×{1, . . . , l} for j = 1, . . . , ni and i = 1, . . . , k are pairwise disjoint. Since µ is G-invariant we have and dividing by µ(B) yields k ≤ l, a contradiction. We conclude that B is completely (G, BX )- nonparadoxical. (cid:3) Lemma 4.4. Let G y X be a continuous action on a zero-dimensional compact metrizable space. Let A be a completely (G, CX )-nonparadoxical clopen subset of X such that G · A = X. Then there exists a G-invariant Borel probability measure µ on X such that µ(A) > 0. {1, . . . , l} so thatSni Xj=1 kµ(B) ≤ k Xi=1 n k n µ(Bi) = Xj=1 Xi=1 µ(si,jBi) = µ(cid:18) k [j=1 n [i=1 µ(si,jBi)(cid:19) ≤ lµ(B), 16 DAVID KERR AND PIOTR W. NOWAK Proof. We claim that there is a state σ on S(X, G, CX ) such that σ([A]) > 0. Indeed suppose that this is not the case. Since G · A = X, by compactness there is a finite set F ⊂ G such that F · A = X and hence [X] ≤ F [A]. By the Goodearl-Handelman theorem [18, Lemma 4.1] we can find n, m ∈ N such that m > nF and m[A] ≤ n[X], in which case m[A] ≤ nF [A], contradicting the complete (G, CX )-nonparadoxicality of A. Thus the desired σ exists. By Lemma 5.1 of [34] there is a G-invariant Borel measure ν on X such that ν(B) = σ([B]) for all clopen sets B ⊆ X. Since ν(X) ≥ ν(V ) > 0 and ν(X) = ν(F · A) ≤ [s∈F ν(sA) = F ν(A) < ∞, we can set µ(·) = ν(·)/ν(X) to obtain a G-invariant Borel probability measure on X. (cid:3) Proposition 4.5. Let G y X be a minimal continuous action on a zero-dimensional compact metrizable space. Then the following are equivalent: (1) there is a G-invariant Borel probability measure on X, (2) X is completely (G, CX )-nonparadoxical, (3) there is a nonempty clopen subset of X which is completely (G, CX )-nonparadoxical. Proof. Lemma 4.3 yields (1)⇒(2), while (2)⇒(3) is trivial. Since for any clopen set A ⊆ X we have G · A = X by minimality, we obtain (3)⇒(1) from Lemma 4.4. (cid:3) Proposition 4.6. Let α be a continuous action of G on a zero-dimensional compact metrizable space X such that C(X) ⋊λ G is stably finite. Then every nonempty clopen subset of X is completely (α, CX )-nonparadoxical. Proof. Write A = C(X) ⋊λ G for economy. Suppose that there is a nonempty clopen subset V of X which fails to be completely (α, CX )-nonparadoxical. Then there are k, l ∈ N with k > l, a clopen partition {V1, . . . , Vn} of V , and si,j ∈ G and mi,j ∈ {1, . . . , l} for i = 1, . . . , n and j = 1, . . . , k such that the sets si,jVi × {mi,j} ⊆ V × {1, . . . , l} for i = 1, . . . , n and j = 1, . . . , k are pairwise disjoint. For i = 1, . . . , n and j = 1, . . . , k set ai,j = emi,j ,j ⊗ usi,j 1Vi ∈ Mk ⊗ A where the first component in the elementary tensor is a standard matrix unit in Mk. Then aij is a partial isometry with Now since k in K0(A) we have Pk Xj=1 j=1Pn (ai,j)∗ai,j = ejj ⊗ 1Vi, ai,j(ai,j)∗ = emi,j mi,j ⊗ 1si,j Vi. (ai,j)∗ai,j = n Xi=1 k Xj=1 ejj ⊗ 1V = 1 ⊗ 1V , i=1[(ai,j)∗ai,j] = k[1V ]. On the other hand, k Xj=1 n Xi=1 ai,j(ai,j)∗ ≤(cid:18) l Xj=1 ejj(cid:19) ⊗ 1V , RESIDUALLY FINITE ACTIONS 17 so that in K0(A) we have k[1V ] ≤ l[1V ] and hence −[1V ] ≥ (l − k − 1)[1V ] ≥ 0. Thus there exists a projection p in some matrix algebra over A such that [p] + [1V ] = 0. We can thus find a d ∈ N and pairwise orthogonal projections p′, q, r ∈ Md ⊗ A such that (1) p′ and q are Murray-von Neumann equivalent to p and 1V , respectively, in some matrix algebra over A (viewing 1V and p as sitting in arbitrarily large matrix algebras over A as consistent with the definition of K0), and (2) p′ + q + r is Murray-von Neumann equivalent to r in Md ⊗ A. Since 1V 6= 0 the projection q is not equal to zero, and so p′ + q + r is an infinite projection. This means that C(X) ⋊λ G fails to be stably finite, contradicting (1). (cid:3) Finally we discuss the relation between residual finiteness and complete nonparadoxicality in the zero-dimensional setting. Proposition 4.7. Let G y X be a residually finite continuous action on a zero-dimensional compact metrizable space. Then every nonempty clopen subset of X is completely (G, CX )- nonparadoxical. Proof. Suppose to the contrary that there is a nonempty clopen set A ⊆ X which is (G, CX , k, l)- nonparadoxical for some integers k > l > 0. Then there exist nonempty clopen sets A1, . . . , An ⊆ i=1 1siAi ≤ l · 1A. Let d be a compatible metric on X. Take an ε > 0 which is smaller than the Hausdorff distance between Ai and X \ Ai for each i = 1, . . . , n and smaller that the Hausdorff distance between siAi and X \ siAi for each i = 1, . . . , n. By residual finiteness there is a finite set E, an action of G on E, and a map ζ : E → X such that ζ(E) is ε-dense in X and d(ζ(siz), siζ(z)) < ε for all z ∈ E and i = 1, . . . , n. Since ζ(E) is ε-dense in X, ζ −1(A) is nonempty. For each i = 1, . . . , n set Ei = ζ −1(Ai). A and s1, . . . , sn ∈ G such thatPn i=1 1Ai ≥ k · 1A andPn Each element of ζ −1(A) is contained in ζ −1(siAi) for at most l values of i, and thus, since ζ(siEi) ⊆ siAi for i = 1, . . . , n by our choice of ε, n n n lζ −1(A) ≥ ζ −1(siAi) ≥ Xi=1 Xi=1 siEi = Ei. Xi=1 that Pn On the other hand, each element of ζ −1(A) is contained in ζ −1(Ai) for at least k values of i, so (cid:3) i=1 Ei ≥ kζ −1(A). Therefore l ≥ k, a contradiction. Residual finiteness is not a necessary condition for the conclusion in Proposition 4.7, as can be seen for G = Z as follows. Recall that the group G is said to be supramenable if for every nonempty set A ⊆ G there is a finitely additive left-invariant measure µ : PX → [0, +∞] with µ(A) = 1 [33] (see Chapter 10 of [40]). Every group of polynomial growth, in particular Z, is supramenable [33]. Supramenability is easily seen to be equivalent to the property that for every action of G on a set X and every nonempty set A ⊆ X there is a finitely additive G-invariant measure µ : PX → [0, +∞] with µ(A) = 1. It follows by a standard observation as in the proof of Lemma 4.3 that, for every action of a supramenable G on a set X, every subset of X is completely (G, PX )-nonparadoxical, where PX is the power set of X. On the other hand, for Z-actions residual finiteness is the same as chain recurrence (see the beginning of Section 7), and one can easily construct a continuous Z-action on the Cantor set which fails to be chain recurrent using Lemma 2 of [31], which shows that if X a compact Haudorff space and T : X → X is a homeomorphism then T is chain recurrent if and only if there does not exist an open set U ⊆ X such that U \ T (U ) 6= ∅ and T (U ) ⊆ U . 18 DAVID KERR AND PIOTR W. NOWAK 5. Minimal actions of Fr Here we give several characterizations of residual finiteness for minimal actions of the free group Fr where r ∈ N ∪ {∞}. This is done by stringing together a variety of known results with a few of the observations from the previous sections. The following lemma is essentially extracted from the proof of Proposition 2.3 in [27]. Note that in that proof it is claimed that, given a continuous action of Fr on compact metric space X preserving a Borel probability measure, for every ε one can partition X into finitely many measurable subsets of equal (and hence rational) measure and diameter less than ε. This is false in general, as X could contain a clopen subset A of irrational measure and such an A would be equal to a union of sets in the partition whenever ε is smaller than the distance between x and y for every x ∈ A and y ∈ X \ A. The argument can be modified to take of this problem, but we will instead give a simpler proof that was supplied to us by Hanfeng Li. Lemma 5.1. Let X be a compact metrizable space and Fr y X a continuous action. Suppose there exists an Fr-invariant Borel probability measure µ on X with full support. Then the action is residually finite. Proof. We may assume that r is finite in view of the definition of residual finiteness. Write S for the standard generating set for Fr. Let ε > 0. Take a finite measurable partition P of X whose elements have nonzero measure and diameter less than ε. Write Q for the collection of sets in the join Ws∈S sP which have nonzero measure. Consider, for each P ∈ P and s ∈ S, the homogeneous linear equation PQ∈Q,Q⊆P xQ = PQ∈Q,Q⊆sP xQ in the variables xQ for Q ∈ Q. This system of equations has the solution xQ = µ(Q) for Q ∈ Q, and, since the rational solutions are dense in the set of real solutions by virtue of the rationality of the coefficients, we can find a solution consisting of rational xQ which are close enough to the corresponding quantities µ(Q) to be all nonzero. Pick a positive integer M such that M xQ is an integer for every Q ∈ Q. For each Q ∈ Q take a set EQ of cardinality M xQ and define E to be the disjoint union of these sets. Take a map ζ : E → X which sends EQ into Q for each Q ∈ Q. Now for every P ∈ P and s ∈ S the sets SQ∈Q,Q⊆P EQ and SQ∈Q,Q⊆sP EQ have the same cardinality and so we can define an action of Fr on E by having a generator s send SQ∈Q,Q⊆P EQ to SQ∈Q,Q⊆sP EQ in some arbitrarily chosen way for each P ∈ P. Then ζ and this action on E witness the definition of residual finiteness with respect to ε and the generating set S. (cid:3) In [13] Cuntz and Pedersen introduced, for an action α of G on a C ∗-algebra A, a notion of G-finiteness, which means that there do not exist distinct elements a, b ∈ A such that (i) a ≤ b and (ii) a and b are G-equivalent in the sense that there exist a collection of elements ui,si ∈ A, where i ranges in an arbitrary index set I and each si is an element of G, such that u∗ i,siui,si and a =Xi∈I αsi(ui,siu∗ i,si) b =Xi∈I where the sums are norm convergent. Theorem 8.1 of [13] asserts that A is G-finite if and only if it has a separating family of G-invariant tracial states, which is equivalent to the existence of a faithful G-invariant tracial state if A is separable. We will say that a continuous action G y on a compact Hausdorff space X is G-finite (in the Cuntz-Pedersen sense) if the induced action on C(X) is G-finite. RESIDUALLY FINITE ACTIONS 19 Theorem 5.2. Let X be a compact metrizable space and Fr y X a minimal continuous action. Then the following are equivalent: (1) the action is residually finite, (2) there is an Fr-invariant Borel probability measure on X, (3) C(X) ⋊λ Fr is an MF algebra, (4) C(X) ⋊λ Fr is stably finite, (5) the action is Fr-finite in the Cuntz-Pedersen sense, If moreover X is zero-dimensional then we can add the following conditions to the list: (6) every nonempty clopen subset of X is completely (Fr, CX)-nonparadoxical. (7) there exists a nonempty clopen subset of X which is completely (Fr, CX)-nonparadoxical. Proof. (1)⇒(2). By Proposition 2.3. (2)⇒(1). By minimality every G-invariant Borel probability measure on X has full support, and so Lemma 5.1 applies. (1)⇒(3). By Theorem 3.4. (3)⇒(4). By Proposition 3.3.8 of [6]. (4)⇒(2). Stable finiteness implies the existence of a quasitrace (see Section V.2 of [2]) and restricting a quasitrace on C(X) ⋊λ Fr to C(X) yields a G-invariant Borel probability measure on X. (2)⇔(5). Apply Theorem 8.1 of [13] as quoted above, using the fact that every Borel proba- bility measure on X has full support by the minimality of the action. Finally, in the case that X is zero-dimensional (2)⇔(6)⇔(7) is a special case of Proposition 4.5. (cid:3) The hypotheses in the above theorem actually imply conditions (1) to (5) in the case of Z, i.e., when r = 1. When r ≥ 2 these conditions may fail, as the action on the Gromov boundary (or any amenable minimal action) shows. Problem 5.3. Suppose in the above theorem that the action is topologically free and X is Cantor set. It then follows by Theorem 5.4 of [34] that if every nonempty clopen subset of X is (Fr, CX)-paradoxical then C(X) ⋊λ Fr is purely infinite. We therefore ask whether there is a dichotomy between the MF property and pure infiniteness within this class of actions. This would be the case if for any such action (Fr, CX)-nonparadoxicality implies complete (Fr, CX)- nonparadoxicality for every nonempty clopen subset of X, i.e., if the type semigroup S(X, G, CX ) is almost unperforated in the sense of Section 5 of [34]. Example 5.4. Let α be a minimal continuous action of Fr on T. Then the following are equivalent. (1) α is residually finite, (2) there is a G-invariant Borel probability measure on T, (3) C(T) ⋊λ Fn is an MF algebra, (4) C(T) ⋊λ Fn is stably finite, (5) αs is residually finite for every s ∈ G. The equivalences (1)⇔(2)⇔(3)⇔(4) are by Theorem 5.2, (1)⇒(5) is trivial, and (5)⇒(2) is a consequence of a result of Margulis [28]. Note that, by Theorem 2 of [28], if a minimal action of a countable group on the circle has an invariant Borel probability measure then it is 20 DAVID KERR AND PIOTR W. NOWAK conjugate to an isometric action with respect to the standard metric and thus factors through a group containing a commutative subgroup of index at most two. Therefore there are no faithful residually finite continuous actions of F2 on the circle. 6. Actions on spaces of probability measures In this section we will use spaces of probability measures in order to construct an example, for every nonamenable G, of a continuous action of G on a compact metrizable space such that α is not residually finite but its restriction to every cyclic subgroup of G is residually finite. The crossed product of such an action fails to be stably finite, but, unlike in the case of integer actions, one cannot witness this failure by using the compression of an open set by a single group element to construct a nonunitary isometry. See the discussion after Proposition 6.4. For a compact Hausdorff space X we will view the space MX of Borel probability measures on X with its weak∗ topology as a topological subspace of MX via the point mass identification. Proposition 6.1. Let X be a compact metrizable space and G y X a residually finite continuous action. Then the induced action G y MX is residually finite. Proof. Let Ω be a finite subset of the unit ball of C(X), and equip MX with the continuous pseudometric dΩ(σ, ω) = maxf ∈Ω σ(f ) − ω(f ). Let ε > 0. By residual finiteness there are a finite set Y , an action of G on Y , and a map ζ : Y → X such that (i) dΩ(sζ(y), ζ(sy)) < ε for all y ∈ Y and s ∈ F and (ii) ζ(Y ) is ε-dense in X with respect to dΩ. Let m ∈ N. Let E be the finite subset of MY consisting of all convex combinations of the form Py∈Y cyδy where cy ∈ {0, 1/m, 2/m, . . . , (m − 1)/m, 1} for each y ∈ Y . The the action of G on Y extends to an action of G on E by setting for s ∈ G. Define ζ : E → MX as the restriction of the push-forward map MY → MX associated cyδy(cid:19) = Xy∈Y to ζ. Then for z =Py∈Y cyδy ∈ E and s ∈ F we have s(cid:18)Xy∈Y dΩ(sζ(z), ζ(sz)) = sup cyδsy cysζ(y)(cid:19) − f(cid:18)Xy∈Y cyζ(sy)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ≤ sup cyf (sζ(y)) − f (ζ(sy)) < ε. f ∈Ω(cid:12)(cid:12)(cid:12)(cid:12) f(cid:18)Xy∈Y f ∈ΩXy∈Y Since the set of convex combinations of point masses is weak∗ dense in MX , given any η > 0 we can take ε sufficiently small and m sufficiently large to guarantee that the image of ζ is η-dense in MX with respect to dΩ. Since the pseudometrics of the form dΩ generate the uniformity on MX , we conclude that the action G y MX is residually finite. (cid:3) Proposition 6.2. Let K be a compact convex subset of a locally convex topological vector space and let T : K → K be a homeomorphism. Then the action of Z generated by T is residually finite. Proof. Let ε > 0. Let Ω be a finite set of continuous affine real-valued functions on K, and equip K with the continuous pseudometric dΩ(x, y) = maxf ∈Ω f (x) − f (y). Let ε > 0. Take a finite set V ⊆ K which is ε-dense for dΩ. Choose an m ∈ N such that 2m−1 maxf ∈Ω kf k < ε. RESIDUALLY FINITE ACTIONS 21 Set E = {(v, k) : v ∈ V and k = −m, . . . , m} and define the bijection S : E → E by setting S(v, k) = (v, k + 1) for v ∈ V and k = −m, . . . , m − 1 and S(v, m) = −m for v ∈ V . Since Z is amenable there is a T -fixed point w ∈ K. Define a map ζ : E → K by setting, for each v ∈ V and k = −m, −m + 1, . . . , m, Then for v ∈ V and k = −m, . . . , m − 1 we have ζ(v, k) =(cid:18)1 − k m(cid:19)T kv + k m w. dΩ(ζ(S(v, k)), T ζ(v, k)) = dΩ(ζ(v, k + 1), T ζ(v, k)) 1 m(cid:12)(cid:12)f (T k+1v) − f (w)(cid:12)(cid:12) kf k < ε, = max f ∈Ω 2 m ≤ max f ∈Ω while ζ(S(v, m)) = w = T ζ(v, m) so that dΩ(ζ(S(v, m)), T ζ(v, m)) = 0. Observe that by the uniform continuity of T we can enlarge m to obtain the same kind of estimates for finitely many powers of T at a time. Since the continuous pseudometrics of the form dΩ generate the topology on K, we conclude that T is residually finite. (cid:3) Proposition 6.3. Let G y X be a continuous action on a compact metrizable space. Suppose that the induced action G y MX is residually finite. Then there exists a G-invariant Borel probability measure on X. Proof. By Proposition 2.3 there is a G-invariant Borel probability measure µ on MX . Consider the barycentre bµ of µ, i.e., the unique element of K satisfying bµ(f ) = µ(f ) for all f ∈ CR(MX ), with f viewed on the right side of the equality as an affine function on MK under the canonical identification. Then bµ is a fixed point for the action G y MX , that is, bµ is a G-invariant Borel probability measure on X. (cid:3) Proposition 6.4. Suppose that G is nonamenable. Then there exists a continuous action α of G on a compact metrizable space such that α is not residually finite but its restriction to every cyclic subgroup of G is residually finite. Proof. Since G is nonamenable there is a continuous action α of G on a compact metrizable space X which does not admit an invariant Borel probability measure. By Proposition 6.3, the induced action α of G on MX is not residually finite. However, for every cyclic subgroup H of G the restriction of α to H is residually finite, which is obvious if H < ∞ and follows from Proposition 6.2 otherwise. (cid:3) Note that if G is nonamenable and G y X is an action as in the statement of Proposition 6.4 then C(X)⋊λG fails to be stably finite, for otherwise it would admit a quasitrace (see Section V.2 of [2]) and restricting a quasitrace to C(X) would produce a G-invariant Borel probability measure on X. Therefore Mk(C(X) ⋊λ G) contains a nonunitary isometry for some k ∈ N. However one cannot verify the existence of such an isometry by using the compression of an open set by a single group element as in the case of crossed products by Z [31] (see the discussion preceding Theorem 7.2 in the next section). Problem 6.5. For an action of the kind just described, exhibit a nonunitary isometry in some matrix algebra over C(X) ⋊λ G. 22 DAVID KERR AND PIOTR W. NOWAK Problem 6.6. In the statement of Proposition 6.4, can the action be taken to be minimal and/or the space taken to be the Cantor set? 7. Z-systems Here we specialize our discussion to Z-actions, for which the C ∗-algebraic notions of quasidi- agonality, strong quasidiagonality, and strong NFness all admit direct dynamical interpretations in terms of residual finiteness. Tacit use will be made of the fact that, due to the amenability of Z, the reduced crossed product associated to a continuous action of Z on a compact Hausdorff space is nuclear and coincides canonically with the full crossed product. For terminological economy we will conceive of a Z-action as a pair (X, T ) where X is a compact Hausdorff space and T : X → X is a homeomorphism associated to the generator 1, and we will call such a pair a Z-system. We will thus speak of residually finite Z-systems. A Z-system (X, T ) is said to be metrizable if X is metrizable. We write orb(x) for the orbit of the point x under T , i.e., orb(x) = {T nx : n ∈ Z}. For Z-systems, residual finiteness coincides with Conley's concept of chain recurrence [12], defined as follows. Let (X, T ) be a Z-system. Let x, y ∈ X and let ε be a neighbourhood of the diagonal in X × X. An ε-chain from x to y is a finite sequence {x1 = x, x2, . . . , xn = y} in X such that n > 1 and (T xi, xi+1) ∈ ε for every i = 1, . . . , n − 1. The point x is said to be chain recurrent if for every neighbourhood of the diagonal in X × X there is an ε-chain from x to itself (if X is a metric space then we can quantify instead over all ε > 0, with ε-chain taking its meaning in relation to the set {(x, y) ∈ X × X : d(x, y) < ε}). This is equivalent to x being pseudo-nonwandering in the sense of [31]. We remark that the set of chain recurrent points is a closed T -invariant subset of X. Finally, the system (X, T ) is said to be chain recurrent if every point in X is chain recurrent. Note that a Z-system for which there is a dense set of recurrent points is chain recurrent. Proposition 7.1. A Z-system (X, T ) is residually finite if and only if it is chain recurrent. Proof. Suppose first that (X, T ) is chain recurrent. Let F be a finite subset of Z. Let ε and ε′ be neighbourhoods of the diagonal in X × X. Take a finite ε-dense subset D of X. For each x ∈ D pick an ε′-chain from x to itself and write Cx for the sequence obtained by omitting x at the end of the ε′-chain. Set E = Fx∈D Cx, and let S : E → E be the bijection which cyclically permutes the points of each Cx according to the sequential order. Then the map ζ : E → X defined by taking the inclusion on each Cx has ε-dense image, and by uniform continuity it will satisfy (ζ(Snz), T nζ(z)) ∈ ε for every n ∈ F if ε′ is taken fine enough. Thus (X, T ) is residually finite. Conversely, suppose that (X, T ) is residually finite. Let x ∈ X and let ε be a neighbourhood of the diagonal in X × X. By residual finiteness there is a finite set E, a bijection S : E → E, a map ζ : E → X such that (ζ(Sz), T ζ(z)) ∈ ε for all z ∈ E, and a z0 ∈ E such that (ζ(z0), x) ∈ ε. By a simple perturbation argument involving a finer choice of ε, we may assume that ζ(z0) = x. Then by applying ζ to the S-cycle in which z0 lies we obtain an ε-chain for x. Hence (X, T ) is chain recurrent. (cid:3) In Lemma 2 of [31] (which is also valid in the nonmetrizable case) Pimsner showed that a Z-system (X, T ) is chain recurrent if and only if there does not exist an open set U ⊆ X which is compressed by T in the sense that U \ T (U ) 6= ∅ and T (U ) ⊆ U . From such an open set RESIDUALLY FINITE ACTIONS 23 one can construct a nonunitary isometry in the crossed product C(X) ⋊λ Z [31, Prop. 8], which establishes the implication (3)⇒(1) in the following theorem of Pimsner from [31]. Theorem 7.2. For a Z-system (X, T ) the following are equivalent: (1) (X, T ) is pseudo-nonwandering (i.e., residually finite), (2) C(X) ⋊λ Z is quasidiagonal, (3) C(X) ⋊λ Z is stably finite. Moreover, if X is metrizable then the following condition can be added to the list: (4) C(X) ⋊λ Z can be embedded into an AF algebra. Actually only the metrizable case of the above theorem was treated in [31], but since residual finiteness passes to factors and quasidiagonality is a local property one can easily derive from [31] the equivalence of conditions (1), (2), and (3) for general Z-actions by passing through (4) as applied to metrizable factors. The most involved implication in the theorem is (1)⇒(4), while (4)⇒(2) follows from the fact that AF algebras are quasidiagonal and quasidiagonality passes to C ∗-subalgebras, and (2)⇒(3) is true for any unital C ∗-algebra (see Section V.4 of [2]). Note that, since crossed products by actions of Z are always nuclear, for metrizable X the crossed product C(X) ⋊λ Z is quasidiagonal if and only if it is an MF algebra [6, Thm. 5.2.2]. The property of quasidiagonality can be strengthened in certain natural ways, and in view of the above theorem we may ask whether residual finiteness can also be used to dynamically characterize the situation in which the crossed product satisfies such a stronger property. Recall for example that a C ∗-algebra is said to be strongly quasidiagonal if all of its representations are quasidiagonal. By Voiculescu's theorem [38], every simple separable quasidiagonal C ∗-algebra is strongly quasidiagonal. Hadwin showed in [22, Thm. 25] that, for a compact metric space (X, d) and a homeomorphism T : X → X, the crossed product C(X) ⋊λ Z is strongly quasidiagonal if and only if for every x ∈ X there are integers m, n ≥ N such that d(T nx, T −mx) < ε. It is easy to verify that the latter condition is equivalent to hereditary residual finiteness, by which we mean that every subsystem of (X, T ) is residually finite. We also remark that C(X) ⋊λ Z is residually finite dimensional (i.e., has a separating family of finite-dimensional representations) if and only if the periodic points are dense in X [37, Thm. 4.6]. The proof of Theorem 25 in [22] can also be used to characterize when the crossed product of a metrizable Z-system is strong NF in terms of residual finiteness. Following some preliminary observations we will give this characterization in Theorem 7.5. Recall that a separable C ∗- algebra A is said to be an NF algebra if it can be expressed as the inductive limit of a generalized inductive system with contractive completely positive connecting maps, and a strong NF algebra if the connecting maps can be chosen to be complete order embeddings [6]. Theorem 5.2.2 of [6] gives various characterizations of NF algebras; in particular, the following are equivalent: (i) A is an NF algebra, (2) A is a nuclear MF algebra, and (iii) A is nuclear and quasidiagonal. In the strong NF case we have by [6, 4, 5] the equivalence of the following conditions: (1) A is a strong NF algebra, (2) for every Ω ∈ F A and ε > 0 there are a finite-dimensional C ∗-algebra B and a complete order embedding ϕ : B → A such that for each x ∈ Ω there is a b ∈ B with kx−ϕ(b)k < ε, (3) A is nuclear and inner quasidiagonal, (4) A is nuclear and has a separating family of irreducible quasidiagonal representations. 24 DAVID KERR AND PIOTR W. NOWAK For a λ > 1, we say that a C ∗-algebra A is an OL∞,λ space if for every finite set Ω ⊆ A and ε > 0 there is a finite-dimensional C ∗-algebra B and an injective linear map ϕ : B → A with Ω ⊆ B such that kϕkcbkϕ−1 : ϕ(B) → Bkcb < λ. We write OL∞(A) for the infimum over all λ > 1 such that A is an OL∞,λ space. These notions were introduced in [23] as a means to analyze and quantify the relationships between properties like nuclearity, quasidiagonality, inner quasidiagonality, and stable finiteness using local operator space structure. If A is a strong NF algebra then OL∞(A) = 1 by a straightforward perturbation argument. Whether the converse is true is an open question, although in [16] it was shown to hold under the assumption that A has a finite separating family of primitive ideals. As part of Theorem 7.5 we will verify that the converse also holds in our situation, so that we obtain a dynamical characterization of both strong NFness and the OL∞ invariant equalling 1 for crossed products of metrizable Z-systems. Lemma 7.3. Let (X, T ) be a transitive metrizable Z-system and let d be a compatible metric on X. Let x be a point in X such that orb(x) = X. Then the following are equivalent: (1) (X, T ) is residually finite, (2) {T kx}k≥0 ∩ {T kx}k<0 6= ∅, (3) for every ε > 0 and N ∈ N there are integers m, n ≥ N such that d(T nx, T −mx) < ε, (4) for every ε > 0 there are m, n ∈ N such that d(T nx, T −mx) < ε. Proof. (1)⇒(2). If (2) does not hold then {T kx}k≥0 equals X \ {T kx}k<0 and hence is a clopen set which is sent to a proper clopen subset of itself under T , so that (X, T ) fails to be residually finite by Pimsner's characterization of chain recurrence in Lemma 2 of [31]. (2)⇒(3). Let ε > 0 and N ∈ N. Take a δ > 0 with δ ≤ ε such that if d(y, z) < δ then d(T ky, T kz) < ε for every k = −N, −N + 1, . . . , N . Whether x is periodic or nonperiodic, it is clear from (2) that there exist integers m, n ≥ 0 such that m + n ≥ 2N and d(T nx, T −mx) < δ. Set r equal to N − n, m − N , or 0 according to whether n < N , m < N , or m, n ≥ N . Then m − r, n + r ≥ N and d(T n+rx, T −m+rx) < ε, yielding (3). (3)⇒(4). Trivial. (4)⇒(1). Given m, n ∈ N, the sequence {x, T x, . . . , T n−1x, T −mx, T −m+1x, . . . , T −1x, x} forms an ε-chain precisely when d(T nx, T −mx) < ε, and so x is chain recurrent. Since the chain recurrent set is closed and Z-invariant, we obtain (1). (cid:3) The following lemma is a local version of Theorem 25 of [22] and essentially follows by the same argument using Berg's technique as in the proof of Lemma 3.6 in [32]. We reproduce this argument below. For background on induced representations see [41]. Lemma 7.4. Let (X, T ) be a transitive metrizable Z-system, and let x be a point in X such that orb(x) = X. Then the following are equivalent: (1) (X, T ) is residually finite, (2) every irreducible representation of C(X) ⋊λ Z induced from the isotropy group of x is quasidiagonal, (3) C(X) ⋊λ Z has a faithful irreducible quasidiagonal representation. Proof. If orb(x) has finite cardinality n then T is a cyclic permutation. As is well known, the crossed product C(X) ⋊λ Z in this case is ∗-isomorphic to Mn ⊗ C(T), whose irreducible representations are all evidently quasidiagonal and, up to unitary equivalence, induced from the isotropy group of x. We may thus assume that orb(x) is infinite. Then the isotropy group of x RESIDUALLY FINITE ACTIONS 25 is trivial and there is only one induced representation, namely the irreducible representation π : C(X) ⋊λ Z → B(ℓ2(Z)) defined by π(f )ξn = f (T nx)ξn and π(u)ξn = ξn+1 for all f ∈ C(X) and n ∈ Z, where u is the canonical unitary associated to T and {ξn}n∈Z is the canonical orthonormal basis of ℓ2(Z). Since Z acts freely on orb(x) and subrepresentations of πC(X) correspond to subsets of orb(x), it follows that πC(X) is G-almost free in the sense of Definition 1.12 of [42], so that π is faithful by Corollary 4.19 of [42]. We thus have (2)⇒(3). Note also that condition (3) implies that C(X) ⋊λ Z is a quasidiagonal C ∗-algebra and thus we get (3)⇒(1) in view of Theorem 7.2. Let us then assume (1) and show that π is quasidiagonal in order to obtain (2). Let Ω be a finite subset of C(X) and let n ∈ N. In view of the definition of a quasidiagonal representation, we need only produce an orthogonal projection p in B(ℓ2(Z)) which dominates the orthogonal projection onto the subspace spanned by {ξj : −n ≤ j ≤ n} and satisfies k[p, π(f )]k < 2/n for all f ∈ Ω and k[p, π(u)]k < 8/n. Fix a compatible metric d on X. By uniform continuity there exists a δ > 0 such that if x1, x2 ∈ X and d(x1, x2) < δ then f (T kx1) − f (T kx2) < 1/n for each f ∈ Ω and k = 0, . . . , n − 1. By Lemma 7.3 we can find r > n and s < −2n such that d(T rx, T sx) < δ, so that f (T r+kx) − f (T s+kx) < 1/n for each f ∈ Ω and k = 0, . . . , n − 1. For each f ∈ Ω we define a perturbation af ∈ B(ℓ2(W )) of π(f ) by af ξn =(cid:26) f (T s+kx)ξn f (T nx)ξn if n = r + k for some k = 0, . . . , n − 1, otherwise, in which case we have kaf − π(f )k < 1/n. Next we apply Berg's technique (see Section VI.4 of [15]) to produce orthogonal unit vectors ζk, ηk ∈ span(ξr+k, ξs+k) for each k = 0, . . . , n − 1 and a unitary v ∈ B(ℓ2(Z)) such that (1) vζk = ζk+1 and vηk = ηk+1 for k = 0, . . . , n − 2, (2) vζn−1 = ξs+n and vηn−1 = ξr+n, (3) v agrees with π(u) on the orthogonal complement of the span of the vectors ξr, ξr+1, . . . , ξr+n−1, ξs, ξs+1, . . . , ξs+n−1, and (4) kπ(u) − vk < 4/n. Let p be the orthogonal projection onto the span of the vectors ξr+n, ξr+n+1, . . . , ξs−1, η0, η1, . . . , ηn−1. Since the unitary v cyclically permutes these vectors we have [p, v] = 0 and hence k[p, π(u)]k ≤ kp(π(u) − v)k + k(v − π(u))pk ≤ 2kπ(u) − vk < 8/n. Also, if f ∈ Ω then span(ξr+k, ξs+k) is an eigenspace for af for every k = 0, . . . , n − 1, so that [p, af ] = 0 and hence k[p, π(f )]k ≤ kp(π(f ) − af )k + k(af − π(f ))pk ≤ 2kaf − f k < 2/n, completing the proof. (cid:3) Theorem 7.5. Let (X, T ) be a metrizable Z-system and let d be a compatible metric on X. Then the following are equivalent: (1) there is a collection {(Xi, T )}i∈I of transitive residually finite subsystems of (X, T ) such (2) there is a dense set D ⊆ X such that for every x ∈ D, ε > 0, and N ∈ N there are that Si∈I Xi is dense in X, integers m, n ≥ N for which d(T nx, T −mx) < ε, (3) C(X) ⋊λ Z is strong NF, (4) OL∞(C(X) ⋊λ Z) = 1. 26 DAVID KERR AND PIOTR W. NOWAK Proof. (1)⇔(2). This is a simple consequence of Lemma 7.3. (2)⇒(3). Since the set W = Si∈I Xi is dense in X the canonical quotient maps C(X) ⋊λ Z → C(Xi) ⋊λ Z for i ∈ I form a separating family. This can be seen by taking the faithful representation π : C(X) ⋊λ Z → B(ℓ2(W ) ⊗ ℓ2(Z)) canonically induced from the multiplication representation of C(X) on ℓ2(W ) and observing that for each i ∈ I the cut-down of π(C(X)⋊λ Z) by the orthogonal projection onto ℓ2(Xi) ⊗ ℓ2(Z) is ∗-isomorphic to C(Xi) ⋊λ Z. Since for each i ∈ I the crossed product C(Xi) ⋊λ Z has a faithful irreducible quasidiagonal representation by Lemma 7.4, we thus conclude using Theorem 4.5 of [4] that C(X) ⋊λ Z is strong NF. (3)⇒(4). This implication holds for general C ∗-algebras, as mentioned prior to the theorem statement. (4)⇒(1). By Theorem 5.4 of [16] C(X) ⋊λ Z has a separating family Π of irreducible represen- tations whose images are stably finite C ∗-algebras. By [19] every primitive ideal of C(X) ⋊λ Z is the kernel of an irreducible representation induced from the isotropy group of some point x ∈ X. Now any two faithful irreducible representations of a separable prime (equivalently, separable primitive) C ∗-algebra are approximately unitarily equivalent. In the antiliminal case this follows from Voiculescu's theorem [38] as every faithful irreducible representation will be essential, while in the non-antiliminal case the C ∗-algebra has an essential ideal ∗-isomorphic to the compact operators and hence has only one faithful irreducible representation up to unitary equivalence. Consequently we may assume that each π ∈ Π is induced from the isotropy group of some xπ ∈ X. If this point xπ has nontrivial isotropy group then it is periodic and the system (orb(xπ), T ) is trivially residually finite. If on the other hand xπ has trivial isotropy group then, as indicated in the proof of Lemma 7.4, there is a unique induced representation of C(orb(xπ)) ⋊λ Z and it is faithful, implying that C(orb(xπ)) ⋊λ Z is stably finite and hence is dense in X, which results from the fact that π(f ) = 0 for all π ∈ Π and f ∈ C(X) whose (cid:3) by Theorem 7.2 that (orb(xπ), T ) is residually finite. It remains to observe that Sπ∈Π orb(xπ) support is contained in the complement of Sπ∈Π orb(xπ). Example 7.6. Using Theorem 7.5 we can produce by dynamical means many examples of NF algebras which are not strong NF (cf. Examples 5.6 and 5.19 of [4]). For instance, take two copies of translation on Z each compactified with two fixed points ±∞ and identify +∞ from each copy with −∞ of the other copy. This system is residually finite, but the transitive subsystems generated by each copy of Z fail to be residually finite, so that C(X) ⋊λ Z is NF but not strong NF. This example is a dynamical analogue of Example 3.2 in [16]. Example 7.7. We can also use Theorem 7.5 to exhibit strong NF crossed products which are not strongly quasidiagonal. For instance, take translation on Z and compactify it so that it spirals around to the example from the previous paragraph in both the forward and backward directions, spending longer and longer intervals near each of the two fixed points in each successive approach. The crossed product is strong NF since the system is transitive and residually finite, but it is not strongly quasidiagonal by the discussion in the second paragraph following Theorem 7.2. In the backward direction we could instead have convergence to one of the fixed points, in which case we have the additional feature that the backward and forward limit sets of the unique dense orbit do not coincide. RESIDUALLY FINITE ACTIONS 27 References [1] C. Anantharaman-Delaroche and J. Renault. Amenable Groupoids. L'Enseignement Math´ematique, Geneva, 2000. [2] B. Blackadar. Operator Algebras. Theory of C∗-Algebras and von Neumann Algebras. Encyclopaedia of Math- ematical Sciences, 122. Operator Algebras and Non-commutative Geometry, III. Springer-Verlag, Berlin, 2006. [3] B. Blackadar and E. Kirchberg. Irreducible representations of inner quasidiagonal C ∗-algebras. arXiv:0711.4949. [4] B. Blackadar and E. Kirchberg. Inner quasidiagonality and strong NF algebras. Pacific J. Math. 198 (2001), 307 -- 329. [5] B. Blackadar and E. Kirchberg. Generalized inductive limits and quasidiagonality. In: C ∗-algebras (Munster, 1999), 23 -- 41, Springer, Berlin, 2000. [6] B. Blackadar and E. Kirchberg. Generalized inductive limits of finite-dimensional C ∗-algebras. Math. Ann. 307 (1997), 343 -- 380. [7] L. Bowen. The ergodic theory of free group actions: entropy and the f -invariant. Groups Geom. Dyn. 4 (2010), 419 -- 432. [8] L. Bowen. Measure conjugacy invariants for actions of countable sofic groups. J. Amer. Math. Soc. 23 (2010), 217 -- 245. [9] N. P. Brown. On quasidiagonal C ∗-algebras. In: Operator algebras and applications, 19 -- 64, Adv. Stud. Pure Math., 38, Math. Soc. Japan, Tokyo, 2004. [10] N. P. Brown, K. J. Dykema, and K. Jung. Free entropy dimension in amalgamated free products. Proc. London Math. Soc. 97 (2008), 339 -- 367. [11] N. P. Brown and N. Ozawa. C∗-Algebras and Finite-Dimensional Approximations. Graduate Studies in Mathematics, 88. Amer. Math. Soc., Providence, RI, 2008. [12] C. Conley. Isolated invariant sets and the Morse index. CBMS Regional Conference Series in Mathematics, 38. Amer. Math. Soc., Providence, RI, 1978. [13] J. Cuntz and G. K. Pedersen. Equivalence and traces on C ∗-algebras. J. Funct. Anal. 33 (1979), 135 -- 164. [14] M. Dadarlat. On the approximation of quasidiagonal C ∗-algebras. J. Funct. Anal. 167 (1999), 69 -- 78. [15] K. R. Davidson. C ∗-algebras by Example. Fields Institute Monographs, 6. Amer. Math. Soc., Providence, RI, 1996. [16] C. Eckhardt. On OL∞ structure of nuclear, quasidiagonal C ∗-algebras. J. Funct. Anal. 258 (2010), 1 -- 19. [17] G. Gong and H. Lin. Almost multiplicative morphisms and almost commuting matrices. J. Operator Theory 40 (1998), 217 -- 275. [18] K. R. Goodearl and D. Handelman. Rank functions and K0 of regular rings. J. Pure Appl. Algebra bf 7 (1976), 195 -- 216. [19] E. C. Gootman. Primitive ideals of C ∗-algebras associated with transformation groups. Trans. Amer. Math. Soc. 170 (1972), 97 -- 108. [20] E. C. Gootman and J. Rosenberg. The structure of crossed product C ∗-algebras: a proof of the generalized Effros-Hahn conjecture. Invent. Math. 52 (1979), 283 -- 298. [21] U. Haagerup and S. Thorbjørnsen. A new application of random matrices: Ext(C ∗ red(F2)) is not a group. Ann. of Math. (2) 162 (2005), 711 -- 775. [22] D. Hadwin. Strongly quasidiagonal C ∗-algebras. With an appendix by J. Rosenberg. J. Operator Theory 18 (1987), 3 -- 18. [23] M. Junge, N. Ozawa, and Z.-J. Ruan. On OL∞ structures of nuclear C ∗-algebras. Math. Ann. 325 (2003), 449 -- 483. [24] D. Kerr and H. Li. Soficity, amenability, and dynamical entropy. To appear in Amer. J. Math. [25] D. Kerr and H. Li. Entropy and the variational principle for actions of sofic groups. To appear Invent. Math. [26] H. Lin. AF-embeddings of the crossed products of AH-algebras by finitely generated abelian groups. Int. Math. Res. Pap. IMRP 2008, Art. ID rpn007, 67 pages. [27] A. Lubotzky and Y. Shalom. Finite representations in the unitary dual and Ramanujan groups. In: Discrete geometric analysis, 173 -- 189, Contemp. Math., 347, Amer. Math. Soc., Providence, RI, 2004. 28 DAVID KERR AND PIOTR W. NOWAK [28] G. Margulis. Free subgroups of the homeomorphism group of the circle. C. R. Acad. Sci. Paris S´er. I Math. 331 (2000), 669 -- 674. [29] G. Margulis and E. B. Vinberg. Some linear groups virtually having a free quotient. J. Lie Theory 10 (2000), 171 -- 180. [30] S. Orfanos. Quasidiagonality of crossed products. To appear in J. Operator Theory. [31] M. V. Pimsner. Embedding some transformation group C ∗-algebras into AF-algebras. Ergod. Th. Dynam. Sys. 3 (1983), 613 -- 626. [32] M. Pimsner and D. Voiculescu. Exact sequences for K-groups and Ext-groups of certain cross-product C ∗-algebras. J. Operator Theory 4 (1980), 93 -- 118. [33] J. M. Rosenblatt. Invariant measures and growth conditions. Trans. Amer. Math. Soc. 193 (1974), 33 -- 53. [34] M. Rørdam and A. Sierakowski. Purely infinite C ∗-algebras arising from crossed products. To appear in Ergod. Th. Dynam. Sys. [35] N. Salinas. Homotopy invariance of Ext(A). Duke Math. J. 44 (1977), 777 -- 794. [36] J.-L. Sauvageot. Id´eaux primitifs de certains produits crois´es. Math. Ann. 231 (1977), 61 -- 76. [37] J. Tomiyama. The Interplay between Topological Dynamics and Theory of C ∗-Algebras. Lecture Notes Series, vol. 2, Res. Inst. Math., Seoul, 1992. [38] D. Voiculescu. A noncommutative Weyl-von Neumann theorem. Rev. Roum. Math. Pures Appl. 21 (1976), 97 -- 113. [39] D. Voiculescu. A note on quasidiagonal C ∗-algebras and homotopy. Duke Math. J. 62 (1991), 267 -- 271. [40] S. Wagon. The Banach-Tarski Paradox. Cambridge University Press, Cambridge, 1993. [41] D. P. Williams. Crossed Products of C ∗-Algebras. Mathematical Surveys and Monographs, 134. Amer. Math. Soc., Providence, RI, 2007. [42] G. Zeller-Meier. Produits crois´es d'une C ∗-alg`ebre par un groupe d'automorphismes. J. Math. Pures Appl. (9) 49 (1968), 101 -- 239. David Kerr, Department of Mathematics, Texas A&M University, College Station TX 77843-3368, U.S.A. E-mail address: [email protected] Piotr W. Nowak, Department of Mathematics, Texas A&M University, College Station TX 77843- 3368, U.S.A. E-mail address: [email protected]
1001.0424	1	1001	2010-01-04T01:42:44	Families of Type {\rm III KMS} States on a Class of $C^*$-Algebras containing $O_n$ and $\mathcal{Q}_\N$	[ "math.OA", "math.FA" ]	We construct a family of purely infinite $C^*$-algebras, $\mathcal{Q}^\lambda$ for $\lambda\in (0,1)$ that are classified by their $K$-groups. There is an action of the circle $\T$ with a unique ${\rm KMS}$ state $\psi$ on each $\mathcal{Q}^\lambda.$ For $\lambda=1/n,$ $\mathcal{Q}^{1/n}\cong O_n$, with its usual $\T$ action and ${\rm KMS}$ state. For $\lambda=p/q,$ rational in lowest terms, $\mathcal{Q}^\lambda\cong O_n$ ($n=q-p+1$) with UHF fixed point algebra of type $(pq)^\infty.$ For any $n>0,$ $\mathcal{Q}^\lambda\cong O_n$ for infinitely many $\lambda$ with distinct KMS states and UHF fixed-point algebras. For any $\lambda\in (0,1),$ $\mathcal{Q}^\lambda\neq O_\infty.$ For $\lambda$ irrational the fixed point algebras, are NOT AF and the $\mathcal{Q}^\lambda$ are usually NOT Cuntz algebras. For $\lambda$ transcendental, $K_1\cong K_0\cong\Z^\infty$, so that $\mathcal{Q}^\lambda$ is Cuntz' $\mathcal Q_{\N}$, \cite{Cu1}. If $\lambda^{\pm 1}$ are both algebraic integers, the {\bf only} $O_n$ which appear satisfy $n\equiv 3(mod 4).$ For each $\lambda$, the representation of $\mathcal{Q}^\lambda$ defined by the KMS state $\psi$ generates a type ${\rm III}_\lambda$ factor. These algebras fit into the framework of modular index (twisted cyclic) theory of \cite{CPR2,CRT} and \cite{CNNR}.	math.OA	math	Families of Type III KMS States on a Class of C∗-Algebras containing On and QN A.L. Careya, J. Phillipsb,c, I.F. Putnamb, A. Renniea a Mathematical Sciences Institute, Australian National University, Canberra, ACT, AUSTRALIA b Department of Mathematics and Statistics, University of Victoria, Victoria, BC, CANADA c Corresponding Author. email: [email protected]; telephone: 250-721-7450; fax: 250-721-8962 Abstract We construct a family of purely infinite C∗-algebras, Qλ for λ ∈ (0, 1) that are classified by their K-groups. There is an action of the circle T with a unique KMS state ψ on each Qλ. For λ = 1/n, Q1/n ∼= On, with its usual T action and KMS state. For λ = p/q, rational in lowest terms, Qλ ∼= On (n = q − p + 1) with UHF fixed point algebra of type (pq)∞. For any n > 0, Qλ ∼= On for infinitely many λ with distinct KMS states and UHF fixed-point algebras. For any λ ∈ (0, 1), Qλ 6= O∞. For λ irrational the fixed point algebras, are NOT AF and the Qλ are usually NOT Cuntz algebras. For λ transcendental, K1 ∼= K0 ∼= Z∞, so that Qλ is Cuntz' QN, [Cu1]. If λ±1 are both algebraic integers, the only On which appear satisfy n ≡ 3(mod 4). For each λ, the representation of Qλ defined by the KMS state ψ generates a type IIIλ factor. These algebras fit into the framework of modular index (twisted cyclic) theory of [CPR2, CRT] and [CNNR]. Keywords: KMS state, IIIλ factor, modular index, twisted cyclic theory, K-Theory. AMS Classification codes: 46L80, 58J22, 58J30. 1. Introduction In this paper we introduce some new examples of KMS states on a large class of purely infinite C∗- algebras that were motivated by the 'modular index theory' of [CPR2, CNNR]. We were aiming to find examples of algebras that were not Cuntz-Krieger algebras (or the CAR algebra) and were not previously known in order to explore the possibilities opened by [CNNR]. These algebras, denoted by Qλ for 0 < λ < 1, are not constructed as graph algebras, but as "corner algebras" of certain crossed product C∗-algebras. The Qλ have similar structural properties to the Cuntz algebras, however there are important new features, such as 1) when λ = p/q is rational in lowest terms, then Qλ ∼= Oq−p+1 as mentioned in the Abstract, 2) when λ is algebraic, the K-groups depend on the minimal polynomial (and its coefficients) of λ, 3) when λ is transcendental, Qλ ∼= QN, Cuntz' algebra, [Cu1]. We prove in Section 3 that the Qλ are purely infinite, simple, separable, nuclear C∗-algebras, so there is no nontrivial trace on them. Also in Section 3 we determine in many cases the K-groups of these algebras and use classification theory to identify them when these algebras have the same K-groups as others found previously (these facts are summarised in the Abstract). As each Qλ comes equipped with a gauge action of the circle, our results thus give an uncountable family of distinct circle actions on QN, each with its own unique KMS state. Indeed, for all 0 < λ < 1, we find a unique KMS state, [BR2], for this gauge action, and we prove in Section 4 that the GNS representation of Qλ associated to our KMS state generates a type IIIλ von Neumann algebra. The result of [CPR2] that motivated this paper was the construction of a 'modular spectral triple' with which one may compute an index pairing using the KMS state. In [CNNR] it was shown how modular spectral triples arise naturally for 1 2 KMS states of circle actions and lead to 'twisted residue cocycles' using a variation on the semifinite residue cocycle of [CPRS2]. It is well known that such twisted cocycles can not pair with ordinary K1. In [CPR2, CRT] a substitute was introduced which is called 'modular K1'. The correct definition of modular K1 was found in [CNNR], and there is a general spectral flow formula which defines the pairing of modular K1 with our 'twisted residue cocycle'. There is a strong analogy with the local index formula of noncommutative geometry in the L1,∞- summable case, however, there are important differences: the usual residue cocycle is replaced by a twisted residue cocycle and the Dixmier trace arising in the standard situation is replaced by a KMS-Dixmier functional. The common ground with [CPRS2] stems from the use of the spectral flow formula of [CP2] to derive the twisted residue cocycle and this has the corollary that we have a homotopy invariant. To illustrate the theory for these examples we compute, for particular modular unitaries in matrix algebras over the algebras Qλ, the precise numerical values arising from the general formalism. 2. The algebras Qλ for 0 < λ < 1. 2.1. The C∗-algebras C∗(Γλ) = C(Γλ) and their K-theory. We will construct our algebras Qλ as "corner" algebras in certain crossed product C∗-algebras but first we need some preliminaries. For 0 < λ < 1, let Γλ be the countable additive abelian subgroup of R defined by: Γλ =( k=NXk=−N nkλk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N ≥ 0 and nk ∈ Z) . Loosely speaking, Γλ consists of Laurent polynomials in λ and λ−1 with integer coefficients. It is not only a dense subgroup of R, but is clearly a unital subring of R. Proposition 2.1. Let 0 < λ < 1. (1) (2) sum where d ≥ 2 is the degree of the minimal (monic) polynomial in Z[x] satisfied by λ. (3) (4) internal direct sum. (5) If λ = p/q where 0 < p < q are integers in lowest terms, then Γλ = Z[1/n], where n = pq. If λ and λ−1 are both algebraic integers, then Γλ = Z + Zλ + ··· + Zλd−1 is an internal direct If λ is transcendental then, Γλ =Lk∈Z Zλk is an internal direct sum. If λ = 1/√n with n ≥ 2 a square-free positive integer, then Γλ = Z[1/n] + Z[1/n] · √n is an In general, if λ is algebraic with minimal polynomial, nλd + ··· + m = 0 over Z, then Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1 ⊆ Γλ ⊆ Z[ 1 mn ] ⊕ Z[ 1 mn ]λ ⊕ ··· ⊕ Z[ 1 mn ]λd−1. Hence, rank(Γλ) := dimQ(Γλ ⊗Z Q) = d. Proof. In case (1), since gcd(p, q) = 1, there exist a, b ∈ Z so that 1 = ap + bq. Therefore, 1 aλ + b ∈ Γλ; and similarly, 1 have: m q = p ∈ Γλ. Since, Γλ is a commutative ring, for any k, m ∈ Z with k ≥ 1 we q = ap+bq nk = m (pq)k is in Γλ. That is, Z[1/n] ⊆ Γλ. On the other hand, for k ≥ 1 we have pk qk = p2k That is, Z[1/n] = Γλ. nk ∈ Z[1/n] and λ−k = (pq)k = p2k 1 nk ∈ Z[1/n]. qk pk = q2k 1 (pq)k = q2k 1 λk = 1 In case (2), it is not hard to see the minimal polynomial of λ in Z[x] is not only monic, but also has constant term = ±1; say, p(λ) = λd + aλd−1 + ··· ± 1 = 0. Clearly, λ ∈ Z + Zλ + ··· + Zλd−1. Since λ−1p(λ) = 0, we also have λ−1 ∈ Z + Zλ + ··· + Zλd−1. By an easy induction, we have λk ∈ Z + Zλ + ··· + Zλd−1, for all k ∈ Z. Hence, Γλ = Z + Zλ + ··· + Zλd−1. The sum is direct by the minimality of the degree of the minimal polynomial. In case (3) the sum is direct because if λ satisfied a Laurent polynomial over Z, then by multipling by a high power of λ it would also satisfy a genuine polynomial over Z. 3 Case (4) is an easy calculation which we leave to the reader. Case (5) is proved by similar methods used in case (2). Again, the sum Z[ 1 mn ]λd−1 is direct by the minimality of the degree of the minimal polynomial. mn ]λ + ··· + Z[ 1 mn ] + Z[ 1 (cid:3) Proposition 2.2. Let 0 < λ < 1. (1) If λ = p/q is rational in lowest terms so that Γλ = Z[1/n], where n = pq, then K0(C(Γλ)) = Z[1Γλ)] and K1(C(Γλ)) = Z[1/n]. (2) sum as above, then If λ and λ−1 are both algebraic integers, so that Γλ = Z + Zλ + ··· + Zλd−1 is an internal direct K0(C(Γλ)) = even^ (Γλ) = dMk=0,k even k^(Γλ) and K1(C(Γλ)) = odd^(Γλ) = dMk=1,k odd k^(Γλ). (3) If λ is transcendental then, k^(Γλ) and K1(C(Γλ)) = K0(C(Γλ)) = If λ = 1/√n with n ≥ 2 a square-free positive integer, then even^ (Γλ) = ∞Mk=0,k even odd^(Γλ) = ∞Mk=1,k odd k^(Γλ). K0(C(Γλ)) ∼= Z ⊕ Z[1/n] and K1(C(Γλ)) ∼= Z[1/n] ⊕ Z[1/n] In general, if λ is algebraic with nλd + ··· + m = 0 over Z then the composition of the inclusions Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1 ⊆ Γλ ⊆ Z[ 1 mn ] ⊕ Z[ 1 mn ]λ ⊕ ··· ⊕ Z[ 1 mn ]λd−1 (4) (5) induces an inclusion on K-Theory, so that both of the following maps are one-to-one odd^(Zd) ∼= K1(C∗(Z⊕··· Zλd−1)) ֒→ K1(C(Γλ)). even^ (Zd) ∼= K0(C∗(Z⊕··· Zλd−1)) ֒→ K0(C(Γλ)) and T. Since Proof. In case (1), Γλ = lim−→ K0(C(T)) = Z[1] is generated by multiples of the trivial rank one bundle, the maps in the direct limit K0(C(Γλ) = lim−→ K0(C(T)) are the identity map in each case, so that K0(C(Γλ)) = Z[1]. On the other hand, K1(C(T)) is generated by the maps on C(T) z 7→ zk, and each map in the direct limit is the same map induced by z 7→ zn. Thus, K1(C(Γλ)) = Z[1/n]. Cases (2) and (3) are well-known facts about the K-Theory of tori. Z where each map is multiplication by n, so that Γλ = lim←− Case (4): first one uses item (4) of the previous Proposition, then the proof of case (1) above in order to apply Proposition 2.11 of [Sc]. The proof is finished off with the easily proved observation that Z[1/n] ⊗ Z[1/n] = Z[1/n]. Case (5) the composed embedding is just containment: mn ]λ ⊕ ··· ⊕ Z[ 1 Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1 ⊆ Z[ 1 mn ]λd−1. Since we know that K∗(C∗(Z)) → mn ] ⊕ Z[ 1 4 K∗(C∗(Z[1/mn])) is one-to-one (even an isomorphism after tensoring with Q), an application of C. Schochet's Kunneth Theorem, [Sc] shows that the induced map on K-Theory: K∗(C∗(Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1)) −→ K∗(C∗(Z[ 1 mn ] ⊕ Z[ 1 mn ]λ ⊕ ··· ⊕ Z[ 1 mn ]λd−1)) is one-to-one (even an isomorphism after tensoring with Q). (cid:3) Corollary 2.3. If λ is algebraic with minimal polynomial of degree d so that rank(Γλ) = d then rank(K0(C(Γλ))) = rank( even^ (Zd)) = 2d−1 = rank( odd^(Zd)) = rank(K1(C(Γλ))). Proof. For each N ≥ d − 1, let ΓN = Zλ−N + ··· + ZλN ⊆ Γλ. Then each ΓN is a finitely generated torsion free (and hence free abelian) subgroup of Γλ. Moreover, 1 mn Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1 ⊆ ΓN ⊆ Γλ ⊆ Z[ ]λ ⊕ ··· ⊕ Z[ ] ⊕ Z[ 1 mn 1 mn ]λd−1, so that by tensoring with Q the induced inclusions are all equalities, and hence all are Q-vector spaces of dimension d. Since ΓN is free abelian, ΓN ∼= Zd. Now, K0(C∗(ΓN )) ∼= K0(C(Td)) ∼= So, each Ki(C∗(ΓN )) ⊗Z Q is a Q-vector space of dimension 2d−1 and the map: K∗(C∗(Z ⊕ Zλ ⊕ ··· ⊕ Zλd−1)) ⊗Z Q −→ K∗(C∗(ΓN )) ⊗Z Q and K1(C∗(ΓN )) ∼= K1(C(Td)) ∼= even^ (Zd) ∼= Z2d−1 odd^(Zd) ∼= Z2d−1 . is one-to-one and hence an isomorphism of Q-vector spaces. Since the corresponding isomorphism onto K∗(C∗(ΓN +1)) ⊗Z Q factors through K∗(C∗(ΓN )) ⊗Z Q the maps K∗(C∗(ΓN )) ⊗Z Q → K∗(C∗(ΓN +1)) ⊗Z Q are all isomorphisms. Now, C∗(Γλ) = limN C∗(ΓN ) and so Ki(C∗(Γλ)) = limN Ki(C∗(ΓN )), and therefore, for each i = 1, 2. (cid:3) Ki(C∗(Γλ)) ⊗Z Q = lim N Ki(C∗(ΓN )) ⊗Z Q ∼= Q2d−1 Now, let Gλ ⊃ G0 λ be the following countable discrete groups of matrices: Of course, G0 λ is isomorphic to the additive group Γλ, and Gλ is semidirect product of Z acting on λ ∼= Γλ. We let Gλ act on R as an "ax+b" group, noting that the action leaves Γλ invariant. That G0 is, Gλ =(cid:26)(cid:18) λn a 0 1 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) a ∈ Γλ, n ∈ Z(cid:27) ⊃ G0 λ =(cid:26)(cid:18) 1 a 0 1 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) a ∈ Γλ(cid:27) . 1 (cid:19) ∈ Gλ define g · t := λnt + a. for t ∈ R and g =(cid:18) λn a 0 Notation. For such an element g ∈ Gλ we will use the notation g := [λn : a] in place of the matrix for g and g := det(g) = λn for the determinant of g. Note: G0 We use this action on R to define the transpose action α of Gλ on L∞(R) : αg(f )(t) = f (g−1t) for f ∈ L∞(R) and t ∈ R. λ = {g ∈ Gλ g = 1} ⊳ Gλ. Now let C λ jections X[a,b) where a, b ∈ Γλ. That is, 0 (R) be the separable C∗-subalgebra of L∞(R) generated by the countable family of pro- C λ 0 (R) = closure ( nXk=1 ckX[ak,bk)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 5 ck ∈ C; ak, bk ∈ Γλ)! . We observe that C λ 0 (R) and since αg(X[a,b)) = X[g(a),g(b)) both are invariant under the action α of Gλ. We define the separable C∗-algebras Aλ ⊃ Aλ as the crossed products: 0 (R) is a commutative AF-algebra. Clearly, C0(R) ⊂ C λ 0 Aλ = Gλ ⋊α C λ 0 (R) = Z ⋊ (G0 λ ⋊α C λ 0 (R)) ⊃ Aλ 0 = G0 λ ⋊α C λ 0 (R). c ⊂ l1 00(R) denote the dense ∗-subalgebra of C λ λ are amenable these equal the reduced crossed products by [Ped, Theorem 7.7.7 ]. 0 (R) consisting of finite linear combinations of the 0 (R)) ⊂ Aλ denote the dense ∗-subalgebra of Aλ 00(R). Similarly we define Aλ Since Gλ and G0 Let C λ generating projections, X[a,b), and let Aλ α(Gλ, C λ consisting of finitely supported functions x : Gλ → C λ Proposition 2.4. For any λ ∈ (0, 1) Aλ Proof. Since Aλ = Z ⋊ Aλ 0 is in Nnuc. By the proof of the previous Corollary, we can write Γλ as an increasing union of finitely generated torsion-free abelian groups ΓN which are free abelian group of finite rank so that Aλ 0 is the direct limit of crossed products of the separable commutative C∗-algebra C λ (cid:3) 0 and Aλ are in the bootstrap class Nnuc. 0, it suffices to see that Aλ 0 (R) by Zmi and hence is in Nnuc. 0,c ⊂ Aλ 0 . Notation: We remind the reader of the crossed product operations in our setting (Definition 7.6.1 of [Ped]) together with some particular notations we use. To this end, let x, y ∈ l1 0 (R)) then we have the product and adjoint formulas: α(Gλ, C λ (x · y)(g) = Xh∈Gλ x∗(g) = αg((x(g−1))∗) for g ∈ Gλ. x(h)αh(y(h−1g)) for g ∈ Gλ; c , then: α(Gλ, C λ c (respectively, Aλ 0,c) is dense in Aλ (respectively, Aλ 0 (R), then we write 0 ) we often do our calculations 0 (R)) is supported on the single element g ∈ Gλ and x(g) = f ∈ C λ If x ∈ l1 x = f · δg. Since Aλ with these elements and we have the following easily verified calculus for them. Lemma 2.5. Let f1 · δg1, f2 · δg2, f · δg ∈ Aλ (1) (2) (3) (4) (5) (6) The product of partial isometries of the form X[a,b) · δg is a partial isometry of the same form. (7) Consider the partial isometry, v = X[a,b) · δg. Any two of the following: vv∗, v∗v, g completely determine the interval [a, b) and the element g. Definition 2.6. Let e ∈ Aλ (f1 · δg1) · (f2 · δg2) = f1αg1(f2) · δg1g2 (f · δg)∗ = αg−1( ¯f ) · δg−1. f · δg is self-adjoint if and only if f is self-adjoint and g = 1. f · δg is a projection if and only if f is a projection and g = 1. f · δg is a partial isometry if and only if f is a projection. 0,c be the projection e = X[0,1)·δ1. We define the separable unital C∗-algebras: Qλ := eAλe ⊃ eAλ 0 e =: F λ. c e, and F λ We will also have occasion to use the dense subalgebras Qλ Proposition 2.7. The orthogonal family of projections en = X[n,n+1) · δ1 ∈ Aλ equivalent by partial isometries in Aλ c := eAλ 0 for n ∈ Z are mutually 0 of the form Vn,k := X[n,n+1) · δgn−k where gn−k = [1 : (n − k)]. c := eAλ 0,ce. 6 Moreover, the finite sums EN :=PN−1 that n=−N en = X[−N,N ) · δ1 form an approximate identity for Aλ so Aλ ∼= Qλ ⊗ K(l2(Z)) and Aλ 0 ∼= F λ ⊗ K(l2(Z)). Proof. By Lemma 2.5, one easily calculates that: for each pair n, k ∈ Z, Vn,kV ∗n,k = en and V ∗n,kVn,k = ek. Now for each positive integer N if we have y ∈ Aλ c that satisfies supp(yh) ⊆ [−N, N ) for all h, then using Lemma 2.5 again we see that EN · y = y. Since the collection of all such elements y ∈ Aλ c is dense in Aλ, we see that the increasing sequence of projections {EN} form an approximate identity for Aλ. Corollary 2.8. It follows from Proposition 2.4.7 of [RS] and Proposition 2.4 that for any λ ∈ (0, 1), Qλ and F λ are both in Nnuc. Lemma 2.9. (cf. [PhR, Proposition 3.1, Lemma 3.6]) The algebra C λ 0 (R) is a commutative separable AF algebra consisting of all functions f : R → C which vanish at ∞ and: are right continuous at each x ∈ Γλ; have a finite left-hand limit at each x ∈ Γλ; and are continuous at each x ∈ (R \ Γλ). Moreover, if φ ∈ \C λ 0 (R) → C) then there exists a unique x0 ∈ R such that: (1) if x0 ∈ (R \ Γλ) then φ(f ) = f (x0) for all f ∈ C λ (2) if x0 ∈ Γλ then either(cid:26) φ(f ) = f (x0) for all f ∈ C λ φ(f ) = f−(x0) = limx→x− 0 (R), (the space of all nonzero ∗-homomorphisms: C λ 0 (R), or f (x) for all f ∈ C λ 0 (R). 0 (R), (cid:3) 0 Proof. Since generating functions for C λ 0 (R) satisfy each of the properties above which are clearly preserved by passing to uniform limits, we see that any function in C λ 0 (R) satisfies these properties. Conversely, it is easy to show that any function satisfying these properties can be uniformly approxi- mated by a finite linear combination of the generators. The remainder of the proof is given in [PhR, Lemma 3.6]. (cid:3) Notation. We denote the dual space, \C λ that is the topology of pointwise convergence on C λ space, and C λ 0 (R) ∼= C0(Rλ). 0 (R) by Rλ and endow it with the relative weak-∗ topology, 0 (R). Of course, Rλ is a locally compact Hausdorff Proposition 2.10. The algebras Aλ and Aλ over, Aλ is purely infinite and hence so is Qλ. Proof. Now, both Gλ and G0 0 (R) as countable discrete groups of outer automorphisms. Thus, we can apply Theorem 3.2 of [E] once we check that neither action has any nontrivial invariant ideals in C λ 0 (and hence Qλ and F λ) are simple C∗-algebras. More- 0 (R) and that the actions are properly outer in the sense of Definition 2.1 of [E]. λ act on C λ g To do this we look at the induced action of Gλ and G0 λ on Rλ. So, for g ∈ Gλ we have g acting on Rλ via g(φ) = φ◦ α−1 so that for φ = φx given by evaluation at x ∈ R, we have as expected g(φx) = φg(x). Now, if x ∈ Γλ we use the notation φx− to denote the ∗-homomorphism φx−(f ) = f−(x) = f (x−) = limy→x− f (y). One easily checks that since g(x) ∈ Γλ, we have g(φx−) = φg(x)−. Next we claim that each of the sets {φm m ∈ Γλ} and {φm− m ∈ Γλ} is dense in Rλ in the relative weak-∗ topology. For example, we show that the second set is dense. To approximate φx for some x ∈ R we let {mn} be a sequence in Γλ converging to x from the right in R. Let f ∈ C λ 0 (R) so that f is right continuous at x. One easily shows that φmn− (f ) − φx(f ) → 0; that is, the sequence {φmn−} converges to φx in the relative weak-∗ topology. 7 λ and Gλ on C λ It is easy to see that the action of G0 λ on Rλ has dense orbits, and so, of course, the action of Gλ has dense orbits also. This implies that the actions of G0 0 (R) have no nontrivial invariant ideals since the induced action on Rλ has no nontrivial invariant closed sets. We complete the proof by showing that the action is properly outer in the sense of Definition 2.1 of [E]. Since there are no nontrivial α-invariant ideals and C λ 0 (R) is commutative this is the condition that for each g 6= 1 and each nonzero closed two sided ideal I invariant under αg we have k(αg − Id)Ik = 2. Since I is nonzero there is a nonempty open subset, O of Rλ so that I = O. But since g 6= 1 and O is not finite there exists y ∈ O such that g(y) 6= y and g(y) ∈ O. Let x = g(y) ∈ O so that g−1(x) = y ∈ O and x 6= g−1(x). So we can choose a continuous compactly supported real-valued function f on O with f (x) = 1, f (g−1(x)) = −1 and kfk = 1. But then f ∈ I and 2 ≥ k(αg − Id)Ik ≥ k(αg − Id)(f )k = kαg(f ) − fk ≥ f (g−1(x)) − f (x) = 2. Now that we know Aλ is simple, we can easily apply Theorem 9 of [LS] to conclude that Aλ satisfies hypothesis (v) of Proposition 4.1.1 (page 66) of [RS]. For simple C∗-algebras, this is equivalent to being purely infinite by Definition 4.1.2 of [RS]: the authors of [LS] had used one of the earlier definitions of purely infinite in their paper (namely, hypothesis (v)). By Proposition 4.1.8 of [RS] Qλ is also purely infinite. (cid:3) Corollary 2.11. It follows from Corollaries 8.2.2 and 8.4.1 (Kirchberg-Phillips) of [RS] and the fact that Aλ is stable that for any λ ∈ (0, 1), Aλ is classified up to isomorphism (among Kirchberg algebras in Nnuc) by its K-theory. Since we need to calculate with elements of Qλ and F λ, we make the following observations. Lemma 2.12. Now, Qλ (respectively, F λ) is the norm closure of finite linear combinations of the ele- ments of the form e(X[a,b) · δg)e, where g ∈ Gλ (respectively, g ∈ G0 λ), henceforth called the generators. Thus, we calculate (1) If f · δg ∈ Aλ, (respectively, f · δg ∈ Aλ 0 (R), then 0 ) where f ∈ C λ e(f · δg)e = X[a,b)f · δg where [a, b) = [0, 1) ∩ [g(0), g(1)). (2) Thus, for g ∈ Gλ, (respectively, g ∈ G0 [g(0), g(1)). In particular, for g ∈ Gλ, (respectively, g ∈ G0 [a, b) ⊆ [0, 1) ∩ [g(0), g(1)). λ) f · δg is in Qλ (respectively, F λ) iff supp(f ) ⊆ [0, 1) ∩ λ) X[a,b) · δg is in Qλ (respectively, F λ) iff Proof. The first item is an easy calculation using part (1) of Lemma 2.5 and the fact that αg(X[a,b)) = X[g(a),g(b)). The second item follows easily from the first. Proposition 2.13. If λ is rational, then Aλ 0 and F λ are AF-algebras. In particular, if λ = p/q where 0 < p < q are in lowest terms, then F λ is the UHF algebra n∞ where n = pq. Moreover, the minimal projections in the finite-dimensional subalgebras can all be chosen from the canonical commutative subalgebra C λ (cid:3) 0 (R) · δI . Proof. We have shown in Proposition 2.1 that if λ = p/q where 0 < p < q are in lowest terms, then Γλ = Z[1/n], where n = pq. Now, any element in Z[1/n] has the form m/nk = m(1/nk) where k ≥ 1. Therefore any of the generating partial isometries X[a,b) · δ[1:c] ∈ Aλ 0 can (by bringing a, b and c to a common denominator) be written (assuming c > 0) as a finite linear combination of partial isometries of the form X[l/nk,(l+1)/nk)·δ[1:1/nk]. For partial isometries in F λ we would have to restrict 0 ≤ l ≤ nk−1 and such partial isometries generate an nk by nk matrix subalgebra of F λ. It should now be clear that F λ is a UHF algebra of type n∞. (cid:3) 8 At this point we define some special elements in Qλ which behave very much like the isometries Sµ ∈ On, except for the fact that some of them are not isometries. Definition 2.14. Fix 0 < λ < 1 and let k be a positive integer. Define mk to be the unique positive integer satisfying: mkλk < 1 ≤ (mk + 1)λk. For 0 ≤ m ≤ mk define partial isometries Sk,m ∈ Qλ via: Sk,m = X[mλk,(m+1)λk ) · δgk,m where gk,m = [λk : mλk]. Note: for m < mk the Sk,m are actually isometries, and Sk,mk is an isometry iff 1 = (mk + 1)λk. Remarks. The defining inequalities mkλk < 1 ≤ (mk + 1)λk for the positive integer mk are equivalent to: 0 < λ−k − mk ≤ 1. In particular, these differences are positive and bounded above by 1. In the case of Q1/n we have mk = nk − 1. Generally we have mk Lemma 2.15. With the previously defined elements we have: 1 ≤ mk < 1 ≤ (mk + 1) ≤ (m1 + 1)k. S∗k,m = X[0,1) · δg−1 k,m and S∗k,mk = X[0,λ−k−mk) · δg−1 k,mk where for all m, g−1 k,m = [λ−k : −m]. Moreover, for 0 ≤ m < mk, S∗k,mSk,m = X[0,1) · δ1 = e while S∗k,mk Finally, for 0 ≤ m < mk, Sk,mS∗k,m = X[mλk,(m+1)λk) · δ1 while Sk,mk S∗k,mk Sk,mk = X[0,λ−k−mk) · δ1. = X[mkλk,1) · δ1, so that mkXm=0 Sk,mS∗k,m = X[0,1) · δ1 = e. Proof. These are just straightforward calculations based on Lemma 2.5 which we leave to the reader. (cid:3) Theorem 2.16. For each λ with 0 < λ < 1, consider the partial isometries S1,m for m = 0, 1, ..., m1 m=0 S1,mS∗1,m = 1. For λ = 1/n, where m1λ < 1 ≤ (m1 + 1)λ. For m < m1, S1,m is an isometry andPm1 m1 = n − 1, S1,m1 is also an isometry, and Q1/n ∼= On, the usual Cuntz algebra. Proof. The first statement is clear. With λ = 1/n we have inside Q1/n, n isometries one for each m = 0, 1, ..., (n − 1) defined by: Sm = X[ m Using Lemma 2.12, we easily see that for each m, Sm ∈ Q1/n. Then, using item (1) of Lemma 2.5 we calculate: n ) · δgm where gm = [1/n : m/n] and so S∗m = X[0,1) · δg−1 m = [n : −m]. n , m+1 where g−1 m S∗mSm = X[0,1) · δ1 = e and SmS∗m = X[ m n , m+1 n ) · δ1 and so n−1Xm=0 SmS∗m = X[0,1) · δ1 = e. Since e is the identity of Q1/n, we have constructed a unital copy of On inside Q1/n. Now one shows by induction that for each k > 0 the product of exactly k of these n isometries has the form Sk,m where Sk,m has the same defining equation as Sm above but with nk in place of n and m = 0, 1, ..., (nk − 1). These new isometries have range projections Sk,mS∗k,m = X[ m nk ) · δ1 which therefore lie in this copy of On. By adding up some of these projections, we can get any projection of the form X[a,b) · δ1 where 0 ≤ a < b ≤ 1 and both a, b have the form m/nk. But any element a ∈ Γ1/n can be written as a = m for a sufficiently large k ≥ 0 and some m ∈ Z depending on k, and any pair a, b can be brought to a common denominator nk. Hence any projection of the form X[a,b) · δ1 in Q1/n is in this copy of On. nk , m+1 nk Now, a straightforward calculation gives us: 9 (1) nk−1Xm=1 Sk,mS∗k,m−1 = nk−1Xm=1 X[ m+1 nk , m+2 nk ) · δ[1 : 1/nk] = X[1/nk,1) · δ[1 : 1/nk] ∈ On. Finally, let X[a,b) · δg ∈ Q1/n be an arbitrary generator. By taking adjoints if necessary we can assume that g has the form g = [nk : ∗] where k ≥ 0. Since Sk,0 is an isometry in On it suffices to prove that Sk,0(X[a,b) · δg) ∈ On. That is, we are reduced to the case g = [1 : c] and again by taking adjoints if necessary we can assume that c ≥ 0. The case c = 0 is done and so we can assume that c > 0. So (with possibly new a, b) we have X[a,b) · δ[1 : c] where 0 < c ≤ 1 and [a, b) ⊆ [0, 1) ∩ [c, c + 1) = [c, 1). But, X[a,b)· δ[1 : c] = X[a,b)X[c,1)· δ[1 : c] = X[a,b)· δ1X[c,1)· δ[1 : c] and we already know that X[a,b) · δ1 ∈ On. Therefore it suffices to see that X[c,1) · δ[1 : c] ∈ On. However, c = l/nk for some 0 < l < nk and so: X[c,1) · δ[1 : c] = X[l/nk,1) · δ[1 : l/nk] =(cid:16)X[1/nk,1) · δ[1 : 1/nk](cid:17)l which is in On by Equation 1. Since all generators for Q1/n are in On we're done. (cid:3) 2.2. K-Theory of Qλ for λ rational. : Since Aλ 0 is stable and stably isomorphic to the UHF algebra Fλ, each of its projections is equivalent to one in some finite-dimensional subalgebra and 0 (R), and in this case the trace induces an isomorphism from K0(Aλ hence to some projection in C λ 0 ) onto Γλ = Z[1/(pq)] ⊂ R. This isomorphism carries the projection e = X[0,1) · δ1 which is the identity of Qλ and F λ onto 1 ∈ Z[1/(pq)]. Now, since Aλ 0 we can use the Pimsner-Voiculescu exact sequence to calculate K∗(Aλ) = K∗(Qλ). When we do this we get: 0 ) = {0}, and since Aλ = Z ⋊λ Aλ 0 is AF, K1(Aλ K1(Qλ) = {0}, and K0(Qλ) = Z[1/(pq)]/(1 − λ)Z[1/(pq)]. Proposition 2.17. For λ rational with λ = p/q in lowest terms, we have K1(Qλ) = {0}, and K0(Qλ) ∼= Z[1/(pq)]/(1 − λ)Z[1/(pq)] ∼= Z(q−p). Proof. By Proposition 2.1, Γλ = Z[1/(pq)], so we must show that Z[1/(pq)]/(1 − (1/(pq))Z[1/(pq)] ∼= Z(q−p). Since (q − p) = (1− p/q)q and every element of Z[1/(pq)] is of the form m/(pq)N , it is easy to see that (q − p)Z[1/(pq)] = (1 − p/q)Z[1/(pq)]. Now, (q − p) and (pq)N are relatively prime for any N and so there exist a, b ∈ Z so that 1 = a(q − p) + b(pq)N and hence m/(pq)N = (q − p)am/(pq)N + mb. That is, m/(pq)N and mb represent the same element in the quotient. So, every element in the quotient has an integer representative. Two integers c, d represent the same element in the quotient if and only if c − d = (p − q)n/(pq)N , or (c − d)(pq)N = n(q − p). But then: (c − d) = (c − d)[a(q − p) + b(pq)N ] = (c − d)a(q − p) + b(c − d)(pq)N = [(c − d)a + bn](q − p). That is, c, d represent the same element in Z/(q − p)Z = Z(q−p). On the other hand if (c − d) is in (q − p)Z then clearly, [c] = [d] in Z[1/(pq)]/(1 − (1/(pq))Z[1/(pq)] and we are done. Corollary 2.18. If λ = p/q in lowest terms, then (cid:3) In particular, if λ = k F λ = F p/q ∼= U HF ((pq)∞) and Qλ = Qp/q ∼= O(q−p+1). k+1 then F λ ∼= U HF ((k(k + 1))∞) and Qλ ∼= O2. 10 Proof. Since each Qλ is separable, nuclear, simple, purely infinite and in the bootstrap category Nnuc once we show that the class of the identity e ∈ Qλ is a generator for K0(Qλ) = Z/(q − p)Z, the Kirchberg-Phillips Classification Theorem, Theorem 8.4.1 of [RS], shows that Qλ ∼= O(q−p+1). To this end we observe that since e is mapped to 1 in Z[1/pq], we must show that [1] is a generator for K0(Qλ) = Z[1/pq]/(1 − (p/q))Z[1/pq]. Now, by the proof of the previous proposition, k[1] = [k · 1] = 0 ∈ Z[1/pq]/(1 − (p/q))Z[1/pq] if and only if [k · 1] = 0 ∈ Z/(q − p)Z if and only if k − 0 = m(q − p) for some m ∈ Z if and only if k is a multiple of (q − p). That is, [1], [2 · 1], . . . , [(q − p − 1) · 1] are all nonzero in K0(Qλ) = Z/(q − p)Z and hence [1] is a generator. 2.3. The K-Theory of the Algebras Aλ 0 for λ irrational. The case λ rational is much simpler, and while it does fit into the following scheme, it does not need this deeper machinery. Initially, we (and others) believed that the algebras Aλ In fact we will show that Aλ 0 is never AF when λ is irrational. We will set up our examples to fit the situation on page 1487 of [Put2] so that we can apply the six-term exact sequence of Theorem 2.1 on page 1489 of [Put2]. We let Γ = Γλ ∼= G0 λ. Thus, Γ ⊂ R is a countable dense subgroup of R which acts on R by translations. Before looking at the crossed product of Γ acting on C λ 0 ) we first consider the crossed product of Γ acting on C0(R). Since Γ acts on R by translation we can Fourier transform to get an isomorphism: 0 were AF algebras when λ is irrational. 0 (R) = C0(Rλ) (which gives us Aλ (cid:3) Then, by Connes' Thom isomorphism we get for i = 0, 1: Γ ⋊ C0(R) ∼= R ⋊ C(Γ). Proposition 2.19. The composition: Ki(Γ ⋊ C0(R)) ∼= Ki( R ⋊ C(Γ)) ∼= Ki+1(C(Γ)). K1(C0(R)) i∗−→ K1(Γ ⋊ C0(R)) b −→ K1( R ⋊ C(Γ)) ∼=−→ K0(C(Γ)) takes the generator [u] ∈ K1(C0(R)) = Z · [u]; where u is the Bott element in C0(R)1 defined by u(t) = 1+it 1−it ; to [1Γ] where 1Γ is the identity function in C(Γ). Proof. We first work on the right hand side of this sequence of maps. Let u(t) = 1 + ε(t), then by the proof of Connes' Thom isomorphism from K0(C(Γ)) ⊗Z K1(C0(R)) −→ K1(R ⋊ C(Γ)) we see that [1Γ] ⊗ [u] gets mapped to the class [1 + (convolution by ε · 1Γ)]. Now in this displayed equation, K0(C(Γ)) ⊗Z K1(C0(R)) = K0(C(Γ)) ⊗Z Z · [u] = K0(C(Γ)) · [u] ∼= K0(C(Γ)). Thus, [1Γ] in K0(C(Γ)) gets mapped to the class [1 + (convolution by ε · 1Γ)] by the Thom isomorphism. On the other hand, the map K1((C0(R)1) −→ K1((Γ ⋊ C0(R))1) takes [u] 7−→ [δ0 · ε + 1] and by the Fourier transform this goes to [(convolution by ε · 1Γ) + 1] in K1(R ⋊ C(Γ)). Combining these we get: 1 ∈ Z 7−→ [u] ∈ Z · [u] = K1((C0(R))1) = K1(C0(R)) 7−→ [1Γ] ∈ K0((C(Γ)). (cid:3) Now, by Proposition 2.1 we know Γ in many cases so that these last groups are quite computable. In the notation of [Put2] we define the transformation groupoids: G := Rλ ⋊ Γ, G′ := R ⋊ Γ, and H := Γ ⋊ Γ. 0 = C∗r (G) is the reduced C∗-algebra of G; Γ ⋊ C0(R) = C∗r (G′) is the reduced C∗-algebra of Then, Aλ G′; and K(l2(Γ)) is the reduced C∗-algebra of H. By the proof of Proposition 2.10 there is a continuous proper surjective map: Rλ → R, where points in R which are not in Γ each have a single pre-image, while points γ ∈ Γ have exactly two pre-images in Rλ, which we denote by γ− and γ+. Thus, there are two disjoint embeddings of Γ in Rλ : 11 i0, i1 : Γ → Rλ : i0(γ) = γ−, i1(γ) = γ+. Now in order to mesh with the notation of [Put2], we let Y := Γ with the equivalence relation, "="; X := Rλ, with the equivalence relation (i0(γ) ∼ i1(γ)); and quotient π : X → X′ := R where X′ = X/(i0(γ) ∼ i1(γ)) = R; while the "factor groupoid" of G = Rλ×Γ = X×Γ is G′ := R×Γ = X′×Γ. We represent each of these three C∗-algebras on H := l2(Γ+) ⊕ l2(Γ−) where Γ± = {γ± γ ∈ Γ} in the following way. First we denote the natural orthonormal basis elements of H by δa+ and δa− for each a ∈ Γ. Now the unitary representation U of Γ on H is Uγ(δa±) = δ(a−γ)± . The actions of C0(Rλ), C0(R), and C0(Γ) on H are as follows for f1 ∈ C0(Rλ), f2 ∈ C0(R), f3 ∈ C0(Γ), and δa± ∈ H π1(f1)(δa± ) = f1(a±)δa± π2(f2)(δa± ) = f2(a)δa± π3(f3)(δa±) = f3(a)δa± . These three covariant pairs of representations, (π1, U ), (π2, U ), and (π3, U ) define representations of 0 , C∗r (G′) = Γ ⋊ C0(R), and C∗r (H) = K(l2(Γ)) respectively on H. Since each of these C∗r (G) = Aλ C∗-algebras is simple these representations are faithful. Now, one checks that the hypotheses of Theorem 2.1 of [Put2] are satisfied. As in [Put1, Put2] one shows that the two mapping cone algebras of the inclusions: C∗r (G′) = Γ ⋊ C0(R) −→ Aλ 0 = C∗r (G) and C∗r (H) −→ C∗r (H) ⊕ C∗r (H) : ( x 7→ (x, x) ) have isomorphic K-Theory. One then pastes these isomorphisms into the mapping cone long exact 0 = C∗r (G). Next one observes that for any C∗-algebra, B sequence for C∗r (G′) = Γ ⋊ C0(R) −→ Aλ the diagonal embedding B −→ B ⊕ B induces the diagonal embedding K∗(B) −→ K∗(B) ⊕ K∗(B) with quotient isomorphic to K∗(B) (this is true for any abelian group). This implies that K∗(B) ∼= K∗+1(M (B, B ⊕ B)) so that we get the six-term exact sequence from [Put2]: K1(C∗r (H)) / K0(C∗r (G′)) / K0(C∗r (G)) In our set-up this becomes: K1(C∗r (G)) K1(C∗r (G′)) K0(C∗r (H)) {0} / K0(Γ ⋊ C0(R)) / K0(Γ ⋊ C0(Rλ)) K1(Γ ⋊ C0(Rλ)) K1(Γ ⋊ C0(R)) Z Which by Connes' Thom isomorphism becomes: {0} / K1(C(Γ)) / K0(Aλ 0 ) K1(Aλ 0 ) K0(C(Γ)) Z By Proposition 2.19, the nonzero element [1Γ] in K0(C0(Γ)) ∼= K1(Γ ⋊ C0(R)) is mapped to the image of the class [u] in K1(Γ ⋊ C0(R)) by Connes' Thom isomorphism, and then the image of [1Γ] in K1(Γ⋊C0(Rλ)) is the same as the image of [u] under the inclusion K1(Γ⋊C0(R)) −→ K1(Γ⋊C0(Rλ)). However, this is clearly the same as the image of [u] under the inclusion K1(C0(R)) → K1(C0(Rλ)) → K1(Γ ⋊ C0(Rλ)). This composition is 0 since C0(Rλ) is an AF-algebra. That is, the element [1Γ] in / / O O o o o o / / O O o o o o / / O O o o o o 12 0 ) and hence is in the image of the map Z −→ K0(C0(Γ)). Since K0(C0(Γ)) is mapped to 0 in K1(Aλ [1Γ] generates a copy of Z in K0(C0(Γ)), we have a nonzero homomorphism from Z to Z[1Γ] which is onto and hence one-to-one. By the exactness, the map K0(Aλ CONCLUSION : K0(Aλ 0 ) ∼= K1(C(Γλ)) and K1(Aλ 0 ) −→ Z is the zero map. 0 ) ∼= K0(C(Γλ))/[1Γλ ]Z. Proposition 2.20. If λ is irrational, then K1(Aλ 0 ) 6= {0} so that Aλ 0 is not an AF-algebra. Proof. By items (3) and (5) of Proposition 2.2 we see that when λ is irrational, K0(C(Γλ)) is not singly generated so that K1(Aλ (cid:3) 2.4. K-theory computations of particular Qλ for λ irrational. Example(s) λ = 1/√n: 0 ) ∼= K0(C(Γλ))/[1Γλ ]Z 6= {0}. n > 1 a square-free integer. Using Proposition 2.1, we get: for K1(F λ) = K1(Aλ K0(F λ) = K0(Aλ 0 ) = K1(C( Γλ)) = Z[1/n] ⊕ Z[1/n] 0 ) = (K0(C( Γλ)))(cid:14)Z[1] = (Z[1] ⊕ Z[1/n])(cid:14)Z[1] = Z[1/n]. To compute the K-theory of Qλ in this case using the Pimsner-Voiculescu exact sequence, one must first compute the induced automorphism λ∗ on K1(C( Γλ)) and on K0(C( Γλ)) by a more detailed analysis of the proof of [Sc, Proposition 2.11]. In the case of K1(C( Γλ)) we get a copy of the group Γλ = Z[1/n] + Z[1/n]√n and the action on Γλ is just multiplication by λ = 1/√n. As an action translated to the abstract group Z[1/n] ⊕ Z[1/n], the automorphism becomes λ∗(a, b) = (b, a/n). Therefore, id∗ − λ∗ on K0(Aλ 0 ) = Z[1/n] ⊕ Z[1/n] to itself is clearly 1 : 1. Now it is an instructive exercise to show that the kernel of the homomorphism is exactly the range of the homomorphism (a, b) ∈ Z[1/n] ⊕ Z[1/n] 7→ [a + b] ∈ Z[1/n](cid:14)(1 − 1/n)Z[1/n] id∗ − λ∗ : Z[1/n] ⊕ Z[1/n] −→ Z[1/n] ⊕ Z[1/n]. Hence, we have the isomorphisms: (Z[1/n] ⊕ Z[1/n])(cid:14)(id∗ − λ∗)(Z[1/n] ⊕ Z[1/n]) ∼= Z[1/n](cid:14)(1 − 1/n)Z[1/n] ∼= Z(cid:14)(n − 1)Z. where the last isomorphism follows from the proof of Proposition 2.17 with p = 1 and q = n. Once we have computed the action of λ∗ on K1(Aλ Now, by Proposition 2.11 of [Sc] we have the isomorphism: 0 ) = Z[1/n] we will be ready to compute K∗(Qλ). K0(C(Γλ)) ∼= (Z[1] ⊗Z Z[1]) ⊕ (Z[1/n] ⊗Z Z[1/n]) = Z[1] ⊕ (Z[1/n] ⊗Z Z[1/n]). The action of λ∗ on Z[1] is of course the identity. However, the action of λ∗ on (Z[1/n] ⊗Z Z[1/n]) is just x ⊗ y 7→ y ⊗ x/n. If one combines this with the multiplication isomorphism x ⊗ y 7→ xy : Z[1/n]⊗Z Z[1/n] −→ Z[1/n] we see that λ∗ acts as multiplication by 1/n on Z[1/n] = Z[1/n]⊗Z Z[1/n]. Thus, λ∗ on the quotient K1(Aλ 0 ) = Z[1/n] is just multiplication by 1/n. Therefore, id∗ − λ∗ becomes multiplication by (1 − 1/n) on Z[1/n] which is clearly 1 : 1. Applying the Pimsner-Voiculescu exact sequence and recalling that Ki(Qλ) = Ki(Aλ) we get the isomorphisms: K0(Qλ) ∼= Z(cid:14)(n − 1)Z, and K1(Qλ) ∼= Z(cid:14)(n − 1)Z, For n > 2 we get K1 6= 0 and so these are not Cuntz algebras, in fact not even Cuntz-Krieger algebras since K1 has nonzero torsion. For λ = 1/√2 however we get K0 = 0 = K1 and by classification theory, we must have Q1/√2 ∼= O2! However, even in this case the fixed point algebra, is NOT AF since it has K1 = Z[1/2], the tape-measure group. So for the simplest irrational number 1/√2 we get the Cuntz algebra, O2 with a strange gauge action of T. for λ = 1/√n. 13 Remarks. In the examples below it is important to note that any polynomial of the form f (x) = xn + axn−1 + ··· + bx ± 1 has at most n − 1 roots in the open interval (0, 1) because the product of all the roots of f must equal ±1. Example(s) quadratic integers and an algorithm: If both λ and λ−1 are quadratic integers with λ ∈ (0, 1), then λ2 + aλ ± 1 = 0 where the integer polynomial f (x) = x2 + ax ± 1 is irreducible over Q. With these restrictions there are two cases, either f (x) = x2 + ax − 1 where a > 0 and λ = 1/2·(√a2 + 4−a) ∈ (0, 1); or f (x) = x2+ax+1 where a ≤ −3 and λ = 1/2·(−√a2 − 4−a) ∈ (0, 1). In the first case, λ2 + aλ − 1 = 0, with a > 0, so that λ + a − λ−1 = 0 and λ−1 = a + λ. For this case we outline an algorithm using the ideas of the Smith Normal Form and the Pimsner-Voiculescu exact sequence to calculate the K-Theory. By Proposition 2.2, and the CONCLUSION before Proposition 0 ) ∼= K1(C(Γλ)) =V1(Γλ) = Γλ ∼= Z2. Giving Γλ its Z-basis {1, λ} we 2.20, Γλ = Z + Zλ and K0(Aλ 1 −a (cid:21) . So, (id − λ∗) = see that the action of the automorphism λ∗ on K0(Aλ (cid:20) 1 −1 (a + 1) (cid:21) := M. To compute the kernel and cokernel of this matrix mapping Z2 → Z2 we D =(cid:20) 1 0 0 a (cid:21) . Hence, on K0(Aλ ker(id − λ∗) = ker(M ) ∼= ker(D) = {0} and coker(id − λ∗) ∼= coker(D) = Z/aZ. row and column-reduce M over Z to obtain matrices P, Q ∈ GL(2, Z) so that P M Q = D where D is diagonal over Z. Then ker(M ) ∼= ker(D) and Z2/M (Z2) ∼= Z2/D(Z2). In this case, we get 0 ) ∼= Γλ has matrix: (cid:20) 0 0 ) we have −1 1 Now we compute (id − λ∗) on K1(Aλ 0 ) ∼= K0(C(Γλ))/Z · 1o = (Z · 1o ⊕ Z(1 ∧ λ))/Z · 1o = Z(1 ∧ λ). 0 ) ∼= Z. Therefore, Now, λ∗(1∧λ) = λ∧λ2 = λ∧(1−aλ) = λ∧1 = (−1)1∧λ. That is, λ∗ = −id on K1(Aλ (id− λ∗) = multiplication by 2 on Z(1∧ λ) which has ker(id− λ∗) = {0} and cokernel(id− λ∗) ∼= Z/2Z. Applying these results to the Pimsner-Voiculescu exact sequence we obtain: K0(Qλ) = Z/aZ and K1(Qλ) = Z/2Z, for λ2 + aλ − 1 = 0, n ≥ 1. λ = (1/2)(√5 − 1) is the inverse of the golden mean, we get K0 = {0} and K1 = Z/2Z. None of these examples are Cuntz-Krieger algebras since K1 is not torsion-free. In particular, when 0 ) ∼= Γλ = Z + Zλ with Z-basis In the second case, λ2 + aλ + 1 = 0, we have as above, K0(Aλ {1, λ}; the diagonal version of (id − λ∗) is D = diag[1, (a + 2)] so that ker(id − λ∗) = {0} and 0 ) ∼= Z(1 ∧ λ) only now, λ∗ = id here so that coker(id − λ∗) ∼= Z/(a + 2)Z. On the other hand, K1(Aλ (id − λ∗) = 0 and hence ker(id − λ∗) ∼= Z while coker(id − λ∗) ∼= Z. By Pimsner-Voiculescu we get K0(Qλ) = Z ⊕ (Z/(a + 2)Z) and K1(Qλ) = Z, for λ2 + aλ + 1 = 0, a ≤ −3. We note that in this case, Qλ has the correct K-theory to be a Cuntz-Krieger algebra (and is therefore stably isomorphic to one), and that in the case a = −3 (i.e., λ = (1/2)(3−√5)) we have K0 = Z = K1. Example cubic integers: If λ and λ−1 are cubic integers with λ ∈ (0, 1), then λ3 + aλ2 + bλ± 1 = 0 where the integer polynomial f (x) = x3 + ax2 + bx ± 1 is irreducible over Q. Such an f is irreducible if and only if f (1) 6= 0 6= f (−1). There are two cases depending on the constant, ±1. First, consider f (x) = x3 + ax2 + bx − 1 = 0 with f (1) = a + b 6= 0 and f (−1) = a − b − 2 6= 0 so that f is irreducible. Now assume a + b is positive (but a 6= b + 2). Then f (0) = −1 and f (1) = a + b > 0 so that f has a unique root in (0, 1) since it is a cubic. 14 Next consider the same polynomial, f (x) = x3 + ax2 + bx− 1 = 0, with a + b negative (but a 6= b + 2). Since both f (0) and f (1) are negative, in order to have a solution the function f must have a local maximum on (0, 1). There are examples with no solutions in (0, 1); for example, f (x) = x3 − 3x − 1. In order to have a unique solution, then considering f′(x), one would need 4a2 − 12b = 0 : while this has many solutions, they all satisfy a ≤ b and so we can not have a + b < 0. So solutions are not unique in this case. But, there are infinitely many cubics with two distinct solutions in (0, 1); eg., f (x) = x3 − (a + k)x2 + ax − 1 for a ≥ k + 4 and k ≥ 1 has two solutions in (0, 1), since f (.5) > 0. We now calculate the K-theory of Qλ assuming that λ satisfies f (x) = x3 + ax2 + bx − 1 = 0, where a+b 6= 0, and a−b 6= 2. Now, λ3+aλ2+bλ−1 = 0, so that λ3 = 1−aλ2−bλ and λ−1 = λ2+aλ+b. Then, 0 ) ∼= K1(C(Γλ)) =Vodd(Γλ) = Γλ⊕(Γλ∧Γλ∧Γλ) = Γλ⊕(Z(1∧λ∧λ2)) ∼= Γλ = Z+Zλ+Zλ2 and K0(Aλ 0 ) ∼= Z4 Z4. Giving Γλ its natural Z-basis {1, λ, λ2} the induced homomorphism (id − λ∗) on K0(Aλ yields the diagonal matrix, D = diag[1, 1, (a + b), 0] so that on K0(Aλ 0 ) we have Now, K1(Aλ computations we get for K1(Aλ 0 ) we have ker(id − λ∗) ∼= ker(D) ∼= Z and coker(id − λ∗) ∼= coker(D) = (Z/(a + b)Z) ⊕ Z. 0 ) ∼= K0(C(Γλ))/Z · 1o = V2(Γλ) = Z(1 ∧ λ) + Z(1 ∧ λ2) + Z(λ ∧ λ2) ∼= Z3. By similar ker(id − λ∗) ∼= ker(D) = {0} and coker(id − λ∗) ∼= coker(D) = Z/(a + b)Z. 0 ) ∼= Z3; the matrix D = diag[1, 1, (a + b)]. Hence, on K1(Aλ Applying these results to the Pimsner-Voiculescu exact sequence we obtain: K0(Qλ) = Z ⊕ (Z/(a + b)Z) and K1(Qλ) = Z ⊕ (Z/(a + b)Z) for λ3 + aλ2 + bλ − 1 = 0. Remarks. In case a + b = 1 (which has infinitely many solutions corresponding to infinitely many distinct invertible cubic integers λ ∈ (0, 1)) we get K0(Qλ) = Z = K1(Qλ), which as noted above is also true for the invertible quadratic integer, λ = (1/2)(3 − √5). In the general cubic case with constant term −1 we always have non-torsion elements in both K0 and K1 : this is the opposite of the case where the constant term is +1, where we see below that K0 and K1 are both torsion groups. A similar phenomenon occurs in the quadratic case above, except that there we get torsion in the −1 case and non-torsion in the +1 case! That this may be a periodic phenomenon is supported by a calculation of two quartic examples: first, the unique solution λ ∈ (0, 1) to the irreducible quartic f (x) = x4 − 3x3 + 1 gives us K0 = Z and K1 = Z ⊕ (Z/3Z) ⊕ (Z/3Z); while, second, the unique solution λ ∈ (0, 1) to the irreducible quartic f (x) = x4 + 3x3 − 1 gives us K0 = (Z/3Z) ⊕ (Z/3Z) and K1 = (Z/9Z) ⊕ (Z/2Z), similar to the quadratic case. Proposition 2.21 is further evidence. When an irreducible polynomial f (x) = xn + axn−1 + ··· ± 1 has two roots, λ1, λ2 ∈ (0, 1), then Γλ1 ∼= Γλ2 as rings (but not as ordered rings, for that would imply equality). Still, Qλ1 ∼= Qλ2 (at least stably) since the calculation of their K-groups are identical. Their KMS states are not equivalent since the type III factors that they generate are not isomorphic, as we will see below. Proposition 2.21. Suppose λ satisfies the irreducible (over Z) polynomial, f (x) = xn + ··· ± 1 = 0. (1) For n odd, if f (x) = xn + ··· + 1 then K0(Qλ) has Z/2Z as a summand. While, if f (x) = xn+···−1 then K0(Qλ) has Z as a summand (so, by the next Proposition, rank(K0) = rank(K1) ≥ 1 in this case). (2) For n even, if f (x) = xn +··· + 1 then K1(Qλ) has Z as a summand (so, by the next Proposition, rank(K0) = rank(K1) ≥ 1 in this case). While, if f (x) = xn + ··· − 1 then K1(Qλ) has Z/2Z as a summand. 0 ) there is a λ∗-invariant summand, (1∧ λ∧ λ2 ∧···∧ λn−1)Z. Depending on n(mod 2) Proof. In K∗(Aλ and the constant term ±1, the action of λ∗ on this summand is ±id. Hence, (id − λ∗) here is either 0 or 2(id). Applying Pimsner-Voiculescu gives a summand in K∗(Qλ) of either Z or Z/2Z. (cid:3) Proposition 2.22. Suppose λ is algebraic. (1) Then, rank(K0(Qλ)) = rank(K1(Qλ)) so that Qλ is not stably isomorphic to O∞. (2) If λ and λ−1 are both algebraic integers and Qλ is stably isomorphic to a Cuntz algebra On, then the minimal polynomial of λ has odd degree and constant term +1. Moreover, n is congruent to 3(mod 4) and all such Cuntz algebras appear this way. Proof. To see part (1) we tensor the Pimsner-Voiculescu exact sequence by Q (which preserves exact- ness) to obtain an exact hexagon of Q-vector spaces: 15 V1 θ1 / V1 τ1 / / K Q 1 µ0 µ1 K Q 0 τ0 V0 V0 θ0 where Vi = Ki(Aλ 0 ) ⊗ Q, and K Q i = Ki(Aλ) ⊗ Q. Then dim(K Q 0 ) = rank(µ0) + nullity(µ0) = rank(µ0) + rank(τ0) and dim(K Q 1 ) = rank(µ1) + rank(τ1) and rank(τ0) + rank(θ0) = dim(V0) = rank(θ0) + rank(µ1) so that rank(τ0) = rank(µ1). Similarly, rank(τ1) = rank(µ0), so that: dim(K Q 0 ) = rank(µ0) + rank(τ0) = rank(µ1) + rank(τ1) = dim(K Q 1 ). That is, rank(K0(Qλ)) = dim K0(Qλ) ⊗Z Q = dim K0(Aλ) ⊗Z Q = ··· rank(K1(Qλ)). By Proposition 2.21, if the minimal polynomial of λ has even degree, then K1(Qλ) 6= {0}, and so Qλ cannot be stably isomorphic to a Cuntz algebra. If Qλ is stably isomorphic to On then n is finite by part (1)and by Proposition 2.21, the order of K0(Qλ) must be even and therefore n must be odd. Furthermore, the minimal polynomial must have constant term +1. In order for K0(Qλ) to be a finite cyclic group of even order, it must be of the form Z/mZ ⊕ Z/2Z where m is odd since Z/2Z is a summand. Let m = 2k + 1 then as claimed. n = ♯[Z/mZ ⊕ Z/2Z] + 1 = 2m + 1 = 4k + 3 In the examples below where λ3 + aλ2 + bλ+ 1 = 0, and either a− b = 1 and b ≤ −2 OR a = b = −1 and b ≤ −1, we obtain (stably, at least) all the Cuntz algebras On where n ≡ 3(mod 4). Now consider the case of irreducible cubics of the form f (x) = x3+mx2+nx+1; so f (1) = m+n+2 6= 0 and f (−1) = m − n 6= 0. Since f (0) = 1, if we have f (1) = m + n + 2 < 0, then we have as above a unique root in (0, 1). (cid:3) If f (1) = m + n + 2 > 0, we can have distinct roots. For example, if n = −4 and m = 3, then, f (x) = x3 + 3x2 − 4x + 1 has two roots in (0, 1). If n << 0, we get several solutions m for each n: eg., n = −7 implies that any m with 6 ≤ m ≤ 9 will yield a polynomial with two roots in (0, 1). We now calculate the K-Theory of Qλ assuming λ satisfies f (λ) = λ3 + mλ2 + nλ + 1 = 0. Again, 0 ) ∼= Z3, the matrix K0(Aλ D = diag[1, 1, (m − n)]. Both matrices are 1 : 1 since m + n + 2 6= 0 6= m − n. We get: 0 ) ∼= Z4, but now the diagonal matrix D = diag[1, 1, (m + n + 2), 2]. On K1(Aλ K0(Qλ) = Z/(n + m + 2)Z ⊕ Z/2Z and K1(Qλ) = Z/(m − n)Z for λ3 + mλ2 + nλ + 1 = 0. To obtain Cuntz algebras, we need m − n = ±1. It turns out f (1) > 0 can not occur, so we must have f (1) = m + n + 2 < 0 hence there is a unique root λ in (0, 1). Combining this inequality with m − n = ±1 we get exactly two infinite families of solutions; m = n + 1 for n ≤ −2, OR m = n − 1 / O O o o o o 16 for n ≤ −1. In either case, the sequence of numbers {m + n + 2} is the same: {2k + 1k ≥ 0}. For this sequence we get the K0 groups: Z/(2k + 1)Z ⊕ Z/2Z ∼= Z/(4k + 2)Z. Since the K1 groups are all {0}, by construction, the algebras Qλ are (at least stably) the Cuntz algebras, O4k+3 for k ≥ 0. That is, O3, O7, O11, etc. Example, λ transcendental: Lemma 2.23. Let ϕ : Ln∈Z Z −→ Z be the surjective homomorphism, φ({an}) := Pn∈Z an; and let S ∈ Aut(Ln∈Z Z) be the shift S({an}n∈Z) := {an−1}n∈Z. Then, (id − S) is 1 : 1 and ker(ϕ) = Im(id − S), so that (Ln∈Z Z)/Im(id − S) ∼= Z. Proof. As a model for Ln∈Z Z we use Z[x, x−1] =Ln∈Z Zxn, the ring of Laurent polynomials over Z (i.e., the group ring over Z of the group {xnn ∈ Z}). Here, ϕ is the augmentation map, S is multiplication by x, and (id − S) is multiplication by (1 − x) which is 1 : 1. Now, anxn+N ∈ ker(ϕ). an = 0 ⇔ NXn=−N where initially q(x) ∈ Q[x]. Since p(x) ∈ Z[x] it is easy to see that in fact, q(x) ∈ Z[x] also. Then, n=−N an = 0. Hence, p(x) factors: p(x) = (1 − x)q(x) NXn=−N anxn ∈ ker(ϕ) ⇔ NXn=−N n=−N anxn+N ∈ Z[x] so p(1) =PN Let p(x) =PN NXn=−N anxn = x−N p(x) = (1 − x)x−N q(x) ∈ (1 − x)Z[x, x−1] = Im(id − S). That is, ker(ϕ) ⊆ Im(id − S), and the other containment is immediate. Proposition 2.24. If λ is transcendental then (cid:3) K0(Qλ) ∼= ∞Mn=1 Z ∼= K1(Qλ). Proof. In this case, by Proposition 2.2 and the CONCLUSION before Proposition 2.20 we have: K0(Aλ 0 ) = ∞Mk=1 2k−1^ (Γλ) and K1(Aλ 0 ) = ∞Mk=1 2k^(Γλ) where Γλ = ∞Mn=−∞ Zλn. integers, we have: sum of (λ∗-invariant) examples of the previous lemma where the action of λ∗ is just the shift. The Now, each individual summandVm(Γλ) is invariant under λ∗ and yields (for m > 1) an infinite direct general case is notation-heavy, so we do the examples, V2 and V3 . Letting Z+ denote the positive 2^(Γλ) = Mk∈Z+ Mn∈Z (λn ∧ λn+k1 ∧ λn+k1+k2)Z! . The case m = 1 is just the group Γλ =Ln∈Z Zλn which yields a single instance of the lemma. Applying the lemma we see that (id − λ∗) is 1 : 1 on both K0(Aλ K0(Aλ of copies of Z. An application of the Pimsner-Voiculescu exact sequence completes the proof. 0 ) and that both 0 )) are isomorphic to a countable direct sum (cid:3) (λn ∧ λn+k)Z! and 3^(Γλ) = M(k1,k2)∈Z2 0 )/(id − λ∗)(K1(Aλ 0 )/(id − λ∗)(K0(Aλ Mn∈Z 0 )) and K1(Aλ 0 ) and K1(Aλ + Remark. The classification theory of Kirchberg algebras implies that for λ transcendental we have a new realisation of the algebras found in [Cu1] and denoted QN there. 2.5. The dual action of T1 on Aλ and its restriction to the gauge action on Qλ. Recall, G0 λ = {g ∈ Gλ g = 1} is a normal subgroup of Gλ. The subgroup of Gλ of elements of the form [λn : 0] is isomorphic to Z and acts on G0 λ by conjugacy: [λn : 0][1 : b][λ−n : 0] = [1 : λnb]. Thus Gλ = Z ⋊ G0 λ is a semidirect product and we can write Aλ as an iterated crossed product: 17 Aλ = Gλ ⋊α C λ 0 (R) = Z ⋊ (G0 λ ⋊α C λ 0 (R)) = Z ⋊ Aλ 0 . The dual action γ of T1 on Aλ is relative to this latter crossed product so that for each z ∈ T1 and x in the Banach ∗-algebra, l1 0 (R)) we have: α(Gλ, C λ γz(x)(g) = znx(g) if x ∈ l1 α(Gλ, C λ 0 (R)); g ∈ Gλ and g = λn. Since Aλ is defined to be the completion of this Banach ∗-algebra in its universal representation, the action γ extends uniquely to an action (also denoted by γ) of T1 as automorphisms of Aλ. The fixed point subalgebra of the dual action is, of course, exactly Aλ 0 = G0 λ ⋊α C λ 0 (R). 0 , the action γ restricts to an action of T1 on Qλ = eAλe, which we will Since the projection e is in Aλ also denote by γ. We call this the gauge action of T1 on Qλ. Now, γ is clearly a strongly continuous action of T1 on Qλ. Averaging over γ with respect to normalised Haar measure gives a positive, faithful expectation Φ of Qλ onto the fixed-point algebra which is clearly F λ: Φ(a) := 1 2πZT1 γz(a) dθ for a ∈ Qλ, and z = eiθ. Proposition 2.25. The fixed point algebra, F λ = eAλ tions of elements of the form: 0 e is the norm closure of finite linear combina- for a, b, c ∈ Γλ. Recall, Aλ X[a,b) · δg where g = [1 : c] and [a, b) ⊆ [0, 1) ∩ [c, 1 + c), 0 ∼= K(l2(Z)) ⊗ F λ. Proof. Applying the integral formula for Φ to a finite linear combination of the generators for Qλ we see that the only terms that survive are those where g = 1 : that is, g has the above form. Then we apply item (2) of Lemma 2.12 to obtain the condition on the interval [a, b). Corollary 2.26. The stabilised algebra Qλ⊗K is a crossed product of the stabilised fixed-point algebra F λ ⊗ K by an action of Z. For λ = 1/n this is a theorem of J. Cuntz. (cid:3) 0 ∼= F λ ⊗ K, and Aλ ∼= Qλ ⊗ K. By the discussion at the beginning of 0 and the proof is complete. See [Cu, Section 2]. Proof. By Proposition 2.7, Aλ subsection 3.1, Aλ ∼= Z ⋊ Aλ Remarks. If we combine the previous observation that F λ is the fixed point subalgebra of Qλ under the gauge action with Corollary 2.18 we get, for example, O2 ∼= Q2/3 with a gauge action whose fixed point subalgebra F 2/3 is a UHF algebra of type 6∞. Interestingly, F 3/4 is UHF of type 12∞ = 6∞ which is therefore isomorphic to F 2/3. So we have two gauge actions on O2 with isomorphic UHF fixed point subalgebras, with distinct, inequivalent KMS states: one where β = log(3/2) and the other where β = log(4/3) by Proposition 2.30 below. Moreover, the two von Neumann algebras generated by the GNS representations of O2 are not isomorphic as they are type IIIλ factors for λ equalling 2/3 and 3/4, respectively, by Theorem 2.35 below. (cid:3) 18 2.6. The γ-invariant semifinite weight on Aλ and its restriction to Qλ. The aim of this subsection is to exhibit the unique KMS states for the gauge action on Qλ. We first recall the definition of KMS states. Definition 2.27. Let A be a C∗-algebra with a continuous action γ : R → Aut(A). Let ψ be a state on A and β ∈ R a real number. We define ψ to be a KMSβ state for the action γ if ψ(x γiβ (y)) = ψ(yx) for all x, y ∈ A a dense γ-invariant ∗-subalgebra of Aγ, the subalgebra of analytic elements for the action γ. We refer to [BR1, Section 2.5] for basic information on the subalgebra of analytic elements, Aγ and to [BR2, Section 5.3] for all the basic information on KMS states. Since Gλ is discrete it is well-known that the map 0 (R)) → C λ 0 (R) extends uniquely to a faithful conditional expectation E : Aλ → C λ defined (norm) lower semicontinuous weight on C λ (norm) lower semicontinuous weight on Aλ which we denote by ¯ψ. In particular, for x ∈ l1 we have: 0 (R). Composing E with the densely 0 (R) given by integration, gives us a densely defined 0 (R)) x 7→ x(1) : l1 α(Gλ, C λ α(Gλ, C λ ¯ψ(xx∗) =ZR xx∗(1)(t)dt =ZRXh∈Gλ x(h)x(h) (t)dt = Xh∈Gλ(cid:18)ZR x(h)(t)2dt(cid:19) . So that ¯ψ is faithful. We observe that ¯ψ is not a trace, since ¯ψ(x∗x) =Ph∈Gλ h−1RR x(h)(t)2dt. Proposition 2.28. The weight ¯ψ on Aλ restricts to a faithful semifinite trace ¯τ on Aλ restricts to a state denoted by ψ on Qλ satisfying: (1) The gauge action γ of T1 on Qλ leaves the state ψ invariant. (2) The state ψ restricted to the fixed point algebra, F λ is a faithful (finite) trace denoted by τ ; which is, of course, the restriction of ¯τ on Aλ (3) With Φ : Qλ → F λ the canonical expectation, we have ψ = τ ◦ Φ. Proof. Since ¯ψ(e) =RR X[0,1)(t)dt = 1, we see that ¯ψ restricted to Qλ is a faithful state. To see item 0 (R)) leaves ¯ψ invariant. To this end, let x ≥ 0 (1), it suffices to see that the gauge action on l1 be in l1 0 and also α(Gλ, C λ 0 to F λ. α(Gλ, C λ 0 (R)), and let z ∈ T1. Then and so E(γz(x)) = γz(x)(1) = z0x(1) = x(1) = E(x) ¯ψ(γz(x)) =ZR γz(x)(1)(t)dt =ZR E(x)(t)dt = ¯ψ(x). To see item (2) we use Proposition 2.25 and the above computation that shows that while ¯ψ is not generally a trace, to see that it is a trace when the group elements all have determinant 1. To see item (3), it suffices to see that for any x ∈ Qλ we have E(x) = E(Φ(x)), but this is the same as x(1) = Φ(x)(1) which is clear since det(1) = 1. (cid:3) Now, since the state ψ is invariant under the action γ, this action is unitarily implemented on L2(Qλ, ψ). For z ∈ T1 and x ∈ Qλ c we define: (uz(x))h = znxh for h ∈ Gλ with h = λn. 19 We define the spectral subspaces of this unitary group on L2(Qλ, ψ) in the usual way. For each k ∈ Z let Φk be the operator on L2(Qλ, ψ) : 2πZT1 z−kuz(x)dθ, z = eiθ, x ∈ L2(Qλ, ψ). We observe that if x = f · δg is a typical generator of Qλ considered as a vector in L2(Qλ, ψ) then we have: Φk(x) = 1 Φk(f · δg) =(cid:26) f · δg 0 if g = λk otherwise More generally, on H := L2(Qλ, ψ), we have Φk(H) = {x ∈ H uz(x) = zkx for all z ∈ T1}. Lemma 2.29. For each k ∈ Z the subspace span{f · δg ∈ Qλ g = λk} is dense in the range of Φk. The operators Φk are mutually orthogonal projections on H which sum to the identity operator 1 = π(e). Proof. The proof of the first statement is similar to the proof of Proposition 2.25. The mutual orthog- onality of the Φk follows from the fact that hf1 · δg1f2 · δg2iψ = 0 unless g1 = g2. (cid:3) Proposition 2.30. The dense ∗-subalgebra of Qλ consisting of finite linear combinations of the partial isometries X[a,b) · δg is contained in the subset of entire elements, Qλ γ , for the action γ considered as an action of R : t 7→ γeit. Moreover, ψ is a KMSβ state for this action where β = log(λ−1). In fact, ψ is the unique KMS state for this action (regardless of β). Proof. Let y = X[a,b) · δg ∈ Qλ where det(g) = λk. Then, t 7→ γeit(y) = eikty; t ∈ R obviously extends to the entire function w 7→ γeiw (y) = eikwy; w ∈ C. For w = log(λ−1)i, this equation becomes γeiw (y) = γλ(y) = λky. Letting β = log(λ−1), we have γβi(y) = λky. Now, let x = X[c,d) · h so we want to see that: λkψ(xy) = ψ(xγβi(y)) = ψ(yx). That is, we want λkψ(xy) = ψ(yx). Now both sides of this equation are zero unless h = g−1. But, when h = g−1, we have xy = X[c,d) · X[g−1(a),g−1(b)) · δI while yx = X[a,b) · X[g(c),g(d)) · δI . Moreover, s ∈ [c, d) ∩ [g−1(a), g−1(b)) ⇐⇒ g(s) ∈ [g(c), g(d)) ∩ [a, b). Since det(g) = λk the transformation g increases the measure by a factor of λk and the result follows. That is, ψ is a KMSβ state for the action γ of R for β = log(λ−1). Now let φ be a KMS state on Qλ for the action γ. Since Qλ is purely infinite it has no nontrivial traces and so φ must be KMS for some nonzero β. Hence by [BR2, Proposition 5.3.3], φ is invariant under the action of γ. Now, if X[a,b) · δg ∈ Qλ with det(g) = λk, then we have for all z ∈ T: φ(X[a,b) · δg) = φ(γz(X[a,b) · δg)) = zkφ(X[a,b) · δg). That is, if det(g) 6= 1 we must have φ(X[a,b) · δg) = 0, and so φ is supported on F λ. Since F λ is γ-invariant and φ is KMS for some nonzero β, φ is a trace on F λ by [BR2, 5.3.28]. Now, if x = X[a,b) · δg ∈ F λ and g 6= I, then we claim that φ(x) = 0. For suppose g = [1 : c] with c > 0. Then there is a positive integer n such that a + nc < b ≤ a + (n + 1)c and so x = X[a,b) · δg = X[a,c) · δg + X[a+c,a+2c) · δg + ··· + X[a+nc,b) · δg := v0 + v1 + ··· + vn. Now each of these partial isometries vk satisfies v2 φ is a trace on F λ. Thus, φ(x) = 0 as claimed. k = 0, and so φ(vk) = φ(vkv∗kvk) = φ(v2 kv∗k) = 0 since 20 Hence φ is supported on the commutative subalgebra C := span{f · δI f ∈ C λ 0 (R) and supp(f ) ⊆ [0, 1)}. 0 ∼= F λ ⊗ K we can define a lower semicontinuous, densely defined trace, T r on Aλ Morever, if f1, f2 are characteristic functions of subintervals of [0, 1) with endpoints in Γλ and having the same length they give equivalent elements fi · δI in F λ and therefore have the same value under φ. Now, since Aλ 0 via T r = φ ⊗ T r, where T r is the trace on K. So, for X[a,b) · δI ∈ F λ we have T r(X[a,b) · δI ) = φ(X[a,b) · δI ). Then, for k1 < k2 ∈ Z the element X[k1,k2) · δI is the sum of (k2 − k1) projections in Aλ 0 each equivalent to X[0,1) · δI which has trace equal to 1; that is, T r(X[k1,k2) · δI ) = (k2 − k1). Now, for any a < b in Γλ, we have X[a,b) · δI ∼ X[0,b−a) · δI and so T r(X[a,b) · δI ) = T r(X[0,b−a) · δI ), and these values are finite since both these projections are dominated by X[−N,N ) · δI for a sufficiently large integer N. It now suffices to prove the following. Claim: T r(X[a,b) · δI ) = b − a for a < b ∈ Γλ. By the previous discussion we can assume that a = 0 so that b > 0. Given ε > 0 we choose positive integers m, n such that 1 m ≤ ε and n − 1 m ≤ b < n m , so that (n − 1) ≤ bm < n and (n − 1), bm, n ∈ Γλ. Hence (n − 1) = T r(X[0,(n−1)) · δI ) ≤ T r(X[0,bm) · δI ) ≤ T r(X[0,n) · δI ) = n. But, and these projections are mutually equivalent in Aλ X[0,bm) · δI = X[0,b) · δI + X[b,2b) · δI + ··· + X[(m−1)b,bm) · δI (n − 1) ≤ m T r(X[0,b) · δI ) ≤ n so that 0 . That is, n − 1 m ≤ T r(X[0,b) · δI ) ≤ n m . m ≤ ε. That is, b = T r(X[0,b) · δI ) = φ(X[0,b) · δI ) and φ agrees with the Hence, T r(X[0,b) · δI ) − b < 1 given trace τ on F λ and therefore φ agrees with ψ on Qλ. Remarks. The above proof shows that the algebra F λ has a unique (faithful) tracial state τ, and that Aλ 0 has a unique (faithful) lower semicontinuous, densely defined trace normalized so that it has value 1 at e = X[0,1) · δI . 2.7. The von Neumann algebra π(Aλ)−wo acting on L2(Aλ, ¯ψ) is a type IIIλ factor. To prove this we will show that it is unitarily equivalent to a version of the Murray-von Neumann "group- measure space" construction of type III factors on l2(Gλ) ⊗ L2(R) : see [D, Chapter 1, Section 9]. We conclude that it is a IIIλ factor by an appeal to Connes' thesis [C0]. In order to be consistent with our use of right C∗-modules later, we will do our GNS constructions so that our inner products are linear in the second variable. Proposition 2.31. The ∗-algebra Aλ c is a Tomita algebra with the inner product: (cid:3) hyxi ¯ψ = ¯ψ(y∗x) = Xh∈Gλ h−1hxhyhiL2(R). Here we denote xh in place of x(h) to simplify notation. In this setting we have for x ∈ Aλ c : (1) Sharp: S(x)h = αh(xh−1); (2) Flat: F (x)h = hαh(xh−1); (3) Delta: ∆(x)h = hxh. 21 Proof. We refer to [Ta] for Takesaki's version of the axioms for a Tomita algebra. Since Sharp is defined to be the adjoint operation on the algebra, item (1) is immediate. A straightforward calculation shows that for all x, y ∈ Aλ c we have that the defining equation for F lat holds, namely:hS(y)xi ¯ψ = hF (x)yi ¯ψ so that item (2) holds. By definition, ∆ = F S and so a simple calculation shows that ∆(x)h = hxh and (3) holds. From this formula for ∆ we see that for each z ∈ C we have ∆z(x)h = hzxh and a straightforward calculation shows that ∆z(x · y) = (∆z(x)) · (∆z(y)) so that each ∆z is an algebra homomorphism of Aλ c as required. That each left multiplication π(x) is bounded when x is supported on a single group element is straightforward and the generalization to finitely supported elements is then trivial. The fact that it is a ∗-representation holds as it does for the GNS representation for any weight. In order to see that products are dense we recall that we have local units. That is, for each positive integer N we have defined EN = X[−N,N ) · δ1, and have noted that for each y ∈ Aλ c that satisfies supp(yh) ⊆ [−N, N ) for all h, we have EN · y = y. Axioms IV, V, VI in [Ta] are simple calculations involving the definitions of S, F , and ∆ which we leave to the reader. Since our inner products are linear in the second variable, we modify Tomita's Axiom VIII to c . We easily calculate that hx∆z(y)i ¯ψ = read: z 7→ hx∆x(y)i ¯ψ is analytic on C for all x, y ∈ Aλ Ph hz−1hyhxhiL2(R). This function is analytic since the sum is finite. c , ¯ψ) decomposes as the integrated form of a covariant (cid:3) Lemma 2.32. The representation of Aλ pair of representations: c on L2(Aλ 00(R) → B(L2(Aλ (1) π : C λ (2) U : Gλ → U (L2(Aλ c , ¯ψ)), where : (π(f )(y))h = f · yh for f ∈ C λ 00(R) and y ∈ Aλ c ; c , ¯ψ)) where : (Ug(y))h = αg(yg−1h) for g ∈ Gλ and y ∈ Aλ c . Proof. It is straightforward to verify that U is a unitary representation of Gλ and that π is a ∗- representation of C λ 00(R). To see the covariance condition: (Ug π(f )Ug−1(y))h = ··· = αg(f · αg−1(ygg−1)) = αg(f ) · yh = (π(αg(f ))y)h. That is, Ug π(f )Ug−1 = π(αg(f )). Now, by Proposition 7.6.4 of [Ped] the integrated form of this covariant pair is the representation: (π × U )(y) =Xh Now, we evaluate this operator on the vector x ∈ Aλ c : [π(yh)Uh(x)]k =Xh yhαh(xh−1k) = (y · x)(k) = (π(y)(x))k. [((π × U )(y))(x)]k =Xh π(yh)Uh for y ∈ Aλ c . That is, (π × U )(y) = π(y) the operator left multiplication by y. (cid:3) 2.7.1. A representation of Aλ on l2(Gλ) ⊗ L2(R). We define a covariant pair of representations of 0 (R) and Gλ on l2(Gλ) ⊗ L2(R) as follows: C λ (1) for f ∈ C λ 0 (R) let π(f ) = 1 ⊗ Mf , and (2) for g ∈ Gλ let U g = Λ(g) ⊗ Vg where Λ is the left regular representation of Gλ on l2(Gλ) : (Λ(g)ξ)(h) = ξ(g−1h) for ξ ∈ l2(Gλ); and V is the unitary action of Gλ on L2(R) induced by the action of Gλ on R : (Vg(f ))(t) = g−1/2f (g−1t); for f ∈ L2(R). 22 Using these equations one easily checks the covariance condition for g ∈ Gλ and f ∈ C λ 0 (R) : U gπ(f )U∗g = π(αg(f )). Clearly the representation π extends uniquely by weak-operator continuity to the usual representation 1 ⊗ M of L∞(R) on l2(Gλ) ⊗ L2(R) and is covariant with the unitary representation U of Gλ for the action α of Gλ on L∞(R). Clearly, the von Neumann algebra on l2(Gλ) ⊗ L2(R) generated by the unitaries U g and the operators 1 ⊗ Mf for g ∈ Gλ and f ∈ C λ 00(R), is the same as the von Neumann algebra generated by the unitaries U g and the operators 1 ⊗ Mf for g ∈ Gλ and f ∈ L∞(R). The second item of the following Proposition is clear. Proposition 2.33. (1) The representation π = (π × U ) of Aλ on L2(Aλ the representation (π × U ) of Aλ on l2(Gλ) ⊗ L2(R). (2) (π × U )(Aλ)′′ is the von Neumann crossed product (in the sense of the group-measure space con- struction of Chapter 1 Section 9 of [D]) Gλ ⋊α L∞(R). (3) This von Neumann algebra is a type III factor. c , ¯ψ) is unitarily equivalent to Proof. To see item (3) we use the proof of [D, Theorem 2, Section 9, Chapter 1] where instead of the ax + b group G with a, b ∈ Q and a > 0 and its subgroup G0 (with a = 1), we use Gλ and its subgroup G0 λ (with g = 1), to conclude that our von Neumann algebra is a type III factor. To see item (1), we first define a unitary W : L2(Aλ fi · δhi! = c , ¯ψ) → l2(Gλ) ⊗ L2(R) as follows: mXi=1 W mXi=1 hi−1/2δhi ⊗ fi. On the left side of this equation we are using the formalism f · δh for singly supported elements in Aλ with f ∈ C λ 00(R) and h ∈ Gλ. On the right of this equation we are using δh to denote the canonical orthonormal basis elements in l2(Gλ) and regarding f ∈ C λ 00(R) ⊂ L2(R). Clearly, W is well-defined and linear with dense range. One easily checks that: for all x, y ∈ Aλ c we have c recalling that the inner product on Aλ inverse (adjoint) defined at first on the elements in l2(Gλ) ⊗ L2(R) of the formPm fi ∈ C λ 00(R), is given by: hyxi ¯ψ = hW (x)W (y)il2⊗L2 c is linear in the second coordinate. Thus W is a unitary and its i=1 δhi ⊗ fi with the δhi ⊗ fi! = One then verifies the following two equations for f ∈ C λ W ∗ mXi=1 mXi=1 hi1/2fi · δhi. 00(R) and g ∈ Gλ : (1) W π(f )W ∗ = 1 ⊗ Mf = π(f ) and (2) W UgW ∗ = Λ(g) ⊗ Vg = U g. The second equation is more subtle and requires the observation: Ug(f · δh) = g1/2Vg(f ) · δgh. This completes the proof of the proposition. (cid:3) c , ¯ψ) is type IIIλ. We work in the unitarily equivalent setting 2.7.2. The factor π(Aλ)′′ acting on L2(Aλ of (π× U )(Aλ)′′ acting on l2(Gλ)⊗L2(R) afforded by Proposition 2.33. Recall that the subgroup of Gλ of matrices of the form [λn : 0] is isomorphic to Z and acts on the normal subgroup G0 λ by conjugacy, and so Gλ = Z ⋊ G0 λ is a semidirect product and we can write a canonical right coset decomposition of Gλ : Gλ = [n∈Z λ · [λn : 0]. G0 23 This gives us an internal orthogonal decomposition of l2(Gλ) : l2(Gλ) =Xn∈Z ⊕ l2(cid:0)G0 Here the latter isomorphism is given explicitly on basis elements by the map which takes the δ-function at g · [λn : 0] to δn ⊗ δg for n ∈ Z and g ∈ G0 λ. One checks that the restriction of the representation (π× U) of Aλ = Gλ ⋊ C λ on l2(Gλ)⊗L2(R) is unitarily equivalent to the representation on l2(Z)⊗l2(G0 pair: λ · [λn : 0](cid:1) ∼= l2(Z) ⊗ l2(G0 0 (R) to Aλ 0 (R) λ)⊗L2(R) via the covariant λ). 0 := G0 λ ⋊ C λ 1Z ⊗ Λ(h) ⊗ Vh = 1Z ⊗ U h for h ∈ G0 1Z ⊗ 1 ⊗ Mf = 1Z ⊗ π(f ) for f ∈ C λ λ and 0 (R). Therefore, the von Neumann subalgebra of (π × U )(Aλ)′′ generated by (π × U )(Aλ the von Neumann algebra on l2(G0 1⊗Mf for f ∈ C λ Λ(h) ⊗ Vh for h ∈ G0 II∞ by the methods of [D, Chapter 1, Section 9]. Thus, (π × U )(Aλ type III factor, (π × U )(Aλ)′′. Moreover, the faithful normal semifinite trace on (π × U )(Aλ by the restriction of ¯ψ. 0 ) is isomorphic to λ and 0 (R). This is clearly the same as the von Neumann algebra generated by the operators λ and 1 ⊗ Mf for f ∈ L∞(R), and this von Neumann algebra is a factor of type 0 )′′ is a type II∞ subfactor of the 0 )′′ is given λ) ⊗ L2(R) generated by the operators Λ(h) ⊗ Vh for h ∈ G0 Finally, conjugation by the unitary, U g for g = [λ : 0], which lies in our type III factor, defines an automorphism β of the type II∞ subfactor which scales the trace by λ. If N0 is our type II∞ factor λ)⊗L2(R) is unitarily acting on l2(G0 equivalent to the von Neumann crossed product Aλ ∼= Z ⋊β N0 and hence is a type IIIλ factor by [C0, λ)⊗L2(R) then our type III factor, say Aλ acting on l2(Z)⊗ l2(G0 Theorem 4.4.1]. We have proved the following Proposition. c , ¯ψ) is a type IIIλ factor. Proposition 2.34. The von Neumann algebra π(Aλ)′′ acting on L2(Aλ Moreover, it is unitarily equivalent to (π × U )(Aλ)′′ acting on l2(Gλ) ⊗ L2(R). The von Neumann subalgebra of (π × U )(Aλ)′′ generated by (π × U )(Aλ 0 ) is a type II∞ factor. The space l2(Gλ) ⊗ L2(R) λ)⊗L2(R) and with this factorization, our II∞ factor has the form N0 = 1Z⊗ N0 factors as l2(Z)⊗l2(G0 where N0 acts on l2(G0 λ)⊗L2(R). Thus, our type IIIλ factor is unitarily equivalent to the von Neumann crossed product Z ⋊β N0 where the automorphism β of N0 is given by β = Ad(U g) where g = [λ : 0]. 2.8. The von Neumann algebra, π0(Qλ)−wo acting on L2(Qλ, ψ) is type IIIλ. Theorem 2.35. The von Neumann algebra, π0(Qλ)−wo acting on L2(Qλ, ψ) is type IIIλ. Moreover, the von Neumann subalgebra, π0(F λ)−wo is a type II1 factor with unique faithful normal state given by the restriction of the vector state, ψ which is the same as τ on F λ. By the general theory of type III factors, π0(Qλ)−wo is isomorphic to π(Aλ)−wo acting on L2(Aλ Proof. Recall that Qλ = eAλe where e = X[0,1) · δ1 ∈ Aλ. Then c , ¯ψ). π(e)(π(Aλ)−wo)π(e) = (π(e)π(Aλ)π(e))−wo = π(Qλ)−wo and the cut-down of the type III factor π(Aλ)−wo (on its separable Hilbert space) by the nonzero projection π(e) is isomorphic to π(Aλ)−wo since π(e) is Murray-von Neumann equivalent to the identity operator. Of course the cut-down mapping by π(e) is not an isomorphism. Moreover, by left Hilbert algebra theory, the operator right multiplication by e which is denoted by π′(e) is in the commutant of π(Aλ)−wo acting on L2(Qλ, ¯ψ) and since we are in a factor the mapping π(Aλ)−wo → π′(e)π(Aλ)−wo is an isomorphism by [D, Chapter 1, Section 2, Prop. 2]. Restricting this isomorphism to π(Qλ)−wo gives us an isomorphism π(Qλ)−wo → π′(e)π(Qλ)−wo which acts on the 24 Hilbert space π′(e)π(e)(L2(Aλ, ¯ψ)), which has as a dense subspace π′(e)π(e)(Aλ c e ⊂ eAλe = Qλ with the inner product given by ¯ψ which is the same as the inner product on eAλ c e given by the state ψ. The completion of this space is, of course, L2(Qλ, ψ) with the action of Qλ being the GNS represen- tation afforded by the state ψ. We denote this representation of Qλ on L2(Qλ, ψ) by π0 to distinguish it from the representation π of Aλ on the larger space, L2(Aλ Similar considerations applied to the type II∞ subfactor, π(Aλ that: c , ¯ψ). 0 )−wo ⊂ π(Aλ)−wo on L2(Aλ c , ¯ψ), show c ) = eAλ π(e)(π(Aλ 0 )−wo)π(e) = (π(e)π(Aλ 0 )π(e))−wo = π(F λ)−wo. 0 )−wo of π(Aλ)−wo and has finite (ψ) Now the projection π(e) is actually in the type II∞ subfactor π(Aλ trace = 1 there. Therefore, π(F λ)−wo is a type II1 factor on L2(Qλ, ψ) with trace given by the vector state ψ. We remark that this is clearly a larger space than the subspace, L2(F λ, τ ) ⊂ L2(Qλ, ψ). (cid:3) c is a Tomita algebra with the inner product: hyxiψ = ψ(y∗x). Proposition 2.36. The ∗-algebra Qλ Again we denote xh in place of x(h) to simplify notation. In this setting we have for x ∈ Qλ c : (1) Sharp: S(x)h = αh(xh−1); (2) Flat: F (x)h = hαh(xh−1); (3) Delta: ∆(x)h = hxh. Proof. This is really a corollary of Proposition 2.7, as Qλ c is just a Tomita-subalgebra of Aλ c . (cid:3) 3. The modular spectral triple of the algebra Qλ Having introduced the main features of the algebras Qλ, we now turn briefly to the modular index theory of [CNNR, CPR2, CRT]. We begin with some semifinite preliminaries. 3.1. Semifinite noncommutative geometry. We need to explain some semifinite versions of stan- dard definitions and results following [CPRS2]. Let φ be a fixed faithful, normal, semifinite trace on a von Neumann algebra N . Let KN be the φ-compact operators in N (that is the norm closed ideal generated by the projections E ∈ N with φ(E) < ∞). Definition 3.1. A semifinite spectral triple (A,H,D) is given by a Hilbert space H, a ∗-algebra A ⊂ N where N is a semifinite von Neumann algebra acting on H, and a densely defined unbounded self-adjoint operator D affiliated to N such that [D, a] is densely defined and extends to a bounded operator in N for all a ∈ A and (λ − D)−1 ∈ KN for all λ 6∈ R. The triple is said to be even if there is Γ ∈ N such that Γ∗ = Γ, Γ2 = 1, aΓ = Γa for all a ∈ A and DΓ + ΓD = 0. Otherwise it is odd. Note that if T ∈ N and [D, T ] is bounded, then [D, T ] ∈ N . We recall from [FK] that if S ∈ N , the t-th generalized singular value of S for each real t > 0 is given by µt(S) = inf{kSEk : E is a projection in N with φ(1 − E) ≤ t}. The ideal L1(N , φ) consists of those operators T ∈ N such that kTk1 := φ(T) < ∞ where T = √T ∗T . In the Type I setting this is the usual trace class ideal. We will denote the norm on L1(N , φ) by k · k1. An alternative definition in terms of singular values is that T ∈ L1(N , φ) if kTk1 := R ∞0 µt(T )dt < ∞. When N 6= B(H), L1(N , φ) need not be complete in this norm but it is complete in the norm k · k1 + k · k∞. (where k · k∞ is the uniform norm). We use the notation L(1,∞)(N , φ) =(cid:26)T ∈ N : kTkL(1,∞) := sup t>0 1 log(1 + t)Z t 0 µs(T )ds < ∞(cid:27) . 25 measurable operators affiliated to N . Our notation is however consistent with that of [C] in the special case N = B(H). With this convention the ideal of φ-compact operators, K(N ), consists of The reader should note that L(1,∞)(N , φ) is often taken to mean an ideal in the algebra eN of φ- those T ∈ N (as opposed to eN ) such that µ∞(T ) := limt→∞ µt(T ) = 0. Definition 3.2. A semifinite spectral triple (A,H,D) relative to (N , φ) with A unital is (1,∞)- summable if (D − λ)−1 ∈ L(1,∞)(N , φ) for all λ ∈ C \ R. It follows that if (A,H,D) is (1,∞)-summable then it is n-summable (with respect to the trace φ) for all n > 1. We next need to briefly discuss Dixmier traces. For more information on semifinite Dixmier traces, see [CPS2, CRSS]. For T ∈ L(1,∞)(N , φ), T ≥ 0, the function FT : t → is bounded. There are certain ω ∈ L∞(R+ by setting 1 log(1 + t)Z t µs(T )ds 0 ∗ )∗, [CPS2, C], which define (Dixmier) traces on L(1,∞)(N , φ) φω(T ) = ω(FT ), T ≥ 0 and extending to all of L(1,∞)(N , φ) by linearity. For each such ω we write φω for the associated Dixmier trace. Each Dixmier trace φω vanishes on the ideal of trace class operators. Whenever the function FT has a limit at infinity, all Dixmier traces return that limit as their value. This leads to the notion of a measurable operator [C, LSS], that is, one on which all Dixmier traces take the same value. 3.2. The Kasparov module and modular spectral triple. We have seen that the algebras Qλ do not possess a faithful gauge invariant trace but that there is a KMSβ where β = − log(λ) for the gauge action, γ, namely ψ := τ ◦ Φ : Qλ → C, where Φ : Qλ → F λ is the expectation and τ : F λ → C In fact, ψ is the only KMS state for the gauge action (for any β), is a faithful normalised trace. by Proposition 2.30. We show below that the generator of the gauge action D acting on a suitable C∗-F λ-module X gives us a Kasparov module (X,D) whose class lies in KK 1,T(Qλ, F λ). In some examples, including the case λ ∈ Q, we have K1(Qλ) = {0} and so pairing with ordinary K1 would be fruitless. However, following [CPR2, CNNR] we may compute a numerical pairing using a 'modular spectral triple' constructed from the Kasparov module. We now review this construction adapted to the present situation. Let H = L2(Qλ) be the GNS Hilbert space given by the faithful state ψ with the inner product on Qλ defined by ha, bi = ψ(a∗b) = (τ ◦ Φ)(a∗b). Then D is a self-adjoint unbounded operator on H, [CPR2]. The representation of Qλ on H by left multiplication (which we now denote by π in place of π0) is bounded and nondegenerate: the left action of an element a ∈ Qλ by π(a) satisfies π(a)b = ab for all b ∈ Qλ. This distinction between elements of Qλ as vectors in L2(Qλ) and operators on L2(Qλ) is sometimes crucial. The dense subalgebra Qλ c e which is the finite span of elements in Qλ of the form X[a,b) · δg is in the smooth domain of the derivation δ = ad(D). We remind the reader that the KMS condition on the modular automorphism group of the state ψ, [Ta], (for t = i) is: ψ(xy) = ψ(σi(π(y))x) = ψ(σ(y)x) for x, y ∈ π(Qλ), where σ(y) = ∆−1(y). Lemma 3.3. The group of modular automorphisms of the von Neumann algebra π(Qλ)′′ is given on the generators by c := eAλ (2) σt(π(f · δg)) := ∆itπ(f · δg)∆−it = π(∆it(f · δg)) = gitπ(f · δg) = det(g)itπ(f · δg). Proof. This is immediate from Lemma 2.8 if we note that g = det(g). (cid:3) 26 Corollary 3.4. With Qλ acting on H := L2(Qλ) and with D the generator of the natural unitary implementation of the gauge action of T1 on Qλ, we have ∆ = λD or eitD = ∆it/ log λ. To simplify notation, we let A = Qλ and F = F λ = Aγ, the fixed point algebra for the T1 gauge action, γ. For convenience we will suppress the notations D ⊗ 1k and so on. The algebras Ac, Fc are defined as the finite linear span of the generators. Right multiplication makes A into a right F -module, and similarly Ac is a right module over Fc. We define an F -valued inner product (··)R on both these modules by (ab)R := Φ(a∗b). Definition 3.5. Let X be the right F C∗-module obtained by completing A (or Ac) in the norm kxk2 X := k(xx)RkF = kΦ(x∗x)kF . The algebra A acting by left multiplication on X provides a representation of A as adjointable operators on X. Let Xc be the copy of Ac ⊂ X. The T1 action on Xc is unitary and extends to X, [CNNR, PR]. For all k ∈ Z, the projection operator onto the k-th spectral subspace of the T1 action is also denoted (somewhat carelessly) Φk on X: Φk(x) = z−kuz(x)dθ, z = eiθ, x ∈ X. Observe that Φ0 restricts to Φ on A and on generators of Qλ we have (3) if g = λk otherwise 1 2πZT1 Φk(f · δg) =(cid:26) f · δg 0 Of course L2(Qλ) and X have a common dense subspace Qλ c on which these projections are identical. Let Ak = Φk(A) and observe from (3) that A∗kAk = F = AkA∗k so that the gauge action γ on Qλ has full spectral subspaces. We quote the following result from [PR], the proof in our case is the same. Lemma 3.6. The operators Φk are adjointable endomorphisms of the F -module X such that Φ∗k = Φk = Φ2 k and ΦkΦl = δk,lΦk. If K ⊂ Z then the sumPk∈K Φk converges strictly to a projection in the endomorphism algebra. The sum Pk∈Z Φk converges to the identity operator on X. For all x ∈ X, the sum x =Pk∈Z Φkx =Pk∈Z xk converges in X. The unbounded operator of the next proposition is of course the generator of the T1 action on X. We refer to Lance's book, [L, Chapters 9,10], for information on unbounded operators on C∗-modules. Proposition 3.7. [PR] Let X be the right C∗-F -module of Definition 3.5. Define D : XD ⊂ X to be the linear space XD = {x =Xk∈Z xk ∈ X : kXk∈Z k2(xkxk)Rk < ∞}. For x ∈ XD define D(x) =Pk∈Z kxk. Then D : XD → X is a is self-adjoint, regular operator on X. This should be compared to the following Hilbert space version. Proposition 3.8. The generator D of the one-parameter unitary group {uz z ∈ T1} on L2(Qλ, ψ) has eigenspaces given by the ranges of the Φk and D(x) = kx iff Φk(x) = x. In particular dom(D) = {x =Xk xk Φk(xk) = xk and Xk k2kxkk2 < ∞}, and D(Pk xk) =Pk kxk. 27 X = kΦ(x∗x)kF λ. Introduce the rank one operators on X : ΘR Remark. On generators in Qλ regarded as elements of either X or L2(Qλ, ψ) we have D(f · δg) = (logλ(g))f · δg. To continue, we recall the underlying right C∗-F λ-module, X, which is the completion of Qλ for the norm kxk2 x,yz = x(yz)R. Then using the operators Sk,m defined above, we obtain formulas for the projections Φk similar to those of [PR, Lemma 4.7] with some important differences. First recall [CPR2, Lemma 3.5]. Lemma 3.9. Any F λ-linear endomorphism T of the module X which preserves the copy of Qλ inside X, extends uniquely to a bounded operator on the Hilbert space H = L2(Qλ). In particular, the finite rank endomorphisms of the pre-C∗ module Qλ condition, and we denote the algebra of all these endomorphisms by End00 Lemma 3.10. Compare [PR, Lemma 4.7]. The following formulas hold in both L(X) and in B(H). (1) For k ≥ 0, we have c (acting on the left) satisfy this F (Qλ c ). x,y by ΘR Φ0 = ΘR e,e while for k > 0, Φk = ΘR Sk,m,Sk,m. mkXm=0 (2) For −k < 0, we have k,m,S ∗ S ∗ Φ−k = ΘR k,m for any m = 0, 1, ..., mk − 1 and also for mk if λ−k = mk + 1. Proof. Since both Φk and the finite rank endomorphisms satisfy the hypotheses of the previous lemma, the first statement of this lemma will follow from calculations done on generators. The following calculations are based on the formulas in Lemma 2.15. (1) Let k > 0 and let x =Pl xl be a finite sum of generators, xl satisfying Φl(xl) = xl. Then mkXm=0 Sk,m,Sk,m(xl) =Xl Sk,mΦ(S∗k,mxl) = mkXm=0 mkXm=0 Sk,mΦ(S∗k,mxk) ΘR ΘR Sk,mS∗k,mxk = exk = xk = Φk(x). mkXm=0 Sk,m,Sk,m(x) = Xl mkXm=0 = For k = 0 this is a similar but far easier calculation. (2) Let −k < 0 and let x =Pl xl be a finite sum of generators as above. Then, for 0 ≤ m < mk S∗k,mΦ(Sk,mxl) = S∗k,mΦ(Sk,mx−k) ΘR k,m,S ∗ S ∗ ΘR k,m,S ∗ S ∗ k,m k,m (x) = Xl (xl) =Xl = S∗k,mSk,mx−k = ex−k = x−k = Φ−k(x). (cid:3) We recall the following result discussed in Section 3 of [CNNR] (a 'bare hands' proof can be given by the method in [CPR2]). c ))′′, where we take the commu- Proposition 3.11. Let N be the von Neumann algebra N = (End00 tant inside B(H). Then N is semifinite, and there exists a faithful, semifinite, normal trace τ : N → C such that for all rank one endomorphisms ΘR F (Qλ x,y of Qλ c , x,y) = (τ ◦ Φ)(y∗x), τ (ΘR x, y ∈ Qλ c . In addition, D is affiliated to N and π(Qλ) is a subalgebra of N . 28 The fact that τ (Φk) = λ−k implies that with respect to the trace τ we can not expect D to satisfy a finite summability criterion. We solve this problem exactly as in [CPR2]. Definition 3.12. We define a new weight on N +: let T ∈ N + then τ∆(T ) := supN τ (∆N T ) where ∆N = ∆(Pk≤N Φk). Remarks. Since ∆N is τ -trace-class, we see that T 7→ τ (∆N T ) is a normal positive linear functional on N and hence τ∆ is a normal weight on N + which is easily seen to be faithful and semifinite. As in [CPR2], we now give another way to define τ∆ which is not only conceptually useful but also makes a number of important properties straightforward to verify. Many proofs require only trivial notation changes and the substitution of n± with λ∓. Notation. Let M be the relative commutant in N of the operator ∆. Equivalently, M is the relative commutant of the set of spectral projections {Φkk ∈ Z} of D. Clearly, M =Pk∈Z ΦkN Φk. Definition 3.13. As τ restricted to each ΦkN Φk is a faithful finite trace with τ (Φk) = λ−k we define bτk on ΦkN Φk to be λk times the restriction of τ . Then,bτ :=Pkbτk on M =Pk∈Z ΦkN Φk is a faithful normal semifinite trace bτ with bτ (Φk) = 1 for all k. We usebτ to give an alternative expression for τ∆ below M. The map Ψ : N → M defined by Ψ(T ) =Pk ΦkT Φk is a conditional expectation onto M and τ∆(T ) =bτ (Ψ(T )) for all T ∈ N +. That is, τ∆ =bτ ◦ Ψ so that bτ (T ) = τ∆(T ) for all T ∈ M+. Finally, if one of A, B ∈ M is bτ -trace-class and T ∈ N then τ∆(AT B) = τ∆(AΨ(T )B) =bτ (AΨ(T )B). Lemma 3.14. An element m ∈ N is in M if and only if it is in the fixed point algebra of the action, στ∆ t (T ) = ∆itT ∆−it. Both π(F λ) and the projections Φk belong to t Proof. The proof is the same as the proof of [CPR2, Lemma 3.9] with λk in place of n−k. Lemma 3.15. The modular automorphism group στ∆ The weight τ∆ is a KMS weight for the group στ∆ t t of τ∆ is inner and given by στ∆ , and στ∆ on N defined for T ∈ N by στ∆ t (T ) = ∆itT ∆−it. (cid:3) t Qλ = στ◦Φ t . Proof. This follows from: [KR, Thm 9.2.38], which gives us the KMS properties of τ∆: the modular group is inner since ∆ is affiliated to N . The final statement about the restriction of the modular group to Qλ is clear. We now have the key lemma: Lemma 3.16. Suppose g is a function on R such that g(D) is τ∆ trace-class in M, then for all f ∈ F λ we have (cid:3) τ∆(π(f )g(D)) = τ∆(g(D))τ (f ) = τ (f )Xk∈Z g(k). Proof. First note that τ∆(g(D)) =bτ (Pk∈Z g(k)Φk) =Pk∈Z g(k)bτ (Φk) =Pk∈Z g(k). We first do the c so that all the sums are finite. Now, computation for f ∈ F λ g(k)bτ (π(f )Φk) g(k)λk τ (π(f )Φk). τ∆(π(f )g(D)) =bτ (π(f )Xk∈Z =Xk∈Z g(k)bτk(π(f )Φk) =Xk∈Z g(k)Φk) =Xk∈Z So it suffices to see for each k ∈ Z, we have τ (π(f )Φk) = λ−kτ (f ). Now, by Theorem 2.35 π(F λ)′′ is a type II1 factor on H whose unique trace say T r (with norm one) extends the trace τ on F λ in the sense that T r(π(f )) = τ (f ). Since the projection Φk is in the commutant of the factor π(F λ)′′ the map T ∈ π(F λ)′′ 7→ T Φk = ΦkT Φk is a normal isomorphism by [D, Chapter 1, section 2, Prop. 2] and so it has a unique normalised trace also given by T race(T Φk) = T r(T ). But τ (T Φk) is a trace on Φkπ(F λ)′′Φk = π(F λ)′′Φk and so must be τ (Φk) = λ−k times the unique norm one trace. That is, we get the required formula: τ (π(f )Φk) = λ−kT race(π(f )Φk) = λ−kT r(π(f )) = λ−kτ (f ). 29 So for f ∈ F λ c , we have the formula: τ∆(π(f )g(D)) = τ∆(g(D))τ (f ) =Xk∈Z g(k)τ (f ). Now, the right hand side is a norm-continuous function of f . To see that the left side is norm-continuous we do it in more generality. Let T ∈ N , then sincebτ is a trace on M we get: τ∆(T g(D)) = bτ (Ψ(T g(D)) = bτ (Ψ(T )g(D)) ≤ kΨ(T )kbτ (g(D)) ≤ kTkbτ ((g(D)) = kTkτ∆(g(D)). That is the left hand side is norm-continuous in T and so we have the formula: τ∆(π(f )g(D)) = τ∆(g(D))τ (f ) =Xk∈Z g(k)τ (f ) 1 2 for all f ∈ F λ. Proposition 3.17. (i) We have (1 + D2)−1/2 ∈ L(1,∞)(M, τ∆). That is, τ∆((1 + D2)−s/2) < ∞ for all s > 1. Moreover, for all f ∈ F λ lim s→1+ (s − 1)τ∆(π(f )(1 + D2)−s/2) = 2τ (f ) so that π(f )(1 + D2)−1/2 is a measurable operator in the sense of [C]. (ii) For π(a) ∈ π(Qλ) ⊂ N the following (ordinary) limit exists and (cid:3) bτω(π(a)) = (s − 1)τ∆(π(a)(1 + D2)−s/2) = τ ◦ Φ(a), lim s→1+ the original KMS state ψ = τ ◦ Φ on Qλ. Proof. (i) This proof is identical to [CPR2, Proposition 3.12]. (ii) This proof is the same as [CPR2, Proposition 3.14] with Qλ, F λ replacing On, F. Definition 3.18. The triple (A,H,D) along with γ, ψ, N , τ∆ satisfying properties (0) to (3) below is called the modular spectral triple of the dynamical system (Qλ, γ, ψ) c of the algebra Qλ is faithfully represented in N with the latter acting on 0) The ∗-subalgebra A = Qλ the Hilbert space H = L2(Qλ, ψ), 1) there is a faithful normal semifinite weight τ∆ on N such that the modular automorphism group of τ∆ is an inner automorphism group σt (for t ∈ C) of (the Tomita algebra of ) N with σiA = σ in the sense that σi(π(a)) = π(σ(a)), where σ is the automorphism σ(a) = ∆−1(a) on A, 2) τ∆ restricts to a faithful semifinite trace bτ on M = N σ, with a faithful normal projection Ψ : N → M satisfying τ∆ =bτ ◦ Ψ on N , 3) with D the generator of the one parameter group implementing the gauge action of T on H we [D, π(a)] extends to a bounded operator (in N ) for all a ∈ A and for λ in the resolvent set of have: D, (λ − D)−1 ∈ K(M, τ∆), where K(M, τ∆) is the ideal of compact operators in M relative to τ∆. In particular, D is affiliated to M. (cid:3) 30 For matrix algebras A = Qλ triple in the obvious fashion. c ⊗ Mk over Qλ c , (Qλ c ⊗ Mk,H ⊗ Mk,D ⊗ Idk) is also a modular spectral We need some technical lemmas for the discussion in the next Section. A function f from a complex domain Ω into a Banach space X is called holomorphic if it is complex differentiable in norm on Ω. The following is proved in [CPR2, Lemma 3.15]. Lemma 3.19. (1) Let B be a C∗-algebra and let T ∈ B+. The mapping z 7→ T z is holomorphic (in operator norm) in the half-plane Re(z) > 0. (2) Let B be a von Neumann algebra with faithful normal semifinite trace φ and let T ∈ B+ be in L(1,∞)(B, φ). Then, the mapping z 7→ T z is holomorphic (in trace norm) in the half-plane Re(z) > 1. (3) Let B, and T be as in item (2) and let A ∈ B then the mapping z 7→ φ(AT z) is holomorphic for Re(z) > 1. Lemma 3.20. In these modular spectral triples (A,H,D) for matrices over the algebras Qλ we have (1+D2)−s/2 ∈ L1(M, τ∆) for all s > 1 and for x ∈ N , τ∆(x(1+D2)−r/2) is holomorphic for Re(r) > 1 and we have for a ∈ Qλ Proof. We include a brief proof since there are some small but important differences from [CPR2, Lemma 3.16]. Since the eigenvalues for D are precisely the set of integers, and the projection Φk on the eigenspace with eigenvalue k satisfies τ∆(Φk) = 1, it is clear that (1 + D2)−s/2 ∈ L1(M, τ∆). Now, lemma. To see the last statement, we observe that τ∆([D, π(a)](1 + D2)−r/2) = τ∆(Ψ([D, π(a)])(1 + D2)−r/2), so it suffices to see that Ψ([D, π(a)]) = 0 for a ∈ A = Qλ c . To this end, let a = f · δg where det(g) = λn is one of the linear generators of Qλ c . Then by calculating the action of the operator Dπ(f · δg) on the linear generators fi · δhi of the Hilbert space, H, we obtain: τ∆(x(1 + D2)−r/2) = bτ (Ψ(x)(1 + D2)−r/2) is holomorphic for Re(r) > 1 by item (3) of the previous c , τ∆([D, π(a)](1 + D2)−r/2) = 0, for Re(r) > 1. Dπ(f · δg) = nπ(f · δg) + π(f · δg)D that is [D, π(f · δg)] = logλ(g)π(f · δg). More generally, [D, π( cifi · δhi)] = ci(logλ(hi))π(fi · δhi). mXi=1 mXi=1 c ) for all a ∈ Qλ If we apply Ψ to this equation, we see that Ψ(π(fi · δhi)) = π(Φ(fi · δhi)) = 0 whenever logλ(hi) 6= 0, and so the whole sum is 0. We also observe that [D, π(a)] ∈ π(Qλ c . This is not too surprising since D is the generator of the action γ of T on Qλ. (cid:3) 3.3. Modular K1. We now make appropriate modifications to [CPR2, Section 4]) using [CNNR] introducing elements of these modular spectral triples (A,H,D) (where A is a matrix algebra over Qλ we say that a unitary (invertible, projection,...) in the n × n matrices over Qλ for some n is a unitary (invertible, projection,...) over Qλ. We write σt for the automorphism σt ⊗ Idn of A. Definition 3.21. Let v be a partial isometry in the ∗-algebra A. We say that v satisfies the modular condition with respect to σ if the operators vσt(v∗) and v∗σt(v) are in the fixed point algebra F ⊂ A for all t ∈ R. Of course, any partial isometry in F is a modular partial isometry. Lemma 3.22. ([CPR2, Lemma 4.8]) Let v ∈ A be a modular partial isometry. Then we have c ) that will have a well defined pairing with our Dixmier functionalbτω. Let A = Qλ. Following [HR] is a modular unitary over A. Moreover there is a modular homotopy uv ∼ uv∗ . uv =(cid:18) 1 − v∗v v v∗ 1 − vv∗ (cid:19) Note that in [CPR2] we used a different approach which is implied by the one given here. In [CPR2] we defined modular unitaries in terms of the regular automorphism: π(σ(a)) = π(∆−1(a)) = ∆−1π(a)∆ = σi(π(a)). 31 That is we said that a unitary in A is modular if uσ(u∗) and u∗σ(u) are in the fixed point algebra. Examples. (1) For k, j > 0 recall Sk,m ∈ Qλ X[mλk,(m+1)λk ) · δ1 which is in clearly F λ. Then for each {k, m},{j, n} we have a unitary c with m < mk (see Definition 2.14) we write Pk,m = Sk,mS∗k,m = u{k,m},{j,n} =(cid:18) 1 − Pk,m Sk,mS∗j,n Sj,nS∗k,m 1 − Pj,n (cid:19) . It is simple to check that this a self-adjoint unitary satisfying the modular condition, and that τ (Pk,m) = λk and τ (Pj,n) = λj. These examples behave very much like the SµS∗ν examples of [CPR2]. (2) For k, j > 0 consider the "leftover" partial isometries Sk,mk and Sj,mj of Definition 3.13 which we will denote by Sk and Sj to lighten the notation. We let vj,k = SjS∗k and calculate its range and initial projections which are both in F λ: Pj = SjS∗kSkS∗j = X[mj λj ,mjλj +λj(λ−k−mk)) · δ1, and Pk = SkS∗j SjS∗k = X[mkλk,mkλk+λk(λ−j−mj )) · δ1. We note for future reference that: We also note that we have a modular unitary uj,k: τ (Pj ) = λj(λ−k − mk) and τ (Pk) = λk(λ−j − mj). uj,k =(cid:18) 1 − Pk SjS∗k SkS∗j 1 − Pj (cid:19) . Define the modular K1 group as follows. Definition 3.23. Let K1(A, σ) be the abelian group with one generator [v] for each partial isometry v over A satisfying the modular condition and with the following relations: 1) 2) 3) [v] = 0 if v is over F, [v] + [w] = [v ⊕ w], if vt, t ∈ [0, 1], is a continuous path of modular partial isometries in some matrix algebra over A then [v0] = [v1]. One could use modular unitaries as in [CPR2] in place of these modular partial isometries. The following can now be seen to hold. Lemma 3.24. (Compare [CPR2, Lemma 4.9]) Let (A,H,D) be our modular spectral triple relative to (N , τ∆) and set F = Aσ and σ : A → A. Let L∞(∆) = L∞(D) be the von Neumann algebra generated by the spectral projections of ∆ then L∞(∆) ⊂ Z(M). Let v ∈ A be a partial isometry with vv∗, v∗v ∈ F . Then π(v)Qπ(v∗) ∈ M and π(v∗)Qπ(v) ∈ M for all spectral projections Q of D, if and only if v is modular. That is, π(v)∆π(v∗) and π(v∗)∆π(v) (or π(v)Dπ(v∗) and π(v∗)Dπ(v)) are both affiliated to M if and only if v is modular. 32 Thus we see that modular partial isometries conjugate ∆ to an operator affiliated to M, and so v∆v∗ commutes with ∆ (and vDv∗ commutes with D). We will next show that there is an analytic pairing between (part of) modular K1 and modular spectral triples. To do this, we are going to use the analytic formulae for spectral flow in [CP2]. 3.4. The mapping cone algebra. Our aim in the remainder is to calculate an index pairing explicitly c of Qλ. In the following few pages we will for the matrix algebras A over the smooth subalgebra Qλ sometimes abuse notation and write a in place of π(a) for a ∈ A in order to make our formulae more readable. Whenever we do this, however, we will use σi(·) = ∆−1(·)∆ the spatial version of the algebra homomorphism, σ. We will generally use the spatial version σi when in the presence of operators not in π(A). We briefly review some results from [CNNR], that provide an interpretation of the modular index pairing given by the spectral flow. If F ⊂ A is a sub-C∗-algebra of the C∗-algebra A, then the mapping cone algebra for the inclusion is M (F, A) = {f : R+ = [0,∞) → A : f is continuous and vanishes at infinity, f (0) ∈ F}. When F is an ideal in A it is known that K0(M (F, A)) ∼= K0(A/F ), [Put1]. In general, K0(M (F, A)) is the set of homotopy classes of partial isometries v ∈ Mk(A) with range and source projections vv∗, v∗v in Mk(F ), with operation the direct sum and inverse −[v] = [v∗]. All this is proved in [Put1]. It is shown in [CNNR] that there is a natural map that injects K1(A, σ) into K T 0 (M, F ), the equivariant K-theory of the mapping cone algebra. Note that the T action on A lifts in the obvious way to the mapping cone. Now, it was shown in [CPR1] that certain Kasparov A, F -modules extend to Kasparov M (F, A), F -modules, and this was extended to the equivariant case in [CNNR]. Importantly the theory applies to the equivariant Kasparov module coming from a circle action. The extension is explicit, namely there is a pair ( X, D) which is a graded unbounded Kasparov module for the mapping cone algebra M (F, A) constructed using a generalised APS construction, [APS3]. If v is a partial isometry in Mk(A), setting ev(t) = 1 − vv∗ iv∗ 1+t2 t 1+t2 −iv t 1+t2 ! , 1+t2 v∗v defines ev as a projection over M (F, A). Then in [CNNR] we showed that if v ∈ A is a modular partial isometry we have h[ev] −(cid:20)(cid:18) 1 0 0 0 (cid:19)(cid:21) , [( X, D)]i = −Index(P vP : v∗vP (X) → vv∗P (X)) ∈ K0(F ) = Index(P v∗P : vv∗P (X) → v∗vP (X)) ∈ K T 0 (F ). (4) We thus obtain an index map K1(A, σ) → K T 0 (F ). The latter may be thought of as the ring of Laurent polynomials K0(F )(χ, χ−1) where we think of χ, χ−1 as generating the representation ring of T. We may obtain a real valued invariant from this map by evaluating χ at e−β where β is the inverse temperature of our KMS state and applying the trace to the resultant element of K0(F ). Then one of the main results of [CNNR] is that the real valued invariant so obtained is identical with the spectral flow invariant of the next subsection. However the general theory of [CNNR] does not tell us the range of this index map and it is the latter that is of interest for these explicit calculations. 33 3.5. A local index formula for the algebras Qλ. Using the fact that we have full spectral subspaces we know from [CNNR] that there is a formula for spectral flow which is analogous to the local index formula in noncommutative geometry. We remind the reader that τ∆ =bτ ◦ Ψ where Ψ : N → M is the canonical expectation, so that τ∆ restricted to M isbτ . Theorem 3.25. (Compare [CPR2, Theorem 5.5]) Let (A,H,D) be the (1,∞)-summable, modular spectral triple for the algebra Qλ we have constructed previously. Then for any modular partial isometry v and for any Dixmier trace bτω associated to bτ , we have spectral flow as an actual limit sfbτ (vv∗D, vDv∗) = The functional on A⊗A defined by a0⊗ a1 7→ 1 b, B-cocycle (see the proof below for the definition). 2bτω(v[D, v∗](1 +D2)−1/2) = τ ◦ Φ(v[D, v∗]). (s− 1)bτ (v[D, v∗](1 +D2)−s/2) = 2 lims→1+(s− 1)τ∆(a0[D, a1](1 +D2)−s/2) is a σ-twisted lim s→1+ 1 2 1 Remark. Spectral flow in this setting is independent of the path joining the endpoints of unbounded self adjoint operators affiliated to M however it is not obvious that this is enough to show that it is constant on homotopy classes of modular unitaries. This latter fact is true but the proof is lengthy so we refer to [CNNR]. Theorem 3.26. We let (Qλ (1) Let u be a modular unitary defined in Section 5 of the form c ⊗ M2). c ⊗ M2,H ⊗ C2,D ⊗ 12) be the modular spectral triple of (Qλ u{k,m},{j,n} =(cid:18) 1 − Pk,m Sk,mS∗j,n Sj,nS∗k,m 1 − Pj,n (cid:19) . Then the spectral flow is positive being given by (2) Let u be a modular unitary defined in Section 5 of the form: sfτ∆(D, uDu∗) = (k − j)(λj − λk) ∈ Z[λ] ⊂ Γλ. uj,k =(cid:18) 1 − Pk SjS∗k SkS∗j 1 − Pj (cid:19) , where SkS∗j = Sk,mkS∗j,mj spectral flow is given by and Pk and Pj are its range and initial projections, respectively. Then the sfτ∆(D, uDu∗) = (k − j)[λj(λ−k − mk) − λk(λ−j − mj)] ∈ Γλ. Proof. We have already observed that these are, in fact modular unitaries. For the computations we use a calculation from the proof of Lemma 3.20 to get in example (1): u[D ⊗ 12, u] =(cid:18) 1 − Pk,m Sk,mS∗j,n = (cid:18) 1 − Pk,m Sk,mS∗j,n Sj,nS∗k,m 1 − Pj,n (cid:19)(cid:18) So using Theorem 3.25 and our previous computation of the Dixmier trace, Proposition 3.17, and the fact that Pk,m = Sk,mS∗k,m = X[mλk,(m+1)λk) · δ1 so that τ (Pk,m) = λk we have Sj,nS∗k,m 1 − Pj,n (cid:19)(cid:18) (j − k)Sj,nS∗k,m (cid:19) = (k − j)(cid:18) −Pk,m (k − j)Sk,mS∗j,n Pj,n (cid:19) . [D, Sk,mS∗j,n] [D, Sj,nS∗k,m] (cid:19) 0 0 0 0 0 0 sfτ∆(D, uk,mDuk,m) = (k − j)τ (Pj,n − Pk,m) = (k − j)(λj − λk). This number is always positive as the reader may check, and is contained in Z[λ]. The computations in example (2) are similar and use the fact that Pk = X[mkλk,mkλk+λk(λ−j−mj )) · δ1, so that τ (Pk) = λk(λ−j − mj) ∈ Γλ. In these examples, the spectral flow is not contained in the smaller polynomial ring, Z[λ]. (cid:3) 34 Remarks. The observation of [CPR2] that the twisted residue cocycle formula for spectral flow is calculating Araki's relative entropy of two KMS states [Ar] also applies to the examples in this subsection. Acknowledgements We would like to thank Nigel Higson, Ryszard Nest, Sergey Neshveyev, Marcelo Laca, Iain Raeburn and Peter Dukes for advice and comments. The first and fourth named authors were supported by the Australian Research Council. The second and third named authors acknowledge the support of NSERC (Canada). References [Ar] H. Araki, Relative entropy of states of von Neumann algebras, Publ. RIMS, Kyoto Univ., 11 (1976) 809 -- 833 and Relative entropy for states of von Neumann algebras II, Publ. RIMS, Kyoto Univ., 13 (1977) 173 -- 192. [APS3] M.F. Atiyah, V.K. Patodi, I.M. Singer, Spectral asymmetry and Riemannian geometry. III, Math. Proc. Camb. [BR1] [BR2] Phil. Soc. 79 (1976) 71 -- 99. O. Bratteli, D. Robinson, Operator algebras and quantum statistical mechanics 1, Springer-Verlag, 2nd Ed, 1987. O. Bratteli, D. Robinson, Operator algebras and quantum statistical mechanics 2, Springer-Verlag, 2nd Ed, 1987. [CNNR] A.L. Carey, R. Nest, S. Neshveyev, A. Rennie, Twisted cyclic theory, equivariant KK-theory and KMS states, [CP1] [CP2] to appear in J. reine angew. Math. A. L. Carey, J. Phillips, Unbounded Fredholm modules and spectral flow, Canadian J. Math., 50 (4) (1998) 673 -- 718. A. L. Carey, J. Phillips, Spectral flow in θ-summable Fredholm modules, eta invariants and the JLO cocycle, K-Theory, 31 (2004) 135 -- 194. [CPR1] A.L. Carey, J. Phillips, A. Rennie, A noncommutative Atiyah-Patodi-Singer index theorem in KK-theory, to appear in J. reine angew. Math. [CPR2] A.L. Carey, J. Phillips, A. Rennie, Twisted cyclic theory and an index theory for the gauge invariant KMS state on Cuntz algebras, to appear in Journal of K-Theory. A.L. Carey, J. Phillips, F. Sukochev, Spectral flow and Dixmier traces, Adv. Math, 173 (2003) 68 -- 113. [CPS2] [CPRS2] A.L. Carey, J. Phillips, A. Rennie, F. Sukochev, The local index formula in semifinite von Neumann algebras I: Spectral Flow, Adv. in Math. 202 (2006) 451 -- 516. [C0] [C] [Cu] [Cu1] [D] [E] [K] [CRSS] A. L. Carey, A. Rennie, A. Sedaev, F. Sukochev, Dixmier traces and asymptotics of zeta functions, somewhere A.L. Carey, A. Rennie, K. Tong, Spectral flow invariants and twisted cyclic theory from the Haar state on [CRT] SUq(2), to appear A. Connes, Une classification des facteurs de type III, Annales Scientifiques de le'Ecole Norm. Sup., 4em serie t. 6 (1973) 18 -- 252. A. Connes, Noncommutative geometry, Academic Press, 1994. J. Cuntz, Simple C ∗-algebras generated by isometries, Commun. Math. Phys, 57 (1977) 173 -- 189. J. Cuntz, C ∗-algebras associated with the ax + b-semigroup over N inK-Theory and Noncommutative Geometry, Eds, G. Cortinas, J. Cuntz, M Karoubi, R. Nest, C.A. Weibel, EMS Series of Congress Reports, Volume 2. J. Dixmier, von Neumann algebras, North-Holland, 1981. G. Elliott, Some simple C ∗-algebras constructed as crossed products with discrete outer automorphism groups, Publ. RIMS Kyoto Univ., 16 (1980) 299-311. T. Fack and H. Kosaki, Generalised s-numbers of τ -measurable operators, Pac. J. Math., 123 (1986) 269 -- 300. N. Higson, J. Roe, Analytic K-homology, Oxford University Press, 2000. J. Kaad, R. Nest, A. Rennie, KK-theory and spectral flow in von Neumann algebras, arXive:math.OA/0701326. R.V. Kadison, J. R. Ringrose, Fundamentals of the theory of operator algebras, Vol. II: advanced theory, Academic Press, 1986. G. G. Kasparov, The operator K-functor and extensions of C ∗-algebras, Math. USSR. Izv. 16 No. 3 (1981) 513 -- 572. J. Kustermans, G. Murphy, L. Tuset, Differential calculi over quantum groups and twisted cyclic cocycles, J. Geom. Phys., 44 (2003) 570 -- 594. E. C. Lance, Hilbert C ∗-modules, Cambridge University Press, Cambridge, 1995. M. Laca, J. Spielberg, Purely infinite C ∗-algebras from boundary actions of discrete groups, Journal fur die reine und ang. Mathematik, 480 (1996) 125 -- 139. [FK] [HR] [KNR] [KR] [L] [LS] [KMT] 35 [LSS] [PR] [PhR] [Ped] [PT] [Put1] [Put2] [RS] [Sc] [Ta] [T] S. Lord, A. Sedaev, F. A. Sukochev, Dixmier traces as singular symmetric functionals and applications to measurable operators, Journal of Functional Analysis, 224 no.1 (2005) 72 -- 106. D. Pask, A. Rennie, The noncommutative geometry of graph C ∗-algebras I: The index theorem, Journal of Functional Analysis, 233 (2006) 92 -- 134. J. Phillips, I. Raeburn, Semigroups of isometries, Toeplitz algebras and twisted crossed products, J. Int. Equat. and Op. Th., 17 (1993) 579-602. G. K. Pedersen, C ∗-algebras and their automorphism groups, London Math. Soc. monographs 14, Academic Press, London 1979. G. K. Pedersen, M. Takesaki, The Radon-Nikodym theorem for von Neumann algebras, Acta Math., 130 (1973) 53 -- 87. I. Putnam, An excision theorem for the K-theory of C ∗-algebras, J. Operator Theory, 38 (1997) 151 -- 171. I. Putnam, On the K-theory of C ∗-algebras of principal groupoids, Rocky Mountain J. Math., 28 no. 4 (1998) 1483 -- 1518. M. Rørdam and E. Størmer, Classification of nuclear C ∗-algebras. Entropy in operator algebras, Encyclopedia of Mathematical Sciences, 126 (2002), Springer, Berlin. C. Schochet, Topological methods for C ∗-algebras II: geometric resolutions and the Kunneth formula, Pac. J. Math., 98 No. 2 (1982) 443 -- 458. M. Takesaki, Tomita's theory of modular Hilbert algebras and its applications, Lecture Notes in Mathematics, 128 (1970), Springer, Berlin. J. Tomiyama, On the projection of norm one in W ∗-algebras, Proc. Japan Acad., 33 (1957) 608 -- 612. E-mail [email protected] address: [email protected], [email protected], [email protected],
1708.07496	3	1708	2019-02-25T13:48:59	Non-classification of free Araki-Woods factors and $\tau$-invariants	[ "math.OA", "math.LO" ]	We define the standard Borel space of free Araki-Woods factors and prove that their isomorphism relation is not classifiable by countable structures. We also prove that equality of $\tau$-topologies, arising as invariants of type III factors, as well as coycle and outer conjugacy of actions of abelian groups on free product factors are not classifiable by countable structures.	math.OA	math	Non-classification of free Araki-Woods factors and τ -invariants by Rom´an Sasyk1, Asger Tornquist2 and Stefaan Vaes3 Abstract We define the standard Borel space of free Araki-Woods factors and prove that their iso- morphism relation is not classifiable by countable structures. We also prove that equality of τ -topologies, arising as invariants of type III factors, as well as cocycle and outer conju- gacy of actions of abelian groups on free product factors are not classifiable by countable structures. 1 Introduction In his celebrated work, Connes [Co75] proved that all amenable factors acting on a separable Hilbert space are hyperfinite and deduced that there is a unique amenable factor for each of the types In with n ∈ N, II1, II∞ and IIIλ with λ ∈ (0, 1). The uniqueness of the hyperfinite III1 factor was proved in Haagerup's fundamental paper [Ha85]. In [Co75, Co74b], Connes also proved that hyperfinite type III0 factors are isomorphic to Krieger factors. So, using [Kr74, CT76], they are completely classified by their flow of weights, which is a properly ergodic action of R on a standard measure space. This last result can equally well be interpreted as a non-classification theorem. Indeed, properly ergodic flows up to conjugacy cannot be classified in any concrete way, and descriptive set theory provides a framework to give a precise meaning to the complexity of such a classification problem. In particular, using Hjorth's concept of turbulence [Hj97], it follows from [FW03] that properly ergodic flows and therefore amenable type III0 factors cannot be classified by countable structures like countable groups or other combinatorial invariants. Even for the subclass of infinite tensor products of matrix algebras, such a non-classification result holds, see [ST09]. For nonamenable factors, complete classification theorems for large families of such factors were obtained in Popa's deformation/rigidity theory [Po06]. In particular, his work allows to "embed" unclassifiable structures (like, for instance, probability measure preserving transfor- mations) into the theory of II1 factors and prove that II1 factors are not classifiable by countable structures, see [ST08]. Among the most natural and most studied nonamenable factors are the free Araki-Woods factors of Shlyakhtenko [Sh96], associated with an orthogonal representation (Ut)t∈R of R on a real Hilbert space HR of dimension at least 2. This construction generalizes Voiculescu's free Gaussian functor. When Ut is the trivial representation, the associated free Araki-Woods factor is of type II1 and isomorphic to the free group factor L(Fn) with n = dimR(HR). When Ut is non trivial, one obtains a full factor of type III that may be viewed as a free probability analog of the classical Araki-Woods factors. The almost periodic free Araki-Woods factors, associated with almost periodic representations Ut, were completely classified by Shlyakhtenko in [Sh96] in terms of Connes' Sd-invariant, which 1Universidad de Buenos Aires, Departamento de Matem´atica and Instituto Argentino de Matem´aticas- CONICET, Buenos Aires (Argentina). E-mail: [email protected]. Supported in part by research grant PICT 2012-1292 (ANPCyT). 2University of Copenhagen, Department of Mathematics, Copenhagen (Denmark). E-mail: [email protected]. Supported in part by the DNRF Niels Bohr Professorship of Lars Hesselholt, and by DFFResearch Project 7014-00145B. 3KU Leuven, Department of Mathematics, Leuven (Belgium). E-mail: [email protected]. Supported by European Research Council Consolidator Grant 614195 RIGIDITY, and by long term structural funding -- Methusalem grant of the Flemish Government. 1 is a countable subgroup of R+ ∗ . While the complete classification of arbitrary free Araki-Woods factors remains wide open, a first result in that direction, classifying a large family of non almost periodic free Araki-Woods factors, was obtained in [HSV16]. As the first main result of our paper, we turn the space of free Araki-Woods factors into a standard Borel space and reinterpret one of the examples in [HSV16] as providing a Borel family Mµ of free Araki-Woods factors indexed by arbitrary non-atomic probability measures on the Cantor set C such that Mµ ∼= Mη if and only if the measures µ and η are mutually absolutely continuous; see Theorem 2.2. In combination with [KS99], it follows that free Araki-Woods factors are not classifiable by countable structures. The finest modular theory invariant for full type III factors is Connes τ -invariant [Co74a], defined as the weakest topology on R that makes the canonical modular homomorphism R → Aut(M )/ Inn(M ) continuous. In [Sh02], Shlyakhtenko proved that there is a continuum of possible τ -topologies and used this to construct a continuum of non almost periodic free Araki- Woods factors. As our second main result, we prove in Theorem 3.1 that τ -topologies are not even classifiable by countable structures. This provides another proof for the non-classification of free Araki-Woods factors, but also applies to other classification questions where a τ -invariant makes sense. As an example, we prove in Theorem 4.2 that cocycle or outer conjugacy of actions of Z, R, or any other non compact, abelian, locally compact group G on a free group factor L(Fn), or any other free product factor M1 ∗ M2, are not classifiable by countable structures. Finally note that a similar non-classification theorem for conjugacy of a single automorphism (i.e. an action of Z) was obtained in [KLP08] for the hyperfinite II1 factor R and in [KLP14] for free product factors, using a spectral invariant. Such spectral invariants are however not preserved under cocycle conjugacy. Moreover, by [Oc85], all outer actions of a discrete amenable group G on R are cocycle conjugate, and this leads to a complete classification of actions of G on R up to cocycle conjugacy. Acknowledgment. Research for this paper was mainly carried out when R. Sasyk was at the Hausdorff Research Institute for Mathematics in Bonn during the Trimester program on "von Neumann Algebras". He wishes to thank the Hausdorff Institute for kind hospitality and support. S. Vaes wants to thank D. Shlyakhtenko for discussions and suggestions that helped to improve this article. 2 Non-classification of free Araki-Woods factors Whenever K is a separable Hilbert space, we denote by vN(K) the set of all von Neumann subalgebras of B(K) and we equip vN(K) with the Effros Borel structure defined in [Ef64] as the smallest σ-algebra such that for every ϕ ∈ B(K)∗, the map vN(K) → R : M 7→ kϕM k is measurable. We start by defining a Borel subset FAW ⊂ vN(K) that contains, up to isomorphism, all free Araki-Woods factors. The correct interpretation of [HSV16, Example 5.4] then provides a Borel map Θ : Probc(C) → FAW from the space of continuous probability measures on the Cantor set C to the space of free Araki-Woods factors such that Θ(µ) ∼= Θ(µ′) if and only µ ≈ µ′, where ≈ denotes the equivalence relation of mutual absolute continuity on Prob(K). So, Θ is a Borel reduction from (Probc(K), ≈) to (FAW, ∼=) and it follows from [KS99, Theorem 2.1] that free Araki-Woods factors are not classifiable by countable structures. To define the Borel set FAW, fix a separable, infinite dimensional Hilbert space H. Define the full Fock space K = F(H) given by F(H) = CΩ ⊕ ∞Mn=1(cid:0)H ⊗ · · · ⊗ H(cid:1) } {z n times 2 . (2.1) We denote by Ba(H) the space of antilinear bounded operators from H to H and we equip both Ba(H) and B(H) with the standard Borel structure induced by the strong topology. We define the following parameter space for FAW ⊂ vN(K). D =(cid:8)(J, T ) ∈ Ba(H) × B(H)(cid:12)(cid:12)J = J ∗, J 2 = 1, T = T ∗, 0 ≤ T ≤ 1, T + JT J = 1, T and 1 − T have trivial kernel(cid:9) . Since the operators with trivial kernel form a Borel subset of B(H), it is easy to see that D is a Borel subset of Ba(H) × B(H). The formula S = J(T −1 − 1)1/2 defines a bijection between the set D and the set of all densely defined, closed, antilinear operators S on H satisfying S2 ⊂ 1, with the inverse being determined by the polar decomposition of S. To every such involution S is associated the free Araki-Woods factor acting on K given by (2.2) Φ(J, T ) =(cid:8)ℓ(ξ) + ℓ(S(ξ))∗(cid:12)(cid:12) ξ ∈ D(S)(cid:9)′′ =(cid:8)ℓ(T 1/2ξ) + ℓ(J(1 − T )1/2ξ)∗(cid:12)(cid:12) ξ ∈ H(cid:9) . It follows that Φ : D → vN(K) is a Borel map. Since the graph of S can be recovered from M = Φ(J, T ) as the closure of the set (2.3) it follows that Φ is injective. (cid:8)(xΩ, x∗Ω)(cid:12)(cid:12) x ∈ M, xΩ ∈ H, x∗Ω ∈ H(cid:9) , Definition 2.1. For K = F(H), we define the Borel set FAW ⊂ vN(K) as the image of the set D defined in (2.2) under the injective Borel map Φ defined in (2.3). By construction, every free Araki-Woods factor is isomorphic with an element of FAW. We can now deduce from [HSV16, Example 5.4] the following result. Theorem 2.2. Let C be the Cantor set. There exists a Borel map Θ : Probc(C) → FAW such that Θ(µ) ∼= Θ(µ′) if and only if µ ≈ µ′, where ≈ denotes the equivalence relation of mutual absolute continuity. The map Θ can be chosen such that all Θ(µ) have as τ -invariant the usual topology on R. In particular, the equivalence relation (FAW, ∼=) is not classifiable by countable structures. Proof. Denote by λ the Lebesgue measure on the interval [0, 1]. Fix the separable Hilbert space H = L2([0, 1], λ) ⊕ L2([0, 1], λ) ⊕ C2 . Denote by e0, e1 ∈ C2 the two standard basis vectors. Define the anti-unitary involutions J0 : C2 → C2 and J : H → H given by J0(e1) = e2 , J0(e2) = e1 and J(ξ1 ⊕ ξ2 ⊕ ξ3) = ξ2 ⊕ ξ1 ⊕ J0(ξ3) . Fix any q ∈ (0, 1). Using the functions ϕµ : [0, 1] → [0, 1] introduced in Lemma 2.3 below, we define for every probability measure µ ∈ Prob([0, 1]), the operator Tµ ∈ L∞([0, 1])⊕L∞([0, 1])⊕ M2(C) ⊂ B(H), given by Tµ =(cid:0)1 + exp(ϕµ)(cid:1)−1 ⊕(cid:0)1 + exp(−ϕµ)(cid:1)−1 ⊕(cid:18)(1 + q)−1 0 0 (1 + q−1)−1(cid:19) . By construction, (J, Tµ) ∈ D and by Lemma 2.3, the map µ 7→ (J, Tµ) is Borel. We define the Borel map Θ1 : Prob([0, 1]) → FAW : Θ1(µ) = Φ(J, Tµ) . 3 Denoting by δa the Dirac measure in a ∈ R, we associate to any nonatomic probability measure µ on [0, 1] the following probability measures on R : the opposite measure µop given by µop(U ) = eµ = in [HSV16]. 1 4(cid:0)µ + µop + δlog q + δ− log q(cid:1) . µ(−U ) and the symmetric probability measure eµ given by Araki-Woods factor Φ(J, Tµ) is isomorphic with the free Araki-Woods factor denoted by Γ(eµ, 1) Using the unitary V : L2([0, 1], µ) → L2([0, 1], λ) given by Lemma 2.3, it follows that the free By [Ru62, Theorems 5.1.4 and 5.2.2] (see also [Ke95, 19.1 and 19.2]), we can choose a copy of the Cantor set C ⊂ [0, 1] such that C is independent as a subset of R. Viewing Probc(C) as a subset of Probc([0, 1]), we define Θ : Probc(C) → FAW : Θ(µ) = Θ1((µ + λ)/2) . By [HSV16, Example 5.4], we get that for all µ, µ′ ∈ Probc(C), the von Neumann algebras Θ(µ) and Θ(µ′) are isomorphic if and only if the measures µ and µ′ are mutually absolutely continuous, and that the τ -invariant of every Θ(µ) equals the usual topology on R. By [KS99, Theorem 2.1], the equivalence relation ≈ of mutual absolute continuity on Probc(C) is not classifiable by countable structures. A fortiori, the equivalence relation ∼= on FAW is not classifiable by countable structures. The following lemma is probably well known. It is for instance essentially contained in [Ke95, Proof of Theorem 17.41]. For completeness, we provide a short proof. Lemma 2.3. For every probability measure µ ∈ Prob([0, 1]), define the increasing function ϕµ : [0, 1] → [0, 1] : ϕµ(y) = min{x ∈ [0, 1] µ([0, x]) ≥ y} . Equip Prob([0, 1]) with the weak topology and denote by λ the Lebesgue measure on [0, 1]. Then, the following holds. 1. Each ϕµ is a Borel function. 2. For every µ ∈ Prob([0, 1]), we have that (ϕµ)∗(λ) = µ. 3. The map Prob([0, 1]) → L1([0, 1], λ) : µ 7→ ϕµ is continuous. 4. If µ has no atoms, the map ϕµ is strictly increasing and V : L2([0, 1], µ) → L2([0, 1], λ) : (V ξ)(y) = ξ(ϕµ(y)) is unitary. Proof. 1. Since ϕµ is increasing, it is a Borel function. 2. For every a ∈ [0, 1], we have that ϕ−1 coincide on [0, a] for all a ∈ [0, 1]. This implies that (ϕµ)∗(λ) = µ. 3. A direct computation gives that for all µ, η ∈ Prob([0, 1]), we have µ ([0, a]) = [0, µ([0, a])]. Therefore, (ϕµ)∗(λ) and µ kϕµ − ϕηk1 =Z 1 0 (cid:12)(cid:12)µ([0, a]) − η([0, a])(cid:12)(cid:12) da . It follows that Prob([0, 1]) → L1([0, 1]) : µ 7→ ϕµ is continuous. 4. When µ has no atoms, the function x 7→ µ([0, x]) is continuous and thus attains all values in [0, 1]. It follows that ϕµ is strictly increasing. Defining Rµ = ϕµ([0, 1]), we have that µ([0, 1] \ Rµ) = 0 and ϕµ : [0, 1] → Rµ is a bijection sending λ to µ. It follows that V is a well defined unitary operator. 4 3 Non-classification of τ -topologies Recall that a factor M is called full when Inn(M ) ⊂ Aut(M ) is closed. Then, Out(M ) has a natural Polish group structure. In [Co74a, Definition 5.1], Connes defined the invariant τ (M ) as the weakest topology on R that makes the (canonical) modular homomorphism R → Out(M ) : t 7→ σϕ t continuous. Given a unitary representation π : R → U (H) on a separable Hilbert space, in [Co74a, Theorem 5.2], a full factor M is constructed such that τ (M ) equals the weakest topology on R that makes π continuous. Let G be a second countable, locally compact, abelian group. Since unitary representations of G on a separable Hilbert space are classified by their spectral measure and multiplicity function, it is natural to define as follows the τ -topology τ (µ) for any Borel probability measure µ on and denote by τ (µ) the weakest topology on G that makes the map g 7→ πµ(g) continuous the Pontryagin dual bG. Consider the unitary representation πµ : G → U (L2(bG, µ)) : (πµ(g)ξ)(ω) = ω(g) ξ(ω) for all g ∈ G, ξ ∈ L2(bG, µ), ω ∈ bG (3.1) from G to the unitary group equipped with the strong topology. We denote by bµ the Fourier bµ : G → C :bµ(g) =Z bG transform of µ defined as ω(g) dµ(ω) . Given a Polish space K, we always equip Prob(K) with the weak topology, i.e. the weakest Note that gn → e in the topology τ (µ) if and only if bµ(gn) → 1. topology making the maps µ 7→ RK F dµ continuous for all bounded continuous functions F ∈ Cb(K). Finally, denote by S∞ the Polish group of all permutations of N. Recall that an equivalence relation E on a Polish space X is said to be generically S∞-ergodic if for any continuous action of S∞ on a Polish space Y , any Baire measurable morphism f : X → Y from E to the orbit equivalence relation of S∞ y Y must map a comeager subset of X to a single S∞-orbit. If moreover E has meager orbits, it follows that E is not classifiable by countable structures. In [Sh02, Theorem 2.3], it was proven that there is a continuum of different τ -topologies on R. We now prove that equality of τ -topologies is not even classifiable by countable structures. Theorem 3.1. Let G be a second countable, locally compact, noncompact, abelian group. Let K ⊂ bG be a nonempty closed subset for which there exists a probability measure µ1 ∈ Prob(bG) with support K such that cµ1 tends to zero at infinity. Then, the equivalence relation on Prob(K) defined by µ ∼ µ′ iff τ (µ) = τ (µ′) has meager equiv- alence classes and is generically S∞-ergodic. In particular, ∼ is not classifiable by countable structures. Note that the theorem applies wheneverbλ(K) > 0, wherebλ is the Haar measure on bG, because it then suffices to choose µ1 ∈ Prob(K) in the same measure class asbλK . So, for G as in Theorem 3.1, the equivalence relation on Prob(bG) given by equality of τ -topologies is not classifiable by countable structures. In particular, it is not smooth, meaning that it is impossible to define a standard Borel space of topologies on G in such a way that µ 7→ τ (µ) becomes a Borel map. Proof. Denote by ≈ the equivalence relation on Prob(K) given by mutual absolute continuity. Note that ≈ ⊂ ∼. Since K is the support of the probability measure µ1 whose Fourier transform cµ1 tends to zero at infinity, it follows in particular that µ1 has no atoms, so that K has no 5 isolated points. By [KS99, Theorem 2], the equivalence relation ≈ is generically S∞-ergodic. A fortiori, the equivalence relation ∼ is generically S∞-ergodic. To prove that ∼ has meager equivalence classes, choose an atomic µ0 ∈ Prob(K) such that the set of atoms of µ0 is dense in K. (3.2) is comeager in Prob(K). Claim. If gn ∈ G is such that gn → ∞ andcµ0(gn) → 1 and if a ∈ [0, 1], then (cid:8)µ ∈ Prob(K)(cid:12)(cid:12) there exists a subsequence gnk such that bµ(gnk ) → a(cid:9) P (k, n0) where P (k, n0) = [n≥n0(cid:8)µ ∈ Prob(K)(cid:12)(cid:12) bµ(gn) − a < 1/k(cid:9) . \k≥1 \n0≥1 Note that the set in (3.2) equals Since P (k, n0) is a union of open sets, to prove the claim, it suffices to prove that P (k, n0) is dense in Prob(K). For i ∈ {0, 1}, denote by Pi ⊂ Prob(K) the set of probability measures that are absolutely continuous w.r.t. µi. Since both µ0 and µ1 have support K, it follows that Pi ⊂ Prob(K) is dense for every i ∈ {0, 1}. Choose ηi ∈ Pi and define η = aη0 + (1 − a)η1. Since cµ0(gn) → 1, we get that πµ0(gn) → 1 strongly and thus, bη0(gn) → 1. Since cµ1 tends to zero at infinity, the representation πµ1 is mixing and it follows that bη1(gn) → 0. So,bη(gn) → a and η ∈ P (k, n0). Thus, aP0 + (1 − a)P1 ⊂ P (k, n0). Since Pi ⊂ Prob(K) is dense for every i ∈ {0, 1}, it follows that P (k, n0) ⊂ Prob(K) is dense as well. Assume that the equivalence class [µ]∼ of some µ ∈ Prob(K) is not meager. Since the closure of πµ0(G) is compact while G is noncompact, we can choose as follows a sequence gn ∈ G such that gn → ∞ and πµ0(gn) → 1 strongly. Starting with any sequence hn ∈ G such that hn → ∞, after passage to a subsequence, we may assume that πµ0(hn) converges strongly. Taking a sequence kn ∈ N that tends to infinity sufficiently fast, the sequence gn = hknh−1 n has the required properties. the claim above to the sequence (gnk ), there exists η1 ∈ [µ]∼ and a subsequence gnkl In particular,cµ0(gn) → 1. Since the nonmeager set [µ]∼ intersects every comeager set, by the claim above, there exists η0 ∈ [µ]∼ and a subsequence gnk such that limk bη0(gnk ) = 1. Applying liml bη1(gnkl such that 6→ e in the topology τ (η1). Since τ (η0) = τ (µ) = τ (η1), we have reached a contradiction. So we have proven that the equivalence relation ∼ has meager equivalence classes. → e in the topology τ (η0), while gnkl ) = 0. It follows that gnkl Remark 3.2. Let M be the free Araki-Woods factor associated with an orthogonal represen- tation (Ut)t∈R of R on a real Hilbert space HR of dimension at least 2. In [Sh97, Corollary 8.6], it is proven that M is full and that τ (M ) coincides with the weakest topology on R that makes the map R → O(HR) : t 7→ Ut continuous (see [Va04, Th´eor`eme 2.7] for the fully general case). So also from Theorem 3.1, one can deduce that free Araki-Woods factors cannot be classified by countable structures. In the case where G = R, the non-classification of τ -topologies in Theorem 3.1 can be made concrete as follows. Define the compact space K = {−1, 0, 1}N and the continuous map θ : K → R : θ(x) = xn2−n . ∞Xn=0 Consider the Polish space X = (0, 1/4)N and define for every a ∈ X, the probability measure νa on K given by νa = ∞Yn=1(cid:16)(1 − an)δ0 + an 2 δ−1 + an 2 δ1(cid:17) . 6 Note that a 7→ νa is a continuous map from X to Prob(K) equipped with the weak topology. Pushing forward with θ, we define the probability measure µa on R given by µa = θ∗(νa). The following proposition describes the τ -topology of the measure µa. Proposition 3.3. For every a ∈ X, the topology τ (µa) is induced by the translation invariant metric da on R defined by da(t, s) = ∞Xn=0 an dZ(cid:0)2−n(t − s)(cid:1)2!1/2 , (3.3) where dZ denotes the distance of a real number to Z ⊂ R. Proof. Fix a ∈ X. For every t ∈ R, we have exp(2πi θ(x) t) dνa(x) = ∞Yn=0(cid:0)1 − an(1 − cos(2πt2−n))(cid:1) . Since an ∈ (0, 1/4) for all n, it follows that for every sequence tk ∈ R, the following holds: cµa(t) =ZK cµa(tk) → 0 if and only if This concludes the proof of the proposition. an(1 − cos(2πtk2−n)) → 0 if and only if da(tk, 0) → 0 . ∞Xn=0 Combining Proposition 3.3 with the following result, we obtain a more concrete proof for the non-classification of τ -topologies on R. Proposition 3.4. Define the equivalence relation ∼ on X given by a ∼ b if and only if the metrics da and db defined in (3.3) induce the same topology on R. Then ∼ has meager equiv- alence classes and is generically S∞-ergodic. In particular, ∼ is not classifiable by countable structures. Proof. Define the equivalence relation ≈ on X given by a ≈ b if and only ifP∞ n=0 an −bn < ∞. Let a ∈ X and tk ∈ R a sequence such that da(tk, 0) → 0. For every fixed n, we have that an > 0 and thus that limk dZ(2−ntk) = 0. It follows that ≈ ⊂ ∼. Define the homeomorphism ϕ : RN → X : ϕ(x)n = ϕ0(xn) where ϕ0(y) = 1 8 + 1 4π arctan(y) . Define the equivalence relation ≈1 on RN given by a ≈1 b if and only if a − b ∈ ℓ1 R(N). Since ϕ0 is a contraction, when a ≈1 b, also ϕ(a) ≈ ϕ(b) and thus, ϕ(a) ∼ ϕ(b). Since ϕ is a homeomorphism and since, by [Hj97, Proposition 3.25], ≈1 is generically S∞-ergodic, also ∼ is generically S∞-ergodic. Let U ⊂ X be a non meager set. By Lemma 3.5 below, we can take a ∈ U and a sequence nk ∈ N such that nk → ∞ and da(2nk , 0) → 0. Also by Lemma 3.5 below, we can take b ∈ U such that db(2nk , 0) 6→ 0. So the topologies on R induced by da and db are different. Since a, b ∈ U and U was an arbitrary non meager set, it follows that ∼ has meager equivalence classes. Lemma 3.5. The set of a ∈ X for which there exists a sequence nk ∈ N such that nk → ∞ and da(2nk , 0) → 0 is comeager. Given a sequence nk ∈ N with nk → ∞, the set of b ∈ X such that db(2nk , 0) 6→ 0 is comeager. 7 Proof. Note that for all a ∈ X and m ∈ N, we have that da(2m, 0)2 = am+k2−k . ∞Xk=1 Using that aj ∈ (0, 1/4) for all j ∈ N, it follows that for all a ∈ X and m, n ∈ N, 1 2 am+1 ≤ da(2m, 0)2 ≤ 2−n−2 + am+k2−k . nXk=1 (3.4) Using the second inequality in (3.4), it follows that the first set in the lemma contains ∞\n=1 ∞[m=nna ∈ X(cid:12)(cid:12)(cid:12) nXk=1 am+k2−k < 2−no , which is a countable intersection of open dense sets. Given a sequence nk ∈ N with nk → ∞, using the first inequality in (3.4), it follows that the second set in the lemma contains ∞\k=1 ∞[j=k(cid:8)b ∈ X(cid:12)(cid:12) bnj+1 > 1/8(cid:9) , which also is a countable intersection of open dense sets. 4 Non-classification of actions on free product II1 factors Let G be a locally compact, second countable group and M a von Neumann algebra with separable predual. An action of G on M is a continuous group homomorphism from G to the Polish group Aut(M ). We denote this space of actions as Hom(G, Aut(M )). Equipped with the topology of uniform convergence on compact sets, Hom(G, Aut(M )) is a Polish space. There are several natural notions of equivalence of actions. Given α, β ∈ Hom(G, Aut(M )), one says that 1. α and β are conjugate if there exists an element θ ∈ Aut(M ) such that θ ◦ βg ◦ θ−1 = αg for all g ∈ G ; 2. α and β are cocycle conjugate if there exists a continuous map u : G → U (M ) and an element θ ∈ Aut(M ) such that θ ◦ βg ◦ θ−1 = (Ad ug) ◦ αg and ugh = ugαg(uh) for all g, h ∈ G ; 3. α and β are outer conjugate if there exists an element θ ∈ Aut(M ) such that for all g ∈ G, θ ◦ βg ◦ θ−1 and αg have the same image in the outer automorphism group Out(M ) = Aut(M )/ Inn(M ). Clearly, conjugacy implies cocycle conjugacy, and cocycle conjugacy implies outer conjugacy, but the converse implications need not hold. We recall from [Sh97] the following definition of the τ -invariant of a group action on a full factor, which is analogous to the τ -invariant for von Neumann algebras (see [Co74a] and the beginning of Section 3). 8 Definition 4.1 ([Sh97, Definition 8.1]). Let G be a locally compact, second countable group and M a full factor with separable predual. For every action α ∈ Hom(G, Aut(M )), the topology τ (α) is defined as the weakest topology on G that makes the map G → Out(M ) : g 7→ αg continuous. Note that τ (α) is invariant under outer conjugacy. We can then combine Theorem 3.1 with the construction of Gaussian actions to prove that for noncompact abelian groups G, cocycle conjugacy and outer conjugacy of actions of G on the free group factors L(Fn), and more generally on arbitrary free products (M, τ ) = (M1, τ1) ∗ (M2, τ2) of a diffuse and a nontrivial tracial von Neumann algebra, are not classifiable by countable structures. In particular, outer conjugacy of a single automorphism of L(Fn) or of such a free product (i.e. an action of G = Z) is not classifiable by countable structures. Note that in [KLP14, Theorem 6.2], it was proven that conjugacy of a single automorphism of a free product of II1 factors is not classifiable by countable structures. Theorem 4.2. Let G be a second countable, locally compact, noncompact, abelian group. Let (M, τ ) = (M1, τ1) ∗ (M2, τ2) be the free product of two von Neumann algebras equipped with a faithful normal tracial state. Assume that M1 is diffuse and that M2 6= C1. Then M is a full II1 factor. None of the equivalence relations of conjugacy, cocycle conjugacy or outer conjugacy on the space Hom(G, Aut(M )) of actions of G on M is classifiable by countable structures. Before proving Theorem 4.2, we need the following lemma. Lemma 4.3. Let (X, µ) be a standard nonatomic probability space and let G be a second countable, locally compact, noncompact, abelian group. Denote by Autpmp(X, µ) the Polish group of probability measure preserving automorphisms of (X, µ). There exists a continuous map with the following properties. β : Probc(bG) → Hom(G, Autpmp(X, µ)) : ρ 7→ βρ that βρ(g) = θ ◦ βη(g) ◦ θ−1 for all g ∈ G. 1. If ρ, η ∈ Probc(bG) are mutually absolutely continuous, there exists a θ ∈ Autpmp(X, µ) such 2. For every ρ ∈ Probc(bG), the weakest topology on G making the map G → Autpmp(X, µ) : g 7→ βρ(g) continuous equals the topology τ (ρ). Proof. Let H be a fixed separable Hilbert space. Since H can be viewed as a real Hilbert space, we can view (X, µ) as the Gaussian probability space associated with H, see e.g. [AEG93, Proof of Proposition 1.2]. We then obtain a continuous and injective homomorphism π : U (H) → Autpmp(X, µ) with the property that for all un, u ∈ U (H), we have that un → u strongly if and only if π(un) → π(u). θ∗(γ) = λ, where λ is the Lebesgue measure on [0, 1]. Using the notation of Lemma 2.3, define To prove the lemma, we choose a probability measure γ on bG such that γ ≈bλ, wherebλ is the Haar measure on bG. Put H = L2(bG, γ). Choose a bijective Borel map θ : bG → [0, 1] such that for every ρ ∈ Probc(bG), the Borel map For every ρ ∈ Probc(bG), we define the unitary representation ζρ of G on H given by ψρ : bG → bG : ψρ = θ−1 ◦ ϕθ∗(ρ) ◦ θ . ζρ : G → U (L2(bG, γ)) : (ζρ(g)ξ)(ω) = (ψρ(ω))(g) ξ(ω) . 9 Using the unitary V in Lemma 2.3, it follows that the unitary representation ζρ is unitarily equivalent to the unitary representation πρ defined in (3.1). In particular, if ρ, η ∈ Probc(bG) are mutually absolutely continuous, the unitary representations ζρ and ζη are unitarily equivalent. It follows that the map has all the required properties. β : Probc(bG) → Hom(G, Autpmp(X, µ)) : βρ(g) = π(ζρ(g)) Proof of Theorem 4.2. By [Ue10, Theorem 3.7], any free product (M, τ ) = (M1, τ1)∗(M2, τ2) of tracial von Neumann algebras with M1 diffuse and M2 nontrivial is a full II1 factor. Moreover, in the proof of [Ue10, Theorem 3.7], the following is shown: if N1 ⊂ M1 is a diffuse von Neumann subalgebra and if un ∈ M is a bounded sequence such that kunx − xunk2 → 0 for all x ∈ N1 ∪ M2, then kun − τ (un)1k2 → 0. We are in exactly one of the following two cases. Case 1. M1 has a direct summand of type In. Case 2. M1 is of type II1. Proof in case 1. Take a nonzero central projection z ∈ Z(M1) such that M1z = A⊗Mn(C) for some n ≥ 1 and some diffuse abelian von Neumann algebra A. Write A = L∞(X, µ) ⊗ B where B is a diffuse abelian von Neumann algebra. Whenever β ∈ Autpmp(X, µ), we denote by Ψ(β) ∈ Aut(M1) the unique automorphism given by the identity on N1 = (B ⊗ Mn(C)) ⊕ M1(1 − z) and given by β on L∞(X, µ) ⊂ M1z. Note that β 7→ Ψ(β) is a continuous homomorphism from Autpmp(X, µ) to the Polish group of trace preserving automorphisms of (M1, τ1). Using the notation of Lemma 4.3, define Θ : Probc(bG) → Hom(G, Aut(M )) : ρ 7→ Θρ with Θρ(g) = Ψ(βρ(g)) ∗ id . By Lemma 4.3, we get that the actions Θρ and Θη are conjugate whenever the probability measures ρ, η ∈ Probc(bG) are mutually absolutely continuous. We claim that the τ -invariant τ (Θρ) equals τ (ρ). To prove the claim, assume that Θρ(gn) → id in Out(M ). We have to prove that Θρ(gn) → id in Aut(M ), because then βρ(gn) → id in Autpmp(X, µ), so that gn → e in τ (ρ) by Lemma 4.3. Take un ∈ U (M ) such that (Ad un) ◦ Θρ(gn) → id in Aut(M ). Since the automorphisms Θρ(gn) act as the identity on N = N1 ∗ M2, we get that kunx − xunk2 → 0 for all x ∈ N . By [Ue10, Theorem 3.7], as explained in the first paragraph of the proof, we get that Ad un → id and the claim follows. We can now complete the proof of the theorem in case 1. Assume that f : Hom(G, Aut(M )) → Z is a Borel reduction of either conjugacy, cocycle conjugacy or outer conjugacy to the orbit equivalence relation E of a continuous action of S∞ on a Polish space Z. Since G is noncompact, Since Θρ and Θη are conjugate when ρ ≈ η, we get that (f (Θρ), f (Θη)) ∈ E whenever ρ ≈ η. the dual group bG is nondiscrete and therefore, bG has no isolated points. By [KS99, Theorem 2], the equivalence relation ≈ of mutual absolute continuity on Prob(bG) is generically S∞-ergodic. By [KS99, Proposition 4.1], Probc(bG) ⊂ Prob(bG) is a dense Gδ-set. So ≈ is also generically S∞-ergodic on Probc(bG). So, there exists a comeager set U ⊂ Probc(bG) such that (f (Θρ), f (Θη)) ∈ E for all ρ, η ∈ U . By Theorem 3.1, equality of τ -topologies on Prob(bG) has meager equivalence classes. So, we find ρ, η ∈ U with τ (ρ) 6= τ (η). Since (f (Θρ), f (Θη)) ∈ E and f is a Borel reduction, we also get that Θρ and Θη are outer conjugate. This implies that τ (Θρ) = τ (Θη). So, τ (ρ) = τ (η) and we have reached a contradiction. 10 Proof in case 2. Fix a projection q ∈ M2 with 0 < τ (q) < 1. By [Dy92b, Proposition 3.2], we can take a large enough integer n ≥ 1 such that the free product (Mn(C), τ ) ∗ (Cq + C(1 − q), τ ) is of type II1. A fortiori, (Mn(C), τ ) ∗ (M2, τ ) is of type II1. Since M1 is of type II1, we can write M1 = N1 ⊗ Mn(C). Fix a minimal projection p ∈ Mn(C). Then, M ∼= pM p ⊗ Mn(C). By [Dy92b, Theorem 1.2], we have pM p ∼= N1 ∗ N2 where N2 = p(Mn(C) ∗ M2)p. Now we can use the same trick as in [KLP14, Theorem 6.2]. Since N1 and N2 are of type II1, we can write Ni = Pi ⊗ M2(C). By [Dy92a, Theorem 3.5(iii)], we can identify N1 ∗ N2 = (P1 ⊗ M2(C)) ∗ (P2 ⊗ M2(C)) = (L(F3) ∗ P1 ∗ P2) ⊗ M2(C) . Further identifying L(F3) = L∞(X, µ) ∗ L(F2), we have found an isomorphism θ : M → (L∞(X, µ) ∗ L(F2) ∗ P1 ∗ P2) ⊗ M2n(C) . Defining Θ : Probc(bG) → Hom(G, Aut(M )) : ρ 7→ Θρ with Θρ(g) = θ−1 ◦(cid:0)(βρ(g) ∗ id ∗ id ∗ id) ⊗ id(cid:1) ◦ θ , the same reasoning as above concludes the proof of the theorem. Remark 4.4. In the specific case of M = L(F∞), one can also consider the so-called free Bogoljubov actions arising from Voiculescu's free Gaussian functor (see [Vo83]). When G is a second countable, locally compact, noncompact, abelian group, the same arguments as in the proof of Theorem 4.2 show that the equivalence relations of conjugacy, cocycle conjugacy and outer conjugacy on the free Bogoljubov actions are not classifiable by countable structures. References [AEG93] S. Adams, G.A. Elliott and T. Giordano, Amenable actions of groups. Trans. Amer. Math. Soc. 344 (1994), 803-822. [Co74a] A. Connes, Almost periodic states and factors of type III1. J. Funct. Anal. 16 (1974), 415-445. [Co74b] A. Connes, On hyperfinite factors of type III0 and Krieger's factors. J. Funct. Anal. 18 (1975), 318-327. A. Connes, Classification of injective factors. Ann. of Math. 104 (1976), 73-115. [Co75] [CT76] A. Connes and M. Takesaki, The flow of weights on factors of type III. Tohoku Math. J. 29 (1977), 473-575. [Dy92a] K. Dykema, Interpolated free group factors. Pacific J. Math. 163 (1994), 123-135. [Dy92b] K. Dykema, Free products of hyperfinite von Neumann algebras and free dimension. Duke [Ef64] Math. J. 69 (1993), 97-119. E.G. Effros, The Borel space of von Neumann algebras on a separable Hilbert space. Pacific J. Math. 15 (1965), 1153-1164. [FW03] M. Foreman and B. Weiss, An anti-classification theorem for ergodic measure preserving [Ha85] [Hj97] transformations. J. Eur. Math. Soc. (JEMS) 6 (2004), 277-292. U. Haagerup, Connes' bicentralizer problem and uniqueness of the injective factor of type III1. Acta Math. 158 (1987), 95-148. G. Hjorth, Classification and orbit equivalence relations. Mathematical Surveys and Mono- graphs 75, American Mathematical Society, Providence, 2000. [HSV16] C. Houdayer, D. Shlyakhtenko and S. Vaes, Classification of a family of non almost periodic [Ke95] free Araki-Woods factors. J. Eur. Math. Soc. (JEMS), to appear. arXiv:1605.06057 A.S. Kechris, Classical descriptive set theory. Graduate Texts in Mathematics 156, Springer- Verlag, New York, 1995. 11 [KS99] A.S. Kechris and N.E. Sofronidis, A strong generic ergodicity property of unitary and self- adjoint operators. Ergodic Theory Dynam. Systems 21 (2001), 1459-1479. [KLP08] D. Kerr, H. Li and M. Pichot, Turbulence, representations, and trace-preserving actions. Proc. Lond. Math. Soc. 100 (2010), 459-484. [KLP14] D. Kerr, M. Lupini and N.C. Phillips, Borel complexity and automorphisms of C ∗-algebras. J. Funct. Anal. 268 (2015), 3767-3789. [Kr74] W. Krieger, On ergodic flows and the isomorphism of factors. Math. Ann. 223 (1976), 19-70. A. Ocneanu, Actions of discrete amenable groups on von Neumann algebras. Lecture Notes [Oc85] in Mathematics 1138, Springer-Verlag, 1985. S. Popa, Deformation and rigidity for group actions and von Neumann algebras. In Pro- ceedings of the International Congress of Mathematicians (Madrid, 2006), Vol. I, European Mathematical Society Publishing House, 2007, pp. 445-477. [Po06] [Ru62] W. Rudin, Fourier analysis on groups. Interscience Tracts in Pure and Applied Mathematics [Sh96] [Sh97] [Sh02] [ST08] [ST09] [Ue10] [Va04] [Vo83] 12, John Wiley and Sons, New York, London, 1962. D. Shlyakhtenko, Free quasi free-states. Pacific J. Math. 177 (1997), 329-368. D. Shlyakhtenko, A-valued semicircular systems. J. Funct. Anal. 500 (1999), 1-47. D. Shlyakhtenko, On the classification of full factors of type III. Trans. Amer. Math. Soc. 356 (2004), 4143-4159. R. Sasyk and A. Tornquist. The classification problem of von Neumann factors, J. Funct. Anal. 256 (2009), 2710-2724. R. Sasyk and A. Tornquist, Turbulence and Araki-Woods factors. J. Funct. Anal. 259 (2010), 2238-2252. Y. Ueda, Factoriality, type classification and fullness for free product von Neumann algebras. Adv. Math. 228 (2011), 2647-2671. S. Vaes, ´Etats quasi-libres libres et facteurs de type III (d'apr`es D. Shlyakhtenko). Ast´erisque 299, (2005), 329-350. D. Voiculescu, Symmetries of some reduced free product C∗-algebras. In Operator algebras and their connections with topology and ergodic theory (Busteni, 1983), Lecture Notes in Math. 1132, Springer, Berlin, 1985, pp. 556-588. 12
1107.2053	1	1107	2011-07-11T15:16:36	Aligned CP-semigroups	[ "math.OA" ]	A CP-semigroup is aligned if its set of trivially maximal subordinates is totally ordered by subordination. We prove that aligned spatial E_0-semigroups are prime: they have no non-trivial tensor product decompositions up to cocycle conjugacy. As a consequence, we establish the existence of uncountably many non-cocycle conjugate E_0-semigroups of type II_0 which are prime.	math.OA	math	ALIGNED CP-SEMIGROUPS CHRISTOPHER JANKOWSKI, DANIEL MARKIEWICZ, AND ROBERT T. POWERS Abstract. A CP-semigroup is aligned if its set of trivially maximal subordinates is totally ordered by subordination. We prove that aligned spatial E0-semigroups are prime: they have no non-trivial tensor product decompositions up to cocycle conjugacy. As a consequence, we establish the existence of uncountably many non-cocycle conjugate E0-semigroups of type II0 which are prime. Let H be a Hilbert space, which we will always assume to be separable and infinite- dimensional, and let B(H) denote the ∗-algebra of all bounded operators over H. A CP- semigroup acting on B(H) is a point-σ-weakly continuous semigroup α = {αt : B(H) → B(H)}t≥0 of normal completely positive contractions such that α0 = id. When αt is an en- domorphism and αt(I) = I for all t ≥ 0, then α is called an E0-semigroup. In the special case when H = K ⊗ L2(0, ∞), a CP-semigroup α acting on B(H) is called a CP-flow over K if αt(A)St = StA for all A ∈ B(H), t ≥ 0 where {St : t ≥ 0} is the right shift semigroup. We will say that a CP-semigroup β is a subordinate of α if it also acts on B(H) and αt − βt is completely positive. If in addition β is a CP-flow, then it is called a flow subordinate of α. We direct the reader to [Arv03] for a general reference on the theory of CP-semigroups and to [Pow03b] for the basic theory of CP-flows. In this paper we study a class of CP-semigroups which has a set of subordinates which is minimal in a natural sense. We call such CP-semigroups aligned. This class is shown to include examples considered previously in [Pow03a, APP06, Jan10, JMP11]. We prove that aligned E0-semigroups have a notable property: they are prime in the sense that they have no non-trivial tensor product decompositions up to cocycle conjugacy. As a consequence, we establish the existence of uncountably many non-cocycle conjugate E0-semigroups of type II0 which are prime. The previously known non-trivial examples of prime E0-semigroups were obtained independently in [MP09] (type II1), [Tsi08] (type II1) and [Lie09] (type IIk for k = 1, 2, . . . ). Powers introduced in [Pow03a] the concept of q-purity which has been valuable for the study of E0-semigroups (see for example [APP06, Jan10]). The notion of q-purity was refined in [JMP11]: a CP-flow is q-pure if and only if its set of flow subordinates is totally ordered by subordination. It is clear that an E0-semigroup which is in addition a q-pure CP-flow must have index 0 or 1, and must be of type I1, II0 or II1. The case of type I0 is excluded because an automorphism group cannot be a CP-flow. The restriction to flow subordinates in the definition of q-pure CP-flows can obscure some useful properties of these CP-semigroups. In order to circumvent this difficulty we consider an alternative concept which is inspired by the notion of q-purity. Definition 1. Let α be a CP-semigroup and let β be a CP-semigroup subordinate of α. We will say that β is trivially maximal if the semigroup ektβt is not a subordinate of α for k > 0. 2000 Mathematics Subject Classification. Primary: 46L55, 46L57. C.J. was partially supported by the Skirball Foundation via the Center for Advanced Studies in Mathematics at Ben-Gurion University of the Negev. D.M. and R.T.P. were partially supported by grant 2008295 from the U.S.-Israel Binational Science Foundation. 1 2 CHRISTOPHER JANKOWSKI, DANIEL MARKIEWICZ, AND ROBERT T. POWERS We will denote by S(α) the set of all trivially maximal subordinates of α partially ordered by subordination. We will say α is aligned if S(α) is totally ordered. Let β be a trivially maximal subordinate of α. We will denote by S(α; β) the set of all trivially maximal CP-semigroup subordinates of α which dominate β, partially ordered by subordination. We will say that α is aligned relative to β if S(α; β) is totally ordered. Notice that if an aligned E0-semigroup is spatial, then it must have index zero. Lemma 2. A unital CP-semigroup is aligned if and only if its minimal dilation is aligned. Proof. Suppose that α is a unital CP-semigroup acting on H with minimal dilation αd acting on B(H1), i.e. there exists an isometry W : H → H1 such that αt(x) = W ∗αd t (W xW ∗)W and W W ∗ is increasing for αd. In order to prove the statement it suffices to show that there is an order isomorphism between S(α) and S(αd). As proved in [Bha01] or by Theorem 3.5 of [Pow03b], there exists an order isomorphism between the set of CP-semigroup subordinates of α and the set of CP-semigroup subordinates of αd. This isomorphism is described as follows. For every subordinate β of α there exists a unique subordinate β′ of αd such that βt(x) = W ∗β′ (1) It is clear that if β is trivially maximal (with respect to α), then β′ is trivially maximal (with respect to αd). Conversely, suppose that β is not trivially maximal, so that there exists k > 0 such that ektβt is a subordinate of α. Then there exists γ a unique subordinate of αd such that t(W xW ∗)W. ektβt(x) = W ∗γt(W xW ∗)W, ∀x ∈ B(H), t ≥ 0. Therefore, by dividing by ekt we obtain that β is also the compression of e−ktγt which is t = e−ktγt obviously also a subordinate since k > 0. By uniqueness of the correspondence, β′ for all t. Hence β′ is not trivially maximal. (cid:3) CP-flow subordinates of a CP-flow are always trivially maximal, therefore if a CP-flow is aligned then it is automatically q-pure. We also note that a CP-flow is q-pure if and only if it is aligned with respect to the subordinate x 7→ StxS∗ t . We omit the elementary proof of this fact. We now show that unital CP-flows are aligned if and only if they are q-pure and induce E0- semigroups of type II0. First let us approach the case when the CP-flow is in also a semigroup of endomorphisms. Proposition 3. Let α be an E0-semigroup which is in addition a CP-flow over K. If α has type II0, then S(α) is precisely the set of flow subordinates of α. Therefore, α is aligned if and only if it is type II0 and it is q-pure. Proof. Suppose that α is an E0-semigroup of type II0 which is also a CP-flow over K. Note that every flow subordinate of α is clearly an element of S(α), as flow subordinates are always trivially maximal. Conversely, let β be a trivially maximal CP-semigroup subordinate of α. By Theorem 3.4 of [Pow03b] there exists a family of operators (C(t))t≥0 in B(H) such that βt(x) = C(t)αt(x), where the family (C(t))t≥0 is a contractive positive local cocycle, i.e. C(t) ∈ αt(B(H))′ and 0 ≤ C(t) ≤ I for all t > 0, C(t + s) = C(t)αt(C(s)) for all t, s ≥ 0 and t 7→ C(t) is strongly continuous for t ≥ 0 with C(t) → I as t → 0+. Let St denote as usual the right shift semigroup on H = K ⊗ L2(0, ∞) and let Vt = C(t)St. Note that Vt is strongly continuous and furthermore it is a semigroup: for all t, r ≥ 0, VtVr = C(t)StC(r)Sr = C(t)αt(C(r))StSr = C(t + r)St+r = Vt+r. ALIGNED CP-SEMIGROUPS 3 Moreover, Vt is a unit of αt, since for all x ∈ B(H), αt(x)Vt = αt(x)C(t)St = C(t)αt(x)St = C(t)Stx = Vtx. Notice, however, that α is type II0, therefore there exists κ ∈ C such that Vt = e−κtSt for all t ≥ 0. Furthermore, since C(t) is a contraction, we have that Vt is a contraction for all t > 0, hence we must have Re(κ) ≥ 0. We now show that in fact κ must be a real number. Recall that βt is a CP-semigroup, and 0 ≤ S∗ t βt(I)St = S∗ t C(t)St = e−κtI, hence κ is real and satisfies κ ≥ 0. We will now prove that κ = 0. Let γt(x) = eκtβt(x). Notice that St is an intertwiner semigroup for γ: for all t > 0 and x ∈ B(H), γt(x)St = eκtβt(x)St = eκtC(t)αt(x)St = eκtC(t)Stx = Stx. It follows that γ is a CPκ-flow in the sense of Definition 4.0 of [Pow03b], that is to say, e−κtγt is a CP-semigroup and γ is intertwined by the shift. However, by Theorem 4.15 of [Pow03b], every CPκ-flow must be a CP-flow. But a CP-flow is contractive, hence we must have γt(I) ≤ I, thus eκtβt(I) = eκtC(t) ≤ I for all t > 0. Therefore, we have that for all positive t > 0, ektβt(x) = eκtC(t)αt(x) ≤ αt(x). Since β is trivially maximal, we must have that κ = 0. Thus we have shown that every element of S(α) is a CP-flow. Therefore, if α is q-pure and of type II0 then it is aligned. On the other hand, it is clear that if α is an aligned E0-semigroup, then it is type II0 and q-pure as discussed before the proposition. (cid:3) Theorem 4. Let α be a unital CP-flow over K. If the minimal dilation of α is type II0, then S(α) is precisely the set of flow subordinates of α. Therefore, α is aligned if and only if it is q-pure and its minimal dilation is type II0. Proof. By Lemma 4.50 of [Pow03b], there exists a minimal dilation of α to an E0-semigroup αd which is a CP-flow over the Hilbert space H1 = K1 ⊗ L2(0, ∞), i.e. there exists an isometry W : H → H1 such that αt(x) = W ∗αd W W ∗ is increasing for αd and if Sd t denotes the right shift semigroup on the space H1, then W St = Sd t W for all t > 0. We use the order isomorphism established in the proof of Lemma 2 that associates to each subordinate β ∈ S(α) a unique subordinate β′ ∈ S(αd) satisfying (1). If αd is type II0 and β ∈ S(α), then it follows from the previous proposition that β′ is a CP-flow. Hence for all t > 0, t (W xW ∗)W, βt(x)St = W ∗β′ t(W xW ∗)W St = W ∗β′ t(W xW ∗)Sd t W = W ∗Sd t W xW ∗W = Stx. In other words, β is a CP-flow. Thus we proved that all elements of S(α) are CP-flows. On the other hand, every flow subordinate of α is clearly an element of S(α). Thus, if α is a unital q-pure CP-flow and its minimal dilation has type II0, then it is aligned. Conversely, it is clear that if α is aligned then its minimal dilation αd as discussed above is also aligned, thus it has index zero. Since αd is a CP-flow, it cannot be an automorphism group, hence it has type II0. Moreover, since α is aligned, it is clearly q-pure. (cid:3) 4 CHRISTOPHER JANKOWSKI, DANIEL MARKIEWICZ, AND ROBERT T. POWERS Prime E0-semigroups Definition 5. An E0-semigroup α is called prime if, whenever α is cocycle conjugate to β ⊗ γ where β and γ are E0-semigroups, then β or γ is type I0. It follows from the complete classification of E0-semigroups of type I in terms of the index, and the additivity of the index with respect to tensoring, that a prime E0-semigroup of type I is cocycle conjugate either to an automorphism group or to the CAR/CCR flow of index 1. It is a corollary of the work on the gauge group of an E0-semigroup by Markiewicz-Powers in [MP09] or Tsirelson in [Tsi08], that prime E0-semigroups of type II1 exist. And, as it has belatedly come to our attention, earlier Liebscher had proven that prime E0-semigroups of type IIk exist for k ≥ 1 (see Proposition 4.32 and Note 4.33 in [Lie09]). We establish that there exist uncountably many prime E0-semigroups of type II0. Theorem 6. Aligned spatial E0-semigroups are prime. Proof. We prove the contrapositive. Suppose that α is an E0-semigroup and α = β⊗γ where β and γ are two spatial E0-semigroups neither of which has type I0. Without loss of generality, by applying an appropriate conjugacy, we may assume that both act on B(H) where H is infinite-dimensional and separable. Let U and V be normalized units of β and γ, respectively. Let us denote by θU and θV the semigroups given by θU t (x) = VtxV ∗ t , which are E-semigroup subordinates of β and γ, respectively (notice these are not unital since β and γ are not automorphism groups). Notice that σV = β ⊗ θV is a subordinate of α. We show that it is trivially maximal. Suppose that k ≥ 0 and ektσV is also a subordinate of α. t Then we have that for all x ∈ B(H), t (x) = UtxU ∗ t and θV ekt(I ⊗ VtV ∗ t ) = ekt(βt ⊗ θV t )(I) = ektσV t (I) ≤ αt(I) = I However kVtk = 1, hence by taking norms on both sides we conclude ekt ≤ 1 for all t > 0. Thus k = 0. Analogously, σU = θU ⊗ γ is trivially maximal. We prove that σV and σU are incomparable. Suppose that σV ≥ σU . Notice that for all x, y ∈ B(H) and t > 0, σV t (x ⊗ y)(Ut ⊗ I) = αt(x)Ut ⊗ θV σU t (x ⊗ y)(Ut ⊗ I) = θU t (y)) t (x)Ut ⊗ γt(x) = (Ut ⊗ I)(x ⊗ γt(y)) t (x) = (Ut ⊗ I)(x ⊗ θV Therefore we have that for all x, y ∈ B(H) and t > 0, (Ut ⊗ I)∗hσV t (x ⊗ y) − σU t (x ⊗ y)i(Ut ⊗ I) = x ⊗ [θV t (y) − γt(y)] Thus in the special case when x = I, we have that the map y 7→ I ⊗ [θV t (y) − γt(y)] is completely positive for all t > 0, hence θV ≥ γ. However θV is a pure element in the cone of completely positive maps, therefore we have that for every t > 0, γt is a multiple of θV t , and we have a contradiction since γ is not type I0. By symmetry, we obtain that σU and σV are incomparable as asserted. Thus we proved that α is not aligned. (cid:3) Corollary 7. There exist uncountably many non-cocycle conjugate E0-semigroups of type II0 which are prime. Proof. By Theorem 6, it suffices to exhibit an uncountable family of non-cocycle conjugate aligned E0-semigroups. In order to do so, we consider a class of E0-semigroups arising from boundary weight doubles as in [Jan10]. Let g(x) be a fixed complex-valued measurable function such that g 6∈ L2(0, ∞) yet (1 − e−x)1/2g(x) ∈ L2(0, ∞), and for each t > 0 let gt = χ(t,∞)g, which is an element of L2(0, ∞). Define the weight on B(L2(0, ∞)) given by ν(A) = limt→0+(gt, Agt). For every 0 < λ < 1/2, let µλ : M2(C) → C be given by µλ(X) = λx11 + (1 − λ)x22. ALIGNED CP-SEMIGROUPS 5 Let us define the boundary weight map from M2(C)∗ to boundary weights on B(C2⊗L2(0, ∞)) given by ω(ρ)(A) = ρ(I)µλ(cid:0)(I ⊗ ν)(A)(cid:1) for all ρ ∈ M2(C) and A in the domain of finiteness of I⊗ν (the so called null boundary algebra of definition 4.16 in [Pow03b]). Then by Corollary 3.3 of [Jan10] and the assumptions on g(x), we have that ω gives rise to an E0-semigroup of type II0. Once one applies Theorem 4.4 of [JMP11] to reconcile the earlier definition of q-purity with the one in this paper, we obtain from Lemma 4.3 and Proposition 5.2 of [Jan10] that ω gives rise to a q-pure unital CP-flow. Therefore by Theorem 4 it gives to a aligned E0- semigroup αλ. Finally, by Theorem 5.4 of [Jan10], given λ, ζ ∈ (0, 1/2), we have that αλ is cocycle conjugate to αζ if and only if λ = ζ. (cid:3) We should point out that it is possible to obtain a different uncountable family of non- cocycle conjugate E0-semigroups by using Theorem 3.22 of [Pow03a]. For details see the discussion in the end of section III therein. This would be indeed be a different family from the one exhibited above in the sense that, by Corollary 5.5 of [Jan10], the E0-semigroups constituting both families are not cocycle conjugate. References [APP06] A. Alevras, R. T. Powers, and G. L. Price, Cocycles for one-parameter flows of B(H), J. Funct. Anal. 230 (2006), no. 1, 1 -- 64. [Arv03] W. Arveson, Noncommutative dynamics and E-semigroups, Springer Monographs in Mathematics, Springer-Verlag, New York, 2003. [Bha01] B. V. R. Bhat, Cocycles of CCR flows, Mem. Amer. Math. Soc. 149 (2001), no. 709, x+114. [Jan10] C. Jankowski, On type II0 E0-semigroups induced by boundary weight doubles, J. Funct. Anal. 258 (2010), no. 10, 3413 -- 3451. [JMP11] C. Jankowski, D. Markiewicz, and R. T. Powers, E0-semigroups and q-purity, preprint [Lie09] arXiv:1106.2304, 2011. V. Liebscher, Random sets and invariants for (type II) continuous tensor product systems of Hilbert spaces, Mem. Amer. Math. Soc. 199 (2009), no. 930, xiv+101. [MP09] D. Markiewicz and R. T. Powers, Local unitary cocycles of spatial E0-semigroups, J. Funct. Anal. 256 (2009), no. 5, 1511 -- 1543. [Pow03a] R. T. Powers, Construction of E0-semigroups of B(H) from CP-flows, Advances in quantum dy- namics (South Hadley, MA, 2002), Contemp. Math., vol. 335, Amer. Math. Soc., Providence, RI, 2003, pp. 57 -- 97. [Pow03b] , Continuous spatial semigroups of completely positive maps of B(H), New York J. Math. 9 [Tsi08] (2003), 165 -- 269 (electronic). B. Tsirelson, On automorphisms of type II Arveson systems (probabilistic approach), New York J. Math. 14 (2008), 539 -- 576. Christopher Jankowski, Department of Mathematics, Ben-Gurion University of the Negev, P.O.B. 653, Be'er Sheva 84105, Israel. E-mail address: [email protected] Daniel Markiewicz, Department of Mathematics, Ben-Gurion University of the Negev, P.O.B. 653, Be'er Sheva 84105, Israel. E-mail address: [email protected] Robert T. Powers, Department of Mathematics, David Rittenhouse Lab., 209 South 33rd St., Philadelphia, PA 19104-6395, U.S.A. E-mail address: [email protected]
1307.6988	1	1307	2013-07-26T10:43:03	Bounded elements of C*-inductive locally convex spaces	[ "math.OA", "math.RA" ]	The notion of bounded element of C-inductive locally convex spaces (or C-inductive partial -algebras) is introduced and discussed in two ways: the first one takes into account the inductive structure provided by certain families of C-algebras; the second one is linked to natural order of these spaces. A particular attention is devoted to the relevant instance provided by the space of continuous linear maps acting in a rigged Hilbert space.	math.OA	math	BOUNDED ELEMENTS OF C-INDUCTIVE LOCALLY CONVEX SPACES GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI Abstract. The notion of bounded element of C-inductive locally con- vex spaces (or C-inductive partial -algebras) is introduced and dis- cussed in two ways: the first one takes into account the inductive struc- ture provided by certain families of C-algebras; the second one is linked to natural order of these spaces. A particular attention is devoted to the relevant instance provided by the space of continuous linear maps acting in a rigged Hilbert space. 1. Introduction Some locally convex spaces exhibit an interesting feature: they contain a large number of C-algebras that often contribute to their topological struc- ture, in the sense that these spaces can be thought as generalized inductive limits of C-algebras. These objects were called C-inductive locally convex spaces in [8] and their structure was examined in detail, also taking in mind that they arise naturally when one considers the operators acting in the joint topological limit of an inductive family of Hilbert spaces as described in [7]. Indeed, a typical instance of this structure is obtained by considering B(D, D×) of operators acting in the rigged Hilbert space canoni- the space L cally associated to an O-algebra of unbounded operators acting on a dense domain D of Hilbert space H. In [8] a series of features of this structure were studied giving a particular attention to the order structure, positive B(D, D×) contains linear functionals and representation theory. The space L a subspace isomorphic to the -algebra B(H) of bounded operators in H whose elements can be in natural way considered as the bounded elements of B(D, D×). The notion of bounded element of a locally convex -algebra A L was first introduced by Allan [1] with the aim of developing a spectral theory for topological -algebras: an element x of the topological -algebra A[τ ] is 2010 Mathematics Subject Classification. 47L60, 47L40. Key words and phrases. bounded elements, inductive limit of C-algebras, partial - algebras. 1 2 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI Allan bounded if there exists λ 6= 0 such that the set {(λ−1x)n; n = 1, 2, . . .} is a bounded subset of A[τ ]. This definition was suggested by the successful spectral analysis for closed operators in Hilbert space H: a complex number λ is in the resolvent set ρ(T ) of a closed operator T if T − λI has an inverse in the -algebra B(H) of bounded operators. There are, however, several other possibilities for defining bounded ele- ments. For instance, one may say that x is bounded if π(x) is a bounded operator, for every (continuous, in a certain sense) -representation π de- fined on a dense domain Dπ of some Hilbert space Hπ. This could be a reasonable definition in itself, provided that A possesses sufficiently many -representations in Hilbert space. Moreover some attempts to extend this notion to the larger set-up of locally convex quasi -algebras [18, 20, 21, 11] or locally convex partial - algebras [4, 5, 6] have been done. But in these cases, Allan's notion cannot be adopted, since powers of a given element x need not be defined. In the case of -algebras, bounded elements in purely algebraic terms have been considered by Vidav [23] and Schmudgen [17] with respect to some (positive) wedge. The aim of this paper is to extend the notion of bounded element to the case of C-inductive locally convex spaces A with defining family of C- algebras {Bα; α ∈ F} (F is an index set directed upward). There are also in this case several possibilities: the first one consists in taking elements that have representatives in every C-algebra Bα of the family whose norms are uniformly bounded; the second one consists into taking into account the order structure of A, in the same spirit of the quoted papers of Vidav and Schmudgen. The paper is organized as follows. After some preliminaries (Section 2), B(D, D×) can be derived we study, in Section 3, how bounded elements of L from its C-inductive structure and from its order structure. We show that these two notions are equivalent and that an element X is bounded if and only if X maps D into H and X ∈ B(H). Finally, in Section 4, we consider the same problem for abstract C-inductive locally convex spaces and give B(D, D×) maintain conditions for some of the characterizations proved for L their validity. Some of these results are then specialized to the case where A is a C-inductive locally convex partial -algebra. BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 3 2. Notations and preliminaries For general aspects of the theory of partial -algebras and of their repre- sentations, we refer to the monograph [2]. For the convenience of the reader, however, we repeat here the essential definitions. A partial -algebra A is a complex vector space with conjugate linear involution ∗ and a distributive partial multiplication ·, defined on a subset Γ ⊂ A× A, satisfying the property that (x, y) ∈ Γ if, and only if, (y∗, x∗) ∈ Γ and (x · y)∗ = y∗ · x∗. From now on, we will write simply xy instead of x · y whenever (x, y) ∈ Γ. For every y ∈ A, the set of left (resp. right) multipliers of y is denoted by L(y) (resp. R(y)), i.e., L(y) = {x ∈ A : (x, y) ∈ Γ}, (resp. R(y) = {x ∈ A : (y, x) ∈ Γ}). We denote by LA (resp. RA) the space of universal left (resp. right) multipliers of A. In general, a partial -algebra is not associative. The unit of partial -algebra A, if any, is an element e ∈ A such that e = e∗, e ∈ RA ∩ LA and xe = ex = x, for every x ∈ A. Let H be a complex Hilbert space and D a dense subspace of H. We denote by L†(D, H) the set of all (closable) linear operators X such that D(X) = D, D(X) ⊇ D. The map X → X † = X ∗ ↾D defines an involution on L†(D, H), which can be made into a partial -algebra with respect to the weak multiplication [2]; however, this fact will not be used in this paper. Let L†(D) be the subspace of L†(D, H) consisting of all its elements which leave, together with their adjoints, the domain D invariant. Then L†(D) is a -algebra with respect to the usual operations. A -subalgebra M of L†(D) is called an O-algebra. Let M be an O-algebra. The graph topology tM on D is the locally convex topology defined by the family {k · kA}A∈M, where kξkA =pkξk2 + kAξk2 = k(I + A∗A)1/2ξk, ξ ∈ D. For A = 0, the null operator of L†(D), k · k0 is exactly the norm of H, thus we will omit the 0 in the notation of the norm. The topology tM is finer than the norm topology, unless M does consist of bounded operators only. If M is an O-algebra, we write A (cid:22) B if kAξk ≤ kBξk, for every ξ ∈ D. Then, M is directed upward with respect to this order relation. If A ∈ M, we denote by HA the Hilbert space obtained by endowing D(A) with the graph norm k · kA. 4 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI If A, B ∈ M and A (cid:22) B, then UBA = (I + B∗B)−1/2(I + A∗A)1/2 is a contractive map of HA into HB; i.e., kUBAξkB ≤ kξkA, for every ξ ∈ HA. If the locally convex space D[tM ] is complete, then M is said to be closed. If M = L†(D) then the corresponding graph topology is denoted by t† instead of tL†(D). As is known, a locally convex topology t on D finer than the topology induced by the Hilbert norm defines, in standard fashion, a rigged Hilbert space (RHS) D[t] ֒→ H ֒→ D×[t×], where D× is the vector space of all continuous conjugate linear functionals on D[t], i.e., the conjugate dual of D[t], endowed with the strong dual topology t× = β(D×, D) and ֒→ denotes a continuous embedding with dense range. The Hilbert space H is identified (by considering the form which puts D and D× as an extension of the inner product of D) with a dense subspace of D×[t×]. Let L(D, D×) denote the vector space of all continuous linear maps from D[t] into D×[t×]. In L(D, D×) an involution X 7→ X † can be introduced by the equality hXξ η i = hX †η ξ i, ∀ξ, η ∈ D. Hence L(D, D×) is a -invariant vector space. To every X ∈ L(D, D×) there corresponds a separately continuous sesquilin- ear form θX on D × D defined by θX(ξ, η) = hXξ η i , ξ, η ∈ D. The vector space of all jointly continuous sesquilinear forms on D × D will B(D, D×) the subspace of all X ∈ be denoted with B(D, D). We denote by L L(D, D×) such that θX ∈ B(D, D) and by L†(D) the -algebra consisting of all operators of L†(D), which together with their adjoints are continuous If t = t†, then L†(D) = L†(D). We will refer to from D[t] into D[t]. the rigged Hilbert space defined by endowing D with the topology t† as to the canonical rigged Hilbert space defined by L†(D) on D. In this case (L B(D, D×), L†(D)) is a quasi -algebra [2]. The spaces L(D, D×) and L B(D, D×) have been studied at length by sev- [12, 13, 14, 22]) and several pathologies concerning eral authors (see, e.g. their multiplicative structure have been considered (see also [2, 3] and ref- erences therein). Recently some spectral properties of operators of these classes have also been studied [10]. BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 5 3. Bounded elements of L B(D, D×) The inductive structure of L B(D, D×), with D endowed with the graph to- pology t†, has been discussed in [8, Section 5]. To keep the paper reasonably self-contained, we sum the main features up. By the definition itself, X ∈ L B(D, D×) if, and only if, there exists γX > 0 and A ∈ L†(D) such that (1) θX(ξ, η) = hXξ η i ≤ γXkξkA kηkA, ∀ξ, η ∈ D. Conversely, if θ ∈ B(D, D), there exists a unique X ∈ L B(D, D×) such that θ = θX. Thus, the map I : X ∈ L B(D, D×) 7→ θX ∈ B(D, D) is an isomorphism of vector spaces and I(θ∗) = X †, where θ∗(ξ, η) = θ(η, ξ), for every ξ, η ∈ D. We denote by BA(D, D) the subspace of B(D, D) consisting of all θ ∈ B(D, D) such that (1) holds for fixed A ∈ L†(D). If θ ∈ BA(D, D), it extends to a bounded sesquilinear form on HA × HA (we use the same symbol for this extension). Hence, there exists a unique operator X θ A ∈ B(HA) such that θ(ξ, η) =DX θ Aξ ηEA , ∀ξ, η ∈ HA. On the other hand, if XA ∈ B(HA), then the sesquilinear form θXA defined by θXA(ξ, η) = hXAξ η iA , ξ, η ∈ D, is an element of BA(D, D) and the map ΦA : XA ∈ B(HA) → θXA ∈ BA(D, D) is a -isomorphism of vector spaces with involution. If B (cid:23) A, then, for ξ, η ∈ D, θXA(ξ, η) = hXAξ η iA ≤ kXAkA,AkξkA kηkA ≤ kXAkA,AkξkB kηkB , where k · kA,A denotes the operator norm in B(HA). Hence, there exists a unique XB ∈ B(HB) such that hXAξ η iA = hXBξ η iB , ∀ξ, η ∈ D. So it is natural to define JBA(XA) = XB, ∀XA ∈ B(HA). 6 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI It is easily seen that JBA = Φ−1 B ΦA. The space LA B (D, D×) := I−1BA(D, D) is a Banach space, with norm kXkA := sup θX (ξ, η) kξkA,kηkA≤1 and L the family of subspaces {LA B(D, D×) can be endowed with the inductive topology τind defined by B (D, D×); A ∈ L†(D)} as in [16, Section 1.2. III]. In conclusion, XA ∈ B(HA) ↔ θXA ∈ BA(D, D) ↔ X ∈ LA B(D, D×) are isometric -isomorphisms of Banach spaces. Hence, to every X ∈ L B(D, D×) one can associate the net {XB; B ∈ L†(D); B (cid:23) A} of its representatives in each of the spaces HB. Definition 3.1. We say that X ∈ L L B(D, D×) if X has a representative XA in every B(HA) and B(D, D×) is a bounded element of kXkb := sup kXAkA,A < +∞. A∈L†(D) The space L B(D, D×)b of all bounded elements of L B(D, D×) is a Banach space with norm k · kb. Proposition 3.2. L algebra of operators. B(D, D×)b is -isomorphic (as Banach space) to a C- Proof. Let H⊕ denote the Hilbert space direct sum of the HA, A ∈ L†(D); i.e., HA H⊕ := MA∈L†(D) =(ξ⊕ = (ξA); ξA ∈ HA, ∀A ∈ L†(D) and XA kξAk2 A < +∞) . If {XA}A∈L†(D) is a net of operators XA ∈ B(HA), A ∈ L†(D), we define X⊕ξ⊕ = {XAξA} provided that PA kXAξAk2 < +∞, ξA ∈ HA. The space constructed in this way is QA The operator X⊕ = {XA} is bounded if and only if supA kXAkA,A < +∞. B(HA) = B(H⊕). To every X ∈ B(D, D×)b we can associate the net {XA} which we have defined above. Clearly, {XA} ∈ B(H⊕). It is easily seen that the map L τ : X ∈ L B(D, D×)b 7→ {XA} ∈ B(H⊕) is isometric. Thus, the statement is proved. (cid:3) BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 7 B(D, D×) having a representative XA for Remark 3.3. An element X ∈ L every A ∈ L†(D) need not be bounded in the sense of Definition 3.1. The spaces {HA; A ∈ L†(D)}, together with their conjugate duals make D× into an indexed PIP-space [3, Ch.2]. In that language, operators having repre- sentatives in every HA are called totally regular operators. For more details on their behavior see [3, Sect. 3.3.3] where also a C-agebra corresponding to our bounded elements has been studied. Our next goal is to characterize bounded elements of L B(D, D×) in sev- eral different ways. For doing this, we need to consider the natural order structure of L B(D, D×). We say that X ∈ L B(D, D×) is positive, and write X ≥ 0, if hXξ ξ i ≥ 0, for every ξ ∈ D. It is easy to see that, if X is positive, then it is symmetric; i.e., X = X †. Proposition 3.4. The following conditions are equivalent. (i) X ≥ 0. (ii) There exists A ∈ L†(D) such that XB ≥ 0, ∀B (cid:23) A. Proof. (i)⇒(ii): Since X ∈ L such that B(D, D×), there exists A ∈ L†(D) and γ > 0 hXξ η i ≤ γkξkBkηkB , B (cid:23) A. If X ≥ 0, then, for every ξ ∈ D, hXBξ ξ iB = hXξ ξ i ≥ 0, ∀B (cid:23) A. Since D is dense in HB, we have hXBξ ξ iB ≥ 0, ∀ξ ∈ HB. (ii)⇒(i): Let XB ≥ 0 for every B (cid:23) A. Then, for every ξ ∈ D, hXξ ξ i = hXBξ ξ iB ≥ 0. (cid:3) Theorem 3.5. Let X ∈ L lent. B(D, D×). The following statements are equiva- (i) X : D → H and X ∈ B(H). (ii) X ∈ L (iii) There exists λ > 0 such that B(D, D×)b. −λI ≤ ℜ(X) ≤ λI, −λI ≤ ℑ(X) ≤ λI where ℜ(X) = X+X † 2 and ℑ(X) = X−X † 2i . Proof. (i)⇒(ii): If X : D → H and X is bounded, then, for every A ∈ L†(D), (2) hXξ η i ≤ kXkkξkkηk ≤ kXkkξkAkηkA. 8 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI This means that X has a bounded representative XA in every B(HA). By (2), kXAkA,A ≤ kXk, for every A ∈ L†(D), so supA∈L†(D) kXAkA,A < +∞. (ii)⇒(i) Let X ∈ L B(D, D×)b. Then, for every A ∈ L†(D) ∀ξ, η ∈ D. hXξ η i ≤ kXAkA,A kξkA kηkA, In particular, for A = 0, (3) hXξ η i ≤ kX0kkξkkηk, ∀ξ, η ∈ D. By (3), for every ξ ∈ D, F (η) = hXξ η i is a bounded conjugate linear functional on D, so by Riesz's lemma Xξ ∈ H. It is, finally easily seen that X ∈ B(H). (iii)⇒(i) Suppose first that X = X †. Note that the operator X sat- is a positive operator and isfies the following: 0 ≤ X+λI 2λ ≤ I; so X+λI 2λ (cid:10) X+λI 2λ ξ ξ(cid:11) ≤ hξ ξ i , ∀ξ ∈ D; this implies that and by Riesz's lemma there exists ζ ∈ H such that (4) (5) 2λ (cid:12)(cid:12)(cid:12)(cid:12) (cid:28) X + λI (cid:28) X + λI 2λ ≤ kξk kηk, ξ η(cid:29)(cid:12)(cid:12)(cid:12)(cid:12) ξ η(cid:29) = hζ η i , ∀ξ, η ∈ D ∀ξ, η ∈ D and then X+λI representative for every A ∈ L†(D). Indeed, 2λ ξ ∈ H. This implies that Xξ ∈ H too. Moreover, X has a hXξ η i ≤ γkξkkηk ≤ γkξkAkηkA ∀A ∈ L†(D), where γ > 0. From (4) it follows that X is bounded and X ∈ B(H). In the very same way one can prove the boundedness of X if X † = −X. The result for a general X follows easily. (i)⇒ (iii): this is a standard result of the C-algebras theory. (cid:3) 4. Bounded elements of C-inductive locally convex spaces The results obtained in Section 3 have an abstract generalization to locally convex spaces that are inductive limits of C-algebras in a generalized sense. These spaces were called C-inductive locally convex spaces in [8]. We begin with recalling the basic definitions. Let A be a vector space over C. Let F be a set of indices directed upward and consider, for every α ∈ F, a Banach space Aα ⊂ A such that: (I.1) Aα ⊆ Aβ, if α ≤ β; (I.2) A =Sα∈F Aα; BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 9 (I.3) ∀α ∈ F, there exists a C-algebra Bα (with unit eα and norm k · kα) and a norm-preserving isomorphism of vector spaces φα : Bα → Aα; β , for every α, β ∈ F with β ≥ α. β φα)(xα) ∈ B+ α ⇒ xβ = (φ−1 (I.4) xα ∈ B+ We put jβα = φ−1 If x ∈ A, there exist α ∈ F such that x ∈ Aα and (a unique) xβ ∈ Bβ β φα, if α, β ∈ F, β ≥ α. such that x = φβ(xβ), for all β ≥ α. Then, we put jβα(xα) := xβ if α ≤ β. By (I.4), it follows easily that jβα preserves the involution; i.e., jβα(x∗ α) = (jβα(xα))∗. The family {Bα, jβα, β ≥ α} is a directed system of C-algebras, in the sense that: (J.1) for every α, β ∈ F, with β ≥ α, jβα : Bα → Bβ is a linear and injective map; jαα is the identity of Bα, (J.2) for every α, β ∈ F, with α ≤ β, φα = φβjβα. (J.3) jγβjβα = jγα, α ≤ β ≤ γ. We assume that, in addition, the jβα's are Schwarz maps (see, e.g. i.e., [15]); (sch) jβα(xα)∗jβα(xα) ≤ jβα(x∗ αxα), ∀xα ∈ Bα, α ≤ β. For every α, β ∈ F, with α ≤ β, jβα is continuous [15] and, moreover, kjβα(xα)kβ ≤ kxαkα, ∀xα ∈ Bα. An involution in A is defined as follows.. Let x ∈ A. Then x ∈ Aα, for α). some α ∈ F, i.e., x = φα(xα), for a unique xα ∈ Bα. Put x∗ := φα(x∗ Then if β ≥ α, we have φ−1 β (x∗) = φ−1 β (φα(x∗ α)) = jβα(x∗ α) = (jβα(xα))∗ = x∗ β. It is easily seen that the map x 7→ x∗ is an involution in A. Moreover, by the definition itself, it follows that every map φα preserves the involution; i.e., φα(x∗ α) = (φα(xα))∗, for all xα ∈ Bα, α ∈ F. Definition 4.1. A locally convex vector space A, with involution ∗, is called a C-inductive locally convex space if (i) there exists a family {{Bα, φα}, α ∈ F}, where F is a direct set and, for every α ∈ F, Bα is a C-algebra and φα is a linear injective map of Bα into A, satisfying the above conditions (I.1) - (I.4) and (sch), with Aα = φα(Bα), α ∈ F; 10 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI (ii) A is endowed with the locally convex inductive topology τind gener- ated by the family {{Bα, φα}, α ∈ F}. The family {{Bα, φα}, α ∈ F} is called the defining system of A. We notice that the involution is automatically continuous in A[τind]. A C-inductive locally convex space has a natural positive cone. An element x ∈ A is called positive if there exists γ ∈ F such that φ−1 α , ∀α ≥ γ. B+ We denote by A+ the set of all positive elements of A. α (x) ∈ Then, (i) Every positive element x ∈ A is hermitian; i.e., x ∈ Ah := {y ∈ A : y∗ = y} . (ii) A+ is a non empty convex pointed cone; i.e. A+ ∩ (−A+) = {0}. (iii) If α ∈ F and xα ∈ B+ Moreover, every hermitian element x = x∗ is the difference of two positive α , φα(xα) is positive. elements, i.e. there exist x+, x− ∈ A+ such that x = x+ − x−. A linear functional ω is said to be positive if ω(x) ≥ 0 for every x = (xα) ∈ A. As shown in [8, Prop. 3.9, 3.10], ω is positive if, and only if, ωα(xα) := ω(φα(xα)) ≥ 0 for every α ∈ F. We write, in this case, ω = lim −→ ωα. 4.1. Bounded elements. Definition 4.2. Let A be a C-inductive locally convex space. An element x = (xα) ∈ A, with xα ∈ Bα, is called bounded if x ∈ Aα, for every α ∈ F and supα∈F kxαkα < ∞. The set of bounded elements of A is denoted by Ab. Proposition 4.3. The set Ab is a Banach space under the norm kxkb = supα∈F kxαkα. Proof. We only prove the completeness. Let {xn} be a Cauchy sequence in Ab. Then for every α ∈ F the sequence {xα n := (xn)α, is Cauchy in Bα so it converges to some xα ∈ Bα. Since the jβα's are continuous, one easily proves that the family {xα} defines an element x = (xα) of A. From the Cauchy condition, for every ǫ > 0, there exists nǫ ∈ N such that n}, with xα (6) sup α∈F kxα n − xα mkα < ǫ BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 11 If m > nǫ, Hence, Thus x ∈ Ab. kxαkα ≤ kxα − xα mkα. mkα + kxα mkα ≤ ǫ + kxα sup α∈F kxαkα ≤ ǫ + sup α∈F kxα mkα < ∞. Fix now n > nǫ and let m → ∞ in (6). Then, kxα n − xαkα ≤ ǫ. sup α∈F This proves that xn → x. (cid:3) In what follows we will consider -representations of a C-inductive locally convex space. We recall the basic definitions. Let F be a set directed upward by ≤. A family {Hα, Uβα, α, β ∈ F, β ≥ α}, where each Hα is a Hilbert space (with inner product h· · iα and norm k · kα) and, for every α, β ∈ F, with β ≥ α, Uβα is a linear map from Hα into Hβ, is called a directed contractive system of Hilbert spaces if the following conditions are satisfied (i) Uβα is injective; (ii) kUβαξαkβ ≤ kξαkα, (iii) Uαα = Iα, the identity of Hα; (iv) Uγα = UγβUβα, α ≤ β ≤ γ. ∀ξα ∈ Hα; A directed contractive system of Hilbert spaces defines a conjugate dual pair (D×, D) which is called the joint topological limit [7] of the directed contractive system {Hα, Uβα, α, β ∈ F, β ≥ α} of Hilbert spaces. Definition 4.4. Let A be the C-inductive locally convex space defined by the system {{Bα, Φα}, α ∈ F} as in Definition 4.1. For each α ∈ F, let πα be a -representation of Bα in Hilbert space Hα. The collection π := {πα} is said to be a -representation of A if (i) for every α, β ∈ F there exists a linear map Uβα : Hα → Hβ such that the family {Hα, Uβα, α, β ∈ F, β ≥ α} is a directed contractive system of Hilbert spaces; (ii) the following equality holds (7) πβ(jβα(xα)) = Uβαπα(xα)U ∗ βα, ∀xα ∈ Bα, β ≥ α. 12 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI In this case we write π(x) = lim −→ π = lim −→ πα. πα(xα) for every x = (xα) ∈ A or, for short, The -representation π is said to be faithful if x ∈ A+ and π(x) = 0 imply x = 0 (of course, π(x) = 0 means that there exists γ ∈ F such that πα(xα) = 0, for α ≥ γ). Remark 4.5. With this definition (which is formally different from that given in [8] but fully equivalent), π(x), x ∈ A, is not an operator but rather a collection of operators. But as shown in [8], π(x) can be regarded as an operator acting on the joint topological limit (D×, D) of {Hα, Uβα, α, β ∈ F, β ≥ α}. The corresponding space of operators was denoted by LB(D, D×); B(D, D×) studied in Section it behaves in the very same way as the space L 3 and reduces to it when the family of Hilbert spaces is exactly {HA; A ∈ L†(D)}. The main difference consists in the fact that the Hα's need not be all subspaces of a certain Hilbert space H. Lemma 4.6. Let π = lim −→ every α ∈ F, πα is a faithful -representation of Bα. πα be a faithful -representation of A. Then, for Proof. Let xα ∈ B+ α with πα(xα) = 0. Let x ∈ A be the unique element of A such that x = φα(xα). Then πβ(xβ) = πβ(jβα(xα)) = Uβαπα(xα)U ∗ βα = 0. Hence π(x) = 0 and, therefore x = 0. Thus there exists γ ∈ F such that xγ = 0, for γ ≥ γ. Let β ≥ α, γ. Then 0 = xβ = jβα(xα). Hence, by the injectivity of jβα, xα = 0. (cid:3) As shown in [8, Proposition 3.16], if a C-inductive locally convex space A fulfills the following conditions (r1) if xα ∈ Bα and jβα(xα) ≥ 0, β ≥ α, then xα ≥ 0; (r2) eβ ∈ jβα(Bα), (r3) every positive linear functional ω = lim −→ ∀α, β ∈ F, β ≥ α; ing property ωα on A satisfies the follow- • if α ∈ F and ωβ(jβα(x∗ α)jβα(xα)) = 0, for some β > α and xα ∈ Bα, then ωα(x∗ αxα) = 0; then, A admits a faithful representation. The conditions (r1), (r2), in fact, guarantee that A possesses sufficiently many positive linear functionals, in the sense that for every x ∈ A+, x 6= 0 there exists a positive linear functional ω on A such that ω(x) > 0 [8, Theorem 3.14]. Theorem 4.7. Let A be a C∗-inductive locally convex space and x = (xα) ∈ A. The following statements hold. BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 13 (i) If x ∈ Ab, then, for every -representation π = lim −→ πα of A, one has kπα(xα)kαα < ∞, sup α∈F where k · kαα denote the norm of B(Hα). (ii) Conversely, if A admits a faithful -representation πf = lim −→ πf α and kπf α(xα)kαα < ∞, sup α∈F then x ∈ Ab. Proof. (i): For every α ∈ F, πα is a -representation of the C-algebra Bα. Hence kπα(xα)kαα ≤ kxαkα. Thus if x ∈ Ab the statement follows immediately from the definition. (ii): Let πf (x) = lim −→ is a faithful representation of Bα. The -representation πf isomorphism of C∗-algebras, for all α ∈ F; hence α(xα). Then, by Lemma 4.6, for every α ∈ F, πf πf α α is an isometric sup α∈F kxαkα = sup α∈F kπf α(xα)kαα < ∞. This proves that x is a bounded element of A. (cid:3) 4.2. Order bounded elements. Let A be a C-inductive locally convex space. If x ∈ A, we put ℜ(x) = x + x∗ 2 and ℑ(x) = x − x∗ 2i . Both ℜ(x) and ℑ(x) are symmetric elements of A. Assume that A has an element u = u∗ such that kuαkα ≤ 1, for every α ∈ F, and there exists γ ∈ F such that uβ = jβγ(eγ) ∀β ≥ γ, (eγ is the unit of Bγ). For shortness we call the element u a pre-unit of A. Remark 4.8. The pre-unit u ∈ A, if any, is unique. Indeed, let suppose there is another v ∈ A satisfying the same properties as u. Then, ∃γ, γ′ ∈ F; uβ = jβγ(eγ), vβ′ = jβ′γ′(e′ γ), ∀β ≥ γ, β′ ≥ γ′ so, if δ ≥ γ, γ′, one has uλ = vλ, ∀λ ≥ δ. Definition 4.9. Let A be a C-inductive locally convex space with pre-unit u. We say that x ∈ A is order bounded (with respect to u) if there exists λ > 0 such that −λu ≤ ℜ(x) ≤ λu − λu ≤ ℑ(x) ≤ λu. 14 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI Theorem 4.10. Let A be a C-inductive locally convex space satisfying condition (r1). Assume that A has a pre-unit u. Then, x ∈ Ab if, and only if, x has a representative for every α ∈ F (i.e. for every α ∈ F, there exists xα ∈ Bα such that x = φα(xα)) and x is order bounded with respect u. Proof. Let us assume that x = x∗ ∈ Ab. Then, x has a representative xα, with x∗ α = xα, in every B(Hα) and λ := supα∈F kxαkα < ∞. Hence, we have −λeα ≤ xα ≤ λeα, ∀α ∈ F, where eα denotes the unit of Bα. By the definition of u, there exists γ ∈ F such that uβ = jβγ(eγ ) for β ≥ γ. Hence, taking into account that the maps jβα preserve the order, we have −λuβ ≤ xβ ≤ λuβ, ∀β ≥ γ. This implies that −λu ≤ x ≤ λu. Now, let us suppose that for some λ > 0, −λu ≤ x ≤ λu. Then, there exists γ ∈ F such that (8) − λuβ ≤ xβ ≤ λuβ, ∀β ≥ γ. Let now α ∈ F. Then, there is δ ≥ α, γ such that (8) holds for δ ≥ α. Hence, by using (r1), we conclude that −λuα ≤ xα ≤ λuα ∀α ∈ F. This implies that, kxαkα ≤ λ, for every α ∈ F. Thus, x ∈ Ab. (cid:3) From the proof of the previous theorem it follows easily that Proposition 4.11. Let x = x∗ ∈ Ab and put p(x) = inf{λ > 0; −λu ≤ x ≤ λu}. Then, p(x) = kxkb. 5. C-inductive partial -algebras As shown in [8], a partial multiplication in A can be defined by a family w = {wα}, wα ∈ Bα. Let w = {wα} be a family of elements, such that each wα ∈ B+ α and jβα(wα) = wβ, for all α, β ∈ F with β ≥ α. Let x, y ∈ A. The partial multiplication x · y is defined by the conditions: ∃γ ∈ F : φβ(φ−1 x · y = φβ(φ−1 β (x)wβφ−1 β (y)) = φβ′(φ−1 β′ (x)wβ′ φ−1 β′ (y)), ∀β, β′ ≥ γ β (x)wβφ−1 β (y)), β ≥ γ. BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 15 Then, A is an associative partial -algebra with respect to the usual op- erations and the above defined multiplication (see [2, Sect. 2.1.1] for the definitions) and we will call it a C-inductive partial -algebra. The partial -algebra A has a unit e (that is, an element e which is a left- and right universal multiplier such that x · e = e · x = x, for every x ∈ A) if, and only if, every element wα of the family {wα} defining the multiplication is invertible and (9) jβα(w−1 α ) = w−1 β , ∀α, β ∈ F, β ≥ α. In this case, e = φα(w−1 α ), independently of α ∈ F. The element e is called a bounded unit if it is a bounded element of A and kekb = 1. Proposition 5.1. Let A be a C-inductive partial -algebra with the mul- tiplication defined by a family {wα}. Assume that e = (w−1 α ) is a bounded unit of A. Then Ab is a Banach partial -algebra; that is, Ab[k · kb] is a Banach space with isometric involution ∗ and there exists C ≥ 1 such that the following inequality holds (10) kx · ykb ≤ Ckxkbkykb, ∀x, y ∈ Ab with x · y well-defined. Remark 5.2. The constant C in (10) can be taken equal to 1 if w−1 α = eα, for each α ∈ F, where eα is the unit of the C-algebra Bα. Under the same assumption, the norm of Ab satisfies the C-property, which in our case reads kx∗ · xkb = kxk2 b , ∀x ∈ Ab with x∗ · x well-defined. This is no longer true in the general case. Remark 5.3. In Example 5.3 of [8] two of us tried to construct a family {WA ∈ B(HA); A ∈ L†(D)} so that the partial multiplication defined in B(D, D×) by the method mentioned above would reproduce the quasi - L B(D, D×), L†(D)) (see Section 2). Unfortunately, the algebra structure of (L conclusion of that discussion is uncorrect (see, [9] for more details). Let A be a C-inductive partial -algebra with the multiplication defined by a family {wα} as above. The spaces RA and LA of the right-, respectively, left universal multipliers (with respect to w) of A are algebras. Hence, A0 := LA ∩ RA is a -algebra and, thus, (i) (A, A0) is a quasi -algebra. (ii) If A is endowed with τind, then the maps x 7→ x∗, x 7→ a · x, x 7→ x · b, a, b ∈ A0 are continuous. 16 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI It is easily seen from the very definition that, if a ∈ RA and x ∈ A+, then a∗xa ∈ A+. Hence, if P(A) denotes the family of all positive linear functionals on A, we have in particular ω(a∗xa) ≥ 0, for every ω ∈ P(A). Theorem 5.4. Let A be a C-inductive partial -algebra with the multipli- cation defined by a family {wα} and with pre-unit u. Assume, moreover, that the following condition (P) holds: (P) y ∈ A, ω(a∗ya) ≥ 0, ∀ω ∈ P(A) and a ∈ RA ⇒ y ∈ A+ ; then, for x ∈ A, the following conditions are equivalent. (i) x is order bounded with respect to u. (ii) There exists γx > 0 such that ω(a∗xa) ≤ γxω(a∗ua), ∀ω ∈ P(A), ∀a ∈ RA. (iii) There exists γx > 0 such that ω(b∗xa)2 ≤ γxω(a∗ua)ω(b∗ub), ∀ω ∈ P(A), ∀a, b ∈ RA. Proof. It is sufficient to consider the case x = x∗; (i)⇒(ii): Let ω ∈ P(A). By the hypothesis, −γu ≤ x ≤ γu, for some γ > 0; then ω(γu − x) ≥ 0 and ω(a∗(γu − x)a) ≥ 0, ∀a ∈ RA. On the other hand, similarly, one can show that ω(a∗(x − γu)a) ≥ 0. (ii)⇒(i): Assume now that u is a pre-unit and there exists γx > 0 such that ω(a∗xa) ≤ γxω(a∗ua), ∀ω ∈ P(A), a ∈ RA. Then γxω(a∗ua) ± ω(a∗xa) ≥ 0 ⇒ ω(a∗(γxu ± x)a) ≥ 0, ∀ω ∈ P(A), a ∈ RA. So, by (P), γxu ± x ≥ 0. (i)⇒(iii): By the assumption, there exists γ > 0 such that −γu ≤ x ≤ γu. Let ω ∈ P(A). Then, the linear functional ωa on A, defined by ωa(x) := ω(a∗xa), is positive. Hence, if x = x∗ i.e., −γωa(u) ≤ ωa(x) ≤ γωa(u); ω(a∗xa) ≤ γω(a∗ua). Now, let x ∈ A+, a, b ∈ RA. Let us define Ωx ω(a, b) := ω(b∗xa). Then, it ω is a positive sesquilinear form on RA × RA. Using is easily checked that Ωx BOUNDED ELEMENTS OF C-INDUCTIVE SPACES 17 the the Cauchy-Schwartz inequality we obtain ω(b∗xa) ≤ ω(a∗xa)1/2ω(b∗xb)1/2 ≤ γω(a∗ua)1/2ω(b∗ub)1/2. The extension to arbitrary x ∈ A goes through as in the proof of Proposition 4.3 of [8]. (iii)⇒(ii) It is trivial. (cid:3) The previous proof shows that if x = x∗ ∈ A is order bounded with respect to u then p(x) ≤ sup{ω(b∗xa); ω ∈ P(A); a, b ∈ RA; ω(a∗ua) = ω(b∗ub) = 1}. where p(x) is the quantity defined in Proposition 4.11. The following statement is an easy consequence of Theorem 4.11 and Theorem 5.4. Theorem 5.5. Let A be a C-inductive partial -algebra with the multi- plication defined by a family {wα} and pre-unit u. Assume that conditions (r1) and (P) are satisfied. For an element x ∈ A, having a representative in every Bα, α ∈ F, the following statements are equivalent. (i) x ∈ Ab. (ii) x is order bounded with respect to u. (iii) For every ω ∈ P(A) ω(b∗xa)2 ≤ γxω(a∗ua)ω(b∗ub), ∀a, b ∈ RA. References [1] G.R. Allan, A spectral theory for locally convex algebras, Proc. London. Math. Soc. 15 (1965) 399 -- 421. [2] J-P. Antoine, A. Inoue, C. Trapani, Partial -algebras and their operator realizations, Kluwer, Dordrecht, 2002. [3] J-P. Antoine, C. Trapani, Partial Inner Product Spaces -- Theory and Applications, Springer Lecture Notes in Mathematics, vol. 1986, Berlin, Heidelberg, 2009. [4] J-P. Antoine, C. Trapani, F. Tschinke, Spectral properties of partial -algebras Mediterr. j. math. 7 (2010) 123 -- 142. [5] J.-P. Antoine, C.Trapani and F. Tschinke, Bounded elements in certain topological partial -algebras, Studia Math. 203 (2011), 223-251. [6] J-P. Antoine, G. Bellomonte, C. Trapani, Fully representable and -semisimple topo- logical partial -algebras, Studia Mathematica, 208 (2012), 167-194. 18 GIORGIA BELLOMONTE, SALVATORE DI BELLA, AND CAMILLO TRAPANI [7] G. Bellomonte, C. Trapani, Rigged Hilbert spaces and contractive families of Hilbert spaces Monatshefte f. Math., 164, (2011) 271-285 (published on line on October 2010 DOI 10.1007/s00605-010-0249-1). [8] G. Bellomonte, C. Trapani, Quasi -algebras and generalized inductive limits of C- algebras, Studia Mathematica 202 (2011), 165-190. [9] G. Bellomonte, C. Trapani, Erratum/Addendum to the paper "Quasi -algebras and generalized inductive limits of C-algebras", [Studia Mathematica 202 (2011), 165- 190], to appear. [10] G. Bellomonte, S. Di Bella, C. Trapani,Operators in Rigged Hilbert spaces: some spectral properties, preprint, Palermo 2013. [11] M. Fragoulopoulou, C. Trapani and S. Triolo, Locally convex quasi -algebras with sufficiently many -representations, J. Math. Anal. Appl. 388 (2012), 1180-1193. [12] K-D. Kursten, The completion of the maximal Op-algebra on a Fr´echet domain, Publ. Res. Inst. Math. Sci., Kyoto Univ. 22 (1986), 151 -- 175. [13] K-D. Kursten, On algebraic properties of partial algebras, Rend. Circ. Mat. Palermo, Ser.II, Suppl. 56 (1998), 111 -- 122. [14] K-D. Kursten and M. Lauter, An extreme example concerning factorization products on the Schwartz space S(Rn) Note Mat. 25 (2005/06), 31 -- 38. [15] T.W. Palmer, Banach Algebras and the General Theory of -Algebras, Volume 2, Cambridge Univ. Press, Cambridge, 2001. [16] K. Schmudgen, Unbounded operator algebras and representation theory, Birkhauser Verlag, Basel, 1990. [17] K. Schmudgen, A strict Positivstellensatz for theWeyl algebra, Math. Ann. 331 (2005) 779 -- 794. [18] C. Trapani, Bounded elements and spectrum in Banach -algebras, Studia Mathe- matica 172 (2006) 249 -- 273. [19] C.Trapani, Unbounded C-seminorms, biweights and -representations of partial - algebras: a review, International J. Math. Math.Sci., Volume 2006 (2006), Article ID 79268, 34 pages. [20] C. Trapani, Bounded and strongly bounded elements of Banach -algebras, Contem- porary Math. 427 (2007) 417 -- 424. [21] C. Trapani, -Representations, seminorms and structure properties of normed quasi -algebras, Studia Mathematica, 186 (2008), 47-75. [22] C. Trapani and F. Tschinke, Partial multiplication of operators in rigged Hilbert spaces, Integral Equations Operator Theory 51 (2005), 583 -- 600. [23] I. Vidav, On some *regular rings, Acad. Serbe Sci. Publ. Inst. Math. 13 (1959) 73 -- 80. Dipartimento di Matematica e Informatica, Universit`a di Palermo, I-90123 Palermo, Italy E-mail address: [email protected] E-mail address: [email protected] E-mail address: [email protected]
1102.2012	2	1102	2011-12-15T16:07:46	Mapping Cones are Operator Systems	[ "math.OA", "quant-ph" ]	We investigate the relationship between mapping cones and matrix ordered *-vector spaces (i.e., abstract operator systems). We show that to every mapping cone there is an associated operator system on the space of n-by-n complex matrices, and furthermore we show that the associated operator system is unique and has a certain homogeneity property. Conversely, we show that the cone of completely positive maps on any operator system with that homogeneity property is a mapping cone. We also consider several related problems, such as characterizing cones that are closed under composition on the right by completely positive maps, and cones that are also semigroups, in terms of operator systems.	math.OA	math	MAPPING CONES ARE OPERATOR SYSTEMS NATHANIEL JOHNSTON AND ERLING STØRMER Abstract. We investigate the relationship between mapping cones and matrix ordered ∗-vector spaces (i.e., abstract operator systems). We show that to every mapping cone there is an associated operator system on the space of n-by-n complex matrices, and furthermore we show that the associated operator system is unique and has a certain homogeneity property. Conversely, we show that the cone of completely positive maps on any operator system with that homogeneity property is a mapping cone. We also consider several related problems, such as characterizing cones that are closed under composition on the right by completely positive maps, and cones that are also semigroups, in terms of operator systems. Keywords: operator systems, mapping cones, dual cones, positive maps AMS Subject Classifications: 15B48, 47D03, 47L99 1. Introduction In operator theory, some of the most important families of linear maps are the positive and k-positive maps, and their dual cones [1] of superpositive and k-superpositive maps. These sets of maps are all specific examples of mapping cones [2], which are closed cones of positive maps that are invariant under left and right composition by completely positive maps -- a property of k-positive and k-superpositive maps that is easily verified. It has recently been shown [3, 4, 5] that the k-positive and k-superpositive maps can be seen as the completely positive maps on certain natural operator system structures. We thus have two settings, seemingly very different, that give rise to the familiar cones of k- positive and k-superpositive maps. A natural question that arises is whether this is simply coincidence, or if there is indeed a fundamental link between mapping cones and operator systems. In this work, we show that there is indeed an extremely strong connection between mapping cones and operator systems. In fact, we show that there is a bijection between mapping cones and operator systems with a property that we refer to as super-homogeneity. If we remove the super-homogeneity property, then the bijection is no longer with mapping cones but rather with cones that are only closed under right (but not necessarily left) composition by completely positive maps. We also answer some related questions involving semigroup cones of positive maps. The paper is organized as follows. In Section 2 we introduce much of our notation and present the basics of cones of linear maps on complex matrices. In Section 3 we present abstract operator systems and derive a simple uniqueness property in the finite-dimensional setting that we are interested in. In Section 4 we present and prove our most general result for right-CP-invariant cones, which shows their intimate link with operator systems, and in Section 5 we investigate the special case of mapping cones and the kinds of operator systems 1 2 N. JOHNSTON AND E. STØRMER that they give rise to. We close in Section 6 by exploring some properties of mapping cones that have the additional property of being semigroups, and we see that they too can be seen as arising from operator systems. 2. Cones of Positive Maps If H is a (finite-dimensional) Hilbert space and L(H) is the space of linear maps on H, then a map Φ : L(H) → L(H) is said to be positive if Φ(X) ∈ L(H)+ whenever X ∈ L(H)+, k-positive if idk ⊗ Φ is positive, and completely positive if Φ is k-positive for all k ∈ N. If A ∈ L(H) then the map AdA : L(H) → L(H) defined by AdA(X) ≡ A∗XA is completely positive, and conversely every completely positive map can be written as a sum of maps of the form AdAi for some {Ai} ∈ L(H) [6, 7]. i,j=1 eiej ∗ ⊗ eiej ∗ (ej Given a fixed orthonormal basis {ei}n L(H) ⊗ L(H), where E := Pn i=1 of H, the Choi-Jamio lkowski isomorphism [8] associates a linear map Φ : L(H) → L(H) with the operator CΦ := (idn ⊗ Φ)(E) ∈ ∗ is the outer product of ei and ej, and CΦ is called the Choi matrix of Φ). For us, it will be useful to know that Φ is completely positive if and only if CΦ is positive, and Φ is positive if and only if CΦ is block-positive -- i.e., (v∗ ⊗ w∗)CΦ(v ⊗ w) ≥ 0 for all v, w ∈ H. Given a cone of positive maps C, we define CC := {CΦ : Φ ∈ C} and C† := {Φ† : Φ ∈ C}, where Φ† : L(H) → L(H) is the adjoint map defined via the Hilbert-Schmidt inner product so that Tr(Φ(X)Y ) = Tr(XΦ†(Y )) for all X, Y ∈ L(H). ∗ is the dual vector of ej, eiej A mapping cone [2] is a nonzero closed cone C of positive maps from L(H) into itself with the property that Φ ◦ Ω ◦ Ψ ∈ C whenever Ω ∈ C and Φ, Ψ : L(H) → L(H) are completely positive. For the remainder of this work, we will generally assume all cones to be convex, though we will still specify if the distinction is important or there is the possibility of confusion. By linearity, it is enough that AdA ◦ Ω ◦ AdB ∈ C whenever Ω ∈ C and A, B ∈ L(H) for a convex cone C to be a mapping cone. It will also sometimes be useful for us to consider (not necessarily closed) cones C such that Ω ◦ Ψ ∈ C whenever Ω ∈ C and Ψ : L(H) → L(H) is completely positive -- that is, cones that are closed under right-composition, but not necessarily left-composition, by completely positive maps. We will call such cones right-CP-invariant. Left-CP-invariant cones can be defined analogously, and it is clear that C is right-CP-invariant if and only if C† is left-CP-invariant. The dual cone C◦ of a cone C ⊆ L(H) of Hermitian operators is defined via the Hilbert- Schmidt inner product as C◦ := {Y ∈ L(H) : Tr(XY ) ≥ 0 for all X ∈ C}. Similarly, the dual cone C◦ of a cone C of maps on L(H) is defined via the Choi-Jamio lkowski isomorphism as C◦ := {Ψ : L(H) → L(H) : Tr(CΦCΨ) ≥ 0 for all Φ ∈ C}. We note that for convex cones C ⊆ L(H), we have C◦◦ = C -- the closure of C. This fact is well-known in convex analysis and follows easily from [9, Theorem 14.1] or [10, Theorem 3.4.3], for example. Throughout the rest of this work, we will associate the n-dimensional Hilbert space H with Cn and L(H) with the space of n × n complex matrices Mn, both for simplicity and to be consistent with standard operator system notation. Then L(Mn) denotes the set of linear maps from Mn into itself, P(Mn) denotes the set of positive maps on Mn, Pk(Mn) denotes the set of k-positive maps on Mn, and CP(Mn) the set of completely positive maps MAPPING CONES ARE OPERATOR SYSTEMS 3 on Mn. We let Pn ⊆ Mn ⊗ Mn denote the cone of block-positive operators, and Sn := P ◦ n is the cone of separable operators [1] -- operators X ∈ Mn ⊗ Mn that can be written in the form Yi ⊗ Zi for positive semidefinite {Yi}, {Zi} ∈ Mn. X =Xi Associated to the cone of separable operators via the Choi-Jamio lkowski isomorphism is the cone of superpositive maps S(Mn) (sometimes called entanglement-breaking maps [11]). Similarly, the cone of k-superpositive maps is the dual cone of the cone of k-positive maps: Sk(Mn) := Pk(Mn)◦. We close this section with a simple lemma (which also appeared as [12, Lemma 3] with a different proof) that allows us to relate the Choi matrices of Φ and Φ†. Note that in the particularly important case Φ = AdA, the lemma says that (idn ⊗ AdA)(E) = (AdAT ⊗ idn)(E), where T denotes the transpose map. Lemma 1. Let Φ : Mn → Mn. Then (idn ⊗ Φ)(E) = ((T ◦ Φ† ◦ T ) ⊗ idn)(E). Proof. Throughout this proof, by a vectorization vec(X) of a matrix X, we mean the vector in Cn ⊗ Cn ∼= Cn2 obtained from X ∈ Mn by stacking the columns of X on top of each other, starting with the leftmost column. Use the singular value decomposition to write ∗. It is easily verified that for any X ∈ Mn, vec(X T ) = F vec(X), where F is the "swap" or "flip" operator that acts on elementary tensors as F (a ⊗ b) = b ⊗ a. The result follows from recalling (see [7] or [13, Proposition 6.2], for example) that we can write (cid:3) i Y Ai, where vec(Ai) = vi and vec(Bi) = wi. CΦ = Pi viwi Φ(X) =Pi AiXB∗ i and Φ†(Y ) =Pi B∗ 3. Operator Systems on Mn An (abstract) operator system on Mn is a family of convex cones {Cm}∞ m=1 ⊆ Mm ⊗ Mn that satisfy the following two properties: • C1 = M + • for each m1, m2 ∈ N and A ∈ Mm1,m2 we have (AdA ⊗ idn)(Cm1 ) ⊆ Cm2 . n , the cone of positive semidefinite elements of Mn; and Abstract operator systems can be defined more generally as matrix ordered ∗-vector spaces on any Archimedean ∗-ordered vector space V , but the above definition with V = Mn is much simpler and suited to our particular needs. The interested reader is directed to [14, Chapter 13] for a more thorough treatment of general abstract operator systems. The fact that matrix ordered ∗-vector spaces can be thought of as operator systems follows from the work of Choi and Effros [15]. Remark 2. Abstract operator systems typically are defined with two additional require- ments that we have not mentioned: • Cm ∩ −Cm = {0} for each m ∈ N; and • for every m ∈ N and X = X ∗ ∈ Mm ⊗Mn, there exists r > 0 such that rI +X ∈ Cm. Both of these conditions follow for free from the fact that, in our setting, C1 = M + n . To see that the first property holds, notice that C1 ∩ −C1 = {0}, and suppose that X ∈ Cm ∩ −Cm for some m ≥ 2. Then (AdA ⊗ idn)(X) ∈ C1 ∩ −C1 for any A ∈ Mm,1. Because C1 ∩ −C1 = {0}, it follows that (v∗ ⊗ w∗)X(v ⊗ w) = 0 for all v ∈ Cm, w ∈ Cn. It follows (via [16, Lemma 2.1], for example) that X = 0, so Cm ∩ −Cm = {0} for all m ∈ N. 4 N. JOHNSTON AND E. STØRMER The second property holds because the smallest family of cones on Mn such that (AdA ⊗ idn)(Cm1) ⊆ Cm2 for all m1, m2 ∈ N are the cones of separable operators in Mm ⊗ Mn [5, Theorem 5]. It is well-known that there always exists r > 0 such that rI + X is separable [17], so the same r ensures that rI + X ∈ Cm. One particularly important operator system is the one constructed by associating Mm ⊗ Mn with Mmn in the natural way and letting Cm ⊆ Mm ⊗ Mn be the cones of positive semidefinite operators. We will denote this operator system simply by Mn, and it will be clear from context whether we mean the operator system Mn or simply the set Mn without regard to any family of cones. Other operator systems on Mn will be denoted like O(Mn) in order to avoid confusion with the operator system Mn itself. If O1(Mn) and O2(Mn) are two operator systems defined by the cones {Cm}∞ m=1 and {Dm}∞ m=1 respectively, then a map Φ : Mn → Mn is said to be completely positive from O1(Mn) to O2(Mn) if (idm ⊗ Φ)(Cm) ⊆ Dm for all m ∈ N. The set of maps that are completely positive from O1(Mn) to O2(Mn) is denoted by CP(O1(Mn), O2(Mn)), or simply CP(O(Mn)) if the target operator system equals the source operator system. It will often be useful for us to consider operator systems with the additional property that (idm ⊗ AdB)(Cm) ⊆ Cm for each m ∈ N and B ∈ Mn -- a property that is equivalent to the fact CP(Mn) ⊆ CP(O(Mn)). We will call operator systems with this property super- homogeneous. We now present a result that shows that operator systems on Mn are in fact characterized completely by their nth cone. That is, there is a unique way to construct an operator system given an appropriate cone Cn ⊆ Mn ⊗ Mn. Proposition 3. Let Cn ⊆ Mn ⊗ Mn be a convex cone such that Sn ⊆ Cn ⊆ Pn and (AdA ⊗idn)(Cn) ⊆ Cn for all A ∈ Mn. Then there exists a unique family of cones {Cm}m6=n such that {Cm}∞ m=1 defines an operator system on Mn, given by Cm :=(cid:8)Xi (AdAi ⊗ idn)(X) : Ai ∈ Mn,m ∀ i, X ∈ Cn(cid:9). Furthermore, the operator system is super-homogeneous if and only if (idn ⊗AdB)(Cn) ⊆ Cn for all B ∈ Mn. Proof. We first prove that the family of convex cones given by the proposition do indeed define an operator system. We first show that (AdB ⊗ idn)(Y ) ∈ Cm2 for any m1, m2 ∈ N, Y ∈ Cm1, and B ∈ Mm1,m2. This is true from the definition of Cm if m1 = n. If m1 6= n then write Y = Pi(AdAi ⊗ idn)(X) for some X ∈ Cn and {Ai} ⊂ Mn,m1. Then AiB ∈ Mn,m2 for all i, so We now show that C1 = M + n , and similarly vv∗ ⊗ X ∈ Pn if and only if X ∈ M + X ∈ M + if and only if X ∈ M + where we have identified R+ ⊗ M + noting that if X ∈ C1 and v ∈ Cn then vv∗ ⊗ X ∈ Cn, so X ∈ M + n . C1 ⊆ M + n . For any v ∈ Cn, note that vv∗ ⊗ X ∈ Sn if and only if n . It follows that vv∗ ⊗ X ∈ Cn n , n . The opposite inclusion follows simply from It follows that n . Then C1 ⊇ {(AdA ⊗ idn)(vv∗ ⊗ X) : A ∈ Mn,1, X ∈ M + n(cid:9) = M + n , so C1 = M + n , so the cones {Cm}∞ m=1 define an operator system on Mn. n with M + (AdB ⊗ idn)(Y ) =Xi (AdAiB ⊗ idn)(X) ∈ Cm2. MAPPING CONES ARE OPERATOR SYSTEMS 5 To prove uniqueness, assume that there exists another family of cones {Dm}∞ m=1 that define an operator system, such that Dn = Cn. It is clear that Cm ⊆ Dm for all m ∈ N, so we only need to prove the other inclusion. If m ≤ n, let X ∈ Dm and let V : Cm → Cn be an isometry (i.e., V ∗V = I). Then Y := (AdV ∗ ⊗ idn)(X) ∈ Dn = Cn, so X = (AdV ⊗ idn)(Y ) ∈ Cm. Thus Dm ⊆ Cm, so Dm = Cm for m ≤ n. If m > n then we recall from [4, Section 2.3] the k-super minimal and k-super maximal operator system structures. In particular, it was shown that if two operator systems on Mn, defined by cones {Cm}∞ and {Dm}∞ m ∈ N. See also [5, Section 4]. m=1 m=1 respectively, are such that Cm = Dm for 1 ≤ m ≤ n, then Cm = Dm for all The "only if" direction of the final claim is trivial from the definition of super-homogeneity, and the "if" direction follows easily from the fact that (AdA ⊗ idn) and (idm ⊗ AdB) com- mute. This completes the proof. (cid:3) We close this section with a result that shows that to determine complete positivity of a map from one operator system on Mn to another, it is enough to look at the action of that map on the nth cone of the operator systems. Corollary 4. Let Φ : Mn → Mn and let O1(Mn) and O2(Mn) be operator systems defined by families of cones {Cm}∞ m=1, respectively. Then Φ ∈ CP(O1(Mn), O2(Mn)) if and only if (idn ⊗ Φ)(Cn) ⊆ Dn. m=1 and {Dm}∞ Proof. The "only if" implication follows trivially from the definition of CP(O1(Mn), O2(Mn)). For the "if" implication, suppose (idn ⊗ Φ)(Cn) ⊆ Dn. Fixing m ∈ N arbitrarily and ap- plying Pi AdAi ⊗ idn for {Ai} ∈ Mn,m to both sides then gives (idm ⊗ Φ)(Cm) = [{Ai}∈Mn,m(Xi ⊆ [{Ai}∈Mn,m(Xi (AdAi ⊗ Φ)(Cn)) (AdAi ⊗ idn)(Dn)) = Dm, where both of the above equalities follow from the form of the cones {Cm}∞ guaranteed by Proposition 3. proof. m=1 It follows that Φ ∈ CP(O1(Mn), O2(Mn)), completing the (cid:3) m=1 and {Dm}∞ 4. Right-CP-Invariant Cones as Operator Systems In this section we establish a link between right-CP-invariant cones and operator systems. Our first result is in the same vein as some known results on mapping cones such as [18, Theorem 1] and [19, Theorem 1]. Here we prove an analogous statement for cones that are just right-CP-invariant. Proposition 5. Let C ⊆ L(Mn) be a right-CP-invariant cone. Then Ψ† ◦ Φ ∈ CP(Mn) for all Φ ∈ C if and only if Ψ ∈ C◦. 6 N. JOHNSTON AND E. STØRMER Proof. To prove the "only if" implication, suppose Ψ ∈ L(Mn) and Ψ† ◦ Φ ∈ CP(Mn) for all Φ ∈ C. Then CΨ†◦Φ ∈ (Mn ⊗ Mn)+ so 0 ≤ Tr(ECΨ†◦Φ) = Tr(E(idn ⊗ (Ψ† ◦ Φ))(E)) = Tr((idn ⊗ Ψ)(E)(idn ⊗ Φ)(E)) = Tr(CΨCΦ) ∀ Φ ∈ C, i,j=1 eiej ∗ ⊗ eiej ∗. It follows that Ψ ∈ C◦. It is perhaps worth where we recall that E :=Pn noting that the proof of this implication did not make use of right-CP-invariance of C. To see why the "if" implication holds, assume Ψ ∈ C◦. Then, because C is right-CP- invariant, it follows that for any Φ ∈ C and Ω ∈ CP(Mn) we have Φ ◦ Ω ∈ C so 0 ≤ Tr(CΨCΦ◦Ω) = Tr((idn ⊗ Ψ)(E)(idn ⊗ (Φ ◦ Ω))(E)) = Tr((idn ⊗ (Φ† ◦ Ψ))(E)(idn ⊗ Ω)(E)) = Tr(CΦ†◦ΨCΩ). It follows via the Choi-Jamio lkowski isomorphism that CΦ†◦Ψ ∈ (Mn ⊗ Mn)+, so Φ† ◦ Ψ ∈ CP(Mn). Then (Φ† ◦ Ψ)† = Ψ† ◦ Φ ∈ CP(Mn), completing the proof. (cid:3) It is not difficult to verify that if O(Mn) is any operator system, then CP (Mn, O(Mn)) is a right-CP-invariant cone. Similarly, CP (O(Mn), Mn) is easily seen to be a closed left- CP-invariant cone. The main result of this section shows that these properties actually characterize the possible cones of completely positive maps to and from Mn, and furthermore that these cones uniquely determine O(Mn). Recall that P(Mn) denotes the cone of positive maps on Mn, S(Mn) denotes the cone of superpositive maps on Mn, and CC denotes the cone of Choi matrices of maps from the cone C. Theorem 6. Let C ⊆ L(Mn) be a convex cone. The following are equivalent: (1) C is right-CP-invariant with S(Mn) ⊆ C ⊆ P(Mn). (2) There exists an operator system O1(Mn), defined by cones {Cm}∞ m=1, such that CC = Cn. (3) There exists an operator system O2(Mn) such that C = CP(Mn, O2(Mn)). (4) There exists an operator system O3(Mn) such that (C◦)† = CP(O3(Mn), Mn). Furthermore, O1(Mn) = O2(Mn) and is uniquely determined by C, and O3(Mn) is uniquely determined by C and can be chosen so that O3(Mn) = O1(Mn). Proof. We prove the result by showing that (1) ⇔ (2), (2) ⇔ (3), and (2) ⇔ (4). To see that (1) ⇒ (2), define Cn := CC. If A ∈ Mn and Φ ∈ C then (1) (AdA ⊗ idn)(CΦ) = (AdA ⊗ Φ)(E) = (idn ⊗ (Φ ◦ AdAT ))(E) ∈ Cn, where the second equality comes from Lemma 1 and the inclusion comes from the fact that C is right-CP-invariant, so Φ ◦ AdAT ∈ C. The implication (1) ⇒ (2) and uniqueness of O1 then follow from Proposition 3. The reverse implication (2) ⇒ (1) also follows from the string of equalities (1), but this time we use the fact that Cn is a cone defining an operator system to get the inclusion. The fact that S(Mn) ⊆ C ⊆ P(Mn) follows from the fact that for the minimal operator system on Mn, Cn is the cone of block-positive operators and for the maximal operator system on Mn, Cn is the cone of separable operators [5, Theorem 5]. To see that (2) ⇒ (3), let O2(Mn) = O1(Mn). We then have to show that if CC = Cn, then C = CP(Mn, O1(Mn)). We already showed that (2) ⇒ (1), so we know that C is MAPPING CONES ARE OPERATOR SYSTEMS 7 right-CP-invariant. So if Φ ∈ C then for any X ∈ (Mn ⊗ Mn)+ there exists Ψ ∈ CP(Mn) such that (idn ⊗ Φ)(X) = (idn ⊗ (Φ ◦ Ψ))(E) ∈ Cn, where the inclusion comes from C being right CP-invariant. It follows via Corollary 4 that Φ ∈ CP(Mn, O1(Mn)), so C ⊆ CP(Mn, O1(Mn)). To see the opposite inclusion, simply note that if Φ ∈ CP(Mn, O1(Mn)) then, because E ∈ (Mn ⊗ Mn)+, we have CΦ = (idn ⊗ Φ)(E) ∈ Cn = CC, so Φ ∈ C. It follows that C = CP(Mn, O1(Mn)). To establish uniqueness of O2 (and simultaneously prove (3) ⇒ (2)), suppose that the cones {Dm}∞ m=1 define an operator system O2(Mn) such that C = CP(Mn, O2(Mn)). Be- cause E ∈ (Mn ⊗ Mn)+, we again have that (idn ⊗ Φ)(E) ∈ Dn for any Φ ∈ C, so CC ⊆ Dn. On the other hand by the equivalence of (1) and (2), Dn = CC′ for some right-CP-invariant cone C′. If Φ ∈ C′ then for any X ∈ (Mn ⊗ Mn)+ there exists Ψ ∈ CP(Mn) such that (idn ⊗ Φ)(X) = (idn ⊗ (Φ ◦ Ψ))(E) ∈ Dn, where the inclusion comes from C′ being right CP-invariant. It follows via Corollary 4 that C′ ⊆ CP(Mn, O2(Mn)). Then C ⊆ C′ ⊆ CP(Mn, O2(Mn)) = C, so C = C′ and hence Cn = Dn. Uniqueness now follows from Proposition 3. The proof that (2) ⇔ (4) mimics the proof that (2) ⇔ (3) and makes use of the fact that Ψ† ◦ Φ ∈ CP(Mn) for all Φ ∈ C if and only if Ψ ∈ C◦ (Proposition 5). To see that (2) ⇒ (4), let O3(Mn) = O1(Mn). Then for any Ψ ∈ C◦ and Φ ∈ C we have Ψ† ◦ Φ ∈ CP(Mn), so CΨ†◦Φ ∈ (Mn ⊗ Mn)+. It follows that (idn ⊗ Ψ†)(Cn) ⊆ (Mn ⊗ Mn)+. Corollary 4 implies that Ψ† ∈ CP(O3(Mn), Mn), so (C◦)† ⊆ CP(O3(Mn), Mn). The opposite inclusion follows by simply reversing this argument. Uniqueness of O3 (up to closure) and the implication (4) ⇒ (2) follow similarly by the fact that Ψ† ∈ CP(OC(Mn), Mn) if and only if Ψ† ◦ Φ ∈ CP(Mn) for all Φ ∈ C if and only if Ψ ∈ C◦, where OC(Mn) is an operator system with its nth cone Cn := CC. (cid:3) The equivalence of conditions (1) and (2) in Theorem 6 can be seen as providing a bijection between right-CP-invariant cones and operator systems on Mn. Given an operator system O(Mn) defined by cones {Cm}∞ m=1, the associated right-CP-invariant cone is given via the maps associated to Cn via the Choi-Jamio lkowski isomorphism. In the other direction, given a right-CP-invariant cone, the associated operator system gets its nth cone from the Choi- Jamio lkowski isomorphism and then gets its remaining cones via the construction given in Proposition 3. 5. Mapping Cones as Operator Systems Before introducing the main results of this section, we present a lemma that shows that the largest cone of completely positive maps between any two operator systems on Mn is the cone of positive maps -- a result that follows from recent work on minimal and maximal operator systems [3, 5]. Lemma 7. Let O1(Mn) and O2(Mn) be operator systems. Then CP(O1(Mn), O2(Mn)) ⊆ P(Mn). 8 N. JOHNSTON AND E. STØRMER Proof. Let O1(Mn) and O2(Mn) be defined by the families of cones {Cm}∞ m=1, respectively. Let Φ ∈ L(Mn) be such that Φ /∈ P(Mn). Because the smallest family of cones defining an operator system on Mn are the separable operators and the largest such family of cones are the block-positive operators [5, Theorem 5], we know that I ⊗ X ∈ Cn for all X ∈ M + n such that Φ(X) /∈ M + n . It is then easily verified that I ⊗ Φ(X) /∈ Pn, so (idn ⊗ Φ)(I ⊗ X) /∈ Dn. It follows that Φ /∈ CP(O1(Mn), O2(Mn)), so CP(O1(Mn), O2(Mn)) ⊆ P(Mn). (cid:3) n and Dn ⊆ Pn. Because Φ /∈ P(Mn), there exists a particular X ∈ M + m=1 and {Dm}∞ The following result shows how the bijection inroduced by Theorem 6 works when the right-CP-invariant cone is in fact a mapping cone -- in this situation the associated operator system is super-homogeneous. Corollary 8. Let C ⊆ L(Mn) be a closed, convex cone. The following are equivalent: (1) C is a mapping cone. (2) There exists a super-homogeneous operator system O1(Mn), defined by cones {Cm}∞ m=1, such that CC = Cn. (3) There exists a super-homogeneous operator system O2(Mn) such that C = CP(Mn, O2(Mn)). (4) There exists a super-homogeneous operator system O3(Mn) such that (C◦)† = CP(O3(Mn), Mn). (5) There exist super-homogeneous operator systems O4(Mn) and O5(Mn) such that C = CP(O4(Mn), O5(Mn)). Furthermore, O1(Mn) = O2(Mn) = O3(Mn) and is uniquely determined by C. Proof. The equivalence of (1), (2), (3), and (4) (and uniqueness of the corresponding op- erator systems) follows immediately from the corresponding statements of Theorem 6 and the fact that C is left-CP-invariant if and only if (idn ⊗ AdB)(CC) ⊆ CC, which then gives super-homogeneity of the corresponding operator system via Proposition 3. Because Mn is a super-homogeneous operator system, it is clear that (3) ⇒ (5). All that remains to do is prove that (5) ⇒ (1). To this end, let O4(Mn) and O5(Mn) be super-homogeneous operator systems defined by families of cones {Cm}∞ m=1, respectively. By the equivalence of conditions (1) and (2), we know that there exist mapping cones C′ and C′′ such that CC′ = Cn and CC′′ = Dn. By Corollary 4, (idn ⊗ Φ)(Cn) ⊆ Dn if and only if Φ ∈ CP(O4(Mn), O5(Mn)), so it follows that Φ ◦ Ψ ∈ C′′ for all Ψ ∈ C′ if and only if Φ ∈ CP(O4(Mn), O5(Mn)). Right-CP-invariance of CP(O4(Mn), O5(Mn)) now follows from left-CP-invariance of C′ and left-CP-invariance of CP(O4(Mn), O5(Mn)) follows from left-CP-invariance of C′′. The fact that CP(O4(Mn), O5(Mn)) ⊆ P(Mn) follows from Lemma 7. (cid:3) m=1 and {Dm}∞ It is natural at this point to consider well-known mapping cones and ask what are the corresponding operator systems via the bijection of Corollary 8. The mapping cone of stan- dard completely positive maps CP(Mn) of course corresponds to the "naive" operator system with positive cones equal to the cones of positive semidefinite operators. It was shown in [3] that S(Mn) = CP(Mn, OM AX(Mn)), where OM AX(Mn) is the maximal operator system structure on Mn. It follows that the operator system associated with the mapping cone S(Mn) is OM AX(Mn), and the cones that define OM AX(Mn) are exactly the cones of separable operators. It was similarly shown that S(Mn) = CP(OM IN (Mn), Mn), where MAPPING CONES ARE OPERATOR SYSTEMS 9 OM IN (Mn) is the minimal operator system structure on Mn. It follows from condition (4) of Corollary 8 (and the fact that S(Mn) = (P(Mn)◦)†) that the operator system associated with the mapping cone P(Mn) is OM IN (Mn), and the cones that define OM IN (Mn) are the cones of block-positive operators. It was shown in [5] that if OM INk(Mn) and OM AXk(Mn) denote the super k-minimal and super k-maximal operator systems on Mn [4], respectively, then we have that Pk(Mn) = CP(Mn, OM INk(Mn)) and Sk(Mn) = CP(Mn, OM AXk(Mn)). Thus the operator systems associated with the mapping cones Pk(Mn) and Sk(Mn) are OM INk(Mn) and OM AXk(Mn), respectively. Finally, consider the mapping cone of completely co-positive maps {Φ◦T : Φ ∈ CP(Mn)}. It is not difficult to see that the associated operator system is the one defined by the cones of operators with positive partial transpose -- i.e., the operators X ∈ Mm ⊗ Mn such that (idm ⊗ T )(X) ≥ 0. We close this section by considering what Corollary 8 says in the case when the mapping cone C is symmetric -- that is, when T ◦ Φ ◦ T ∈ C and Φ† ∈ C whenever Φ ∈ C. The concept of symmetric mapping cones was seen to be important in [12], and it is worth noting that all of the specific mapping cones considered so far, such as the cones of k-positive and completely co-positive maps, are in fact symmetric. It will be useful for us to define a linear operator F ∈ Mn ⊗ Mn by F (v ⊗ w) = w ⊗ v and extending linearly. The operator F is sometimes called the swap or flip operator, and we observe that F = F T . Theorem 9. Let C ⊆ L(Mn) be a convex mapping cone and let O(Mn) be the operator system, defined by cones {Cm}∞ m=1, associated to C via the bijection of Corollary 8. Then C is symmetric if and only if Cn is closed under the transpose map and the map X 7→ F XF . Proof. The proof relies on Lemma 1 which tells us that CT ◦Φ†◦T = F CΦF , and [19, Lemma 4] which tells us that CT ◦Φ◦T = C T Φ F . It then follows immediately that T ◦ Φ ◦ T ∈ C whenever Φ ∈ C if and only if Cn (which equals CC) is closed under the transpose map T . Similarly, Φ† ∈ C whenever Φ ∈ C if and only if Cn is closed under the map X 7→ F X T F . The result follows. (cid:3) Φ . Combining these two results shows that CΦ† = F C T 6. Semigroup Cones as Operator Systems Theorem 6 and Corollary 8 provide characterizations of completely positive maps to and from Mn, and completely positive maps between two different super-homogeneous operator systems on Mn. However, they say nothing about completely positive maps from a super- homogeneous operator system back into itself. Toward deriving a characterization for this situation, we will say that a cone C ⊆ L(Mn) is a semigroup if it is closed under composition -- i.e., if Φ ◦ Ψ ∈ C for all Φ, Ψ ∈ C. Notice that many of the standard examples of mapping cones, such as the k-positive maps and the k-superpositive maps, are semigroups (however, the cone of completely co-positive maps is not). The following proposition is a generalization of the fact that Φ is k-positive if and only if Φ ◦ Ψ is k-superpositive for all k-superpositive Ψ [1, Theorem 3.8]. Note that it is similar to Proposition 5, but by using the fact that C is a semigroup instead of just right-CP-invariant or a mapping cone we are able to show that Φ† ◦ Ψ ∈ C◦, not just that Φ† ◦ Ψ ∈ CP(Mn). Proposition 10. Let C ⊇ CP(Mn) be a closed convex cone semigroup. Then Φ ∈ C if and only if Φ† ◦ Ψ ∈ C◦ for all Ψ ∈ C◦. 10 N. JOHNSTON AND E. STØRMER Proof. To show the "only if" direction, it is enough to show that Tr(CΦ†◦ΨCΩ) ≥ 0 for all Ω ∈ C. To this end, simply note that Tr(CΦ†◦ΨCΩ) = Tr(CΨCΦ◦Ω) ≥ 0, where the final inequality follows from the fact that Φ, Ω ∈ C so Φ ◦ Ω ∈ C. To see the "if" direction, suppose Φ† ◦ Ψ ∈ C◦ for all Ψ ∈ C◦. Then, because idn ∈ CP(Mn) ⊆ C, we have It follows that Φ ∈ C◦◦ = C. (cid:3) 0 ≤ Tr(CΦ†◦ΨE) = Tr(CΨCΦ) ∀ Ψ ∈ C◦. If O(Mn) is an operator system defined by cones {Cm}∞ m=1, then the dual cones {C ◦ m}∞ m=1 define an operator system as well, which we will denote O◦(Mn). For simplicity, we will only consider this operator system as a family of dual cones, in keeping with our focus throughout the preceding portion of the paper, and not the associated dual operator space structure. The interested reader is directed to [20] for a more thorough treatment of dual operator systems. It is easily verified that O(Mn) is super-homogeneous if and only if O◦(Mn) is super-homogeneous, and the "naive" operator system on Mn is easily seen to be self-dual: M ◦ n = Mn. By the duality of the cones of k-positive maps and k-superpositive maps we know that OM IN ◦ k (Mn) = OM AXk(Mn) and OM AX ◦ k (Mn) = OM INk(Mn). We now consider what types of cones can be completely positive from a super-homogeneous operator system back into itself. By using [5, Theorem 5] and the fact that Pk(Mn) is a semigroup, it is not difficult to see that CP(OM INk(Mn)) = Pk(Mn). By using [1, Theorem 3.8] we can similarly see that CP(OM AXk(Mn)) = Pk(Mn), so we can't possibly hope for a uniqueness result as strong as that of Theorem 6 or Corollary 8 in this setting. Nonethe- less, we have the following result, which shows that duality plays a strong role here and the fact that CP(OM INk(Mn)) = CP(OM AXk(Mn)) follows from the duality of OM INk(Mn) and OM AXk(Mn). Furthermore, there is a unique operator system that gives the cone CP(O(Mn)) that is "large enough" to contain (Mn ⊗ Mn)+ as a subset of its nth cone -- in this case it is OM INk(Mn). Theorem 11. Let C ⊆ L(Mn) be a convex cone. The following are equivalent: † (1) C is a semigroup cone with CP(Mn) ⊆ C ⊆ P(Mn). (2) There exists a super-homogeneous operator system O(Mn) such that C = CP(O(Mn)). and O(Mn) can be chosen so that its nth cone Additionally, CP(O◦(Mn)) = CP(O(Mn)) Cn = CC. Furthermore, O(Mn) is unique up to the condition (Mn ⊗ Mn)+ ⊆ Cn. Proof. We first prove that (2) ⇒ (1). Let {Cm}∞ m=1 be the cones associated with the operator system O(Mn). If X ∈ Cm and Φ, Ψ ∈ CP(O(Mn)) then (idm ⊗ Φ)(X) ∈ Cm. But then applying idm ⊗ Ψ shows (idm ⊗ (Ψ ◦ Φ))(X) ∈ Cm as well, so it follows that Ψ ◦ Φ ∈ CP(O(Mn)) and thus CP(O(Mn)) is a semigroup. Because O(Mn) is super-homogeneous, we know that AdB ∈ CP(O(Mn)) for all B ∈ Mn, and so CP(Mn) ⊆ CP(O(Mn)). To see that CP(O(Mn)) ⊆ P(Mn), simply use Lemma 7. To see that (1) ⇒ (2), we argue much as we did in Theorem 6. It is clear, via the Choi-Jamio lkowski isomorphism, that Sn ⊆ CC ⊆ Pn. Now note that C is left- and right- CP-invariant (but perhaps not a mapping cone because it may not be closed) because MAPPING CONES ARE OPERATOR SYSTEMS 11 Φ ◦ Ψ ∈ C for any Φ ∈ C and Ψ ∈ CP(Mn) ⊆ C (and similarly for composition on the left by Ψ ∈ CP(Mn)). Thus, if A ∈ Mn and Φ ∈ C then (AdA ⊗ AdB)(CΦ) = (AdA ⊗ (AdB ◦ Φ))(E) = (idn ⊗ (AdB ◦ Φ ◦ AdAT ))(E) ∈ CC, where the second equality comes from Lemma 1. It follows from Proposition 3 that there exists a super-homogeneous operator system O(Mn), defined by cones {Cm}∞ m=1, such that Cn = CC. Because C is a semigroup, it follows that (idn ⊗ (Φ ◦ Ψ))(E) ∈ Cn for any Φ, Ψ ∈ C. Then (idn ⊗ Φ)(CΨ) ∈ Cn, so (idn ⊗ Φ)(Cn) ⊆ Cn, which implies C ⊆ CP(O(Mn)) by Corollary 4. To see the other inclusion, note that idn ∈ CP(Mn), so idn ∈ C. It follows that (idn ⊗ idn)(E) = E ∈ Cn. Thus, if Φ ∈ CP(O(Mn)) then (idn ⊗ Φ)(E) ∈ Cn = CC, so Φ ∈ C, which implies that C = CP(O(Mn)). To see the claim about CP(O◦(Mn)), suppose that CP(Mn) ⊆ C ⊆ P(Mn) is a closed con- vex cone semigroup. Then for any Φ, Ψ ∈ C◦ ⊆ CP(Mn) and Ω ∈ C we have Tr(CΦ◦ΨCΩ) = Tr(CΨCΦ†◦Ω). We know from Proposition 10 that Φ† ◦ Ω = (Ω† ◦ Φ)† ∈ (C◦)† ⊆ CP(Mn)† = CP(Mn) ⊆ C. It follows that Tr(CΨCΦ†◦Ω) ≥ 0, so Φ ◦ Ψ ∈ C◦, which implies that C◦ is also a semigroup. † † † Now by repeating our argument from earlier, we see from Proposition 3 that there is an operator system on Mn defined by the cone Cn := CC◦ = C ◦ C, and this is the dual operator system O◦(Mn) of the operator system defined by CC. For any Φ ∈ C◦◦, Ψ ∈ C◦, we have (idn ⊗ Φ†)(CΨ) = CΦ†◦Ψ ∈ CC◦ by Proposition 10. It follows via Corollary 4 that ⊆ CP(O◦(Mn)). To see the other inclusion, suppose Φ ∈ CP(O◦(Mn)). Then (C◦◦)† = C Φ ◦ Ψ ∈ C◦ for all Ψ ∈ C◦, so Proposition 10 tells us that Φ ∈ (C◦◦)† = C . It follows that C = CP(O◦(Mn)). Finally, to see the uniqueness condition, suppose that the cones {Dm}∞ m=1 define an operator system O2(Mn) such that C = CP(O(Mn)) = CP(O2(Mn)), where O(Mn) is the operator system with nth cone Cn := CC already introduced. We furthermore require that (Mn ⊗ Mn)+ ⊆ Dn, and in particular that E ∈ Dn. Then (idn ⊗ Φ)(E) ∈ Dn for any Φ ∈ C, so CC ⊆ Dn. On the other hand by the equivalence of (1) and (2), Dn = CC′ for some semigroup cone C′. If Φ ∈ C′ then for any X ∈ Dn there exists Ψ ∈ C′ such that (idn ⊗ Φ)(X) = (idn ⊗ (Φ ◦ Ψ))(E) ∈ Dn, where the inclusion comes from C′ being a semigroup. It follows via Corollary 4 that C′ ⊆ CP(O2(Mn)). Then C ⊆ C′ ⊆ CP(O2(Mn)) = C, so C = C′ and hence Cn = Dn. Uniqueness now follows from Proposition 3. (cid:3) It is worth noting that if C is closed and condition (1) of Theorem 11 holds, then C is necessarily a mapping cone. It follows that if O(Mn) is a super-homogeneous operator system defined by closed cones then CP(O(Mn)) is always a mapping cone (which can also be seen from Corollary 8), although the converse does not hold. That is, there exist mapping cones C such that there is no operator system O(Mn) with C = CP(O(Mn)) -- the simplest example being the mapping cone of completely co-positive maps. Acknowledgements. Thanks are extended to Vern Paulsen for valuable suggestions and comments on an early draft. N.J. was supported by an NSERC Canada Graduate Scholar- ship and the University of Guelph Brock Scholarship. 12 N. JOHNSTON AND E. STØRMER References [1] L. Skowronek, E. Størmer, and K. Math. Phys. 50 (2009), 062106. Zyczkowski, Cones of positive maps and their duality relations, J. [2] E. Størmer, Extension of positive maps into B(H), J. Funct. Anal. 66 (1986), 235 -- 254. [3] V. Paulsen, I. Todorov, M. Tomforde, Operator system structures on ordered spaces, Proc. Lond. Math. Soc. (2010), doi:10.1112/plms/pdq011. [4] B. Xhabli, Universal operator system structures on ordered spaces and their applications, Ph.D. Thesis, University of Houston (2009). [5] N. Johnston, D. W. Kribs, V. I. Paulsen, and R. Pereira, Minimal and maximal operator spaces and operator systems in entanglement theory, J. Funct. Anal. 260 8 (2011), 2407 -- 2423. [6] M. Nielsen and I. Chuang, Quantum computation and quantum information, Cambridge University Press (2000). [7] M.-D. Choi, Completely positive linear maps on complex matrices, Linear Algebra Appl. 10 (1975), 285-290. [8] A. Jamio lkowski, Linear transformations which preserve trace and positive semidefiniteness of operators, Rep. Math. Phys. 3 (1972). [9] R. Rockafellar, Convex analysis, Princeton University Press (1997). [10] C. J. Goh and X. Q. Yang, Duality in optimization and variational inequalities, Taylor & Francis, London (2002). [11] M. Horodecki, P. W. Shor, and M. B. Ruskai, General entanglement breaking channels, Rev. Math. Phys. 15 (2003), 629 -- 641. [12] E. Størmer, Mapping cones of positive maps, Math. Scand. 108 (2011), 223 -- 232. [13] J. Watrous, Theory lecture notes, http://www.cs.uwaterloo.ca/~watrous/lecture-notes.html (2004). of quantum information published electronically at [14] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge University Press, Cambridge (2003). [15] M.-D. Choi and E. G. Effros, Injectivity and operator spaces, J. Funct. Anal. 24 (1977), 156-209. [16] N. Johnston, Characterizing operations preserving separability measures via linear preserver problems, Linear Multilinear Algebra 59 (2011), 1171 -- 1187. [17] L. Gurvits and H. Barnum, Size of the separable neighborhood of the maximally mixed bipartite quantum [18] state, Los Alamos National Laboratory unclassified technical report LAUR (2002) 02-2414. L. Skowronek, Cones with a mapping cone symmetry in the finite-dimensional case, Linear Algebra Appl. 435 (2011), 361 -- 370. [19] E. Størmer, Duality of cones of positive maps, Munster J. Math. 2 (2009), 299 -- 310. [20] D. Blecher and B. Magajna, Dual operator systems, Bull. London Math. Soc. (2010), doi: 10.1112/blms/bdq103. Department of Mathematics and Statistics, University of Guelph, Guelph, Ontario N1G 2W1, Canada E-mail address: [email protected] Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, NO-0316 Oslo, Norway E-mail address: [email protected]
1306.0945	1	1306	2013-06-04T23:53:02	Notes on extremality of the Choi map	[ "math.OA", "quant-ph" ]	It is widely believed that the Choi map generates an extremal ray in the cone $\mathcal P(M_3)$ of all positive linear maps between $C^*$-algebra $M_3$ of all $n\times n$ matrices over the complex field. But the only proven fact is that the Choi map generates the extremal ray in the cone of all positive linear map preserving all real symmetric $3\times 3$ matrices. In this note, we show that the Choi map is indeed extremal in the cone $\mathcal P(M_3)$. We also clarify some misclaims about the correspondence between positive semi-definite biquadratic real forms and postive linear maps, and discuss possible positive linear maps which coincide with the Choi map on symmetric matrices.	math.OA	math	NOTES ON EXTREMALITY OF THE CHOI MAP KIL-CHAN HA Abstract. It is widely believed that the Choi map generates an extremal ray in the cone P(M3) of all positive linear maps between C ∗-algebra M3 of all n × n matrices over the complex field. But the only proven fact is that the Choi map generates the extremal ray in the cone of all positive linear map preserving all real symmetric 3 × 3 matrices. In this note, we show that the Choi map is indeed extremal in the cone P(M3). We also clarify some misclaims about the correspondence between positive semi-definite biquadratic real forms and postive linear maps, and discuss possible positive linear maps which coincide with the Choi map on symmetric matrices. 3 1 0 2 n u J 4 ] . A O h t a m [ 1 v 5 4 9 0 . 6 0 3 1 : v i X r a 1. Introduction Let Mn be the C ∗-algebra of all n × n matrices over the complex field. Because the convex structure of the positive cone P(Mn) of all positive linear maps between Mn is highly complicated even in lower dimensions, it would be very useful to find extreme rays of this cone. Another approach to understand the convex structure of P(Mn) is to considered the possibility of decomposition of P(Mn) into subcones. For example, a positive linear map between matrix algebra is said to be decomposable if it is the sum of a completely positive linear map and a completely copositive linear map. In the sixties, it was shown that every positive linear map in P(M2) is decomposable, and all extreme points of the convex set of unital positive linear maps in P(M2) had been found [10]. The first example of indecomposable positive linear map was given by Choi [1, 2]. This Choi map is defined by Φ(X) =   −x21 −x31 x11 + x33 −x12 x22 + x11 −x13 −x23   , −x32 x33 + x22 where X = (xij) ∈ M3. It is widely believed [8, 7, 6, 9] that the Choi map generates an extremal ray in P(M3). It is not as trivial as one may think. We note that extremality of some variants of Choi map can be confirmed from their exposedness [4, 5]. But the only proven fact on the Choi map is that, for x = (x1, x2, x3)t, y = (y1, y2, y3)t ∈ R3, the corresponding real form ytΦ(xxt)y is extremal in the convex cone of all positive semi-definite biquadratic real forms [3]. Let Sn be the real vector space of all n × n real Date: August 2, 2021. 1991 Mathematics Subject Classification. 15A30, 46L05. Key words and phrases. Positive semi-definite biquadratic real form, Positive linear map, Convex cone, Extremal. partially supported by NRFK 2012-0002600. 1 symmetric matrices in Mn. Then we see that the restriction map ΦS3 of Φ is extremal in the cone P(S3) of all positive linear map between S3. We note that the extremality of ΦS3 in the cone P(S3) does not imply the ex- tremality of Φ in P(M3). To explain this, we consider a positive linear map Ψ1 defined by (1) Ψ1 = 1 2 (Φ + Φ ◦ t). Since Φ(S3) ⊂ S3, we see that Ψ1(S3) ⊂ S3 and Ψ1S3 = ΦS3. Thus the restriction map Ψ1S3 is extremal in P(S3). But Ψ1 is not extremal in P(Mn). The purpose of this note is to clarify this situation. In the next section, we show that the Choi map is indeed extremal in P(M3). In section 3, we explain briefly the correspondence between positive semi-definite real biquadratic forms and positive linear maps, and clarify some misclaims about this correspondence in the literatures. We also discuss possible extensions of ΦS3 to positive linear maps in P(M3). Throughout this note, Mn(R) denotes the real vector space consisting of n × n real matrices, and {Ekℓ} the usual matrix units in Mn. For a n × n matrix A, det(A) denotes the determinant of A, and dk(A) denotes the determinant of the submatrix formed by deleting the k-th row and k-th column of A. 2. Extremality of the Choi map In this section, we show that the Choi map Φ is extremal in P(M3). Suppose that (2) Φ = φ1 + φ2 for φ1, φ2 ∈ P(Mn). For each φk (k = 1, 2), we define two linear maps φk1, φk2 by φk1(X) = 1 2 (φk(X) + φk(X)), φk2(X) = 1 2i (φk(X) − φk(X)), X ∈ M3 where φk(X) denotes the matrix whose entries are conjugates of the corresponding entries of the matrix φk(X). Then, we see that φkℓ(Mn(R)) ⊂ Mn(R) (k = 1, 2 and ℓ = 1, 2) and φk1 is positive linear map for k = 1, 2. Therefore, both φ11 and φ21 are positive linear maps preserving S3. We note that (3) φk = φk1 + iφk2 (k = 1, 2) and φ(S3) ⊂ S3. So we see that φS3 = φ11S3 + φ21S3. 2 Thus, we can conclude that φ11S3 = λΦS3 and φ21S3 = (1 − λ)ΦS3 for some 0 ≤ λ ≤ 1 because Φ is extremal in P(S3). From this, we have that (4) φ11(E11) = λ(E11 + E22), φ21(E11) = (1 − λ)(E11 + E22), φ11(E22) = λ(E22 + E33), φ21(E22) = (1 − λ)(E22 + E33), φ11(E33) = λ(E33 + E11), φ21(E33) = (1 − λ)(E33 + E11), φ11(S12) = −λS12, φ11(S23) = −λS23, φ11(S31) = −λS31, φ21(S12) = −(1 − λ)S12, φ21(S23) = −(1 − λ)S23, φ21(S31) = −(1 − λ)S31, where Skℓ = Ekℓ + Eℓk for 1 ≤ k < ℓ ≤ 3. For each 1 ≤ k < ℓ ≤ 3, we define hermitian matrices Hkℓ = (Ekℓ − Eℓk)i. Then M3 is generated by B = {E11, E22, E33, S12, S13, S23, H12, H13, H23}. We note that a positive linear map in P(M3) is uniquely determined by its value on B. Now, we examine φ1(X) and φ2(X) for each X ∈ B to determine two positive linear maps φ1, φ2. First, we consider the positive semi-definite (PSD) matrices φk(E11), φk(E22) and φk(E33) for k = 1, 2. Lemma 2.1. φ1(Ekk) and φ2(Ekk) are of the following form (5) φ1(E11) =   φ1(E22) =   φ1(E33) =   −a1i λ 0 a1i 0 0 λ 0 0 0 λ 0 0 0 −a2i 0 a2i λ λ 0 0 a3i 0 0 −a3i 0 λ   ,   ,   , φ2(E11) =   φ2(E22) =   φ2(E33) =   1 − λ −a1i 0 1 − λ 0 a1i 0 0 0 0 0 0 0 1 − λ −a2i 0 1 − λ a2i 1 − λ 0 −a3i 0 a3i 0 0 0 1 − λ   ,   ,   for real numbers a1, a2 and a3. Proof. Since (3, 3)-entry of the PSD matrix Φ(E11) is equal to zero, (3, 3)-entries of both φ1(E11) and φ2(E11) are also equal to zero from the equation (2) and the positivity of φk. Again, the positivity of φk(E11) implies that φk(E11) is of the form in equation (5). The rest can be checked similarly. (cid:3) Note that φ1(Skℓ) and φ2(Skℓ) are hermitian matrices. So, all diagonal entries of both φ12(Skℓ) and φ22(Skℓ) are equal to zero for all Skℓ ∈ B by the identity (3) since φ12(Skℓ) and φ22(Skℓ) are real matrices. Thus, from (3) and (4), we see that φ1(Skℓ) 3 and φ2(Skℓ) are of the form φ1(S12) =   φ1(S13) =   φ1(S23) =   (6) (7) 0 −λ + b1i 0 −b3i −λ − b1i −b2i 0 −b4i b4i −λ + b5i 0 −λ − b5i −b6i 0 b7i 0 −b7i −b8i −λ − b9i −λ + b9i 0 b2i b3i 0 b6i 0 b8i   ,   ,   , φ2(S12) =   φ2(S13) =   φ2(S23) =   0 −1 + λ + b1i b2i 0 b4i −1 + λ + b5i −b7i 0 b7i b8i −1 + λ + b9i 0 −1 + λ − b1i −b2i −b3i 0 b3i −b4i −1 + λ − b5i 0 b6i −b6i 0 −b8i −1 + λ − b9i 0 0     ,   for real numbers b1, b2, · · · , b9. Now, for any x = (x1, x2, x3)t ∈ C3, we define a rank one PSD matrix (8) X[x1, x2, x3] := xx∗ = (xkxℓ). From the positivity of φk(X[x1, x2, x3]), we can correlate the variables bk's in (6) with the variables aℓ's in (5). Lemma 2.2. Let b1, b2, · · · , b9 be the variables in (6), and a1, a2, a3 be the variables in (5). Then we have (9) b1 = 0, b2 = −a2, b3 = −a3, b4 = a2, b5 = 0, b6 = a1, b7 = −a3, b8 = −a1, b9 = 0. Proof. For a real number t, we consider two PSD matrices φk(X[1, t, 0]) = φ1(E11) + tφ1(S12) + t2φ1(E22) (k = 1, 2). We know that the principal minors dk (φℓ(X[1, t, 0]))'s are nonnegative. In particular, we have d3 (φ1(X[1, t, 0])) = −b2 1t2 − 2a1b1t + (λ2 − a2 1) ≥ 0 for all real number t. Thus we get b1 = 0. Similarily, by considering the principal minors d2 (φ1(X[1, 0, t])) and d1 (φ1(X[0, 1, t])), we can show that b5 = 0 and b9 = 0 respectively. 4 We see that determinants det (φk(X[1, t, 0])) (k = 1, 2) are quartic polynomial in t and divisible by t2. Therefore, the coefficients of t4 in det (φk(X[1, t, 0])) (k = 1, 2) should be nonnegative. So, we have −λ(a2 + b2)2 ≥ 0 and − (1 − λ)(a2 + b2)2 ≥ 0. Therefore, we can conclude that b2 = −a2. By the same method, we can show that b7 = −a3 by considering two quartic polynomials det (φk(X[0, 1, t])) (k = 1, 2) in variable t. Two quartic polynomials det (φℓ(X[1, 0, t])) in t are also divisible by t2. When b5 = 0, the coefficients of t2 are −λ(a1 − b6)2 and − (1 − λ)(a1 − b6)2. Therefore, we get b6 = a1. Up to now, we have shown that b1 = b5 = b9 = 0, b2 = −a2, b6 = a1 and b7 = −a3. From the equations (5), (6) and (7), we can compute that det (φ1(X[1, 1, 1])) + det (φ1(X[1, 1, −1])) + det (φ1(X[1, −1, 1])) det (φ2(X[1, 1, 1])) + det (φ2(X[1, 1, −1])) + det (φ2(X[1, −1, 1])) + det (φ1(X[1, −1, −1])) = −8λ(cid:0)(a1 + b8)2 + (a2 − b4)2 + (a3 + b3)2(cid:1) , + det (φ2(X[1, −1, −1])) = −8(1 − λ)(cid:0)(a1 + b8)2 + (a2 − b4)2 + (a3 + b3)2(cid:1) . Since the above two values must be nonnegative, we can conclude that b3 = −a3, b4 = a2 and b8 = −a1. This completes the proof. (cid:3) Now, we consider hermitian matrices φ1(Hkℓ) and φ2(Hkℓ) for Hkℓ ∈ B. For real numbers ci's and complex numbers αi's, we may write φk(H12) (k = 1, 2) as φ1(H12) =   c1 α1 α2 c2 α3 α1 α2 α3 c3   , φ2(H12) =   −c1 −i − α1 −α2 −α3 i − α1 −α2 −c3 −c2 −α3   . We consider PSD matrices φk(X[1, −ti, 0]) = φk(E11) + tφk(H12) + t2φk(E22) (k = 1, 2). We see that d3 (φk(X[1, −ti, 0])) (k = 1, 2) are cubic polynomial in t, and the coeffi- cients of t3 are Therefore, we have c1 = 0. Then we have the following: λc1 and (λ − 1)c1. d3 (φ1(X[1, −ti, 0])) =λ2 − a2 d3 (φ2(X[1, −ti, 0])) =(λ − 1)2 − a2 1 + [λc2 − 2a1Im(α1)]t + (λ2 − α12)t2 ≥ 0, 1 + [(λ − 1)c2 − 2a1(1 + Im(α1))]t + [(λ − 1)2 − Re(α1)2 − (1 + Im(α1))2]t2 ≥ 0, 5 for all t ∈ R, where α1 = Re(α1) + Im(α1)i. So we get the conditions (10) (11) λ2 ≥ α12 (λ − 1)2 ≥ Re(α1)2 + (1 + Im(α1))2 Since 0 ≤ λ ≤ 1, we see that Im(α1) < 0 by (11). Then, we get that −λ ≤ Im(α1) < 0 from the condition (10), that is, 0 ≤ 1 − λ ≤ 1 + Im(α1) < 1. Therefore, we have that (1 − λ)2 ≤ (1 + Im(α1))2 ≤ Re(α1)2 + (1 + Im(α1))2 ≤ (1 − λ)2. Consequently, we conclue α1 = −λi. We note that this implies the coefficient of t in d3 (φ1(X[1, −ti, 0])) should be zero. Therefore, we get c2 = −2a1. When c1 = 0, d2 (φk(X[1, −ti, 0])) (k = 1, 2) are quadratic polynomials in t divisible by t. Therefore, the coefficients of t should be zero. From this observation, we have that λc3 = 0 and (λ − 1)c3 = 0. Consequently, we see that c3 = 0. Finally, we show that α2 = a2 by considering the determinant of φk(X[1, −ti, 0]). Under the conditions c1 = 0 and α1 = −λi, we can show that the determinants det (φk(X[1, −ti, 0])) (k = 1, 2) are quartic polynomials in t, and the coefficients of t4 are −λα2 − a22 and − (1 − λ)α2 − a22. Since both coefficients should be nonnegative, we get α2 = a2. To sum up, we have correlated all entries of φk(H12) except α3 with ai in (5). That is, c1 = c3 = 0, c2 = −2a1, α1 = −λi and α2 = a2. Lemma 2.3. Let a1, a2, a3 be the real variables in (6). Then φ1(Hkℓ) and φ2(Hkℓ) are of the following forms (12) φ1(H12) =   φ1(H13) =   φ1(H23) =   0 −λi a2 λi −2a1 α 0 a2 −2a3 α β λi −a1 0 −a3 β −λi 0 −a1 −a3 γ −λi 0 λi −2a2   , 0 γ φ2(H12) =     , φ2(H13) =     , φ2(H23) =   0 (1 − λ)i −a2 2a3 −β (1 − λ)i (λ − 1)i −a2 −α 0 2a1 −α −β (λ − 1)i 0 a1 a1 0 a3 0 0 a3 −γ (1 − λ)i (λ − 1)i −γ 2a2   ,   ,   . Proof. For the case of φk(H12), we have done it with α3 = α. Through the same process with principal minors d3 (φk(X[0, 1, −ti])) , d1 (φk(X[0, 1, −ti])) and the determinant det (φk(X[0, 1, −ti])), we can show that φk(H23) is of the form in (12) for k = 1, 2. 6 For the case of φk(H13), it suffices to consider d3 (φk(X)) , d2 (φk(X)) and then det (φk(X)) with X = X[1, 0, −ti]. But, in this case, each d2 (φk(X)) is quartic poly- nomial in t divisible by t2, and we can determine (1, 3) and (1, 1) entries of φk(H13) by examining the coefficients of t2 as in (10) and (11). (cid:3) Now, we are ready to prove that the Choi map is indeed extremal in the cone P(M3). We note that a positive linear map ψ satisfies ψ(Ekℓ) = 1 2 (ψ(Skℓ) − iψ(Hkℓ)) , ψ(Eℓk) = 1 2 (ψ(Skℓ) + iψ(Hkℓ)) for 1 ≤ k < ℓ ≤ 3. Therefore, positive linear maps φ1 and φ2 in (2) are uniquely determined by the Lemma 2.1, 2.2 and 2.3. For the convenience of readers, we make up a list of entries of φ1(X) and φ2(X) for 3 × 3 matrix X = (xkℓ). In the following list, [A]kℓ denotes the entry in the k-th row and ℓ-th column of a matrix A. (13) [φ1(X)]12 = −x12λ + a1x11i − a3x32i + [φ1(X)]21 = −x21λ − a1x11i + a3x23i − [φ1(X)]13 = −x13λ + a3x33i − a2x12i − [φ1(X)]11 = (x11 + x33)λ + a3(x13 − x31)i, 1 2 1 2 1 2 [φ1(X)]22 = (x22 + x11)λ + a1(x12 − x21)i, 1 2 1 2 1 2 [φ1(X)]33 = (x33 + x22)λ + a2(x23 − x32)i, [φ1(X)]32 = −x32λ − a1x31i − a2x22i + [φ1(X)]23 = −x23λ + a1x13i + a2x22i − [φ1(X)]31 = −x31λ − a3x33i + a2x21i + (a2 − β)x13i + (a1 + γ)x23i − (a2 − β)x31i − (a3 + α)x12i − (a1 + γ)x32i + (a3 + α)x21i + 1 2 1 2 1 2 1 2 1 2 1 2 (a2 + β)x31i, (a1 − γ)x32i, (a2 + β)x13i, (a3 − α)x21i, (a1 − γ)x23i, (a3 − α)x12i, [φ2(X)]21 = −x21(1 − λ) + a1x11i − a3x23i + [φ2(X)]13 = −x13(1 − λ) − a3x33i + a2x12i + [φ2(X)]12 = −x12(1 − λ) − a1x11i + a3x32i − [φ2(X)]11 = (x11 + x33)(1 − λ) − a3(x13 − x31)i, 1 2 1 2 1 2 [φ2(X)]22 = (x22 + x11)(1 − λ) − a1(x12 − x21)i, 1 2 1 2 1 2 [φ1(X)]33 = (x33 + x22)(1 − λ) − a2(x23 − x32)i [φ2(X)]31 = −x31(1 − λ) + a3x33i − a2x21i − [φ2(X)]23 = −x23(1 − λ) − a1x13i − a2x22i + [φ2(X)]32 = −x32(1 − λ) + a1x31i + a2x22i − (a2 − β)x13i − (a1 + γ)x23i + (a2 − β)x31i + (a3 + α)x12i + (a1 + γ)x32i − (a3 + α)x21i − 1 2 1 2 1 2 1 2 1 2 1 2 (a2 + β)x31i, (a1 − γ)x32i, (a2 + β)x13i, (a3 − α)x21i, (a1 − γ)x23i, (a3 − α)x12i, where a1, a2, a3 ∈ R, α, β, γ ∈ C and 0 ≤ λ ≤ 1. 7 First, we determine real variables a1, a2 and a3. To do this, we consider principal minors d3 (φk(X[1, t, si])) (k = 1, 2) for real numbers t and s, . They are quadratic polynomials in t with the following coefficients of t2. 4λa3s + (λ2 − a2 3)s2, 4(λ − 1)a3s +(cid:0)(λ − 1)2 − a2 3(cid:1) s2 These coefficients should be nonnegative for all s because principal minors of a PSD matrix are nonnegative. Therefore, we have a3 = 0. And then, by considering d2 (φk(X[t, 1, si])) as a polynomial in t for each k = 1, 2, we can show that a2 = 0 similarly. Finally, we get a1 = 0 from the condition that the coefficients of t2 in qua- dratic polynomials d1(φk(X[si, 1, t])) (k = 1, 2) should be nonnegative. Consequently, we have (14) a1 = a2 = a3 = 0. Now, we show that all complex variables α, β and γ should be zero when a1 = a2 = a3 = 0 in (13). (Case 1: λ = 0) If λ = 0, then we have PSD matrices φ1(X[1, 1, i]) =   0 −β −γ −β 0 −γ 0 0 0   , φ1(X[1, i, 1]) =   0 0 γ −α 0 γ 0 −α 0   from (13). Therefore, we have α = β = γ = 0 when λ = 0. In this case, we conclude that φ2 is the Choi map Φ and φ1 is the zero map. (Case 2: λ = 1) In this case, by considering PSD matrices φ2(X[1, 1, i]) and φ2(X[1, i, 1]), we can show that φ1 is the Choi map Φ and φ2 is the zero map as in (Case 1). (Case 3: 0 < λ < 1) Since the determinant of a PSD is nonnegative, we have the following inequalities: (15) = − 4λ(α2 + β2) + 12λ2Im(β) ≥ 0, det(cid:0)φ1(X[1, e det(cid:0)φ2(X[1, e π 2 i, e π 2 i, e π 2 i])(cid:1) + det(cid:0)φ1(X[1, e− 2 i])(cid:1) + det(cid:0)φ2(X[1, e− π π 2 i, e− π 2 i, e− π 2 i])(cid:1) 2 i])(cid:1) π = − 4(1 − λ)(α2 + β2) − 12(1 − λ)2Im(β) ≥ 0, These are equivalent to the inequalities 0 ≤ α2 + β2 3λ ≤ Im(β) ≤ − α2 + β2 3(1 − λ) ≤ 0 because of 0 < λ < 1. Therefore, we see Im(β) = 0, and so α = β = 0. Finally, we show that γ = 0 when a1 = a2 = a3 = 0 and α = β = 0 in (13). For a PSD matrix X[1, 1, i], we get two inequalities det (φ1(X[1, 1, i])) = −2λγ2 + 6λ2Im(γ) ≥ 0, det (φ2(X[1, 1, i])) = −2(1 − λ)γ2 − 6(1 − λ)2Im(γ) ≥ 0. 8 Then we see that 0 ≤ γ2 3λ ≤ Im(γ) ≤ − γ2 3(1 − λ) ≤ 0. Thus, we get γ = 0. Consequently, we have shown that φ1 = λΦ and φ2 = (1 − λ)Φ. In any cases, we have the following. Theorem 2.4. The Choi map generates an extreme ray in the cone P(M3). 3. Correspondence between positive semi-definite biquadratic real forms and positive linear maps We note [2] that the Choi map Φ is originated from a real biquadratic form B(x, y) with the relation B(x, y) =(x2 1 + x2 3)y2 1 + (x2 2 + x2 1)y2 2 + (x2 3 + x2 2)y2 3 − 2(x1x2y1y2 + x2x3y2y3 + x3x1y3y1) =ytΦ(xxt)y for real column vector x = (x1, x2, x3)t, y = (y1, y2, y3)t ∈ R3. In general, for any positive real linear map φ : Sn → Sn, we get the correspond- ing positive semi-definite real biquadratic form Bφ(x, y) with x, y ∈ Rn defined by Bφ(x, y) = ytφ(xxt)y. On the other hand, let B(x, y) be a positive sem-definite real biquadratic form with x, y ∈ Rn. Then, for any fixed x ∈ Rn, B(x, y) is a positive semi-definite real quadratic form with respect to y ∈ Rn. So we can write B(x, y) in the form ytS(x)y for some n × n symmetric matrix S(x). Thus we get a map which take any one dimensional projection xxt to S(x). Consequently, we get the corresponding positive linear map φ : Sn → Sn by linearity, which satisfy the relation Bφ(x, y) = B(x, y). Therefore, we see that there is a one-to-one correspondence between the set of positive semi-definite real biquadratic forms and the set P(Sn) consisting of positive real linear maps between Sn. We can find some misclaims on the above correspondence in the literatures [7, 9]. They claim that (i) Let Ψ be a positive complex linear map with Ψ(Mn(R)) ⊂ Mn(R). If the corresponding real biquadratic form BΨ(x, y) is extremal in the cone of all positive semi-definte real biquadratic forms, then Ψ is extremal in the cone P(Mn). (ii) Using linearity and hermicity, the above correspondence can be extended to the correspondence between the set of positive semi-definite real biquadratic forms and the set P(Mn) trivially. For the first claim, we have shown there exists a counter example Ψ1 in (1). Here, we clarify the claim (ii). 9 As before, we define n × n symmetric matrices Skℓ and antisymmetric matrices Akℓ for 1 ≤ k < ℓ ≤ n by Skℓ = Ekℓ + Eℓk, Akℓ = Ekℓ − Eℓk. We note that Sn is generated by G = {Ekk : 1 ≤ k ≤ n} ∪ {Skℓ : 1 ≤ k < ℓ ≤ n} and real matrix algebra Mn(R) is generated by G ∪ {Akℓ : 1 ≤ k < ℓ ≤ n}. Thus, a map φ ∈ P(Sn) can be extended to a positive linear map eφ in P(Mn(R)) by defining eφ(Akℓ) for 1 ≤ k < ℓ ≤ n. It is worthy to note that the positivity of eφ is not affected by eφ(Akℓ)'s. That is, the positivity of eφ is determined by eφSn = φ. On the other hand, we can uniquely extend eφ ∈ P(Mn(R)) to the complex linear map eφ between Mn by the linearity eφ(X + iY ) = eφ(X) + ieφ(Y ) for X, Y ∈ Mn(R). But, in this extension, the positivity of the complex linear map eφ is not determined by the positivity of the real linear map eφ. In general, let Ψ be a positive linear map in P(Mn) with Ψ(Mn(R)) ⊂ Mn(R). So, we can regard Ψ as an extension of real linear map. Then, it is well known that Ψ preserve hermitian matrices. That is, Ψ(iAkℓ) must be a hermitian matrix. Since any hermitian matrix H ∈ Mn can be written by H = S+iA with a symmetric matrix S ∈ Mn(R) and an antisymmetric matrix A ∈ Mn(R), we observe the following. Proposition 3.1. Let Ψ be a positive linear map in P(Mn) with Ψ(Mn(R)) ⊂ Mn(R). Then Ψ preserve hermitian matrices if and only if Ψ preserve symmetric matrices and antisymmetric matrices respectively. Proof. For a symmetric matrix S and an antisymmetric matrix A, we have Ψ(S + iA)∗ = Ψ(S)t − iΨ(A)t = Ψ(S) + iΨ(A) = Ψ(S + iA) since Ψ preserve hermitian matrices and Ψ(Mn(R)) ⊂ Mn(R). Therefore, Ψ(S)t = Ψ(S) and Ψ(A)t = −Ψ(A). This completes the proof. (cid:3) We note that there exists non-positive linear map Ψ2 between M3, which satisfies the following conditions • Ψ2 preserves hermitian matrices. • Ψ2M3(R) is a positive linear map between M3(R) and Ψ2S3 = ΦS3 for the Choi map Φ. From the condition Ψ1S3 = ΦS3, Ψ2 is determined by the values Ψ2(Akℓ). We put Ψ2(A12) = −A12, Ψ2(A13) = −A13, Ψ2(A23) = −A12. Then, Ψ2 is defined by (16) Ψ2(X) =   x11 + x33 −x21 + 1 2(x23 − x32) −x31 −x12 − 1 2 (x23 − x32) x11 + x22 2(x23 + x32) − 1 10 −x13 − 1 2(x23 + x32) x22 + x33   for X = (xkℓ) ∈ Mn. We know that Ψ2M3(R) is a positive linear map between M3(R) and Ψ2S3 = ΦS3. But, this map is not positive map between M3 because det (Ψ2(X)) = −25 for a PSD matrix X = X[1, 2 − i, −1 − i] in (8). Now, we give an example of extremal positive linear map Ψ3 in P(M3), which is not equal to the Choi map Φ but Ψ3S3 = ΦS3. This example explains the claim (ii) is nonsense, and we can conclude that extremal extension of ΦS3 is not unique. We define Ψ3 by Ψ3(X) = Φ(X) for all X ∈ {E11, E22, E33, S12, S13, S23, A12, A13}, Ψ3(A23) = A23. Then, we obtain a complex linear map Ψ3 preserving hermitian matrices by Proposi- tion 3.1. We will show that this map is indeed positive. We recall that the positivity of the Choi map is easily proven through the positivity of the real map ΦS3 with the following relation: Φ(X[x1, x2, x3]) = V Φ(X[r1, r2, r3])V ∗ with V =   eiθ1 0 0 0 eiθ2 0 0 0 eiθ3   where xi = rieiθi for i = 1, 2, 3. But, this method is not applicable to general cases. In fact, it is easy to show that there exists no matrix V satisfying Ψ3(X) = V Ψ3(Y )V ∗ for rank one PSD matrices X = X[r1eiθ1, r2eiθ2, r3eiθ3] ∈ M3 and Y = X[r1, r2, r3] ∈ M3(R). So, we will show that Ψ3(X) is a PSD matrix for any rank one PSD matrix X. Since Φ is positive, we see that dk (Ψ3(X)) = dk (Φ(X)) ≥ 0 for any rank one PSD matrix X = X[x1, x2, x3] and k = 1, 2, 3. By the arithmetic mean-geometric mean inequality, we also see that det (Ψ3(X)) =x12x34 + x22x4 1 + x32x24 − x12x22x32 − 2x12Re(x2 2x2 3) ≥2x12(cid:0)x22x32 − Re(x2 3)(cid:1) ≥ 0. 2x2 Therefore, Ψ3(X) is a PSD matrix for any rank one PSD matrix X = X[x1, x2, x3]. That is, Ψ3 is a positive linear map in P(M3). The extremality of Ψ3 in P(M3) is similarly checked as that of the Choi map. Finally, we discuss extremality for various convex cones of positive linear maps. Let Ψ be a positive linear map preserving Mn(R) in P(Mn). Then ΨMn(R) belongs to P(Mn(R)). Furthermore, we see that Ψ(Sn) ⊂ Sn from the positivity of Ψ and Ψ(Mn(R)) ⊂ Mn(R), that is, ΨSn ∈ P(Sn). Now, let PR(Mn) be the set of all positive complex linear maps preserving Mn(R). Then PR(Mn) is a convex subcone contained in P(Mn). Since any real linear map between Mn(R) can be uniquely extended to the complex linear map between Mn, we may think PR(Mn) = P(Mn) ∩ P(Mn(R)). So, we 11 may consider extremality of a map in each convex cones P(Mn), PR(Mn), P(Mn(R)) and P(Sn). We note that the Choi map is extremal in each of these convex cones. Owing to the map Ψ1 in (1), we know that the extremality of ΨSn in P(Sn) implies neither the extremality of ΨMn(R) in P(Mn(R)) nor the extremality of Ψ in P(Mn). On the other hand, for Ψ ∈ PR(Mn), it is easy to see that the extremality of ΨMn(R) in P(Mn(R)) implies the extremality of Ψ in PR(Mn). But, it seems that the converse does not hold since there exists a non-positive complex linear map Ψ satisfying two properties Ψ(Mn(R)) ⊂ Mn(R) and ΨMn(R) ∈ P(Mn(R)) as a map Ψ2 in (16). So, it is interesting to find an extremal positive linear map Ψ in PR(Mn) whose restriction ΨMn(R) is not extremal in P(Mn(R)). References [1] M.-D. Choi, Positive semidefinite biquadratic forms, Linear Algebra and Appl. 12 (1975), 95 -- 100. [2] M.-D. Choi, Some assorted inequalties for positive linear maps on C ∗-algebras, J. Operator The- ory 4 (1980), 271-285. [3] M.-D. Choi, T.-Y. Lam, Extremal positive semidefinite forms, Math. Ann. 231 (1977), 1 -- 18. [4] K.-C. Ha, S.-H. Kye, Entanglement witnesses arising from exposed positive linear maps, Open Syst. Inf. Dyn. 18 (2011), 323 -- 337. [5] K.-C. Ha, S.-H. Kye, Exposedness of Choi type entanglement witnesses and applications to lengths of separable states, arXiv:1211.5675 (2013). [6] H.-J. Kim, S.-H. Kye, Indecomposable extreme positive linear maps in matrix algebras, Bull. London Math. Soc. 26 (1994), 575 -- 581. [7] H. Osaka, A class of extremal positive maps in 3 × 3 matrix algebras, Publ. Res. Inst. Math. Sci. 28 (1992), 747 -- 756. [8] A. G. Robertson, Positive projections on C ∗-algebras and an extremal positive map, J. London Math. Soc. (2) 32 (1985), 133 -- 140. [9] R. Sengupta, Arvind, Extremal extensions of entanglement witnesses: Finding new bound entan- gled states, Phys. Rev. A 84 (2011), 032328 [10] E. Størmer, Positive linear maps of operator algebras, Acta Math. 110 (1963), 233 -- 278. Faculty of Mathematics and Applied Statistics, Sejong University, Seoul 143-747, Korea 12
1612.05127	1	1612	2016-12-15T16:18:54	Graph products and the absence of property (AR)	[ "math.OA" ]	We discuss the internal structure of graph products of right LCM semigroups and prove that there is an abundance of examples without property (AR). Thereby we provide the first examples of right LCM semigroups lacking this seemingly common feature. The results are particularly sharp for right-angled Artin monoids.	math.OA	math	GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) NICOLAI STAMMEIER Abstract. We discuss the internal structure of graph products of right LCM semi- groups and prove that there is an abundance of examples without property (AR). Thereby we provide the first examples of right LCM semigroups lacking this seemingly common feature. The results are particularly sharp for right-angled Artin monoids. 1. Introduction The starting point of a number of recent breakthroughs in the theory of semigroup C ∗-algebras is the seminal work [Li12, Li13], in which a universal C ∗-algebra C ∗(S) is associated to every left cancellative monoid S. In the last years, a particular line of research focused on left cancellative monoids for which the intersection of two principal right ideals is either empty, or another principal right ideal again. Such monoids are called right LCM semigroups, and they form an intriguing and tractable class of examples in between positive cones in quasi-lattice ordered groups and general left cancellative monoids, see [BLS17, Lemma 3.3 and Corollary 3.6] for details. Inspired by the treatment of the quasi-lattice ordered case in [CL07], a boundary quotient Q(S) of C ∗(S) was introduced for right LCM semigroups S in [BRRW14]. Soon thereafter, Starling provided an in-depth analysis of Q(S) in [Star15], relying on major advances in the understanding of the connections between inverse semigroups, groupoids, and C ∗-algebras stemming from [EP17, EP16, Ste16]. In [BS16], it was shown that the boundary quotient has a more accessible presentation if the right LCM semigroup has the so-called accurate refinement property, henceforth abbreviated property (AR). This property is an analogue of 0-dimensionality for topological spaces in the context of semigroups, and is enjoyed by various examples, see [BS16, Section 2 and Corollary 3.11]. The presence of property (AR) was found to be useful in the construction of a bound- ary quotient diagram for right LCM semigroups in the spirit of [BaHLR12], see [Sta]. This diagram sets the grounds for a unifying approach to the study of equilibrium states on C ∗-algebras in [ABLS], where remarkable results on the structure of KMS-states on C ∗(S) were obtained for right LCM semigroups satisfying an admissibility condition which implies property (AR), see Subsection 2.1. Working with abstract right LCM semigroups as opposed to explicit classes of examples allowed for a unification of the in- spiring case studies [LR10, BaHLR12, LRR11, LRRW14, CaHR16], and also for coverage of a substantial amount of new examples, most notably, algebraic dynamical systems. Moreover, the techniques in [BS16, Sta, ABLS] raise several questions on the structure of right LCM semigroups, perhaps most notably: (a) Are there right LCM semigroups without property (AR)? 2010 Mathematics Subject Classification. 20M10 (Primary) 20F36, 46L05 (Secondary). The author was supported by RCN through FRIPRO 240362. 1 2 NICOLAI STAMMEIER (b) Which right LCM semigroups are admissible? The aim of the present work is to investigate in how far graph products of right LCM semigroups as considered in [VdC01, FK09] provide answers to these two questions. In addition, we also address structural aspects related to the distinguished subsemigroups S∗, Sc and Sci. We apply our results to the classical case of right-angled Artin monoids A+ Γ given by an undirected graph Γ since many graph related phenomena can already be witnessed here. Indeed, the explicit presentation of the boundary quotient in [CL07, Corollary 8.5] involving only the vertex sets of the finite coconnected components of the graph Γ may be regarded as an indication for a particularly accessible structure of foundation sets. Another motivation comes from the elegant solution to the isomorphism problem for C ∗(A+ Γ ), see [ELR16]. Since property (AR) is known for various kinds of right LCM semigroups, we were struck by surprise to find that a right-angled Artin monoid A+ Γ has property (AR) if and only if all of its finitely generated direct summands are free, see Corollary 4.6. In terms of the Γ, this means that all finite coconnected components Γi do not contain any edges. The result follows from more general graph product considerations in Corollary 4.5 that rely on Theorem 4.3, where we show that graph products over infinite coconnected graphs have no foundation set other than the obvious ones containing an invertible element, while the analogous statement holds in the finite case for accurate foundation sets. The characterisation of property (AR) for right-angled Artin monoids A+ Γ in Corol- lary 4.6 allows us to determine when A+ Γ is admissible in the sense of [ABLS]. It turns out that admissibility and the existence of a generalised scale coincide for right-angled Artin monoids, see Corollary 4.9 and Corollary 4.10. If existent, the generalised scale on A+ Γ is unique and arises as the product of the unique generalised scales on its non-abelian direct summands A+ Γi, see Proposition 4.8 and Corollary 4.9. Thus we are lead to the conclusion that graph products of right LCM semigroups mostly lack property (AR), and are therefore not admissible in the sense of [ABLS]. While this rules out the possibility of applying [ABLS] to graph products of right LCM semigroups in great generality, we obtain a fairly detailed description of the behaviour of graph products with respect to the subsemigroups Sc and Sci, see Theorem 3.4. These result show that the graph product represents a useful tool to construct new, and potentially very interesting examples of right LCM semigroups that are well-behaved to some degree, but demand more sophisticated techniques then those applicable to right LCM semigroups that have property (AR) or even a generalised scale. That is why we feel that this work might stimulate further research in the direction of inverse semigroups and groupoids related to (right LCM) semigroups and their C ∗-algebras. Acknowledgements: The author thanks Nathan Brownlowe, Nadia Larsen, and Adam Sørensen for helpful conversations. 2. Preliminaries Here we provide the prerequisites we shall need concerning right LCM semigroups and graph products. 2.1. Right LCM semigroups. A left cancellative semigroup S is called right LCM if the intersection of two principal right ideals in S is either empty or a principal right GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 3 ideal. For s, t ∈ S, we say that s and t are orthogonal and write s ⊥ t if sS ∩ tS = ∅. Unless specified otherwise, we will always assume that a right LCM semigroup S has an identity, i.e. S is a monoid. Let us first discuss property (AR). A finite subset F ⊂ S is called a foundation set for S if for every s ∈ S there is f ∈ F such that f 6⊥ s, see [BRRW14, Section 5]. A subset F ⊂ S is accurate if f ⊥ f ′ for all f, f ′ ∈ F, f 6= f ′, see [BS16, Definition 2.1]. If F, F ′ are foundation sets such that F ′ ⊂ F S, then F ′ is called a refinement of F . We then say that S has the accurate refinement property, or property (AR), if every foundation set for S has an accurate refinement, see [BS16, Definition 2.3]. For a right LCM semigroup S, its subgroup of invertible elements shall be denoted by S∗. This subgroup lies inside the core subsemigroup Sc := {a ∈ S a 6⊥ s for all s ∈ S}, which was first considered for right LCM semigroups in [Star15], but stems from [CL07, Definition 5.4]. We remark that Sc is again a right LCM semigroup. Furthermore, it induces an equivalence relation s ∼ t :⇔ sa = tb for some a, b ∈ Sc called the core relation. In contrast to Sc, we also consider the subsemigroup Sci of core irreducible elements, that is, the collection of all elements s ∈ S \ Sc for which every factorization s = ta with t ∈ S, a ∈ Sc satisfies a ∈ S∗. While Sci does not have an identity by construction, its unitisation S1 ci := Sci ∪ {1} and S′ ci := Sci ∪ S∗ do. A right LCM semigroup S is called core factorable if S = S1 ciSc. We say that Sci ⊂ S is ∩-closed if sS ∩ tS = rS implies r ∈ Sci whenever s, t ∈ Sci. To provide some indication why this property is of interest, let us mention that Sci ⊂ S is ∩-closed if and only if S′ ci is right LCM and its inclusion into S is a homomorphism of right LCM semigroups, i.e. it preserves intersections of principal right ideals, see [ABLS, Proposition 3.3]. Finally, a nontrivial homomorphism N : S → N× is called a generalised scale if N −1(n)/∼ = n and every minimal complete set of representatives for N −1(n)/∼ forms an accurate foundation set for S for all n ∈ N(S). Every generalised scale N satisfies ker N = Sc by [ABLS, Proposition 3.6(i)], and the existence of a generalised scale entails vital information on the structure of S. For instance, it implies that the right LCM semigroup has property (AR), see [ABLS, Proposition 3.6(v)]. Finally, we recall from [ABLS, Definition 3.1] that a right LCM semigroup S is called admissible, if it is core factorable, Sci ⊂ S is ∩-closed, and S admits a generalised scale N such that N(S) ⊂ N× is freely generated by its irreducible elements. 2.2. Graph products. Within this work, a graph will mean a countable, undirected graph Γ = (V, E) without loops or multiple edges. The concept of a graph product of groups emerged in [Gre90] as a generalization of graph groups, and has been transferred to the setting of monoids in [VdC01]: For a graph Γ = (V, E) and a family of monoids (Sv)v∈V , the graph product is the monoid SΓ obtained as the quotient of the direct sum Lv∈V Sv by the congruence generated by the relation (st, ts) if s ∈ Sv, t ∈ Sw with (v, w) ∈ E, see [VdC01, Section 2] and [FK09, Section 1]. Given a graph Γ, its right- angled Artin monoid A+ Γ is the graph product with Sv = N for all v ∈ V . These monoids have also been studied under the names of graph monoids, free partially commutative monoids, and trace monoids, see for instance [Die90]. If one switches the vertex monoids from the natural numbers to the integers, the resulting graph product is the right-angled Artin group AΓ associated to Γ, see [Cha07] for more. 4 NICOLAI STAMMEIER It was shown in [CL02] that the graph product is well-behaved with respect to quasi- lattice orders. Invoking a characterization of the right LCM property via the inverse hull semigroup, Fountain and Kambites showed that this can be generalised to right LCM semigroups, see [FK09, Theorem 2.6], where we note that we can move back and forth between right cancellative, left LCM semigroups (used in [FK09]) and left cancellative, right LCM semigroups by passing to the opposite semigroup. According to [FK09, Theorem 1.1], which is an adaptation of the corresponding result in [Gre90], every element s in a graph product SΓ is represented by an essentially unique reduced expression sv(1)sv(2) · · · sv(n), that is, sv(k) ∈ Sv(k), v(k) 6= v(k + 1), and whenever there are 1 ≤ k < m ≤ n such that v(k) = v(m), then there exists k < ℓ < m such that (v(k), v(ℓ)) /∈ E. The analogous result had been proven in the quasi-lattice ordered case before, see [CL02, Theorem 2]. The reduced expression is unique in the sense that any two reduced expressions for the same element are shuffle equivalent, i.e. we can move from one to the other by a finte number of switches of neighbouring factors whose vertices are adjacent in Γ. Thus there exists a subadditive function ℓ : SΓ → N that assigns the length of any reduced expression to the element in question. A graph Γ is said to be coconnected if there exists no partition V = V1 ⊔ V2 with Vi 6= ∅ and V1 × V2 ⊂ E. Equivalently, Γ is coconnected if the opposite graph Γopp := (V, V × V \ (E ∪ {(v, v) v ∈ V })) is connected. The decomposition of Γ into its coconnected components is the initial step in the analysis of SΓ, see for instance [ELR16], where the synonym co-irreducible is used. Every graph Γ has a unique decomposition into coconnected components, which we denote by (Γi)i∈I with Γi = (Vi, Ei). The original graph can be recovered from (Γi)i∈I as V = Fi∈I Vi and E = {(v, w) ∈ V × V (v, w) ∈ Ei or w /∈ Vi ∋ v for some i ∈ I}. It follows from this observation that SΓ coincides with the direct sum Li∈I SΓi over the graph products obtained from its coconnected components. A vertex v ∈ V is called isolated if v does not emit any edge, and universal if v is connected to every other vertex in Γ. We note that the only coconnected graph with a universal vertex v is V = {v}, and that any graph containing an isolated edge is necessarily coconnected. For convenience, we let Vu denote the set of universal vertices, and I2 := {i ∈ I Vi ≥ 2}. We will make use of the following notion of a blocking path, that is actually a path in the opposite graph. Definition 2.1. Let Γ = (V, E) be a graph and C ⊂ V . A blocking path for C is a finite sequence of vertices w(1), . . . , w(n) ∈ V such that (a) w(1) /∈ C, (w(k), w(k + 1)) /∈ E for all 1 ≤ k ≤ n − 1, and (b) for every u ∈ C there exists 1 ≤ k ≤ n such that (w(k), u) /∈ E. It turns out that blocking paths are almost always available whenever the graph is coconnected, and we will frequently make use of this elementary observation in the course of this work. Lemma 2.2. If Γ is a coconnected graph with at least two vertices, then every finite proper subset C of V admits a blocking path ending in any prescribed vertex. Proof. Let C = {v(1), . . . , v(m)} ⊂ V be finite and proper, that is, V \C 6= ∅. If (v, u) ∈ E for all v ∈ C, u ∈ V \ C, then we would get a contradiction to Γ being coconnected. GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 5 Thus there exists w(1) ∈ V \ C such that (v(k), w(1)) /∈ E for some 1 ≤ k ≤ m. Without loss of generality, we can assume k = 1. Since Γ is coconnected, we can choose w′(k) ∈ V for 2 ≤ k ≤ m such that (v(k), w′(k)) /∈ E. Again by coconnectedness, there exists a finite path w(1), . . . , w(n) in Γopp that visits every w′(k), 2 ≤ k ≤ n. This is a blocking path for C, and since Γopp is connected, we can attach to this blocking path a path leading to any prescribed vertex without loosing the blocking property for C. (cid:3) Remark 2.3. Let Γ = (V, E) be a graph and (Sv)v∈V a family of right LCM semi- groups. Suppose w(1), . . . , w(n) is a blocking path for some nonempty C, and we can choose sn, tn ∈ Sw(n) \ S∗ w(n). Then for all s0, t0 whose reduced expressions only contain parts from vertex semigroups of vertices in C, and all sk, tk ∈ Sw(k), 1 ≤ k < n, we have ℓ(s0s1 · · · sn) = ℓ(s0) + n and s0s1 · · · sn ⊥ t0t1 · · · tn, unless sk = tk for 0 ≤ k < n and sn 6⊥ tn. Thus blocking paths allow for the construction of shuffle inert elements in graph products, which turns out to be quite useful. 3. The internal structure of graph products In this section we show that many of the properties of SΓ that are of interest to us, e.g. in connection with [ABLS], can be understood from a study of the corre- sponding graph products for the coconnected components (Γi)i∈I of Γ. The reason is SΓ = Li∈I SΓi and the following list of straightforward observations, where we write s = ⊕i∈I si for s ∈ Li∈I Si: Proposition 3.1. Let (Si)i∈I be a family of right LCM semigroups. Then S := Li∈I Si has the following features: i , Sc = Li∈I(Si)c, and S′ (i) S∗ = Li∈I S∗ (ii) s, t ∈ S are core related if and only if si and ti are core related in Si for all i ∈ I. (iii) The following statements hold for S if and only if their analogues hold for all Si: S is core factorable, Sci ⊂ S is ∩-closed, α : Sc y S/∼, a.[s] := [as] is faithful, and S has finite propagation. ci = Li∈I(Si)′ ci. (iv) The action α : Sc y S/∼, a.[s] := [as] is almost free if and only if one of the following conditions holds: (a) Si is left reversible for all i ∈ I, that is, S = Sc so that S/ ∼ is a singleton. (b) There exists a unique i ∈ I such that Si is not left reversible, αi : (Si)c y Si/∼ is almost free, and Sj is left reversible for all j ∈ I \ {i}. In view of the direct sum decomposition for graph products over the coconnected components, we need to understand the behaviour of the graph product in the case of a coconnected graph with at least two vertices. To do this, we will need to consider a variant of the action α for S∗, i.e. α∗ : S∗ y S/S∗, x.[s] := [xs]. Also, we will assume that all vertex semigroups Sv, v ∈ V are nontrivial in order to avoid pathological cases. For instance, if Γ is the union of a complete graph and an isolated vertex v, and Sv is trivial, then the graph product will be the direct sum of the right LCM semigroups attached to the vertices of the complete graph, even though the original graph was larger and coconnected. Theorem 3.2. If Γ = (V, E) is coconnected, V ≥ 2, and (Sv)v∈V is a family of nontrivial right LCM semigroups, then the following assertions hold: 6 NICOLAI STAMMEIER Γ is the graph product of (S∗ (i) S∗ (ii) For s, t ∈ SΓ, s ∼ t is equivalent to s ∈ tS∗ Γ. (iii) SΓ is core factorable and (SΓ)ci ⊂ SΓ is ∩-closed. (iv) The action α : S∗ Γ (v) The action α is almost free if and only if v )v∈V , (SΓ)c = S∗ y SΓ/∼ is faithful if and only if SΓ is not a group. Γ, and (SΓ)ci = SΓ \ S∗ Γ. v : S∗ (a) α∗ (b) for every connected component U ⊂ V with U ≥ 2, either Su is a group v is almost free for every isolated vertex v ∈ V , and y Sv /S∗ v for all u ∈ U or S∗ u is trivial for all u ∈ U. v )v∈V to SΓ (resulting from the universal property) is bijective, so that S∗ Proof. For (i), let sv(1)sv(2) · · · sv(n) be a reduced expression for s ∈ SΓ. Clearly, s is invertible in SΓ if and only if sv(k) ∈ S∗ v(k) for all k. The homomorphism from the graph product of (S∗ Γ is the graph product with respect to Γ and (S∗ v )v∈V . Now assume that there is 1 ≤ m ≤ n such that sv(k) ∈ S∗ v(m). Since Γ is coconnected, there is w ∈ V with w 6= v(m) and (v(m), w) /∈ E. For every t ∈ Sw \ {1}, we thus have sv(m)sv(m+1) · · · sv(n) ⊥ tsv(m)sv(m+1) · · · sv(n). By left cancellation, this yields s ⊥ sv(1)sv(2) · · · sv(m−1)tsv(m)sv(m+1) · · · sv(n), so that s /∈ (SΓ)c. This proves (SΓ)c = S∗ Γ, and the claims (SΓ)ci = SΓ \ S∗ Γ, (ii), and (iii) are immediate consequences of this. v(k) for 1 ≤ k < m but sv(m) /∈ S∗ v . Every x ∈ S∗ For (iv), we note that α is not faithful if Sv is a group for all v ∈ V because then SΓ/∼ is a singleton while S∗ Γ = SΓ is nontrivial. So let us assume that there exists v ∈ V with Sv 6= S∗ Γ \ {1} has a reduced expression xu(1)xu(2) · · · xu(m) with xu(k) ∈ S∗ u(k) \ {1}. Since Γ is coconnected and V ≥ 2, there exists a blocking path w(1), . . . , w(n) for {u(m)} with w(n) = v, see Lemma 2.2. Choose sw(k) ∈ Sw(k) \ {1} for 1 ≤ k < n and sw(n) ∈ Sw(n) \ S∗ w(n). Then s := sw(1)sw(2) · · · sw(n) ∈ SΓ satisfies xu(m)s ⊥ s. If 1 ≤ k ≤ m − 1 satisfies (u(k), u(ℓ)) ∈ E for all k < ℓ ≤ m, then (u(k), u(m)) ∈ E in particular implies u(k) 6= v(1). For the same reason, (u(k), v(1)) ∈ E implies u(k) 6= v(2), and so on. Thus xu(1)xu(2) · · · xu(m)sw(1)sw(2) · · · sw(n) is a reduced expression for xs and we conclude that orthogonality is not destroyed by xu(1)xu(2) · · · xu(m−1), i.e. xs ⊥ s. In particular, [xs] 6= [s] and therefore α is faithful. To prove (v), we first observe that (a) is necessary for α to be almost free: If v ∈ V is isolated, then [xs] = [s] for x ∈ S∗ v \ {1} and [s] ∈ SΓ implies s ∈ Sv. Suppose next that (b) does not hold, i.e. there exists a connected component U ⊂ V of Γ with U ≥ 2 such that there are u, v ∈ U with Sv 6= S∗ u 6= {1}. If u = v, then we can pick w ∈ U \ {v} with (v, w) ∈ E. If there is x ∈ S∗ w 6= {1}, then [xs] = [sx] = [s] for all s ∈ Sv, and since Sv /S∗ v is infinite, α fails to be almost free for x. On the other hand, Sw is nontrivial, so S∗ w is infinite, and then almost freeness fails for every x ∈ S∗ w = {1} implies that Sw/S∗ v and S∗ v 6= {1}. Now suppose u 6= v. As U is connected, we can find a path v(0) := u, v(1), . . . , v(n) := v from u to v inside U, i.e. (v(k), v(k + 1)) ∈ E for all 0 ≤ k < n. Then there exists 0 ≤ k < n such that S∗ v(k+1), and we can apply the above argument to deduce that α is not almost free. We have thus proven that almost freeness of α implies (a) and (b). v(k) 6= {1} and Sv(k+1) 6= S∗ Conversely, assume that (a) and (b) hold. If SΓ is a group, then there is nothing to show, so we may suppose that SΓ 6= S∗ Γ \ {1} be presented by a reduced expression xu(1)xu(2) · · · xu(m) with xu(k) ∈ S∗ Γ with reduced expression sv(1) · · · sv(n), sv(k) ∈ Sv(k). Let 1 ≤ j ≤ n be the smallest number such that u(k) \ {1}. Fix s ∈ SΓ \ S∗ Γ. Let x ∈ S∗ GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 7 sv(j) /∈ S∗ v(j). By (b), j is invariant under shuffling and we know that v(j) does not belong to the connected component of any u(k) that emits an edge. Therefore, xs ⊥ s and then [xs] 6= [s], unless j = m = 1 and u(1) = v(1) = v for some isolated vertex v ∈ V . In this case, (a) says that there are only finitely many fixed points for x in Sv /S∗ (cid:3) Remark 3.3. The graph product SΓ for a coconnected graph Γ with V ≥ 2 has finite propagation if S∗ v . Thus α is almost free if (and only if) (a) and (b) hold. v is a finite group for all v ∈ V . Let us now summarise what Proposition 3.1 and Theorem 3.2 imply for graph products of right LCM semigroups. Theorem 3.4. Let Γ = (V, E) be a graph and (Sv)v∈V a family of nontrivial right LCM semigroups. Then: Γ = Lv∈Vu S∗ v ⊕ Li∈I2 S∗ Γi. (i) S∗ (ii) (SΓ)c = Lv∈Vu(Sv)c ⊕ Li∈I2 S∗ Γi. (iii) (SΓ)′ ci ⊕ Li∈I2 SΓi. (iv) Two elements s, t ∈ SΓ are core related if and only if sv ∼v tv for all v ∈ Vu and ci = Lv∈Vu(Sv)′ si ∈ tiS∗ Γi for all i ∈ I2. (v) SΓ is core factorable if and only if Sv is core factorable for every v ∈ Vu. (vi) (SΓ)ci ⊂ SΓ is ∩-closed if and only if (Sv)ci ⊂ Sv is ∩-closed for every v ∈ Vu. (vii) The action α : (SΓ)c y SΓ/∼ is faithful if and only if αv : (Sv)c y Sv/∼ is faithful for every v ∈ Vu, and for every i ∈ I2 there exists w ∈ Vi such that Sw is not a group. (viii) The action α : (SΓ)c y SΓ/∼ is almost free if and only if one of the following conditions holds: (a) (Sv)c = {1} for all v ∈ Vu and S∗ (b) (Sv)c 6= {1} for a unique v ∈ Vu with αv : (Sv)c → Sv/∼ almost free, while w = {1} for all w ∈ V \ Vu. Sw = (Sw)c for all w ∈ Vu \ {v} and Sw′ = S∗ (c) S∗ 6= {1} for a unique i ∈ I2 with αi : S∗ Γi Sw = (Sw)c for all w ∈ Vu and SΓj = S∗ Γj for all j ∈ I2 \ {i}. w′ for all w′ ∈ V \ Vu. Γi → SΓi/∼ almost free, while (ix) SΓ has finite propagation if Sv has finite propagation for every v ∈ Vu and S∗ w is a finite group for all w ∈ V \ Vu. The conditions for almost freeness in Theorem 3.4 correspond to (SΓ)c = {1}, (SΓ)c = ∼== SΓi/∼, respectively. Hence they are quite restrictive, and we (Sv)c, and SΓ/∼ view this as an indication that finite propagation might be much more useful for graph products than almost freeness of α, see [ABLS, Theorem 4.2(2)] for details. When applied to right-angled Artin monoids, Theorem 3.4 takes a simpler form: Corollary 3.5. For a graph Γ = (V, E), the right-angled Artin monid A+ Γ satisfies: N, and (A+ Γ )1 ci = Li∈I2 A+ Γi. Γ )c = Lv∈Vu Γ )∗ = {1}, (A+ (i) (A+ (ii) Two elements s, t ∈ A+ (iii) A+ (iv) The action α : (A+ (v) The action α : (A+ Γ is core factorable, (A+ Γ )c y A+ Γ )c y A+ Γ is trivial or A+ core of A+ Γ are core related if and only if si = ti for all i ∈ I2. Γ is ∩-closed, and A+ Γ )ci ⊂ A+ Γ has finite propagation. Γ /∼ is faithful if and only if Γ has no universal vertex. Γ /∼ is almost free if and only if Vu ∈ {∅, V }, i.e. the Γ is the free abelian monoid in V . 8 NICOLAI STAMMEIER 4. The absence of property (AR) In this section, we will prove that for many graph products of right LCM semigroups SΓ, the only accurate foundation sets are given by elements of S∗ Γ. In particular, we obtain the an abundance of right LCM semigroups that lack property (AR). Again, the starting point is a basic observation for direct sums of right LCM semigroups, which allows us to boil the analysis down to the coconnected case: Proposition 4.1. Let (Si)i∈I be a family of right LCM semigroups. property (AR), then Si has property (AR) for all i ∈ I. Proof. Fix i ∈ I and let S := Li∈I Si. Every foundation set F for Si is a foundation set for S. Suppose that F has an accurate refinement Fa in S. For s ∈ SΓ, we let s = si + si with si ∈ Si and si ∈ Lj∈I\{i} Sj. If s ∈ Fa, then {fi ∈ Si f ∈ Fa : fi 6⊥ si} is an accurate refinement for F inside Si. Corollary 4.2. If a graph product SΓ has property (AR), then SΓi has property (AR) for each coconnected component Γi of Γ. If Li∈I Si has (cid:3) Theorem 4.3. Let Γ = (V, E) be a coconnected graph with at least two vertices and suppose (Sv)v∈V is a family of nontrivial right LCM semigroups. (i) If Γ is infinite, then every foundation set for SΓ contains an invertible element. In particular, SΓ has property (AR) and C ∗(SΓ) = Q(SΓ). (ii) If Γ is finite and E 6= ∅, then the accurate foundation sets for SΓ correspond Γ. In particular, SΓ has property (AR) if and only if SΓ does not admit a to S∗ foundation set without invertible elements. Proof. Both (i) and (ii) hold for trivial reasons if SΓ is a group, so we can assume that there exists w ∈ V with Sw 6= S∗ w. Let F ⊂ SΓ be a finite subset without invertible elements. For every f ∈ F , we choose a reduced expression f = fv(1) · · · fv(mf ) with mf ∈ N× and fv(k) ∈ Sv(k). Suppose first that Γ is infinite. As f ∈ SΓ \ S∗ Γ, there is a least 1 ≤ kf ≤ mf such that fv(kf ) /∈ S∗ v(kf ). Then C := {v ∈ V fv(k) ∈ Sv for some f ∈ F, 1 ≤ k ≤ kf }. Then C is a finite set of vertices so that Lemma 2.2 grants us a blocking path w(1), . . . , w(n) for C ending in w. If we choose any sk ∈ Sw(k) \ {1} for 1 ≤ k < n and sn ∈ Sw \ S∗ w, then s1 · · · sn ⊥ f for all f ∈ F as s1 · · · sn ⊥ fv(1) · · · fv(kf ) by construction, see Remark 2.3. Therefore F is not a foundation set. We conclude that every foundation set for SΓ contains an invertible element x, which clearly gives an accurate refinement {x}. So SΓ has property (AR), but if the only accurate foundation sets come from invertible elements, then the boundary relation Pf ∈F ef SΓ = 1 becomes trivial so that C ∗(S) = Q(S). Now let Γ be finite, E 6= ∅, and assume F to be accurate as well. We need to show that F is not a foundation set. Without loss of generality, we can require that fv(mf ) is not invertible for all f ∈ F because invertible ends do not play a role when it comes to intersections of right ideals. Since F does not contain any invertibles, we have ℓ(f ) ≥ 1 for all f ∈ F . Let L := maxf ∈F ℓ(F ), and choose f ∈ F with ℓ(f ) = L. Then we have f = stv for some v ∈ V, tv ∈ Sv \ {1}, and s ∈ SΓ with ℓ(s) = L − 1. We will first show that v is isolated, and then use this together with E 6= ∅ to conclude that F cannot be a foundation set. GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 9 If (v, u) ∈ E for some u ∈ V , we employ Lemma 2.2 to obtain a blocking path w(1), . . . , w(n) for C := {u} ∪ Nu, and set w(0) := u. Next, choose bk ∈ Sw(k) \ S∗ w(k) for each 1 ≤ k ≤ n, and let r ∈ Su \ {1}. It then follows that srb ⊥ f for b := b1 · · · bn. Moreover, we have ℓ(srb) ≥ m + 1. This could be assumed by extending the path w(0), . . . , w(n) in Γopp, but actually holds true in any case. It then follows that whenever f ′ ∈ F satisfies f ′ 6⊥ f 6= f ′ so that F would not be accurate. The blocking path then forces f ′ = srb1 · · · bk for some 1 ≤ k ≤ n. However, we then get f ′ ⊥ sr′b for every r′ ∈ Su \ {r}. Since Su is a left cancellative semigroup that is not a group, it is infinite. Thus there is r ∈ Su \ {1} such that srb ⊥ f ′ for all f ′ ∈ F . 6⊥ srb, we have srb ∈ f ′SΓ. If sr ∈ f ′Γ, then f ′ We deduce from this that F cannot be a foundation set if there exists f ∈ F with ℓ(f ) = L that does not end in a part from an isolated vertex. In particular, if Γ does not have any isolated vertices, no accurate finite subset F without invertible elements is a foundation set. Now suppose Γ has an isolated vertex v, and let F ′ := {f ∈ F fv(k) ∈ Sv for some k ⇒ v is not isolated.}, that is, the subset of F consisting of those elements whose reduced expressions do not contain any part coming from an isolated vertex. As E 6= ∅ and the vertex semigroups are all nontrivial, the finite accurate set F ′ is also non-empty. Suppose first that there is f ∈ F ′ with f ∈ Sv \ S∗ v for some v ∈ V . Since F ′ is accurate and (v, u) ∈ E for some u ∈ V , we have s /∈ f ′SΓ for all s ∈ Su and f ′ ∈ F ′. Thus we get str ⊥ f ′ for all f ′ ∈ F ′ whenever s ∈ Su, t ∈ Sv, and r ∈ Sw \ S∗ w, compare Remark 2.3. For f ∈ F \ F ′, we have strtr ⊥ f unless f ∈ strtSΓ because v is isolated and r is not invertible. Since F is finite while Sw \ S∗ w is infinite, we conclude that there are s ∈ Su, t ∈ Sv, and r ∈ Sw \ S∗ w such that strtr ⊥ f for all f ∈ F . So F is not a foundation set. On the other hand, if we have ℓ(f ) ≥ 2 for every f ∈ F ′, we pick a vertex v that emits an edge. Then s /∈ f SΓ for all s ∈ Sv, f ∈ F ′, and thus str ⊥ f for all f ∈ F ′ whenever s ∈ Sv, t ∈ Sv, and r ∈ Sw \ S∗ w. As in the previous case, there are s, t, r such that strtr ⊥ f for all f ∈ F , and thus F is not a foundation set. Finally, if F is a foundation set for SΓ with F ∩ S∗ F satisfies F ′ ∩ S∗ every foundation set F with x ∈ F ∩ S∗ Γ = ∅, then every refinement F ′ of Γ = ∅ as well, and thus can never be accurate. On the other hand, (cid:3) Γ has an accurate refinement {x}. We point out that the assumptions in Theorem 4.3 are modest means to avoid the somewhat pathological cases: SΓ = Sv, the free product SΓ = ∗v∈V Sv, and the graph product of groups. Remark 4.4. By Theorem 4.3 (i), foundation sets of SΓ are governed by parts from the finite coconnected components in the following sense: Let F be a foundation set for SΓ such that no propert subset of F is a foundation set. If s = sv(1) · · · sv(n) ∈ F with sv(k) ∈ Sv(k), then sv(k) /∈ S∗ v(k) implies that v(k) ∈ Vi for some finite coconnected component Γi = (Vi, Ei) of Γ. Corollary 4.5. Let Γ be a graph and (Sv)v∈V a family of nontrivial right LCM semi- groups. If there is i ∈ I2 for which Γi = (Vi, Ei) is finite with Ei 6= ∅, Sv is not a group 10 NICOLAI STAMMEIER for some v ∈ Vi, and there exists a foundation set F for SΓi without invertible elements, then SΓ does not have property (AR). Proof. The claim follows from combining Theorem 4.3 with Corollary 4.2. (cid:3) The previous results apply nicely to right-angled Artin monoids. Corollary 4.6. For graph Γ, the following conditions are equivalent: (1) Every finite coconnected component Γi of Γ is edge-free. (2) Every finitely generated direct summand of A+ (3) The right-angled Artin monoid A+ Γ is free. Γ has property (AR). Proof. The equivalence of (1) and (2) is clear from the direct sum description of A+ Γ in Subsection 2.2. From Remark 4.4 we infer that it suffices to obtain accurate refinements of foundation sets F for A+ Γi. But if (2) holds, then the latter is just a direct sum of finitely generated free monoids, and clearly admits accurate refinements. So (2) implies (3). Finally, if (3) is valid and Γi = (Vi, Ei) is a coconnected component of Γ with 2 ≤ Vi < ∞, then {av v ∈ Vi} is a foundation set for A+ Γi without invertible elements, so Corollary 4.5 forces Ei = ∅, that is, (1) holds. (cid:3) Γ with F ⊂ Lv∈Vu Sv ⊕ Li∈I2:Vi<∞ A+ By Corollary 4.6, there exist countably many mutually non-isomorphic, finitely gen- erated right LCM semigroups without property (AR). As a final part of this section, we address the existence of a generalised scale for right-angled Artin monoids associated to finite graphs. The existence of a generalised scale turned out to be relevant for a standardised approach to study KMS-states on the semigroup C ∗-algebra C ∗(A+ Γ ), see [ABLS]. We first note that free monoids have a generalised scale only if they are finitely generated and nonabelian, in which case it is unique: Proposition 4.7. The free monoid F+ generalised scale N : F+ of w ∈ F+ m. m in 2 ≤ m < ∞ generators admits a unique m → N× given by N(w) = mℓ(w), where ℓ denotes the word length Proof. The map N is a generalised scale. On the other hand, let N be a generalised m = ha1, . . . , ami, and fix 1 ≤ i ≤ m. Then N (ai) > 1 as ai is not part of scale on F+ m)c = {1}. By definition of N and since ∼ is trivial, the set N −1( N (ai)) is an accurate (F+ m of cardinality N(ai) that contains ai. If there was 1 ≤ j ≤ m, j 6= i foundation set for F+ such that N (aj) 6= N(ai), then the foundation set property would give a w ∈ N −1( N (ai)) m. As this forces N (w) ≥ N (ai) N (aj) > N (ai) = N(w), we arrive such that w ∈ ajaiF+ at a contradiction. Thus N (aj) = N(ai) for all j 6= i. But as {a1, . . . , am} is an accurate foundation set for F+ (cid:3) m, we conclude that N (ai) = m for all 1 ≤ i ≤ m. We call m ∈ Li∈I{k ∈ N 2 ≤ k < ∞} rationally independent if for all distinct N, the supernatural numbers Qi∈I mki k, k′ ∈ Li∈I Proposition 4.8. Let M be a free abelian monoid, m ∈ Li∈I{k ∈ N 2 ≤ k < ∞} for mi admits a generalised scale N : S → N× if F+ some nonempty set I. Then S := M ⊕Li∈I and only if m is rationally independent. In this case, N restricts to the unique generalised scales Ni on F+ mi, and is therefore unique. i and Qi∈I mk′ i i are distinct. GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 11 Proof. As M = Sc = ker N for any generalised scale N on S, see [ABLS, Proposi- tion 3.6(i)], we can focus on (F+ mi is the free monoid in mi generators, which we denote by ai,1, . . . , ai,mi. The strategy is to prove that mi, and (a) any generalised scale N on S restricts to Ni on F+ (b) the homomorphism N : S → N× arising from (Ni)i∈I is a generalised scale if and mi)i∈I. Recall that F+ only if m is rationally independent. For (a), suppose S admits a generalised scale N and fix i ∈ I, 1 ≤ k ≤ mi. Then N(ai,k) > 1 and there are w1, . . . , wN (ai,k)−1 ∈ S such that {ai,k, w1, . . . , wN (ai,k)−1} is an accurate foundation set for S contained in N −1(N(ai,k)). Let us decompose wℓ F+ as wℓ = wℓ ⊕ wℓ ∈ F+ mj(cid:1). Then {ai,k, w1, . . . , wN (ai,k)−1} is a foundation set for F+ mi with ai,k ⊥ wℓ and N( wℓ) ≤ N(ai,k) for all ℓ. This forces {ai,k, w1, . . . , wN (ai,k)−1} ⊃ {ai,1, . . . , ai,mi}, and thus N(ai,ℓ) ≤ N(ai,k) for all 1 ≤ ℓ ≤ mi, just like in the proof of Proposition 4.7. As k was arbitrary, we deduce N(ai,k) = mi = Ni(ai,k) for all i, k. mi ⊕ (cid:0)Mn ⊕ Lj∈I\{i} In view of (a), the question behing the main claim becomes: Under which condition is the homomorphism N : S → N× arising from the family of generalised scales (Ni)i∈I itself a generalised scale? If m is rationally independent, then every k ∈ N(S) has a factorization k = Qi∈I mki i with uniquely determined ki ∈ N. This implies that N −1(k) = {t ⊕ L i∈I wi t ∈ M, wi ∈ F+ mi with ℓi(wi) = ki}. N, k 6= k′ such that K := Qi∈I mki Therefore, N −1(k)/∼ = k, and any transversal of N −1(k)/∼ is an accurate foundation set for S, that is, N is a generalised scale. On the other hand, if there are k, k′ ∈ i , then both k and k′ yield a set of Li∈I K mutually orthogonal elements s1, . . . , sK ∈ S and t1, . . . , tK ∈ S, respectively, with N(sj) = K = N(tj) for all j. Since there is i ∈ I with ki 6= k′ i, the i-th components of sj and tj ′ have different length for all j, j′. Thus sj 6∼ tj ′ for all j, j′, and we get N −1(K)/∼ ≥ 2K. Therefore N is not a generalised scale in this case. (cid:3) i = Qi∈I mk′ i We can now state our conclusions for right-angled Artin monoids. Corollary 4.9. For every graph Γ, the right-angled Artin monoid A+ Γ admits a gener- alised scale N if and only if Vu 6= V , all coconnected components Γi = (Vi, Ei) are finite and edge-free, and Li∈I2Vi is rationally independent. In this case, N is unique. Proof. The condition Vu 6= V is equivalent to saying that A+ So if all coconnected components Γi = (Vi, Ei) are finite and edge-free, then A+ Γ Lv∈Vu scale N if and only if Li∈I2Vi is rationally independent. Γ is non-abelian, i.e.I2 6= ∅. ∼= Γ has a (unique) generalised F+ Vi. Hence, Proposition 4.8 implies that A+ N⊕Li∈I2 Conversely, suppose A+ Γ admits a generalised scale N. Since N is a nontrivial ho- momorphism with ker N = Lv∈Vu N, we need to have Vu 6= V so that the set I2 is non-empty. Moreover, A+ Γ has property (AR) by [ABLS, Proposition 3.6], so Corol- lary 4.6 implies that all finite coconnected components Γi of Γ are edge-free. If there was an infinite coconnected component Γi = (Vi, Ei), then 1 < N(av) < ∞ for all v ∈ Vi, and the defining property of a generalised scale would yield an accurate foundation set 12 NICOLAI STAMMEIER of the form {av, f1, . . . , fN (av )−1} for suitable fk ∈ A+ mark 4.4, and we conclude that Γi is finite for all i ∈ I2. But then A+ Proposition 4.8, and it follows that Li∈I2Vi is rationally independent. Γ . However, this contradicts Re- Γ is covered by (cid:3) Corollary 4.10. For every graph Γ, the right-angled Artin monoid A+ and only if it admits a generalised scale. Γ is admissible if Proof. According to Corollary 3.5 (iii), A+ Γ is ∩- closed, no matter what Γ is. By Corollary 4.9, the conditions characterising the existence of a generalised scale N include rational independence of Li∈I2Vi. This feature implies Irr(N(A+ Γ ), which is the last extra condition for admissibility. (cid:3) Γ )) = {Vi i ∈ I2} and that this set freely generates N(A+ Γ is core factorable and (A+ Γ )ci ⊂ A+ References [ABLS] Zahra Afsar, Nathan Brownlowe, Nadia S. Larsen, and Nicolai Stammeier, Equilibrium states on right LCM semigroup C ∗-algebras. preprint, arxiv:1611.01052. [BaHLR12] Nathan Brownlowe, Astrid an Huef, Marcelo Laca, and Iain Raeburn, Boundary quotients of the Toeplitz algebra of the affine semigroup over the natural numbers, Ergodic Theory Dynam. Systems 32 (2012), no. 1, 35–62, DOI 10.1017/S0143385710000830. [BLS17] Nathan Brownlowe, Nadia S. Larsen, and Nicolai Stammeier, On C ∗-algebras associ- ated to right LCM semigroups, Trans. Amer. Math. Soc. 369 (2017), no. 1, 31–68, DOI 10.1090/tran/6638. [BRRW14] Nathan Brownlowe, Jacqui Ramagge, David Robertson, and Michael F. Whittaker, Zappa- Sz´ep products of semigroups and their C ∗-algebras, J. Funct. Anal. 266 (2014), no. 6, 3937–3967, DOI 10.1016/j.jfa.2013.12.025. [BS16] Nathan Brownlowe and Nicolai Stammeier, The boundary quotient for algebraic dynamical systems, J. Math. Anal. Appl. 438 (2016), no. 2, 772–789, DOI 10.1016/j.jmaa.2016.02.015. [Cha07] Ruth Charney, An introduction to right-angled Artin groups, Geom. Dedicata 125 (2007), 141–158, DOI 10.1007/s10711-007-9148-6. [CaHR16] Lisa Orloff Clark, Astrid an Huef, and Iain Raeburn, Phase transitions on the Toeplitz algebras of Baumslag-Solitar semigroups, Indiana Univ. Math. J. 65 (2016), no. 6, 2137 – 2173, DOI 10.1512/iumj.2016.65.5934. [VdC01] Ant´onio Veloso da Costa, Graph products of monoids, Semigroup Forum 63 (2001), no. 2, 247–277, DOI 10.1007/s002330010075. [CL02] John Crisp and Marcelo Laca, On the Toeplitz algebras of right-angled and finite-type Artin groups, J. Aust. Math. Soc. 72 (2002), no. 2, 223–245, DOI 10.1017/S1446788700003876. , Boundary quotients and ideals of Toeplitz C ∗-algebras of Artin groups, J. Funct. [CL07] Anal. 242 (2007), no. 1, 127–156, DOI 10.1016/j.jfa.2006.08.001. [Die90] Volker Diekert, Combinatorics on traces, Lecture Notes in Computer Science, vol. 454, Springer-Verlag, Berlin, 1990. With a foreword by Wilfried Brauer. [ELR16] Søren Eilers, Xin Li, and Efren Ruiz, The isomorphism problem for semigroup C ∗-algebras of right-angled Artin monoids, Doc. Math. 21 (2016), 309–343. electronic version. [EP17] Ruy Exel and Enrique Pardo, Self-similar graphs, a unified treatment of Katsura and Nekrashevych C ∗-algebras, Adv. Math. 306 (2017), no. 1, 1046–1129, DOI 10.1016/j.aim.2016.10.030. [EP16] , The tight groupoid of an inverse semigroup, Semigroup Forum 92 (2016), no. 1, 274–303, DOI 10.1007/s00233-015-9758-5. [FK09] John Fountain and Mark Kambites, Graph products of right cancellative monoids, J. Aust. Math. Soc. 87 (2009), no. 2, 227–252, DOI 10.1017/S144678870900010X. [Gre90] Elisabeth R. Green, Graph products of groups, University of Leeds, 1990. GRAPH PRODUCTS AND THE ABSENCE OF PROPERTY (AR) 13 [LR10] Marcelo Laca and Iain Raeburn, Phase transition on the Toeplitz algebra of the affine semigroup over the natural numbers, Adv. Math. 225 (2010), no. 2, 643–688, DOI 10.1016/j.aim.2010.03.007. [LRR11] Marcelo Laca, Iain Raeburn, and Jacqui Ramagge, Phase transition on Exel crossed prod- ucts associated to dilation matrices, J. Funct. Anal. 261 (2011), no. 12, 3633–3664, DOI 10.1016/j.jfa.2011.08.015. [LRRW14] Marcelo Laca, Iain Raeburn, Jacqui Ramagge, and Michael F. Whittaker, Equilibrium states on the Cuntz-Pimsner algebras of self-similar actions, J. Funct. Anal. 266 (2014), no. 11, 6619–6661, DOI 10.1016/j.jfa.2014.03.003. [Li12] Xin Li, Semigroup C ∗-algebras and amenability of semigroups, J. Funct. Anal. 262 (2012), no. 10, 4302–4340, DOI 10.1016/j.jfa.2012.02.020. [Li13] , Nuclearity of semigroup C*-algebras and the connection to amenability, Adv. Math. 244 (2013), 626–662, DOI 10.1016/j.aim.2013.05.016. [Sta] Nicolai Stammeier, A boundary quotient diagram for right LCM semigroups, to appear in Semigroup Forum. arxiv:1604.03172. [Star15] Charles Starling, Boundary quotients of C ∗-algebras of right LCM semigroups, J. Funct. Anal. 268 (2015), no. 11, 3326–3356, DOI 10.1016/j.jfa.2015.01.001. [Ste16] Benjamin Steinberg, Simplicity, primitivity and semiprimitivity of ´etale groupoid algebras with applications to inverse semigroup algebras, J. Pure Appl. Algebra 220 (2016), no. 3, 1035–1054, DOI 10.1016/j.jpaa.2015.08.006. Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, NO-0316 Oslo, Norway E-mail address: [email protected]
1111.6927	3	1111	2012-11-15T20:43:55	C*-algebras for categories of paths asociated to the Baumslag-Solitar groups	[ "math.OA" ]	In this paper we describe the C-algebras associated to the Baumslag-Solitar groups with the ordering defined by the usual presentations. These are Morita equivalent to the crossed product C-algebras obtained by letting the group act on its directed boundary. We use the method of categories of paths to define the algebras, and to deduce the presentation by generators and relations. We obtain a complete description of the Toeplitz algebras, and we compute the K-theory of the Cuntz-Kreiger algebras.	math.OA	math	C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS JACK SPIELBERG Abstract. In this paper we describe the C ∗-algebras associated to the Baumslag-Solitar groups with the ordering defined by the usual presentations. These are Morita equivalent to the crossed product C ∗-algebras obtained by letting the group act on its directed boundary. We use the method of categories of paths to define the algebras, and to deduce the presentation by generators and relations. We obtain a complete description of the Toeplitz algebras, and we compute the K-theory of the Cuntz-Kreiger algebras. 1. Introduction In this paper we study the C ∗-algebras associated to the Baumslag-Solitar groups with the ordering defined by the usual presentation. We show that in most cases this is a quasi-lattice ordering in the sense of [9]. However we use the notion of category of paths of [13] to describe the C ∗-algebras. This is a construction of Toeplitz and Cuntz-Krieger algebras in a very general setting that includes ordered groups, higher-rank graphs, and many other examples of C ∗-algebras obtained from oriented combinatorial objects. The chief virtue of this construction is that a presentation of the algebras by generators and relations is obtained naturally from the category, eliminating the guesswork typically needed for their identification. For example, if a free semigroup acts on its ℓ2-space, it is easy to observe that the generators define isometries having pairwise orthogonal ranges spanning a codimension-one subspace. The rank-one projection onto this subspace generates the compact operators when pushed around by the isometries, and in the quotient, the classes of isometries have range projections adding to the identity. Cuntz's theorem ([4]) shows that this quotient algebra is uniquely defined by this presentation. For other semigroups, however, it may not be obvious what is the correct quotient, even if the compact operators are present in the algebra generated by the isometries on ℓ2. Moreover, it may also be unclear what relations ought to be used to define the quotient. In the present case, we obtain generators and relations for a C ∗-algebra Morita equivalent to the crossed product algebra associated to the action of a Baumslag-Solitar group on its directed boundary. These presentations turn out to coincide with certain examples obtained by Katsura in his work on topological graphs ([6]). Our method gives a new approach to the description of these algebras by generators and relations, and also gives the ideal structure of the Toeplitz versions. In the case of the solvable examples BS(1, d), this is a well-known example with core isomorphic to a Bunce-Deddens algebra ([1]). The presentations of the Baumslag-Solitar groups can be thought of as arising from their status as the fundamental group of a graph of groups: the graph consists of one vertex and one edge, with infinite cyclic groups attached to both vertex and edge. The Bass-Serre theory requires the choice of an orientation for this graph; this amounts to a direction on the Cayley graph ([12]). In this case, the directed Cayley graph is a category of paths, and we use the theory developed in [13] to study the action of the group on the boundary (see [13], Example 8.5). The organization of this paper is as follows. We first use the HNN structure of a Baumslag-Solitar group to prove that, as an ordered group (relative to the subsemigroup defined by the presentation), it is suitable for our constructions. In the process we define an associated odometer-like action by the second generator. We then identify the maximal directed hereditary subsets of the category, and hence the boundary of the semigroup. We prove that the restriction of the groupoid to the boundary of the category is amenable. We then identify all directed hereditary subsets, and prove that the entire groupoid is amenable. From this we deduce the generators and relations for the Cuntz-Krieger algebra. We then compute the K-theory of the algebra by first studying the fixed-point algebra by the gauge action, and then applying the Pimsner- Voiculescu exact sequence. We mention that the degree functors for these categories are degenerate, and Date: 28 November 2011. 2010 Mathematics Subject Classification. Primary 46L05; Secondary 46L80, 46L55. Key words and phrases. Baumslag-Solitar group, Cuntz-Krieger algebra, Toeplitz Cuntz-Krieger algebra, K-theory. 1 2 JACK SPIELBERG thus the fixed-point algebras are not AF. Finally we use theorems from [13] to establish the fundamental structural properties of these algebras. We conclude this introduction with a description of the definitions and results from [13] needed for the rest of the paper. We identify the objects of a category with the identity morphisms in that category, and we use juxtaposition to indicate composition of morphisms. Morphisms are referred to as paths, and objects as vertices. We use s and r to denote the source and range of morphisms, and Λ0 for the vertices in the category Λ. A category of paths is a small category satisfying only consider (certain) finitely aligned examples. (1) αβ = αγ implies β = γ (left-cancellation). (2) βα = γα implies β = γ (right-cancellation). (3) αβ = s(β) implies α = β = s(β) (no inverses). For any α ∈ Λ we define the left shift σα : αΛ → s(α)Λ by σα(αβ) = β (σα is well-defined by left- cancellation). The right shift map β ∈ s(α)Λ 7→ αβ ∈ r(α)Λ is the inverse of σα. We say that β extends α if there exists α′ ∈ Λ such that β = αα′. (We may express this by writing β ∈ αΛ.) It follows from the definition that this is a partial order on Λ. If β is an extension of α, we call α an initial segment of β. The set of initial segments of β is denoted [β]. We write α ⋓ β (α meets β) if αΛ ∩ βΛ 6= ∅, and α ⊥ β (α is disjoint from β) otherwise. We let α ∨ β denote the set of minimal common extensions of α and β, i.e. the of the elements of F . We say that Λ is finitely aligned if for every pair of elements α, β ∈ Λ, there is a finite minimal elements of αΛ ∩ βΛ. For a subset F ⊆ Λ we letW F denote the set of minimal common extensions subset G of Λ such that αΛ ∩ βΛ =Sε∈G εΛ. It follows that we may take G = α ∨ β. In this paper we will A subset C ⊆ Λ is directed if for all α, β ∈ C there is γ ∈ C extending both α and β. C is hereditary if [α] ⊆ C for every α ∈ C. The collection of all directed hereditary subsets of Λ is denoted Λ∗; the set of maximal elements of Λ∗ is denoted Λ∗∗. A directed hereditary set is finite if it contains a maximal element; in this case it must be of the form [α] for some α ∈ Λ. Otherwise it is infinite. We define a topology on Λ∗ as follows. For α ∈ Λ, and β1, . . ., βn ∈ αΛ \ {α}, let E = αΛ \Sn Then the bE form a base of compact-open sets for a locally compact Hausdorff topology. The boundary of Λ bE = {C ∈ Λ∗ : E ⊇ C ∩ γΛ for some γ ∈ C}. We define a groupoid with unit space Λ∗ as follows. First define a relation on Λ × Λ × Λ∗ by (α, β, x) ∼ (α′, β′, x′) if there are y ∈ Λ∗ and γ, γ′ ∈ Λ such that x = γy, x′ = γ′y, αγ = α′γ′, and βγ = β′γ′. Then ∼ is an equivalence relation. The set of equivalence classes becomes a locally compact Hausdorff ´etale groupoid, where the set of composable pairs is is defined to be ∂Λ = Λ∗∗. i=1 βiΛ. Set G2 =(cid:8)(cid:0)[α, β, x], [γ, δ, y](cid:1) : βx = γy(cid:9), y = ηz, and βξ = γη. Then and inversion is given by [α, β, x]−1 = [β, α, x]. Multiplication G2 → G is given as follows. Let (cid:0)[α, β, x], [γ, δ, y](cid:1) ∈ G2. Since βx = γy, it follows from Lemma 4.12 of [13] that there are z, ξ, and η such that x = ξz, A base of compact-open sets for G is given by the sets [α, β, bE] = {[α, β, x] : x ∈ bE}. The C ∗-algebras of Λ are defined as T C ∗(Λ) = C ∗(G) (the Toeplitz C ∗-algebra), and C ∗(Λ) = C ∗(G∂Λ) (the Cuntz-Krieger algebra). We have the following theorem giving generators and relations for T C ∗(Λ) ([13], Theorem 6.3). [α, β, x] [γ, δ, y] = [αξ, δη, z]. Theorem 1.1. Let Λ be a finitely aligned category of paths. The representations of T C ∗(Λ) are in one-to-one correspondence with the families {Tα : α ∈ Λ} of Hilbert space operators satisfying the relations αTα = Ts(α). (1) T ∗ (2) TαTβ = Tαβ, if s(α) = r(β). (3) TαT ∗ γ . β =Wγ∈α∨β TγT ∗ αTβT ∗ In order to describe the analogous theorem for C ∗(Λ), we need the notion of exhaustive set. Let v ∈ Λ0. A subset F ⊆ vΛ is exhaustive (at v) if for every α ∈ vΛ there exists β ∈ F such that α ⋓ β. The exhaustive sets can be used to characterize the points of ∂Λ ([13], Theorem 7.8). We have the following theorem giving generators and relations for C ∗(Λ) ([13], Theorem 8.2). C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 3 Theorem 1.2. Let Λ be a countable finitely aligned category of paths. Assume that G is amenable. The representations of C ∗(Λ) are in one-to-one correspondence with the families {Sα : α ∈ Λ} of Hilbert space operators satisfying the relations αSα = Ss(α). (1) S∗ (2) SαSβ = Sαβ, if s(α) = r(β). (3) SαS∗ γ. αSβS∗ β =Wγ∈α∨β SγS∗ (4) Sv =Wβ∈F SβS∗ β if F is finite exhaustive at v. (Equivalently, 0 =Qδ∈F (Sv − SδS∗ The last theorem requires amenability of the groupoid. The usual way of obtaining this is to decompose C ∗(Λ) by means of a cocycle to a discrete abelian group (called a degree functor in [13], Definition 9.1). We cite the following summary of the standard argument ([13], Proposition 9.3). δ ).) Proposition 1.3. Let G be a locally compact Hausdorff ´etale groupoid, Q a countable abelian group, and c : G → Q a continuous homomorphism. Let Gc = c−1(0), also a locally compact Hausdorff ´etale groupoid. Suppose that Gc is amenable. Then G is amenable. An ordered group (Γ, Λ) is called quasi-lattice ordered if for each t ∈ ΛΛ−1 there is an element α ∈ tΛ ∩ Λ such that tΛ ∩ Λ = αΛ ([9]). This idea was generalized in [13], Definition 8.6, as follows. The ordered group (Γ, Λ) is called finitely aligned if for each t ∈ Γ there is a finite set F ⊆ tΛ ∩ Λ such that tΛ ∩ Λ =Sα∈F αΛ. It is possible that Λ is a finitely aligned category of paths even if (Γ, Λ) is not a finitely aligned ordered group. A weaker notion is given in [13], Definition 8.11: (Γ, Λ) is locally finitely exhaustible if for each t ∈ Γ, there is a finite set F ⊆ tΛ ∩ Λ such that every element of tΛ ∩ Λ meets some element of F . In Lemma 8.10 of [13] it is shown how to define a locally compact Γ-space, ∂(Γ, Λ), which is the directed boundary of the ordered group. The following appears in [13] as part of Theorem 8.13 and Corollary 8.17. Theorem 1.4. Let (Γ, Λ) be a countable ordered group. Suppose that Λ is finitely aligned as a category of paths. Then ∂(Γ, Λ) is locally compact Hausdorff if and only if (Γ, Λ) is locally finitely exhaustible. Moreover, C ∗(Λ) is Morita equivalent to the crossed product algebra C0(∂(Γ, Λ)) × Γ. 2. The category of paths of a Baumslag-Solitar group equivalent relation, we may as well assume that c and d are not both negative. We will consider separately the cases cd > 0 and cd < 0. Thus we formulate the situation for a pair of positive integers c and d, and let the group be defined by the relation abc = bda, or by the relation abc = b−da. The case cd > 0 will be further divided accordingly as c ≥ d or c < d, giving three cases overall: For nonzero integers c and d we consider the group Γ = ha, b(cid:12)(cid:12) abc = bdai ([3]). Since ab−c = b−da is an Case (BS1): Γ = ha, b(cid:12)(cid:12) abc = bdai where c ≥ d ≥ 1. Case (BS2): Γ = ha, b(cid:12)(cid:12) abc = bdai where d > c ≥ 1. Case (BS3): Γ = ha, b(cid:12)(cid:12) abc = b−dai where c, d ≥ 1. We denote by θ : G → Z the homomorphism given by θ(a) = 1 and θ(b) = 0. We sometimes refer to θ(g) as the height of g. Let Λ be the submonoid generated by a and b, and let B = {bi : i ≥ 0} denote the submonoid generated by b. The following may be found in [11]. Proposition 2.1. Each element of Γ has a unique represention in the form bi1 aε1 · · · bin aεn bq, where εµ ∈ {±1}, iµ ∈ [0, d) if εµ = +1, and iµ ∈ [0, c) if εµ = −1. The standard from of the proposition is obtained by moving b's to the right via bkda = abkc and bkca−1 = a−1bkd. Corollary 2.2. Let t ∈ Γ have the form in Proposition 2.1. Then t ∈ Λ if and only if εµ = +1 for all µ, and q ≥ 0 in cases (BS1) and (BS2), and in case (BS3) if n = 0. We note the following proposition for later use. It follows from the same kind of arguments as gives Proposition 2.1 and its corollary. Proposition 2.3. Each element α ∈ Λ has unique representations in the two forms (L) α = bi0 abi1 a · · · bik−1 abp, iµ ∈ [0, d), p ∈ Z; (R) α = bqabj1abj2 · · · abjk , jµ ∈ [0, c), q ∈ Z, 4 JACK SPIELBERG where k = θ(α). Remark 2.4. In cases (BS1) and (BS2), in forms (L) and (R) we have p, q ≥ 0. In case (BS3), if θ(α) = 0 we have p = q ≥ 0. It also follows that b has infinite order in G. The following corollary follows easily. Corollary 2.5. If 0 < k < θ(α), or if k = θ(α) > 0 and we are in case (BS1) or (BS2), then α has a unique initial segment of height k in the form (L) of Proposition 2.3 with p = 0. If k = θ(α) > 0 and we are in case (BS3), then there is a unique element α0 in the form (L) of Proposition 2.3 with p = 0, and p0 ∈ Z, such that every initial segment of height k has form (L) equal to α0bq with q ≤ p0. Lemma 2.6. In case (BS3), if α ∈ Λ and θ(α) > 0, then αbp ∈ Λ for all p ∈ Z. Moreover, B ⊆ [α]. Proof. We can write α = α0abm for α0 ∈ Λ and m ≥ 0. Choose n so that nc + m + p ≥ 0. Then using the relation a = bdabc, we find that αbp = α0abm+p = α0bndabnc+m+p ∈ Λ. For the second statement, we write α = biaα1 for some i ∈ N and α1 ∈ Λ. Then α = bi+mdabmcα1 for all m ∈ N. (cid:3) The following begins our study. Lemma 2.7. Λ is a category of paths ([13], Example 8.5). Proof. Note that θ(α) ≥ 0 for all α ∈ Λ. We note that if α, β ∈ Λ are such that αβ = e, then θ(α)+θ(β) = 0, and hence θ(α) = θ(β) = 0. But then α, β ∈ B. But then we must have α = β = e. Hence Λ is a category of paths. (cid:3) We will prove that Λ is finitely aligned, and that in cases (BS1) and (BS2), the ordered group (Γ, Λ) is quasi-lattice ordered in the sense of [9] (Theorem 2.11 below). (The case (BS3) is slightly different -- see Lemma 2.12.) The argument varies by case. We require lemmas for cases (BS1) and (BS2). Lemma 2.8. Suppose that we are in case (BS2). Fix k ≥ 1. For i = (i0, . . . , ik) ∈ [0, c)k+1, let α(i) = bi0abi1 a · · · bik a. There are maps ψ : [0, c)k+1 → [0, c)k+1 and r : [0, c)k+1 → Z+ such that (1) bdr(i)α(ψ(i)) = α(i)bc and ψ(i)0 = i0. (2) If q < c then bi0+1 6∈ [α(i)bq]. (3) If i0 < m ≤ i0 + dr(i), then α(i)bc = bdr(i)α(ψ(i)) is the unique minimal common extension of bm and α(i). (4) For h ≥ 0, if then i0 + d h−1Xµ=0 r(ψµ(i)) < m ≤ i0 + d r(ψµ(i)) hXµ=0 α(i)b(h+1)c = bd Ph µ=0 r(ψµ(i))α(ψh+1(i)) is the unique minimal common extension of bm and α(i). Proof. For (1), the maps ψ and r are defined by the unique form (R) of α(i)bc. That r(i) ≥ 1 follows from the assumption c < d. It follows from the defining relation of G that if i0 = 0 then α(i)0 = 0. Hence by the uniqueness of form (L), ψ(i)0 = i0 for all i. For (2), note that if q < c then α(i)bq is already in form (R). We next prove (3). Let m be as in the statement. By (1) we know that α(i)bc ∈ bmΛ ∩ α(i)Λ. Let β ∈ bmΛ ∩ α(i)Λ. Then β = α(i)γ for some γ. Write γ = bqabℓ1 · · · abℓn with ℓµ ∈ [0, c). Then the form (R) of β equals (form (R) of α(i)bq) · abℓ1 · · · abℓn. Therefore bm ∈ [β] if and only if bm ∈ [α(i)bq]. This occurs if and only if q ≥ c, by (2). Therefore β ∈ α(i)bcΛ. Finally, we prove (4), by induction on h. The case h = 0 is (3). Suppose it is true for h, and let i0 + d hXµ=0 r(ψµ(i)) < m ≤ i0 + d r(ψµ(i)). h+1Xµ=0 First we note that Thus α(i)b(h+2)c is a common extension of bm and α(i). α(i)b(h+2)c = bd Ph µ=0 r(ψµ(i))α(ψh+1(i))bc = bd Ph+1 µ=0 r(ψµ(i))α(ψh+2(i)) ∈ bmΛ. C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 5 Now let β ∈ bmΛ ∩ α(i)Λ be any common extension. Then β ∈ bm−dr(ψh(i))Λ ∩ α(i)Λ. µ=0 r(ψµ(i)) < m − dr(ψh(i)), the inductive hypothesis implies that β ∈ α(i)b(h+1)cΛ. Define Since i0 + dPh−1 β′ by β = bd Ph µ=0 r(ψµ(i))β′. Then β′ ∈ bi0+1Λ ∩ α(ψ(h+1)(i))Λ. By (3), β′ ∈ α(ψh+1(i))bcΛ. Therefore β ∈ α(i)b(h+2)cΛ. (cid:3) We next consider case (BS1). We will use the following notation: for i ∈ [0, d)N and k ∈ N, let αk(i) = bi0a · · · bik a. Lemma 2.9. Suppose that we are in case (BS1). There are maps φ : [0, d)N → [0, d)N and r : [0, d)N → (Z+)N such that for all k ∈ N, (1) bdαk(φ(i)) = αk(i)bcr(i)k. φ and r are uniquely determined by this condition, and φ(i)k, r(i)k depend only on i0, . . ., ik. (2) If q < cr(i)k then bi0+1 6∈ [αk(i)bq]. (3) If i0 < m ≤ i0 + d then αk(i)bcr(i)k is the unique minimal common extension of αk(i) and bm. (4) If i0 + (h − 1)d < m ≤ i0 + hd then αk(i)bc Ph−1 µ=0 r(φµ(i))k = bhdαk(φh(i)) is the unique minimal common extension of αk(i) and bm. (5) The map φ is a homeomorphism (for the product topology on [0, d)N). Proof. (1): We define φ and r inductively. Put φ(i)0 = i0 and r(i)0 = 1. Then bdα0(φ(i)) = bdbi0 a = bi0 abc = α0(i)bcr(i)0. It is clear that φ(i)0 and r(i)0 are determined uniquely. Suppose that φ(i)µ and r(i)µ have been defined for µ < k so that bdαµ(φ(i)) = αk(i)bcr(i)µ . Define φ(i)k ∈ [0, d) by Then there is r(i)k ≥ 0 such that (∗) We note that φ(i)k ≡ ik − cr(i)k−1 (mod d). φ(i)k − dr(i)k = ik − cr(i)k−1. and hence that r(i)k ≥ 1. Now, dr(i)k = φ(i)k − ik + cr(i)k−1 ≥ φ(i)k − ik + c > φ(i)k, bdαk(φ(i)) = bdαk−1(φ(i))bφ(i)k a = αk−1(i)bcr(i)k−1bφ(i)k a = αk−1(i)bik+dr(i)k a = αk(i)bcr(i)k . It is clear that φ(i)k and r(i)k are uniquely determined by the equation bφ(i)k+cr(i)k−1a = bik abcr(i)k. More- over, the construction shows that they depend only on i0, . . ., ik. (2): We again prove this by induction on k. Let q < cr(i)0 = c. Then α0(i)bq = bi0 abq is in form (R), and hence does not extend bi0+1. Suppose that the result is true for k − 1. If q < cr(i)k, write q = sc + t, with t ∈ [0, c). Then s < r(i)k. We have αk(i)bq = αk−1(i)bik abscbt = αk−1(i)bik bsdabt. By (∗), ik + sd ≤ ik + (r(i)k − 1)d = ik + r(i)kd − d = φ(i)k + cr(i)k−1 − d < cr(i)k−1. Then the inductive hypothesis implies that αk−1(i)bik+sd 6∈ bi0+1Λ. By uniqueness of form (R), it follows that αk(i)bq 6∈ bi0+1Λ. (3): Let β ∈ αk(i)Λ. Write σαk(i)β = bqabe1 · · · abeℓ in form (R), so eµ ∈ [0, c). By uniqueness of form (R), β ∈ bmΛ if and only if αk(i)bq ∈ bmΛ. By (1) and (2) this occurs if and only if q ≥ cr(i)k, that is, if and only if β ∈ αbcr(i)k Λ. (4): The case h = 1 is given by (3). Suppose the result is true for h − 1. Let i0 + (h − 1)d < m ≤ i0 + hd. The inductive hypothesis implies that b(h−1)dαk(φh−1(i)) is the minimal common extension of bm−d and αk(i). Let β ∈ bmΛ ∩ αk(i)Λ. Then β ∈ b(h−1)dΛ ∩ αk(i)Λ. Therefore β ∈ b(h−1)dαk(φh−1(i)). Writing β = b(h−1)dβ′, we have that β′ ∈ bm−(h−1)dΛ ∩ αk(φh−1(i))Λ. Since i0 < m − (h − 1)d ≤ i0 + d, (3) implies that β′ ∈ αk(φh−1(i))bcr(φh−1(i))k = bdαk(φh(i)). Therefore β ∈ bhdαk(φh(i)). Since bhdαk(φh(i)) extends bm and αk(i), the result is proved. 6 JACK SPIELBERG (5): Given j ∈ [0, d)N, the unique form (L) gives i ∈ [0, d)N and r ∈ [0, d)Z+ such that bdαk(j) = αk(i)bcrk . It follows that j = φ(i), and so φ is onto. By the uniqueness of forms (L) and (R), j and i uniquely determine each other, so that φ is bijective. The continuity of φ follows from the last part of (1). Since [0, d)N is compact, φ is a homeomorphism. (cid:3) Proposition 2.10. Let α = be0 a · · · bes abes+1 and β = bf0a · · · bftabft+1 in form (L), i.e. with eµ, fν ∈ [0, d) for µ ≤ s and ν ≤ t. Then α ⋓ β if and only if eµ = fµ for µ ≤ min{s, t}. In this case, α and β have a unique minimal common extension. (In particular, Λ is finitely aligned.) Proof. The "only if" statement follows from Corollary 2.5. For the converse, suppose that s ≤ t and eµ = fµ for µ ≤ s. We first consider case (BS2). Let bfs+1a · · · bft+1 = bvabi1 · · · bik abq in form (R), i.e. with q, iµ ∈ [0, c) for 1 ≤ µ ≤ k. If es+1 ≤ v, then β ∈ αΛ, and the result holds. Suppose instead that v < es+1. Applying σbe0 a···bes abv to α and β, and letting m = es+1 − v, we find that it is enough to show that the paths bm and abi1a · · · bik abq meet, and have a unique minimal common extension. Choose h as in Lemma 2.8(4). Then abi1 a · · · bik ab(h+1)c is the unique minimal common extension of bm and abi1 a · · · bik a. Since q < c, this is also the unique minimal common extension of bm and abi1a · · · bik abq. Now we consider case (BS1). If es+1 ≤ fs+1 then β ∈ αΛ, and the result holds. Suppose instead that es+1 > fs+1. Applying σbe0 a···bes abfs+1 to α and β, and letting m = es+1 − fs+1, k = t − s − 1, iµ = fµ+s+1 for 1 ≤ µ ≤ k, and q = ft+1, it is enough to show that bm and abi1 · · · bik abq meet and have a unique minimal common extension. Choose h such that (h − 1)d < m ≤ hd. By Lemma 2.9 there are j1, . . ., jk ∈ [0, d), and r ≥ 0, such that bhdabj1 · · · bjk a = abi1 · · · bik abcr is the unique minimal common extension of bm and abi1 · · · bik a. If q ≤ cr this is also the unique minimal common extension of bm and abi1 · · · bik abq. If q > cr, then abi1 · · · bik abq is the unique minimal common extension. Finally we consider case (BS3). If s < t, choose m ≥ 0 such that fs+1 − es+1 + md ≥ 0. Then β = be0 a · · · besabfs+1a · · · = αbfs+1−es+1a · · · = αbfs+1−es+1+mdabmc · · · ∈ αΛ. If s = t, then β ∈ αΛ if es+1 ≤ fs+1, and α ∈ βΛ if es+1 > fs+1. (Thus in case (BS3), α ⋓ β if and only if α and β are comparable.) (cid:3) Theorem 2.11. In cases (BS1) and (BS2), the ordered group (Γ, Λ) is quasi-lattice ordered. Proof. Let t ∈ ΛΛ−1. Let t = αβ−1 with α, β ∈ Λ. Write α = αk(i)bp and β = αℓ(j)bq in form (L). Then t = αk(i)bp−qαℓ(j)−1. If p − q ≡ 0 (mod c), then abp−qa−1 = bn, with n ∈ Z. Then t = αk−1(i)bnαℓ−1(j)−1. If n ≡ 0 (mod c), we may repeat this procedure. Continue this until no more such cancellation is possible. Thus we may assume that t = αk(i)bnαℓ(j)−1 with n 6≡ 0 (mod c). We first consider the case that n ≤ 0. Let µ ∈ tΛ ∩ Λ. There is ν ∈ Λ such that tν = µ; that is, αk(i)bnαℓ(j)−1ν = µ. Since the reduced form of the left hand side is obtained by moving b's to the right (see the remarks after Proposition 2.1), and since no cancellation is possible across bn, it follows that bnαℓ(j)−1ν ∈ Λ. Then by Corollary 2.5 we have µ ∈ αk(i)Λ. Since αk(i) ∈ tΛ ∩ Λ, we have tΛ ∩ Λ = αk(i)Λ. Finally, if n > 0 we have t−1 = αℓ(j)b−nαk(i)−1. By the previous argument we know that t−1Λ ∩ Λ = αℓ(j)Λ. Let µ ∈ tΛ ∩ Λ, and let ν ∈ Λ with tν = µ. Then t−1µ = ν, so ν = αℓ(j)γ for some γ ∈ Λ. Then µ = tαℓ(j)γ = αk(i)bnγ ∈ αk(i)bnΛ. Since αk(i)bn ∈ tΛ ∩ Λ, we have tΛ ∩ Λ = αk(i)bnΛ. (cid:3) In case (BS3), it turns out that the ordered group (Γ, Λ) is not usually finitely aligned, even though Λ is finitely aligned (with unique minimal common extensions). However (Γ, Λ) is always locally finitely exhaustible. Lemma 2.12. Suppose we are in case (BS3). Then (Γ, Λ) is locally finitely exhaustible. It is finitely aligned if and only if c = 1 (and in this case is quasi-lattice ordered). Proof. Let t ∈ ΛΛ−1. As in the proof of Theorem 2.11, we may assume that t = αk(i)bnαℓ(j)−1, where n 6≡ 0 (mod c). We first assume that ℓ ≥ 0, (so that θ(αℓ(j)) > 0). Then αk(i) = tαℓ(j)b−n ∈ tΛ ∩ Λ. If µ ∈ tΛ ∩ Λ, and tν = µ with ν ∈ Λ, then we have αk(i)bnαℓ(j)−1ν = µ. By Corollary 2.5 there is p0 ∈ Z such that αk(i)bq0 ∈ [µ]. Therefore αk(i) ⋓ µ. Now assume that ℓ = −1 (that is, that αℓ(j) ∈ B). If k ≥ 0, then t ∈ αk(i)B ⊆ Λ. If k = −1, then t ∈ B ∪ B−1. Then tΛ ∩ Λ = tΛ, if t ∈ B, and = Λ if t ∈ B−1. To see that (Γ, Λ) is not finitely aligned if c > 1, consider t = aba−1. Then tΛ ∩ Λ =Sn∈Z abnΛ, but for all n ∈ Z, abn 6∈Sk>n abkΛ. C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 7 In the case that c = 1, it is easy to see that ΛΛ−1 = Λ ∪ Λ−1, so that (Γ, Λ) is totally ordered, hence (cid:3) quasi-lattice ordered. We remark that the definition of C ∗(Λ), and hence the rest of the paper, requires only that Λ be a finitely aligned category of paths (Proposition 2.10). The previous results show that the interpretation as an algebra Morita equivalent to the crossed product of Γ acting on its directed boundary is valid. Lemma 2.8 served the purpose of proving that (Γ, Λ) is quasi-lattice ordered in case (BS2). We next give results analogous to Lemma 2.9 for cases (BS2) and (BS3). Lemma 2.13. Assume that we are in case (BS2). (1) There are maps φ : [0, d)N → [0, d)N and r : [0, d)N → {0, 1}N such that for all k ≥ 0 we have bdαk(φ(i)) = αk(i)bcr(i)k. (2) Let ℓ = inf{µ ≥ 1 : iµ ≥ c} (and ℓ = ∞ if iµ < c for µ ≥ 1). Then r(i)k = 1 if and only if k < ℓ. (3) φ is a homeomorphism of [0, d)N for the product topology. Proof. (1) and (2): We define φ(i) by φ(i)µ = i0, iµ + d − c, iℓ − c, iµ, if µ = 0 if 0 < µ < ℓ if µ = ℓ if µ > ℓ. Then bdα0(φ(i)) = bdbi0 a = bi0abc = α0(i)bc. Inductively, for 0 < µ < ℓ, we have bdαµ(φ(i)) = αµ−1(i)bcbd−c+iµa = αµ−1(i)biµ abc = αµ(i)bc. bdαℓ(φ(i)) = αℓ−1(i)bcbiℓ−ca = αℓ(i), If ℓ < ∞, and for µ > ℓ, (3): Let j ∈ [0, d)N. Define i ∈ [0, d)N and r ∈ NN by form (L) of bdαk(j) for all k: bdαµ(φ(i)) = bdαℓ(φ(i))biℓ+1 a · · · biµ a = αµ(i). bdαk(j) = αk(i)bcrk . That iµ is well-defined independently of the choice of k ≥ µ follows from Corollary 2.5. We first show that rk ∈ {0, 1} for all k. When k = 0 we have bdbj0 a = bj0 abc, so r0 = 1. Suppose rk−1 ∈ {0, 1}. Then bdαk(j) = bdαk−1(j)bjk a = αk−1(i)bcrk−1+jk a. If rk−1 = 0, then bdαk(j) = αk−1(i)bjk a, so that ik = jk and rk = 0. If rk−1 = 1, then bdαk(j) = αk−1(i)bc+jk a. Then, if c + jk ≥ d we have ik = jk + c − d and rk = 1, while if c + jk < d we have ik = jk + c and rk = 0. Notice that in the course of the last argument, we showed also that if rk−1 = 0 then rk = 0. Let rµ = 1 for µ ≤ k. Then αk−1(i)bik abc = αk(i)bc = bdαk(j) = bdαk−1(j)bjk a = αk−1(i)bc+jk a. Therefore bik+da = bik abc = bjk+ca, and hence ik + d = jk + c. Thus ik = c + jk − d < c. It follows that j = φ(i), and thus φ is onto. To see that φ is one-to-one, let φ(i) = φ(i′). Then αk(i′)bcr′ k = bdαk(φ(i′)) = bdαk(φ(i)) = αk(i)bcrk . By the uniqueness of form (L) we have that i′ = i. From the definition of φ we see that φ(i)k is locally constant, so that φ is continuous, and hence a homeomorphism. (cid:3) Lemma 2.14. Assume we are in case (BS3). There are maps φ : [0, d)N → [0, d)N and r : [0, d)N → NN such that for all k ≥ 0 we have bdαk(φ(i)) = αk(i)b(−1)kcr(i)k . Moreover, φ is a homeomorphism of [0, d)N for the product topology. 8 JACK SPIELBERG Proof. For t ∈ R let f0(t) = ⌊t⌋ and f1(t) = ⌈t⌉. Let τ (k) = 1 r(i)0 = 1, φ(i)0 = i0 and 2(cid:0)1 − (−1)k(cid:1). We define r and φ recursively by (cid:19) r(i)k = fτ (k)(cid:18) (−1)kik + r(i)k−1c φ(i)k = ik + (−1)k(cid:0)r(i)k−1c − r(i)kd(cid:1). d − ik We first show that r(i)k ≥ 0 for all k. This is true for k = 0; suppose it is true for k − 1. If k is even, d > −1. We next show that φ(i)k ∈ [0, d) for all k. Again, it is true for k = 0. Suppose it is true for k−1. If k is even, then r(i)k = (cid:4) 1 d(cid:0)ik + r(i)k−1c(cid:1)(cid:5) ≥ ⌊0⌋ = 0. If k is odd, r(i)k = (cid:6) 1 d(cid:7) ≥ 0, since d(cid:0)−ik + r(i)k−1c(cid:1)(cid:7) ≥ (cid:6)− ik r(i)k =(cid:4) 1 d(cid:0)ik +r(i)k−1c < r(i)k +1, and hence 0 ≤ ik +r(i)k−1c−r(i)kd < d, d(cid:0)−ik + r(i)k−1c(cid:1)(cid:7), hence r(i)k − 1 < If k is odd, r(i)k = (cid:6) 1 d(cid:0)−ik + r(i)k−1c(cid:1) ≤ r(i)k, and hence 0 ≤ ik −(cid:0)r(i)k−1c − r(i)kd(cid:1) < d, which is the statement that d(cid:0)ik +r(i)k−1c(cid:1)(cid:5), hence r(ik) ≤ 1 which is the statement that φ(i)k ∈ [0, d). 1 Now we check the equation of the statement. It is true for k = 0. Suppose it is true for k − 1. Then φ(i)k ∈ [0, d). bdαk(φ(i)) = bdαk−1(φ(i))bφ(i)k a = αk−1(i)b(−1)k−1cr(i)k−1 bφ(i)k a. Note that (−1)k−1cr(i)k−1 + φ(i)k = ik − (−1)kr(i)kd. Thus bdαk(φ(i)) = αk−1(i)bik −(−1)kr(i)kda = αk−1(i)bik ab(−1)kcr(i)k = αk(i)b(−1)kcr(i)k . From the definition of φ we see that φ(i)k is locally constant, so that φ is continuous. We show that φ is bijective, and hence is a homeomorphism. For this, we construct the inverse of φ. We define maps ψ : [0, d)N → [0, d)N and s : [0, d)N → {0, 1}N recursively by s(i)0 = 1, ψ(i)0 = i0 and (cid:19) s(i)k = fτ (k−1)(cid:18) (−1)k−1ik + s(i)k−1c ψ(i)k = ik + (−1)k−1(cid:0)s(i)k−1c − s(i)kd(cid:1). d The proof that s(i)k ≥ 0 and ψ(i)k ∈ [0, d) for all k is similar to the proof of the analogous facts for r and φ above. We show that r ◦ ψ = s and s ◦ φ = r. We have r(ψ(i))k = fτ (k)(cid:18) (−1)kψ(i)k + r(ψ(i))k−1c d (cid:19) = fτ (k) (−1)k(cid:0)ik + (−1)k−1s(i)k−1c + (−1)ks(i)kd(cid:1) + s(i)k−1c = fτ (k)(cid:18) (−1)kik (cid:19) + s(i)k d d ! = s(i)k; s(φ(i))k = fτ (k−1)(cid:18) (−1)k−1φ(i)k + s(φ(i))k−1c d (cid:19) = fτ (k−1) (−1)k−1(cid:0)ik + (−1)kr(i)k−1c − (−1)kr(i)kd(cid:1) + ri())k−1c = fτ (k−1)(cid:18) (−1)k−1ik d d ! (cid:19) + r(i)k since fτ (n)(cid:0)(−1)nt(cid:1) = 0 for 0 ≤ t < 1. Now we have = r(i)k, ψ(φ(i))k = φ(i)k + (−1)n−1(cid:0)s(φ(i))k−1c − s(φ(i))kd(cid:1) = φ(i)k + (−1)k−1(cid:0)r(i)k−1c − r(i)kd(cid:1) = ik, and similarly, φ(ψ(i))k = ik. (cid:3) C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 9 3. Directed hereditary subsets of Λ We pause a moment to describe the standard picture of Λ. This is just the portion of the Cayley graph of G given by Λ, with direction given by the generators. The set Λ/B then becomes a directed tree -- the portion of the Bass-Serre tree for G corresponding to Λ. We think of the Cayley graph as made up of branching sheets, which we will describe precisely in a moment. The tree is then the side view, where each sheet becomes an infinite path. (See e.g. [5] for a sketch of the case (BS2), c = 1, d = 2.) The precise description of these objects is found in the identification of the directed hereditary subsets of Λ, i.e. of Λ∗. For α ∈ Λ we have the directed hereditary subset [α]. Moreover, since αΛ \ αaΛ \ αbΛ = {α}, it follows that [α] is an open point of Λ∗. Thus Λ is a discrete subset of Λ∗ (where we identify Λ with the set of finite directed hereditary subsets of Λ). n=0[αbn]. It is clear that [αB] is hereditary and directed, hence is an element of Λ∗. Letting α = bi0 a · · · bik a in form (L), it follows from Corollary 2.5 that α is uniquely determined µ=0 αbµaΛ. Then Next, for αB ∈ Λ/B we let [αB] =S∞ by [αB]. Thus we may identify {[αB] : α ∈ Λ} with Λ/B. For m ≥ 0 set Um = αbmΛ \Sd−1 Um = {αbj : j ≥ m} defines a neighborhood cUm of [αB] in Λ∗. Thus Λ/B is a relatively discrete subset of Definition 3.1. An element C ∈ Λ∗ is of finite height if θ is bounded on C. An element of Λ∗ not of finite height is said to be of infinite height. Λ∗ \ Λ. Lemma 3.2. The elements of Λ∗ of finite height are {[α] : α ∈ Λ} ∪ {[ξ] : ξ ∈ Λ/B}. Proof. Let C ∈ Λ∗ have finite height. Let α ∈ C have maximal height. By the hereditary property of C we may assume that α = bi0 a · · · bik a in form (L). By Corollary 2.5, every element of C of height k is of the form αbp for some p. If {p : αbp ∈ C} has a maximum element q, then C = [αbq]. Otherwise, we have C = [αB]. (cid:3) Remark 3.3. The elements Λ/B form a directed tree when ordered by containment; this follows from Corollary 2.5. For α = bi0 a · · · bik a, the immediate successors of [αB] in this tree are [αaB], [αbaB], . . ., [αbd−1aB]. Thus if we orient the tree so as to be directed upward, each vertex has d upward edges (and, apart from B, one downward edge). We now consider the the directed hereditary subsets having infinite height. Definition 3.4. Let i ∈ [0, d)N. C0(i) = C∞(i) = ∞[k=0 ∞[k,p=0 [αk(i)] [αk(i)bp]. Lemma 3.5. (1) C0(i) and C∞(i) are directed hereditary subsets of Λ. (2) C0(i) ⊆ C∞(i). (3) (cid:8)C∞(i) : i ∈ [0, d)N(cid:9) = Λ∗∗. (4) If C ∈ Λ∗ is of infinite height, there is a unique i ∈ [0, d)N such that C ⊆ C∞(i). Moreover C0(i) ⊆ C. Proof. (1) C0(i) is directed since it is an increasing union of directed sets. C∞(i) is directed by Lemmas 2.8(4) and 2.9(4). Both are hereditary since they are unions of hereditary sets. (2) is immediate. (3) To see that C∞(i) is a maximal directed hereditary subset, let β ∈ Λ \ C∞(i). Write β = bj0 a · · · bjk abp in form (L). Then there is ℓ ≤ k such that jℓ 6= iℓ. By Proposition 2.10 we have that β ⊥ αk(i), and hence there cannot exist a directed hereditary subset containing β and C∞(i). Now we show that these are all of the maximal elements of Λ∗∗. Let C ∈ Λ∗. Since any two elements of C have a common extension, Proposition 2.10 implies that there is i ∈ [0, d)N such that C ⊆S∞ (4) If C ∈ Λ∗ is of infinite height, the sequence i in the proof of part (3) is uniquely determined by C. (cid:3) For each k there is p ≥ 0 such that αk(i)bp ∈ C. Therefore αk(i) ∈ C for all k, and hence C0(i) ⊆ C. k,p=0[αk(i)bp] = C∞(i). 10 JACK SPIELBERG We now describe precisely the sheets mentioned above. Namely, the sheets are the subsets C∞(i) of Λ, for i ∈ [0, d)N. The cosets(cid:8)αk(i)B : k ∈ N(cid:9) define an infinite path in the tree Λ/B. Thus we may identify the boundary of Λ with the boundary of the directed tree Λ/B. Lemma 3.6. Let i ∈ [0, d)N and let C ∈ Λ∗ with C0(i) ⊆ C. Then C = C∞(i) if and only if B ⊆ C. Proof. Of course, the hypothesis C0(i) ⊆ C is equivalent to assuming that C ⊆ C∞(i) and that C is of infinite height. We first consider case (BS1). By Lemma 2.9(4) we have Thus B ∨ C0(i) ⊇ C∞(i). Conversely, given h ≥ 0 let m = i0 + hd. Then B ∨ C0(i) ⊇(cid:2)bm ∨ αk(i)(cid:3) =(cid:2)αk(i)bc Ph−1 bm ∈(cid:2)bm ∨ αk(i)(cid:3) =(cid:2)αk(i)bc Ph−1 µ=0 r(φµ(i))k(cid:3). µ=0 r(φµ(i))k(cid:3) ⊆ C∞(i). Hence B ⊆ C∞(i). Now we consider case (BS2). Fix k. There are i′ ∈ [0, c)k and q ≥ 0 such that αk(i) = bqα(i′) in form µ=0 r(ψµ(i)) (where ψ is as in Lemma 2.8). By Lemma 2.8(4) we (R). Let h ≥ 0, and choose m = q + dPh have bm ∨ αk(i) = bq(cid:0)bm−q ∨ α(i′)(cid:1) = bqα(i′)b(h+1)c = αk(i)b(h+1)c. If B ⊆ C, then since h was arbitrary, we see that C∞(i) ⊆ B ∨ C0(i) ⊆ C. Conversely, for all h we have Thus B ⊆ C∞(i). bhd ∈(cid:2)bd Ph−1 µ=0 r(ψµ(i))α(ψh(i))(cid:3) =(cid:2)α(i)bhc(cid:3) ⊆ C∞(i). Finally we consider case (BS3). By Lemma 2.6, [αk+1(i)] = [αk(i)bik+1 a] ⊇ [αk(i)B]. Thus C0(i) = C∞(i), (cid:3) and of course, B ⊆ C∞(i). Corollary 3.7. Λ∗∗ is a closed subset of Λ∗ (and thus ∂Λ = Λ∗∗). The map C∞(i) 7→ i of ∂Λ → [0, d)N is a homeomorphism, equivariant for the maps bd· on ∂Λ, and φ−1 on [0, d)N (from Lemmas 2.9, 2.13, and 2.14). Proof. It was pointed out in the remarks before Definition 3.1 that the subset of Λ∗ consisting of elements of finite height is an open subset of Λ∗. Thus we must show that if C 6∈ Λ∗∗ is of infinite height, then C has a neighborhood disjoint from Λ∗∗. By Lemma 3.6 we know that m = max{ℓ : bℓ ∈ C} is finite. Then Λ∗ \ \bm+1Λ is a neighborhood of C disjoint from Λ∗∗. It is clear that C∞(i) 7→ i is a bijection of ∂Λ with [0, d)N. Given (j0, . . . , jk) ∈ [0, d)k+1, the open set (bj0 a · · · bjk aΛ)b corresponds to the cylinder set defined by (j0, . . . , jk), showing that the map is a homeomor- phism. The equivariance is clear from the three lemmas mentioned. (cid:3) Remark 3.8. It follows from Corollary 3.7 that left-concatenation by b defines a homeomorphism of ∂Λ. Definition 3.9. Let Σ ≡ Σ(Λ) = Λ∗ \(cid:0)Λ ∪ (Λ/B) ∪ Λ∗∗(cid:1). Lemma 3.10. Σ is a relatively open subset of Λ∗ \ Λ. Proof. Let C ∈ Σ. Then m = sup{j : bj ∈ C} < ∞. Then Λ∗ \ \bm+1Λ is a neighborhood of C disjoint from (Λ/B) ∪ Λ∗∗. (cid:3) We now see that G0 = Λ∗ is the disjoint union of four invariant subsets: Λ∗ = Λ ⊔ (Λ/B) ⊔ ∂Λ ⊔ Σ, where the meanings of the first two subsets were specified in the remarks before Definition 3.1 (in case (BS3), the set Σ is empty). The subset Λ is discrete. It was shown in the remarks before Definition 3.1 that the subset Λ/B is relatively discrete, hence open, in Λ∗ \ Λ. We now have seen that Σ is also open in Λ∗ \ Λ. Let G = G(Λ) be the groupoid of Λ. We wish to prove that G is amenable. We note that since the maximal degree functor θ is degenerate, the fixed-point groupoid is not AF. However, we may still use Proposition 1.3. The first step is to show that the restriction of G to ∂Λ is amenable. Theorem 3.11. G∂Λ is amenable. C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 11 Proof. By Proposition 1.3, it suffices to show that Gθ∂Λ is amenable. Let H = Gθ∂Λ. Thus H =(cid:8)[α, β, x] : θ(α) = θ(β), x ∈ ∂Λ(cid:9). For k ∈ N let Hk =(cid:8)[α, β, x] ∈ H : θ(α) = θ(β) = k(cid:9). Then H =S∞ show that Hk is a subgroupoid of H. It is clear that Hk is closed under inversion. Let(cid:0)[α, β, x], [γ, δ, y](cid:1) ∈ k . Since βx = γy, we have that in form (L), β = bi0a · · · βik abp and γ = bi0a · · · bik abq. Thus y = bp−qx H (2) (cf. Remark 3.8). Therefore, letting ε = bi0 a · · · bik a, we have k=0 Hk. First we [α, β, x][γ, δ, y] = [α, εbpx][εbq, δ, bp−qx] = [αbq, εbp+q, b−qx][εbp+q, δbp, b−qx] = [αbq, δbp, b−qx] ∈ Hk. Next we show that Hk ⊆ Hk+1. Let θ(α) = θ(β) = k, and [α, β, x] ∈ Hk. Write x = C∞(i) and y = C∞(σ(i)), where we use σ to denote the (noninvertible) left shift on [0, d)N. Then [α, β, x] = [αbi0 a, βbi0 a, y] ∈ Hk+1. Note also that Hk = [i,j∈[0,d)k [p,q∈N [bi0 a · · · bik−1 abp, bj0a · · · bjk−1 abq, ∂Λ] is an open subgroupoid of H, hence also of Hk+1. Now we observe that the map [bp, bq, x] 7→ (p − q, x) is an isomorphism of H0 onto the transformation groupoid Z ⋉ ∂Λ. Thus H0 is amenable. Finally, if we write the multiplication in Hk as [bi1 a · · · bik abp, bj1 a · · · bjk abq, x][bj1 a · · · bjk abq, bℓ1a · · · bℓk abr, x] = [bi1 a · · · bik abp, bℓ1a · · · bℓk abr, x], we see that Hk is isomorphic to the product groupoid(cid:0)[0, d)k × [0, d)k(cid:1) × H0, which is amenable. Therefore H is amenable (by [10], III.1). (cid:3) Before proving that G is amenable, we require a detailed description of the remaining elements of Λ∗ having infinite height, that is, the elements of Σ. We first consider case (BS1). Let i ∈ [0, d)N. Let n0, n1, . . . ∈ N satisfy (∗∗) cnℓ−1 − iℓ d ≤ nℓ < c(nℓ−1 + 1) − iℓ d , ℓ ≥ 1. Note that since the outside terms of these inequalities differ by c d ≥ 1, such sequences exist for any choice of n0. We let n = (n0, n1, . . .), and set Cn(i) = [αℓ−1(i)bnℓd+iℓ]. ∞[ℓ=0 Lemma 3.12. The sets in the above union increase, and hence Cn(i) is directed and hereditary. Proof. Since cnℓ ≤ nℓ+1d + iℓ+1, we have αℓ−1(i)bnℓd+iℓ ∈ [αℓ−1(i)bnℓd+iℓa] = [αℓ(i)bnℓc] ⊆ [αℓ(i)bnℓ+1d+iℓ+1]. (cid:3) We remark that what we have denoted C0(i) equals the directed hereditary set defined as above for the sequence n consisting entirely of zeros. Lemma 3.13. αℓ−1(i)b(nℓ+1)d 6∈ Cn(i). Proof. We will show that for m ≥ ℓ we have αℓ−1(i)b(nℓ+1)d 6∈ [αm−1(i)bnmd+im ]. Since iℓ < d, this is true when m = ℓ. Let m ≥ ℓ, and suppose inductively that biℓa · · · bim−1 abnmd+im = bnℓd+iℓabjℓ+1 · · · abjm , where jℓ+1, . . ., jm ∈ [0, c) (this is true vacuously when m = ℓ). Since cnm ≤ dnm+1 + im+1 < c(nm + 1), we have that abnm+1d+im+1 = bnmdabjm+1, where jm+1 ∈ [0, c). Then biℓa · · · bimabnm+1d+im+1 = biℓa · · · bim−1 abnmd+imabjm+1 = bnℓd+iℓabjℓ+1 · · · abjm+1. Since this is in form (R), we see that b(nℓ+1)d 6∈ [biℓa · · · bnm+1d+im+1], since (nℓ + 1)d > nℓd + iℓ. Therefore αℓ−1(i)b(nℓ+1)d 6∈ [αm(i)bnm+1d+im+1]. (cid:3) Corollary 3.14. nℓ = max(cid:8)m : αℓ−1(i)bmd ∈ Cn(i)(cid:9). Lemma 3.15. Let n and n′ both satisfy the inequalities (∗∗). Suppose that nk = n′ nℓ < n′ ℓ. Then Cn(i) ( Cn′ (i). k for k < ℓ, and that 12 JACK SPIELBERG Proof. By Corollary 3.14 we know that αℓ−1(i)bn′ n′ ℓ ≥ nℓ + 1, we have ℓd ∈ Cn′ (i) \ Cn(i), and hence Cn(i) 6= Cn′ (i). Since c(nℓ + 1) − iℓ+1 nℓ+1 < ≤ cn′ ℓ − iℓ+1 d ≤ n′ ℓ+1. d Inductively we find that nk < n′ k for k ≥ ℓ, and hence that Cn(i) ⊆ Cn′ (i). (cid:3) Lemma 3.16. Let C ∈ Λ∗ with C0(i) ⊆ C ( C∞(i). Then there exists n = (n0, n1, . . .) satisfying the inequalities (∗∗) such that C = Cn(i). Proof. Since C 6= C∞(i), Lemma 3.6 implies that B 6⊆ C. Thus we may define nℓ = max(cid:8)m : αℓ−1(i)bmd ∈ C(cid:9). Thus αℓ−1(i)bnℓd ∈ C and αℓ−1(i)b(nℓ+1)d 6∈ C. Let C ′ = σαℓ−1(i)C. Then bnℓd ∈ C ′ and biℓa ∈ C ′. Therefore bnℓd+iℓa = bnℓd ∨ biℓ a ∈ C ′. Thus bnℓd+iℓ ∈ C ′, and hence αℓ−1(i)bnℓd+iℓ ∈ C. We claim that the sequence n = (n0, n1, . . .) satisfies the inequalities (∗∗). For the strict inequality, suppose that iℓ + dnℓ ≥ c(nℓ−1 + 1). Then αℓ−1(i)bc(nℓ−1+1) ∈ C, and hence suppose otherwise; i.e. αℓ−2(i)biℓ−1+d(nℓ−1+1) ∈ C. But this contradicts the definition of nℓ−1. For the weak inequality, we already know that αℓ−2(i)biℓ−1+dnℓ−1 ∈ C. Thus biℓ−1+dnℓ−1, biℓ−1 a ∈ σαℓ−2(i)C. Therefore Hence αℓ−1(i)bcnℓ−1 ∈ C. On the other hand, we know that αℓ−1(i)biℓ+nℓd ∈ C. Moreover, by Lemma 2.9(4), we have that biℓ−1 abcnℓ−1 = biℓ−1+dnℓ−1 ∨ biℓ−1 a ∈ σαℓ−2(i)C. bnℓd+iℓ+1 ∨ biℓa = biℓabc(nℓ+1). Hence if αℓ−1(i)bnℓd+iℓ+1 ∈ C, then αℓ−1(i)b(nℓ+1)d+iℓa = αℓ(i)bc(nℓ+1) ∈ C, contradicting the definition of nℓ. Therefore nℓd + iℓ = max(cid:8)m : αℓ−1(i)bm ∈ C(cid:9). Thus we conclude that cnℓ−1 ≤ nℓd + iℓ. This finishes the demonstration that n satisfies the inequalities (∗∗). By its definition, we have Cn(i) ⊆ C. Finally, we claim that Cn(i) = C. For this consider a typical element β ∈ C. Then β = αℓ−1(i)bq for (cid:3) some ℓ and q. Then from the above argument we must have q ≤ nℓd + iℓ, and hence that β ∈ Cn(i). Corollary 3.17. Σ = {Cn(i) : i ∈ [0, d)N, n satisfies the inequalities (∗∗) for i}. We will now study the restriction of G to Σ. Lemma 3.18. Recall the maps φ and r from Lemma 2.9. Let i ∈ [0, d)N. (1) Let n satisfy the inequalities (∗∗) for φ(i). Set n′ = n + r(i). Then n′ satisfies the inequalities (∗∗) for i. Moreover bdCn(φ(i)) = Cn′ (i). (2) Let n′ satisfy the inequalities (∗∗) for i, and suppose that n′ 0 ≥ 1. Then n′ ≥ r(i), and n = n′ − r(i) satisfies the inequalities (∗∗) for φ(i). Proof. (1): n satisfies the inequalities (∗∗) for φ(i) if and only if (for all ℓ,) Using equation (∗) from Lemma 2.9, this is equivalent to cnℓ−1 ≤ dnℓ + φ(i)ℓ ≤ cnℓ−1 + c. cnℓ−1 ≤ dnℓ + iℓ + dr(i)ℓ − cr(i)ℓ−1 ≤ cnℓ−1 + c, and hence to Then we have bdCn(φ(i)) = = = cn′ ℓ−1 ≤ dn′ ℓ + iℓ ≤ cn′ ℓ−1 + c. ∞[ℓ=0(cid:2)αℓ−1(i)bnℓd+φ(i)ℓ+cr(i)ℓ−1(cid:3), by Lemma 2.9(1), ∞[ℓ=0(cid:2)bdαℓ−1(φ(i))bnℓd+φ(i)ℓ(cid:3) = ∞[ℓ=0(cid:2)αℓ−1(i)bnℓd+iℓ+dr(i)ℓ(cid:3), by equation (∗), ∞[ℓ=0(cid:2)αℓ−1(i)bn′ ℓd+iℓ(cid:3) = Cn′ (i). C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 13 (2): We show that n′ ℓ ≥ r(i)ℓ for all ℓ ∈ N. By assumption this holds for ℓ = 0. Suppose it holds for ℓ − 1. Then using the inequalities (∗∗) for n′, and equation (∗), we have dn′ ℓ ≥ cr(i)ℓ−1 − iℓ = dr(i)ℓ − φ(i)ℓ > dr(i)ℓ − d. Therefore n′ satisfies the inequalities (∗∗) for φ(i). ℓ > r(i)ℓ − 1, and hence n′ ℓ ≥ r(i)ℓ. Thus n = n′ − r(i) ≥ 0. The proof of (1) shows that n (cid:3) Theorem 3.19. Assume that we are in case (BS1). Then GΣ is amenable. Proof. Since the action of b on Σ is not surjective, the argument differs in a few places from that of Theorem 3.11. By Proposition 1.3, it suffices to show that GθΣ is amenable. Let M = GθΣ = (cid:8)[β, γ, x] : x ∈ k=0 Mk, where Mk =(cid:8)[β, γ, x] ∈ M : θ(β) = θ(γ) = k(cid:9). Letting σ denote Σ, θ(β) = θ(γ)(cid:9). Write M =S∞ the left shift on sequences (as well as the left shift on Λ), we have σbi0 aCn(i) = σbi0 a ∞[ℓ=0(cid:2)αℓ−1(i)bnℓd+iℓ(cid:3) = ∞[ℓ=0(cid:2)bi1a · · · biℓabnℓ+1d+iℓ+1(cid:3) ∞[ℓ=0(cid:2)αℓ−1(σ(i))bσ(n)ℓd+σ(i)ℓ(cid:3) = Cσ(n)(σ(i)). = Thus if θ(β) = θ(γ) = k, [β, γ, Cn(i)] =(cid:2)β, γ, bi0aCσ(n)(σ(i))(cid:3) =(cid:2)βbi0 a, γbi0a, Cσ(n)(σ(i))(cid:3) ∈ Mk+1. It is clear that Mk is closed under inversion. If [α, β, Cm(i)] and [γ, δ, Cn(j)] are composable elements of Mk, then βCm(i) = γCn(j) and θ(β) = θ(γ) = k. Then in form (L) we have β = bµ0 a · · · bµk−1 abp and γ = bµ0 a · · · bµk−1 abq, with p, q ≥ 0. Without loss of generality, assume that p ≤ q. Then γ = βbq−p, and Cm(i) = bq−pCn(j). Then [α, β, Cm(i)] [γ, δ, Cn(j)] = [α, β, bq−pCn(j)] [βbq−p, δ, Cn(j)] = [αbq−p, δ, Cn(j)] ∈ Mk. In the same way as in the proof of Theorem 3.11 we have that Mk is isomorphic to(cid:0)[0, d)k × [0, d)k(cid:1) × M0, and also that Mk is an open subgroupoid of Mk+1. Thus it suffices to prove that M0 is amenable. We have M0 =(cid:8)[bp, bq, x] : x ∈ Σ, p, q ≥ 0(cid:9). We claim that [bp, bq, x] 7→ p − q is a continuous homomorphism from M0 to Z. Since the kernel is just the diagonal {(x, x) : x ∈ Σ}, hence amenable, Proposition 1.3 will imply that M0 is amenable. We have only to show that the homomorphism is well-defined. Let [bp, bq, Cm(i)] = [br, bs, Cn(j)]. Then comparing sources gives bqCm(i) = bsCn(j). Without loss of generality we suppose that q ≤ s. Then Cm(i) = bs−qCn(j). Comparing ranges, we then have brCn(j) = bpCm(i) = bp−q+sCn(j). Since the action of b on Λ∗ is injective, we have that r = p − q + s, hence r − s = p − q. (cid:3) We now turn to case (BS2). Lemma 3.20. Suppose that we are in case (BS2). Let i ∈ [0, d)N. For n ∈ N, let Cn(i) =S∞ Put s = lim supµ→∞ iµ. k=0(cid:2)αk−1(i)bnc(cid:3). (1) Cn(i) is directed and hereditary, and Cn(i) ⊆ Cn+1(i) ⊆ C∞(i) for n ∈ N. (2) If s ≥ c then C0(i) = C∞(i). (3) If s < c then C0(i) 6= C1(i). If C is a directed hereditary set with C0(i) ( C then C1(i) ⊆ C. (4) Suppose s < c, and let m =(cid:6) c−s d−c(cid:7). Then C0(i) ( C1(i) ( · · · ( Cm−1(i) ( C∞(i), and these are the only directed hereditary sets of infinite height contained in C∞(i). Proof. (1): We have αk(i)bnc = αk−1(i)bik abnc = αk−1(i)bik+nda = αk−1(i)bnc · bik+n(d−c)a. Therefore [αk−1(i)bnc] ⊆ [αk(i)bnc], and therefore the union defining Cn(i) is increasing. Hence Cn(i) is a hereditary directed set. Since [αk(i)bnc] ⊆ [αk(i)b(n+1)c], it follows that Cn(i) ⊆ Cn+1(i). It is clear that Cn(i) ⊆ C∞(i). 14 JACK SPIELBERG (2): We note that for any β ∈ Λ we have bd ∈ [βabc] (e.g. from Lemma 2.13(1), since r(i)k ≤ 1 in that lemma). Thus if 1 ≤ l1 < · · · < lm are such that iℓµ ≥ c for 1 ≤ µ ≤ m, then bmd ∈ [αℓm(i)]. If p ≥ 0 is given, there is m such that md ≥ p. Since s ≥ c there are ℓµ as above, 1 ≤ µ ≤ m. Then bp ∈ [bmd] ⊆ [αℓm(i)] ⊆ C0(i). Therefore B ⊆ C0(i), and by Lemma 3.6 it follows that C0(i) = C∞(i). (3): Since s < c there is ℓ such that iµ < c for µ ≥ ℓ. Let β = αℓ−1(i). Since biℓ a · · · bik a is in form (R) for k ≥ ℓ, we have On the other hand, bd 6∈ [k≥ℓ [biℓ a · · · bik a] = σβ [k≥ℓ [αk(i)] = σβ(C0(i)). βbd+iℓa = βbiℓabc = αℓ(i)bc ∈ C1(i), so that bd ∈ σβ(C1(i)). Therefore C0(i) 6= C1(i). Let C ) C0(i) be a directed hereditary set. Choose γ ∈ C \ C0(i). We may assume that k = θ(γ) ≥ ℓ. Then γ = bi0 a · · · bik abq, where q > ik+1. Let i′ = σk+1(i) and C ′ = σαk(i)(C). Then C0(i′) ( C ′ and bq ∈ C ′, where q > i′ 0 = ik+1. By Lemma 2.8(3), for each µ, Therefore C1(i′) ⊆ C ′. Hence αµ(i′)bc ∈ [bq ∨ αµ(i′)] ⊆ C ′. C1(i) =[µ [αµ(i)bc] = αk(i)[µ [αµ(i′)bc] ⊆ αk(i)C ′ = C. (4): We have that s + (m − 1)(d − c) < c ≤ s + m(d − c). Let h be such that ik ≤ s for k ≥ h. Set j = σh(i) ∈ [0, s]N. Then for each n, Cn(i) = αh−1(i)Cn(j). The definition of φ in the proof of Lemma 2.13 implies that φn(j)k = jk + n(d − c), k ≥ 1, for 0 ≤ n ≤ m, and that φn(j) ∈ [0, c)N for n < m, while φm(j)k ≥ c for infinitely many k. Hence by Lemma 2.13(1), for 0 ≤ n < m we have αk(j)bnc = bndαk(φn(j)), for k ≥ 1, and hence by (3) that σbnd Cn(j) = C0(φn(j)) ( C1(φn(j)) = σbnd Cn+1(j), with no directed hereditary set strictly between them. Therefore Cn(j) ( Cn+1(j) with no directed hereditary set strictly between them. Since φm(j)k ≥ c infinitely often, it follows from (2) that Cm(j) = C∞(j), finishing the proof. (cid:3) Theorem 3.21. Suppose that we are in case (BS2). Then GΣ is amenable. (In fact, it is Morita equivalent to the standard groupoid of Oc.) Proof. We let Z = (cid:8)C0(i) : i ∈ [0, c)N(cid:9) ⊆ Σ. Note that for C ∈ Σ, C ∈ Z if and only if bd 6∈ C. Thus Z = (dbdΛ)c is a compact-open subset of Σ. We claim that it is a transversal, in the sense of [8]. To see this, let C ∈ Σ. Then m = sup{µ : bµ ∈ C} < ∞. Let C ′ = σbm has source C and range in Z. It follows from [8] that GΣ is equivalent to GZ . (C). Then bd 6∈ C ′, so C ′ ∈ Z. Then [e, bm, C ′] To analyze GZ we first consider a pair β ∈ Λ and C ∈ Σ such that βC ∈ Z. Since bd 6∈ βC, we must have bd 6∈ C, hence C ∈ Z. Moreover, writing β = bm0a · · · bmk abmk+1 in form (L), and C = C0(i) with i ∈ [0, c)N, we must have m0, . . ., mk ∈ [0, c) and mk+1 ∈ [0, c − i0). Then βC = C0(f (m, i)), where f (m, i) = (m0, . . . , mk, mk+1 + i0, i1, i2, . . .). Now, if [β, γ, C] ∈ GZ, let β and C be as above, and let γ = bn0a · · · bnℓabnℓ+1 in form (L). Then the map [β, γ, C] 7→(cid:0)f (m, i), k − ℓ, f (n, i)(cid:1) defines an isomorphism of GZ onto (cid:8)(x, p, y) ∈ [0, c)N × Z × [0, c)N : xµ−p = yµ for all large enough µ(cid:9), namely the standard groupoid for the Cuntz algebra Oc. This groupoid is well-known to be amenable (e.g. [10]). Therefore GZ is amenable. (cid:3) Theorem 3.22. G is amenable. C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 15 Proof. Since G∂Λ is amenable (Theorem 3.11), by [10] it suffices to prove that G(∂Λ)c is amenable. It is easy to see that GΛ is amenable -- for any finitely aligned category of paths, GΛ is the direct sum over v ∈ Λ0 of the elementary groupoids s−1(v) × s−1(v). Since Λ/B and Σ are disjoint relatively open subsets of Λc, it remains to show that GΛ/B is amenable (GΣ was shown to be amenable in Theorems 3.19 and 3.21). This follows since GΛ/B is a transitive groupoid, with isotropy isomorphic to Z (generated by b). (cid:3) We now apply Theorem 1.2 to give generators and relations for C ∗(Λ). Theorem 3.23. Let Λ be the category of paths associated to the Baumslag-Solitar group G. The represen- tations of C ∗(Λ) are in one-to-one correspondence with pairs {Sa, Sb} of Hilbert space operators satisfying the relations b Sa in cases (BS1) and (BS2); Sd b SaSc b = Sa in case (BS3). (1) Sa and Sb are isometries. (2) Sb is a unitary. (3) SaSc b = Sd (4) Si bSaS∗ aS−i b = 1. d−1Xi=0 Moreover, in case (BS3), relation (2) is redundant. Proof. By Theorem 1.2, we know that representations of C ∗(Λ) are in one-to-one correspondence with families {Sα : α ∈ Λ} satisfying the relations (1) - (4) of Theorem 1.2. Given such a family, we check that the operators Sa and Sb satisfy conditions (1) - (4). From Theorem 1.2(1) it follows that (1) holds, and that Sb is an isometry. Then (2) follows from Theorem 1.2(4) applied to the finite exhaustive set {b}. From Theorem 1.2(2) we conclude that (3) holds. Since bia ⊥ bja when 0 ≤ i 6= j < d, it follows from Theorem 1.2(3) that Si b Sa have orthogonal ranges when 0 ≤ i 6= j < d. Theorem 1.2(4) applied to the finite exhaustive set {bia : 0 ≤ i < d} verifies (4). bSa and Sj Conversely, suppose that Sa and Sb are given satisfying (1) - (4). We define Sα for all α ∈ Λ by setting Se = 1, and for α = bi0a · · · bik abp in form (L), setting Sα = Si0 b . The proof (e.g. in [11]) of Proposition 2.3 uses only the relation abc = b±da. Thus relation (3) (and relation (2) in case (BS3)) implies that Theorem 1.2(2) holds. Relations (1) and (2) imply that Theorem 1.2(1) holds. b Sa · · · Sik b SaSp Now we verify Theorem 1.2(3). Let α, β ∈ Λ. If α ⊥ β, let them be written as in the statement of Proposition 2.10. Then by that Proposition, there is ℓ ≤ min{s, t} such that eµ = fµ for µ < ℓ, and eℓ 6= fℓ. Then we may compute: S∗ b Sa have orthogonal ranges. Since α ⊥ β, this verifies Theorem 1.2(3) in this case. Suppose that α ⋓ β. If e.g. β ∈ αΛ, then we find that SαS∗ β, and β = α ∨ β. Suppose instead that neither of α and β extends the other. The proof of Proposition 2.10 shows that (without loss of generality) we may assume that α = γbm and β = γabi1a · · · bik abq, and that α ∨ β = γabi1a · · · bik abp for some p. Since SbS∗ b = 1 by (2), the final factors of Sb do not affect the computation of SS∗. Hence SαS∗ αSβ = · · · (beℓa)∗(bfℓ a) · · · = 0, since (4) implies that Seℓ b Sa and Sfℓ β = SβS∗ αSβS∗ β = SβS∗ β = Sα∨βS∗ αSβS∗ α∨β. Finally we verify Theorem 1.2(4). Let F be a finite exhaustive set. If v ∈ F then Theorem 1.2(4) is immediately satisfied. Suppose v 6∈ F . First we suppose that bm ∈ F for some m ≥ 1. Then the right-hand side of Theorem 1.2(4) dominates SbmS∗ αb = SαS∗ α, we may assume that elements of F have the form bi0a · · · bik a (in form (L)). Let us identify such elements with cylinder sets in [0, d)N via the sequences (i0, . . . , ik). Moreover, because F is exhaustive we have that bm = 1, by (2). Now suppose that F ∩ B = ∅. Since SαbS∗ these cylinder sets form a cover of [0, d)N. Thus we see thatWα∈F SαS∗ b = S∗ a(S∗ b )dSa. Now we have from (3): it follows from (3) that Sc α = 1. For the final statement of the theorem, assume that we are in case (BS3). Since Sa and Sb are isometries, and hence SaS∗ a = Sd b SaS∗ a(S∗ b )dSaS∗ a. Thus SaS∗ b SaS∗ a(S∗ Sa = Sd b SaSc b SaS∗ a(S∗ b = Sd a ≤ Sd b )dSa, b )d. Now using (4) gives SbS∗ b = dXi=1 This proves (2). Si bSaS∗ a(S∗ b )i = d−1Xi=1 Si bSaS∗ a(S∗ b )i + Sd b SaS∗ a(S∗ b )d ≥ Si bSaS∗ a(S∗ b )i + SaS∗ a = 1. d−1Xi=1 (cid:3) 16 JACK SPIELBERG Remark 3.24. The "Cuntz-Krieger" relation (Theorem 1.2(4)), defining the C ∗-algebra from the Toeplitz C ∗-algebra of Λ, is represented by (2) and the equality (as opposed to ≤) in (4). We point out here that if the group falls under case (BS1) or (BS2), then both of these relations are necessary. To see this, we consider the representations of T C ∗(Λ) on ℓ2(Λ/B) and on ℓ2(Σ). The first of these satisfies (2) but not (4); the second satisfies (4) but not (2). Remark 3.25. The relations 1.2(1) - (4) are the same as those found by Katsura ([6], Example A.6). Specifically, our C ∗(Λ) is isomorphic to Katsura's O(En,m) for m 6= 0, with the identifications d = n, and c = m in case (BS1) or (BS2) when m > 0, and c = −m in case (BS3) when m < 0. We remark that in the case where m < 0, our analysis shows that one of the relations for the C ∗-algebra turns out to be redundant. (In the case that m = 0, the group is not one of those discussed in [3]. In fact, if c = 0, the group becomes Z ∗ Z/dZ. If d > 1 then Λ is not a category of paths, since (3) fails. If d = 1, then G = Z, and Λ is the path category of the directed graph having one vertex and one edge, giving the same result as [6].) 4. K-theory Our next task is to compute the K-theory of C ∗(Λ) = C ∗(G∂Λ). We give a different calculation than that of [6]. Thus we also compute the K-theory of the core algebra. Let A = C ∗(Λ). We let γ denote the gauge action of T on A induced by the cocycle θ : Λ → Z. Then Takai-Takesaki duality implies that A ⊗ K ∼= (A ×γ T) ×bγ Z. Lemma 4.1. The fixed-point algebra Aγ is Morita equivalent to A ×γ T. Proof. Let ζ ∈ C(T) be the function ζ(z) = z. Then the collection {ζnSαS∗ β : n ∈ Z, α, β ∈ Λ} ⊆ C(T, A) ⊆ A×γ T is a total set. A short calculation in the convolution algebra C(T, A) shows that (ζmSαS∗ ν ) = β : θ(α) = θ(β)} is a total set in Aγ. Now let δm,n−θ(µ)+θ(ν)ζnSαS∗ n ∈ Z and α, β ∈ Λ. Choose k ≥ 0 such that n + θ(β) + k ≥ 0, and let ν ∈ Λ with θ(ν) = n + θ(β) + k. Let M = {bi1a · · · bik a : ij ∈ [0, d) for 1 ≤ j ≤ k}. Then SαS∗ ν . Of course, the collection {SαS∗ β)(ζnSµS∗ βSµS∗ βµ. We have that ζnSαS∗ ζnSαµS∗ β = Xµ∈M β =Pµ∈M SαµS∗ (SαµS∗ ν )(Sν S∗ ν )(ζnSν S∗ βµ) βµ = Xµ∈M is in the ideal generated by Aγ. Thus Aγ is a full hereditary subalgebra of A ×γ T. (cid:3) We next compute the K-theory of Aγ. We know that Aγ = C ∗(H), where H = Gγ∂Λ. Recall from the n=0 Hn, that H0 ∼= Z ⋉ ∂Λ, and that Hn ∼= ([0, d)n × [0, d)n) × H0. So we begin with the computation of K∗(C ∗(H0)). Since C ∗(H0) = C(∂Λ) ×b Z, and ∂Λ is totally disconnected, we obtain from the Pimsner-Voiculescu exact sequence: proof of Theorem 3.11 that H =S∞ 0 −→ K1(C ∗(H0)) −→ K0(C(∂Λ)) id−b∗−−−−→ K0(C(∂Λ)) −→ K0(C ∗(H0)) −→ 0 Since K0(C(∂Λ)) ∼= C(∂Λ, Z), it follows that K1(C ∗(H0)) ∼= Z-span{χE : E ⊆ ∂Λ is compact-open and b-invariant}, and K0(C ∗(H0)) ∼= C(∂Λ, Z)(cid:14)Z-span{χE − χbE : E ⊆ ∂Λ compact-open}. We recall the action of b on ∂Λ = [0, d)N from Lemmas 2.9, 2.13 and 2.14: add 1 in the 0th coordinate; d in the jth coordinate carries as ±c to the (j + 1)st coordinate, using +c in cases (BS1) and (BS2), and −c in case (BS3) (bd acts as φ−1). Lemma 4.2. Let e = (c, d) be the greatest common divisor of c and d and let c = c′e, d = d′e. For µ1, . . ., µk ∈ [0, e), let U (µ1, . . . , µk) = {i ∈ [0, d)N : iℓ ≡ µℓ (mod e), 1 ≤ ℓ ≤ k}. Then U (µ1, . . . , µk) is b-invariant, and every open b-invariant set is a union of such sets. Proof. Since the complement of U (µ1, . . . , µk) is a union of such sets, it is enough to show that b·U (µ1, . . . , µk) ⊆ U (µ1, . . . , µk); moreover, since the 0th coordinate is unrestricted, it is enough to check that bd·U (µ1, . . . , µk) ⊆ U (µ1, . . . , µk). Let i ∈ U (µ1, . . . , µk), and let i′ = bd · i. Then for each ℓ there is kℓ such that i′ ℓ ≡ iℓ + kℓc ℓ ≡ iℓ ≡ µℓ (mod e) for 1 ≤ ℓ ≤ k. Thus (mod d). Since e divides both c and d, i′ ∈ U (µ1, . . . , µk). Therefore U (µ1, . . . , µk) is b-invariant. it follows that i′ Let us write Z(j0, . . . , jk) for the cylinder set {i ∈ ∂Λ : iℓ = jℓ for 0 ≤ ℓ ≤ k} in ∂Λ. We claim that for µ1, . . ., µk ∈ [0, e), and for any j0, U (µ1, . . . , µk) =Sn∈Z bn · Z(j0, µ1, . . . , µk). The containment ⊇ follows from the fact that U (µ1, . . . , µk) is invariant and contains Z(j0, µ1, . . . , µk). To see the containment ⊆, notice first C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 17 that we may adjust the 0th coordinate arbitrarily. Viewing the remaining coordinates as copies of Z/dZ, note that addition by c has d′ orbits. Thus we may adjust the first coordinate to any element congruent to µ1 modulo e. Then adding a multiple of cd′ will not further change the first coordinate. Since cd′ = c′d, this amounts to adding cc′ in the second coordinate. Again, since (cc′, d) = e, we may adjust the second coordinate to any element congruent to µ2 modulo e without changing the first coordinate. Repeating this argument, we see that we may fill up U (µ1, . . . , µk) by applying b repeatedly to Z(j0, µ1, . . . , µk), proving the claim. Now let V be an open b-invariant set. We write V =Sp Z(p) as a union of cylinder sets, where the p are tuples from [0, d). Then V = [n∈Z bn · V = [n∈Z[p bn · Z(p) =[p U (p′), where if p = (j0, j1, . . . , jk) then p′ = (µ1, . . . , µk) for µℓ ∈ [0, e), µℓ ≡ jℓ (mod ℓ), 1 ≤ ℓ ≤ k. (cid:3) Definition 4.3. We will use the following notation. For j ∈ [0, d)k+1 let QZ(j0,...,jk) = Sbj0 a···bjk aS∗ bj0 a···bjk a in C ∗(H0). If Z ⊆ ∂Λ is a compact-open subset, then Z is a finite disjoint union of cylinder sets. Since Z(i0, . . . , ik) =Fd−1 j=0 Z(i0, . . . , ik, j), and QZ(i0,...,ik) =Pd−1 Z =Fj Zj for any finite disjoint collection of cylinder sets {Zj}. j=0 QZ(i0,...,ik,j), we may define QZ =Pj QZj if Note that if Z ⊆ ∂Λ is an invariant compact-open set, then QZ and Sb commute. In the next few items , though it is only the coordinate-wise group structure that will be we write Ze for (Z/eZ)Z+ ∼= [0, e)Z+ convenient (and we omit the 0th coordinate). Corollary 4.4. K1(C ∗(H0)) ∼= C(Ze, Z). We note that under this isomorphism, we have that [SbQU(µ1,...,µk)] corresponds to χZ(µ1+eZ,...,µk+eZ). Lemma 4.5. K0(C ∗(H0)) ∼= C(Ze, Z[ 1 d′ ]). Proof. We define a homomorphism C(∂Λ, Z) → C(Ze, Z[ 1 see that it is well-defined, we note that we have to check that the relation Z(j0, . . . , jk) =Sd−1 is respected. For this we compute d′ ]) by χZ(j0,...,jk) 7→ (d′)−kχZ(j1+eZ,...,jk+eZ). To ℓ=0 Z(j0, . . . , jk, ℓ) (d′)−k−1χZ(j1+eZ,...,jk+eZ,ℓ+eZ) (d′)−k−1d′χZ(j1+eZ,...,jk+eZ,ℓ′+eZ) = (d′)−kχZ(j1+eZ,...,jk+eZ). χZ(j0,...,jk,ℓ) 7→ d−1Xℓ=0 = d−1Xℓ=0 e−1Xℓ′=0 The map is clearly surjective, and its kernel contains χZ − χb·Z for every cylinder set Z. We claim that its kernel is generated by the functions of the form χZ − χb·Z; this will conclude the proof. Let f be an element of the kernel. We may choose k such that f is a linear combination of characteristic functions of cylinder sets of length k: f =Pj njχZ(j), where j ranges over [0, d)k+1. Let [0, d)k+1 =Fp Ep be the equivalence classes nj(cid:1)χZ(p). defined by congruence modulo e in coordinates 1 through k. Then f maps to (d′)−kPp(cid:0)Pj∈Ep njχZ(j)(cid:1), and it Since f is in the kernel, we have thatPj∈Ep nj = 0 for each p. But then f =Pp(cid:0)Pj∈Ep is easy to see that each inner sum is in the span of the functions of the form χZ − χb·Z. (cid:3) The computation of the K-theory of Aγ uses the following elementary lemma. Lemma 4.6. Let M and N be abelian groups, and η : M → M , ξ : N → N , and I : M → N homo- that morphisms, such that I ◦ η = ξ ◦ I. Let fM = lim (1) ker(I) =Sn ker(ηn). (2) N =Sn ξ−n(I(M )). Then eI is an isomorphism. −→ η M , eN = lim −→ ξ N , and eI = lim −→ I : fM → eN . Suppose 18 JACK SPIELBERG with xn 7→ x. There is k such that ξk(I(xn)) = 0. Let xn+k = ηk(xn). Then xn+k ∈ ker(I), so by (1) there is ℓ such that xn+k ∈ ker ηℓ. Then 0 = ηℓ(xn+k) 7→ x, so x = 0. Proof. We let M (n) denote the nth copy of M in the inductive limit, etc. Let x ∈ ker(eI). Choose xn ∈ M (n) Let y ∈ eN . Choose yn ∈ N (n) with yn 7→ y. By (2) there is k such that yn ∈ ξ−k(I(M )). Then there is zn+k ∈ M (n+k) such that ξk(yn) = I(zn+k). Let zn+k 7→ z ∈ fM . Then yn 7→ eI(z), so that y = eI(z). Theorem 4.7. The K-theory of Aγ is given by (cid:3) K0(Aγ) ∼= Z[ 1 d ] and K1(Aγ) ∼= Z[ 1 c ]. Moreover, the generator d−k is represented in K0 by [Sbj0 a···bjk aS∗ sented in K1 by [Sbi1 a···bik aSbS∗ Proof. Recall from the proof of Theorem 3.11 that C ∗(Hn) ∼= Mdn ⊗ C ∗(H0). Explicitly, we have that C ∗(H0) ∼= C(∂Λ) ×b Z via bj0 a···bjk a], and the generator c−k is repre- bi1 a···bik a]. χ[e,e,α∂Λ] ←→ SαS∗ α and χ[b,e,∂Λ] ←→ Sb (where we view C(∂Λ) ×b Z ⊆ C ∗(Λ) by means of the generators of Theorem 3.23. The inclusion C ∗(H0) ֒→ C ∗(H1) ∼= Md ⊗ C ∗(H0) is described on these generators as follows. Let α = bi0 a · · · bim a, and α′ = bi1a · · · bima. Then SαS∗ α ←→ χ[e,e,α∂Λ] = χ[bi0 a,bi0 a,α′∂Λ] ←→ ei0,i0 ⊗ Sα′ S∗ α′ , Sb ←→ χ[b,e,∂Λ] = = d−2Xℓ=0 χ[b,e,bℓa∂Λ] = d−1Xℓ=0 d−1Xℓ=0 χ[bℓ+1a,bℓa,∂Λ] χ[bℓ+1a,bℓa,∂Λ] + χ[ab±c,bd−1a,∂Λ] ←→ eℓ+1,ℓ ⊗ 1 + e0,d−1 ⊗ S±c b , d−2Xℓ=0 where the exponent +c is used in cases (BS1) and (BS2), while −c is used in case (BS3). This dichomoty In general, the inclusion Mdn ⊗ C ∗(H0) ∼= C ∗(Hn) ֒→ C ∗(Hn+1) ∼= will continue throughout the proof. Mdn+1 ⊗ C ∗(H0) is given by tensoring by Mdn on the left of the above inclusion. For the computation of K0(Aγ), we consider the map on C(Ze, Z[ 1 The generator χZ(µ1+eZ,...,µk+eZ) corresponds to (d′)k[Sbj0 abµ1 a···bµk aS∗ The above inclusion sends this to d′ ]) induced by the above inclusion. bj0 abµ1 a···bµk a], for any choice of j0. (d′)k[ej0,j0 ⊗ Sbµ1 a···bµk aS∗ bµ1 a···bµk a] = (d′)k[Sbµ1 a···bµk aS∗ bµ1 a···bµk a], which corresponds to (d′)k(d′)−k+1χZ(µ2+eZ,...,µk+eZ) = d′χZ(µ2+eZ,...,µk+eZ). Thus we have K0(Aγ) = lim −→ η0 C(Ze, Z[ 1 d′ ]), 0 ). Let Ze have the usual product measure, assigning measure e−k to the set Z(µ1 + eZ, . . . , µk + eZ). We claim that K = {f ∈ C(Ze, Z[ 1 via the map η0 : χZ(µ1+eZ,...,µk+eZ) 7→ d′χZ(µ2+eZ,...,µk+eZ). We will find K = Sn ker(ηn cylinder set of length k. Thus f =Pµ1,...,µk ∈[0,e) cµ1,...,µk χZ(µ1+eZ,...,µk+eZ). Then d′ ]) : R f = 0}. To see this, first let R f = 0. Choose k so that f is constant on each and hence cµ1,...,µk , cµ1,...,µk = 0. ηk 0 =Z f = e−k Xµ1,...,µk ∈[0,e) 0 (f ) = e−k(d′)k Xµ1,...,µk ∈[0,e) f = Xµ1,...,µk∈[0,e) Conversely, let f ∈ ker(ηn Thus 0 ) for some n. Choose k ≥ n such that f is constant on cylinder sets of length k. cµ1,...,µk χZ(µ1+eZ,...,µk+eZ), C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 19 and hence 0 = ηn 0 (f ) = (d′)n Xµ1,...,µk∈[0,e) = (d′)n Xµn+1,...,µk∈[0,e)(cid:0) Xµ1,...,µn∈[0,e) cµ1,...,µk χZ(µn+1+eZ,...,µk+eZ) cµ1,...,µk(cid:1)χZ(µn+1+eZ,...,µk+eZ). cµ1,...,µk(cid:1) = 0. Thus the inner sum vanishes for each choice of µn+1, . . ., µk. Therefore Z f = e−k Xµ1,...,µk ∈[0,e) cµ1,...,µk = e−k Xµn+1,...,µk ∈[0,e)(cid:0) Xµ1,...,µn∈[0,e) Thus Lemma 4.6(1) holds, where M = C(Ze, Z[ 1 ξ is surjective, Lemma 4.6(2) holds. Therefore Lemma 4.6 implies that K0(Aγ) ∼= Z[ 1 we have that [Sbj0 a···bjk aS∗ is (d′)−ke−k = d−k, thus identifying generators of K0(Aγ). d ], I is integration, η = η0, and ξ = d·. Since d ]. From Lemma 4.5 bj0 a···bjk a] = (d′)−kχZ(j1+eZ,...,jk+eZ) in K0(C ∗(H0)). The integral of this function d′ ]), N = Z[ 1 For the computation of K1(Aγ) we consider the map on C(Ze, Z) induced by the inclusion of C ∗(Hn) into C ∗(Hn+1). Recalling Definition 4.3, the inclusion gives 7→ ei0,i0 ⊗ QU(µ1,...,µk) = d′−1Xj1,...,jk=0 d−1Xi0=0 d−1Xi0=0 d′−1Xj1,...,jk=0 where eU (µ1, . . . , µk) =Sd′−1 SbχU(µ1,...,µk) 7−→ = 1 ⊗ 0 . . . 0 1 Hence Sbi0 abµ1 +cj1 a···bµk +cjk aS∗ bi0 abµ1 +cj1 a···bµk +cjk a d′−1Xj1,...,jk=0 Sbµ1 +cj1 a···bµk +cjk aS∗ bµ1 +cj1 a···bµk +cjk a Sbµ1 +cj1 a···bµk +cjk aS∗ bµ1 +cj1 a···bµk +cjk a = 1 ⊗ QeU(µ1,...,µk), S±c b . . . 1 0   QeU(µ1,...,µk) . . . . . . QeU(µ1,...,µk) .  j1,...,jk=0 Z(µ1 + cj1, . . . , µk + cjk). Thus the inclusion gives (in d × d matrices) (SbQU(µ1,...,µk))d 7−→ 1 ⊗ S±c b QeU(µ1,...,µk) = 1 ⊗ (Se b QeU(µ1,...,µk))±c′ , i=0 Ei = U (µ2, . . . , µk). We claim that [Se this, we first define a *-homomorphism τ : Me(C) → Aγ by τ (eij ) = Si b QeU(µ1,...,µk)]. Let Ei = bieU (µ1, . . . , µk) for 0 ≤ i < e. Then the Ei b QeU(µ1,...,µk)] = [SbQU(µ2,...,µk)]. To see bQE0S−j i=0 ei+1,i + e0,e−1 . Let vt, 0 ≤ t ≤ 1 be a continuous path b + QE0S−e+1 b QE0S−j b b i=0 Si+1 and therefore [SbQU(µ1,...,µk)] 7−→ ±c′[Se are pairwise disjoint, andSe−1 be the shift matrix. Then τ (v) =Pe−2 e−1Xi=0 of unitary matrices from 1 to v. We have SbQU(µ2,...,µk) = e−1Xi=0 SbQEi = b QE0S−i Si+1 b = τ (ei+1,i) + Se b τ (e0,e−1) = (Se QEi)τ (v). e−2Xi=0 Thus SbQU(µ2,...,µk)τ (vt)∗ is a continuous path from SbQU(µ2,...,µk) to Se Now we find that i=1 QEi, proving the claim. χZ(µ1+eZ,...,µk+eZ) ←→ [SbQU(µ1,...,µk)] 7−→ ±c′[Se b QeU(µ1,...,µk)] = ±c′[SbQU(µ2,...,µk)] ←→ ±c′χZ(µ2+eZ,...,µk+eZ). Thus we have K1(Aγ) = lim −→ η1 C(Ze, Z), . Let v =Pe−2 e−1Xi=1 b QE0 + b QE0 +Pe−1 20 JACK SPIELBERG via the map η1 : χZ(µ1+eZ,...,µk+eZ) 7→ c′χZ(µ2+eZ,...,µk+eZ). Note that Z η1(χZ(µ1+eZ,...,µk+eZ)) =Z ±c′χZ(µ2+eZ,...,µk+eZ) = ±c′e−k+1 = ±ce−k = ±cZ χZ(µ1+eZ,...,µk+eZ). Thus I ◦ η = ξ ◦ I, where M = C(Ze, Z), N = Z[ 1 c ], η = η1, ξ = ±c·, and I is given by integration. Essentially the same computation as for K0 shows that Lemma 4.6(1) holds. Since Lemma 4.6(2) clearly holds, we have that K1(Aγ) ∼= Z[ 1 c ]. Finally, we have M (n) = K1(C ∗(Hn)) ∼= C(Ze, Z), where C(Ze, Z) ∋ 1 = [ei1,i1 ⊗ · · · ⊗ ein,in ⊗ Sb] = [Sbi1 a···bin aSbS∗ bi1 a···bin a] 7−→ ±c−n, thus identifying the generators of K1(Aγ). Theorem 4.8. The K-theory of C ∗(Λ) is given as follows. (cid:3) (1) If either c > 1 or we are in case (BS3), and if d > 1, then K0(C ∗(Λ)) ∼= Z/(d−1)Z and K1(C ∗(Λ)) ∼= Z/(±c − 1)Z, where the minus sign is used in case (BS3). (2) If either c > 1 or we are in case (BS3), and if d = 1, then K0(C ∗(Λ)) ∼= Z and K1(C ∗(Λ)) ∼= Z/(±c − 1)Z ⊕ Z, where the minus sign is used in case (BS3). (3) If c = 1 in case (BS2) (so d > 1), then K0(C ∗(Λ)) ∼= Z/(d − 1)Z ⊕ Z and K1(C ∗(Λ)) ∼= Z. (4) If c = 1 in case (BS1) (so d = 1), then K0(C ∗(Λ)) ∼= Z ⊕ Z and K1(C ∗(Λ)) ∼= Z ⊕ Z. In the first two cases, the class [1] of the identity is given by 1, while in the last two cases it is given by (1, 0). Proof. Recall from Lemma 4.1 that Aγ is Morita equivalent to A ×γ T. In A ×γ T we have partial isometries β) = ζn+1SαS∗ ζSbia, 0 ≤ i < d, with (ζSbia)∗(ζSbia) = ζ1 and (ζSbia)(ζSbia)∗ = SbiaS∗ β, the partial isometries ζSbia, 0 ≤ i < d. We have it follows thatbγ∗ is given by multiplication by d−1 in K0. To calculate the effect ofbγ∗ on K1, we consider bia]; i.e. bγ∗(1) = (±c)−1. Thus bγ∗ is given on K1 by multiplication by Thus bγ∗([Sb]) = [ζSb] = [SbiaSbS∗ (±c)−1. The Pimsner-Voiculescu exact sequence for A ∼ (A ×γ T) ×bγ Z gives (ζSbi a)(ζSb)(ζSbia)∗ = (ζSbia)(ζSb)(S∗ bia) = SbiaSbS∗ bia. bia. Sincebγ(ζnSαS∗ Z[ 1 c ] ↑ ±c−1−−−→ Z[ 1 c ] −→ K1(A) K0(A) ←− Z[ 1 d ] ↓ d−1←−− Z[ 1 d ] The various cases of the theorem follow from this diagram. The identification of the class of the identity (cid:3) in K0 follows from the form of the generators given in Theorem 4.7. We end by deriving the essential properties of C ∗(Λ) from properties of the groupoid G∂Λ. Theorem 4.9. (1) G∂Λ is minimal. (2) G∂Λ is contractive if and only if d > 1. (3) G∂Λ is topologically free if and only if d ∤ c. Proof. (1): This follows from Theorem 10.14 of [13], since Λ has only one vertex. (2): We use Theorem 10.16 of [13]. Since Λ has only one vertex, every nontrivial path is a cycle. Any element which does not by itself form an exhaustive set will be a non-exhaustive cycle in Λ. If d > 1, then a is such an element. Conversely, if d = 1, then the boundary of Λ reduces to a point, and then it is clear that G∂Λ is not contractive (or even locally contractive). (3): For the only if direction, note that if dc, then bda = abc = a(bd)( c d ). Hence for any γ ∈ Λ we have c bdγ = γbθ(γ) d . Thus bdγ ⋓ γ for all γ, so that Λ has {bd, e}-periodicity (as in [13], Definition 10.8). By [13], Theorem 10.10, G∂Λ is not topologically free. For the converse, suppose that d ∤ c. Let α 6= β. We must find γ such that αγ ⊥ βγ (as in [13], Remark 10.11). If α ⊥ β, we may take γ = e. So suppose that α ⋓ β. By Proposition 2.10 and left-cancellation we may assume that, say, θ(α) = 0. We treat three cases. (In the following, when we write ±c we mean +c in cases (BS1) and (BS2), and −c in case (BS3).) First, suppose that α = e and θ(β) > 0. Then β = biaβ′, where i ∈ [0, d). Let j ∈ [0, d) with j 6= i (since d 6= 1). Then αbja = bja ⊥ biaβ′bja = βbja. Second, suppose that α = e and θ(β) = 0. Then β = bq with q > 0. Since C ∗-ALGEBRAS FOR CATEGORIES OF PATHS ASSOCIATED TO THE BAUMSLAG-SOLITAR GROUPS 21 = j + rd, where 0 < j < d. We have d ∤ c, there is a least positive integer k such that q(cid:0) c d(cid:1)k d (cid:1)k−1 bqak = ak−1bq(cid:0) ±c 6∈ Z. Then q(cid:0) ±c d (cid:1)k−1 a = ak−1bjabrc ⊥ ak. Finally we suppose that α = bp and θ(β) > 0. Then β = aβ′ (otherwise we could cancel some element of B on the left). Let p = i + rd with i ∈ [0, d). If i 6= 0, then αa = biab±rc ⊥ aβ′a = βa. If i = 0, then αba = bab±rc ⊥ aβ′ba = βba. (cid:3) Corollary 4.10. C ∗(Λ) is a Kirchberg algebra if and only if d ∤ c. References [1] H. Al-Sulami, C ∗-algebras for boundary actions of solvable Baumslag-Solitar groups, Int. J. Math. Anal. (Ruse) 1 (2007), no. 21-24, 10671080. [2] C. Anantharaman-Delaroche, Purely infinite C ∗-algebras arising from dynamical systems, Bull. Soc. Math. France 125 (1997), no. 2, 199 -- 225. [3] G. Baumslag and D. Solitar, Some two-generator one-relator non-Hopfian groups, Bull. Amer. Math. Soc. 68 1962 199 -- 201. [4] J. Cuntz, Simple C ∗-algebras generated by isometries, Comm. Math. Phys. 57 (1977), no. 2, 173 -- 185. [5] B. Farb, L. Mosher, A rigidity theorem for the solvable Baumslag-Solitar groups (with an appendix by Daryl Cooper), Invent. Math. 131 (1998), no. 2, 419451. [6] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗-algebras IV, Pure infiniteness. J. Funct. Anal. 254 (2008), no. 5, 1161 -- 1187. [7] M. Laca and I. Raeburn, Semigroup crossed products and the Toeplitz algebras of nonabelian groups, J. Funct. Anal. 139 (1996), no. 2, 415 -- 440. [8] P. Muhly, J. Renault and D. Williams, Equivalence and isomorphism for groupoid C ∗-algebras, J. Operator Theory 17 (1987), no. 1, 3 -- 22. [9] A. Nica, C ∗-algebras generated by isometries and Wiener-Hopf operators, J. Operator Theory 27 (1992), no. 1, 17 -- 52. [10] J. Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, 793, Springer, Berlin, 1980. [11] P. Scott and T. Wall, Topological methods in group theory, Homological group theory (Proc. Sympos., Durham, 1977), pp. 137 -- 203, London Math. Soc. Lecture Note Ser., 36, Cambridge Univ. Press, Cambridge-New York, 1979. [12] J-P. Serre, Trees, Springer-Verlag, Berlin Heidelberg, 1980. [13] J. Spielberg, Groupoids and C ∗-algebras for categories of paths, preprint 2011, arXiv:1111.6924v2. School of Mathematical and Statistical Sciences, Arizona State University, P.O. Box 871804, Tempe, AZ 85287- 1804 E-mail address: [email protected]
1001.3182	2	1001	2011-07-12T05:05:34	The structure of an isometric tuple	[ "math.OA", "math.FA" ]	An $n$-tuple of operators $(V_1,...,V_n)$ acting on a Hilbert space $H$ is said to be isometric if the operator $[V_1\...\ V_n]:H^n\to H$ is an isometry. We prove a decomposition for an isometric tuple of operators that generalizes the classical Lebesgue-von Neumann-Wold decomposition of an isometry into the direct sum of a unilateral shift, an absolutely continuous unitary and a singular unitary. We show that, as in the classical case, this decomposition determines the weakly closed algebra and the von Neumann algebra generated by the tuple.	math.OA	math	THE STRUCTURE OF AN ISOMETRIC TUPLE MATTHEW KENNEDY Abstract. An n-tuple of operators (V1, . . . , Vn) acting on a Hilbert space H is said to be isometric if the operator [V1 · · · Vn] : H n → H is an isometry. We prove a decomposition for an isometric tuple of operators that generalizes the classical Lebesgue-von Neumann-Wold decomposition of an isometry into the direct sum of a unilateral shift, an absolutely continuous unitary and a singular unitary. We show that, as in the classical case, this decomposition determines the weakly closed algebra and the von Neumann algebra generated by the tuple. 1. Introduction This paper concerns the structure of an isometric tuple of operators, an object that appears frequently in mathematics and mathematical physics. From the perspective of an operator theorist, the notion of an isometric tuple is a natural higher-dimensional generalization of the notion of an isometry. An n-tuple of operators (V1, . . . , Vn) acting on a Hilbert space H is said to be isometric if the row operator [V1 ··· Vn] : H n → H is an isometry. This is equivalent to requiring that the operators V1, . . . , Vn satisfy the algebraic relations These relations are often referred to as the Cuntz relations. The main result in this paper is a decomposition of an isometric tuple that generalizes the classical Lebesgue-von Neumann-Wold decomposition of an isometry into the direct sum of a unilateral shift, an absolutely continuous unitary and a singular unitary. We show that, as in the classical case, this decomposition determines the structure of the weakly closed algebra and the von Neumann algebra generated by the tuple. The existence of a higher-dimensional Lebesgue-von Neumann-Wold de- composition was conjectured by Davidson, Li and Pitts in [DLP05]. They 2000 Mathematics Subject Classification. Primary 47A13; Secondary 47L55, 46L10. Research partially supported by an NSERC Canada Graduate Scholarship. 1 V ∗ I 0 i Vj =  if i = j, if i 6= j. THE STRUCTURE OF AN ISOMETRIC TUPLE 2 observed that the measure-theoretic definition of an absolutely continuous operator was equivalent to an operator-theoretic property of the functional calculus for that operator. Since this property naturally extends to the higher-dimensional setting, this allowed them to define the notion an abso- lutely continuous isometric tuple. The key technical result in this paper is a more effective operator-algebraic characterization of an absolutely continuous isometric tuple. The lack of such a characterization had been identified as the biggest obstruction to establishing the conjecture in [DLP05] (see also [DY08]). As we will see, the difficulty here can be attributed to the lack of a higher-dimensional analogue of the spectral theorem. In this paper, we overcome this difficulty by extending ideas from the commutative theory of dual algebras to the noncommutative setting. A similar approach was used in [Ken11] to prove that certain isometric tuples are hyperreflexive. In the present paper, the assumptions on the isometric tuples we consider are much weaker, and the problem is substantially more difficult. The idea to use this approach was inspired by results of Bercovici in [Ber98]. In Section 2, we review the Lebesgue-von Neumann-Wold decomposition In Section 3, of a single isometry that is the motivation for our results. we provide a brief review of the requisite background material on higher- dimensional operator theory, and we introduce the notions of absolute conti- nuity and singularity. In Section 4, we prove an operator-algebraic character- ization of an absolutely continuous isometric tuple. In Section 5, we prove an operator-algebraic characterization of a singular isometric tuple. In Section 6, we prove the Lebesgue-von Neumann-Wold decomposition of an isometric tuple, and we obtain some consequences of this result. The present exposition was inspired by the perspective of Muhly and Solel in [MS10], which appeared shortly after the first version of this paper. They consider the notion of absolute continuity in a more general setting. 2. Motivation The structure of a single isometry V is well understood. By the Wold decomposition of an isometry, V can be decomposed as V = Vu ⊕ U, THE STRUCTURE OF AN ISOMETRIC TUPLE 3 where Vu is a unilateral shift of some multiplicity, and U is a unitary. By the Lebesgue decomposition of a measure applied to the spectral measure of U , we can decompose U as U = Va ⊕ Vs, where Va is an absolutely continuous unitary and Vs is a singular unitary, in the sense that their spectral measures are absolutely continuous and singular respectively with respect to Lebesgue measure. This allows us to further decompose V as V = Vu ⊕ Va ⊕ Vs. We will refer to this as the Lebesgue-von Neumann-Wold decomposi- tion of an isometry. It will be convenient to consider the above notions of absolute continuity and singularity from a different perspective. Let A(D) denote the classical disk algebra of analytic functions on the complex unit disk D with continuous extension to the boundary. An isometry V induces a contractive representa- tion of A(D), namely the A(D) functional calculus for V , given by f → f (V ), f ∈ A(D). Recall that the algebra A(D) is a weak-* dense subalgebra of the algebra H ∞ of bounded analytic functions on the complex unit disk. In certain cases, the representation of A(D) induced by V is actually the restriction to A(D) of a weak-* continuous representation of H ∞, namely the H ∞ functional calculus for V , given by f → f (V ), f ∈ H ∞. It follows from Theorem III.2.1 and Theorem III.2.3 of [SF70] that this occurs if and only if Vs = 0 in the Lebesgue-von Neumann-Wold decomposition of V . This motivates the following definitions. Definition 2.1. Let V be an isometry. We will say that V is absolutely continuous if the representation of A(D) induced by V extends to a weak-* continuous representation of H ∞. If V has no absolutely continuous restric- tion to an invariant subspace, then we will say that V is singular. The importance of the Lebesgue-von Neumann-Wold decomposition of an isometry V is that it determines the structure of the weakly closed algebra W(V ) and the von Neumann algebra W∗(V ) generated by V . Recall that THE STRUCTURE OF AN ISOMETRIC TUPLE 4 W(V ) is the weak closure of the polynomials in V , and W∗(V ) is the weak closure of the polynomials in V and V ∗. Let α denote the multiplicity of Vu as a unilateral shift, and let µa and µs be scalar measures equivalent to the spectral measures of Va and Vs re- spectively. Since a unilateral shift of multiplicity one is irreducible, W∗(V ) is given by W∗(V ) ≃ B(ℓ2)α ⊕ L∞(Va) ⊕ L∞(µs)(Vs). It was established by Wermer in [Wer52] that W(V ) can be self-adjoint, depending on α and µa. If α 6= 0 or if Lebesgue measure is absolutely continuous with respect to µa, then W(V ) is given by W(V ) ≃ H ∞(Vu ⊕ Va) ⊕ L∞(µs)(Vs). Otherwise, if neither of these conditions holds, then W(V ) = W∗(V ). The following example shows that it is possible for the weakly closed alge- bra generated by an absolutely continuous isometry to be self-adjoint. We will see later that there is no higher-dimensional analogue of this phenome- non. Example 2.2. Let U denote the operator of multiplication by the coordinate function on L2(T, m), where m denotes Lebesgue measure. Let m1 and m2 denote Lebesgue measure on the upper and lower half of the unit circle respectively, and let U1and U2 denote the operator of multiplication by the coordinate function on L2(T, m1) and L2(T, m2) respectively. Since the spectral measure of U ≃ U1 ⊕ U2 is equivalent to Lebesgue measure, U is absolutely continuous. Thus U1 and U2 are also absolutely continuous. From above, W∗(U ) ≃ L∞(U ), W(U ) ≃ H ∞(U ). However, since Lebesgue measure is not absolutely continuous with respect to m1 or m2, W(Ui) = W∗(Ui) = L∞(Ui), i = 1, 2. In particular, the weakly closed algebras W(U1) and W(U2) generated by U1 and U2 respectively are self-adjoint. THE STRUCTURE OF AN ISOMETRIC TUPLE 5 3. Background and preliminaries 3.1. The noncommutative function algebras. The noncommutative Hardy space F 2 n is defined to be the full Fock-Hilbert space over Cn, i.e. n = ⊕∞ F 2 k=0(Cn)⊗k, where we will write ξ∅ to denote the vacuum vector, so that (Cn)⊗0 = Cξ∅. Let ξ1, . . . , ξn be an orthonormal basis of Cn and let F∗ n denote the unital free semigroup on n generators {1, . . . , n} with unit ∅. For a word w = w1 ··· wk in F∗ n, it will be convenient to write ξw = ξw1 ⊗ ··· ⊗ ξwk . We can identify F 2 n with the set of power series in n noncommuting variables ξ1, . . . , ξn with square-summable coefficients, i.e. F 2 n =  Xw∈F∗ n awξw : Xw∈F∗ n . aw2 < ∞  In particular, we can identify the noncommutative Hardy space F 2 1 with the classical Hardy space H 2 of analytic functions having power series expansions with square-summable coefficients. The left multiplication operators L1, . . . , Ln are defined on F 2 n by Liξw = ξi ⊗ ξw = ξiw, w ∈ F∗ n. It is clear that the n-tuple L = (L1, . . . , Ln) is isometric. We will call it the unilateral n-shift since, for n = 1, L1 can be identified with the unilateral shift on H 2. For a word w = w1 ··· wk in F∗ n, it will be convenient to write Lw = Lw1 ··· Lwk . The noncommutative disk algebra An is the norm closed unital alge- bra generated by L1, . . . , Ln and the noncommutative analytic Toeplitz algebra Ln is the weakly closed unital algebra generated by L1, . . . , Ln. These algebras were introduced by Popescu in [Pop96], and have subse- quently been studied by a number of authors (see for example [DP98] and [DP99]). The noncommutative disk algebra An and the noncommutative analytic Toeplitz algebra Ln are higher-dimensional analogues of the classical disk algebra A(D) and the classical algebra H ∞ of bounded analytic functions. In particular, the algebra An is a proper weak-* dense subalgebra of the algebra Ln. If we agree to identify functions in H ∞ with the corresponding multiplication operators on H 2, then we can identify A(D) with A1 and H ∞ with L1. THE STRUCTURE OF AN ISOMETRIC TUPLE 6 As in the classical case, an element A in Ln is uniquely determined by its Fourier series awLw, A ∼ Xw∈F∗ n where aw = (Aξ∅, ξw) for w in F∗ n. The Cesaro sums of this series converge strongly to A, and it is often useful heuristically to work directly with this representation. We will also need to work with the right multiplication operators R1, . . . , Rn defined on F 2 n by Riξw = ξw ⊗ ξi = ξwi, w ∈ F∗ n. The n-tuple R = (R1, . . . , Rn) is unitarily equivalent to L = (L1, . . . , Ln). The unitary equivalence is implemented by the "unitary flip" on F 2 n that, for a word w1 ··· wk in F∗ n, takes ξw1···wk to ξwk···w1. We will let Rn denote the weakly closed algebra generated by R1, . . . , Rn. 3.2. Free semigroup algebras. Let V = (V1, . . . , Vn) be an isometric n- tuple. The weakly closed unital algebra W(V ) generated by V1, . . . , Vn is called the free semigroup algebra generated by V . As in Section 3.1, for a word w = w1 ··· wk in the unital free semigroup F∗ n, it will be convenient to write Vw = Vw1 ··· Vwk . Example 3.1. The noncommutative analytic Toeplitz algebra Ln intro- duced in Section 3.1 is a fundamental example of a free semigroup algebra. We will see that it plays an important role in the general theory of free semigroup algebras. The study of free semigroup algebras was initiated by Davidson and Pitts in [DP99]. They observed that information about the unitary invariants of an isometric tuple can be detected in the algebraic structure of the free semigroup algebra it generates, and used this fact to classify a large fam- ily of representations of the Cuntz algebra. Free semigroup algebras have subsequently received a great deal of interest (see for example [Dav01]). It was shown in [DP98] that Ln has a great deal of structure that is analogous to the analytic structure of H ∞. This motivates the following definition. Definition 3.2. An isometric n-tuple V = (V1, . . . , Vn) is said to be an- alytic if the free semigroup algebra generated by V is isomorphic to the noncommutative analytic Toeplitz algebra Ln. THE STRUCTURE OF AN ISOMETRIC TUPLE 7 The notion of analyticity is of central importance in the theory of free semigroup algebras. This is apparent from the work of Davidson, Katsoulis and Pitts in [DKP01]. They proved the following general structure theorem. Theorem 3.3 (Structure theorem for free semigroup algebras). Let V = W(V ) be a free semigroup algebra. Then there is a projection P in V with range invariant under V such that (1) if P 6= 0, then the restriction of V to the range of P is an analytic free semigroup algebra, (2) the compression of V to the range of P ⊥ is a von Neumann algebra, (3) V = PVP + (W∗(V ))P ⊥. The analytic structure of a free semigroup algebra reveals itself in the form of wandering vectors. Let V = (V1, . . . , Vn) be an isometric n-tuple acting on a Hilbert space H. A vector x in H is said to be wandering for V if the set of vectors {Vwx : w ∈ F∗ n} is orthonormal. In this case we will also say that x is wandering for the free semigroup algebra generated by V . The existence of wandering vectors for an analytic free semigroup algebra was established in [Ken11], settling a conjecture first made in [DKP01] (see also [DLP05] and [DY08]). Examples show that the structure of an analytic free semigroup algebra can be quite complicated, making this result far from obvious. 3.3. Dilation theory. Recall that an operator T is said to be contractive if kTk ≤ 1. An n-tuple of operators T = (T1, . . . , Tn) acting on a Hilbert space H is said to be contractive if the row operator [T1 ··· Tn] : H n → H is contractive. Sz.-Nagy showed that every contractive operator T acting on a Hilbert space H has a unique minimal dilation to an isometry V , acting on a bigger Hilbert space K (see for example [SF70]). This means that H ⊆ K, H is cyclic for V and T k = PHV k H , k ≥ 1. Sz.-Nagy's dilation theorem was generalized in the work of Bunce, Frazho and Popescu in [Bun84], [Fra82] and [Pop89a] respectively. They showed that every contractive n-tuple of operators T = (T1, . . . , Tn) acting on a Hilbert space H has a unique minimal dilation to an isometric n-tuple V = (V1, . . . , Vn), acting on a bigger Hilbert space K. This means that H ⊆ K, THE STRUCTURE OF AN ISOMETRIC TUPLE 8 H is cyclic for V1, . . . , Vn and PH Vi1 ··· Vik H = Ti1 ··· Tik , i1, . . . , ik ∈ {1, . . . , n} and k ≥ 1. 3.4. The Wold decomposition. The classical Wold decomposition decom- poses a single isometry into the direct sum of a unilateral shift of some multiplicity and a unitary. In order to state the Wold decomposition of an isometric tuple, we need to generalize these notions. In Section 3.1, we introduced the unilateral n-shift L = (L1, . . . , Ln), and we saw that it is the natural higher-dimensional generalization of the classical unilateral shift. An isometric n-tuple is said to be a unilateral shift of multiplicity α if it is unitarily equivalent to the ampliation L(α) = (L(α) 1 , . . . , L(α) The higher-dimensional generalization of a unitary is based on the fact that a unitary is the same thing as a surjective isometry. An n-tuple of operators U = (U1, . . . , Un) is said to be unitary if the operator [U1 ··· Un] : H n → H is a surjective isometry. This is equivalent to requiring that the operators U1, . . . , Un satisfy n ), for some positive integer α. n UiU ∗ i = I. Xi=1 Note that a unilateral shift is not unitary. This is because the "vacuum" vector ξ∅ in F 2 n is not contained in the range of the unilateral n-shift L = (L1, . . . , Ln). In [DP99], Davidson and Pitts studied a family of "atomic" isometric tuples that arise from certain infinite directed trees. As the following example shows, this family contains a large number of unitary tuples. Example 3.4. Fix an infinite directed n-ary tree B with vertex set V such that every vertex has a parent. For a vertex v in V , let ci(v) denote the i-th child of v. Let H = ℓ2(V ), so that the set {ev : v ∈ V } is an orthonormal basis for H. Define operators S1, . . . , Sn on H by Siev = eci(v), 1 ≤ i ≤ n. It's clear that S1, . . . , Sn are isometries, and the fact that B is an infinite directed n-ary tree implies that the range of Si and the range of Sj are orthogonal for i 6= j. Thus S = (S1, . . . , Sn) is an isometric n-tuple. The fact that every vertex has a parent implies that every basis vector is in the range of some Si. Thus S is a unitary n-tuple. THE STRUCTURE OF AN ISOMETRIC TUPLE 9 Let V = (V1, . . . , Vn) be an arbitrary isometric n-tuple. If V is unitary, then the C∗-algebra C∗(V1, . . . , Vn) generated by V is isomorphic to the Cuntz algebra On. Otherwise, it is isomorphic to the extended Cuntz al- gebra En, the extension of the compacts by On. Since the only irreducible -representation of the compacts is the identity representation, and since On is simple, a -representation of En can be decomposed into a multiple of the identity representation and a representation of On. The Wold decomposition of an isometric n-tuple, which was proved by Popescu in [Pop89a], can be obtained as a consequence of these C∗-algebraic facts, based on the observa- tion that the C∗-algebra generated by a unilateral n-shift is isomorphic to En. Proposition 3.5 (The Wold decomposition). Let V = (V1, . . . , Vn) be an isometric n-tuple. Then we can decompose V as V = Vu ⊕ U, where Vu is a unilateral n-shift and U is a unitary n-tuple. 3.5. Absolutely continuous and singular isometric tuples. As in the classical case, an isometric n-tuple V = (V1, . . . , Vn) induces a contractive representation of the noncommutative disk algebra An, called the An func- tional calculus for V , determined by Li1 ··· Lik → Vi1 ··· Vik , i1, . . . , ik ∈ {1, . . . , n} and k ≥ 1. This is a consequence of Popescu's generalization of von Neumann's inequal- ity in [Pop91]. Recall from Section 3.1 that An is a proper weak-* dense subalgebra of the noncommutative analytic Toeplitz algebra Ln. The following definition is the natural generalization of Definition 2.1. Definition 3.6. Let V = (V1, . . . , Vn) be an isometric n-tuple. We will say that V is absolutely continuous if the representation of An induced by V is the restriction to An of a weak-* continuous representation of Ln. We will say that V is singular if V has no absolutely continuous restriction to an invariant subspace. It is clear from Definition 3.2 and Definition 3.6 that an analytic iso- metric tuple is absolutely continuous. In order to obtain the Lebesgue-von Neumann-Wold decomposition of an isometric tuple, we will prove the con- verse result that an absolutely continuous isometric tuple is analytic. THE STRUCTURE OF AN ISOMETRIC TUPLE 10 4. Absolutely continuous isometric tuples The main result in this section is an operator-algebraic characterization of an absolutely continuous isometric tuple. Specifically, we will show that for n ≥ 2, every absolutely continuous isometric n-tuple is analytic. For n ≥ 2, fix an absolutely continuous isometric n-tuple S = (S1, . . . , Sn) acting on a Hilbert space H. Let Φ denote the corresponding representation of the noncommutative disk algebra An, given by Φ(Lw) = Sw, w ∈ F∗ n. Since S is absolutely continuous, Φ extends to a representation of Ln that is weak-* continuous. It was shown in Corollary 1.2 of [DY08] that Φ is actually a completely isometric isomorphism and a weak-* homeomorphism from Ln to the weak-* closed algebra generated by S1, . . . , Sn. This is equivalent to the fact that an infinite ampliation of S is an analytic isometric tuple. Evidently, it is much more difficult to show that S is analytic. As an explanation, we offer the aphorism that things are generally much nicer in the presence of infinite multiplicity. Showing that S is analytic amounts to showing that the free semigroup algebra (i.e. the weakly closed algebra) W(S) generated by S1, . . . , Sn is isomorphic to the noncommutative analytic Toeplitz algebra Ln. Since we know from above that the weak-* closed algebra generated by S1, . . . , Sn is isomorphic to Ln, our strategy will be to show that this algebra is actually equal to W(S). 4.1. The noncommutative Toeplitz operators. Let S denote the weak- * closed algebra generated by S1, . . . , Sn. The map Φ introduced at the beginning of this section is a completely isometric isomorphism and a weak-* homeomorphism from Ln to S. It will be useful for what follows to extend Φ even further. Let Mn denote the weak-* closure of the operator system Ln + L∗ n. We will call the elements of Mn the noncommutative Toeplitz operators, because they are a natural higher-dimensional generalization of the classical Toeplitz operators. The noncommutative Toeplitz operators were introduced by Popescu in [Pop89b]. It was shown in Corollary 1.3 of [Pop09] that A belongs to Mn if THE STRUCTURE OF AN ISOMETRIC TUPLE 11 and only if A if i = j, 0 otherwise, R∗ i ARj =  where R1, . . . , Rn are the right multiplication operators introduced in Section 3.1. A short proof of this fact was also given in Lemma 3.2 of [Ken11]. It follows from this characterization that Mn is weakly closed. Let T denote the weak-* closure of the operator system S +S ∗. The proof of the following proposition is nearly identical to the proof of Theorem 3.6 of [Ken11]. Proposition 4.1. Let S = (S1, . . . , Sn) be an absolutely continuous isomet- ric n-tuple. The representation Φ of Ln induced by S extends to a completely isometric and weak-* homeomorphic -map from Mn to T . We will need to exploit the fact that Mn and T are dual spaces. Let T∗ denote the predual of T , i.e. the set of weak- continuous linear functionals on T . Similarly, let Mn∗ denote the predual of Mn. Basic functional analysis implies that the inverse map Φ−1 is the dual of an isometric isomorphism φ from Mn∗ to T∗. Moreover, since Φ−1 is isometric, so is φ. We can identify the predual of B(F 2 n ), with the set of trace class operators C 1(F 2 n ), i.e. the set of weak-* continuous n ) on linear functionals on B(F 2 F 2 n , where K in C 1(F 2 n ) corresponds to the linear functional (T, K) = tr(T K), T ∈ B(F 2 n ). If we let (Mn)⊥ denote the preannihilator of Mn, i.e. (Mn)⊥ = {K ∈ C 1(F 2 n ) : tr(AK) = 0, ∀A ∈ Mn}, then we can identify the predual (Mn)∗ with the quotient space C 1(F 2 Similarly, we can identify the predual T∗ with the quotient space C 1(H)/T⊥. n , it will be convenient to let [ξ ⊗ η]Mn denote the weak-* For ξ and η in F 2 n )/(Mn)⊥. continuous linear functional on Mn given by (A, [ξ ⊗ η]Mn) = (Aξ, η), A ∈ Mn. In other words, [ξ⊗η]Mn denotes the equivalence class of the rank one tensor x ⊗ y in (Mn)∗. Similarly, for x and y in H, let [x ⊗ y]T denote the weak-* continuous linear functional on T given by (T, [x ⊗ y]T ) = (T x, y), T ∈ T . THE STRUCTURE OF AN ISOMETRIC TUPLE 12 4.2. Intertwining operators. An operator X : F 2 n → H is said to in- tertwine the isometric n-tuple S = (S1, . . . , Sn) and the unilateral n-shift L = (L1, . . . , Ln) if it satisfies XLi = SiX, 1 ≤ i ≤ n. Observe that if X intertwines S and L, then the operator JX∗XJ is a noncommutative Toeplitz operator, where J is the unitary flip introduced in Section 3.1. Indeed, using the fact that JRi = LiJ for 1 ≤ i ≤ n, we compute R∗ i JX∗XJRj = JL∗ i X∗XLjJ i SjXJ = JX∗S∗ JX∗XJ if i = j, 0 otherwise. =   Since S is absolutely continuous, it follows from Theorem 2.7 of [DLP05] that every vector x in H is in the range of an operator that intertwines S and L. 4.3. Dual algebra theory. Recall that to prove the isometric n-tuple S = (S1, . . . , Sn) is analytic, our strategy is to show that the weak-* closed al- gebra S = W∗(S1, . . . , Sn) is actually equal to the weakly closed algebra W(S1, . . . , Sn). This amounts to showing that S is already weakly closed. However, instead of working directly with S, it will be necessary to work with the operator system T . In fact, we will need to consider the general structure of the predual of T . In Section 4.1, we saw that an element in the predual T∗ of the operator system T can be identified with an equivalence class of trace class operators. We will show that T satisfies a very powerful predual "factorization" property, in the sense that the equivalence class of an element in the predual T∗ always contains "nice" representatives. We will see that S inherits this property from T , and that this will imply the desired result. The idea of studying factorization in the predual of an operator algebra is the central idea in dual algebra theory, which has been applied with great success to a number of problems in the commutative setting (see for ex- ample [BFP85]). As we will see, many of the factorization properties that were introduced in the commutative setting make sense even in the present noncommutative setting. THE STRUCTURE OF AN ISOMETRIC TUPLE 13 Definition 4.2. A weak-* closed subspace A of operators acting on a Hilbert space H is said to have property A1(1) if, given a weak-* continuous linear functional τ on A with kτk ≤ 1 and ǫ > 0, there are vectors x and y in H such that kxk ≤ (1 + ǫ)1/2, kyk ≤ (1 + ǫ)1/2 and τ = [x ⊗ y]A. If a weak-* closed subspace of B(H) has property A1(1), then the equiva- lence class of any weak-* continuous linear functional on the subspace con- tains an operator of rank one. Note that in this case, every weak-* continuous linear functional on the subspace is actually weakly continuous. It was shown in [DP99] that Ln has property A1(1), and the same proof also shows that Mn has property A1(1). Of course, the main difficulty with a predual factorization property like property A1(1) is that it is often extremely difficult to show that it holds. The next factorization property turns out to be much stronger than property A1(1), but it is sometimes easier to show that it holds due to its approximate nature. Definition 4.3. A weak-* closed subspace A of operators acting on a Hilbert space H is said to have property X0,1 if, given a weak-* continuous linear functional τ on A with kτk ≤ 1, z1, ..., zq in H and ǫ > 0, there are vectors x and y in H such that (1) kxk ≤ 1 and kyk ≤ 1, (2) k[x ⊗ zj]Ak < ǫ and k[zj ⊗ y]Ak < ǫ for 1 ≤ j ≤ q, (3) kτ − [x ⊗ y]Ak < ǫ. It's easy to see that the infinite ampliation of a weak-* closed subspace of B(H) has property X0,1. Thus, intuitively, a weak-* closed subspace of B(H) that has property X0,1 can be thought of as having "approximately infinite" multiplicity. It was shown in [BFP85] that property X0,1 implies property A1(1). We will show that T has property X0,1. Since this property is inherited by weak-* closed subspaces, it will follow that S has property X0,1, and hence that S has property A1(1). It is easy to show that any weak-* closed subspace of operators with property A1(1) is weakly closed (see for example Proposition 59.2 of [Con00]). Thus this will imply the desired result that S is weakly closed. 4.4. Approximate factorization. THE STRUCTURE OF AN ISOMETRIC TUPLE 14 Lemma 4.4. Given unit vectors x, z1, ..., zq in H and ǫ > 0, there are vectors ξ, ζ1, ..., ζq in F 2 n such that (1) kξk < √q(1 + ǫ)1/2, (2) kζik < (1 + ǫ)1/2 for 1 ≤ i ≤ q, (3) [x ⊗ zi]T = φ([ξ ⊗ ζi]Mn) for 1 ≤ i ≤ q. ik < (1+ǫ)1/2, kζ ′ q, ζ ′ ik < (1+ǫ)1/2 and [x⊗zi]T = φ([υ′ Proof. Since Mn has property A1(1), there are vectors υ′ n such that kυ′ F 2 for 1 ≤ i ≤ q. Let Vi = R12k for 1 ≤ i ≤ q, so that V1, ..., Vq are isometries in Rn with pairwise orthogonal ranges. Set ξ = Pq i and ζi = Viζ ′ i for 1 ≤ i ≤ q. Then kξk < √q(1 + ǫ)1/2, kζik < (1 + ǫ)1/2 and for T in T , 1, ..., ζ ′ i⊗ζ ′ q in i]Mn) i=1 Viυ′ 1, ..., υ′ (φ([ξ ⊗ ζi]Mn), T ) = (Φ−1(T )ξ, ζi) Xj=1 = (Φ−1(T ) q Vjυ′ j, Viζ ′ i) i, ζ ′ i) i]Mn), T ) = (Φ−1(T )υ′ = (φ([υ′ i ⊗ ζ ′ = ([x ⊗ zi]T , T ). Hence [x ⊗ zi]T = φ([ξ ⊗ ζi]Mn). Lemma 4.5. Let η be a unit vector contained in the algebraic span of {ξw : w ∈ F∗ n}. Then there are words u and v in F∗ n such that (cid:3) LuRvη = Lξ∅ = Rξ∅, where L is an isometry in Ln, and R is an isometry in Rn with range or- thogonal to the range of R1. Proof. Expand η as η = Pw≤m awξw for some m ≥ 0. Let u = 12m and let v = 1m2. Then LuRvη = Pw≤m awξuwv. Set L = Pw≤m awLuwv and R = Pw≤m awRuwv. Then LuRvη = Lξ∅ = Rξ∅, and it's clear that the w ≤ m and w′ ≤ m, It remains to show that L and R are isometries. For w and w′ in F+ range of R is orthogonal to the range of R1. n with L∗ vL∗ wLw′Lv =  I 0 if w = w′, otherwise. THE STRUCTURE OF AN ISOMETRIC TUPLE 15 This gives awaw′L∗ uwvLuw′v awaw′L∗ vL∗ wLw′Lv L∗L = Xw≤m Xw′≤m = Xw≤m Xw′≤m = Xw≤m aw2I = I, where the last equality follows from the fact that η is a unit vector. Thus L is an isometry, and it follows from a similar computation that R is an isometry. (cid:3) Lemma 4.6. Given unit vectors z1, ..., zq in H and ǫ > 0, there exists a unit vector x in H and vectors ξ, ζ1, ..., ζq in F 2 n such that (1) kξk < √q(1 + ǫ)1/2, (2) kζik < (1 + ǫ)1/2 for 1 ≤ i ≤ q, (3) ξ = kξkLξ∅ = kξkRξ∅, where L is an isometry in Ln, and R is an isometry in Rn with range orthogonal to the range of R1, (4) k[x ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k < ǫ for 1 ≤ i ≤ q. Proof. Let x′ be any unit vector in H. By Lemma 4.4, there are vectors ξ′, ζ ′ q in F 2 1, ..., ζ ′ n such that (1) kξ′k < √q(1 + ǫ)1/2, (2) kζ ′ (3) [x′ ⊗ zi]T = φ([ξ′ ⊗ ζ ′ ik < (1 + ǫ)1/2 for 1 ≤ i ≤ q, i]Mn) for 1 ≤ i ≤ q. Let η be a vector contained in the algebraic span of {ξw : w ∈ F∗ kηk < √q(1 + ǫ)1/2 and kξ′ − ηk < ǫ/(1 + ǫ)1/2. Then n} such that k[x′ ⊗ zi]T − φ([η ⊗ ζ ′ i]Mn)k ≤ k[x′ ⊗ zi]T − φ([ξ′ ⊗ ζ ′ i]Mnk +k[(ξ′ − η) ⊗ ζ ′ i]Mn)k ≤ kξ′ − ηkkζ ′ ik < ǫ for 1 ≤ i ≤ q. By Lemma 4.5, there are words u and v in F+ n such that LuRvη = kηkLξ∅ = kηkRξ∅, THE STRUCTURE OF AN ISOMETRIC TUPLE 16 where L is an isometry in Ln, and R is an isometry in Rn with range or- thogonal to the range of R1. Set x = Sux′, ξ = LuRvη and ζi = Rvζ ′ i for 1 ≤ i ≤ q. Then for T in T , ([x ⊗ zi]T − φ([ξ ⊗ ζi]Mn), T ) = ([Sux′ ⊗ zi]T − φ([η ⊗ ζ ′ = ([x′ ⊗ zi]T − φ([η ⊗ ζ ′ ≤ k[x′ ⊗ zi]T − φ([η ⊗ ζ ′ < ǫkTk. i]Mn), T ) i]Mn), T Su) i]Mn)kkT Suk Hence k[x ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k < ǫ. The following result is implied by Lemma 1.2 in [Kri01]. (cid:3) Lemma 4.7. Given a proper isometry R in Rn, vectors ζ1, ..., ζq in F 2 ǫ > 0, there exists k ≥ 1 such that k(R∗)kζik < ǫ for 1 ≤ i ≤ q. Lemma 4.8. Given a proper isometry S in S, vectors u and v in H and ǫ > 0, there exists k ≥ 1 such that k[u ⊗ (S∗)kv]Sk < ǫ. Proof. Since Ln has property A1, there are vectors µ and ν in F 2 [u ⊗ v]S = φ([µ ⊗ ν]Ln). Thus for A in S, n such that n and ([u ⊗ (S∗)kv]S , A) = ([µ ⊗ (Φ−1(S)∗)kν]Ln, Φ−1(A)) = (Φ−1(A)µ, (Φ−1(S)∗)kν) ≤ kAkkµkk(Φ−1(S)∗)kνk, which gives k[u⊗(S∗)kv]Sk ≤ kµkk(Φ−1(S)∗)kνk. Since Φ−1(S) is a a proper isometry in Ln, and since Ln and Rn are unitarily equivalent, the result now follows by Lemma 4.7. (cid:3) Lemma 4.9. Given unit vectors z1, ..., zq in H and ǫ > 0, there exists a unit vector x in H and vectors ξ, ζ1, ..., ζq in F 2 n such that (1) kξk < √q(1 + ǫ)1/2, (2) kζik < (1 + ǫ)1/2 for 1 ≤ i ≤ q, (3) ξ = kξkLξ∅ = kξkRξ∅, where L is an isometry in Ln, and R is an isometry in Rn with range orthogonal to the range of R1, (4) kR∗ζik < ǫ for 1 ≤ i ≤ q, (5) (Φ(L)kx, x) < ǫ for k ≥ 1, (6) k[x ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k < ǫ for 1 ≤ i ≤ q. Proof. By Lemma 4.6, there exists a unit vector x′ in H and vectors ξ′, ζ1, ..., ζq in F 2 n such that THE STRUCTURE OF AN ISOMETRIC TUPLE 17 (1) kξ′k < √q(1 + ǫ)1/2, (2) kζik < (1 + ǫ)1/2 for 1 ≤ i ≤ q, (3) ξ′ = kξ′kL′ξ∅ = kξ′kR′ξ∅, where L′ is an isometry in Ln and R′ is an isometry in Rnwith the range of R′ orthogonal to the range of R1, (4) k[x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]Mn)k < ǫ for 1 ≤ i ≤ q. 1)m(R′)∗ζik < By Lemma 4.7 and Lemma 4.8, there exists m ≥ 1 such that k(R∗ 1 L′ ǫ for 1 ≤ i ≤ q and k[x′ ⊗ (S∗ 1 )mΦ(L′)∗x′]Sk < ǫ. Set ξ = Lm and R = R′Rm 1 . Then ξ = kξkLξ∅ = kξkRξ∅, L is an isometry in Ln, and R is an isometry in Rn with range orthogonal to the range of R1. For 1 ≤ i ≤ q, this gives kR∗ζik = k(R∗ 1)m(R′)∗ζik < ǫ. 1 ξ′, L = Lm Let x = Sm 1 x′. Then for k ≥ 1, we compute (Φ(L)kx, x) = (Φ(Lm 1 L′)kSm 1 x′, Sm 1 )kx′, Sm 1 x′) 1 x′) 1 Φ(L′Lm = (Sm = (Φ(L′Lm = (Φ(L′Lm = ([x′ ⊗ (S∗ ≤ k[x′ ⊗ (S∗ < ǫ. 1 )kx′, x′) 1 )k−1x′, (S∗ 1 )mΦ(L′)∗x′) 1 )mΦ(L′)∗x′]S , Φ(L′Lm 1 )mΦ(L′)∗x′]Skk(L′Lm 1 )k−1) 1 )k−1k Finally, for T in T we have ([x ⊗ zi]T − φ([ξ ⊗ ζi]Mn), T ) = ([x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]Mn), T Sm 1 ) ≤ k[x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]MnkkT Sm 1 k < ǫkTk. Thus k[x ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k < ǫ. 4.5. Approximately orthogonal vectors. The following lemma is extracted from the proof of Theorem 4.3 in [Ber98]. (cid:3) Lemma 4.10. Given two isometries R and R′ in Rn with orthogonal ranges and vectors ξ and µ in F 2 n with µ in the kernel of R∗, define Then µk = 1 √k k[ξ ⊗ µk]Mnk ≤ k RjR′µ. Xj=1 1 √kkµkkDk − 1k1, THE STRUCTURE OF AN ISOMETRIC TUPLE 18 where Dk denotes the k-th Dirichlet kernel and k · k1 denotes the L1 norm. Lemma 4.11. Given unit vectors z1, ..., zq in H and ǫ > 0, there exists a unit vector x in H such that k[x ⊗ zi]T k < ǫ for 1 ≤ i ≤ q. Proof. We may suppose that ǫ < 1. Using the fact that lim k−1/2kDkk1 = 0, where Dk denotes the k-th Dirichlet kernel and k · k1 denotes the L1 norm, choose k ≥ 1 such that 2(q/k)−1/2kDk − 1k1 < ǫ/(3(1 + ǫ)). Next choose ǫ′ > 0 such that ǫ′ < min(cid:26)1, ǫ(1 − ǫ) 3√k , ǫ(1 − ǫ) 6√q , kǫ k2 − k(cid:27) . By Lemma 4.9, there exists a unit vector x′ in H and vectors ξ′, ζ1, ..., ζq in F 2 n such that (1) kξ′k < √q(1 + ǫ′)1/2, (2) kζik < (1 + ǫ′)1/2 for 1 ≤ i ≤ q, (3) ξ′ = kξ′kLξ∅ = kξ′kRξ∅, where L is an isometry in Ln, and R is an isometry in Rn with range orthogonal to the range of R1, (4) kR∗ζik < ǫ′ for 1 ≤ i ≤ q, (5) (Φ(L)kx′, x′) < ǫ′ for k ≥ 1, (6) k[x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]Mn)k < ǫ′ for 1 ≤ i ≤ q. By (4) we can write ζi = µi + νi, where µi is in the kernel of R∗ and kνik < ǫ′. Let ξ = k−1/2Pk−1 j=0 L1Ljξ′. Then by (3) we can write ξ as k−1 ξ = L1Ljξ′ k−1 1 Xj=0 √k = kξ′k√k Xj=0 = kξ′k√k Xj=1 k L1Lj+1ξ∅ L1Ljξ∅, THE STRUCTURE OF AN ISOMETRIC TUPLE 19 which implies kξk = kξ′k. Applying (3) again, we can also write ξ as ξ = k−1 L1Ljξ′ k−1 1 Xj=0 √k = kξ′k√k Xj=0 = kξ′k√k Xj=0 = kξ′k√k Xj=1 k−1 k L1LjRξ∅ Rj+1R1ξ∅ RjR1ξ∅. By Lemma 4.10 and the choice of k, this gives k[ξ ⊗ µi]Mnk ≤ 1 √kkξkkµikkDk − 1k1 (1 + ǫ′)1/2kµikkDk − 1k1 ≤ r q < ǫ(1 − ǫ)/3. k j=0 Φ(L1Lj). Then kSk ≤ √k, so for T Let y = Sx′, where S = k−1/2Pk−1 in T , ([y ⊗ zi]T − φ([ξ ⊗ ζi]Mn), T ) = ([x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]Mn), T S) ≤ k[x′ ⊗ zi]T − φ([ξ′ ⊗ ζi]Mn)kkT Sk < ǫ′√kkTk < (ǫ(1 − ǫ)/3)kTk, which gives k[y ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k < ǫ(1 − ǫ)/3. Since k[ξ ⊗ νi]Mnk ≤ kξkkνik < √q(1 + ǫ′)1/2ǫ′ < ǫ(1 − ǫ)/3, this gives k[y ⊗ zi]T k ≤ k[y ⊗ zi]T − φ([ξ ⊗ ζi]Mn)k + k[ξ ⊗ µi]Mnk + k[ξ ⊗ νi]Mnk < ǫ(1 − ǫ). THE STRUCTURE OF AN ISOMETRIC TUPLE 20 k−1 Finally, we compute kyk2 = kSx′k2 Xj=0 k X0≤i<j≤k−1 = kx′k2 + 1 Φ(L1Lj)x′(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 √k 1 2 k X0≤i<j≤k−1 k2 − k ǫ′. k ≥ 1 − ≥ 1 − > 1 − ǫ. (x′, Φ(L)j−ix′) + (Φ(L)i−jx′, x′) (x′, Φ(L)j−ix′) − (Φ(L)i−j x′, x′) 1 k X0≤j<i≤k−1 k X0≤j<i≤k−1 1 Hence taking x = (1 − ǫ)−1y, kxk ≥ 1 and k[x ⊗ zi]T k < ǫ for 1 ≤ i ≤ q. (cid:3) Lemma 4.12. Given unit vectors z1, ..., zq in H and ǫ > 0, there exists an intertwining operator X : F 2 n → H such that kXξ∅k = 1 and k[Xξ∅⊗zi]T k < ǫ for 1 ≤ i ≤ q. Proof. By Lemma 4.11, there exists a unit vector x in H such that k[x ⊗ zi]T k < ǫ for 1 ≤ i ≤ q. By Theorem 2.7 of [DLP05], x is in the range of an intertwining operator X′ : F 2 n such that X′ξ = x. The result now follows from the fact that the set of vectors {Rξ∅ : R ∈ Rn} is dense in F 2 n , and the fact that for R in Rn, the operator X′R is intertwining. n → H. Hence there is a vector ξ in F 2 (cid:3) Lemma 4.13. Let X : F 2 Then given ǫ > 0, there is a word v in F∗ n → H be an intertwining operator with kXξ∅k = 1. n such that k[XRvξ∅ ⊗ XRvξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)k < ǫ. Proof. Since X∗X is an L-Toeplitz operator, by Lemma 4.5 of [Ken11], there is a word v in F∗ vX∗XRv is also an L-Toeplitz operator. Let ξ = (R∗ vX∗XRv−I)ξ∅, so that kξk < ǫ/2. For w in F∗ vX∗XRvξ∅−ξ∅k < ǫ/2. Note that R∗ n such that kR∗ n, since (Lwξ, ξ∅) = 0 we can write (SwXRvξ∅, XRvξ∅) = (Lwξ∅, R∗ vX∗XRvξ∅) = (Lwξ∅, ξ∅) + (Lwξ∅, ξ) + (Lwξ, ξ∅). THE STRUCTURE OF AN ISOMETRIC TUPLE 21 Similarly, (S∗ wXRvξ∅, XRvξ∅) = (L∗ = (L∗ vX∗XRvξ∅, ξ∅) wR∗ wξ∅, ξ∅) + (L∗ wξ∅, ξ) + (L∗ wξ, ξ∅). This gives [XRvξ∅ ⊗ XRvξ∅]T = φ([ξ∅ ⊗ ξ∅]Mn + [ξ ⊗ ξ∅]Mn + [ξ∅ ⊗ ξ]Mn), so we conclude that k[XRvξ∅ ⊗ XRvξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)k ≤ k[ξ ⊗ ξ∅]Mn + [ξ∅ ⊗ ξ]Mnk ≤ 2kξkkξ∅k < ǫ, as required. (cid:3) Lemma 4.14. Given unit vectors z1, ..., zq in H and ǫ > 0, there exists an intertwining operator X : F 2 n → H such that kXξ∅k = 1, k[Xξ∅ ⊗ zi]T k < ǫ for 1 ≤ i ≤ q and k[Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)k < ǫ. Proof. By Lemma 4.12, there exists an intertwining operator X′ : F 2 n → H such that kX′ξ∅k = 1 and k[X′ξ∅ ⊗ zi]T k < ǫ for 1 ≤ i ≤ q. By Lemma 4.13, there is a word v in F∗ n such that k[X′Rvξ∅ ⊗ X′Rvξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)k < ǫ. Let X = X′Rv. Then kXξ∅k = kX′Rvξ∅k = kX′Lvξ∅k = kSvX′ξ∅k = kX′ξ∅k = 1. For T in T , ([Xξ∅ ⊗ zi]T , T ) = ([X′Rvξ∅ ⊗ zi]T , T ) = ([X′Lvξ∅ ⊗ zi]T , T ) = ([SvX′ξ∅ ⊗ zi]T , T ) = ([X′ξ∅ ⊗ zi]T , T Sv) ≤ k[X′ξ∅ ⊗ zi]T kkTk. Hence k[Xξ∅ ⊗ zi]T k < ǫ for 1 ≤ i ≤ q. 4.6. The strong factorization property. (cid:3) THE STRUCTURE OF AN ISOMETRIC TUPLE 22 Theorem 4.15. Given a weak-* continuous linear functional τ on T with kτk ≤ 1, unit vectors z1, ..., zq in H and ǫ > 0, there are vectors x and y in H such that (1) kxk ≤ 1 and kyk ≤ 1, (2) kτ − [x ⊗ y]T k < ǫ, (3) k[x ⊗ zi]T k < ǫ and k[zi ⊗ y]T k < ǫ for 1 ≤ i ≤ q. In other words, T has property X0,1. Proof. Choose ǫ′ > 0 such that ǫ′ < ǫ and 1 − (1 + 2ǫ′)−2(1 − ǫ′) < ǫ. Since n with kξk ≤ 1 + ǫ′/2 Mn has property A1(1), there are vectors ξ and υ in F 2 and kυk ≤ 1 + ǫ′/2 such that τ = φ([ξ ⊗ υ]Mn). Since ξ∅ is cyclic for Ln, there are A and B in Ln such that kAξ∅ − ξk < ǫ′/(4(1 + ǫ′)) and kBξ∅ − υk < ǫ′/(4(1 + ǫ′)). Then kAξ∅k ≤ kAξ∅ − ξk + kξk < 1 + ǫ′, and similarly kBξ∅k < 1 + ǫ′. This gives k[Aξ∅ ⊗ Bξ∅]Mn − [ξ ⊗ υ]Mnk ≤ k[(Aξ∅ − ξ) ⊗ Bξ∅]k +k[ξ ⊗ (Bξ∅ − υ)]Mnk ≤ kAξ∅ − ξkkBξ∅k + kξkkBξ∅ − υk < ǫ′/2. By Lemma 4.14, there is an intertwining operator X : F 2 n → H such that kXξ∅k = 1, k[Xξ∅ ⊗ zi]T k < ǫ′/(kAk + kBk) for 1 ≤ i ≤ q and k[Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)k < ǫ′/(2(kAk +kBk)2). Note that since T is self-adjoint, we also have k[zi ⊗ Xξ∅]T k < ǫ′/(kAk + kBk) for 1 ≤ i ≤ q. Define vectors x′ and y′ in H by x′ = Φ(A)Xξ∅ and y′ = Φ(B)Xξ∅. Then kx′k2 = kΦ(A)Xξ∅k2 = kΦ(A)Xξ∅k2 − kAξ∅k2 + kAξ∅k2 = ([Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn), Φ(A∗A)) + kAξ∅k2 ≤ k[Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)kkAk2 + kAξ∅k2 < 1 + 2ǫ′, and similarly, ky′k2 < 1 + 2ǫ′. For T in T , THE STRUCTURE OF AN ISOMETRIC TUPLE 23 ([x′ ⊗ y′]T − φ([Aξ∅ ⊗ Bξ∅]Mn), T ) = ([Φ(A)Xξ∅ ⊗ Φ(B)Xξ∅]T − φ([Aξ∅ ⊗ Bξ∅]Mn), T ) = ([Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn), Φ(A)∗T Φ(B)) ≤ k[Xξ∅ ⊗ Xξ∅]T − φ([ξ∅ ⊗ ξ∅]Mn)kkAkkBkkTk < ǫ′ 2 kTk, which implies k[x′ ⊗ y′]T − φ([Aξ∅ ⊗ Bξ∅]Mn)k < ǫ′/2. Thus k[x′ ⊗ y′]T − τk = k[x′ ⊗ y′]T − φ([ξ ⊗ υ]Mn)k ≤ k[x′ ⊗ y′]T − φ([Aξ∅ ⊗ Bξ∅]Mn)k +k[Aξ∅ ⊗ Bξ∅]Mn − [ξ ⊗ υ]Mnk < ǫ′. For 1 ≤ i ≤ q, k[x′ ⊗ zi]T k = k[AXξ∅ ⊗ zi]T k ≤ kAkk[Xξ∅ ⊗ zi]T k < ǫ′, and similarly, k[zi ⊗ y′]T k < ǫ′. Now take x = (1 + 2ǫ′)−1x′ and y = (1 + 2ǫ′)−1y′. Then by choice of ǫ′ we get kxk ≤ 1 and kyk ≤ 1. Similarly, k[x ⊗ zi]T k < ǫ and k[zi ⊗ y]T k < ǫ for 1 ≤ i ≤ q. Finally, we have k[x ⊗ y]T − τk ≤ (1 + 2ǫ′)−2k[x′ ⊗ y′]T − τk + (1 − (1 + 2ǫ′)−2)kτk < 1 − (1 + 2ǫ′)−2(1 − ǫ′) < ǫ, as required. (cid:3) 4.7. Absolute continuity and analyticity. Theorem 4.16. For n ≥ 2, every absolutely continuous isometric n-tuple is analytic. Proof. For n ≥ 2, let S = (S1, . . . , Sn) be an absolutely continuous isomet- ric n-tuple, and let S denote the weak-* closed unital algebra generated by S1, . . . , Sn. By Corollary 1.2 of [DY08], S is isomorphic to the noncommu- tative analytic Toeplitz algebra Ln. By Theorem 4.15, S has property X0,1, and hence has property A1(1). Therefore, by the discussion in Section 4.3, S is weakly closed, and hence S is actually the free semigroup algebra (i.e. THE STRUCTURE OF AN ISOMETRIC TUPLE 24 the weakly closed algebra) generated by S1, . . . , Sn. Since S is isomorphic to Ln, this implies that S is analytic. (cid:3) The next result follow from Theorem 4.12 of [Ken11]. Corollary 4.17. For n ≥ 2, let S = (S1, . . . , Sn) be an absolutely continuous isometric n-tuple acting on a Hilbert space H. Then the wandering vectors for S span H. It was shown in Corollary 5.8 of [Ken11] that every analytic isometric tuple is hyperreflexive with hyperreflexivity constant at most 3, but the next result can also be proved directly using Theorem 4.15 of the present paper and Theorem 3.1 of [Ber98]. Corollary 4.18. Absolutely continuous row isometries are hyperreflexive with hyperreflexivity constant at most 3. 5. Singular isometric tuples In Theorem 4.16, we showed that for n ≥ 2, an isometric n-tuple is abso- lutely continuous if and only if it is analytic. With this operator-algebraic characterization of an absolutely continuous isometric tuple, we are now able to give an operator-algebraic characterization of a singular isometric tuple. Theorem 5.1. For n ≥ 2, an isometric n-tuple is singular if and only if the free semigroup algebra it generates is a von Neumann algebra. Proof. Let V = (V1, . . . , Vn) be an isometric n-tuple, and let V denote the free semigroup algebra (i.e. the weakly closed algebra) generated by V . If V is a von Neumann algebra, then V has no absolutely continuous part since, by Theorem 4.16, an absolutely continuous isometric tuple is analytic, and the noncommutative analytic Toeplitz algebra Ln is not self-adjoint by Corollary 1.5 of [DP99]. Conversely, if V is singular then it has no analytic restriction to an in- variant subspace since, by Theorem 4.16, an absolutely continuous isometric tuple is analytic. Thus by Theorem 3.3, V is a von Neumann algebra. (cid:3) Example 2.2 showed that it is possible for an absolutely continuous unitary to generate a von Neumann algebra. Theorem 5.1 implies that there is no higher-dimensional analogue of this phenomenon. Recall that a family of operators is said to be reductive if every subspace invariant for the family is also coinvariant. THE STRUCTURE OF AN ISOMETRIC TUPLE 25 Corollary 5.2. For n ≥ 2, every reductive unitary n-tuple is singular. Proof. Let V = (V1, . . . , Vn) be a reductive isometric n-tuple, and let V denote the free semigroup algebra generated by V . By the dichotomy for free semigroup algebras, Corollary 4.13 of [Ken11], if V is not a von Neumann algebra, then there is a vector x that is wandering for V . Let V[x] denote the cyclic invariant subspace generated by x. Then the subspace Pn i=1 ViV[x] is invariant for V but not coinvariant, which would contradict that V is reductive. Thus V is a von Neumann algebra and V is singular by Theorem 5.1. (cid:3) Example 5.3. By Theorem 5.1, for n ≥ 2 an isometric n-tuple is singular if and only if the free semigroup algebra it generates is a von Neumann algebra. The existence of a self-adjoint free semigroup algebra on two or more generators was conjectured in [DKP01], but it took some time for the first example to be constructed. In [Read05], Read showed that B(ℓ2) is generated as a free semigroup algebra on two generators. In [Dav06], Davidson gave an exposition of Read's construction and showed that it could be generalized to show that B(ℓ2) is generated as a free semigroup algebra on n generators for every n ≥ 2. By our characterization of singularity, this gives an example of a singular isometric n-tuple for every n ≥ 2. 6. The Lebesgue-von Neumann-Wold decomposition In Theorem 4.16, we showed that for n ≥ 2, an isometric n-tuple is abso- lutely continuous if and only if it is analytic. In Theorem 5.1, we showed that for n ≥ 2, an isometric n-tuple is singular if and only if the free semigroup algebra (i.e. the weakly closed algebra) it generates is a von Neumann alge- bra. With these operator-algebraic characterizations of absolute continuity and singularity, we will be able to prove the Lebesgue-von Neumann-Wold decomposition of an isometric tuple. In the classical case, the Lebesgue decomposition of a measure guarantees that every unitary splits into absolutely continuous and singular parts. For n ≥ 2, it turns out that it is possible for a unitary n-tuple to be irreducible and neither absolutely continuous nor singular. Definition 6.1. An isometric n-tuple V = (V1, . . . , Vn) is said to be of dilation type if it has no summand that is absolutely continuous or singular. Note that by the Wold decomposition of an isometric tuple, Proposition 3.5, an isometric n-tuple of dilation type is necessarily unitary. The next THE STRUCTURE OF AN ISOMETRIC TUPLE 26 result provides a characterization of an isometric tuple of dilation type as a minimal dilation, in the sense of Section 3.3. Proposition 6.2. Let V = (V1, . . . , Vn) be an isometric n-tuple of dilation type. Then there is a subspace H coinvariant under V such that H is cyclic for V and the compression of V to H ⊥ is a unilateral n-shift. In other words, V is the minimal isometric dilation of its compression to H. Proof. Note that since V has no summand that is absolutely continuous, by Proposition 3.5 V is necessarily a unitary n-tuple. Let V denote the free semigroup algebra generated by V , and let P be the projection from Theorem 3.3 applied to V. Let H be the range of P , so that H is coinvariant under V . Let K = (H +Pn i=1 ViH) ⊖ H. Then K is wandering for the compres- sion of V to H ⊥. If K = 0, then by Theorem 3.3, V can be decomposed into the direct sum of a self-adjoint free semigroup algebra and an analytic free semigroup algebra. By the characterization of singular isometric tuples, Corollary 5.1, this would contradict that V is of dilation type. Thus K 6= 0. The fact that K is cyclic follows from the fact that H is cyclic. (cid:3) Example 6.3 (An irreducible isometric tuple of dilation type). For n ≥ 2, define isometries V1, . . . , Vn on ℓ2(N) by Vkel = en(l−1)+k, where {el}∞ each Vk is spanned by the orthonormal set l=1 is the standard orthonormal basis of ℓ2(N). Then the range of {Vkel}∞ l=1 = {en(l−1)+k}∞ l=1 = {el : l ≡ k mod n}. Therefore, the operators V1, . . . , Vn are isometries with mutually orthogonal ranges, meaning V = (V1, . . . , Vn) is an isometric tuple. We will show that V is an irreducible isometric tuple of dilation type. Since the vector e1 is fixed by V1, it is straightforward to check that the k=1 is weakly convergent to the rank one projection e1e∗ 1 }∞ sequence {V k 1. In 1 is contained in the von Neumann algebra W∗(V ) generated particular, e1e∗ by V . (In fact, this is the projection provided by Theorem 3.3). Since the vector e1 is cyclic for V , it follows that W∗(V ) = B(ℓ2(N)), and hence that V is irreducible. To see that V is of dilation type, it suffices to show that V is neither singular nor absolutely continuous. Since V is irreducible from above, if V THE STRUCTURE OF AN ISOMETRIC TUPLE 27 was singular then by Theorem 5.1, the free semigroup algebra W(V ) gen- erated by V would be B(ℓ2(N). However, the vector e2 is wandering for V and the vector e1 is orthogonal to the wandering subspace spanned by {Vwe2 : w ∈ F∗ n}, which implies that W(V ) is not transitive, and hence that W(V ) is properly contained in B(ℓ2(N). Thus V is not singular. The fact that V is not absolutely continuous follows from Theorem 4.16 and the observation made above that the sequence {V k k=1 is weakly convergent to the projection e1e∗ 1. 1 }∞ Example 6.4 (A family of irreducible isometric tuples of dilation type). It was shown in Corollary 6.6 of [DKS01] that the minimal isometric dilation of a contractive n-tuple A = (A1, . . . , An) acting on a finite-dimensional space i = I and C∗(A) has a minimal coinvariant subspace that is cyclic for C∗(A). These conditions are satisfied, for example, by the contractive tuple A = (A1, A2), where is an irreducible unitary n-tuple if and only if both Pn i=1 AiA∗ A1 = 0 1 0 0 ! , A2 = 0 0 1 0 ! . Thus the minimal isometric dilation of A is an example of an irreducible isometric tuple of dilation type. A similar construction can be carried out for all n ≥ 2. Theorem 6.5 (Lebesgue-von Neumann-Wold Decomposition). Let V = (V1, . . . , Vn) be an isometric n-tuple. Then V decomposes as V = Vu ⊕ Va ⊕ Vs ⊕ Vd, where Vu is a unilateral n-shift, Va is an absolutely continuous unitary n- tuple, Vs is a singular unitary n-tuple, and Vd is a unitary n-tuple of dilation type. Proof. The case for n = 1 follows by the discussion in Section 2. Thus we can suppose that n ≥ 2. By the Wold decomposition of an isometric tuple, Proposition 3.5, we can decompose V as where Vu is a unilateral n-shift and U is a unitary n-tuple. V = Vu ⊕ U, THE STRUCTURE OF AN ISOMETRIC TUPLE 28 By the characterization of an absolutely continuous isometric n-tuple as analytic, Theorem 4.16, and the characterization of a singular isometric n- tuple, Corollary 5.1, an isometric n-tuple cannot be both absolutely contin- uous and singular. Therefore, we can decompose U as U = Va ⊕ Vs ⊕ Vd, where Va is an absolutely continuous isometric n-tuple, Vs is a singular iso- metric n-tuple, and Vd is of dilation type. Thus we can further decompose V as V = Vu ⊕ Va ⊕ Vs ⊕ Vd, as required. (cid:3) The next result follows from combining Proposition 6.2 and Theorem 3.3. Proposition 6.6. Let V = (V1, . . . , Vn) be an isometric n-tuple of dilation type acting on a Hilbert space H. Then there is a projection P and α ≥ 1 and such that the weakly closed algebra W(V ) generated by V is of the form W(V1, . . . , Vn) = W∗(V )P + P ⊥W(V )P ⊥, where P ⊥W(V1, . . . , Vn) P ⊥H≃ L(α) n . The next result follows from the Lebesgue-von Neumann-Wold decompo- sition of an isometric tuple, Proposition 3.5, and Proposition 6.6. Theorem 6.7. Let V = (V1, . . . , Vn) be an isometric n-tuple acting on a Hilbert space H, and let V = Vu⊕Va⊕Vs⊕Vd be the Lebesgue-von Neumann- Wold decomposition of V as in Theorem 6.5. Then there is a projection P and α, β ≥ 0 such that the weakly closed algebra W(V ) generated by V is W(V ) ≃ (Ln(Vu ⊕ Va))(α) ⊕ W∗(Vs) ⊕(cid:16)W∗(Vd)P + P ⊥W(Vd)P ⊥(cid:17) , n . The von Neumann algebra W∗(V1, . . . , Vn) where P ⊥W(V1, . . . , Vn) P ⊥H≃ L(β) generated by V is W∗(V ) ≃ (B(ℓ2))(α) ⊕ W∗(Vs) ⊕ W∗(Vd). Acknowledgement. The author is grateful to his advisor, Ken Davidson, for his support and encouragement. References [Arv69] W.B. Arveson, Subalgebras of C∗-algebras, Acta Mathematica 123 (1969), 141 -- 224. [Arv75] [Ber88] [Ber98] [BFP85] [Bro78] [Bun84] [Cho74] [Con00] [Dav87] [Dav01] [Dav06] [DKS01] [DKP01] [DLP05] [DP98] [DP99] [DY08] [Fra82] [Ken11] THE STRUCTURE OF AN ISOMETRIC TUPLE 29 W.B. Arveson, Interpolation problems in nest algebras, Journal of Functional Analysis 20 (1975), No. 3, 208 -- 233. H. Bercovici, Factorization theorems and the structure of operators on Hilbert space, Annals of Mathematics 128 (1988), No. 2, 399 -- 413. H. Bercovici, Hyper-reflexivity and the factorization of linear functionals, Journal of Functional Analysis 158 (1998), No. 1, 242 -- 252. H. Bercovici, C. Foias, C. Pearcy, Dual algebras with applications to invariant subspaces and dilation theory, CBMS Regional Conference Series in Mathe- matics 56 (1985), American Mathematical Society, Providence. S.W. Brown, Some invariant subspaces for subnormal operators, Integral Equations and Operator Theory 1 (1978), No. 3, 310 -- 333. J. Bunce, Models for n-tuples of non-commuting operators, Journal of Func- tional Analysis 57 (1984), No. 1, 21 -- 30. M.D. Choi, A Schwarz inequality for positive linear maps on C∗-algebras, Illinois Journal of Mathematics 18 (1974), No. 4, 565 -- 574. J.B. Conway, A course in operator theory, Graduate Studies in Mathematics 21 (2000), American Mathematical Society, Providence. K.R. Davidson, The distance to the analytic Toeplitz operators, Illinois Jour- nal of Mathematics 31 (1987), No. 2, 265 -- 273. K.R. Davidson, Free semigroup algebras: a survey, Operator Theory: Ad- vances and Applications 129 (2000), Birkhauser, Bordeaux. K.R. Davidson, B(H) is a free semigroup algebra, Proceedings of the Ameri- can Mathematical Society 134 (2006), No. 2, 1753 -- 1757. K.R. Davidson, D.W. Kribs, M.E. Shpigel, Isometric dilations of non- commuting finite rank n-tuples, Canadian Journal of Mathematics 53 (2001), 506 -- 545. K.R. Davidson, E. Katsoulis, D.R. Pitts, The structure of free semigroup algebras, Journal für die reine und angewandte Mathematik 533 (2001), 99 -- 125. K.R. Davidson, J. Li, D.R. Pitts, Absolutely continuous representations and a Kaplansky density theorem for free semigroup algebras, Journal of Functional Analysis 224 (2005), No. 1, 160 -- 191. K.R. Davidson, D.R. Pitts, The algebraic structure of noncommutative ana- lytic Toeplitz algebras, Mathematische Annalen 311 (1998), 275 -- 303. K.R. Davidson, D.R. Pitts, Invariant subspaces and hyper-reflexivity for free semigroup algebras, Proceedings of the London Mathematical Society 78 (1999), No. 2, 401 -- 430. K.R. Davidson, D. Yang, A note on absolute continuity in free semigroup algebras, Houston Journal of Mathematics 34 (2008), 283 -- 288. A. Frahzo, Models for non-commuting operators, Journal of Functional Anal- ysis 48 (1982), No. 1, 1 -- 11. M. Kennedy, Wandering vectors and the reflexivity of free semigroup algebras, Journal für die reine und angewandte Mathematik 653 (2011), 47 -- 73. [Kri01] [LM78] [LS75] [MS10] [Pop89a] [Pop89b] [Pop09] [Pop91] [Pop96] [Read05] [SF70] [Wer52] THE STRUCTURE OF AN ISOMETRIC TUPLE 30 D.W. Kribs, Factoring in non-commutative analytic Toeplitz algebras, Jour- nal of Operator Theory 45 (2001), No. 1, 175 -- 193. R.I. Loebl, P.S. Muhly, Analyticity and flows in von Neumann algebras, Jour- nal of Functional Analysis 29 (1978), No. 2, 214 -- 252. A.N. Loginov and V.S. Shulman, Hereditary and intermediate reflexivity of W∗-algebras, Izvestiya Rossiiskoi Akademii Nauk Seriya Matematicheskaya 39 (1975), No. 6, 1260 -- 1273. P. Muhly, B. Solel, Representations of Hardy algebras: absolute continuity, intertwiners and superharmonic operators, preprint (2010), arXiv:1006.1398. G. Popescu, Isometric dilations for infinite sequences of noncommuting oper- ators, Transactions of the American Mathematical Society 316 (1989), No. 2, 523 -- 536. G. Popescu, Multi-analytic operators and some factorization theorems, Indi- ana University Mathematics Journal 38 (1989), No. 3, 693 -- 710. G. Popescu, Noncommutative transforms and free pluriharmonic functions, Advances in Mathematics 220 (2009), No. 3, 831 -- 893. G. Popescu, von Neumann inequality for (B(H)n)1, Mathematica Scandinav- ica 68 (1991), No. 2, 292 -- 304. G. Popescu, Non-commutative disc algebras and their representations, Pro- ceedings of the American Mathematical Society 124 (1996), No. 7, 2137 -- 2148. C.J. Read, A large weak operator closure for the algebra generated by two isometries, Journal of Operator Theory 54 (2005), No. 2, 305 -- 316. B. Sz-Nagy, C. Foias, Harmonic analysis of operators on Hilbert space, Uni- versitext (1970), Springer, North Holland. J. Wermer, On invariant subspaces of normal operators, Proceedings of the American Mathematical Society 3 (1952), 270 -- 277. School of Mathematics and Statistics, Carleton University, 1125 Colonel By Drive, Ottawa, Ontario K1S 5B6, Canada E-mail address: [email protected]
1707.04552	2	1707	2018-11-24T20:54:49	Relative commutant pictures of Roe algebras	[ "math.OA", "math-ph", "math-ph" ]	Let X be a proper metric space, which has finite asymptotic dimension in the sense of Gromov (or more generally, straight finite decomposition complexity of Dranishnikov and Zarichnyi). New descriptions are provided of the Roe algebra of X: (i) it consists exactly of operators which essentially commute with diagonal operators coming from Higson functions (that is, functions on X whose oscillation tends to 0 at infinity) and (ii) it consists exactly of quasi-local operators, that is, ones which have finite epsilon propogation (in the sense of Roe) for every epsilon>0. These descriptions hold both for the usual Roe algebra and for the uniform Roe algebra.	math.OA	math	RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS J ´AN SPAKULA AND AARON TIKUISIS Abstract. Let X be a proper metric space, which has finite asymptotic dimension in the sense of Gromov (or more gener- ally, straight finite decomposition complexity of Dranishnikov and Zarichnyi). New descriptions are provided of the Roe algebra of X: (i) it consists exactly of operators which essentially commute with diagonal operators coming from Higson functions (that is, functions on X whose oscillation tends to 0 at ∞), and (ii) it con- sists exactly of quasi-local operators, that is, ones which have finite ǫ-propogation (in the sense of Roe) for every ǫ > 0. These descrip- tions hold both for the usual Roe algebra and for the uniform Roe algebra. 1. Introduction The Roe algebra is a C-algebra constructed from a proper metric space, which encodes "coarse" or "large-scale" properties of the space (in the sense of Gromov). In typical applications, the space may be a complete, non-compact Riemannian manifold with bounded geometry, or a finitely generated group with the word metric. The origins of this construction come from index theory, reflecting the insight that the Roe algebra is large enough to contain indices of many operators with which one wants to do index theory -- such as geometric differential operators -- yet small enough to have interesting and informative K-theory. It plays a central role in the coarse Baum -- Connes conjecture, the study (and partial confirmation) of which has been a fruitful endeavor, leading to significant results concerning the Novikov conjecture and the scalar curvature of Riemannian manifolds [3, 12, 15, 21, 22, 24, 25, 29, 30, 31, 32]. It furthermore appears in work on the essential spectrum of Hamiltonian operators of quantum systems, Schrodinger operators, and various other operators, which are affiliated to the appropriate versions of Roe algebras [7, 8, 9, 19]. Roughly, the Roe algebra consists of bounded, locally compact op- erators on something like L2(X) (where X is the underlying space) which can be approximated by those with finite propogation. Here an operator a has "finite propogation" if it is localized near to the diag- onal; one way of making this precise is, that there exists R > 0 such that for any f, f ′ ∈ Cb(X) (acting on L2(X) as diagonal operators -- by multiplication), if the supports of f and f ′ are separated by a distance 1 2 J ´AN SPAKULA AND AARON TIKUISIS of at least R, then f af ′ = 0. Operators in the Roe algebra are required to be approximated in the operator norm by these finite propogation operators. The Roe algebra is an interesting object to study from an operator algebraic perspective: that is, asking about its structure as an operator algebra, and how this structure relates to the proper metric space from which it is constructed. For example, Ozawa showed that exactness of a group can be characterized by amenability of the corresponding uni- form Roe algebra ([18]). The question of how much of the large -- scale structure of a space is remembered by the Roe algebra, was partially answered by JS and Rufus Willett: given two uniformly discrete proper metric spaces with Yu's property A, their Roe algebras are ∗-isomorphic if and only if the spaces are coarsely equivalent ([26]). In [28], Winter and Zacharias showed an interesting one-way connection between the asymptotic dimension of a metric space and the nuclear dimension of the corresponding uniform Roe algebra; the latter is a numerical invari- ant for amenable C-algebras which is crucial in recent results in the classification of amenable C-algebras. Their result is that the nuclear dimension of the Roe algebra is at most the asymptotic dimension of the underlying uniformly discrete proper metric space, and they asked the (still open) question of whether the reverse inequality also holds. In this paper, we look at a fundamental question: exactly which operators are in the Roe algebra? In [22], Roe defined the concept of finite ǫ-propogation for an operator a on L2(X), as the following variant of finite propogation: a has finite ǫ-propogation if there exists R > 0 such that for any f, f ′ ∈ Cb(X), if the supports of f and f ′ are separated by a distance of at least R, then kf af ′k ≤ ǫkf k · kf ′k. Operators with finite ǫ-propogation for all ǫ > 0 have also been called quasi-local operators in the literature (originally from [21, Page 100]). It is a straightforward observation that, although limits of finite propogation operators need not have finite propogation, limits of finite ǫ-propogation operators have finite ǫ-propogation (that is, the set of quasi-local operators is norm-closed). Therefore, all operators in the Roe algebra are quasi-local. The question we address is the converse: if an operator is quasi-local, is it in the Roe algebra, i.e., is it approximated by operators with finite propogation? We provide an affirmative answer in the situation that the space has finite asymptotic dimension (as predicted by Roe), and more generally under the hypothesis of straight finite decomposition complexity of Dranishnikov and Zarichnyi [4]. The latter is a weaker version of the "classical" finite decomposition complexity, as defined by Guentner, Tessera, and Yu [13, 14]. A motivation for asking whether quasi-local operators are in the Roe algebra, pointed out to the authors by Alexander Engel, is that whereas indices of genuine differential operators are known to be in RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 3 the Roe algebra, corresponding arguments only shows that indices of pseudo-differential operators (using their most natural representative) are quasi-local (see [5, Section 2]). Since the Roe algebra is better studied and understood than the C-algebra of all quasi-local operators, it is interesting and useful to know that a pseudo-differential operator belongs to it; indeed, our result answers [5, Question 6.4] under the assumption of straight finite decomposition complexity (this sort of assumption is anticipated in the question). Additionally, we expose that the above question can be reformulated in terms of essential commutation with Higson functions, or in terms of relative commutants. So far we have been a bit vague about what we mean by the Roe algebra (hiding behind the phrase "something like L2(X)"). This is because in the literature there are two different versions of the Roe al- gebra, the "Roe algebra" and the "uniform Roe algebra". Our results apply to both of these C-algebras, and indeed our main theorem is formulated in a way that encompasses both, as well as the "uniform al- gebra" UC ∗(X). The main result was proven by Lange and Rabinovich for the uniform Roe algebra of Zd in [17]. Engel proved a special case of the result, namely that for discrete groups G that are lattices in a Riemannian manifold with bounded geometry and polynomial volume growth, quasi-local operators that decay sufficiently quickly are in the Roe algebra ([6, Corollary 2.33]).1 Let us now summarize the argument behind the main result: that quasi-local operators are in the Roe algebra (assuming straight finite decomposition complexity). Suppose for simplicity that X is a discrete space with asymptotic dimension at most 1 -- for example a finitely generated free group. This case is much more restricted than finite de- composition complexity, but still difficult enough to allow us to convey the main ideas. Let a be a quasi-local operator. Asymptotic dimension at most 1 will allow us to decompose the space X into 2 pieces, X (0) and X (1), each piece being a disjoint union of sets that are far apart from each other and uniformly bounded in diameter. The characteristic functions e(0), e(1) of these pieces produce a partition of unity, and di- vides a into a sum of four pieces: e(i)ae(j) over i, j = 0, 1. Each e(i)ae(i) looks roughly like an infinite block matrix, indexed by the pieces from X (i). The hypothesis that a is quasi-local (finite ǫ-propogation) gives 1In fact, Engel proved the result for quasi-local operators that decay sufficiently on any Riemannian manifold with bounded geometry and polynomial volume growth. For groups, polynomial growth implies virtual nilpotency ([11]), which in turn implies finite asymptotic dimension ([1, Corollary 68]). To our knowledge, it is not known whether polynomial volume growth implies finite asymptotic di- mension (or even (straight) finite decomposition complexity) outside of the case of groups. 4 J ´AN SPAKULA AND AARON TIKUISIS us a lot of control over the norm of the non-diagonal entries of this ma- trix, and a conditional expectation argument allows us to conclude that e(i)ae(i) is not far away from its "restriction" to the diagonal (provided that the pieces in X (i) are sufficiently well separated), see Corollary 4.3. Since the pieces of the X (i) are uniformly bounded, the operator we get by expecting onto the diagonal has genuinely finite propogation. An algebraic trick allows us to view the asymmetric pieces e(i)ae(j) as matrices in a similar way, so that we can likewise approximate each of them by finite propogation operators. In this way, we approximate a as a sum of four operators with finite propogation. Outline. In Section 2 we introduce our general setup, with the Roe algebra, the uniform Roe algebra, and the uniform algebra as exam- ples. We then state the main result, Theorem 2.8, in the language of our general setup. We give some background on asymptotic dimension and (straight) finite decomposition complexity in Section 2.1. The equiva- lence between quasi-locality and the relative commutant-type property is fairly straightforward, and laid out in Section 3. We use a more tech- nical formulation of quasi-locality as a stepping stone towards proving that it implies being in the Roe algebra (assuming straight finite de- composition complexity), a proof that is carried out in Section 4. In Section 5, we prove that the relative commutant-type property is equiv- alent to essential commutation with Higson functions. The final sec- tion, Section 6, is concerned with the commutative (but non-separable) C-algebra VL∞(X) that arises in our relative commutant-type prop- erty, looking at how well it determines X (up to coarse equivalence), and at its nuclear dimension (roughly, the covering dimension of its spectrum). Acknowledgments. AT was supported by EPSRC EP/N00874X/1. JS was supported by Marie Curie FP7-PEOPLE-2013-CIG Coarse Anal- ysis (631945). We would like to thank Ulrich Bunke, Alexander Engel, John Roe, Thomas Weighill, Stuart White, and Rufus Willett for com- ments and discussion relating to this piece. 2. Definitions and the main result Let A be a C-algebra. We denote by A1 the closed unit ball of A. For a, b ∈ A and ǫ > 0, we write a ≈ǫ b to mean ka − bk ≤ ǫ. Define A∞ := l∞(N, A)/{(an)∞ n=1 ∈ l∞(N, A) : lim n→∞ kank = 0}, which is a C-algebra. We now set up a general situation to which our main result applies, encompassing both Roe algebras and uniform Roe algebras, as well as uniform algebras (see Example 2.5). Subsequently, we will state our main result in its full generality (Theorem 2.8) RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 5 Definition 2.1. Let X be a proper metric space. By an X-module, we mean a Hilbert space H and an injective unital ∗-homomorphism Cb(X) → B(H), which is strictly continuous when viewing Cb(X) and B(H) as multiplier algebras of C0(X) and K(H) respectively. We shall suppress the ∗-homomorphism Cb(X) → B(H), and treat Cb(X) as a C-subalgebra of B(H). For R ≥ 0, an operator a ∈ B(H) has propogation at most R if for any f, f ′ ∈ Cb(X), if the supports of f and f ′ are R-disjoint then f af ′ = 0. For R ≥ 0 and ǫ > 0, an operator a ∈ B(H) has ǫ- propogation at most R if for any f, f ′ ∈ Cb(X)1, if the supports of f and f ′ are R-disjoint then kf af ′k < ǫ. An operator a ∈ B(H) is quasi-local if for every ǫ > 0, it has finite ǫ-propogation. Definition 2.2. Let X be a proper metric space and let H be an X- module. Given an equicontinuous family (ej)j∈J of positive contractions in Cb(X) with pairwise disjoint supports, define the block cutdown map θ(ej )j∈J : B(H) → B(H) by θ(ej )j∈J (a) :=Xj∈J ejaej (using disjointness of the supports and the fact that the family is con- tractive, the right-hand sum converges in the strong operator topology). Let B ⊆ B(H) be a C-subalgebra such that Cb(X)BCb(X) = B. B is closed under block cutdowns if θ(ej )j∈J (B) ⊆ B for every equicontinuous family (ej)j∈J of positive contractions from Cb(X) with pairwise disjoint supports. For an equicontinuous family (ej)j∈J of positive contractions from Cb(X) with pairwise disjoint supports, the block cutdown map θ(ej )j∈J defined above is evidently completely positive and contractive (c.p.c.). Note that multiplication by Cb(X) commutes with block cutdowns: f θ(ej )j∈J (a) = θ(ej )j∈J (f a) and θ(ej )j∈J (a)f = θ(ej )j∈J (af ) for f ∈ Cb(X) and a ∈ B(H). Also note that (2.1) kθ(ej )j∈J (a)k = sup j∈J kejaejk. Note that, if (ej)j∈J is an equicontinuous family of positive contrac- tions from Cb(X) with uniformly bounded, pairwise disjoint supports, then θ(ej )j∈J (a) has finite propogation, for every a ∈ B(H). Definition 2.3. Let X be a proper metric space, H an X-module, and let B ⊆ B(H) be a C-subalgebra such that Cb(X)BCb(X) = B, and which is closed under block cutdowns. Define (i) Roe(X, B) := {b ∈ B : b has finite propogation}) (ii) K(X, B) := C0(X)BC0(X). k·k , and 6 J ´AN SPAKULA AND AARON TIKUISIS If, in addition, we have (2.2) [C0(X), B] ⊆ K(X, B), we shall call Roe(X, B) a Roe-like algebra of X. Remark 2.4. The condition (2.2) implies that K(X, B) is an ideal in Roe(X, B). It is automatically satisfied in all the examples below, where in fact C0(X)B ⊆ K(X, B) (and K(X, B) turns out to be the ideal of compact operators). Finally, it is not needed for the substantial part of this piece, so we shall explicitly refer to it when needed. Example 2.5. Let X be a uniformly discrete proper metric space. Let H′ be an infinite dimensional, separable Hilbert space. Set Hu := l2(X) and H := l2(X, H′); Cb(X) acts on both of these by pointwise multiplication, making them X-modules. (i) With Bu := B(Hu), we see that Cb(X)BuCb(X) = Bu, and Bu is closed under block cutdowns. In this case, Roe(X, Bu) = C ∗ u(X), the uniform Roe algebra, and K(X, Bu) = K(Hu). Since C0(X) ⊆ K(Hu), it follows that C0(X)Bu ⊆ K(Hu) = K(X, Bu). (ii) Set B equal to the set of all b ∈ B(H) which are locally compact, in the sense that for every f ∈ C0(X), f b, bf ∈ K(H). We see that Cb(X)BCb(X) = B, and B is closed under block cutdowns. Then Roe(X, B) = C ∗(X), the Roe algebra, and K(X, B) = K(H). (iii) Assume that X has bounded geometry. Set B0 equal to the closure of the set of all b = (bx,y)x,y∈X ∈ B(H) for which the rank of bx,y ∈ B(H′) is uniformly bounded. When b = (bx,y)x,y∈X ∈ B(H) has entries with rank bounded by k, then so does any block cutdown map applied to b. Since each block cutdown map is continuous, it follows that B0 is closed under block cutdowns. Continuity of multiplication ensures that Cb(X)B0Cb(X) = B0. When X has bounded geometry, then Roe(X, B0) = UC ∗(X), the uni- form algebra of X, defined as the closure of finite propogation operators b = (bx,y)x,y∈X ∈ B(H) for which the rank of bx,y is uniformly bounded. To see this, it is clear that Roe(X, B0) contains UC ∗(X). To show Roe(X, B0) ⊆ UC ∗(X), it suffices to check that every finite propoga- tion operator a ∈ B0 is contained in UC ∗(X). For such a, say its propogation is less than R > 0. Set K := sup x∈X BR(x), RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 7 which is finite due to the hypothesis of bounded geometry. Define ER : B(H) → B(H) by ER ((bx,y)x,y∈X) := (cx,y)x,y∈X where cx,y :=(bx,y, 0, d(x, y) < R; d(x, y) ≥ R. Note that kER ((bx,y)x,y∈X)k ≤ K k(bx,y)x,y∈Xk (this is a straightfor- ward argument, see e.g., the proof of [28, Lemma 8.1]), so that in particular, EK is continuous. Also note that EK(a) = a. Since a ∈ B0, it is a limit of a sequence of operators bn = (bn x,y)x,y∈X such that for each n, there exists Kn bounding the rank of bn x,y over all x, y ∈ X. Thus the same bound Kn applies to EK(bn) so that EK(bn) ∈ UC ∗(X). By continuity of EK, a = limn→∞ EK(bn) ∈ UC ∗(X). In this example, we also have K(X, B0) = K(H), and since B0 ⊆ B (from (ii)), C0(X)B ⊆ K(H) = K(X, B0) (iv) Generalizing (ii), let X be any proper metric space and let H be an adequate X-module in the sense of [22, Definition 3.4]. Recall that an operator b ∈ B(H) is locally compact if C0(X)b, bC0(X) ⊆ K(H). Set B equal to the set of all locally compact, bounded operators. One can easily see that Cb(X)BCb(X) = B; it is also true that B is closed under block cutdowns. To see this, let b ∈ B(H) be locally compact with kbk ≤ 1, let (ej)j∈J be an equicontinuous family of positive contractions in Cb(X) with pairwise disjoint supports, and set b′ := θ(ej )j∈J (b), which we must prove is locally compact. As K(H) is closed, it suffices to show that for any f ∈ Cc(X) with kf k ≤ 1, f b′, b′f ∈ K(H). Given ǫ > 0, note that b′ ≈2ǫ θ((ej −ǫ)+)j∈J (b), where (ej − ǫ)+ ∈ Cb(X) is given by (ej − ǫ)+(x) := max{ej(x) − ǫ, 0}. By equicontinuity and pairwise disjointness of the family (ej), we may choose δ such that if d(x, y) < δ and j 6= j′, then at most one of ej(x) or ej ′(y) can be nonzero. Thus if f ∈ Cc(X), then by compactness of its support, there are only finitely many j for which f (ej − ǫ)+ 6= 0. Consequently, f b′ ≈2ǫ f θ((ej −ǫ)+)j∈J (b) (ej − ǫ)+b(ej − ǫ)+ f (ej − ǫ)+b(ej − ǫ)+, = fXj∈J =Xj and as this is a finite sum of elements of K(H), it is itself in K(H). As K(H) is closed and ǫ > 0 is arbitrary, it follows that f b′ ∈ K(H). Like- wise, b′f ∈ K(H), establishing that b′ is locally compact, and therefore that B is closed under block cutdowns. 8 J ´AN SPAKULA AND AARON TIKUISIS In this example, we get Roe(X, B) = C ∗(X), the Roe algebra, and K(X, B) = K(H) = C0(X)B. Definition 2.6. Let X be a metric space. A bounded sequence (fn)∞ n=1 from Cb (X) is very Lipschitz if, for every L > 0, there exists n0 such that fn is L-Lipschitz for all n ≥ n0. Let VL (X) denote the set of all very Lipschitz bounded sequences from Cb (X). Define VL∞ (X) := VL (X) /{(fn)∞ n=1 ∈ VL (X) lim n→∞ kfnk = 0}. VL (X) is a C∗-subalgebra of l∞ (N, Cb (X)),2 and therefore the quo- tient VL∞ (X) is a C-subalgebra of (Cb(X))∞. E.g., if X is a finitely generated group G with the word metric, then VL∞ (X) can be identified with the fixed point algebra of l∞ (G)∞ under the action of G induced by left translation on l∞ (G). Recall the following definition from [23]. Definition 2.7. Let X be a proper metric space. A function g ∈ Cb(X) is a Higson function (also called a slowly oscillating function) if, for every R > 0 and ǫ > 0, there exists a compact set A ⊆ X such that for x, y ∈ X\A, if d(x, y) < R then g(x) − g(y) < ǫ. The set of all Higson functions on X is denoted Ch(X). E.g., if X is a finitely generated group G with the word metric, then Ch(X) ⊆ l∞(X) is the preimage of the fixed point algebra of l∞ (G) /c0(G) under the action of G induced by left translation on l∞ (G). In the following, H is an X-module, and we view both VL∞ (X) and B ⊆ B (H) as C-subalgebras of B (H)∞, and consider the relative commutant B ∩ VL∞ (X)′ . It is easy to see (at least in the standard cases of Example 2.5) that any finite propogation operator commutes with VL∞ (X), and by taking limits it follows that Roe(X, B) ⊆ B ∩ VL∞ (X)′ . The main result is as follows. Recall that straight finite decomposi- tion complexity, as introduced in [4], is a weakening of finite asymptotic dimension ([14, Theorem 4.1]). Both properties are defined in the fol- lowing subsection. 2To check that the product of two very Lipschitz sequences is itself very Lipschitz, use the fact that if f, g are bounded functions, such that f is L-Lipschitz and g is L′-Lipschitz, then f g is (kf kL′ + kgkL)-Lipschitz. RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 9 Theorem 2.8. Let X be a proper metric space, H an X-module, and let B ⊆ B(H) be a C-subalgebra such that Cb(X)BCb(X) = B, which is closed under block cutdowns, and such that (2.2) holds. For b ∈ B, the following are equivalent. (i) [b, f ] = 0 for all f ∈ VL∞(X); (ii) b is quasi-local (it has finite ǫ-propogation for every ǫ > 0); (iii) [b, g] ∈ K(X, B) (i.e., b essentially commutes with g) for all g ∈ Ch(X). If X has straight finite decomposition complexity, then these are also equivalent to (iv) b ∈ Roe(X, B). The equivalence of (i) and (ii) is fairly straightforward, and the equiv- alence of these conditions with (iii) (at least in the standard cases of Example 2.5) seems to be known by coarse geometers; we shall provide a detailed proof for completeness. The implication (iv) =⇒ (ii) is straightforward and holds in complete generality. The implication (i) ⇒ (iv) was proven by Lange and Rabinovich for the uniform Roe algebra of Zd (i.e., the case X = Zd, H = l2(X), and B = B(H) as in Example 2.5 (i)) in [17] (see [19, Proposition 8] for a proof in English). The result (ii) ⇒ (iv) was claimed by Roe in a remark on page 20 of [22] under a "finite dimensionality" assumption, but it was later found that his supposed proof was incomplete ([20]). The present paper is to the authors' knowledge the first complete proof of a more general case (which is even more general than finite asymptotic dimension). Question 2.9. Is there a uniformly discrete countable metric space with bounded geometry, for which (i)-(iii) does not imply (iv) of Theo- rem 2.8? 2.1. Coarse geometric notions. We collect some terminology from [13, 14, 4]. Definition 2.10. Let X be a proper metric space, let Z, Z ′ ⊆ X, let X and Y be metric families (i.e. at most countable sets of subsets of X), and finally let R ≥ 0. • We shall say that X is uniformly bounded, if supY ∈X diam(Y ) < ∞. • We shall denote the metric neighbourhood of Z of radius R by NR(Z) := {z ∈ X d(z, Z) ≤ R}. We further set NR(X ) := {NR(Y ) : Y ∈ X }. • The distance between Z and Z ′ is d(Z, Z ′) := inf{d(z, z′) : z ∈ Z, z′ ∈ Z ′}. 10 J ´AN SPAKULA AND AARON TIKUISIS • A family (Yj)j∈J is R-disjoint if d(Yj, Yj ′) > R for all j 6= j′; we write Yj GR-disjoint for the union of the Yj to indicate that the family is R-disjoint. • We say that Z R-decomposes over Y, if we can decompose Z = X0 ∪ X1 and Xi = GR-disjoint Xij, i = 0, 1, such that Xij ∈ Y for all i, j. • We say that X R-decomposes over Y, denoted X R−→ Y, if every Y ∈ X R-decomposes over Y. • We say that X has asymptotic dimension at most n, if for every r ≥ 0, we can decompose X = X0 ∪ · · · ∪ Xn and Xi = Gr-disjoint Xij, i = 0, . . . , n, such that the metric family {Xij i, j} is uniformly bounded. • We say that X has straight finite decomposition complexity, if for any sequence 0 ≤ R1 < R2 < · · · , there exists m ∈ N and Ri−−→ Xi metric families {X} = X0, X1, . . . , Xm, such that Xi−1 for i = 1, . . . , m, and the family Xm is uniformly bounded. The notion of straight finite decomposition complexity (sFDC) [4] is apriori weaker than the original notion of finite decomposition com- plexity of Guentner, Tessera and Yu [13, 14], see [4, Proposition 2.3]. The definition of finite decomposition complexity uses a certain "de- composition game", which effectively means that the choices of Ri can depend on the previous decompositions X1, . . . , Xi−1. Already finite decomposition complexity is weaker than finite asymp- totic dimension ([14, Theorem 4.1]). 3. Proof of (i) ⇔ (ii) To prove the main result, we begin with a technical-looking charac- terization of condition (ii). Lemma 3.1. Let X be a proper metric space, let H be an X-module, and let a ∈ B(H). Then k[a, f ]k < ǫ for every f ∈ VL∞ (X)1 if and only if there exists L > 0 such that k[a, f ]k < ǫ whenever f ∈ Cb(X)1 is L-Lipschitz. Remark 3.2. As we shall need to refer to the conclusion of the above lemma later, we shall fix the following notation. In the setup as in the above lemma, we write a ∈ Commut(L, ǫ) if k[a, f ]k < ǫ whenever f ∈ Cb(X)1 is L-Lipschitz. RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 11 Proof. The reverse implication is immediate from the definition of VL (X). For the forward direction, we use a proof by contradiction. Suppose for a contradiction that, for every n there exists fn ∈ Cb (X)1 that is (1/n)-Lipchitz and k[a, fn]k ≥ ǫ. Then evidently, (fn)∞ tradicts the hypothesis that k[a, VL∞ (X)1]k < ǫ. n=1 ∈ VL (X) yet lim∞ k[a, fn]k ≥ ǫ. This con- (cid:3) Proof of Theorem 2.8 (i) ⇒ (ii). Suppose that [b, VL∞ (X)1] = 0 and let ǫ > 0. By Lemma 3.1, let b ∈ Commut(L, ǫ) (in the notation of Remark 3.2) for some L > 0. We claim that b has ǫ-propogation at most L−1. Certainly, suppose that f, f ′ ∈ Cb (X)1 have L−1-disjoint supports. We may define g ∈ Cb (X) such that gsuppf ≡ 1, gsuppf ′ ≡ 0 and g is L-Lipschitz. Hence, k[b, g]k < ǫ. Consequently, kf bf ′k = kf gbf ′k ≤ k[b, g]k + kf bgf ′k < ǫ + 0, as required. (cid:3) Proof of Theorem 2.8 (ii) ⇒ (i). Suppose that b has finite ǫ-propogation for all ǫ > 0. Assume that b is a contraction. We shall verify the con- dition in Lemma 3.1. Therefore, let ǫ > 0 be given. Pick N such that 6/N < ǫ/2. By the hypothesis, let b have (ǫ/ (2N 2))-propogation at most R > 0. Let f ∈ Cb (X)1 be (2RN)−1-Lipschitz. We claim that k[b, f ]k < ǫ. Surely, define sets A1 := f −1(cid:0)[0, 1 N ](cid:1) , Ai := f −1(cid:0)(cid:0) i−1 N , i N(cid:3)(cid:1) , These sets partition X and, for i − j > 1, Ai is (2R)-disjoint from Aj. We may find a partition of unity e1, . . . , eN ∈ Cb(X) such that ei is supported in NR/2(Ai). It follows that the supports of ei and ej are R-disjoint when i − j > 1. i = 2, . . . , N. Thus, (3.1) Also, (3.2) keibejk < ǫ 2N 2 . f ≈1/N i N ei N Xi=1 12 and so k[f, b]k (3.2) ≤ 2 J ´AN SPAKULA AND AARON TIKUISIS N = 2 i N N + k[ i N ei, b]k Xi=1 N +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N N eib! N Xi=1 Xj=1 N(cid:1) eibej(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi,j=1(cid:0) i keibejk +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi−j≤1(cid:0) i N + Xi−j>1 N − j = 2 ≤ 2 ej! − N Xi=1 ei! N Xj=1 j N bej!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N − j . N(cid:1) eibej(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ǫ The terms of the first sum are each dominated by 2N 2 by (3.1), so this entire sum is less than ǫ/2. The second sum can be broken into 4 sums with orthogonal terms (namely, note that when i = j, the terms vanish; what remains is j = i + 1 and j = i − 1, and we break each of these into even and odd parts). Each of the terms of the second sum has norm at most 1/N; thus, we have N + ǫ k[f, b]k < 2 N < ǫ, 2 + 4 as required. (cid:3) 4. Proof of (i) ⇒ (iv) In this section, we prove that (i) ⇒ (iv) in Theorem 2.8. We begin by establishing a few general functional analytic facts. Recall that the strong* topology on B(H) is the one in which a net α → a∗ in (aα) converges to a ∈ B(H) if and only if both aα → a and a∗ the strong operator topology (i.e., kaαξ −aξk → 0 and ka∗ αξ −a∗ξk → 0 for every ξ ∈ H). A conditional expectation from C-algebra A to a C-subalgebra B is a completely positive and contractive projection E from A to B satisfying for all b1, b2 ∈ B and a ∈ A. E(b1ab2) = b1E(a)b2 Lemma 4.1. Let H be a Hilbert space and let G be a subgroup of the group of unitary operators, which is compact in the strong∗ topology. Then there is a unique conditional expectation EG : B(H) → G′ whose restriction to the unit ball is weak operator topology continuous. It satisfies (4.1) kEG(a) − ak ≤ sup u∈G k[a, u]k , a ∈ B(H). RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 13 Proof. Let µG be the normalized Haar measure on G (under the strong∗ topology). Fix a ∈ B(H), and consider the map G → B(H) defined by u 7→ u∗au. Then, with the strong∗ topology on the domain G and the weak operator topology on the range B(H), this map is continuous. We may therefore integrate, defining EG(a) := WOT-ZG u∗au dµG(u). (Here, WOT-RG · dµG indicates the Pettis integral, unique operator satisfying i.e., EG(a) is the hEG(a)ξ, ηi =ZG hu∗auξ, ηi, for all ξ, η ∈ H.) Using invariance of the Haar measure µG, one easily sees that EG(a) commutes with all of G. We now check (4.1); for this, set γ := supu∈G k[a, u]k. For η, ξ ∈ H, h(EG(a) − a) η, ξi =(cid:12)(cid:12)(cid:12)(cid:12) ZG ≤ZG ≤ZG =ZG ≤ γ kηk kξk . h(u∗au − a) η, ξi dµG(u)(cid:12)(cid:12)(cid:12)(cid:12) h(u∗au − a) η, ξi dµG(u) ku∗au − ak kηk kξk dµG(u) k[u, a]k kηk kξk dµG(u) Therefore, (4.1) follows. In particular, we conclude that if a ∈ G′ then EG(a) = a. It is also straightforward to see that the function EG is c.p.c., and therefore it is a conditional expectation. On the unit ball of B(H), the integral defining EG can be uniformly approximated in the weak operator topology by (finite) Riemann sums, which themselves are continuous in the weak operator topology. It follows that the restriction of EG to the unit ball is continuous using the weak operator topology. If E : B(H) → G′ is another conditional expectation whose restric- tion to the unit ball is weak operator topology continuous, then for a 14 J ´AN SPAKULA AND AARON TIKUISIS contraction a ∈ B(H), EG(a) = E (EG(a)) (E fixes G′) u∗au dµG(u)(cid:17) E(u∗au) dµG(u) u∗E(a)u dµG(u) = E(cid:16)WOT-ZG = WOT-ZG = WOT-ZG = EG(E(a)) = E(a) (WOT-continuity of EB(H)1) (E is a conditional expectation) (EG fixes G′). Thus, E = EG. (cid:3) Recall that an atomic abelian von Neumann algebra is a von Neu- mann algebra isomorphic to l∞(X), for some set X. In the following, when H = l2(X), then the conditional expectation B(l2(X)) → l∞(X) consists simply of taking an operator to its diagonal. Corollary 4.2. Let D ⊂ B(H) be an atomic abelian von Neumann al- gebra. Then there is a unique conditional expectation ED : B(H) → D′ whose restriction to the unit ball is weak operator topology continuous. It satisfies (4.2) kED(a) − ak ≤ sup k[a, x]k , a ∈ B(H). x∈D,kxk≤1 Proof. Without loss of generality, D contains the identity operator. D is generated by a family of orthogonal projections (pj)j∈J, whose sum converges strongly to 1. Define G :=(Xj∈J (−1)αj pj : (αj)j∈J ∈ (Z/2)J) . This is a strong∗ compact subgroup of the unitary group of D (it is homeomorphic to (Z/2)J with the product topology), so that Lemma 4.1 applies to it. It is clear that it generates D as a von Neumann algebra, so that G′ = D′. The conclusion follows from Lemma 4.1. (cid:3) Corollary 4.3. Let X be a proper metric space, let H an X-module, and let a ∈ B (H). Suppose a ∈ Commut(L, ǫ) for some L, ǫ > 0 (in the notation of Remark 3.2). Let (ej)j∈J be a family of positive contractions from Cb(X) with (2L−1)-disjoint supports, and define e := Pj∈J ej. Then, with θ(ej )j∈J from Definition 2.2, keae − θ(ej )j∈J (a)k ≤ ǫ. RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 15 Proof. Set Aj equal to the support of ej for each j ∈ J. We may find pairwise disjoint projections pj ∈ B(H), for j ∈ J, such that pj acts as a unit on ej. Define D to be the von Neumann subalgebra generated the unique conditional expectation provided by Corollary 4.2. Then by {pj : j ∈ J} (with unit 1D = Pj pj), and let ED : B(H) → D′ be one finds that for x ∈ B(H), ED(x) = Pj∈J pjxpj (converging in the strong operator topology), and therefore ED(eae) = θ(ej )j∈J (a). Using (2L−1)-disjointness of the family (Aj)j∈J , for f ∈ D1, there exists a function f ∈ Cb (X)1 that is L-Lipschitz such that Therefore, f = 1D f . f (eae) = f eae = e f ae ≈ǫ ea f e = (eae)f. Hence, by Corollary 4.2, eae ≈ǫ ED(eae) = θ(ej )j∈J (a), as required. (cid:3) Definition 4.4. Let X be a proper metric space, H be an X-module, a ∈ B(H) and let X be a metric family (of subsets of X). We say that a is block diagonal with respect to X , if there exists an equicontinuous family (ej)j∈J of positive contractions in Cb(X) with pairwise disjoint supports, such that a = θ(ej )j∈J (a), and the support of each ej is con- tained in some set Yj ∈ X . Furthermore, in this case we shall denote aYj := ejaej and call these operators blocks of a. The next lemma sets up the "induction step" to be applied in the context of the decomposition game, in the proof of Theorem 2.8 (i) ⇒ (iv). Lemma 4.5. Let X be a proper metric space, and let Y be a metric family such that {X} 4L−1+4 −−−−−→ Y for some L > 0. Let H be an X- module, and let a ∈ B (H). Let ǫ > 0 be such that a ∈ Commut(L, ǫ). Then we can write a ≈8ǫ a00 + a01 + a10 + a11, where each aii′ is of the form θ(fk)k∈K (gag′) for some contractions g, g′ ∈ Cb(X) and some family (fk)k∈K of 1-Lipschitz positive contractions 16 J ´AN SPAKULA AND AARON TIKUISIS in Cb(X) with disjoint supports, such that the support of each fk is contained in a set in NL−1+1(Y). Proof. By the decomposition assumption, we can write X = X (0) ∪ X (1), X (i) = Gj∈Ji (4L−1+4)-disjoint X (i) j , i = 0, 1, j ∈ Y for each i, j. We may find a partition of unity consisting j ∈ Cb(X), for i = 0, 1 and j ∈ Ji, j ). It follows that with X (i) of 1-Lipschitz positive contractions e(i) such that the support of e(i) j for each i, the supports of (e(i) is contained in N1(X (i) j )j∈Ji are (4L−1 + 2)-disjoint. For each i = 0, 1, define e(i) :=Pj∈Ji a = e(0)ae(0) + e(0)ae(1) + e(1)ae(0) + e(1)ae(1), e(i) j . Since it suffices to find aii′ ≈ e(i)ae(i′) for each i, i′ ∈ {0, 1} with the required properties. (We will be precise about the degree of approximation -- in short, it depends on whether i and i′ are equal.) For the case i = i′, Corollary 4.3 shows that e(i)ae(i) ≈ǫ θ(e(i) j )j∈Ji (e(i)ae(i)) =: aii. (cid:16)NL−1+1(X (i) j )(cid:17)j∈Ji The latter operator is clearly block diagonal with respect to N1(Y) (hence also with respect to NL−1+1(Y)). Turning now to the case i 6= i′, note that for fixed i, the family is (2L−1 + 2)-disjoint. For each i, j, there exists e(i) j ∈ Cb (X) that is L-Lipschitz, that acts as the identity on e(i) j , and is supported on NL−1+1(X (i) e(i) j . We have j ). For each i, define e(i) :=Pj∈Ji e(i)ae(i′) = e(i)e(i)ae(i′)e(i′) ≈2ǫ e(i)e(i′)ae(i)e(i′). For each j ∈ Ji and j′ ∈ Ji′, there exists a 1-Lipschitz positive con- traction fj,j ′ ∈ Cb(X) that is 1 on NL−1+1(X (i) j ′ ), and is supported on the metric neighbourhood of this set of radius 1. In particular, the support of each fj,j ′ is contained in a set in NL−1+2(Y), the family of supports of the family (fj,j ′)j,j ′∈Ji′ is (2L−1)-disjoint, and j ) ∩ NL−1+1(X (i′) f := Pj,j ′ fj,j ′ acts as an identity on (both sides of) e(i)e(i′)ae(i)e(i′). Applying Corollary 4.3, we obtain f e(i)e(i′)ae(i)e(i′)f ≈ǫ θ(fj,j′ )j∈Ji, j′ ∈Ji′ (cid:16)e(i)e(i′)ae(i)e(i′)(cid:17) . RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 17 Thus, e(i)ae(i′) ≈2ǫ e(i)e(i′)ae(i)e(i′) = f e(i)e(i′)ae(i)e(i′)f ≈ǫ θ(fj,j′ )j∈Ji, j′ ∈Ji′ (cid:16)e(i)e(i′)ae(i)e(i′)(cid:17) =: aii′. By construction, it is clear that aii′ is block diagonal with respect to NL−1+2(Y). Summarizing, we have a ≈ǫ+3ǫ+3ǫ+ǫ a00 + a10 + a01 + a11, and all the (cid:3) aii′ are of the right form. We now strengthen the previous lemma by allowing an arbitrary met- ric family in place of {X} and with a a correspondingly block diagonal operator. Lemma 4.6. Let X be a proper metric space, and let X and Y be metric families, such that X 4L−1+4 −−−−−→ Y for some L > 0. Let H be an X-module, and let a ∈ B (H) be block diagonal with respect to X . Let ǫ > 0 be such that a ∈ Commut(L, ǫ). Then we can write (4.3) a ≈8ǫ a00 + a01 + a10 + a11, where each aii′ is of the form θ(fk)k∈K (gag′) for some contractions g, g′ ∈ Cb(X) and some equicontinuous family (fk)k∈K of positive contractions in Cb(X) with disjoint supports, such that the support of each fk is contained in a set in NL−1+1(Y). In particular: (i) each aii′ is block diagonal with respect to NL−1+1(Y), (ii) if a ∈ Commut(L′, ǫ′) for some L′, ǫ′ > 0, then each aii′ is in Commut(L′, ǫ′) as well, and (iii) if B ⊆ B(H) is a C-subalgebra such that Cb(X)BCb(X) = B and B is closed under block cutdowns, and if a is in B, then each aii′ is in B as well. Proof. Without loss of generality, both X and Y are closed under tak- ing subsets. Start by letting (ej) be an equicontinuous family of pos- itive contractions in Cb(X) with disjoint supports, such that Yj := supp(ej) ∈ X . Applying Lemma 4.5 to each aYj yields aYj ≈8ǫ aj 00 + aj 01 + aj 10 + aj 11, satisfying the conclusions of that lemma. Set aj ii′. aii′ :=Xj Then (4.3) follows from (2.1) To see that each aii′ has the right form, fix i and i′. For each j, there j ∈ Cb(X) and a family (fj,k)k∈Kj of 1-Lipschitz exist contractions gj, g′ 18 J ´AN SPAKULA AND AARON TIKUISIS positive contractions in Cb(Yj) with disjoint supports, such that aj ii′ = θ(fj,k)k∈Kj (gjag′ j). and the support of each fj,k is of the form Y ∩ Z for some Z ∈ NL−1+1(Y). Then observe that aii′ = θ(e1/4 j fj,k)j∈J,k∈Kj ((Xj e1/4 j gj)a(Xj e1/4 j gj)), where the family appearing in this block-cutdown formula, namely (e1/4 j fj,k)j∈J,k∈Kj is equicontinuous and contained in NL−1+1(Y). (i)-(ii) are immediate consequences of the form that aii′ takes. Since multiplication by Cb(X) preserves block structure, and using (2.1), (iii) can also be seen to be a consequence of the form that aii′ takes. (cid:3) We have seen in the previous lemma that we will need to work with "thickened" metric families, so we record the following straightforward observation. Lemma 4.7. Let X and Y be metric families, R, S ≥ 0. Assume that X R−→ Y and that R − 2S ≥ 0. Then NS(X ) R−2S−−−−→ NS(Y). Proof of Theorem 2.8 (i) ⇒ (iv). Recall that we are given an operator b on an X-module H satisfying [b, f ] = 0 for all f ∈ VL∞(X). Given ǫ > 0, our task is to produce a finite propogation operator in B which is ǫ-far from b. Lemma 3.1 provides us with Ln for every ǫn := ǫ/(2 · 8n), such that b ∈ Commut(Ln, ǫn). Let n + 1) + 2(L−1 Rn := 4(L−1 n−1 + 1) + · · · + 2(L−1 1 + 1). As X has straight finite decomposition complexity (see Definition 2.10), Rn−−→ we obtain metric families {X} = X0, X1, . . . , Xm, such that Xn−1 Xn for n ∈ {1, . . . , m} and Xm is uniformly bounded. Note that Lemma 4.7 gives us that N(L−1 n−1+1)+···+(L−1 1 +1)(Xn−1) n +1) 4(L−1 −−−−−−→ N(L−1 n +1)+···+(L−1 1 +1)(Xn). Thus, we can inductively apply Lemma 4.6, with Ln, ǫn, the operators obtained in the previous iteration, and metric families from the above display. After m steps, we will have approximated the operator b by an operator b′ which is a sum of finitely many (4m to be precise) operators in B which are block diagonal with respect to N(L−1 m +1)+···+(L−1 1 +1)(Xm). RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 19 Since Xm is uniformly bounded, so is the above family, and therefore operators which are block diagonal with respect to it have finite pro- pogation; consequently, b′ ∈ Roe(B, X). Tracing through the estimates given by Lemma 4.6, we compute that the distance from b to b′ is at most 8ǫ1 + 4 (8ǫ2 + 4 (8ǫ3 + 4 (. . . ))) = ǫ(cid:0) 1 This finishes the proof. 2 + 1 4 + 1 8 + . . .(cid:1) = ǫ. (cid:3) Remark 4.8. When the asymptotic dimension of X is at most d < ∞, the induction component of the above proof can be removed: one can use the idea of Lemma 4.5 with a decomposition of X into d + 1 (instead of 2) uniformly bounded, (4L−1 + 4)-disjoint families, and correspondingly approximate a be a sum of (d + 1)2 block diagonal operators. 5. Higson functions To prepare to prove (i) ⇔ (iii) of Theorem 2.8, we begin by consid- ering a special class of Higson functions which are more closely related to our definition of VL∞(X). Definition 5.1. Let X be a proper metric space. A function g ∈ Cb(X) is a Lipschitz -- Higson function if, for every L > 0, there exists a compact set A ⊆ X such that gX\A is L-Lipschitz. The set of all Lipschitz -- Higson functions on X is denoted Clh(X). Fix a proper metric space X and a point x0 ∈ X. For R > 0, define eR ∈ C0(X) by eR(x) := max{0, 1 − d(x, BR(x0))/R}. (5.1) Observe that eR is R−1-Lipschitz, is 1 on ¯BR(x0), and vanishes outside of B2R(x0). Lemma 5.2. Ch(X) = Clh(X) + C0(X). Proof. The inclusion ⊇ is straightforward. To go the other direction, let g ∈ Ch(X). We shall produce f ∈ Clh(X) such that f − g ∈ C0(X). Without loss of generality, g is a positive contraction. Fix a point x0 ∈ X. Recursively define R0 := 0 and Rn ≥ max{2(n+1), 2Rn−1} such that 2(n+1) . if x, y ∈ X\BRn(x0) and d(x, y) < n + 1 then g(x) − g(y) < 1 Using eR from (5.1), set g1 := eR1g, gn :=(cid:0)eRn − eRn−1(cid:1) g, n ≥ 2. 20 J ´AN SPAKULA AND AARON TIKUISIS Note that, for n ≥ 2, gn is such that if d(x, y) < n then gn(x)−gn(y) < 1 n , and that ∞ (5.2) g = gn, Xn=1 converging pointwise (as at each point, at most two terms of the sum are nonzero). Define f1 := g1. Fix n ≥ 2; we shall define a function fn which approximates gn, but is more Lipschitz. Define i = 1, . . . , n, i = 1, . . . , n, and define ci ∈ Cb(X) by Ai := g−1 n (cid:2) i n, 1(cid:3) , , 0o , ci(x) := maxn1 − d(x,Ai) Xi=1 fn := 1 n ci, n which is (1/n)-Lipschitz. Set n which is also (1/n)-Lipschitz, as it is an average of such. Moreover, kfn − gnk ≤ 1 n , and the support of fn is contained in the support of gn. n=1 fn; as in (5.2), at each point, at most two terms of fn is 2n−1 0 - Lipschitz, for all n0 ≥ 2. Moreover, f agrees with this tail sum outside of B2Rn0−1, which proves that f ∈ Clh(X). Set f := P∞ the sum are nonzero. Using this fact, one sees that P∞ Similarly, since f − g agrees with the tail P∞ of B2Rn0−1, and this tail has norm at most 2 C0(X). (fn − gn) outside n , it follows that f − g ∈ (cid:3) n=n0 n=n0 Remark 5.3. We note in passing that, due to the previous lemma, the Higson corona νX (defined as the compact Hausdorff space satisfying C(νX) ∼= Ch(X)/C0(X)) satisfies C(νX) ∼= Clh(X)/C0(X). Now we set out two constructions to be used, producing a Lipschitz -- Higson function from a very Lipschitz sequence and vice versa. Neither construction is canonical: both depend on a number of choices. Let (fk)∞ k=1 ∈ VL(X), let (fki)∞ i=0 ⊂ (0, ∞) be a sequence such that Ri+1 ≥ 6Ri for each i. From these, and using eR from (5.1), define i=1 be a subsequence, and let (Ri)∞ (5.3) gx0,(fki )∞ i=1,(Ri)∞ i=0 := fki(eRi − e3Ri−1) ∈ Cb(X). ∞ Xi=1 RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 21 Note that the supports of the functions in the summation are pair- wise disjoint, so we can treat the series as converging pointwise. It is straightforward to see that kgx0,(fki )∞ Lemma 5.4. Fix a proper metric space X and a point x0 ∈ X. For (fk)∞ i=0 ⊂ be as defined (0, ∞) such that Ri+1 ≥ 6Ri for each i, let gx0,(fki )∞ in (5.3). Then k=1 ∈ VL(X), a subsequence (fki)∞ i=1, and a sequence (Ri)∞ k ≤ k(fk)∞ i=1,(Ri)∞ i=0 i=1,(Ri)∞ i=0 k=1k. gx0,(fki )∞ i=1,(Ri)∞ i=0 ∈ Clh(X). Proof. Without loss of generality, we may assume kfkk ≤ 1 for all k. For ease of notation, we set g := gx0,(fki )∞ i=1,(Ri)∞ i=0 . As in the definition of a Lipschitz -- Higson function, let L > 0 be given. Pick i0 such that Ri0−1 > 2/L and such that fi is (L/2)-Lipschitz for all i ≥ i0. We will be done when we show that the restriction of g to X\BRi0 (x0) is L-Lipschitz. For i ≥ i0, eRi − e3Ri−1 is (L/2)-Lipschitz, so that the product hi := fki(eRi − e3Ri−1) is L-Lipschitz. For x, y ∈ X\BRi0 (x0), note that, by the definition of eR and the condition Ri+1 ≥ 6Ri, that at least one of the following conditions holds. (i) d(x, y) ≥ Ri0, or (ii) There exists i such that g(x) = hi(x) and g(y) = hi(y). In the first case, g(x) − g(y) ≤ 2kgk = 2 ≤ 3LRi0−1 ≤ 3LRi0 ≤ Ld(x, y). In the second case, since hi is L-Lipschitz, it follows that g(x) − (cid:3) g(y) ≤ Ld(x, y). Next, let g ∈ Ch(X) be given, along with a sequence (Rk) ∈ (0, ∞) such that limk→∞ Rk = ∞. From this data, define (using eR from (5.1)) (5.4) Fx0,g,(Rk)∞ Lemma 5.5. Fix a proper metric space X and a point x0 ∈ X. For a Higson function g ∈ Clh(X) and a sequence (Rk) ∈ (0, ∞) such that limk→∞ Rk = ∞, let Fx0,g,(Rk)∞ k=1 ∈ l∞(N, l∞(X)) where fk := (1 − eRk )g. be as defined in (5.4). Then := (fk)∞ k=1 k=1 Fx0,g,(Rk)∞ k=1 ∈ VL(X). Proof. Without loss of generality, we may assume kgk ≤ 1. As in the definition of VL(X), let L > 0 be given. Pick M ≥ 2/L such that g is (L/2)-Lipschitz on X\BM (x0). Pick k0 such that Rk ≥ M for all k ≥ k0. For k ≥ k0, using the fact that (1 − eRk ) is (L/2)-Lipschitz and vanishes on BM (x0), it is easy to see that fk = g(1 − eRk ) is L-Lipschitz. (cid:3) 22 J ´AN SPAKULA AND AARON TIKUISIS Proof of Theorem 2.8 (i) ⇒ (iii). By Lemma 5.2, (iii) is equivalent to [g, b] ∈ K(X, B) for all g ∈ Clh(X), which is the statement we will prove assuming (i). Assume that [b, g] 6∈ K(X, B) for some g ∈ Clh(X). Set ǫ := k[b, g]+K(X, B)k. Fix a point x0 ∈ X, let (Rk)∞ i=1 ⊂ (0, ∞) be any sequence such that limi→∞ Ri = ∞, and define (fk)∞ k=1 := Fx0,g,(Rk)∞ as in (5.4), that is, k=1 fk := (1 − eRk )g. By Lemma 5.5, (fk)∞ the condition (2.2) we obtain k=1 ∈ VL(X). Since eRk ∈ C0(X) for each k, using [fk, b] = [g, b] − [eRk g, b] ∈ [g, b] + K(X, B) and therefore by the definition of ǫ, k[fk, b]k ≥ ǫ. Consequently, limk→∞ k[fk, b]k ≥ ǫ, so that b does not commute with the image of (fk)∞ (cid:3) k=1 in VL∞(X). Before we embark on the proof of Theorem 2.8 (iii) =⇒ (i), note that as a consequence of the Spectral Theorem, we may extend the X-module structure on B(H) from bounded continuous functions to bounded Borel functions on X. This is convenient in the following proof, since it allows us to easily "cut up" operators on H using charac- teristic functions of Borel sets in X. (Of course, we cannot assume the algebra B is closed under multiplication by these bounded Borel func- tions.) We opt for this approach for the sake of readability, although it is possible to modify the proof to only use continuous functions for the price of more approximations. Proof of Theorem 2.8 (iii) ⇒ (i). Since each of VL∞(X), Ch(X), and K(X, B) is ∗-closed, it suffices to prove the theorem in the case that b is self-adjoint. We henceforth assume that b = b∗. Fix a point x0 ∈ X, and to shorten notation in this proof, set BR := BR(x0). For each R > 0, we will use χBR to denote the support projection of a function whose cozero set is BR. Assume that [b, f ] 6= 0 for some f ∈ VL∞(X). Then in fact, [b, f ] 6= 0 k=1 for which each fk is a self-adjoint contraction; we for some f = (fk)∞ fix this sequence. Let 0 < ǫ < k[b, f ]k. Consider now two cases. Case 1. There exists R0 > 0 such that for all S > 0 there are infinitely many k for which kχBR0 [b, fk](1 − χBS )k > ǫ 5 . RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 23 Roughly, for this case, we will construct some g (of the form (5.3)) such that the block-column of [b, g] corresponding to BR0 doesn't con- verge to 0 at ∞. Note that as k → ∞, fkBR0 tends towards being constant; so without is constant. Adding a loss of generality, we may assume that fkBR0 scalar to each fk, we arrive at another sequence(cid:16) fk(cid:17)∞ with the same n=1 (that is, self-adjoint and satisfying the Case 1 [b, fk] = b fk for all n, so that we have: for all S > 0 there exist infinitely properties as (fk)∞ condition), such that fkBR0 χBR0 many k such that ≡ 0. From this it follows that χBR0 n=1 kχBR0 b fk(1 − χBS )k > ǫ 5 . Using R0 as above and k0 := 0, recursively choose k1 < k2 < · · · and R1 ≥ 6R0, R2 ≥ 6R1, . . . as follows. Having chosen Ri−1, pick ki > ki−1 such that kχBR0 a fki(1 − χB6Ri−1 )k > ǫ 5 and k fkiχB6Ri−1 k ≤ ǫ 10kbk . (The second inequality can be arranged as fk converge to 0 on any given bounded subset of X.) Then, since (eR)∞ R=1 converges strongly to 1, pick Ri ≥ 6Ri−1 such that kχBR0 b fki(1 − χB6Ri−1 )eRik > ǫ 5 . Since eRi − e3Ri−1 differs from (1 − χB6Ri−1 that )eRi only on B6Ri−1, it follows kχBR0 ǫ 10 Using these recursive choices, define g := gx0,( fki )∞ b fki(eRi − e3Ri−1)k > ∈ Clh(X) ⊆ Ch(X) by Lemma 5.4. If [b, g] ∈ K(X, B), then k[b, g](1 − χBS )k → 0 as S → ∞. However, given S ≥ R0, there exists i such that 3Ri−1 > S. Then i=1,(Ri)∞ i=0 . k[b, g](1 − χBS )k ≥ kχBR0 = kχBR0 > , ǫ 10 [b, g](1 − χBS )(χB2Ri b fki(eRi − e3Ri−1)k − χB3Ri−1 )k which is a contradiction. This concludes the proof in Case 1. Case 2. For every R > 0, there exists S > 0 such that, for all but finitely many k ∈ N, kχBR[b, fk](1 − χBS )k ≤ ǫ 5 . Without loss of generality, we may assume that S > R. 24 J ´AN SPAKULA AND AARON TIKUISIS Roughly, for this case, we will construct some g (of the form (5.3)) such that the certain blocks on the diagonal of [b, g] don't converge to 0 at ∞. In preparation for this, suppose we are given R > 0 and K ∈ N. Let S be given by the Case 2 property. Then there exists k ≥ K such that kχBR[b, fk](1 − χBS )k ≤ ǫ 5 , and in addition, k[b, fk]k > ǫ and fkBS is ǫ 5-approximately constant. From the latter property, it follows that there is a scalar γ such that fkBS ≈ǫ/10 γ, so that (5.5) kχBS [b, fk]χBS k ≤ 2 · ǫ 10 kbk ≤ ǫ 5 . Since b and fk are self-adjoint, (5.6) k(1 − χBS )[b, fk]χBR)k = kχBR[b, fk](1 − χBS )k ≤ ǫ 5 . We now cut up the operator T = [b, fk] as follows: ❅ ❅ BS ❅ BR ❅ ❅ ❅ BR ❅ BS ❅ χBRT χBS χBRT (1 − χBS ) (χBS − χBR)T χBR (1 − χBS )T χBR (1 − χBR)T (1 − χBR) That is, we use the equality T = (1 − χBR)T (1 − χBR) + χBRT χBS + χBRT (1 − χBS ) + (χBS − χBR)T χBR + (1 − χBS )T χBR RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 25 and the reverse triangle inequality to deduce k(1 − χBR)[b, fk](1 − χBR)k ≥ (5.5),(5.6) ≥ > (5.7) = k[b, fk]k −(cid:0)kχBR[b, fk]χBS k + kχBR[b, fk](1 − χBS )k +k(χBS − χBR)[b, fk]χBRk + k(1 − χBS )[b, fk]χBRk(cid:1) k[b, fk]k − 4 · ǫ 5 4ǫ 5 ǫ − ǫ 5 . In summary, we have shown that for every R > 0 and K ∈ N, there exists k ≥ K such that (5.8) k(1 − χBR)[b, fk](1 − χBR)k > ǫ 5 . Now, start with R0 := 1 and k0 := 0, and (as in Case 1) we will choose k1 < k2 < · · · and R1 ≥ 6R0, R2 ≥ 6R1, . . . recursively. Given Ri−1, pick k > ki−1 satisfying (5.8) for R := 6Ri−1, and set ki equal to this k. That is, ki > ki−1 satisfies k(1 − χB6Ri−1 )[b, fki](1 − χB6Ri−1 )k > ǫ 5 . R=1 converges strongly to 1, there exists Ri ≥ 6Ri−1 Then, since (χBR)∞ such that kχBRi Note that χBRi (1 − χB6Ri−1 (1 − χB6Ri−1 )[b, fki](1 − χB6Ri−1 . ) = χBRi − χB6Ri−1 )χBRi k > ǫ 5 . Using these recursive choices, define g := gx0,(fki )∞ ∈ Clh(X) ⊆ Ch(X) by Lemma 5.4. If [b, g] ∈ K(X, B), then k[b, g](1 − χBS )k → 0 as S → ∞. However, given S > 0, there exists i such that 6Ri−1 > S. Then i=1,(Ri)∞ i=0 k[b, g] (1 − χBS )k ≥(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13) > ǫ 5 , (χBRi − χB6Ri−1 )[b, g](χBRi − χB6Ri−1 (χBRi − χB6Ri−1 )[b, fki](χBRi − χB6Ri−1 )(cid:13)(cid:13)(cid:13) )(cid:13)(cid:13)(cid:13) again a contradiction. This concludes the proof. (cid:3) 6. More about VL∞ (X) 6.1. To what extent does VL∞ (X) determine X? If X is a metric space, then we say that a set E ⊆ X × X is uniformly bounded (also called a metric entourage) if there exists R > 0 such that E ⊆ {(x, y) ∈ X × X : d(x, y) ≤ R}. 26 J ´AN SPAKULA AND AARON TIKUISIS Definition 6.1. Let X and Y be metric spaces. We say that a (not necessarily continuous) function φ : X → Y is: • bornologous, if for every R ≥ 0 there exists S ≥ 0, such that for all x, y ∈ X, d(x, y) ≤ R implies d(φ(x), φ(y)) ≤ S; • cobounded, if f −1(y) is bounded for every y ∈ Y ; • coarse, if it is both cobounded and bornologous; • a coarse equivalence, if it is bornologous, and there exists a bornologous ψ : Y → X, such that both ψ ◦ φ and φ ◦ ψ are uniformly close to the identity maps, i.e., their graphs are uni- formly bounded subsets of X × X and Y × Y respectively. Note that in this case both φ and ψ are automatically coarse; • locally Lipschitz, if there exist δ > 0 and T ≥ 0, such that d(x, y) ≤ δ implies d(f (x), f (y)) ≤ T d(x, y), x, y ∈ X. • a Lip-coarse equivalence, if it is a coarse equivalence, it is lo- cally Lipschitz, and in the definition of coarse equivalence, ψ can be chosen to be locally Lipschitz as well. Note that traditionally, coarse geometry does not concern itself with local behaviour. However, as our main tool in this piece are Lipschitz functions, we will insist that the maps involved are locally Lipschitz. On the other hand, in the key setting in which the metric spaces in- volved are uniformly discrete, this requirement is automatic, and thus can be ignored. Proposition 6.2. Let X be a proper metric space. VL∞ (X) is a Lip-coarse invariant for X. More precisely, if X and Y are proper metric spaces which are Lip-coarsely equivalent via a (locally Lipschitz) map φ : X → Y , then composition by φ induces a ∗-isomorphism VL∞ (Y ) → VL∞ (X). Proof. Suppose that X, Y are Lip-coarsely equivalent, so that there are locally Lipschitz coarse maps φ : X → Y and ψ : Y → X such that the graphs of φ ◦ ψ and ψ ◦ φ are uniformly bounded. Denote by δ > 0 and T ≥ 0 the constants of local Lipschitzness of φ. As locally Lips- chitz maps are continuous, φ and ψ induce maps φ∗ : l∞ (N, Cb(Y )) → l∞ (N, Cb(X)) and ψ∗ : l∞ (N, Cb(X)) → l∞ (N, Cb(Y )). First we show that φ∗ (VL (Y )) ⊆ VL (X). Surely, let (fn)∞ n=1 ∈ VL (Y ) with k(fn)∞ n=1k ≤ 1. Let L > 0. Since φ is a coarse map, there exists S > 0 such that if x, y ∈ X satisfy d(x, y) ≤ 2/L then d(φ(x), φ(y)) < S. Without loss of generality, we can assume that S > δT . Since (fn)∞ n=1 ∈ VL(Y ), there exists n0 such that fn is (Lδ/S)- Lipschitz for all n ≥ n0. For n ≥ n0, let us show that fn ◦ φ is L-Lipschitz. Let x, y ∈ X, there are three cases: • If d(x, y) > 2/L then fn(x) − fn(y) ≤ 2kfnk < Ld(x, y). RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 27 • If δ ≤ d(x, y) ≤ 2/L, then d(φ(x), φ(y)) < S. Since fn is (Lδ/S)-Lipschitz, kfn ◦ φ(x) − fn ◦ φ(y)k ≤ Lδ ≤ Ld(x, y). • If d(x, y) < δ, then kfn(φ(x)) − fn(φ(y))k ≤ Lδ S d(φ(x), φ(y)) ≤ LδT S d(x, y) < Ld(x, y). Thus, φ∗ induces a map θφ : VL∞ (Y ) → VL∞ (X); likewise, ψ∗ induces a map θψ : VL∞ (X) → VL∞ (Y ). Let us show that these maps are inverses. By the symmetry of their definition, it suffices to show that θψ ◦ θφ = idVL∞(Y ). Let (fn)∞ n=1 ∈ VL∞ (Y ). Let Γ (φ ◦ ψ) ⊂ Y × Y denote the graph of φ ◦ ψ. Since this is uniformly bounded, given ǫ > 0, we may find n0 such that, for n ≥ n0 and (x, y) ∈ Γ (φ ◦ ψ), fn (x) − fn (y) ≤ ǫ. In other words, for n ≥ n0, fn (x) − fn (φ (ψ (x))) ≤ ǫ, and thus, kfn − ψ∗ ◦ φ∗ (fn) k ≤ ǫ. Consequently, k(fn)∞ trary, θψ ◦ θφ ((fn)∞ n=1 − θψ ◦ θφ ((fn)∞ n=1) = (fn)∞ n=1 in VL∞(Y ). n=1)kVL∞(Y ) ≤ ǫ. Since ǫ is arbi- (cid:3) Lemma 6.3. Let X be a metric space, and let E ⊆ X × X. The following are equivalent: (i) E is uniformly bounded. (ii) For every (fn)∞ n=1 ∈ VL (X) and ǫ > 0, there exists n0 such that, if n ≥ n0 and (x, y) ∈ E then fn (x) − fn (y) < ǫ. (iii) For every (fn)∞ n=1 ∈ VL (X), there exists n1 such that, if (x, y) ∈ E then fn1 (x) − fn1 (y) < 1. Proof. (i) ⇒ (ii) is by the definition of VL (X). (ii) ⇒ (iii) is trivial. For (iii) ⇒ (i), let E ⊆ X × X be a set that isn't uniformly bounded, n=1 ∈ VL (X) such that, for all n and let us show that there exists (fn)∞ there exists (xn, yn)∞ n=1 ∈ E such that fn (x) − fn (y) = 1. For each n, there exists (xn, yn) ∈ E such that d(x, y) > n. Thus there exists a (1/n)-Lipschitz function fn : X → R such that fn(xn) = 0, and fn(yn) = 1. It follows that (fn)∞ n=1 ∈ VL (X), showing that (iii) doesn't hold. (cid:3) Proposition 6.4. Let X, Y be metric spaces. Let φ : Y → X be a function. (i) If φ∗ (VL (X)) ⊆ VL (Y ) then φ is a bornologous map; (ii) If, moreover, the induced map θφ : VL∞ (X) → VL∞ (Y ) is an isomorphism, then φ is a coarse equivalence. 28 J ´AN SPAKULA AND AARON TIKUISIS Proof. (i): Let R > 0. We must show that E := {(φ(x), φ(y)) : x, y ∈ X, d(x, y) < R} is uniformly bounded. To this end, we will verify Lemma 6.3 (iii) for this set. Therefore, let (fn)∞ n=1 ∈ VL (X). Since φ∗ ((fn)∞ n=1) ∈ VL (Y ), there exists n1 such that fn1 ◦ φ is R−1-Lipschitz. Thus for (φ(x), φ(y)) ∈ E, i.e., d(x, y) < R, fn1(φ(x)) − fn1(φ(x)) = fn1 ◦ φ(x) − fn1 ◦ φ(y) < 1, as required. (ii): We must show two things: (a) for every R > 0, the set {(x, y) ∈ Y × Y : d(φ(x), φ(y)) < R} is uniformly bounded, and (b) there ex- ists R > 0 such that, for all x ∈ X, there exists y ∈ Y such that d (x, φ (y)) < R. (a): Let R ≥ 0 be given. We will verify Lemma 6.3 (iii) for E := n=1 ∈ VL (Y ). {(x, y) ∈ Y × Y : d(φ(x), φ(y)) < R}. Therefore, let (fn)∞ Since θφ is surjective, there exists (gn)∞ n=1 ∈ VL (X) such that lim∞ kfn− gn◦φk = 0. By Lemma 6.3 (i) ⇒ (ii), there exists n0 such that, if n ≥ n0 and w, z ∈ X satisfy d(w, z) < R then gn(w) ≈1/3 gn(z). Pick n ≥ n0 such that kfn − gn ◦ φk < 1 3 . Now, let (x, y) ∈ E, i.e., d(φ(x), φ(y)) < R. Then fn (x) ≈1/3 gn (φ (x)) ≈1/3 gn (φ (y)) ≈1/3 fn (y) , as required. (b): Proof by contradiction. Suppose for a contradiction that, for every n there exists xn ∈ X such that for all y ∈ Y , (xn, φ (y)) ≥ n. Thus, there exists a (1/n)-Lipschitz function fn : X → [0, 1] such that fn(xn) = 1 and fn(φ(y)) = 0 for all y ∈ Y . Putting these together, we obtain (fn)∞ n=1 ∈ VL (X) and kfnk = 1 (since fn (xn) = 1), so that k(fn)∞ n=1k = 1 in VL∞ (X). However, since fn (φ (y)) = 0 for all y ∈ Y , it follows that θφ ((fn)∞ n=1) = 0, which contradicts injectivity of θφ. (cid:3) In other words, when VL∞ (X) ∼= VL∞ (Y ), and the isomorphism comes from a map between Y and X, it follows that X and Y are coarsely equivalent. Here is the more interesting question: Question 6.5. Let X, Y be uniformly discrete metric spaces. If VL∞ (X) ∼= VL∞ (Y ), must X and Y be coarsely equivalent? 6.2. The nuclear dimension of VL∞ (X). VL (X) and VL∞ (X) are commutative unital C-algebras, and therefore by Gelfand's Theorem, each are algebras of continuous functions on a compact Hausdorff space, namely the Gelfand spectrum of the respective algebras. As these C- algebras are nonseparable, their spectra are nonmetrizable. Here we show a relationship between the asymptotic dimension of X and the RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 29 covering dimension (suitably interpreted) of these spectra. In fact we use the nuclear dimension of the algebras, which (for the spectra) cor- responds to a version of covering dimension which is slightly modified (in this nonseparable case) from the original definition. Definition 6.6. [28, Definition 2.1] Let A be a C-algebra and let d ∈ N. We say that the nuclear dimension of A is at most d if there exists a net (Fλ, ψλ, φλ) where Fλ is a finite dimensional C-algebra, ψλ : A → Fλ and φλ : Fλ → A are completely positive maps such that lim λ ψλ(φλ(a)) = a, a ∈ A, ψλ is contractive, and Fλ decomposes into direct summands as Fλ = λ ⊕ · · ·⊕ F (d) F (0) is contractive and order zero, for each i. such that φλF (i) λ λ Since we will be considering nuclear dimension for the commutative and nonseparable C-algebras VL(X) and VL∞(X), let us explain ex- actly the modification to covering dimension that is entailed by nuclear dimension. Let Y be a locally compact Hausdorff space; call an open set U ⊆ Y a preimage-open set if it is the preimage of an open subset of R under a continuous function Y → R. Then the nuclear dimension of C0(Y ) is equal to the smallest number d such that every finite cover of Y consisting of preimage-open sets has a (d + 1)-colourable refinement consisting of preimage-open sets (see the proof of [16, Proposition 3.3], and [28, Proposition 2.4]; this fact is alluded to in the discussion be- fore [28, Proposition 2.4]). (In the second countable situation, or more generally when Y is a normal space, all open sets are preimage-open, which is why nuclear dimension coincides with the usual definition of covering dimension in this case.) Definition 6.7. ([10, §1.E]) Let X be a metric space. Then the as- ymptotic dimension of X is at most d ∈ N, written asdim(X) ≤ d, if for every R > 0, there exists a cover of X of the form (U (i) j )i=0,...,d; j∈J , such that for each i = 0, . . . , d, the family (U (i) j )j∈J is R-disjoint and uniformly bounded. Proposition 6.8. Let X be a metric space. dimnuc VL (X) ≤ asdim(X) and dimnuc VL∞ (X) ≤ asdim(X). Proof. As the nuclear dimension decreases when passing to quotients, it suffices to prove the first statement. Set d := asdim(X). Let F ⊂ VL(X) be a finite set and let ǫ > 0 be given. Using the definition of asymptotic dimension in a fairly straightfor- ward way, for each n ∈ N, we may find an infinite partition of unity (e(i) j (n))j∈J(n); i=0,...,d, such that: (i) for each i, (e(i) j (n))j∈J(n) are pairwise orthogonal, 30 J ´AN SPAKULA AND AARON TIKUISIS (ii) each e(i) j (iii) there is a uniform bound, S(n), on the diameters of the sup- is (1/n)-Lipschitz, and ports of e(i) j (n) (allowed to depend only on n). Let us also pick a point x(i) i, j, and n. For n = 0, define J(0) := X, e(0) for i > 0, S(0) := 0, and x(i) j (0) = j ∈ X. j (n) inside the support of e(i) j (0)(x) := δj,x, e(i) j (n), for each j (0) := 0 Define a ∗-homomorphism ψn = (ψ(0) n , . . . , ψ(d) coordinatewise by evaluation at x(i) homomorphism φ(i) n : l∞(J(n)) → Cb(X) by j (n). For i = 0, . . . , n, define a ∗- n ) : Cb(X) →Ld i=0 l∞(J(n)) φ(i) n (λ) = Xj∈J(n) λ(j)e(i) j (n), with the sum converging pointwise, since in fact at each point x ∈ X, at most one summand is nonzero (by condition (i)). Note that if f ∈ Cb(X) is (ǫ/S(n))-Lipschitz then (6.1) Let φ(i) n ◦ ψ(i) n (f ) ≈ǫ f. d Xi=0 F = {(fi,k)∞ k=1 : i = 1, . . . , m}. For each k, we may find some nk ≥ 0 such that fi,k has (ǫ/S(nk))- Lipschitz for all i = 1, . . . , m. Since F ⊆ VL(X), we can pick these nk such that they converge to ∞. We now define a ∗-homomorphism Ψ := (ψnk )∞ k=1 : l∞(N, l∞(X)) →Yk d Mi=0 l∞(J(nk)) ∼= d Mi=0 Yk l∞(J(nk)); we may write Ψ = (Ψ(0), . . . , Ψ(d)). For i = 0, . . . , d, define a ∗- homomorphism Φ(i) := (φ(i) nk)∞ k=1 :Yk l∞(J(nk)) → l∞(N, l∞(X)). Since nk → ∞, we see that the image of Φ(i) is in fact contained in VL(X). Moreover, by (6.1) and our choice of nk, we find that d Φ(i) ◦ Ψ(i)(f ) ≈ǫ f Xi=0 for f ∈ F . Since Qk l∞(J(nk)) has nuclear dimension zero, this is sufficient to prove that VL(X) has nuclear dimension at most d. (cid:3) RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 31 We have an argument to get inequalities in the other direction, under the hypothesis that X has finite asymptotic dimension. For this, we begin with the following lemma. Lemma 6.9. Let X be a set and let η > 0. Let f1, . . . , fm, e1, . . . , en : X → [0, ∞), and λi,j ∈ [0, ∞) for i = 1, . . . , n, j = 1, . . . , m. Suppose j=1 λi,j = 1 for each i. Then that fj ≈η Pn i=1 λi,jei for each j and Pm for each i, there exists j(i) such that {x : ei(x) > mη} ⊆ {x : fj(i)(x) > 0}. λi,j ≥ 1/m. For x ∈ X such that fj(x) = 0, it follows that j=1 λi,j = 1, there exists some j = j(i) such that Proof. Fix i. Since Pm 1 m ei(x) ≤ λi,jei(x) n ≤ λi′,jei′(x) Xi′=1 ≤ fj(x) + η = η. This shows that ei(x) ≤ mη, as required. (cid:3) Theorem 6.10. If X has finite asymptotic dimension, then asdim(X) = dimnuc VL(X) = dimnuc VL∞(X). Proof. Set d := dimnuc VL∞(X). By Proposition 6.8, it suffices to show that asdim(X) ≤ d. Let R > 0 be given, and we will partition X into (d + 1) uniformly bounded, R-disjoint families. By hypothesis, let asdim(X) ≤ m − 1. Then from this, there exists a partition of unity g1, . . . , gm ∈ VL∞(X), such that gj = (gj,l)∞ l=1 where for each j, l, the support of gj,l decomposes as an l-disjoint, uniformly bounded family of subsets of X. Set η := 1 3(d + 1)m . The only nonzero order zero maps from a matrix algebra into a com- mutative algebra occur when the matrix algebra is one-dimensional (this follows from [27, Proposition 3.2 (a)]). Hence, dimnuc VL(X) ≤ d implies that there exists s ∈ N, a c.p.c. map ψ = (ψ(0), . . . , ψ(d)) : C⊕s, and c.p.c. order zero maps φ(i) : Cs → VL∞(X) i=0 such that VL∞(X) →Ld Xi=0 gj ≈η m φ(i) ◦ ψ(i)(gj) and 1VL∞(X) ≈1/2 φ(i) ◦ ψ(i)(1). m Xi=0 32 J ´AN SPAKULA AND AARON TIKUISIS By rescaling, we may assume that ψ(i)(1VL∞(X)) = (1, . . . , 1). Write ψ(i)(gj) = (λ(i,1),j, . . . , λ(i,s),j) ∈ [0, ∞)s (since ψ is positive). By linear- ity, looking at the i′ component of ψ(i)(1), we have m λ(i,i′),j = 1. Xj=1 Each map φ(i) lifts to a c.p.c. order zero map (φ(i) l )∞ l=1 : Cs → VL(X), by [16, Remark 2.4]. For all but finitely many l, we have gj,l ≈η φ(i) l (λ(i,1),j, . . . , λ(i,s),j) and (6.2) 1Cb(X) ≈1/2 φ(i) l (1, . . . , 1). d Xi=0 Xi=0 d Xi=0 Xi′=1 Fix l ≥ R for which (6.2) holds, and such that the image of each mini- is (mη/L)-Lipschitz. Write e(i,1), . . . , e(i,s) l of the minimal projections in Cs, and write mal projection in Cs under φ(i) l for these images under φ(i) fj := gj,l; thus (6.2) becomes (6.3) fj ≈η λ(i,i′),je(i,i′) and 1 ≈1/2 e(i,i′). d s d s Xi=0 Xi′=1 We now apply Lemma 6.9 with (i, i′), i = 0, . . . , d, i′ = 1, . . . , s in place of the index i = 1, . . . , n. This tells us that for each i = 0, . . . , d and i′ = 1, . . . , s, there exists some j(i, i′) such that Bi,i′ := {x ∈ X : e(i,i′) > mη} ⊆ {x ∈ X : fj(i,i′)(x) > 0}. Since the support of fj(i,i′) (= gj(i,i′),l) decomposes as a union of an l-disjoint uniformly bounded family, and l ≥ R, we can partition Bi,i′ into an R-disjoint, uniformly bounded family, say A(i) i′,t Bi,i′ =at∈T Fixing i ∈ {0, . . . , d}, we now consider the family (A(i) i′,t)i′=1,...,s, t∈T . This family is a finite union of uniformly bounded families, whence it is uniformly bounded. Let us check that it is R-disjoint. Since for fixed i′ we already have R-disjointness of (A(i) i′,t)t∈T , we need to show that for i′ 1 6= i′ is at least R. In other words, we need to show that the minimal distance between Bi,i′ 2 and t1, t2 ∈ T , the minimal distance between A(i) i′ 1,t1 and A(i) i′ 2,t2 is at least R. and Bi,i′ We have that ei,i′ 1(x) 6= 0, it must be the case that ei,i′ and ei,i′ ei,i′ Bi,i′ that d(x, y) ≥ R, as required. , so that ei,i′ (y) > mη. Since ei,i′ 2 2 1 2 2 are orthogonal, so if x ∈ Bi,i′ then since 2(x) = 0. Consider now y ∈ is (mη/R)-Lipschitz, it follows 1 1 2 RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 33 Finally let us show that (A(i) i′,t)i=0,...,d, i′=1,...,s, t∈T is a cover of X, i.e., that X = Si,i′ Bi,i′. For x ∈ X, from the second part of (6.3), we have d s 1/2 < e(i,i′)(x). Xi=0 Xi′=1 At most d+1 terms in this sum are nonzero, due to the pairwise orthog- onality withing each family (e(i,i′))s i′=1. Therefore, there exists some i, i′ such that e(i,i′)(x) > 1 2(d+1) = mη. Thus, x ∈ Bi,i′ (by definition) as required. (cid:3) We have the following consequence. Corollary 6.11. Suppose X is a metric space, and for all m ∈ N, X contains a subspace Ym such that asdim(Ym) ∈ [m, ∞). Then dimnuc VL(X) = dimnuc VL∞(X) = ∞. Proof. It is not too hard to see that restriction to Ym produces a surjec- tive ∗-homomorphism VL∞(X) → VL∞(Ym) (for surjectivity, the key point is that an L-Lipschitz function on a closed subset of Ym extends to an L-Lipschitz function on X). Hence we have dimnuc VL∞(X) ≥ dimnuc VL∞(Ym) ≥ m, using [28, Proposition 2.3(iv)] for the first inequality and Theorem 6.10 for the second. (cid:3) In [2, Theorem 7.2], it was shown that the asymptotic dimension of X is equal to the covering dimension of the Higson corona νX, likewise provided that asdim(X) < ∞. Question 6.12. Is dimnucVL(X) = asdim(X) always? Is dimnucVL(X) = dim(νX) always? In light of Corollary 6.11 and [2, Theorem 7.2], the above question is only open in the case of a metric space X of infinite dimension, which does not contain subspaces of arbitrarily large finite dimension. References [1] G. Bell and A. Dranishnikov. Asymptotic dimension. Topology Appl., 155(12):1265 -- 1296, 2008. [2] A. N. Dranishnikov. Asymptotic topology. Uspekhi Mat. Nauk, 55(6(336)):71 -- 116, 2000. [3] Alexander Dranishnikov. On Gromov's positive scalar curvature conjecture for duality groups. J. Topol. Anal., 6(3):397 -- 419, 2014. 34 J ´AN SPAKULA AND AARON TIKUISIS [4] Alexander Dranishnikov and Michael Zarichnyi. Asymptotic dimension, de- composition complexity, and Haver's property C. Topology Appl., 169:99 -- 107, 2014. [5] Alexander Engel. Index theory of uniform pseudodifferential operators. arXiv preprint 1502.00494v1. [6] Alexander Engel. Rough index theory on spaces of polynomial growth and contractibility. arXiv preprint 1505.03988v3. [7] Vladimir Georgescu. On the structure of the essential spectrum of elliptic op- erators on metric spaces. J. Funct. Anal., 260(6):1734 -- 1765, 2011. [8] Vladimir Georgescu. On the essential spectrum of the operators in certain crossed products, 2017. arXiv preprint 1705.00379. [9] Vladimir Georgescu and Andrei Iftimovici. Localizations at infinity and essen- tial spectrum of quantum Hamiltonians. I. General theory. Rev. Math. Phys., 18(4):417 -- 483, 2006. [10] M. Gromov. Asymptotic invariants of infinite groups. In Geometric group the- ory, Vol. 2 (Sussex, 1991), volume 182 of London Math. Soc. Lecture Note Ser., pages 1 -- 295. Cambridge Univ. Press, Cambridge, 1993. [11] Mikhael Gromov. Groups of polynomial growth and expanding maps. Inst. Hautes ´Etudes Sci. Publ. Math., (53):53 -- 73, 1981. [12] Erik Guentner and Jerome Kaminker. Exactness and the Novikov conjecture. Topology, 41(2):411 -- 418, 2002. [13] Erik Guentner, Romain Tessera, and Guoliang Yu. A notion of geometric com- plexity and its application to topological rigidity. Invent. Math., 189(2):315 -- 357, 2012. [14] Erik Guentner, Romain Tessera, and Guoliang Yu. Discrete groups with finite decomposition complexity. Groups Geom. Dyn., 7(2):377 -- 402, 2013. [15] Gennadi Kasparov and Guoliang Yu. The coarse geometric Novikov conjecture and uniform convexity. Adv. Math., 206(1):1 -- 56, 2006. [16] Eberhard Kirchberg and Wilhelm Winter. Covering dimension and quasidiag- onality. Internat. J. Math., 15(1):63 -- 85, 2004. [17] B. V. Lange and V. S. Rabinovich. Noethericity of multidimensional discrete convolution operators. Mat. Zametki, 37(3):407 -- 421, 462, 1985. [18] Narutaka Ozawa. Amenable actions and exactness for discrete groups. C. R. Acad. Sci. Paris S´er. I Math., 330(8):691 -- 695, 2000. [19] V. S. Rabinovich, S. Roch, and B. Silbermann. Fredholm theory and finite sec- tion method for band-dominated operators. Integral Equations Operator The- ory, 30(4):452 -- 495, 1998. Dedicated to the memory of Mark Grigorievich Krein (1907 -- 1989). [20] John Roe. Personal communication. [21] John Roe. An index theorem on open manifolds. I. J. Differential Geom., 27(1):87 -- 113, 1988. [22] John Roe. Index theory, coarse geometry, and topology of manifolds, volume 90 of CBMS Regional Conference Series in Mathematics. Published for the Con- ference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1996. [23] John Roe. Lectures on coarse geometry, volume 31 of University Lecture Series. American Mathematical Society, Providence, RI, 2003. [24] Thomas Schick. The topology of positive scalar curvature. In Proceedings of the International Congress of Mathematicians Seoul 2014, volume II, pages 1285 -- 1308. Kyung Moon SA Co. Ltd., 2014. [25] G. Skandalis, J. L. Tu, and G. Yu. The coarse Baum-Connes conjecture and groupoids. Topology, 41(4):807 -- 834, 2002. RELATIVE COMMUTANT PICTURES OF ROE ALGEBRAS 35 [26] J´an Spakula and Rufus Willett. On rigidity of Roe algebras. Adv. Math., 249:289 -- 310, 2013. [27] Wilhelm Winter. Covering dimension for nuclear C-algebras. J. Funct. Anal., 199(2):535 -- 556, 2003. [28] Wilhelm Winter and Joachim Zacharias. The nuclear dimension of C-algebras. Adv. Math., 224(2):461 -- 498, 2010. [29] Nick Wright. C0 coarse geometry and scalar curvature. J. Funct. Anal., 197(2):469 -- 488, 2003. [30] Guoliang Yu. Zero-in-the-spectrum conjecture, positive scalar curvature and asymptotic dimension. Invent. Math., 127(1):99 -- 126, 1997. [31] Guoliang Yu. The Novikov conjecture for groups with finite asymptotic dimen- sion. Ann. of Math. (2), 147(2):325 -- 355, 1998. [32] Guoliang Yu. The coarse Baum-Connes conjecture for spaces which admit a uniform embedding into Hilbert space. Invent. Math., 139(1):201 -- 240, 2000. J´an Spakula, Mathematical Sciences, University of Southampton, Highfield, Southampton, SO17 1BJ, United Kingdom E-mail address: [email protected] Aaron Tikuisis, Department of Mathematics and Statistics, Univer- sity of Ottawa, Ottawa, Ontario, K1N 6N5, Canada E-mail address: [email protected]
1512.03347	1	1512	2015-12-10T17:59:23	Cohomology of Jordan triples via Lie algebras	[ "math.OA", "math.RA" ]	We develop a cohomology theory for Jordan triples, including the infinite dimensional ones, by means of the cohomology of TKK Lie algebras. This enables us to apply Lie cohomological results to the setting of Jordan triples. Some preliminary results for von Neumann algebras are obtained.	math.OA	math	Cohomology of Jordan triples via Lie algebras Cho-Ho Chu and Bernard Russo Abstract. We develop a cohomology theory for Jordan triples, including the infinite dimensional ones, by means of the cohomology of TKK Lie algebras. This enables us to apply Lie cohomological results to the setting of Jordan triples. Some preliminary results for von Neumann algebras are obtained. . A O h t a m [ 1 v 7 4 3 3 0 . 2 1 5 1 : v i X r a Contents Introduction 1. 2. Jordan triples and TKK Lie algebras 3. Cohomology of Lie algebras with involution 4. Cohomology of Jordan triples 5. Examples 6. Proofs of Theorems 2.3 and 5.5 References 1 3 6 7 13 16 26 1. Introduction A veritable army of researchers took the theory of derivations of operator algebras to dizzying heights -- producing a theory of cohomology of operator algebras as well as much information about automorphisms of operator algebras -- Richard Kadison [19] In addition to associative algebras, cohomology groups are defined for Lie algebras and, to some extent, for Jordan algebras. Since the structures of Jordan derivations and Lie derivations on von Neumann algebras are well understood, and in view of the above quotation, isn't it time to study the higher dimensional non associative cohomology of a von Neumann algebra? The present paper is motivated by this rhetorical question. In this paper we develop a cohomology theory for Jordan triples, including the infinite dimensional ones, by means of the cohomology of TKK Lie algebras. This enables us to apply Lie cohomological results to the setting of Jordan triples. Several references, which will be mentioned below, use Lie theory as a tool to study Jordan cohomology. The outline of the paper is the following. In the rest of this introduction, we give an overview of various cohomology theories, both classical and otherwise. (For a more detailed survey see [29].) In section 2, the definitions of Jordan triple module and Lie algebra module, as well as the Tits-Kantor- Koecher (TKK) construction are reviewed, basically following [5]. It is shown in Theorem 2.3 that 1991 Mathematics Subject Classification. Primary 17C65, 18G60; Secondary 46L70, 16W10. Key words and phrases. Jordan triple, cohomology, TKK algebra, derivation, cocycle, structural transformation, von Neumann algebra. 1 2 CHO-HO CHU AND BERNARD RUSSO a Jordan triple module gives rise to a Lie module for the corresponding TKK algebra. The proof of Theorem 2.3 is deferred to subsection 6.1. After reviewing the cohomology of Lie algebras (with or without an involution) in section 3, two infinite families of cohomology groups are defined for a Jordan triple system V in section 4, one using the Lie cohomology of the TKK algebra of V and the other using the Lie cohomology of the TKK algebra with its canonical involution θ. A complete analysis is given for the first cohomology groups in Proposition 4.6, which shows that structural transformations on V correspond to derivations of the TKK Lie algebra, and triple derivations on V correspond to the θ-invariant derivations. Section 5 contains examples of Jordan cocycles and TKK algebras, and applications, including a characterization of certain 3-cocycles in Theorem 5.5, the proof of which appears in subsection 6.2. The applications to von Neumann algebras appear in Theorem 4.7 and Corollary 5.7. 1.1. Brief survey of cohomology theories. The starting point for the cohomology theory of associative algebras is the paper of Hochschild from 1945 [12]. The standard reference of the theory is [3]. Two other useful references are due to Weibel ([35],[36]). Shortly after the introduction of cohomology for associative algebras, there appeared in [4] a cor- responding theory for Lie algebras. We follow [16] for the definitions and initial results. Applications can be found in [7] and [20]. The cohomology theory for Jordan algebras is less well developed than for associative and Lie algebras. A starting point would seem to be the papers of Gerstenhaber in 1964 [8] and Glassman in 1970 [10], which concern arbitrary nonassociative algebras. A study focussed primarily on Jordan algebras is [9]. We next recall two fundamental results, namely, the Jordan analogs of the first and second Whitehead lemmas as described in [15]. Theorem 1.1 (Jordan analog of first Whitehead lemma [14]). Let J be a finite dimensional semisimple Jordan algebra over a field of characteristic 0 and let M be a J -module. Let f be a linear mapping of J into M such that f (ab) = f (a)b + af (b). Then there exist vi ∈ M, bi ∈ J such that f (a) =Xi ((via)b − vi(abi)). Theorem 1.2 (Jordan analog of second Whitehead lemma [27]). Let J be a finite dimensional separable1 Jordan algebra and let M be a J -module. Let f be a bilinear mapping of J × J into M such that and f (a2, ab) + f (a, b)a2 + f (a, a)ab = f (a2b, a) + f (a2, b)a + (f (a, a)b)a Then there exist a linear mapping g from J into M such that f (a, b) = f (b, a) f (a, b) = g(ab) − g(b)a − g(a)b Two proofs of Theorem 1.2 are given in [17]. One of them, which uses the classification of finite dimensional Jordan algebras, is outlined in [29, 4.3.1]. The other proof uses Lie algebras and is contained in [17, pp. 324 -- 336]. A study of low dimensional cohomology for quadratic Jordan algebras is given in [24]. Since quadratic Jordan algebras (which coincide with "linear" Jordan algebras over characteristic 0 fields) can be considered a bridge from Jordan algebras to Jordan triple systems, this would seem to be a good place to look for exploring cohomology theory for Jordan triples. Indeed, this is hinted at in 1Separable, in this context, means that the algebra remains semisimple with respect to all extensions of the ground field. For algebraically closed fields, this is the same as being semisimple COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 3 [25], since although [24] is about Jordan algebras, the concepts are phrased in terms of the associated triple product {abc} = (ab)c + (cb)a − (ac)b. However, both papers stop short of defining higher dimensional cohomology groups. The paper [24], which is mostly concerned with representation theory, proves, for the only cohomology groups defined, the linearity of the functor H n: H n(J, ⊕iMi) = ⊕iH n(J, Mi), n = 1, 2. The paper [25], which is mostly concerned with compatibility of tripotents in Jordan triple systems, proves versions of the linearity of the functor H n, n = 1, 2, corresponding to the Jordan triple structure. The earliest work on cohomology of triple systems seems to be [11] (Lie triple systems), which is discussed in section 3. Four decades later, the second paper on the cohomology of Lie triple systems appeared [13]. The following is from the review [32] of [1] (associative triple systems). "A cohomology for associative triple systems is defined, with the main purpose to get quickly the cohomological triviality of finite-dimensional separable objects over fields of characteristic 6= 2, i.e., in particular the Whitehead lemmas and the Wedderburn principal theorem." The authors of the present paper know of only two other references dealing with the Wedderburn principal theorem in the context of triple systems, namely, [2] (alternative triple systems) and [23] (Jordan triple systems). In the latter paper, the well-known Koecher-Tits-construction of a Lie algebra from a Jordan algebra is generalized to Jordan pairs. The radical of this Lie algebra is calculated in terms of the given Jordan pair and a Wedderburn decomposition theorem for Jordan pairs (and triples) in the characteristic zero case is proved. Finally, we mention that a more general approach to cohomology of algebras and triple systems appears in the paper of Seibt [31]. 2. Jordan triples and TKK Lie algebras By a Jordan triple, we mean a real or complex vector space V , equipped with a Jordan triple product {·, ·, ·} : V 3 → V which is linear and symmetric in the outer variables, conjugate linear in the middle variable, and satisfies the Jordan triple identity {x, y, {a, b, c}} = {{x, y, a}, b, c} − {a, {y, x, b}, c} + {a, b, {x, y, c}} for a, b, c, x, y ∈ V . Given two elements a, b in a Jordan triple V , we define the box operator a b : V → V by a b(·) = {a, b, ·}. All Lie algebras in this paper are real or complex. We construct a cohomology theory of Jordan triples using the Tits-Kantor-Koecher (TKK) Lie algebras associated with them. Although we could develop the theory for all Jordan triples, we focus on the nondegenerate ones, which will be assumed throughout, to avoid unnecessary complication. For degenerate Jordan triples, the construction is exactly the same albeit more computation is involved. A Jordan triple is called nondegenerate if for each a ∈ V , the condition {a, a, a} = 0 implies a = 0. Given that V is nondegenerate, one has dk ck (aj, bj, ck, dk ∈ V ) Xj aj bj =Xk ck dk ⇒ Xk bj aj =Xj which facilitates a simple definition of the TKK Lie algebra L(V ) of V , with an invoultion θ (cf. [5, p.45]), where V0 = {Pj aj (2.1) L(V ) = V ⊕ V0 ⊕ V, bj : aj, bj ∈ V }, the Lie product is defined by [(x, h, y), (u, k, v)] = (hu − kx, [h, k] + x v − u y, k♮y − h♮v), 4 CHO-HO CHU AND BERNARD RUSSO and for each h =Pi ai bi in the Lie subalgebra V0 of L(V ), the map h♮ : V → V is well defined by h♮ =Xi bi ai. The involution θ : L(V ) → L(V ) is given by θ(x, h, y) = (y, −h♮, x) ((x, h, y) ∈ L(V )). Identifying V with the subspace {(x, 0, 0) : x ∈ V } of L(V ), we have the following relationship between the triple and Lie products: {a, b, c} = [ [a, θ(b)], c] (a, b, c ∈ V ). If no confusion is likely, we often simplify the notation {a, b, c} to {abc}. Given a Lie algebra L and a module X over L, we denote the action of L on X by so that (ℓ, x) ∈ L × X 7→ ℓ.x ∈ X [ℓ, ℓ′].x = ℓ′.(ℓ.x) − ℓ.(ℓ′.x). Definition 2.1. Let V be a Jordan triple. A vector space M over the same scalar field is called a Jordan triple V -module (cf. [29]) if it is equipped with three mappings {·, ·, ·}1 : M × V × V → M, {·, ·, ·}2 : V × M × V → M, {·, ·, ·}3 : V × V × M → M such that (i) {a, b, c}1 = {c, b, a}3; (ii) {·, ·, ·}1 is linear in the first two variables and conjugate linear in the last variable, {·, ·, ·}2 is conjugate linear in all variables; (iii) denoting by {·, ·, ·} any of the products {·, ·, ·}j (j = 1, 2, 3), the identity {a, b, {c, d, e}} = {{a, b, c}, d, e} − {c, {b, a, d}, e} + {c, d, {a, b, e}} is satisfied whenever one of the above elements is in M and the rest in V . For convenience, we shall omit the subscript j from {·, ·, ·}j in the sequel. A V -module M is called nondegenerate if for each m ∈ M , each one of the conditions {m, V, V } = {0}; {V, m, V } = {0} implies m = 0. A nondegenerate Jordan triple V is a nondegenerate module over itself. For a JB*- triple V , its dual V ∗ is a nondegenerate V -module. All Jordan triple modules throughout the paper are assumed to be nondegenerate. Given a, b ∈ V , the box operator a b : V → V can also be considered as a mapping from M to M . Similarly, for u ∈ V and m ∈ M , the "box operators" are defined in a natural way as v 7→ {u, m, v} and v 7→ {m, u, v} respectively. Given a, b ∈ V , the identity (iii) in Definition 2.1 implies u m, m u : V −→ M and [a b, u m] = {a, b, u} m − u {m, a, b} [a b, m u] = {a, b, m} u − m {u, a, b}. for u ∈ V and m ∈ M . We also have [u m, a b] = {u, m, a} b − a {b, u, m} and similar identity for [m u, a b]. COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 5 (2.2) Using similar arguments to the proof in [5, Lemma 1.3.7], one can show that Xi ⇒ Xi ui mi +Xj mi ui +Xj nj vj =Xk vj nj =Xk k m′ u′ m′ k u′ k +Xℓ k +Xℓ n′ ℓ v′ ℓ ℓ n′ v′ ℓ for ui, vj, u′ k, v′ ℓ ∈ V and mi, njm′ Let M0 be the linear span of k, n′ ℓ ∈ M . {u m, n v : u, v ∈ V, m, n ∈ M } in the vector space L(V, M ) of linear maps from V to M . Then M0 is the space of inner structural transformations Instrl (V, M ) (see [25, Section 7]) . Extending the above product by linearity, we can define an action of V0 on M0 by (h, ϕ) ∈ V0 × M0 7→ [h, ϕ] ∈ M0. Lemma 2.2. M0 is a V0-module of the Lie algebra V0. Proof. We are required to show that (2.3) [[h, k], ϕ] = [h, [k, ϕ]] − [k, [h, ϕ]]. We can assume that h = a b, k = c d and ϕ = w m or m w. We assume ϕ = w m, the other case being similar. For the left side of (2.3), we have [[a b, u v], w m] = [{abu} v − u {vab}, w m] = {{abu}vw} m − w {m{abu}v} − {u{vab}w} m + w {mu{vab}} = ({{abu}vw} − {u{vab}w}) m − w ({m{abu}v} − {mu{vab}}). For the right side of (2.3), we have [a b, [u v, w m]] − [u v, [a b, w m]] = [a b, {uvw} m − w {muv}] − [u v, {abw} m − w {mab}] = {ab{uvw}} w − {uvw} {mab} − {abw} {muv} + w {{muv}ab} − {uv{abw}} m + {abw} {muv} + {uvw} {mab} − w {{mab}uv} = ({ab{uvw}} − {uv{abw}}) m − w ({{mab}uv} − {{muv}ab}). (2.3) now follows from the main identity for Jordan triples. (cid:3) Let V be a Jordan triple and L(V ) its TKK Lie algebra. Given a triple V -module M , we now construct a corresponding Lie module L(M ) of the Lie algebra L(V ) as follows. Let and define the action L(M ) = M ⊕ M0 ⊕ M ((a, h, b), (m, ϕ, n)) ∈ L(V ) × L(M ) 7→ (a, h, b).(m, ϕ, n) ∈ L(M ) by (2.4) (a, h, b).(m, ϕ, n) = (hm − ϕa, [h, ϕ] + a n − m b, ϕ♮b − h♮(n) ), where, for h =Pi ai bi and ϕ =Pi ui mi +Pj nj {ui, mi, a} +Xj {ai, bi, m}, ϕa =Xi hm =Xi vj, we have the following natural definitions {nj, vj , a}, ϕ♮ =Xi mi ui +Xj vj nj 6 CHO-HO CHU AND BERNARD RUSSO in which ϕ♮ is well-defined by (2.2). Theorem 2.3. Let V be a Jordan triple and let L(V ) be its TKK Lie algebra. Let M be a triple V -module. Then L(M ) is a Lie L(V )-module. The proof of Theorem 2.3 consists of straightforward but tedious calculations. Details can be found in subsection 6.1. 3. Cohomology of Lie algebras with involution Let T be a Lie triple system. Harris [11, p. 155] has developed a cohomology theory for T in which the cohomology groups are derived from the ones of its enveloping Lie algebra Lu = T + [T, T ] where Lu is equipped with an involution θ and the cochains in the cohomology complex are invariant under θ. Our Jordan triple cohomology makes use of TKK Lie algebras which are involutive. To pave the way, we review briefly the cohomology for Lie algebras, with or without an involution. Let L be a (real or complex) Lie algebra with involution θ. Definition 3.1. Given an involutive Lie algebra (L, θ), an (L, θ)-module is a (left) L-module M, equipped with an involution eθ : M → M satisfying eθ(ℓ.µ) = θ(ℓ).eθ(µ) We also call M an involutive L-module if θ is understood. (ℓ ∈ L, µ ∈ M). For ℓ ∈ L and µ ∈ M, we define [ℓ, µ] := ℓ.µ and [µ, ℓ] := −ℓ.µ. L × · · · × L be the k-fold cartesian product of L. A k-linear map ψ : Lk → M is called Let Lk = θ-invariant if z k−times { } ψ(θx1, · · · , θxk) = eθψ(x1, · · · , xk) Let (L, θ) be an involutive Lie algebra and M an (L, θ)-module. We define A0(L, M) = M and A0 θ(L, M) to be the 1-eigenspace of eθ: A0 For k = 1, 2, . . ., we let θ(L, M) = {µ ∈ M : eθµ = µ}. Ak(L, M) = {ψ : Lk → M ψ is k-linear and alternating} and for (x1, · · · , xk) ∈ L × · · · × L. For k = 0, 1, 2, . . ., we define the coboundary operator dk : Ak(L, M) → Ak+1(L, M) by d0m(x) = Ak θ (L, M) = {ψ ∈ Ak(L, M) : ψ is θ-invariant}. x.m and for k ≥ 1, (dkψ)(x1, . . . , xk+1) = k+1Xℓ=1 (−1)ℓ+1xℓ.ψ(x1, . . . ,bxℓ, . . . , xk+1) + X1≤i<j≤k+1 (3.1) (−1)i+jψ([xi, xj], . . . ,bxi, . . . ,bxj, . . . , xk+1) where the symbol bz indicates the omission of z. The restriction of dk to the subspace Ak θ (L, M), θ (L, M) since a simple verification shows that dkψ is θ-invariant and [20, p. 167]) and the two still denoted by dk, has range Ak alternating whenever ψ is. Also, we have dkdk−1 = 0 for k = 1, 2, . . . (cf. cochain complexes d0 A0(L, M) −→ d1 A1(L, M) −→ d2 A2(L, M) −→ · · · A0 d0 θ(L, M) −→ A1 d1 θ(L, M) −→ A2 d2 θ(L, M) −→ · · · . COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 7 We often omit the subscript k from dk if there is no ambiguity. As usual, we define the k-th cohomology group of L with coefficients in M to be the quotient H k(L, M) = ker dk/dk−1(Ak−1(L, M)) = ker dk/im dk−1 for k = 1, 2, . . . and define H 0(L, M) = ker d0. We define the k-th involutive cohomology group of (L, θ) with coefficients in an (L, θ)-module M to be the quotient H k θ (L, M) = ker dk/dk−1(Ak−1 θ (L, M)) = ker dk/im dk−1 for k = 1, 2, . . . and define H 0 For k = 1, 2, . . ., the map ψ + dk−1(Ak−1 θ θ (L, M) = ker d0 ⊂ H 0(L, M). (L, M)) ∈ H k θ (L, M) 7→ ψ + dk−1(Ak−1(L, M)) ∈ H k(L, M) identifies H k θ (L, M) as a subgroup of H k(L, M). 4. Cohomology of Jordan triples 4.1. The cohomology groups. Let V be a Jordan triple and let L(V ) = V ⊕ V0 ⊕ V be its TKK Lie algebra with the involution θ(a, h, b) = (b, −h♮, a). Given a V -module M , we have shown in Theorem 2.3 that L(M ) = M ⊕ M0 ⊕ M is an L(V )-module. We define an induced involution eθ : L(M ) → L(M ) by for (m, ϕ, n) ∈ M ⊕ M0 ⊕ M. eθ(m, ϕ, n) = (n, −ϕ♮, m) Lemma 4.1. L(M ) is an (L(V ), θ)-module, that is, we have eθ(ℓ.µ) = θ(ℓ).eθ(µ) for ℓ ∈ L(V ) and µ ∈ L(M ). Let k(V ) = {(v, h, v) ∈ L(V ) : h♮ = −h} be the 1-eigenspace of the involution θ (see [5, p.48]), which is a real Lie subalgebra of L(V ), and let k(M ) = {(m, ϕ, m) ∈ L(M ) : ϕ = −ϕ♮} be the 1- eigenspace of eθ. Then k(M ) is a Lie module over the Lie algebra k(V ). We will construct cohomology groups of a Jordan triple V with coefficients in a V -module M using the cohomology groups of L(V ) with coefficients L(M ). For a real Jordan triple V , one can also make use of the cohomology groups of the real Lie algebra k(V ) with coefficients k(M ). Let V be a Jordan triple. As usual, V is identified as the subspace {(v, 0, 0) : v ∈ V } of the TKK Lie algebra L(V ). For a triple V -module M , there is a natural embedding of M into L(M ) = M ⊕ M0 ⊕ M given by and we will identify M with ι(M ). We denote by ιp : L(M ) → ι(M ) the natural projection ι : m ∈ M 7→ (m, 0, 0) ∈ L(M ) We define A0(V, M ) = M and for k = 1, 2, . . ., we denote by Ak(V, M ) the vector space of all ιp(m, ϕ, n) = (m, 0, 0). alternating k-linear maps ω : V k = Given m ∈ M , we define k−times V × · · · × V → M . z } { and view L0(m) as an extension of m ∈ A0(V, M ) to an element in A0(L(V ), L(M )) = L(M ). L0(m) = (m, 0, 0) ∈ L(M ) 8 CHO-HO CHU AND BERNARD RUSSO To motivate the definition of an extension Lk(ω) ∈ Ak(L(V ), L(M )) of an element ω ∈ Ak(V, M ), for k ≥ 1, we first consider the case k = 1 and note that ω ∈ A1(V, M ) is a Jordan triple derivation if and only if ω ◦ (a b) − (a b) ◦ ω = ω(a) b + a ω(b). Let us call a linear transformation ω : V → M extendable if the following condition holds: Xi ai bi = 0 ⇒Xi (ω(ai) bi + ai ω(bi)) = 0. Thus a Jordan triple derivation is extendable, and if ω is any extendable transformation in A1(V, M ), then the map L1(ω)(x1 ⊕ a1 b1 ⊕ y1) := ( ω(x1), ω(a1) b1 + a1 ω(b1), ω(y1) ) is well defined and extends linearly to an element L1(ω) ∈ A1(L(V ), L(M )), in which case we call L1(ω) the Lie extension of ω on the Lie algebra L(V ). Now for k > 1, given a k-linear mapping ω : V k → M , we say that ω is extendable if it satisfies the following condition under the assumption Pi ui vi = 0: (ω(ui, a2, . . . , ak) (vi + b2 + · · · + bk) + (ui + a2 + · · · + ak) ω(vi, b2, . . . , bk)) = 0, Xi for all a2, . . . , ak, b2, . . . , bk ∈ V . For an extendable ω, we can unambiguously define a k-linear map Lk(ω) : L(V )k → L(M ) as the linear extension (in each variable) of (4.1) Lk(ω)(x1 ⊕ a1 b1 ⊕ y1, x2 ⊕ a2 b2 ⊕ y2, · · · , xk ⊕ ak bk ⊕ yk) = ( ω(x1, . . . , xk), kXj=1 ω(a1, . . . , ak) bj + kXj=1 aj ω(b1, . . . , bk), ω(y1, . . . , yk) ). We call Lk(ω) the Lie extension of ω and often omit the subscript k if no confusion is likely. The following lemma is easy to verify. Lemma 4.2. Given an extendable ω ∈ Ak(V, M ), we have Lk(ω) ∈ Ak(L(V ), L(M )). Moreover, Lk(ω) ∈ Ak θ (L(V ), L(M )) if and only if k is odd. This lemma enables us to define the following extension map on the subspace Ak(V, M )′ of extendable maps in Ak(V, M ): Lk : ω ∈ Ak(V, M )′ 7→ Lk(ω) ∈ Ak(L(V ), L(M )). Conversely, given ψ ∈ Ak(L(V ), L(M )) for k = 1, 2, . . ., one can define an alternating map J k(ψ) : V k → M by J k(ψ)(x1, . . . , xk) = ιpψ( (x1, 0, 0), . . . , (xk, 0, 0) ) for (x1, . . . , xk) ∈ V k. We define J 0 : L(M ) → ι(M ) ≈ M = A0(V, M ) by ((m, ϕ, n) ∈ L(M )). J 0(m, ϕ, n) = (m, 0, 0) We call J k(ψ) the Jordan restriction of ψ in Ak(V, M ) and sometimes write J for J k if the index k is understood. Example 4.3. Given a map ψ ∈ Ak θ (L(V ), L(M )), we need not have ψ((x1, 0, 0), . . . , (xk, 0, 0)) ∈ ι(M ). Consider the inner derivation ad(m, ϕ, m) ∈ A1 θ(L(V ), L(M )) defined by For x ∈ V , we have ad(m, ϕ, m)(x ⊕ a b ⊕ y) = (x ⊕ a b ⊕ y).(m, ϕ, m). ad(m, ϕ, m)(x, 0, 0) = (x, 0, 0) · (m, ϕ, m) = (−ϕ(x), x m, 0) /∈ ι(M ). COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 9 With the identification of M and ι(M ), the map J k : ψ ∈ Ak(L(V ), L(M )) 7→ Ak(V, M ) can be viewed as the left inverse of Lk : Ak(V, M )′ → Ak(L(V ), L(M )) since for an extendable ω, we have J kLk(ω)(x1, . . . , xk) = ιpL(ω)((x1, 0, 0), . . . , (xk, 0, 0)) = ω(x1, . . . , xk). We can now define the cohomology groups for a Jordan triple V with coefficients M by means of the cochain complexes for the Lie algebra L(V ) and the involutive Lie algebra (L(V ), θ): d0 L(M ) = A0(L(V ), L(M )) −→ d1 A1(L(V ), L(M )) −→ d2 A2(L(V ), L(M )) −→ · · · ↓ J 0 ↓ J 1 ↓ J 2 M = A0(V, M ) A1(V, M ) A2(V, M ) A0 d0 θ(L(V ), L(M )) −→ A1 d1 θ(L(V ), L(M )) −→ A2 d2 θ(L(V ), L(M )) −→ ↓ J 0 ↓ J 1 ↓ J 2 A2(V, M ) For k = 0, 1, 2, . . ., the k-th cohomology groups H k(V, M ) are defined by M = A0(V, M ) A1(V, M ) · · · · · · · · · · · · · · · H 0(V, M ) = J 0(ker d0) = J 0{(m, ϕ, n) : (u, h, v).(m, ϕ, n) = 0, ∀(u, h, v) ∈ L(V )} = {m ∈ M : m v = 0, ∀v ∈ V } = {0} H k(V, M ) = Z k(V, M )/Bk(V, M ) (k = 1, 2, . . .) Z k(V, M ) = J k(Z k(L(V ), L(M ))), Z k(L(V ), L(M )) = ker dk and where and Bk(V, M ) = J k(Bk(L(V ), L(M ))), Bk(L(V ), L(M )) = im dk−1. For k = 0, 1, 2 . . ., the k-th involutive cohomology groups H k H 0(V, M ) = J 0(ker d0) = {0} θ (V, M ) are defined by and where and H k θ (V, M ) = Z k θ (V, M )/Bk θ (V, M ) (k = 1, 2, . . .) Z k θ (V, M ) = J k(Z k θ (L(V ), L(M ))), Z k θ (L(V ), L(M ))) = ker dkAk θ (L(V ),L(M)) Bk θ (V, M ) = J k(Bk θ (L(V ), L(M ))), Bk θ (L(V ), L(M )) = dk−1(Ak−1 θ (L(V ), L(M ))). We see that the map ω + Bk θ (V, M ) ∈ H k θ (V, M ) 7→ ω + Bk(V, M ) ∈ H k(V, M ) identifies H k k-cocycles, and the ones in H k in Bk(V, M ) are called the coboundaries. θ (V, M ) as a subgroup of H k(V, M ). We call elements in H k(V, M ) the Jordan triple θ (V, M ) the involutive Jordan triple k-cocycles. Customarily, elements 10 CHO-HO CHU AND BERNARD RUSSO 4.2. Triple derivations. Definition 4.4. Let V be a Jordan triple and M a triple V -module. A mapping ω : V → M is called an inner triple derivation if it is of the form ω = kXi=1 (mi vi − vi mi) ∈ M0 for some m1, . . . , mk ∈ M and v1, . . . , vk ∈ V . Note that ω♮ = −ω and (0, ω, 0) ∈ k(M ). Let us compute the first involutive cohomology group H 1 θ (V, M ) = Z 1 θ (V, M )/B1 θ (V, M ). First, we show that B1 θ (V, M ) coincides with the space of inner triple derivations from V to M . Let ω be an inner triple drivation on V . We show that its Lie extension L(ω) is a Lie inner derivation on the Lie algebra L(V ). Indeed, we have L(ω)(x ⊕ a b ⊕ y) = (ω(x), ω(a) b + a ω(b), ω(y)) = (x ⊕ a b ⊕ x).(0, −ω, 0). Hence ω = J 1(L1(ω)) ∈ B1 A1 θ(L(V ), L(M )) be a Lie inner derivation. Then for x ∈ V , we have θ (V, M ), where (0, −ω, 0) ∈ k(M ). Conversely, let ψ = ad(m, ϕ, n) ∈ J 1(ψ)(x) = ιpψ(x, 0, 0) = ιp(x, 0, 0).(m, ϕ, n) = ιp(−ϕ(x), x n, 0) = −ϕ(x), where eθ(ϕ) = ϕ implies that ϕ : V → M is an inner triple derivation. θ (V, M ) coincides with the set of triple derivations of V . We now show that Z 1 Lemma 4.5. Let ω : V → M be a triple derivation. Then L(ω) : L(V ) → L(M ) is a θ-invariant Lie derivation. Proof. For notation's sake we denote L(ω) by D. Thus D(x, a b, y) = (ω(x), ω(a) b + a ω(b), ω(y)), and it is clear that D is θ-invariant. We need to verify D[(x, a b, y), (u, c d, v)] = (x, a b, y) · D(u, c d, v) − (u, c d, v) · D(x, a b, y). for (x, a b, y), (u, c d, v) ∈ L(V ). By writing D[(x, a b, y), (u, c d, v)] = D[(x, 0, y), (u, 0, v)] + D[(0, a b, 0), (u, 0, v)] + D[(0, a b, 0), (0, c d, 0)] + D[(x, 0, y), (0, c d, 0)], we only need to verify the three identities (4.2) (4.3) and D[(x, 0, y), (u, 0, v)] = (x, 0, y) · D(u, 0, v) − (u, 0, v) · D(x, 0, y), D[(0, a b), (u, 0, v)] = (0, a b, 0) · D(u, 0, v) − (u, 0, v) · D(0, a b, 0), (4.4) D[(0, a b, 0), (0, c d, 0)] = (0, a b, 0) · D(0, c d, 0) − (0, c d, 0) · D(0, a b, 0). These are easy consequences of the definitions. For completeness we include details. The left side of (4.2) is D[(x, 0, y), (u, 0, v)] = D(0, x v − u y, 0) = (0, ω(x) v + x ω(v) − ω(u) y − u ω(y), 0), COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 11 and the right side is (x, 0, y) · D(u, 0, v) − (u, 0, v) · D(x, 0, y) = (x, 0, y) · (ω(u), 0, ω(v)) − (u, 0, v) · (ω(x), 0, ω(y)) = (0, x ω(v) − ω(u) y, 0) − (0, u ω(y) − ω(x) v, 0), proving (4.2). The left side of (4.3) is D[(0, a b, 0), (u, 0, v)] = D({abu}, 0, −{bav}) = (ω{abu}, 0, −ω{bav}) and the right side is (0, a b, 0) · D(u, 0, v) − (u, 0, v) · D(0, a b, 0) = (0, a b, 0) · (ω(u), 0, ω(v)) − (u, 0, v) · (0, ω(a) b + a ω(b), 0) = ({abω(u)}, 0, −{baω(v)}) − ({−ω(a)bu} − {aω(b)u}, 0, {bω(a)v} + {ω(b)av}), proving (4.3). The left side of (4.4) is D[(0, a b, 0), (0, c d, 0)] = D(0, [a b, c d], 0) = D(0, {abc} d − c {dab}, 0) = (0, ω{abc} d + {abc} ω(d) − ω(c) {dab} − c ω{dab}, 0) = (0, {ω(a)bc} d + {aω(b)c} d + {abω(c)} d + {abc} ω(d) −ω(c) {dab} − c {ω(d)ab} − c {dω(a)b} − c {daω(b)}, 0), and the right side is (0, a b, 0) · D(0, c d, 0) − (0, c d, 0) · D(0, a b, 0) = (0, a b, 0) · (0, ω(c) d + c ω(d), 0) − (0, c d, 0) · (0, ω(a) b + a ω(b), 0) = (0, [a b, ω(c) d] + [a b, c ω(d)] − [c d, ω(a) = (0, [a b, ω(c) d] + [a b, c ω(d)] + [ω(a) b] − [c d, a ω(b)], 0) b, c d] + [a ω(b), c d], 0 = (0, {abω(c)} d − ω(c) {dab} + {abc} ω(d) − c {ω(d)ab} +{ω(a)bc} d − c {dω(a)b} + {aω(b)c} d − c {daω(b)}, 0) proving (4.4). (cid:3) The previous lemma shows that all triple derivations ω on V are contained in Z 1 θ (V, M ). Con- versely, given a Lie derivation ψ ∈ A1 θ(L(V ), L(M )), we show below that J(ψ) is a triple derivation on V . This shows that every element in Z 1 θ (V, M ) is the space of triple derivations modulo the inner triple derivations of V into M . This will be generalized in the next subsection. θ (V, M ) is a triple derivation and hence H 1 4.3. Structural Transformations. A (conjugate-) linear transformation S : V → M is said to be a structural transformation if there exists a (conjugate-) linear transformation S ∗ : V → M such that and S{xyx} + {x(S ∗y)x} = {xySx} S ∗{xyx} + {x(Sy)x} = {xyS ∗x}. A triple derivation D is a special case of a structural transformation with D∗ = −D. By polarization, this property is equivalent to and S{xyz} + {x(S ∗y)z} = {zySx} + {xySz} S ∗{xyz} + {x(Sy)z} = {zyS ∗x} + {xyS ∗z}. 12 CHO-HO CHU AND BERNARD RUSSO As noted earlier, the space of inner structural transformations coincides, by definition, with the space M0. Triple derivations which are inner structural transformations are inner triple derivations. Also, if ω is a structural transformation, then ω −ω∗ is a triple derivation and if ω is a triple derivation, then iω is a structural transformation which is inner if ω is inner. Proposition 4.6. Let ψ be a Lie derivation of L(V ) into L(M ). Then (i) J(ψ) : V → M is a structural transformation with (Jψ)∗ = −Jψ′ where ψ′ = eθψθ. (ii) If ψ is θ-invariant, then ψ′ = ψ and Jψ is a triple derivation. (iii) If ψ is an inner derivation then Jψ is an inner structural transformation. In particular, if ψ is a θ-invariant inner derivation then Jψ is an inner triple derivation. Conversely, let ω be a structural transformation. (iv) The mapping D = 1 2 L1(ω − ω∗) : L(V ) → L(M ) defined by D(x, a b, y) = 1 2 (ω(x) − ω∗(x), ω(a) b − a ω∗(b) − ω∗(a) b + a ω(b), ω(y) − ω∗(y)) is a derivation of the Lie algebra L(V ) into L(M ). (v) D is θ-invariant if and only if ω is a triple derivation, that is, ω∗ = −ω. (vi) If ω is an inner structural transformation then D is an inner derivation. In particular, if ω is an inner triple derivation then D is a θ-invariant inner derivation. Proof. Let ψ be a Lie derivation of L(V ) into L(M ). We show first that (4.5) Jψ{abc} = {abJψ(c)} + {a, Jψ′(b), c} + {Jψ(a)bc}. Let us define n : V → M , and n1 : V → M by the formulas ψ(0, 0, x) = (m(x), ϕ(x), n(x)), and ψ(x, 0, 0) = (Jψ(x), ϕ1(x), n1(x)). Then (Jψ{abc}, ϕ1{abc}, n1{abc}) = ψ({abc}, 0, 0) = ψ[[(a, 0, 0), (0, 0, b)], (c, 0, 0)] = [(a, 0, 0), (0, 0, b)] · ψ(c, 0, 0) − (c, 0, 0) · ψ[(a, 0, 0), (0, 0, b)] = (0, a b, 0) · (Jψ(c), ϕ1(c), n1(c)) − (c, 0, 0) · ((a, 0, 0) · ψ(0, 0, b) − (0, 0, b) · ψ(a, 0, 0)) = ({abJψ(c)}, [a b, ϕ1(c)], −{ban1(c)}) −(c, 0, 0) · ((a, 0, 0) · (m(b), ϕ(b), n(b)) − (0, 0, b) · (Jψ(a), ϕ1(a), n1(a))) = ({abJψ(c)}, [a b, ϕ1(c)], −{ban1(c)}) −(c, 0, 0) · (−ϕ(b)(a), a n(b), 0) − (c, 0, 0) · (0, −Jψ(a) b, ϕ1(a)♮b) = ({abJψ(c)}, [a b, ϕ1(c)], −{ban1(c)}) −(−{an(b)c}, 0, 0)) − ({Jψ(a)bc}, c ϕ1(a)♮b, 0). Note that (m(b), ϕ(b), n(b)) = ψ(0, 0, b) = eθψ′θ(0, 0, b) = eθψ′(b, 0, 0) = eθ(Jψ′(b), ϕ′(b), n′(b)) = (n′(b), −ϕ′♮(b), Jψ′(b)), so that n(b) = Jψ′(b), proving (4.5). Applying (4.5) to ψ′ = eθψθ, we have, since ψ′′ = ψ Jψ′{abc} = {abJψ′(c)} + {a, Jψ(b), c} + {Jψ′(a)bc} (4.6) proving (i). COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 13 If ψ is θ-invariant, then ψ′ = ψ so that Jψ is a triple derivation, proving (ii). Example 4.3 provides a proof of (iii). (iv) is immediate from Lemma 4.5 since ω − ω∗ is a triple derivation. The definitions show that eθDθ = D if and only if ω = (ω −ω∗)/2, proving (v). Finally, if ω is an inner structural transformation, then ω − ω∗ is an inner triple derivation, so that L1(ω − ω∗) is an inner derivation, proving (vi). (cid:3) The following theorem provides some significant infinite dimensional examples of Lie algebras in which every derivation is inner. Its proof is in the spirit of [28]. Theorem 4.7. Let V be a von Neumann algebra considered as a Jordan triple system with the triple product {xyz} = (xy∗z + zy∗x)/2. Then every structural transformation on V is an inner structural transformation. Hence, every derivation of the TKK Lie algebra L(V ) is inner. Proof. Let S be a structural transformation on the von Neumann algebra V and to avoid cumbersome notation, denote S ∗ by S. From the defining equations, S(1) = S(1)∗, and if S(1) = 0, then S is a Jordan derivation. For an arbitrary structural transformation S, write S = S0 + S1 where S0 = S − 1 S(1) is therefore a Jordan derivation and S1 = 1 S(1) is an inner structural transformation. By the theorem of Sinclair [33], S0 is a derivation and by the theorems of Kadison and Sakai, [18, 30], S0 is an inner derivation, say S0(x) = ax − xa for some a ∈ V . By well known structure of the span of commutators in von Neumann algebras due to Pearcy-Topping, Halmos, Halpern, Fack-de la Harpe, and others (see [28] for the references), a = z +P[ci, di], where ci, di ∈ V and z belongs to the center of V . It follows that S0 = 2Xi ci d∗ i − 2Xi di c∗ i and is therefore also an inner structural transformation. The second statement follows from Proposi- tion 4.6. (cid:3) We determine the structure of L(V ) when V is a finite von Neumann algebra in Corollary 5.7 below. 5. Examples We conclude the paper with some examples of TKK Lie algebras and some Jordan triple cocycles. Let us first note the following immediate consequences of our construction. Theorem 5.1. Let V be a Jordan triple with TKK Lie algebra (L(V ), θ). If the k-th Lie coho- mology group H k(L(V ), L(M )) vanishes, then H k(V, M ) = {0} and H k θ (V, M ) = {0}. We have noted the one-to-one correspondence between the triple derivations of a Jordan triple V and the θ-invariant Lie derivations of the TKK Lie algebra (L(V ), θ), as well as the one-to-one correspondence between the Jordan inner derivations of V and the Lie inner derivations of (L(V ), θ). Corollary 5.2. Let V be a finite dimensional Jordan triple with semisimple TKK Lie algebra L(V ). Then for any finite dimensional V -module M , we have H 1(V, M ) = H 2(V, M ) = {0}. In particular, every triple derivation from V to M is inner. Proof. This follows from Whitehead's lemmas H 1(L(V ), L(M )) = H 2(L(V ), L(M )) = {0}. (cid:3) In fact, in the above corollary, we have H k(L(V ), L(M )) = {0} for all k ≥ 3 if L(M ) is a nontrivial irreducible module over L(V ). We refer to [37] for a converse of this result. 14 CHO-HO CHU AND BERNARD RUSSO 5.1. Examples of cocycles. Let V be a Jordan triple with TKK Lie algebra (L(V ), θ). We discuss examples of Jordan triple cocycles in Z k(V, M ), where M is a triple V -module, and compare them with the Lie cocycles in Z k(L(V ), L(M )). We have shown in the previous section that the space of Jordan triple derivations is exactly the space of 1-cocycles Z 1 θ (L(V ), L(M ))), where the θ-invariant Lie 1-cocycles Z 1 θ (L(V ), L(M )) are exactly the θ-invariant Lie derivations from L(V ) to L(M ). We have also shown that B1 θ (L(V ), L(M ))) is the space of triple inner derivations on V , coming from the θ-invariant Lie inner derivations B1 θ (V, M ) = J 1(B1 θ (V, M ) = J 1(Z 1 θ (L(V ), L(M )). Examples of triple 2-cocycles can be constructed from Jordan restrictions of Lie 2-cocycles. Example 5.3. If ω ∈ A2(V, M ) is extendable with L2(ω) ∈ Z 2(L(V ), L(M )), then ω = 0. Proof. For x, y, z ∈ V , 0 = d2L2(ω)((x, 0, 0), (y, 0, 0), (0, 0, z)) = (x, 0, 0) · (L(ω)((y, 0, 0), (0, 0, z) − (y, 0, 0) · (L(ω)((x, 0, 0), (0, 0, z) +(0, 0, z) · (L(ω)((x, 0, 0), (y, 0, 0) − L(ω)([(x, 0, 0), (y, 0, 0], (0, 0, z)) +L(ω)([(x, 0, 0), (0, 0, z], (y, 0, 0)) − L(ω)([(y, 0, 0), (0, 0, z], (x, 0, 0)) = −(0, ω(x, y) z, 0), hence ω(x, y) z = 0 for all x, y, z and ω = 0. (cid:3) Example 5.4. Let ϕ ∈ M0 be an inner triple derivation, and let b ∈ V . Define a linear map ψ : L(V ) → L(M ) by ψ(z) = [ [z, (0, ϕ, 0))] , (0, 0, b)] (z ∈ L(V )). Observe that ψ is not θ-invariant. Indeed, it can be seen readily that θψ(x, 0, 0) = (0, b ϕ(x), 0) while ψ(θ(x, 0, 0)) = 0. Nevertheless d1ψ ∈ B2(L(V ), L(M )) and the triple 2-coboundary Jd1ψ ∈ B2(V, M ) is given by Jd1ψ(x, y) = ιpd1ψ((x, 0, 0), (y, 0, 0)) = ιp((x, 0, 0) · ψ(y, 0, 0) − (y, 0, 0) · ψ(x, 0, 0)) = ιp((x, 0, 0) · (0, −ϕ(y) = ιp(({ϕ(y), b, x}, 0, 0) − ({ϕ(x), b, y}, 0, 0)) = {ϕ(y), b, x} − {ϕ(x), b, y} b, 0) − (y, 0, 0) · (0, −ϕ(x) b, 0)) showing that B2(V, M ) 6= 0. We note that d1ψ is not θ-invariant since eθd1ψ((x, 0, 0), (y, 0, 0)) = (0, 0, {ϕ(y), b, x}) − (0, 0, {ϕ(x), b, y}), d1ψ((0, 0, x), (0, 0, y)) = (0, 0 − {b, ϕ(y), x} + {b, ϕ(x), y}). Also Jd1ψ need not be extendable. Let V = M2(C) be the Jordan triple of 2 × 2 complex matrices. Let v =(cid:18)1 0 0 0(cid:19) and u =(cid:18)0 0 0 1(cid:19). Then we have u v = 0 and one can find a, c ∈ V such that To see this, let c = v. Then it suffices to find a ∈ V such that Jd1ψ(u, a) 2v 6= 0, where Jd1ψ(u, a) (v + c) + (u + a) Jd1ψ(v, c) 6= 0. Jd1ψ(u, a) = {ϕ(u), b, a} − {ϕ(a), b, u}. Let ϕ = m v − v m ∈ M0 where m = (cid:18) 1 −1 v(x) =(cid:26)(cid:26)(cid:18)0 Now let b = v. Then we have 1 (cid:19). Then we have ϕ(u) = −{v, m, u} = (cid:18)0 0(cid:19) ,(cid:18)1 0(cid:19) , a(cid:27) , v, x(cid:27) . Jd1ψ(u, a) −1 1 1 0 1 0(cid:19). 1 0 COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 15 Finally let a = v, then Jd1ψ(u, a) v(v) = 1 4(cid:18)0 1 1 0(cid:19) 6= 0. We have seen in Example 5.3 that there are no non-zero extendable elements ω ∈ Z 2(V, M ) with L2(ω) ∈ Z 2(L(V ), L(M )). The next example examines this phenomenon for extendable ω ∈ A3(V, M ) with L3(ω) ∈ Z 3 θ (L(V ), L(M )). We state it now as a theorem, in the statement of which, for a, b ∈ V and m ∈ M , [a, b] denotes a b − b a and [m, a] denotes m a − a m. The proof is provided in subsection 6.2. Theorem 5.5. Let ω be an extendable element of A3(V, M ). Then its Lie extension L3(ω) is a Lie 3-cocycle in A3 (5.1) θ(k(V ), k(M )) if and only if ω satisfies the following three conditions: [a, b]ω(x, y, z) = ω([a, b]x, y, z) + ω(x, [a, b]y, z) + ω(x, y, [a, b]z) for all a, b, x, y, z ∈ V ; (5.2) [ω(a, b, c), d] = [ω(d, b, c), a] = [ω(a, b, d), c] = [ω(a, d, c), b] for all a, b, c, d ∈ V ; and (5.3) [ω(x, y, [a, b]z), c] = 0. for all x, y, z, a, b, c ∈ V . 5.2. Examples of TKK algebras. We begin with the following construction from [26, Chapter 12], which has its genesis in [21, pp. 809 -- 810]. Let A be a unital associative algebra with Lie product the commutator [x, y] = xy − yx, Jordan product the anti-commutator x ◦ y = (xy + yx)/2 and Jordan triple product {xyz} = (xyz + zyx)/2 (or {xyz} = (xy∗z + zy∗x)/2 if A has an involution). Denote by Z(A) the center of A and by [A, A] the set of finite sums of commutators. Proposition 5.6. Let A be a unital associative algebra with or without an involution considered as a Jordan triple system. If Z(A) ∩ [A, A] = {0}, then the mapping (x, a b, y) 7→ (cid:20) ab y −ba (cid:21) is x an isomorphism of the TKK Lie algebra L(A) onto the Lie subalgebra (5.4) (cid:26)(cid:20) u +P[vi, wi] y x −u +P[vi, wi] (cid:21) : u, x, y, vi, wi ∈ A(cid:27) of the Lie algebra M2(A) with the commutator product. Corollary 5.7. Let V be a finite von Neumann algebra. Then L(V ) is isomorphic to the Lie algebra [M2(V ), M2(V )]. Proof. The center valued trace of V is zero on [V, V ] and the identity on Z(V ), so the theorem applies. Since M2(V ) is also a finite von Neumann algebra, [M2(V ), M2(V )] coincides with the elements of M2(V ) of central trace zero (by [6, Theoreme 3.2]), so it remains to show that every such element has the form (5.4). For this one can use the argument from [26, pp. 129 -- 130] as follows: if (cid:20) a b c d (cid:21) ∈ M2(V ) has central trace zero, then tr (a) = −tr (d) and −b′ + c′ (cid:21) , d (cid:21) =(cid:20) b′ + c′ (cid:20) a b c c b where c′ = (a + d)/2 and b′ = (a − d)/2. (cid:3) In a properly infinite von Neumann algebra, the assumption Z(A) ∩ [A, A] = {0} fails since A = [A, A]. This assumption also fails in the Murray-von Neumann algebra of measurable operators affiliated with a factor of type II1 ([34]). For a finite factor of type In, Corollary 5.7 states that the classical Lie algebras sl(2n, C) of type A are TKK Lie algebras. Similarly, the TKK Lie algebra of 16 CHO-HO CHU AND BERNARD RUSSO a Cartan factor of type 3 on an n-dimensional Hilbert space is the classical Lie algebra sp(2n, C) of type C ([26, Theorem 3,p. 131]). More examples of TKK Lie algebras can be found in [5, 1.4] and [22, Chapter III]. 6. Proofs of Theorems 2.3 and 5.5 6.1. Proof of Theorem 2.3. For the convenience of the reader, we repeat the statement of Theorem 2.3. Theorem. Let V be a Jordan triple and let L(V ) be its TKK Lie algebra. Let M be a triple V -module. Then L(M ) is a Lie L(V )-module. For the proof, we are required to show that (6.1) [(a, h, b), (c, k, d)] · (m, ϕ, n) = (a, h, b) · ((c, k, d) · (m, ϕ, n)) − (c, k, d) · ((a, h, b)) · (m, ϕ, n)). Let L denote the left side of (6.1). Then (6.2) L = (hc − ka , [h, k] + a d − c , k♮b − h♮d ) · (m, ϕ, n) = (Hm − ϕA , [H, ϕ] + A n − m B {z }A {z L1 } H {z L2 {z b } } {z } B , ϕ♮B − H ♮n ). L3 {z } We can assume that h = x y, k = u v so that • A = {xyc} − {uva} • H = {xyu} v − u {vxy} + a d − c • B = {vub} − {yxd}. b Let R denote the right side of (6.1). Then (6.3) R = (a, h, b) · (km − ϕc , [k, ϕ] + c n − m d , ϕ♮d − k♮n ) − (c, k, d) · (hm − ϕa , [h, ϕ] + a n − m b , ϕ♮b − h♮n ) = (hC − Φa Φ′ , [h, Φ] + a D − C b , Φ♮b − h♮D ) C {z {z C ′ } } } } {z {z R′ 2 R2 R1 {z {z R′ 1 } } Φ {z {z } } } } R3 {z R3 {z D {z {z D′ } } − (kC ′ − Φ′c , [k, Φ′] + c D′ − C ′ d , Φ′♮ − k♮D′ ) = (R1 − R′ 1, R2 − R′ 2, R3 − R′ 3). As above, with h = x y, k = u v and with ϕ = w p + q z, with p, q ∈ M , we have • C = {uvm} − {wpc} − {qzc} • D = {pwd} + {zqd} − {vun} • Φ = {uvw} p − w {puv} + {uvq} z − q • C ′ = {xym} − {wpa} − {qza} • D′ = {pwb} + {zqb} − {yxn} • Φ′ = {xyw} p − w {pxy} + {xyq} z − q {zuv} + c n − m d {zxy} + a n − m b COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 17 We now show that L1 = R1 − R′ 1. We have from (6.2) (6.4) L1 = Hm − ϕA = {{xyu}vm} − {u, vxy, m} + {adm} − {cbm} − {wpA} −{qzA} = {{xyu}vm} − {u, vxy, m} + {adm} − {wp{xyc}} + {wp{uva}} − {qz{xyc}} + {qz{uva}} } 1 {z } {z 6 } 1 {z } 3 {z {z }5 } − {cbm} {z }8 } {z 7 2 {z and from (6.3) (6.5) R1 = hC − Φa = {xy{uvm}} − {xy{wpc}} − {xy{qzc}} − {cna} + {mda} − {{uvw}pa} + {w, puv, a} − {{uvq}za} + {q, zuv, a} } } 2 {z {z 6 } } 3 {z {z 7 } {z }4 7 {z {z }5 } } 1 {z {z 6 = {uv{xym}} − {uv{wpa}} − {uv{qza}} − {{xyw}pc} + {w, pxy, c} − {{xyq}zc} + {q, zxy, c} 1 {z {z 2 } } 6 {z {z 3 } } 7 {z {z 3 } } 2 − {anc} {z {z }4 + {mbc} } {z }8 and (6.6) R′ 1 = kC ′ − Φ′c From (6.4)-(6.6), we have L1 = R1 − R′ 1. In (6.4)-(6.6) we have indicated which terms cancel. To see that the terms labeled 6 cancel, replace {uv{wpa}} by {{uvw}pa} − {w, vup, a} + {wp{uva}}. Similarly, to see that the terms labeled 7 cancel, replace {uv{qza}} by {{uvq}za} − {q, vuz, a} + {qz{uva}}. We next show that L2 = R2 − R′ 2. We have from (6.2) (6.7) L2 = [H, ϕ] + a n − m B = [H, w p + q z] + {xyc} n − {uva} n − m {vub} + m {yxd} = [{xyu} v, w p] − [u {vxy}, w p] + [a d, w p] − [c z] − [c z] − [u {vxy}, q +[{xyu} v, q z] + [a d, q b, w p] z] b, q and from (6.3) (6.8) +{xyc} n − {uva} n − m {vub} + m {yxd} R2 = [h, Φ] + a D − C b = [x y, [u v, w p + q z]] + [x y, c n] − [x y, m d] +a {pwd} + a {zqd} − a {vun} − {uvm} b +{wpc} b + {qzc} b = [x y, {uvw} p] − [x y, w {puv}] + [x y, {uvq} z] −[x y, q −a {vun} − {uvm} b + {wpc} b + {qzc} b, {zuv}] + [x y, c n] − [x y, m d] + a {pwd} 18 and (6.9) CHO-HO CHU AND BERNARD RUSSO R′ 2 = [k, Φ′] + c D′ − C ′ d = [u v, [x y, w p + q z]] + [u v, a n] − [u v, m b] +c {pwd} + c {zqb} − c {yxn} − {xym} d +{wpa} d + {qza} d = [u v, {xyw} p] − [u v, w {pxy}] + [u v, {xyq} z] −[u v, q {zxy}] + [u v, a n] − [u v, m b] + c {pwb} +c {zqb} − {xym} d + {wpa} d + {qza} d − c {yxn} From (6.7) we have (6.10) L2 = {{xyu}vw} p − w {v, xyu, p} − {u, vxy, w} p +w {{vxy}up} + {adw} p − w {dap} − {cbw} p +w {bcp} + {{xyu}vq} z − q {v, xyu, z} − {u, vxy, q} z {{vxy}uz} + {adq} z − q +q +{xyc} n − {uva} n − m {vub} + m {yxd} {daz} − {cbq} z + q {bcz} = ({{xyu}vw} − {u, vxy, w} + {adw} − {cbw} ) p 19 +w (−{v, xyu, p} + {{vxy}up} 1 {z 3 {z {z }9 2 {z 4 {z {z }11 } } + {bcp} ) {z}6 z −{dap} 19 − {cbq} ) {z }6 {z } {z } } {z}7 {z }20 } {z } {z }13 −{daz} + {yxd} + {bcz} ) {z}7 {z }12 ). +({{xyu}vq} − {u, vxy, q} + {adq} +q (−{v, xyu, z} + {{vxy}uz} +({xyc} − {uva} 20 ) n + m (− {vub} From (6.8) we have (6.11) R2 = {xy{uvw}} p − {uvw} {yxp} − {xyw} {puv} − {xyq} {zuv} +q {yx{zuv}} +w {yx{puv}} + {xy{uvq}} z − {uvq} {yxz} 1 {z } {z 2 18 − {xym} {z {z } {z }8 14 {vun} −a } } 15 {z {z 3 {z }12 } {z 4 +a } + {xyc} } {z }11 {z } {pwd} 19 +a d + m {yxd} − {uvm} b + {wpc} b + {qzc} {z }5 {z }6 {z}7 16 {z {z 17 n − c } } {zqd} {z }20 b. {yxn} {z } 10 } } +w {vu{pxy}} + {uv{xyq}} z − {xyq} {vuz} 1 {z } {z 2 } 16 {z {z 3 } − {uvq} {yxz} +q {vu{yxz}} + {uva} n − a 17 − {uvm} {z {z }5 {z } − {xym} b + m {vub} {z 4 +c {z }13 {pwb} } {z }6 {z }20 {z } d + {wpa} d + {qza} d − c {vun} {z }8 . {z }9 +c 15 {z {z 18 } } {zqb} {yxn} {z}7 {z } 10 COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 19 From (6.9) we have (6.12) R′ 2 = {uv{xyw}} p − {xyw} {vup} − {uvw} {pxy} 19 From (6.7)-(6.9), we have L2 = R2 − R′ 2. In (6.7)-(6.9) we have indicated which terms cancel. The terms labeled 1 -- 4 cancel by the main identity. The terms labeled 5 and 8-18 cancel in pairs. The terms labeled 6,7,19,20 all cancel because of the following identity: 14 {abc} d − c {bad} = a {dcb} − {cda} b. which follows from the main identity {ab{cde}} − {cd{abe}} = {{abc}de} − {c, bad, e} by interchanging (a, b) with (c, d) and noticing that the left side changes sign. It remains to show that L3 = R3 − R′ 3. We leave this as an exercise for the reader. 6.2. Proof of Theorem 5.5. For the reader's convenience, we repeat the statement of Theorem 5.5, recalling that we write [a, b] for a b − b a and [m, a] for m a − a m for a, b ∈ V and m ∈ M . Theorem. Let ω be an extendable element of A3(V, M ). Then its Lie extension L3(ω) is a Lie 3-cocycle in A3 θ(k(V ), k(M )) if and only if ω satisfies the following three conditions: (6.13) [a, b]ω(x, y, z) = ω([a, b]x, y, z) + ω(x, [a, b]y, z) + ω(x, y, [a, b]z) for all a, b, x, y, z ∈ V ; (6.14) [ω(a, b, c), d] = [ω(d, b, c), a] = [ω(a, b, d), c] = [ω(a, d, c), b] for all a, b, c, d ∈ V ; and (6.15) [ω(x, y, [a, b]z), c] = 0. for all x, y, z, a, b, c ∈ V . Let ω ∈ A3(V, M ) be extendable and let ψ = d3L3(ω) (ψ is θ-invariant since 3 is odd). Write aj, xj) ∈ k(V ) as Xj = (xj , 0, xj) + (0, [aj, bj], 0). By the alternating character Xj = (xj , aj of ψ, it is a Lie 3-cocycle, that is, ψ(X1, X2, X3, X4) = 0 for Xj ∈ k(V ), if and only if the following five equations hold for ai, bi, xi ∈ V . bj − bj (6.16) (6.17) (6.18) (6.19) (6.20) ψ((x1, 0, x1), (x2, 0, x2), (x3, 0, x3), (x4, 0, x4)) = 0, (4 variables) ψ((x1, 0, x1), (x2, 0, x2), (x3, 0, x3), (0, [a4, b4], 0)) = 0, (5 variables) ψ((x1, 0, x1), (x2, 0, x2), (0, [a3, b3], 0), (0, [a4, b4], 0)) = 0, (6 variables) ψ((x1, 0, x1), (0, [a2, b2], 0), (0, [a3, b3], 0), (0, [a4, b4], 0)) = 0, (7 variables) ψ((0, [a1, b1], 0), (0, [a2, b2], 0), (0, [a3, b3], 0), (0, [a4, b4], 0)) = 0. (8 variables) 20 CHO-HO CHU AND BERNARD RUSSO Note that (6.13)-(6.15) involve 5,4 and 6 variables respectively, so there is an additional amount of redundancy in (6.16)-(6.20). We shall begin by showing that (6.16)-(6.20) imply (6.13)-(6.15). Straightforward calculation of (6.16), using (3.1) and (4.1), shows that it is equivalent to (6.21) [x1, ω(x2, x3, x4)] − [x2, ω(x1, x3, x4)] +[x3, ω(x1, x2, x4)] − [x4, ω(x1, x2, x3)] = 0. We shall see shortly that (6.21) is redundant since it will follow from the identity (6.14), which will be proved using (6.18). However, (6.21) will be used later, in the proof that (6.13)-(6.15) imply (6.16)-(6.20). Similarly, (6.17) is equivalent to −{a4b4ω(x1, x2, x3)} + {b4a4ω(x1, x2, x3)} +ω({a4b4x1}, x2, x3) − ω({b4a4x1}, x2, x3) −ω({a4b4x2}, x1, x3) + ω({b4a4x2}, x1, x3) +ω({a4b4x3}, x1, x2) − ω({b4a4x3}, x1, x2) = 0. which can be rewritten as (6.22) [a4, b4](ω(x1, x2, x3)) = ω([a4, b4]x1, x2, x3) +ω(x1, [a4, b4]x2, x3) + ω(x1, x2, [a4, b4]x3), proving (6.13) (assuming only (6.17)). An interpretation of (6.22) is that the inner triple derivation [a, b] (for the triple product {·, ·, ·} of V ) is also a "triple derivation" for the (ad hoc M -valued) triple product (x, y, z) 7→ ω(x, y, z) of V . In order to proceed efficiently, it is convenient to state the following formulas. First, for ai and bi in V , by (4.1), (6.23) where (6.24) L3(ω)((0, [a1, b1], 0), (0, [a2, b2], 0), (0, [a3, b3], 0)) = (0, Λ, 0) Λ = [ω(a1, a2, a3), b1 + b2 + b3] − [ω(b1, a2, a3), a1 + b2 + b3] + [ω(b1, b2, a3), a1 + a2 + b3] − [ω(b1, b2, b3), a1 + a2 + a3] + [ω(b1, a2, b3), a1 + b2 + a3] + [ω(a1, b2, b3), b1 + a2 + a3] − [ω(a1, b2, a3), b1 + a2 + b3] − [ω(a1, a2, b3), b1 + b2 + a3]. Second, for a, b, c ∈ V and m ∈ M , by (2.4), (6.25) (0, [a, b], 0) · (0, [m, c], 0) = (0, [[a, b]m, c] + [m, [a, b]c], 0), and, for ai and bi in V , by (2.1), (6.26) [(0, [a1, b1], 0), (0, [a2, b2], 0)] = (0, [[a1, b1]a2, b2] + [[b1, a1]b2, a2], 0). Returning to (6.18)-(6.20) and observing that a straightforward calculation of (6.18) shows that it is equivalent to L3(ω)((∗, 0, ∗), (0, ∗, 0), (∗, ∗, ∗)) = 0, 0 = −L3(ω)((0, [x1, x2], 0), (0, [a3, b3], 0), (0, [a4, b4], 0)), which by (6.23) and (6.24) is equivalent to (6.27) 0 = [ω(x1, a3, a4), x2 + b3 + b4] − [ω(x2, a3, a4), x1 + b3 + b4] + [ω(x2, b3, a4), x1 + a3 + b4] − [ω(x2, b3, b4), x1 + a3 + a4] + [ω(x2, a3, b4), x1 + b3 + a4] + [ω(x1, b3, b4), x2 + a3 + a4] − [ω(x1, b3, a4), x2 + a3 + b4] − [ω(x1, a3, b4), x2 + b3 + a4]. COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 21 We shall now see that (6.27) simplifies considerably and gives the same information as (6.19), namely (6.27) is equivalent to (6.28) [ω(a, b, c), d] = [ω(d, b, c), a] = [ω(a, b, d), c] = [ω(a, d, c), b], which is (6.14). Assuming that this has been done, we will have proved that (6.17) is equivalent to (6.13); and that (6.18), (6.19) and (6.14) are equivalent. We shall complete the proof by showing that (6.20), together with (6.13) and (6.14), implies (6.15); and then proving that (6.13)-(6.15) imply (6.16)-(6.20). Note that (6.13), (6.15) and the alternating character of ω imply [[a, b]ω(x, y, z), c] = 0, and that (6.14) and (6.15) imply (6.29) [ω(x, y, z), [a, b]c] = 0. We continue the proof of Theorem 5.5 by showing that (6.28) follows from (6.27) and that (6.19) does not contribute any new properties of ω. After that, we shall deal with (6.20). Since we are assuming (6.18), we may set x1 = 0 and a3 = 0 in (6.27). The result is [ω(x2, b3, a4), b4] − [ω(x2, b3, b4), a4] = 0. If one repeats this process with (x1 = 0 and) a3 = 0 replaced successivly by a4 = 0, b3 = 0, b4 = 0, one obtains three more such equations. Next, replace x1 = 0 by x2 = 0 to obtain four more such equations. Finally, setting a3 = 0 and a4 = 0 in (6.27), and repeating with (a3, a4) replaced successively with (a3, b4), (b3, a4), (b3, b4) results in four more such equations. By changing the names of the variables, the resulting twelve equations reduce to (6.28) (which is (6.14)). We next show that (6.19) yields the same information as (6.18). Straightforward calculation of (6.19) shows that it is equivalent to 0 = (x1, 0, x1) · L3(ω)((0, [a2, b2], 0), (0, [a3, b3], 0), (0, [a4, b4], 0)), which by (6.23) equals (x1, 0, x1) · (0, Λ, 0) = (−Λx1, 0, −Λx1) where (6.30) Λ = [ω(a2, a3, a4), b2 + b3 + b4] − [ω(b2, a3, a4), a2 + b3 + b4] + [ω(b2, b3, a4), a2 + a3 + b4] − [ω(b2, b3, b4), a2 + a3 + a4] + [ω(b2, a3, b4), a2 + b3 + a4] + [ω(a2, b3, b4), b2 + a3 + a4] − [ω(a2, b3, a4), b2 + a3 + b4] − [ω(a2, a3, b4), b2 + b3 + a4], Thus, (6.19) results in (6.31) Λx1 = 0. where Λ is given by (6.30). Comparing this with (6.27) shows that (6.19) is equivalent to (6.18). We now have that (6.18), (6.19), (6.27), (6.28) and (6.14) are equivalent, and that (6.17) and (6.13) are equivalent. It remains, for this part of the proof, to establish (6.15) using (6.16)-(6.20). This will take some perseverance! In order to process (6.20) we shall adopt the following self-explanatory notation. For distinct elements i, j, k, l ∈ {1, 2, 3, 4}, set (6.32) and (6.33) ijkl1 = (0, [ai, bi], 0) · L3(ω)((0, [aj, bj], 0), (0, [ak, bk], 0), (0, [al, bl], 0)) ijkl2 = L3(ω)([(0, [ai, bi], 0), (0, [aj, bj], 0)], (0, [ak, bk], 0), (0, [al, bl], 0)). Then equation (6.20) for ψ = d3L3(ω) is restated as: (6.34) 0 = 12341 − 21341 + 31241 − 41231 −12342 + 13242 − 14232 − 23142 + 24132 − 34122. By (6.32), using (6.23)-(6.24), ijkl1 = (0, [ai, bi], 0) · (0, Λj,k,l, 0) 22 where (6.35) and by (6.25), CHO-HO CHU AND BERNARD RUSSO Λj,k,l = [ω(aj, ak, al), bj + bk + bl] − [ω(bj, ak, al), aj + bk + bl] + [ω(bj, bk, al), aj + ak + bl] − [ω(bj, bk, bl), aj + ak + al] + [ω(bj, ak, bl), aj + bk + al] + [ω(aj, bk, bl), bj + ak + al] − [ω(aj, bk, al), bj + ak + bl] − [ω(aj, ak, bl), bj + bk + al], (6.36) where (6.37) ijkl1 = (0, Γi,j,k,l, 0) Γi,j,k,l = [[ai, bi]ω(aj, ak, al), bj + bk + bl] + [ω(aj, ak, al), [ai, bi](bj + bk + bl)] − [[ai, bi]ω(bj, ak, al), aj + bk + bl] − [ω(bj, ak, al), [ai, bi](aj + bk + bl)] + [[ai, bi]ω(bj, bk, al), aj + ak + bl] + [ω(bj, bk, al), [ai, bi](aj + ak + bl)] − [[ai, bi]ω(bj, bk, bl), aj + ak + al] − [ω(bj, bk, bl), [ai, bi](aj + ak + al)] + [[ai, bi]ω(bj, ak, bl), aj + bk + al] + [ω(bj, ak, bl), [ai, bi](aj + bk + al)] + [[ai, bi]ω(aj, bk, bl), bj + ak + al] + [ω(aj, bk, bl), [ai, bi](bj + ak + al)] − [[ai, bi]ω(aj, bk, al), bj + ak + bl] − [ω(aj, bk, al), [ai, bi](bj + ak + bl)] − [[ai, bi]ω(aj, ak, bl), bj + bk + al] − [ω(aj, ak, bl), [ai, bi](bj + bk + al)]. By (6.33), using (6.26) and (6.23)-(6.24), (6.38) where (6.39) ijkl2 = L3(ω)((0, [[ai, bi]aj, bj] + [[bi, ai]bj, aj], 0), (0, [ak, bk], 0), (0, [al, bl], 0)) = (0, ∆i,j,k,l, 0), ∆i,j,k,l = [ω([ai, bi]aj, ak, al), bj + bk + bl] + [ω([bi, ai]bj, ak, al), aj + bk + bl)] − [ω(bj, ak, al), [ai, bi]aj + bk + bl] − [ω(aj, ak, al), [bi, ai](bj + bk + bl)] + [ω(bj, bk, al), [ai, bi]aj + ak + bl] + [ω(aj, bk, al), [bi, ai](bj + ak + bl)] − [ω(bj, bk, bl), [ai, bi]aj + ak + al] − [ω(aj, bk, bl), [bi, ai](bj + ak + al)] + [ω(bj, ak, bl), [ai, bi]aj + bk + al] + [ω(aj, ak, bl), [bi, ai](bj + bk + al)] + [ω([ai, bi]aj, bk, bl), bj + ak + al] + [ω([bi, ai]bj, bk, bl), aj + ak + al] − [ω([ai, bi]aj, bk, al), bj + ak + bl] − [ω([bi, ai]bj, bk, al), aj + ak + bl] − [ω([ai, bi]aj, ak, bl), bj + bk + al] − [ω([bi, ai]bj, ak, bl), aj + bk + al]. We next analyze (6.37) and (6.39). First, applying (6.13) to the first bracket on each line of (6.37) and applying (6.14) to the expansion of those brackets results in 72 terms, 24 of which cancel with all of the terms in the second bracket on each line of (6.37). Thus the 96 terms in (6.37) are reduced to the 48 terms in (6.40) Γi,j,k,l = COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 23 [ω([ai, bi]aj, ak, al), (bk + bl)] + [ω(aj, [ai, bi]ak, al), (bj + bl)] + [ω(aj, ak, [ai, bi]al), (bj + bk)] − [ω([ai, bi]bj, ak, al), (bk + bl)] − [ω(bj, [ai, bi]ak, al), (aj + bl)] − [ω(bj, ak, [ai, bi]al), (aj + bk)] + [ω([ai, bi]bj, bk, al), (ak + bl)] + [ω(bj, [ai, bi]bk, al), (aj + bl)] + [ω(bj, bk, [ai, bi]al), (aj + ak)] − [ω([ai, bi]bj, bk, bl), (ak + al)] − [ω(bj, [ai, bi]bk, bl), (aj + al)] − [ω(bj, bk, [ai, bi]bl), (aj + ak)] + [ω([ai, bi]bj, ak, bl), (bk + al)] + [ω(bj, [ai, bi]ak, bl), (aj + al)] + [ω(bj, ak, [ai, bi]bl), (aj + bk)] − [ω([ai, bi]aj, bk, bl), (ak + al)] − [ω(aj, [ai, bi]bk, bl), (bj + al)] − [ω(aj, bk, [ai, bi]bl), (bj + ak)] + [ω([ai, bi]aj, bk, al), (ak + bl)] + [ω(aj, [ai, bi]bk, al), (bj + bl)] + [ω(aj, bk, [ai, bi]al), (bj + ak)] − [ω([ai, bi]aj, ak, bl), (bk + al)] − [ω(aj, [ai, bi]ak, bl), (bj + al)] − [ω(aj, ak, [ai, bi]bl), (bj + bk)]. Second, the 8 first brackets on the lines of (6.39) sum to zero, as can be seen by expanding and noting that the resulting terms cancel in pairs by applying (6.14). Thus (6.39) reduces (initially) to the sum of the 8 second brackets on the lines of (6.39), namely, (6.41) ∆i,j,k,l = [ω([bi, ai]bj, ak, al), aj + bk + bl)] − [ω(aj, ak, al), [bi, ai](bj + bk + bl)] + [ω(aj, bk, al), [bi, ai](bj + ak + bl)] − [ω(aj, bk, bl), [bi, ai](bj + ak + al)] + [ω(aj, ak, bl), [bi, ai](bj + bk + al)] + [ω([bi, ai]bj, bk, bl), aj + ak + al] − [ω([bi, ai]bj, bk, al), aj + ak + bl] − [ω([bi, ai]bj, ak, bl), aj + bk + al]. However, there is still more cancellation in (6.41) using (6.14), and what remains is (6.42) ∆i,j,k,l = −[ω(aj, ak, al), [bi, ai](bk + bl)] +[ω(aj, bk, al), [bi, ai](ak + bl)] −[ω(aj, bk, bl), [bi, ai](ak + al)] +[ω(aj, ak, bl), [bi, ai](bk + al)] . The equation (6.34) is thus equivalent to (6.43) 0 = Γ1234 − Γ2134 + Γ3124 − Γ4123 −∆1234 + ∆1324 − ∆1423 − ∆2314 + ∆2413 − ∆3412, where Γijkl and ∆ijkl are given by (6.40) and (6.42). We are now going to decompose each term in (6.43) into "irreducible pieces" as follows. First some notation. Let Σ denote the right side of (6.43), let Γijkl(a1 = 0) denote the sum of the terms of Γijkl which do not involve the variable a1, and Γijkl(a1 6= 0) the sum of the terms of Γijkl which contain the variable a1, with similar notation for other variables, for more then one variable, and for ∆ijkl. With Σ(a1 = 0) denoting the sum of the terms of Σ not containing a1, etc., we have (and this is the first of two underlying principles in what follows) Σ = 0 if and only if Σ(a1 = 0) = 0 and Σ(a1 6= 0) = 0. We shall use (6.40) to process the Γijkl in (6.43) and in parallel use (6.42) to process the ∆ijkl in (6.43). Here we go! By (6.40), (6.44) (6.45) Γi,j,k,l(ai = 0) = 0, Γi,j,k,l(aj = 0) = 24 CHO-HO CHU AND BERNARD RUSSO − [ω([ai, bi]bj, ak, al), (bk + bl)] − [ω(bj, [ai, bi]ak, al), bl] − [ω(bj, ak, [ai, bi]al), bk] + [ω([ai, bi]bj, bk, al), (ak + bl)] + [ω(bj, [ai, bi]bk, al), bl] + [ω(bj, bk, [ai, bi]al), ak] − [ω([ai, bi]bj, bk, bl), (ak + al)] − [ω(bj, [ai, bi]bk, bl), al] − [ω(bj, bk, [ai, bi]bl), ak] + [ω([ai, bi]bj, ak, bl), (bk + al)] + [ω(bj, [ai, bi]ak, bl), al] + [ω(bj, ak, [ai, bi]bl), bk], (6.46) Γi,j,k,l(ak = 0) = [ω([ai, bi]bj, bk, al), bl] + [ω(bj, [ai, bi]bk, al), (aj + bl)] + [ω(bj, bk, [ai, bi]al), aj] − [ω([ai, bi]bj, bk, bl), al] − [ω(bj, [ai, bi]bk, bl), (aj + al)] − [ω(bj, bk, [ai, bi]bl), aj] − [ω([ai, bi]aj, bk, bl), al] − [ω(aj, [ai, bi]bk, bl), (bj + al)] − [ω(aj, bk, [ai, bi]bl), bj] + [ω([ai, bi]aj, bk, al), bl] + [ω(aj, [ai, bi]bk, al), (bj + bl)] + [ω(aj, bk, [ai, bi]al), bj], and (6.47) Γi,j,k,l(al = 0) = − [ω([ai, bi]bj, bk, bl), ak] − [ω(bj, [ai, bi]bk, bl), aj] − [ω(bj, bk, [ai, bi]bl), (aj + ak)] + [ω([ai, bi]bj, ak, bl), bk] + [ω(bj, [ai, bi]ak, bl), aj] + [ω(bj, ak, [ai, bi]bl), (aj + bk)] − [ω([ai, bi]aj, bk, bl), ak] − [ω(aj, [ai, bi]bk, bl), bj] − [ω(aj, bk, [ai, bi]bl), (bj + ak)] − [ω([ai, bi]aj, ak, bl), bk] − [ω(aj, [ai, bi]ak, bl), bj] − [ω(aj, ak, [ai, bi]bl), (bj + bk)]. On the other hand, by (6.42), (6.48) (6.49) (6.50) and (6.51) ∆i,j,k,l(ai = 0) = 0, ∆i,j,k,l(aj = 0) = 0, ∆i,j,k,l(ak = 0) = [ω(aj, bk, al), [bi, ai]bl] − [ω(aj, bk, bl), [bi, ai]al], ∆i,j,k,l(al = 0) = −[ω(aj, bk, bl), [bi, ai]ak] +[ω(aj, ak, bl), [bi, ai]bk]. Returning to (6.40), by (6.44) (6.52) By (6.45) (6.53) Γ1234(a1 = 0) = 0. Γ2134(a1 = 0) = − [ω([a2, b2]b1, a3, a4), (b3 + b4)] − [ω(b1, [a2, b2]a3, a4), b4] − [ω(b1, a3, [a2, b2]a4), b3] + [ω([a2, b2]b1, b3, a4), (a3 + b4)] + [ω(b1, [a2, b2]b3, a4), b4] + [ω(b1, b3, [a2, b2]a4), a3] − [ω([a2, b2]b1, b3, b4), (a3 + a4)] − [ω(b1, [a2, b2]b3, b4), a4] − [ω(b1, b3, [a2, b2]b4), a3] + [ω([a2, b2]b1, a3, b4), (b3 + a4)] + [ω(b1, [a2, b2]a3, b4), a4] + [ω(b1, a3, [a2, b2]b4), b3], (6.54) Γ3124(a1 = 0) = − [ω([a3, b3]b1, a2, a4), (b2 + b4)] − [ω(b1, [a3, b3]a2, a4), b4] − [ω(b1, a2, [a3, b3]a4), b2] + [ω([a3, b3]b1, b2, a4), (a2 + b4)] + [ω(b1, [a3, b3]b2, a4), b4] + [ω(b1, b2, [a3, b3]a4), a2] − [ω([a3, b3]b1, b2, b4), (a2 + a4)] − [ω(b1, [a3, b3]b2, b4), a4] − [ω(b1, b2, [a3, b3]b4), a2] + [ω([a3, b3]b1, a2, b4), (b2 + a4)] + [ω(b1, [a3, b3]a2, b4), a4] + [ω(b1, a2, [a3, b3]b4), b2], COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 25 and (6.55) Γ4123(a1 = 0) = − [ω([a4, b4]b1, a2, a3), (b2 + b3)] − [ω(b1, [a4, b4]a2, a3), b3] − [ω(b1, a2, [a4, b4]a3), b2] + [ω([a4, b4]b1, b2, a3), (a2 + b3)] + [ω(b1, [a4, b4]b2, a3), b3] + [ω(b1, b2, [a4, b4]a3), a2] − [ω([a4, b4]b1, b2, b3), (a2 + a3)] − [ω(b1, [a4, b4]b2, b3), a3] − [ω(b1, b2, [a4, b4]b3), a2] + [ω([a4, b4]b1, a2, b3), (b2 + a3)] + [ω(b1, [a4, b4]a2, b3), a3] + [ω(b1, a2, [a4, b4]b3), b2]. On the other hand, by (6.48) (6.56) By (6.50), (6.57) (6.58) and (6.59) ∆1234(a1 = 0) = 0, ∆1324(a1 = 0) = 0, ∆1423(a1 = 0) = 0. ∆2314(a1 = 0) = [ω(a3, b1, a4), [b2, a2]b4] − [ω(a3, b1, b4), [b2, a2]a4], ∆2413(a1 = 0) = [ω(a4, b1, a3), [b2, a2]b3] − [ω(a4, b1, b3), [b2, a2]a3], ∆2413(a1 = 0) = [ω(a4, b1, a2), [b3, a3]b2] − [ω(a4, b1, b2), [b3, a3]a2]. By (6.43), and (6.52)-(6.59), 0 = Σ(a1 = 0) = −(6.53) + (6.54) − (6.55) − (6.57) + (6.58) − (6.59), and each of the terms on the right side must be decomposed further. Here, we are using the notation (6.53) to denote Γ2134(a1 = 0) and similarly for (6.54), etc. We shall analyze (6.53) first. By (6.53), (6.60) and (6.61) Γ2134(a1 = 0, a3 = 0) = + [ω([a2, b2]b1, b3, a4), b4] + [ω(b1, [a2, b2]b3, a4), b4] − [ω([a2, b2]b1, b3, b4), a4] − [ω(b1, [a2, b2]b3, b4), a4] Γ2134(a1 = 0, a3 6= 0) = − [ω([a2, b2]b1, a3, a4), (b3 + b4)] − [ω(b1, [a2, b2]a3, a4), b4] − [ω(b1, a3, [a2, b2]a4), b3] + [ω([a2, b2]b1, b3, a4), a3] + [ω(b1, b3, [a2, b2]a4), a3] − [ω([a2, b2]b1, b3, b4), a3] − [ω(b1, b3, [a2, b2]b4), a3] + [ω([a2, b2]b1, a3, b4), (b3 + a4)] + [ω(b1, [a2, b2]a3, b4), a4] + [ω(b1, a3, [a2, b2]b4), b3]. The identity given by (6.60) is "irreducible" in the sense that if any of its variables is zero, then it vanishes identically (This is the second of the two underlying principles mentioned earlier). However, since it is a consequence of (6.14), it does not give any new identities and can be ignored. We proceed to decompose (6.61) as follows. (6.62) (6.63) Γ2134(a1 = 0, a3 6= 0, a4 = 0) = − [ω([a2, b2]b1, b3, b4), a3] − [ω(b1, b3, [a2, b2]b4), a3] + [ω([a2, b2]b1, a3, b4), b3] + [ω(b1, a3, [a2, b2]b4), b3]. Γ2134(a1 = 0, a3 6= 0, a4 6= 0) = 26 CHO-HO CHU AND BERNARD RUSSO − [ω([a2, b2]b1, a3, a4), (b3 + b4)] − [ω(b1, [a2, b2]a3, a4), b4] − [ω(b1, a3, [a2, b2]a4), b3] + [ω([a2, b2]b1, b3, a4), a3] + [ω(b1, b3, [a2, b2]a4), a3] + [ω([a2, b2]b1, a3, b4), a4] + [ω(b1, [a2, b2]a3, b4), a4] The identity given by (6.62) is irreducible and can also be ignored, so we proceed to decompose (6.63) as follows. (6.64) Γ2134(a1 = 0, a3 6= 0, a4 6= 0, b3 = 0) = and (6.65) − [ω([a2, b2]b1, a3, a4), b4] − [ω(b1, [a2, b2]a3, a4), b4] + [ω(b1, b3, [a2, b2]a4), a3] + [ω([a2, b2]b1, a3, b4), a4] + [ω(b1, [a2, b2]a3, b4), a4], Γ2134(a1 = 0, a3 6= 0, a4 6= 0, b3 6= 0) = − [ω([a2, b2]b1, a3, a4), b3] + [ω([a2, b2]b1, b3, a4), a3] + [ω(b1, b3, [a2, b2]a4), a3]. By using (6.14), each of (6.64) and (6.65) gives the new identity (6.66) [ω(b1, b3, [a2, b2]a4), a3] = 0, which establishes (6.15), and at the same time shows that (6.57), (6.58) and (6.59) produce no new identities. This completes the analysis of (6.53), which has produced (6.15). Since (6.54) is obtained from (6.53) by interchanging the indices 3 and 2, no new information is provided by (6.54). Similarly, since (6.55) is obtained from (6.54) by interchanging the indices 3 and 4, no new information is provided by (6.55). Thus we have found all irreducible expressions which sum to Σ(a1 = 0), resulting in only one identity, namely (6.15). This completes the proof that (6.16)-(6.20) imply (6.13)-(6.15). (See the paragraph following (6.43).) It is now a simple matter to prove that, conversely, (6.13)-(6.15) imply (6.16)-(6.20). Note that by (6.15), (6.29), and (6.37),(6.39), Γijkl and ∆ijkl vanish, showing that Σ(a1 6= 0) = 0, hence (6.13)-(6.15) imply (6.20). Since earlier arguments have shown that • (6.14) ⇒ (6.21) ⇔ (6.16), • (6.13) = (6.22) ⇔ (6.17), • (6.14) ⇔ (6.27) ⇔ (6.18)⇔ (6.19), this completes the proof that (6.13)-(6.15) imply (6.16)-(6.20), and hence the proof of Theorem 5.5. References [1] R. Carlsson, Cohomology of associative triple systems. Proc. Amer. Math. Soc. 60 (1976), 1 -- 7. Erratum and supplement: Proc. Amer. Math. Soc. 67, no. 2 (1977), 361. [2] R. Carlsson, Der Wedderburnsche Hauptsatz fur alternative Tripelsysteme und Paare. Math. Ann. 228, no. 3 (1977), 233 -- 248. [3] H. Cartan and S. Eilenberg, Homological algebra. Princeton University Press, Princeton, N. J., 1956. xv+390 pp. [4] Chevalley, Claude; Eilenberg, Samuel Cohomology theory of Lie groups and Lie algebras. Trans. Amer. Math. Soc. 63, (1948). 85 -- 124. [5] C-H. Chu, Jordan structures in geometry and analysis. Cambridge Tracts in Math. 190, Cambridge Univ. Press, Cambridge, 2012. [6] Th. Fack and P. de la Harpe, Sommes de commutateurs dans les algbres de von Neumann finies continues. Ann. Inst. Fourier (Grenoble) 30, no. 3 (1980) 49 -- 73. [7] D. B. Fuks, Cohomology of infinite-dimensional Lie algebras. Translated from the Russian by A. B. Sosinski. Contemporary Soviet Mathematics. Consultants Bureau, New York, 1986. xii+339 pp. [8] M. Gerstenhaber, A uniform cohomology theory for algebras. Proc. Nat. Acad. Sci. U.S.A. 51 (1964), 626 -- 629. [9] N. D. Glassman, Cohomology of Jordan algebras. J. Algebra 15 (1970), 167 -- 194. COHOMOLOGY OF JORDAN TRIPLES VIA LIE ALGEBRAS 27 [10] N. D. Glassman, Cohomology of nonassociative algebras. Pacific J. Math. 33 (1970), 617 -- 634. [11] B. Harris, Cohomology of Lie triple systems and Lie algebras with involution. Trans. Amer. Math. Soc. 98 (1961), 148 -- 162. [12] G. Hochschild, On the cohomology groups of an associative algebra. Ann. of Math. (2) 46 (1945), 58 -- 67. [13] T. L. Hodge and B. J. Parshall, On the representation theory of Lie triple systems. Trans. Amer. Math. Soc. 354, no. 11 (2002), 4359 -- 4391. [14] N. Jacobson, General representation theory of Jordan algebras. Trans. Amer. Math. Soc. 70 (1951), 509 -- 530. [15] N. Jacobson, Jordan algebras. 1957 Report of a conference on linear algebras, June, 1956 pp 12 -- 19 National Academy of Sciences National Research Council, Washington, Public. 502 [16] N. Jacobson, Lie algebras. Interscience Tracts in Pure and Applied Mathematics, No. 10 Interscience Publishers (a division of John Wiley & Sons), New York-London 1962 ix+331 pp. [17] N. Jacobson, Structure and representations of Jordan algebras. American Mathematical Society Colloquium Pub- lications, Vol. XXXIX American Mathematical Society, Providence, R.I. 1968 x+453 pp. [18] R, V. Kadison, Derivations of operator algebras. Ann. of Math. (2) 83 (1966), 280 -- 293. [19] R. V. Kadison, Which Singer is that? Surveys in differential geometry, 347 -- 373, Surv. Differ. Geom., VII, Int. Press, Somerville, MA, 2000. [20] A. W. Knapp, Lie groups, Lie algebras, and cohomology. Mathematical Notes, 34. Princeton University Press, Princeton, NJ, 1988. xii+510 pp. [21] M. Koecher, Imbedding of Jordan algebras in Lie algebras. I, Amer. J. Math 89, no. 3 (1967), 787 -- 816. [22] M. Koecher, An elementary approach to bounded symmetric domains, Lecture Notes, Rice University, 1969. [23] O. Kuhn and A. Rosendahl, Wedderburnzerlegung fur Jordan-Paare. (German. English summary) Manuscripta Math. 24 (1978), no. 4, 403 -- 435. [24] K. McCrimmon, Representations of quadratic Jordan algebras. Trans. Amer. Math. Soc. 153 (1971), 279 -- 305. [25] K. McCrimmon, Compatible Peirce decompositions of Jordan triple systems. Pacific J. Math. 103 (1982), no. 1, 57 -- 102. [26] K. Meyberg, Lectures on algebras and triple systems. Notes on a course of lectures given during the academic year 1971 -- 1972. The University of Virginia, Charlottesville, Va., 1972. v+226 pp. [27] A. J. Penico, The Wedderburn principal theorem for Jordan algebras. Trans. Amer. Math. Soc. 70 (1951). 404 -- 420. [28] R. Pluta and B. Russo, Triple derivations on von Neumann algebras. Preprint 2014, arXiv:1309.3526 [29] B. Russo, Derivations and Projections on Jordan triples. Nonassociative algebra, continuous cohomology and quantum functional analysis, Proceedings of V CIDAMA, Almeria, Spain, September 12-16, 2011. World Scientific, to appear. [30] S. Sakai, Derivations of W ∗-algebras. Ann. of Math. (2) 83 (1966), 273 -- 279. [31] P. Seibt, Cohomology of algebras and triple systems. Comm. Algebra 3 (1975), no. 12, 1097 -- 1120. [32] P. Seibt, Review of [1], Mathematical Reviews MR0430026. [33] A. M. Sinclair, Jordan homomorphisms and derivations on semisimple Banach algebras. Proc. Amer. Math. Soc. 24 (1970), 209 -- 214. [34] A. Thom, A note on commutators in the Murray-von Neumann algebra., Preprint 2013. [35] C. A. Weibel, An introduction to homological algebra. Cambridge Studies in Advanced Mathematics, 38. Cambridge University Press, Cambridge, 1994. xiv+450 pp. [36] C. A. Weibel, History of homological algebra. History of topology, 797 -- 836, North-Holland, Amsterdam, 1999. [37] P. Zusmanovich, A converse to the Whitehead theorem, J. Lie Theory 18 (2008), 811-815. School of Mathematical Sciences, Queen Mary, University of London, London E1 4NS, UK E-mail address: [email protected] Department of mathematics, University of California, Irvine, USA E-mail address: [email protected]
1107.0552	1	1107	2011-07-04T07:46:05	Absolute continuity, Interpolation and the Lyapunov order	[ "math.OA" ]	We extend our Nevanlinna-Pick theorem for Hardy algebras and their representations to cover interpolation at the absolutely continuous points of the boundaries of their discs of representations. The Lyapunov order plays a crucial role in our analysis.	math.OA	math	ABSOLUTE CONTINUITY, INTERPOLATION AND THE LYAPUNOV ORDER PAUL S. MUHLY AND BARUCH SOLEL Abstract. We extend our Nevanlinna-Pick theorem for Hardy algebras and their representations to cover interpolation at the absolutely continuous points of the boundaries of their discs of representations. The Lyapunov order plays a crucial role in our analysis. 1. Introduction The celebrated theorem of Nevanlinna and Pick asserts that if n distinct points, z1, z2, . . . , zn, are given in the open unit disc D and if n other complex numbers are also given, w1, w2, . . . , wn, then there is a function f in the Hardy algebra H ∞(T), with norm at most one, such that f (zi) = wi, i = 1, 2, . . . , n, if and only if the Pick matrix (cid:18) 1 − wiwj 1 − zizj (cid:19)n i,j=1 is positive semidefinite. In [4, Theorem 5.3], we generalized this Nevanlinna-Pick theorem to the setting of Hardy algebras over W ∗-correspondences. Here we intend to push the work in [4] further, using tools developed in [2]. In a sense that we shall make precise, we show that there is a condition similar to the positivity of the Pick matrix that allows one to interpolate at "absolutely continuous" points of the boundaries of the domains considered in [4]. Before stating our main theorem, we want to see how one might try to extend the Nevanlinna-Pick theorem to cover families of absolutely continuous contraction operators along the lines suggested by [4, Theorem 6.1]. For this purpose, suppose H is a Hilbert space and Z := (Z1, Z2, . . . , Zn) is an n-tuple of operators in B(H). Then Z defines a completely positive operator ΦZ on the n × n matrices over B(H) via the formula Z1 Z2   ΦZ((aij )) : = =(Ziaij Z ∗ j ).     a11 a21 ... an1 a12 a22 an2 · · · a1n . . . · · · ann     . . . Zn Z1 Z2 ∗   . . . Zn 2000 Mathematics Subject Classification. Primary: 15A24, 46H25, 47L30, 47L55, Secondary: 46H25, 47L65. Key words and phrases. Nevanlinna-Pick interpolation, representations, Hardy algebra, abso- lute continuity, Lyapunov order. The research of both authors was supported in part by a U.S.-Israel Binational Science Foun- dation grant. The second author was also supported by the Technion V.P.R. Fund. 1 2 PAUL S. MUHLY AND BARUCH SOLEL If all the Zi's have norm less than 1, then I − ΦZ is an invertible map on Mn(B(H)) and (I −ΦZ)−1 is also completely positive. The following theorem, then, is a special case of [4, Theorem 6.1]. Theorem 1.1. Suppose Z1, Z2, · · · , Zn are n distinct operators in B(H), each of norm less than 1, and suppose W1, W2, · · · , Wn are n operators in B(H). Then there is an function f in H ∞(T), with supremum norm at most 1, such that f (Zi) = Wi, i = 1, 2, . . . , n, where f (Zi) is defined through the Riesz functional calculus, if and only if the Pick map (1) (I − ΦW ) ◦ (I − ΦZ )−1 defined on Mn(B(H)) is completely positive. Observe that when H is one-dimensional this theorem recovers the classical the- orem of Nevanlinna and Pick that we stated at the outset. Now Sz.-Nagy and Foiaş have shown that the proper domain for their H ∞-functional calculus is the collec- tion of all absolutely continuous contractions. One way to say that a contraction T is absolutely continuous is to say that when T is decomposed as T = Tcnu + U , where Tcnu is completely non-unitary and U is unitary, then the spectral measure of U is absolutely continuous with respect to Lebesgue measure on the circle. The content of [10, Theorems III.2.1 and III.2.3] is that a contraction T is absolutely continuous if and only if the H ∞(T)-functional calculus may be evaluated on T . It is therefore of interest to modify the hypothesis of Theorem 1.1 and ask for con- ditions that allow one to interpolate in the wider context where the variables Zi are assumed to be merely absolutely continuous contractions. In that setting the map I − ΦZ need no longer be invertible, and so it may not be possible to form the generalized Pick operator (I − ΦW ) ◦ (I − ΦZ)−1, let alone determine whether or not it is completely positive. However, there is a notion from matrix analysis, called the "Lyapunov order", which suggests a replacement for the condition that the Pick operator (I − ΦW ) ◦ (I − ΦZ)−1be completely positive. To formulate it we require an idea from the theory of completely positive maps that we analyzed in [2]. Definition 1.2. Let Φ be a completely positive map on a W ∗-algebra A. An element a ∈ A is called superharmonic for Φ in case a ≥ 0 and Φ(a) ≤ a; a is called pure superharmonic in case a is superharmonic and Φn(a) ց 0 as n → ∞. The superharmonic elements for a completely positive map evidently form a convex subset in the cone of all non-negative elements in the W ∗-algebra A. Definition 1.3. Let B be a W ∗-algebra and suppose A is a sub-W ∗-algebra of B. Suppose Φ : A → A is a completely positive map and that Ψ : B → B is also com- pletely positive. Then we say Ψ completely dominates Φ in the sense of Lyapunov in case every pure superharmonic element of Mn(A) for Φn is superharmonic for Ψn, where Φn (resp. Ψn) is the usual promotion of Φ (resp. Ψ) to Mn(A) (resp. Mn(B)). The following proposition links the notion of complete Lyapunov domination to the complete positivity of (1). Proposition 1.4. Suppose that A is a sub-W ∗-algebra of a W ∗-algebra B and suppose Φ and Ψ are completely positive maps on A and B, respectively. Assume that kΦk < 1, so I − Φ is invertible. Then the Pick operator, P := (I − Ψ) ◦ (I − ABSOLUTE CONTINUITY, INTERPOLATION AND THE LYAPUNOV ORDER 3 Φ)−1, is completely positive if and only if Ψ completely dominates Φ in the sense of Lyapunov. Proof. Note that the hypothesis that kΦk < 1 implies that every superharmonic element of A is pure superharmonic. Also note that it suffices to prove that P is positive if and only if {a ∈ A a ≥ 0, Φ(a) ≤ a} ⊆ {b ∈ B b ≥ 0, Ψ(b) ≤ b}, since the same argument will work for every n. Suppose, then, that P is positive and suppose that a ≥ 0 and Φ(a) ≤ a. Then (I − Φ)(a) ≥ 0. Consequently, 0 ≤ P ((I − Φ)(a)) = (I − Ψ)(a), showing that a ≥ Ψ(a). Suppose, conversely, that b ≥ 0. Then since kΦk < 1 and Φ is positive, (I − Φ)−1 = Pn≥0 Φn is positive. Consequently, a = (I −Φ)−1(b) is positive. But also, since (I −Φ)(a) = b is positive, a ≥ Ψ(a), by hypothesis. That is, P (b) = a − Ψ(a) ≥ 0, which is what we want to show. (cid:3) Our extension of Theorem 1.1 can now be formulated as Theorem 1.5. Suppose Z1, Z2, · · · , Zn are n distinct absolutely continuous con- tractions on a Hilbert space H and suppose W1, W2, · · · , Wn are n contractions on H, then there is a function f ∈ H ∞(T), of norm at most 1, such that f (Zi) = Wi, i = 1, 2, . . . , n, if and only if ΦW completely dominates ΦZ in the sense of Lya- punov. The technology we use to prove Theorem 1.5 works in the more general context of Hardy algebras over W ∗-correspondences, as we mentioned earlier. This is the arena in which our analysis takes place. But first, we must provide some background from [4, 2]. We shall follow terminology and most of the notation from [2]. In particular, we shall cite the second section of [2] for further background because it gives a fairly detailed birds-eye view of the theory as of 2010. We are very grateful to Nir Cohen for introducing us to the Lyapunov order. Acknowledgment 2. Background and the Main Theorem Throughout this note, M will denote a fixed W ∗-algebra. We will treat M as an abstract C ∗-algebra that is a dual space and we will not think of it as acting concretely on Hilbert space except through representations that we will specify. Also, E will denote a W ∗-correspondence over M . This means first that E is a right Hilbert C ∗-module over M which is self-dual. Consequently, the algebra of all bounded adjointable M -module maps on E, L(E), is all the bounded module maps and L(E) is a W ∗-algebra. To say that E is a W ∗-correspondence over M means, then, that there is a normal representation ϕ : M → L(E), making E a left M -module [2, Paragraph 2.2]. To eliminate technical digressions we assume that ϕ is faithful and unital. The tensor powers of E, E⊗n, will be the self-dual completions of the usual C ∗-Hilbert module tensor powers, and the Fock space F (E) will be the self-dual completion of the C ∗-direct sum of the E⊗n. Then F (E) is a W ∗-correspondence over M and we denote by ϕ∞ the left action of M on F (E) [2, Paragraph 2.7]. If ξ ∈ E, then Tξ will denote the creation operator it determines: Tξη := ξ ⊗ η, η ∈ F (E). The norm-closed subalgebra generated ϕ∞(M ) and {Tξ ξ ∈ E} is called the tensor algebra of E and will be denoted by 4 PAUL S. MUHLY AND BARUCH SOLEL T+(E) [2, Paragraph 2.7]. The ultra weak closure of T+(E) in L(F (E)) is called the Hardy algebra of E and is denoted H ∞(E) [2, Definition 2.1]. Suppose σ : M → B(Hσ) is a normal representation and let σE : L(E) → B(E ⊗σ Hσ) be the induced representation of L(E) in the sense of Rieffel [7, 8]: σE(T ) := T ⊗ I, T ∈ L(E). We write I(σE ◦ ϕ, σ) for the set of all operators C : E ⊗σ Hσ → Hσ that satisfy the equation Cσ ◦ ϕ(a) = σ(a)C for all a ∈ M ; i.e., I(σE ◦ ϕ, σ) denotes all the intertwiners of σE ◦ ϕ and σ. Also, we write D(E, σ) for the set of all elements of I(σE ◦ ϕ, σ) that have norm less than 1, and we write D(E, σ) for its norm closure. In [6] we proved Lemma 2.1. (See [2, Paragraph 2.8].) Given z ∈ D(E, σ) , define z × σ by z × σ(ϕ∞(a)) := σ(a) and z × σ(Tξ)(h) := z(ξ ⊗ h), a ∈ M , ξ ∈ E, and h ∈ Hσ. Then z × σ extends to a completely contractive (c.c.) representation of T+(E) on Hσ. Conversely, given a c.c. representation ρ of T+(E), then ρ = z × σ, where σ := ρ ◦ ϕ∞ and z(ξ ⊗ h) := ρ(Tξ)h. Further, for F ∈ H ∞(E), the B(Hσ)-valued function bFσ, defined on D(E, σ) by bFσ(z) := z × σ(F ), is bounded analytic and it extends to be continuous on D(E, σ) when F ∈ T+(E). Remark 2.2. We note here that our D(E, σ) is denoted D(Eσ)∗ in [2, Paragraph 2.8], where Eσ := I(σE ◦ ϕ, σ)∗ = I(σ, σE ◦ ϕ) is the σ-dual of E [2, Paragraph 2.6]. This dual space plays an important role in our theory, as we shall see, but we have opted for the change of notation in order to eliminate numerous unnecessary and often confusing adjoints from our formulas. Definition 2.3. A point z ∈ D(E, σ) and the representation z × σ are called abso- lutely continuous in case z × σ extends to be an ultra weakly continuous represen- tation of H ∞(E) in B(Hσ). We write AC(E, σ) for all the absolutely continuous points of D(E, σ). Our choice of terminology is inspired by the fact that when M = E = C, then z is absolutely continuous in our sense if and only if z, which is just an ordinary contraction operator on Hσ, is absolutely continuous in the sense described in the Introduction. In general, D(E, σ) ⊆ AC(E, σ) ⊆ D(E, σ), and both inclusions are proper. If M = E = C, and if σ is the one-dimensional representation of C on C, then D(E, σ) is the open unit disc in the complex plane and D(E, σ) = AC(E, σ). In every other setting of which we are aware, D(E, σ) ( AC(E, σ). Also, we know of no situation where D(E, σ) = AC(E, σ). We have been able to identify AC(E, σ) explicitly in numerous instances [2, Sections 4 and 5] and we know a lot about this space, but there is still much that remains mysterious. The σ-dual of E, Eσ := I(σ, σE ◦ ϕ), is important in this study for several reasons. The first is that it is a W ∗-correspondence over σ(M )′ in a very natural way. For ξ, η ∈ Eσ, hξ, ηi is defined to be ξ∗η - the product being the ordinary operator product, which makes sense as an operator on Hσ since ξ and η both map from Hσ to E ⊗σ Hσ. The actions of σ(M )′ on Eσ are given by the formula: a · ξ · b := (IE ⊗ a)ξb, a, b ∈ σ(M )′, ξ ∈ Eσ. Again, the products on the right hand side of the equation are ordinary operator products. The concept of the σ- dual of a W ∗-correspondence was formalized in [4], but it appeared, implicitly, in a number of places. A key role that it will play ABSOLUTE CONTINUITY, INTERPOLATION AND THE LYAPUNOV ORDER 5 here is in the identification of the commutant of an induced representation, which we will describe in the next section. But here we can already see its relevance for the present considerations by virtue of the following observation: Let z1, z2, . . . , zn be points in D(E, σ). Then they define a map Φz on the n × n matrices over σ(M )′ by the formula (2) Φz((aij )) := (hzi, aij · zj i) (aij ) ∈ Mn(σ(M )′). A moment's reflection reveals that Φz is completely positive, as it is the composition of manifestly completely positive maps. Our objective in this note is the proof of the following theorem, which will occupy the next section. Theorem 2.4. Suppose E is a W ∗-correspondence over a W ∗-algebra M and that σ is a faithful normal representation of M on the Hilbert space Hσ. Suppose, too, that n distinct points z1, z2, . . . , zn ∈ AC(E, σ) are given and that n operators in B(Hσ), W1, W2, · · · , Wn, are given. Define the map Φz on Mn(σ(M )′) by the formula Φz((aij )) = (hzi, aij · zji) and define the map ΦW on Mn(B(Hσ)) by the j(cid:1). Then there is an element F in H ∞(E), with formula ΦW ((Tij)) := (cid:0)WiTijW ∗ kF k ≤ 1, such that bF (zi) = Wi, i = 1, 2, . . . , n, if and only if ΦW completely dominates Φz in the sense of Lyapunov. Proof of Theorem 1.5. That theorem is an immediate consequence of Theorem 2.4. Indeed, in the setting of the former, M = E = C, and σ is just a multiple of the identity representation, the multiple being the Hilbert space dimension of Hσ. Since we may safely identify C ⊗σ Hσ with Hσ, D(E, σ) may be identified with the closed unit ball in B(Hσ), i.e., with all contractions on Hσ. As we noted, the celebrated theorems of Sz.-Nagy and Foiaş identify AC(C, σ) with the absolutely continuous contractions on Hσ in the classical sense. When these identifications are made, Φz of Theorem 2.4 becomes the ΦZ of Theorem 1.5 and, of course, the two ΦW 's are the same. (cid:3) 3. The Proof of Theorem 2.4 We will first show that if ΦW completely dominates Φz in the sense of Lyapunov, then we can find an interpolating F ∈ H ∞(E) of norm at most 1. The route we shall follow is similar, in certain respects, to the route followed in the proof of [4, Theorem 5.3] and is based, ultimately, on the commutant lifting approach to the classical Nevanlinna-Pick theorem pioneered by Sarason [9]. For this purpose, we need another way to express Lyapunov dominance that reflects the fact that the zi's involved all lie in AC(E, σ). The key tool in our approach is the notion of an induced representation for T+(E) and the connection such representations have with the concept of absolute continuity. They are defined as follows: Let τ be a normal representation of M on the Hilbert space Hτ . Then we may induce τ to F (E), obtaining a normal representation τ F (E) of L(F (E)) on the Hilbert space F (E) ⊗τ Hτ . The restriction of τ F (E) to T+(E), then, is called the representation of T+(E) induced by τ . It is clearly an absolutely continuous representation of T+(E), since H ∞(E) is contained in L(F (E)) by definition and τ F (E) is ultraweakly continuous. As we showed in [2], and will discuss in a moment, induced representations are the architypical absolutely continuous representations. We continue to use the notation τ F (E) for its restrictions to T+(E) and H ∞(E). 6 PAUL S. MUHLY AND BARUCH SOLEL In [5] we develop at length the analogies between induced representations of T+(E) and H ∞(E) and unilateral shifts. Indeed, a unilateral shift arises from an induced representation of T+(E), where M = E = C. Definition 3.1. Let π be a faithful representation of M on Hπ and assume that π has infinite multiplicity. Then πF (E) is called the universal induced representation of T+(E) and H ∞(E) determined by π. Any two faithful π's with infinite multiplicity give unitarily equivalent induced representations. Further, every induced representation of T+(E) is unitarily equiv- alent to a subrepresentation of πF (E) obtained by restricting πF (E) to a subspace of the form F (E) ⊗π K, where K is a subspace of Hπ that reduces π [2, Paragraphs 2.5 and 2.11]. This explains the terminology, allowing us to use the definite article. The representation π and the induced representation πF (E) will be fixed for the remainder of this note. We also shall extend the notation I(σE ◦ ϕ, σ), and write I(ρ1, ρ2) for the set of operators C : Hρ1 → Hρ2 that intertwine ρ1 and ρ2, where ρ1 and ρ2 are any two completely contractive representations of T+(E). If ρi is written as zi × σi, i = 1, 2, then it is easy to see that an operator C : Hρ1 → Hρ2 lies in I(ρ1, ρ2) if and only if C ∈ I(σ1, σ2) and z2(IE ⊗ C) = C z1. An important linkage among the universal induced representation, intertwiners, and absolute continuity is the following theorem. Theorem 3.2. [2, Theorem 4.7] A point z ∈ D(E, σ) is absolutely continuous if and only if _{Ran(c) c ∈ I(πF (E), z × σ)} = Hσ. Further, for F ∈ H ∞(E), z ∈ AC(E, σ), and c ∈ I(πF (E), z × σ), (3) bF (z)c = cπF (E)(F ). Proof. The first assertion is explicitly in [2] as Theorem 4.7. The second assertion is easily checked on generators of H ∞(E) of the form ϕ∞(a), a ∈ M , and Tξ, ξ ∈ E. That is all that is necessary to check. (cid:3) Recall that if z ∈ D(E, σ), then z ∗ lies in the W ∗-correspondence Eσ over σ(M )′. It therefore defines a completely positive map Θz on σ(M )′ by the formula Θz(a) = hz ∗, a · z ∗i = z(IE ⊗ a)z ∗, a ∈ σ(M )′. Indeed, Θz is just a special case of the map Φz in the statement of Theorem 2.4. We are going to use the following theorem from [2] to obtain an alternate formulation of the complete Lyapunov dominance assertion in that theorem. Theorem 3.3. [2, Theorem 4.6] If z ∈ D(E, σ), then an operator q ∈ σ(M )′ is a pure superharmonic operator for Θz if and only if q can be written as q = cc∗ for an element c ∈ I(πF (E), z × σ). Corollary 3.4. We adopt the notation of Theorem 2.4. The map ΦW completely dominates Φz in the sense of Lyapunov if and only if the following condition is satisfied: For every integer m ≥ 1, for every choice of function l : {1, 2, . . . , m} → {1, 2, . . . , n}, and for any choice of m operators cj ∈ I(πF (E), zl(j) × σ) the operator matrix inequality (4) (Wl(i)cic∗ j W ∗ l(j))m i,j=1 ≤ (cic∗ j )m i,j=1 ABSOLUTE CONTINUITY, INTERPOLATION AND THE LYAPUNOV ORDER 7 is satisfied. σ Proof. Fix m and a function l : {1, 2, . . . , m} → {1, 2, . . . , n}. Write H (m) for the direct sum of m copies of Hσ and let σm be the inflated representation of M on H (m) , i.e., σm(a) := diag{a, a, · · · , a}. Then σm(M )′ = Mm(σ(M )′). Also, write E(m) for the direct sum of copies of E, which is also a W ∗-correspondence over M in the obvious way, and set z := (zl(1), zl(2), · · · , zl(m)). Then we may view z as a map from E(m) ⊗σm H (m) , which clearly belongs to I(σE(m) ◦ ϕ, σm). Consequently, z defines a completely positive map Θz on σm(M )′ = Mm(σ(M )′), and it is easy to see that σ = (E ⊗σ Hσ)(m) to H (m) m σ σ Θz((bl(i),l(j)))m i,j=1 =(cid:16)zl(i)(IE ⊗ bl(i),l(j))z ∗ l(j)(cid:17) , (bl(i),l(j))m i,j=1 ∈ Mm(σ(M )′). Moreover, by Theorem 3.3, the pure superharmonic elements Mm(σ(M )′) for Θz i,j=1, where ci ∈ I(πF (E), (zl(i) × σ)). On the other hand, the are of the form (cic∗ Wi's may be used to define the completely positive map ΨW,l on Mm(σ(M )′) by the formula j )m ΨW,l((bl(i),l(j))) :=(cid:16)Wl(i)bl(i),l(j)W ∗ l(j)(cid:17) , (bl(i),l(j))m i,j=1 ∈ Mm(σ(M )′). The inequality 4 is the statement that the condition of the corollary is equivalent to the assertion that ΨW,l dominates Θz in the sense of Lyapunov for each choice of m and l. It is now evident that the condition of the corollary implies that ΦW completely dominates Φz in the sense of Lyapunov by choosing m and l judiciously. On the other hand, if ΦW completely dominates Φz, then given m and l, one can clearly choose a k so that the domination of (Φz)kby (ΦW )k in the sense of Lyapunov gives the desired inequalities of the condition for that m and l. (cid:3) In order to follow the commutant lifting approach pioneered by Sarason, we require the description of the commutant of πF (E)(H ∞(E)) that we developed in [4]. The description there works for any induced representation, but we formulate it here specifically for πF (E). Theorem 3.5. [4, Theorem 3.9] Write ι for the identity representation of π(M )′ on Hπ, and let τ be the induced representation of L(Eπ) acting on F (Eπ) ⊗ι Hπ, i.e., let τ = ιF (Eπ). Then the map U : F (Eπ) ⊗ι Hπ → F (E) ⊗π Hπ, defined by the formula (5) U (ξ1 ⊗ ξ2 ⊗ · · · ⊗ ξn ⊗ h) := (IE⊗(n−1) ⊗ ξ1)(IE⊗(n−2) ⊗ ξ2) · · · (IE ⊗ ξn−1)ξn(h), ξ1 ⊗ ξ2 ⊗ · · · ⊗ ξn ⊗ h ∈ (Eπ)⊗n ⊗ι Hπ, is a Hilbert space isomorphism and U τ (H ∞(Eπ))U ∗ = πF (E)(H ∞(E))′. Likewise, U ∗πF (E)(H ∞(E))U = τ (H ∞(Eπ))′, and the double commutant relations hold: πF (E)(H ∞(E))′′ = πF (E)(H ∞(E)), and τ (H ∞(Eπ))′′ = τ (H ∞(Eπ)). We are now ready to show how the complete domination of Φz in the sense of Lyapunov by ΦW implies that we can interpolate the W 's at the z's in Theorem 2.4. 8 PAUL S. MUHLY AND BARUCH SOLEL Lemma 3.6. Let M = span{U ∗c∗h h ∈ Hσ, c ∈ I(πF (E), zi × σ), 1 ≤ i ≤ n}. Then M is a closed subspace of F (Eπ) ⊗ι Hπ that is invariant under τ (H ∞(Eπ))∗. Proof. For X ∈ τ (H ∞(Eπ)), U τ (X)U ∗ lies in the commutant of πF (E)(H ∞(E)) by Theorem (3.5). Consequently cU τ (X)U ∗ ∈ I(πF (E), zi × σ) for every c ∈ I(πF (E), zi × σ). But then τ (X)∗U ∗c∗h = U ∗(U τ (X)∗U ∗)c∗h = U ∗(cU τ (X)U ∗)∗h lies in M for all U ∗c∗h ∈ M. (cid:3) i , c ∈ I(πF (E), zi × σ), de- Lemma 3.7. The correspondence, U ∗c∗h → U ∗c∗W ∗ fined on the generators of M extends to a well-defined contraction operator on M, say R, if and only if for every integer m ≥ 1, for every choice of func- tion l : {1, 2, . . . , m} → {1, 2, . . . , n}, and for every choice of m operators cj ∈ I(πF (E), zl(j) × σ) the operator matrix inequality (Wl(i)cic∗ j W ∗ l(j))m i,j=1 ≤ (cic∗ j )m i,j=1 is satisfied. In this event, R commutes with the restriction of τ (H ∞(Eπ))∗to M. Proof. A linear combination of generators of M is a vector of the form k = j=1 U ∗c∗ j hj, where cj ∈ I(πF (E), zl(j)×σ) for some m and function l : {1, 2, . . . , m} → Pm {1, 2, . . . n}. Since hcic∗ j hj, hii, kkk2 =Xj,i l(j)hjk2 =Xi,j U ∗c∗ j W ∗ hWi(l)cic∗ j W ∗ i(j)hj, hli, while k mXj=1 the first assertion is immediate. But the second is also immediate since R is "right i on a generator of the form U ∗c∗h, c ∈ I(πF (E), zi × σ), i.e., multiplication" by W ∗ i h, while the restriction of τ (X)∗ to M acts by left multiplication RU ∗c∗h = U ∗c∗W ∗ for all X ∈ H ∞(Eπ): τ (X)∗U ∗c∗h = U ∗(U τ (X)∗U ∗)c∗h. (cid:3) Since M is invariant for τ (H ∞(Eπ))∗, we obtain an ultra weakly continuous com- pletely contractive representation ρ of H ∞(Eπ) on M by compressing τ (H ∞(Eπ)) to M, i.e., ρ(X) := PMτ (X)M, X ∈ H ∞(Eπ). Since τ is isometric in the sense of [6] and since R∗ commutes with ρ(H ∞(Eπ)), we may apply our commutant lifting theorem [6, Theorem 4.4] to conclude that there is an operator Y ∈ B(F (Eπ) ⊗ι Hπ) of norm at most one such that PMY M = R∗, Y M⊥ ⊆ M⊥, and Y commutes with τ (H ∞(Eπ)) (see [2, Theorem 2.6], also). By Theorem 3.5, there is an F ∈ H ∞(E), kF k ≤ 1, such that Y = U ∗πF (E)(F )U . We conclude from the properties of Y and the definition of R that U ∗πF (E)(F )∗c∗h = (U ∗πF (E)(F )∗U )U ∗c∗h = Y ∗U ∗c∗h = RU ∗c∗h = U ∗c∗W ∗ i h for all c ∈ I(πF (E), zi × σ). This, in turn, implies that cπF (E)(F ) = Wic for all such c. But cπF (E)(F ) = bF (zi)c for all c ∈ I(πF (E), zi × σ), by equation (3) in Theorem 3.2. Therefore, bF (zi)c = Wic ABSOLUTE CONTINUITY, INTERPOLATION AND THE LYAPUNOV ORDER 9 for all i and all c ∈ I(πF (E), zi × σ). However, by hypothesis, all the zi lie in AC(E, σ). Consequently, by the first assertion of Theorem 3.2, the closed span of the ranges of the c's in I(πF (E), zi × σ) is all of Hσ, for every i. We conclude that bF (zi) = Wi for every i. This completes the proof that if ΦW completely dominates Φz in the sense of Lyapunov, then there is an F ∈ H ∞(E) that interpolates Wi at zi. Proof of the Converse. Part of the argument just given is reversible. Suppose F is an element of H ∞(E) of norm at most one such that bF (zi) = Wi, i = 1, 2, · · · , n. Then for each c ∈ I(πF (E), zi × σ) by equation (3). But then cπF (E)(F ) = bF (zi)c = Wic, (U ∗πF (E)(F )∗U )U ∗c∗ = U ∗c∗W ∗ i for all c ∈ I(πF (E), zi × σ). Since the norm of F is at most 1 we conclude from Lemma 3.7 that for every integer m ≥ 1, for every choice of function l : {1, 2, . . . , m} → {1, 2, . . . , n}, and for every choice of m operators cj ∈ I(πF (E), zl(j)× σ) the operator matrix inequality (Wl(i)cic∗ i,j=1 ≤ (cic∗ j )m j W ∗ l(j))m i,j=1 is satisfied. So by Corollary 3.4, we conclude that ΦW completely dominates Φz in the sense of Lyapunov. (cid:3) References [1] N. Cohen and I. Lewkowicz, The Lyapunov order for real matrices, Linear Algebra and Its Applications, 430, (2009), 1849-1866. [2] P. S. Muhly and B. Solel, Representations of Hardy algebras: absolute continuity, inter- twiners, and superharmonic operators, Integral Equations and Operator Theory 70 (2011), 151-203 (arXiv:1006.1398). [3] P. S. Muhly and B. Solel, Schur Class Operator Functions and Automorphisms of Hardy Algebras, Documenta Math. 13 (2008), 365 -- 411. [4] P. S. Muhly and B. Solel, Hardy algebras, W ∗-correspondences and interpolation theory, Math. Ann. 330 (2004), 353-415. [5] P. S. Muhly and B. Solel , Tensor algebras, induced representations, and the Wold decompo- sition, Canad. J. Math. 51 (1999), 850-880. [6] P. S. Muhly and B. Solel, Tensor algebras over C ∗-correspondences (Representations, dila- tions, and C ∗-envelopes), J. Functional Anal. 158 (1998), 389 -- 457. [7] M. Rieffel, Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure Appl. Alg. 5 (1974), 51 -- 96. [8] M. Rieffel, Induced representations of C ∗-algebras, Adv. in Math. 13 (1974), 176 -- 257. [9] D. Sarason, Generalized interpolation in H∞, Trans. Amer. Math. Soc. 127 (1967), 179 -- 203. [10] B. Sz.-Nagy, C. Foiaş, H. Bercovici, and L. Kérchy, Harmonic Analysis of Operators on Hilbert Space, Springer, New York, 2010. Department of Mathematics, University of Iowa, Iowa City, IA 52242 E-mail address: [email protected] Department of Mathematics, Technion, 32000 Haifa, Israel E-mail address: [email protected]
1807.11425	3	1807	2019-11-26T04:55:06	The non-selfadjoint approach to the Hao-Ng isomorphism	[ "math.OA", "math.FA" ]	In an earlier work, the authors proposed a non-selfadjoint approach to the Hao-Ng isomorphism problem for the full crossed product, depending on the validity of two conjectures stated in the broader context of crossed products for operator algebras. By work of Harris and Kim, we now know that these conjectures in the generality stated may not always be valid. In this paper we show that in the context of hyperrigid tensor algebras of C-correspondences, each one of these conjectures is equivalent to the Hao-Ng problem. This is accomplished by studying the representation theory of non-selfadjoint crossed products of C-correspondence dynamical systems; in particular we show that there is an appropriate dilation theory. A large class of tensor algebras of C-correspondences, including all regular ones, are shown to be hyperrigid. Using Hamana's injective envelope theory, we extend earlier results from the discrete group case to arbitrary locally compact groups; this includes a resolution of the Hao-Ng isomorphism for the reduced crossed product and all hyperrigid C-correspondences. A culmination of these results is the resolution of the Hao-Ng isomorphism problem for the full crossed product and all row-finite graph correspondences; this extends a recent result of Bedos, Kaliszewski, Quigg and Spielberg.	math.OA	math	THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM ELIAS G. KATSOULIS AND CHRISTOPHER RAMSEY Abstract. In an earlier work, the authors proposed a non-self- adjoint approach to the Hao-Ng isomorphism problem for the full crossed product, depending on the validity of two conjectures stated in the broader context of crossed products for operator algebras. By work of Harris and Kim, we now know that these conjec- tures in the generality stated may not always be valid. In this paper we show that in the context of hyperrigid tensor algebras of C∗-correspondences, each one of these conjectures is equiva- lent to the Hao-Ng problem. This is accomplished by studying the representation theory of non-selfadjoint crossed products of C∗-correspondence dynamical systems; in particular we show that there is an appropriate dilation theory. A large class of tensor alge- bras of C∗-correspondences, including all regular ones, are shown to be hyperrigid. Using Hamana's injective envelope theory, we extend earlier results from the discrete group case to arbitrary locally compact groups; this includes a resolution of the Hao-Ng isomorphism for the reduced crossed product and all hyperrigid C∗- correspondences. A culmination of these results is the resolution of the Hao-Ng isomorphism problem for the full crossed product and all row-finite graph correspondences; this extends a recent result of Bedos, Kaliszewski, Quigg and Spielberg. 1. Introduction Let ((X,C),G, α) be a C∗-correspondence dynamical system where G is a locally compact group and α is a generalized gauge action. This action can be extended uniquely to the Cuntz-Pimsner algebra OX The Hao-Ng isomorphism problem asks whether OX ⋊α G ≃ OX⋊αG in the reduced or full crossed products. This problem is named after Hao and Ng who proved the validity of this formula when G is amenable 2010 Mathematics Subject Classification. 46L07, 46L08, 46L55, 47B49, 47L40, 47L65. Key words and phrases: C∗-correspondence, crossed product, Cuntz-Pimsner algebra, tensor algebra, Hao-Ng isomorphism, C∗-envelope, operator algebra. 1 2 E.G. KATSOULIS AND C. RAMSEY [14, Theorem 2.10]. However, this formula was first studied by Abadie in the context of Takai duality for equivalence bimodules. Indeed, in Abadie's proof for the Takai duality, the Hao-Ng isomorphism forms the crucial step of the proof and corresponds to the key isomorphism of [34, Lemma 7.2] in the classical case. In general, the Hao-Ng isomorphism has proved to be a significant stimulant to research as versions of it appear in many different contexts, e.g. in Schafhauser's work [33] on AF-embedability, or in Deaconu's work [7, 8] on group actions on graph C∗-algebras. In its full generality, the problem remains open and under investigation by several authors [3, 17, 18, 24, 27]. The authors initiated a study in [22, 20, 23] of non-selfadjoint crossed products of operator algebra dynamical systems (A,G, α) where α acts by completely isometric isomorphisms of A. The main thrust of [22, Chapter 7] and [20] is that the Hao-Ng isomorphism problem can and should be thought of as a non-selfadjoint problem. For the reduced crossed product this kind of approach has been and continues to be quite successful. For instance, we now know that the Hao-Ng iso- morphism for the reduced crossed product holds for all discrete groups [20], a fact that resolves an open problem from [3] (and more is ac- complished in this paper). The Hao-Ng isomorphism for the full crossed product seems to be In [22] we envisioned the following line of a much harder problem. attack. First one verifies (1) C∗ env(A ⋊α G) ≃ C∗ env(A) ⋊α G in the full crossed product case for an arbitrary non-selfadjoint dynami- cal system (A,G, α); this is Problem 1 in [22]. Subsequently, one solves Problem 2 in [22] by showing that all relative crossed products coin- cide. Assuming that both problems have been resolved in the positive, now one specializes on tensor algebra dynamical systems and obtains (2) T + X ⋊α G ≃ T + X⋊αG by invoking the solution of Problem 2 and the remarks following [22, Theorem 7.13]. Recalling that C∗ X ) = OX , one recovers now the Hao-Ng isomorphism by combining equations (1) and (2). Note that even though a positive answer for both Problems 1 and 2 leads to a positive resolution for the Hao-Ng isomorshism, the exact relation of each one of these problems with the Hao-Ng isomorphism was never clarified in [22]. env(T + The central result of this paper, Theorem 4.9, clarifies that rela- tion and shows that the Hao-Ng problem actually leads to equivalent statements in non-selfadjoint operator algebra theory, whose validity THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 3 or refutation will therefore resolve the isomorphism. Specifically, for a non-degenerate C∗-correspondence X we show that validity of (1) for A = T + X is equivalent to the validity of the Hao-Ng isomorphism OX ⋊αG ≃ OX⋊αG. In addition, for a large class of C∗-correspondences, including all regular ones, we show that the validity of the Hao-Ng iso- morphism OX ⋊α G ≃ OX⋊αG is equivalent to the fact that all relative crossed products for (T + X ,G, α) coincide, where α is a generalized gauge action. (For a general C∗-correspondence X this last statement is im- plied by any of the previous two.) Theorem 4.9 relates to exciting new work by Harris and Kim [15]. Indeed these authors have answered both Problems 1 and 2 from [22, Chapter 7] by producing finite dimensional, hyperrigid dynamical sys- tems (A,G, α) with distinct relative crossed products and failing (1). However the examples of Harris and Kim [15] do not concern tensor algebras of C∗-correspondences and so the Hao-Ng problem remains open. Theorem 4.9 shows now that the resolution of the Hao-Ng prob- lem will lead to or will follow from the existence or the absence of Harris-Kim type examples but in the realm of tensor algebras. Need- less to say that the quest for such examples, or the refutation of their existence, becomes now a project of high priority. To test our new results, we study the Hao-Ng isomorphism for a class of C∗-correspondences that plays a central role in the theory: graph C∗-correspondences. In Theorem 5.4 we show that the Hao-Ng isomorphism problem is true in the case of row-finite graph correspon- dences, thus showing that the crossed product of such a Cuntz-Krieger algebra is the Cuntz-Pimsner algebra of a crossed product (of a graph) correspondence. This is done by showing that in the case of a dynam- ical system (A,G, α) where A is the tensor algebra of any graph, G any locally compact group and α a generalized gauge action, all rela- tive crossed products coincide. Then, Theorem 4.9 finishes the proof for row-finite graphs. Note that in the special case where G is dis- crete, Theorem 5.4 has also been obtained independently by Bedos, Kaliszewski, Quigg and Spielberg using different methods [4, Corol- lary 6.8 and Remark 6.10]. It is worth mentioning here that Theorem 5.4 is essentially obtained by dilating representations in a wholly con- structive manner and should prove of much interest to those who study the representation theory of C∗-correspondences. At the moment, the lack of a constructive dilation proof of C∗ X ) = OX seems to be a barrier to establishing (2) for the full crossed product in general. On the way to proving the above theorems we obtain several results of independent interest. First we resolve Problem 3 from our monograph env(T + 4 E.G. KATSOULIS AND C. RAMSEY X [22, Chapter 8]. Specifically, we show if (X,C) is a non-degenerate C∗- correspondence and α : G → (X,C) is the generalized gauge action of a locally compact group, then T + ⋊α G is necessarily the tensor algebra of some C∗-correspondence. Another result of independent interest is Theorem 3.1, which identi- fies a large class of hyperrigid C∗-correspondences, i.e., C∗-correspon- dences whose tensor algebras are hyperrigid. Indeed our central Theo- rem 4.9 actually applies to all hyperrigid C∗-correspondences. To make that result usable, we show that any C∗-correspondence (X,C, ϕX ) with ϕX(JX)X = X is hyperrigid (here JX denotes Katsura's ideal). This includes all previous known examples of hyperrigid C∗-correspondences and many more, e.g., all regular ones. An interesting byproduct of our techniques on the full crossed prod- uct version of the Hao-Ng problem, is the resolution of the same prob- lem for the reduced crossed product and all hyperrigid C∗-correspon- dences. In [20] the first named author verified the Hao-Ng isomorphism for the reduced crossed product and all discrete groups. Because here we are addressing locally compact groups which may not be discrete, we have to use an approach different from that of [20]. In particular, the algebra A does not embed in either A⋊α G or A⋊r α G and so restricting a maximal map of the crossed product on the core algebra A (as we did in [20]) is no longer an option. Instead we use Hamana's injective envelope theory, an approach towards the Hao-Ng isomorphism which is used in this paper for the first time. This approach was adopted after illuminating discussions with S. Echterhoff and we are grateful to him for that. We denote by N the set of positive integers, while Z+ 0 = N∪{0}. We denote by span{· · ·} the closure of the linear span of {· · ·}. An ideal of a C∗-algebra always means a closed two-sided ideal. 2. Crossed products and C∗-covers Let (A,G, α) be an operator algebra dynamical system, meaning that A is an approximately unital operator algebra and G is a locally compact (Hausdorff) group acting continuously on A by completely isometric automorphisms, α : G → Aut(A). The aim of this section is to better understand the relationship of α-admissible C∗-covers. Recall that a C∗-cover (C, ι) of A is a C∗-algebra C and a complete isometry ι : A → C such that C ∗(ι(A)) = C. The two nicest C∗-covers of A are the "biggest" and the "smallest" covers C∗ env(A). These are defined by their universal properties. Namely, whenever (C, ι) is a C∗-cover there are (unique) max(A) and C∗ THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 5 surjective ∗-homomorphisms ϕ : C∗ env(A) such that ϕ(a) = ι(a) and ψ(ι(a)) = a, for all a ∈ A. From [22], a C∗-cover (C, ι) is called α-admissible if there exists a group representation β : G → Aut(C) acting on C by ∗-automorphisms such that max(A) → C and ψ : C → C∗ βs(ι(a)) = ι(αs(a)), ∀s ∈ G, a ∈ A. In [22, Lemma 3.3] we established that both C∗ max(A) are always α-admissible. However, in [22] we did not provide any examples of C∗-covers which fail to be α-admissible. We thank David Sherman for bringing this to our attention and asking us whether such covers do exist. env(A) and C∗ Proposition 2.1. Not all C∗-covers are α-admissible. Proof. Let C = C(T)⊕M2 and ι : A(D) → C be given by z 7→ z⊕[ 0 0 1 0 ]. By von Neumann's inequality it is straightforward that ι is a complete isometry. Now ι(z) − ι(z2)ι(z)∗ = (z ⊕ [ 0 0 1 0 ]) − (z2 ⊕ [ 0 0 = 0 ⊕ [ 0 0 1 0 ] . 0 0 ])(¯z ⊕ [ 0 1 0 0 ]) Thus, C ∗(ι(A(D))) = C and (C, ι) is a C∗-cover of A(D). Consider the Mobius transformation ϕ(z) = z− 1 2 1− z 2 which gives ϕ ∈ Aut(D). From this define the dynamical system (A(D), α, Z) where αn(f ) = f ◦ ϕn which is the same as z 7→ ϕn(z). It is well known that composition with a Mobius map is a completely isometric auto- morphism of the disc algebra. Suppose that there exists α : Z → Aut(C) such that αn(i(f )) = i(αn(f )),∀f ∈ A(D). Calculating =(cid:18)z − = − z − 1 1 − z ϕ(z) = 1 2 2 2 + 1 2(cid:19)(cid:18)1 + + z2 z 2 4 z2 + · · · 3 8 3 4 z + + · · ·(cid:19) we get that ι(ϕ(z)) = ϕ(z) ⊕h − 1 3 2 0 4 − 1 2i . 6 E.G. KATSOULIS AND C. RAMSEY Hence, α1 (0 ⊕ [ 0 0 1 0 ]) = α1(ι(z) − ι(z2)ι(z)∗) 2i −(cid:16)ϕ(z)2 ⊕h 1 4 − 3 4 0 1 4i(cid:17)(cid:16)ϕ(z) ⊕h − 1 3 2 4 0 − 1 2i(cid:17) 2 3 8 − 3 0 4 − 1 = ι(α1(z)) − ι(α1(z2))ι(α(z))∗ = ι(ϕ(z)) − ι(ϕ(z))2ι(ϕ(z))∗ = ϕ(z) ⊕h − 1 = 0 ⊕h − 3 16 i . = √2(cid:13)(cid:13)(cid:13)h 3 16i(cid:13)(cid:13)(cid:13) 16 i(cid:13)(cid:13)(cid:13) 8 − 3 = 16 3 16 3 3 8 8 3 3√10 16 But then 0 ⊕h − 3 3 8 (cid:13)(cid:13)(cid:13) < 1 = k0 ⊕ [ 0 0 1 0 ]k , a contradiction as ∗-automorphisms are isometric. Therefore, no such α exists and (C, ι) is a non α-admissible C∗-cover of (A(D), α, Z). If we do have an α-admissible cover then we can abuse the notation and call the group representation α again because of the next result. Lemma 2.2. Let (A,G, α) be an operator algebra dynamical system. If (C, ι) is an α-admissible C∗-cover then there is a unique group rep- resentation of G on C acting by ∗-automorphisms extending α. Proof. Let β1 and β2 be two such extensions. Then for any s ∈ G we have β1,s ◦ β−1 2,s = id on A and we just need to prove this is the identity map. To this end assume that β : G → Aut(C) extends the identity map on A. That is, βsA = idA, for all s ∈ G. max(A) there is a unique surjective ∗- homomorphism ϕ : C∗ max(A) → C such that ϕ(a) = ι(a), for all a ∈ A. Notice that βs ◦ ϕ(a) = βs ◦ ι(a) = ι(a), for all a ∈ A. Thus, βs ◦ ϕ = ϕ by uniqueness which implies that βs = id on C, for any s ∈ G. By the universal property of C∗ Now we turn to crossed products of non-selfadjoint operator algebras. Definition 2.3 ([22]). Let (A,G, α) be an operator algebra dynamical system and let (C, ι) be an α-admissible C∗-cover. The relative reduced and full crossed products are denoted by A ⋊r (C,ι),α G and A ⋊(C,ι),α G and are defined to be the closure of Cc(G,A) in C ⋊r α G and C ⋊α G, respectively. All relative reduced crossed products are in fact completely isomet- rically isomorphic [22, Theorem 3.12] and so we define the reduced crossed product, denoted A ⋊r α G, to be this unique object. Lastly, we define the full crossed product to be A ⋊α G := A ⋊C∗ max(A),α G. THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 7 In fact, this is the universal algebra for all covariant representations of (A,G, α) [22, Proposition 3.7]. Finally, it should be noted that, as in the selfadjoint case, if G is amenable then the full and reduced crossed products coincide [22, Theorem 3.14]. Now we are able to state and prove the main theorem of this section. Theorem 2.4. Let (A,G, α) be an operator algebra dynamical system. Then for every α-admissible C∗-cover (C, ι) there are surjective com- pletely contractive homomorphisms A ⋊α G env(A),α G env(A), (C, ι) max(A) respectively αenv, αC and αmax. By the universal proper- qmax−−−→ A ⋊(C,ι),α G qmin−−→ A ⋊C∗ such that they are just the identity on Cc(G,A). Proof. Label the unique extensions of α to the C∗-covers C∗ and C∗ ties there exists surjective ∗-homomorphisms env(A) and ϕmax : C∗ max(A) → C such that ϕenv(ι(a)) = a and ϕmax(a) = ι(a), for all a ∈ A. the following commutative diagram: By uniqueness of the quotient maps and of the extensions we have ϕenv : C → C∗ ϕmax−−−→ C ϕenv−−−→ C∗ env(A) ϕmax−−−→ C ϕenv−−−→ C∗ env(A) αenvy C∗ max(A) αmaxy C∗ max(A) αCy Thus, ker ϕmax is an αmax-invariant ideal and ker ϕenv is an αC-invariant ideal. By [34, Proposition 3.19], full C∗-crossed products preserve exact sequences by α-invariant ideals. Hence, we have the following surjective ∗-homomorphisms max(A) ⋊αmax G ϕmax⋊id C∗ −−−−−→ C ⋊αC G ϕenv⋊id env(A) ⋊αenv G. −−−−→ C∗ So qmax = ϕmax ⋊ idA⋊αG and qmin = ϕenv ⋊ idA⋊(C,ι),αG are completely contractive homomorphisms which amount to the iden- tity on Cc(G,A). The benefit of this theorem, as will be used later, is that one needs only to show that the map qmin ◦ qmax is a completely isometric isomor- phism to establish that all relative crossed products are the same. 8 E.G. KATSOULIS AND C. RAMSEY 3. Hyperrigidity and the Hao-Ng isomorphism A not necessarily unital operator algebra A is said to be hyperrigid if given any (non-degenerate) ∗-homomorphism env(A) −→ B(H) τ : C∗ then τ is the only completely positive, completely contractive extension of the restricted map τA. By adding an injective direct summand if necessary, it is easy to see that in order to verify hyperrigidity, one needs to consider only injective ∗-representations τ but this need not concern us here. The term hyperrigid was coined by Arveson in [2] but the concept had been floating around in various forms before this, e.g. [9]. Our definition is slightly weaker than that of Duncan's [9, Section 4] as Duncan requests that τ be the only completely contractive extension of the restricted map, i.e., no requirement of positivity in the non-unital case. In any case [9, Proposition 4] shows that the graph algebra of any row-finite graph is hyperrigid. Actually we are about to provide a much stronger result but first we need to remind the reader the definition and some of the basic notation regarding C∗-correspondences. A C∗- correspondence (X,C, ϕX ), or just (X,C), consists of a C∗- algebra C, a Hilbert C-module (X,h , i) and a (non-degenerate) ∗- homomorphism ϕX : C → L(X) into the C∗-algebra of adjointable op- erators on X. Equivalently, a (represented) C∗-correspondence (X,C) consists of a C∗-algebra C ⊆ B(K), K a Hilbert space, and a norm- closed C-bimodule X ⊆ B(K) satisfying X ∗X ⊆ C (this allows us to define the inner product h , i) and span{CX} = X (this is the non- degeneracy of the left action of C). The equivalence of the two defini- tions follows from the fact that an abstract C∗-correspondence embeds in the Toeplitz C∗-algebra TX that we will define below and therefore can be represented on a Hilbert space. A representation (ρ, t) of a C∗-correspondence into B(H), is a pair consisting of a non-degenerate ∗-homomorphism ρ : C → B(H) and a linear map t : X → B(H), such that ρ(c)t(x) = t(ϕX(c)(x)), for all c ∈ C and x ∈ X. It is called an isometric (Toeplitz) represen- tation when t(x)∗t(x′) = ρ(hx, x′i), for all c ∈ C and x, x′ ∈ X. THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 9 By the relations above, the C∗-algebra generated by an isometric representation (ρ, t) equals the closed linear span of t(x1)· · · t(xn)t(y1)∗ · · · t(ym)∗, xi, yj ∈ X. For any isometric representation (ρ, t) there exists a ∗-homomorphism ψt : K(X) → B, such that ψt(θx,y) = t(x)t(y)∗, where K(X) is the sub- algebra of L(X) of so-called compact operators generated by θx,y(z) = xhy, zi. (See [19, Chapter 3] for more details on this topic.) There exists a universal Toeplitz representation, denoted as (ρ∞, t∞), so that any other representation of (X,C) is equivalent to a direct sum of sub-representations of (ρ∞, t∞). The Cuntz-Pimsner-Toeplitz C∗-algebra TX is defined as the C∗-algebra generated by the image of (ρ∞, t∞). X of a C∗-correspondence [28] (X,C) is the norm-closed subalgebra of TX generated by all elements of the form ρ∞(c), t∞(x), c ∈ C, x ∈ X. The tensor algebra T + X contains a faithful copy of the C∗-correspondence (X,C). Thus X inherits an operator space from T + X ; we can now say that a representation (ρ, t) of (X,C) is completely contractive whenever t is a completely contractive map with respect to that operator space structure. The tensor algebra T + Consider the ideal JX ≡ ϕ−1 X (K(X)) ∩ ker ϕ⊥ X . (which we will call Katsura's ideal.) An isometric representation (ρ, t) of (X,C, ϕX) is said to be covariant (Cuntz-Pimsner) if and only if ψt(ϕX(c)) = ρ(c), for all c ∈ JX. The universal C∗-algebra for all iso- metric covariant representations of (X,C) is the Cuntz-Pimsner algebra OX. The algebra OX contains (a faithful copy of) C and (a unitarily equivalent) copy of X. The first author and Kribs [21, Lemma 3.5] have shown that the non-selfadjoint algebra of OX generated by these copies of C and X is completely isometrically isomorphic to T + X ) ≃ OX. See [21, 28] for more details. X . Furthermore, C∗ env(T + Now to the hyperrigidity of tensor algebras. X is a hyperrigid operator algebra. Theorem 3.1. Let (X,C) be a C∗-correspondence. If ϕX(JX) acts non-degenerately on X, then (X,C) is a hyperrigid C∗-correspondence, i.e., T + Proof. Let τ : OX −→ B(H) be a ∗-homomorphism and let τ ′ : OX −→ B(H) be a completely contractive and completely positive map that agrees with τ on T + X . We are to prove that τ ′ is multiplicative and so it agrees with τ . Since τ ′ is a completely contractive and completely 10 E.G. KATSOULIS AND C. RAMSEY positive map, we can use multiplicative domain arguments [6, Propo- sition 1.5.7]. Let (ρ, t) be the universal Cuntz-Pimsner representation of (X,C). X is a C∗-algebra, the multiplicative domain of τ ′ con- Since ρ(C) ⊆ T + tains ρ(C). We claim that it also contains t(X). Indeed, for any x ∈ X we have τ ′(t(x))∗τ ′(t(x)) = τ (t(x))∗τ (t(x)) = τ (t(x)∗t(x)) (3) = τ (ρ(hx, xi)) = τ ′(ρ(hx, xi)) = τ ′(t(x)∗t(x)), where the equation on the second line holds because ρ(C) ⊆ T + the two maps agree there. X and Let a ∈ JX and x ∈ X. Since ϕX(a) ∈ K(X), we have zm,k, wm,k ∈ X, m, k ∈ N, so that (4) ϕX(a) = lim θzm,k,wm,k m→∞Xk is a limit of finite rank operators in K(X). Let X0 ⊆ X be the C- submodule generated by x and all zm,k, wm,k ∈ X, m, k ∈ N. Since X0 is countably generated, Kasparov's Stabilization Theorem implies the n=1 in X0 so that kPl existence of {xn}∞ n=1 θxn,xnk ≤ 1, for all l ∈ N, and Xn=1 θxn,xn(ξ) = ξ, for all ξ ∈ X0. From this, a standard approximation argument involving (4) shows that ∞ ∞ ∞ (5) θxn,xnϕX(a) = ϕX(a)θxn,xn = ϕX (a), Xn=1 Xn=1 with the convergence in the norm topology1. Then ϕX(aa∗) = lim k = By the Schwarz inequality ∞ Xn=1 ϕX(a)(cid:0) k Xn=1 θxn,xn(cid:1)ϕX (a)∗ θϕX (a)xn,ϕX (a)xn. (6) τ ′(cid:0)t(ϕX(a)xn)t(ϕX(a)xn)∗(cid:1) ≥ τ ′(t(ϕX (a)xn))τ ′(t(ϕX(a)xn))∗, 1This is exactly the same argument one uses on non-separable Hilbert space to write any compact operator as a (perhaps infinite) sum of rank-one operators. THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 11 for all n ∈ N, and so τ ′(ρ(aa∗)) = τ ′(ψt(ϕX(aa∗)) ∞ = ≥ = τ ′(cid:0)t(ϕX (a)xn)t(ϕX (a)xn)∗(cid:1) τ ′(t(ϕX (a)xn))τ ′(t(ϕX(a)xn))∗ τ (t(ϕX (a)xn))τ (t(ϕX (a)xn))∗ ∞ Xn=1 Xn=1 Xn=1 ∞ = τ (ψt(ϕX(aa∗)) = τ (ρ(aa∗)) = τ ′(ρ(aa∗)). Hence (6) is actually an equality. Combining this with (3), we conclude that t(ϕX(a)xn) belongs to the multiplicative domain of τ ′, for all a ∈ JX and n ∈ N. Since ρ(C) is also contained in the multiplicative domain of τ ′, we have that ∞ t(ϕX (a)x) = t(ϕX(a)xn)ρ(hxn, xi) Xn=1 belongs to the multiplicative domain of τ ′, for all a ∈ JX and x ∈ X. Since ϕX(JX) acts non-degenerately on X, the multiplicative domain of τ ′ contains t(X), as desired. This completes the proof. Recall that a C∗-correspondence (X,C) is said to be regular iff C acts faithfully on X by compact operators, i.e., JX = C. The following is immediate. Corollary 3.2. A regular C∗-correspondence is necessarily hyperrigid. We are about to see that the assumption of injectivity cannot be removed from the Corollary above. But first we need criterion for the failure of hyperrigidity. Proposition 3.3. Let (X,C) be a C∗-correspondence with JX = {0}. Then (X,C) fails to be hyperrigid. Proof. Let (π, t) be any isometric representation of (X,C) on a Hilbert space H. If V1, V2 are the unilateral and bilateral (forward) shift re- spectively, then the associations (7) C ∋ a −→ a ⊗ I, X ∋ x −→ x ⊗ Vi, i = 1, 2, 12 E.G. KATSOULIS AND C. RAMSEY determine isometric representations of X, which are neccesarilly Cuntz- Pimsner covariant, since JX = {0}. Therefore they promote to repre- sentations ϕ1 and ϕ2 of OX ≃ C∗ 0 ) and H⊗ℓ2(Z) re- spectively. Now notice that when ϕ2 is being compressed on H⊗ℓ2(Z+ 0 ), it produces a completely positive contractive map ϕ2 6= ϕ1, which how- ever agrees with ϕ1 on T + X . Hence (X,C) is not hyperrigid. X ) on H⊗ℓ2(Z+ env(T + Recall that if α is an endomorphism of a C∗-algebra A, then the semicrossed product A⋊α Z+ 0 (also denoted as A⋊α Z+ in the literature) is simply the tensor algebra of the C∗-correspondence Aα, where the left action on A is coming from α. In the case where both A and α are unital and α is injective, such algebras are always hyperrigid. This has already been noted in the literature, eg. [16], but it is also an immediate consequence of Corollary 3.2. It is worth noting that the requirement of α being injective cannot be dropped from neither Corollary 3.2 nor the discussion above. Example 3.4. Let X be a compact Hausdorff space which is not a singleton and consider some x ∈ X which is not an isolated point. Let ϕ : X → X with ϕ(y) = x, for all y ∈ X . Then the semicrossed product C(X ) ⋊ϕ Z+ Indeed, in that case, the kernel of the right action equals C0(X\{x}). Hence Katsura's ideal is trivial and Proposition 3.3 applies. 0 is not hyperrigid. env(A), ι) where C∗ Finally recall that the C-envelope of a non-unital operator alge- bra can be computed from the C-envelope of its unitization. More precisely, as the pair (C∗ env(A) is the C-subalgebra env(A1), ι) of the (unique) generated by ι(A) inside the C-envelope (C∗ unitization A1 of A. By the proof of [5, Proposition 4.3.5] this C- envelope of an operator algebra A has the desired universal property, that for any C-cover (ι′,B′) of A, there exists a (necessarily unique and surjective) ∗-homomorphism π : B′ → C∗ env(A), such that π◦ι′ = ι. We start with an elementary result regarding crossed products. Lemma 3.5. Let (C,G, α) be a C∗-dynamical system and let D ⊆ C be the C∗-subalgebra of C generated by some selfadjoint approximate unit for C. Then (i) CCc(G,D) is dense in C ⋊α G. (ii) If π : C ⋊αG → B(H) is a non-degenerate representation, then its restriction on Cc(G,D) is also non-degenerate. Proof. Let {ei}i∈I be the selfadjoint approximate unit generating D. Then any elementary tensor h⊗c ∈ Cc(G,D), where (h⊗c)(s) = h(s)c, THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 13 h ∈ Cc(G), c ∈ C, can be written as h ⊗ c = lim i∈I c (h ⊗ ei) ∈ CCc(G,D). This implies (i). For (ii) notice that by taking adjoints in (i), Cc(G,D)C is also dense in C ⋊α G. Hence π (Cc(G,D))H = π (Cc(G,D)) π(C)H = π (Cc(G,D)C)H which is dense in π(C)H = H and the conclusion follows. Our next result has been established for all discrete groups in [20]. Here we extend it to arbitrary locally compact groups provided that the pertinent algebras are hyperrigid. One of the key ingredients of the proof is the use of injectivity for operator spaces. We briefly review the key definitions and the results used in the proof. We follow [29] in our presentation; most of the material first appeared in [13]. An operator space I is said to be injective provided that for any pair of operator spaces E ⊆ F and completely contractive map ϕ : E → I, there exists a completely contractive map ψ : F → I that extends ϕ. Given an operator space F , we say that (E, κ) is an injective envelope of F provided that (i) E is injective, (ii) κ : F → E is a complete isometry, (iii) if E1 is injective with κ(F ) ⊆ E1 ⊆ E, then E1 = E. Hamana essentially showed that every operator space F ⊆ B(H) admits an injective envelope (E, κ), with E ⊆ B(H) and κ being the inclusion map [29, Theorem 15.4]. The proof of [29, Theorem 15.4] shows that E materializes as the range of a completely contractive idempotent ϕ : B(H) → B(H). If B(H) ∋ I ∈ F then the completely contractive idempotent ϕ is unital and therefore completely positive. Hence the range of ϕ, i.e., E, is an operator system and the Choi- setting a ◦ b = ϕ(ab) Effros Theorem [29, Theorem 15.2] applies: defines a multiplication on E = ϕ(B(H)), and E equipped with this multiplication and its usual ∗-operation becomes a C∗-algebra. If on top of beiing unital, F happens to be an operator algebra as well, then the C∗-subalgebra of (E,◦) generated by F , gives the C∗-envelope of F [29, Theorem 15.16]. Theorem 3.6. Let A be a hyperrigid operator algebra which possesses a contractive approximate unit {ei}i∈I consisting of selfadjoint operators. Let α : G → AutA be a continuous action of a locally compact group. Then (8) C∗ env (A ⋊r α G) ≃ C∗ env(A) ⋊r α G 14 and (9) via canonical embeddings. Proof. Let ρ : C∗ sentation and let E.G. KATSOULIS AND C. RAMSEY C∗ env(A),α G(cid:1) ≃ C∗ env(cid:0)A ⋊C∗ env(A) → B(H) be a faithful (non-degenerate) repre- env(A) ⋊α G ρ : C∗ env(A) −→ B(cid:0)H ⊗ L2(G)(cid:1) u : G −→ B(cid:0)H ⊗ L2(G)(cid:1) so that ρ ⋊ u (which we will denote as π) is the regular representation induced by ρ. (See [34, Section 2.2] for notation and additional infor- mation.) Since ρ is non-degenerate, [34, Lemma 2.17] implies that the induced representation π = ρ ⋊ u is also non-degenerate. Let ϕ : B(cid:0)H ⊗ L2(G)(cid:1) −→ B(cid:0)H ⊗ L2(G)(cid:1) be a completely contractive idempotent map whose range is the injec- α G)1. Let D be the closed (selfadjoint) subalge- tive envelope of π (A ⋊r bra of A generated by {ei}i∈I. Then Cc(G,D) is a selfadjoint subalgebra of A ⋊r (10) α G and so [5, 1.3.12] implies that ϕ (Sπ(f )) = ϕ(S)π(f ), for any S ∈ B(cid:0)H ⊗ L2(G)(cid:1) and f ∈ Cc(G,D). In particular ϕ(cid:0)ρ(a)π(f )(cid:1) = ϕ(ρ(a))π(f ), for all a ∈ A, f ∈ Cc(G,D). On the other hand, ρ(a)π(f ) = π(af ) ∈ π(A ⋊r α G) and so ϕ(ρ(a)π(f )) = ρ(a)π(f ), a ∈ A, f ∈ Cc(G,D). Hence (cid:0)ϕ(ρ(a)) − ρ(a)(cid:1)π(f ) = 0, for all f ∈ Cc(G,D). By Lemma 3.5(ii), π(cid:0)Cc(G,D)(cid:1) acts non-degenerately on H ⊗ L2(G) and so (11) ϕ(ρ(a)) = ρ(a), for all a ∈ A. Hence the mapping ϕ is a completely positive and completely contrac- tive extension of the identity map on ρ(A). However ρ(A) is hyper- rigid and according to the discussion in the beginning of the section, the identity map on ρ(A) is the only such completely contractive and completely positive extension to C∗ env(A)(cid:1). Therefore env(ρ(A)) = ρ(cid:0)C∗ env(A). ϕ(ρ(c)) = ρ(c), for all c ∈ C∗ THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 15 Appealing again to (10), with S = ρ(c), we obtain for all c ∈ C∗ ϕ(π(cf )) = ϕ(cid:0)ρ(c)π(f )(cid:1) = ϕ(ρ(c))π(f ) = ρ(c)π(f ) = π(cf ), env(A) and f ∈ Cc(G,D). By Lemma 3.5(i) we have ϕ(S) = S, for all S ∈ π(cid:0)C∗ π(cid:0)C∗ α G(cid:1). α G(cid:1)1 ⊆ ϕ(cid:0)B(cid:0)H ⊗ L2(G)(cid:1)(cid:1) env(A) ⋊r env(A) ⋊r But this implies that the Choi-Effros multiplication on . Hence env(A) ⋊r is actually the original one coming from B(cid:0)H ⊗ L2(G)(cid:1) and so the C∗-algebra generated by π(A ⋊r α G(cid:1)1 env(cid:0)π(A ⋊r env(A) ⋊r α G)1 ⊆ ϕ(cid:0)B(cid:0)H ⊗ L2(G)(cid:1)(cid:1) equals π(cid:0)C∗ α G(cid:1)1. env(A) ⋊r α G)1(cid:1) = π(cid:0)C∗ Furthermore, the C∗-algebra generated by π(A ⋊r α G) ⊆ π(cid:0)C∗ G(cid:1)1 α G(cid:1). This establishes (8). equals π(cid:0)C∗ In order to prove (9), let this time π := ρ ⋊ u, where (ρ, u) is the env(A),G, α). With this π, a universal covariant representation of (C∗ verbatim repetition of the proof of (8) establishes (9). env(A) ⋊r C∗ α α G,C⋊r α G) is the completion of Cc(G, X) and Cc(G,C) in T + Now we turn to crossed product correspondences. Let G be a locally compact group acting on a non-degenerate C∗-correspondence (X,C) by a generalized gauge action α : G → Aut(TX ), i.e., αs(X) = X and αs(C) = C, for all g ∈ G. The reduced crossed product correspondence (X ⋊r ⋊r α G, which can be thought of as living in T + ⋊r α G but equivalently can be considered as living in OX ⋊r α G. The left and right module actions are given by multiplication and hS, Ti = S∗T , for S, T ∈ Cc(G, X). In a similar manner, one defines the full crossed product corre- spondence (X ⋊α G,C ⋊α G) by completing the spaces in TX ⋊α G. This was shown to be unitarily equivalent to the abstract characteri- zation of the full crossed product correspondence in [22, Remark 7.8]. Lastly, we recall the definition of the crossed product correspondence (X ⋊α G,C ⋊α G) which is the completion of the spaces in OX ⋊α G. In general, it is unknown whether these two correspondences are unitarily equivalent or not. X X Our next result has been established in [22] in the case where G is discrete. In [22] it was also noted that the proof carries over to the general locally compact case. We are about to explain how this is done. 16 E.G. KATSOULIS AND C. RAMSEY In the proof we will use the language of product systems, which we now discuss briefly. Let G be a countable group with unit e ∈ G and let P ⊆ G be a positive cone. A product system X = {Xp}p∈P over (G, P ) consists of a C∗-algebra Xe ⊆ B(K), K a Hilbert space, and a family of (represented) Xe-correspondences Xp ⊆ B(K), p ∈ P\{e}, satisfying the semigroup rule span{XpXq} = Xpq, for all p, q ∈ P . For instance, if (X,C) is a (represented) C∗-correspondence, then by taking G = Z, P = Z+ 0 , X0 = C and Xn = span{X n}, n = 1, 2, . . . , we obtain a product system {Xn}∞ 0 ). As with C∗-correspondences, one can define product systems abstractly but we will not do that here. See [11] for more details. n=0 over (Z, Z+ If X = {Xp}p∈P is a poduct system over (G, P ), then an isomet- ric (Toeplitz) representation ψ = {ψp}p∈P of X on a Hilbert space H consists of a ∗-representation ψe : Xe → B(H) and isometric C∗- correspondence representations ψp : Xp → B(H), p ∈ P , so that ψpq(xpxq) = ψp(xp)ψq(xq) for all xp ∈ Xp, xq ∈ Xq, p, q ∈ P . If (ρ, t) is a representation of a C∗-correspondence (X,C) on H, then by setting ψ0 = ρ and ψn : Xn −→ B(H); x1x2 . . . xn 7−→ t(x1)t(x2) . . . t(xn) we obtain a representation of the product system {Xn}∞ n=0 discussed in the previous paragraph. These representations are exactly the com- pactly aligned, Nica covariant representations of {Xn}∞ n=0 in the lan- guage of [11] and so the Toeplitz C∗-algebra of {Xn}∞ n=0 generated by the theory of [11] coincides with our TX. The use of the product system language in the proof of the next re- sult allows us to import an important result from the theory of product systems, Fowler's Theorem [11, Theorem 7.2]. This result has no ana- logue within the theory of C∗-correspondences as it allows us to check whether a representation of the Toeplitz algebra of a C∗-correspondence is faithful without using gauge actions. Theorem 3.7. Let G be a locally compact group acting by a generalized gauge action α on a non-degenerate C∗-correspondence (X,C). Then Therefore, X T + env(cid:0)T + X C∗ ⋊r α G ≃ T + X⋊r α G. ⋊r α G(cid:1) ≃ OX⋊r α G. Proof. Let ρ : TX → B(H) be some faithful ∗-representation and let V ∈ B(ℓ2(Z+ 0 )) be the forward shift. As we did earlier, set Xn := X n THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 17 for n ≥ 1, X0 := C and for the rest of the proof let X denote the product system {Xn}∞ n=0. Then the map Xn ∋ x 7−→ ρ(x) ⊗ V n ∈ B(H ⊗ l2(Z+ 0 )), n ∈ Z+ 0 , 0 )). Since X ⋊r α G, n ∈ Z+ 0 . is a representation of X that satisfies the requirements of Fowler's The- orem [11, Theorem 7.2]. Therefore it establishes a faithful representa- tion π : TX → B(H ⊗ ℓ2(Z+ α G ⊆ TX ⋊r α G, we may consider the regular representation Indπ, when restricted on X ⋊r α G, as a representation of the product system X ⋊r α G, which we denote as ψ. We furthermore write ψn := ψ Xn⋊r We claim that ψ satisfies the requirements of Fowler's Theorem [11, Theorem 7.2]. Indeed let Qn ∈ l2(Z+ 0 ) be the projection on the one dimensional subspace corresponding to the characteristic function of 0 and let Qn ≡ (I ⊗ Qn) ⊗ I ∈ B(cid:0)H ⊗ l2(Z+ n ∈ Z+ (constant) B(H⊗l2(Z+ to any s ∈ G. Then given any f ∈ Cc(G, Xn), n ≥ 1, we have (cid:0) Q0Indπf(cid:1)h(t) = Q0ZG t (f (s))(cid:1)h(s−1t)ds 0 ) ⊗ L2(G)(cid:1) be the 0 ))-valued function that assigns the value I ⊗Qn π(cid:0)α−1 t (f (s))) ⊗ V n(cid:1)h(s−1t)ds = (I ⊗ Q0) ⊗ IZG(cid:0)ρ(α−1 =ZG(cid:0)ρ(α−1 the product Qn∈F \{0}(I − P ψ n , always dominates Q0 and so factor of the form I − P ψ (I − P ψ (12) n denotes the range space of ψ(Xn ⋊r t (f (s))) ⊗ Q0V n(cid:1)h(s−1t)ds = 0 α G), n ∈ Z+ 0 , then n ), which in our case collapses to a single Therefore if P ψ n )(cid:13)(cid:13) ≥(cid:13)(cid:13)ψ0(f ) Q0(cid:13)(cid:13) =(cid:13)(cid:13)Indπ(f ) Q0(cid:13)(cid:13), for any f ∈ Cc(G,C). However, each I ⊗ Qn reduces π C and therefore (I ⊗ Qn)π(cid:1) C≃(cid:0) ⊕ (I ⊗ Q0)π(cid:1) C, i.e., the restriction of π on C is unitarily equivalent to a direct sum of countably many copies of (I ⊗ Q0)π restricted on C. From this we obtain, (cid:13)(cid:13)ψ0(f ) Yn∈F \{0} π C≃(cid:0) ⊕n∈Z+ 0 ψ0 = Indπ C⋊r α G ≃ ⊕IndQ0π C⋊r Combining the above with (12) we now obtain α G . α G ≃ ⊕ Q0Indπ C⋊r n )(cid:13)(cid:13) =(cid:13)(cid:13)Indπ(f ) Q0(cid:13)(cid:13) = kψ0(f )k (cid:13)(cid:13)ψ0(f ) Yn∈F \{0} (I − P ψ 18 E.G. KATSOULIS AND C. RAMSEY Since the claim is valid, Fowler's Theorem shows now that the in- α G. Note for any f ∈ Cc(G,C), which establishes the claim. duced representation ψ∗ is a faithful representation of TX⋊r now that ψ∗(T + ψ∗(X ⋊r α G is equal to the closed linear span of α G) ≃ T + α G) = [n∈Z+ ⋊r α G is also isomorphic to the closed linear span of α G) = [n∈Z+ IndπCc(G, Xn). However, T + X ψn(Xn ⋊r X⋊r X⋊r 0 0 IndπCc(G, Xn). [n∈Z+ 0 Hence, T + X Finally ⋊r α G. X⋊r α G ≃ T + C∗ env(cid:0)T + X ⋊r α G(cid:1) ≃ C∗ with the last identification following from [21, Theorem 3.7]. env(T + X⋊r α G) ≃ OX⋊r α G. Remark 3.8. The use of the language of product systems in the pre- vious proof has an additional benefit. By switching from (Z, Z+ 0 ) to an arbitrary totally ordered group (G, P ), the same exact proof as above establishes the more general result NT + α G, where NT + X denotes the Nica tensor algebra of X. Actually the proof works for any quasi-lattice ordered group (G, P ) provided that certain issues involv- ing compact alignment and Nica covariance are worked out first. We are recording this fact here for future reference. α G ≃ NT + X⋊r ⋊r X In [20] the Hao-Ng isomorphism problem was resolved for the re- duced crossed product and all discrete groups. In the next result we address the case of an arbitrary locally compact group and we resolve the Hao-Ng problem for the reduced crossed product provided that the C∗-correspondence is hyperrigid. Note that our result subsumes an earlier result [24, Proposition 5.5] which was posted on the arXiv but has not appeared in print. Theorem 3.9. Let G be a locally compact group acting by a general- ized gauge action α on a non-degenerate hyperrigid C∗-correspondence (X,C), e.g. ϕX(JX)X = X. Then OX ⋊r Proof. By Theorem 3.7 we have C∗ other hand, Theorem 3.6 implies that C∗ Hence OX ⋊r α G ≃ C∗ env(T + X ) ⋊r α G. α G ≃ OX⋊r env(cid:0)T + env(cid:0)T + α G ≃ OX⋊r X α G. X ⋊r ⋊r α G(cid:1) ≃ OX⋊r α G(cid:1) ≃ C∗ α G. On the env(T + α G. X )⋊r THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 19 It is important to us that an analogous results holds for the full crossed product. Theorem 3.10. Let G be a locally compact group acting by a general- ized gauge action α on a non-degenerate hyperrigid C∗-correspondence (X,C). Then T + X OX ⋊α G ≃ OX ⋊α G. Proof. In [22, Theorem 7.13] we proved that env(cid:0)T + env(T + x ),α G(cid:1) ≃ C∗ ⋊OX ,α G ≃ T + env(cid:0)T + Now (9) in Theorem 3.6 and the above imply that and C∗ OX ⋊α G ≃ C∗ C∗ and the conclusion follows. ⋊ X ⋊α G X env(T + X ⋊OX ,α G(cid:1) ≃ OX ⋊α G. X ) ⋊α G ≃ OX ⋊α G At this point one might think that the above theorem is the final word regarding the Hao-Ng isomorphism for hyperrigid correspondences. As it turns out, this couldn't be further from the truth. It is indeed the case that we have expressed the crossed product OX ⋊αG as the Cuntz- Pimsner algebra of a C∗-correspondence, namely X ⋊α G, but this is not the C∗-correspondence that the authors of [3] ask for. Is this a big deal? Most definitely yes, and we devote the next section explaining the reasons why. 4. Isometric coextensions X The goal of this section is to answer Problem 3 in Chapter 8 of [22]: Is T + ⋊α G the tensor algebra of some C∗-correspondence? The (affir- mative) answer is one of the key ingredients in the proof of Theorem 4.9, one of the central results of the paper. Definition 4.1. Let (X,C) be a C∗-correspondence and G a locally compact group. A generalized gauge action α : G → Aut((X,C)) is a map from G into the completely isometric module automorphisms. In particular, for each s ∈ G, αs is an isometric automorphism of X and a ∗-automorphism of C such that αs(ξc) = αs(ξ)αs(c), αs(ϕX(c)ξ) = ϕX(αs(c))αs(ξ) for all ξ, η ∈ X and c ∈ C. and αs(hξ, ηi) = hαs(ξ), αs(η)i Earlier we said that α : G → AutTX forms a generalized gauge action of TX if αs(X) = X and αs(C) = C, for all g ∈ G. The following result says that the two definitions are equivalent and it was observed in [20, pg 5760]. (See also [14, Lemma 2.6] for the analogous result with OX). 20 E.G. KATSOULIS AND C. RAMSEY Proposition 4.2. Let (X,C) be a non-degenerate C∗-correspondence and G a locally compact group. If α : G → Aut(TX) is a generalized gauge action of TX then it restricts to a generalized gauge action of (X,C). Conversely, a generalized gauge action α of (X,C) extends uniquely to a generalized gauge action of TX . The fundamental object of study in this section is the C∗-correspon- dence dynamical system ((X,C),G, α) which is given by a non-degenerate C∗-correspondence (X,C) and a generalized gauge action α of the lo- cally compact group G acting on (X,C). Definition 4.3. A representation of the C∗-correspondence dynami- cal system ((X,C),G, α) is a quadruple (ρ, t, u,H) consisting of a com- pletely contractive representation (ρ, t,H) of (X,C) and a strongly con- tinuous unitary representation u : G → U(H) satisfying the covariance relations u(s)t(ξ) = t(αs(ξ))u(s) and u(s)ρ(c) = ρ(αs(c))u(s) for all s ∈ G, ξ ∈ X and c ∈ C. Moreover, (ρ, t, u,H) is said to be isometric if (ρ, t,H) is an isometric (Toeplitz) representation of (X,C). The following theorem is an extension of [28, Theorem 2.12]. Theorem 4.4. Let ((X,C),G, α) be a C∗-correspondence dynamical system. The isometric representations (ρ, t, u,H) of ((X,C),G, α) are in bijective correspondence with the isometric representations (π, u,H) of (TX ,G, α). Specifically, they are related by π = ρ ⋊ t. Proof. If (π, u,H) is an isometric representation of (TX,G, α) then [28, Theorem 2.12] proves that there exists an isometric representation (ρ, t,H) of (X,C) such that ρ ⋊ t = π. Proposition 4.2 gives that α is a generalized gauge action of G on (X,C) and so the covariance relations between ρ, t and u are automatic. Hence, (ρ, t, u,H) is an isometric representation of ((X,C),G, α). Conversely, suppose (ρ, t, u,H) is an isometric representation of ((X,C),G, α). Again by [28, Theorem 2.12] this gives that (ρ ⋊ t,H) is an isometric representation of TX and by Proposition 4.2 α ex- tends uniquely to a generalized gauge action of G on TX. Because u(s)(ρ ⋊ t(·))u(s)∗ and (ρ ◦ αs) ⋊ (t ◦ αs) = (ρ ⋊ t) ◦ αs agree on X and C then the uniqueness of [28, Theorem 2.12] gives that u(s)ρ ⋊ t(a) = ρ ⋊ t(αs(a))u(s) for all s ∈ G and a ∈ TX. Therefore, (ρ ⋊ t, u,H) is an isometric representation of (TX ,G, α). THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 21 If (ρ, t, u,H) and (ρ1, t1, u1,H1) are completely contractive represen- tations of ((X,C),G, α) we say that the latter is a dilation of the former if H ⊆ H1 and (i) H reduces π1 and u1 with π1(c)H = π(c), c ∈ C and u1(g)H = (ii) H is a semi-invariant subspace for t1 and PHt1(ξ)H = t(ξ), ξ ∈ u(g), g ∈ G, and X. We call such a dilation an extension if H is an invariant subspace for t1 and a coextension if H is a coinvariant subspace for t1. The dilation (ρ1, t1, u1,H1) is called minimal when H1 is the smallest reducing subspace for t1 containing H. Now we need a lemma relating to the step-by-step dilation tech- niques of Muhly and Solel of [28, Section 3]. Their proof is modelled after Popescu's step-by-step dilation technique [30] using the Schaeffer matrix construction [32]. The following is also a slight simplification of the original proof in [28]. To this end suppose (ρ, t, u,H) is a completely contractive represen- tation of the C∗-correspondence dynamical system ((X,C),G, α). The ultimate goal is to prove that every such representation dilates to an isometric representation. As in [28], define the Hilbert space HX = X ⊗ρ H h·,·i where ξc ⊗ h = ξ ⊗ ρ(c)h and hξ ⊗ h, η ⊗ ki = hh, ρ(hξ, ηi)ki for ξ, η ∈ X, c ∈ C and h, k ∈ H. As well, define σX : X → B(H,HX) by σX (ξ)h = ξ ⊗ h and t : HX → H by t(ξ ⊗ h) = t(ξ)h. From here one defines the one step dilation to H1 = H ⊕ HX given by t1(ξ) =(cid:20) and t(ξ) (I − t∗t)1/2σX(ξ) 0 (cid:21) 0 ρ1(c) =(cid:20) ρ(c) 0 0 ρ(c) (cid:21) where ρ : C → B(HX ) is given by ρ(c)(ξ ⊗ h) = ϕX (c)ξ ⊗ h. Lemma 4.5. Consider u1(s) =(cid:20) u(s) 0 0 u(s) (cid:21) 22 E.G. KATSOULIS AND C. RAMSEY where u : G → B(HX) is given by u(s)(ξ ⊗ h) = αs(ξ) ⊗ u(s)h, which is well-defined. Then (ρ1, t1, u1,H1) is a completely contractive repre- sentation of ((X,C),G, α) such that t1(ξ)∗t1(η) =(cid:20) ρ(hξ, ηi) 0 0 (cid:21) 0 for all ξ, η ∈ H. Proof. By [28, Lemma 3.7] (ρ1, t1) is a completely contractive repre- sentation of (X,C) on H1 which satisfies the last two statements in the lemma. First note that u is in fact well-defined since it respects the internal C-modularity of HX, u(s)(ξc ⊗ h) = αs(ξc) ⊗ u(s)h = αs(ξ)αs(c) ⊗ u(s)h = αs(ξ) ⊗ ρ(αs(c))u(s)h = αs(ξ) ⊗ u(s)ρ(c)h = u(s)(ξ ⊗ ρ(c)h), for all s ∈ G, c ∈ C, ξ ∈ X, h ∈ H. and h, k ∈ H we have that Additionally, observe that u(s) is unitary since for all s ∈ G, ξ, η ∈ X hu(s)(ξ ⊗ h), u(s)(η ⊗ k)i = hαs(ξ) ⊗ u(s)h, αs(η) ⊗ u(s)ki = hu(s)h, ρ(hαs(ξ), αs(η)i)u(s)ki = hu(s)h, ρ(αs(hξ, ηi))u(s)ki = hu(s)h, u(s)ρ(hξ, ηi)ki = hh, ρ(hξ, ηi)ki = hξ ⊗ h, η ⊗ ki. Next we need to make the following calculations: σX(αs(ξ))u(s)h = αs(ξ) ⊗ u(s)h = u(s)(ξ ⊗ h) = u(s)σX (ξ)h THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 23 and ht∗tσX (αs(ξ))u(s)h, η ⊗ ki = ht(αs(ξ) ⊗ u(s)h), t(η ⊗ k)i = ht(αs(ξ))u(s)h, t(η)ki = hu(s)t(ξ)h, t(η)ki = ht(ξ)h, t(αs−1(η))u(s)∗ki = ht∗t(ξ ⊗ h), αs−1(η) ⊗ u(s−1)ki = hu(s)t∗t(ξ ⊗ h), η ⊗ ki = hu(s)t∗tσX (ξ)h, η ⊗ ki. Combining these one gets u(s)(I − t∗t)σX (ξ) = (I − t∗t)σX(αs(ξ))u(s) = (I − t∗t)u(s)σX(ξ). Hence, u(s)(I − t∗t) = (I − t∗t)u(s) and by a standard trick often attributed to Halmos u(s)(I − t∗t)1/2 = (I − t∗t)1/2 u(s). Now, we need to establish the covariance relations between (ρ1, t1) and u1. From the previous paragraph we have that u(s)(I − t∗t)1/2σX (ξ) = (I − t∗t)1/2σX(αs(ξ))u(s) and thus u1(s)t1(ξ) = t1(αs(ξ))u1(s). Second, it is much more straightforward to calculate that u(s)ρ(c)(ξ ⊗ h) = u(s)(ϕX(c)ξ ⊗ h) = αs(ϕX(c)ξ) ⊗ u(s)h = ϕX (αs(c))αs(ξ) ⊗ u(s)h = ρ(αs(c))u(s)(ξ ⊗ h). Therefore, u1(s)ρ1(c) = ρ1(αs(c))u1(s) and the conclusion follows. Theorem 4.6. Every completely contractive representation of the non- degenerate C∗-correspondence dynamical system ((X,C),G, α) has a minimal isometric coextension. Moreover, the minimal isometric coex- tension is unique up to unitary equivalence. Proof. Let (ρ, t, u,H) be a completely contractive representation of ((X,C),G, α). Following the proof of [28, Theorem 3.3] repeatedly use Lemma 4.5 to get a sequence of completely contractive representations 24 E.G. KATSOULIS AND C. RAMSEY (ρn, tn, un,Hn) of ((X,C),G, α) in the obvious manner: use (ρ, t, u,H) to produce (ρ1, t1, u1,H1) and then recursively (ρn, tn, un,Hn) = ((ρn−1)1, (tn−1)1, (un−1)1, (Hn−1)1) from the previous lemma and the discussion preceding it. Let H′ = ∪n≥1Hn and define ρ′ = lim−→ ρn, t′ = lim−→ tn and u′ = lim−→ un. By [28, Theorem 3.3] (ρ′, t′,H′) is an isometric representation of (X,C). Note that u′ is a strongly continuous unitary representation of G since it is the direct sum of such representations. Now to the covariance relations: u′(s)ρ′(c)PHn = un(s)ρn(c)PHn = ρn(αs(c))un(s)PHn = ρ′(αs(c))u′(s)PHn for all s ∈ G, c ∈ C and n ∈ N. As well, u′(s)t′(ξ)PHn = un+1tn+1(ξ)PHn = tn+1(αs(ξ))un+1(s)PHn = tn+1(αs(ξ))un(s)PHn = t′(αs(ξ))u′(s)PHn for all s ∈ G, ξ ∈ X and n ∈ N. Therefore, the covariance relations are satisfied and (ρ′, t′, u′,H′) is an isometric coextension of (ρ, t, u,H). Now let K = span{t′(ξ1)· · · t′(ξn)h : ξi ∈ X, h ∈ H, n ≥ 1} ⊂ H′. Because of the covariance relations of (ρ′, t′, u′,H′) one can see that K is a reducing subspace of (ρ′, t′, u′,H′) that contains H. Thus, (ρ′, t′, u′,K) is a minimal isometric coextension of (ρ, t, u,H). In regard to uniqueness, suppose that (ρ′′, t′′, u′′,H′′) is another iso- metric coextension of (ρ, t, u,H). [28, Proposition 3.2] proves that there exists a unitary W : K → H′′ such that W ρ′(·) = ρ′′(·)W , W t′(·) = t′′(·)W and W h = h, for all h ∈ H. Now W u′(s)t′(ξ1)· · · t′(ξn)h = W t′(αs(ξ1))· · · t′(αs(ξn))u′(s)h = t′′(αs(ξ1))· · · t′′(αs(ξn))W u(s)h = t′′(αs(ξ1))· · · t′′(αs(ξn))u(s)h = t′′(αs(ξ1))· · · t′′(αs(ξn))u′′(s)W h = u′′(s)t′′(ξ1)· · · t′′(ξn)W h = u′′(s)W t′(ξ1)· · · t′(ξn)h THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 25 Therefore, by minimality W u′(·) = u′′(·)W and so (ρ′, t′, u′,K) and (ρ′′, t′′, u′′,H′′) are unitarily equivalent. Theorem 4.7. Let (X,C) be a non-degenerate C∗-correspondence and let α be a generalized gauge action of a locally compact group G. Then T + X X⋊αG ⋊α G ≃ T + X ⋊TX ,α G ≃ T + X X ⋊TX ,α G ≃ T + Towards proving the remaining isomorphism, let ϕ : T + Proof. It is already proven in [22], in a discussion following Theorem 7.13, that T + X⋊αG. (It also follows from [3, Theorem 3.1] as the isomorphism Φ of that theorem maps generators to generators.) ⋊α G → B(H) be a completely contractive representation. In the same way as in the proof of [22, Theorem 4.1] one can assume that ϕ is nondegenerate. Now by [22, Proposition 3.8] there exists a representation (π, u,H) of (T + X ,G, α) so that ϕ = π ⋊ u. By [28, Theorem 3.10] there is a completely contractive representa- tion, (ρ, t), of (X,C) such that π = ρ ⋊ t. Hence, in the same way as the first part of the proof of Theorem 4.4, (ρ, t, u,H) is a completely contractive representation of ((X,C),G, α). By Theorem 4.6 (ρ, t, u,H) has a unique minimal isometric coextension (ρ′, t′, u′,H′) and thus by Theorem 4.4 (ρ′ ⋊ t′, u′,H′) is an isometric representation of (TX,G, α). As discussed in the proof of [28, Theorem 3.10] H ⊂ H′ is sem- invariant for ρ′ and t′ and thus PHρ′ ⋊ t′H is a completely contractive representation of T + X . Moreover, that same theorem gives that π = ρ ⋊ t = PHρ′ ⋊ t′H Therefore, every completely contractive representation of T + because ρ(c) = PHρ′(c)H and t(ξ) = PHt′(ξ)H for all c ∈ C and ξ ∈ X. ⋊α G dilates to a completely contractive representation of T + ⋊TX ,α G and thus they are completely isometrically isomorphic. Corollary 4.8. Let ((X,A),G, α) be a C∗-correspondence dynamical system and assume that JX = {0}. Then all relative crossed prod- ucts for (T + X ,G, α) are canonically isomorphic via completely isometric maps. X X In particular, the above applies to the non-commutative disc algebra A∞ ⊆ O∞. To obtain the same conclusion for the non-commutative disc algebras An, n < ∞, we need to work much harder. (See the next section.) For the moment, we can put together all previous results to obtain the following, which summarizes our knowledge on the Hao-Ng isomor- phism problem for the full crossed product. 26 E.G. KATSOULIS AND C. RAMSEY Theorem 4.9. Let ((X,A),G, α) be a non-degenerate C∗-correspondence dynamical system. Then the following two statements are equivalent X (i) C∗ env(T + generators to generators, ⋊α G) ≃ OX ⋊α G via a ∗-isomorphism that sends (ii) OX ⋊α G ≃ OX⋊α G via a ∗-isomorphism that sends generators to generators, (Hao-Ng isomorphism) and both imply (iii) all relative crossed products for (T + metrically isomorphic via canonical maps. X ,G, α) are completely iso- If (X,C) is hyperrigid, e.g., ϕX (JX)X = X, then all of the above statements are equivalent. Proof. Assume first that (13) env(T + C∗ X ⋊α G) ≃ OX ⋊α G canonically. Theorem 4.7 shows now that T + X ⋊α G ≃ T + X ⋊TX ,α G ≃ T + X⋊αG canonically and so by taking C∗-envelopes we have a canonical isomor- phism (14) C∗ X env(cid:0)T + ⋊α G(cid:1) ≃ C∗ env(cid:0)T + X ⋊TX ,α G(cid:1) ≃ OX⋊αG. By "equating" the right sides of (13) and (14), we obtain (ii). Conversely, assume that (ii) holds. Then by taking C∗-envelopes in the isomorphisms of Theorem 4.7 we obtain C∗ env(T + env(T + ⋊TX ,αG) ≃ C∗ by (ii). Therefore (i) holds, as desired. ⋊αG) ≃ C∗ env(T + X X X⋊αG) ≃ OX⋊αG ≃ OX ⋊α G Assume now that (ii) is valid, i.e., the Hao-Ng isomorphism is im- plemented via a canonical map. The same map establishes (15) T + X ⋊OX ,α G ≃ T + X⋊α G. By Theorem 4.7 we also have T + (16) From (15) and (16), we obtain T + ⋊α G ≃ T + T + X X X X ⋊α G ≃ T + X⋊α G. ⋊α G ≃ T + env(T + C∗ ⋊ X X ),α G ⋊OX ,α G, or, canonically. By Theorem 2.4, all relative crossed products for (T + are canonically isomorphic, which is (iii). X ,G, α) THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 27 Assume now that (X,C) is hyperrigid and all relative crossed prod- X ,G, α) are canonically isomorphic. Therefore (9) in Theo- ucts for (T + rem 3.6 implies X env(cid:0)T + C∗ ⋊OX ,α G ≃ T + ⋊OX ,α G(cid:1) ≃ OX ⋊α G. ⋊α G and so (i) is valid. By assumption T + with ϕX(JX)X = X is always hyperrigid. X X Finally recall that Theorem 3.1 shows that a C∗-correspondence X The importance of the previous result can not be understated. First, note that condition (i) in Theorem 4.9 is just the equivalence env(A ⋊α G) ≃ C∗ C∗ env(A) ⋊α G of [22, Problem 1] with A = T + X . Hence the equivalence of (i) and (ii) in Theorem 4.9 shows that the Hao-Ng isomorphism and Problem 1 in [22] are actually equivalent problems in the context of tensor algebras of C∗- correspondences. Furthermore, if the Hao-Ng isomorphism holds, then we automatically have from (iii) that OX⋊α G ≃ OX ⋊α G. Therefore a positive resolution for the Hao-Ng isomorphism conjecture also implies a positive resolution for the modified conjecture of [22, page 70]. Also notice that according to condition (iii), the verification of the Hao-Ng isomorphism for hyperrigid C∗-correspondences depends on the canonical identification of two non-selfadjoint operator algebras. We pursue this direction successfully in the next section where we verify the Hao-Ng isomorphism for all graph correspondences of row finite graphs. Furthermore, unlike condition (ii) (Hao-Ng isomorphism), both conditions (i) and (iii) are applicable to arbitrary dynamical systems (T + X ,G, α), i.e., α does not have to be a gauge action. Thus in a sense, a generalization of the Hao-Ng isomorphism problem beyond the realm of gauge actions is possible but only in the language of non-selfadjoint operator algebras. In light of the recent results of Harris and Kim [15], this seems to be a direction worth pursuing. 5. Graph correspondences Following from the last section, we would like to prove that every iso- metric (Toeplitz) representation of ((X,C),G, α) dilates to a covariant (Cuntz-Pimsner) representation. However, the standard proofs that the C∗-envelope of the tensor algebra is the Cuntz-Pimsner algebra are non-constructive [21, 28] which at the moment is a barrier to our method of proof. Significantly. in the case of graph correspondences such a constructive dilation proof is shown to exist. 28 E.G. KATSOULIS AND C. RAMSEY Let (E, V, s, r) be a directed graph, where both E and V are sepa- rable, with associated graph correspondence (X,C, ϕX ). Recall this is where C = c0(V ), X is the completion of cc(E) under the right module structure 0, hαδe, βδfi =(cid:26) αβδs(e), e = f δe · δv =(cid:26) δe, s(e) = v otherwise and the left action of C on X is given by ϕX(δv)δe =(cid:26) δe, r(e) = v otherwise 0, otherwise 0, , . When (X,C) is the graph correspondence for a directed graph (V, E) then OX is ∗-isomorphic to the Cuntz-Krieger algebra of the graph. We wish to find Cuntz-Pimsner representations of these graph cor- respondences. As usual, the main concern is looking at which elements of C are mapped into K(X) by ϕX. In the case of a graph correspondence, Raeburn [31, Proposition 8.8] gives that ϕX(δv) ∈ K(X) if and only if r−1(v) < ∞. Furthermore, δv ∈ ker ϕX if and only if r−1(v) = ∅. Thus, let Vfin = {v ∈ V : 1 ≤ r−1(v) < ∞} be the set of vertices generating Katsura's ideal JX and let K = {(v, w) ∃e ∈ E with s(e) = w, r(e) = v, v ∈ Vfin}. For each pair (v, w) ∈ K, let E(v, w) be the collection of all edges starting from w and ending on v and let [E((v, w)] = CE(v,w). In what follows we will identify the canonical basis of [E(v, w)] with the elements of E(v, w) and use the same symbol for both. Suppose α is a generalized gauge action of (X,C) then it is clear that this induces a permutation of V and in particular that Vfin is invariant under this permutation. By abuse of notation we call this permutation α : V → V . Furthermore, the action α maps [E(v, w)] Indeed, if E(v, w) = {e1, e2, . . . , en} unitarily onto [E(α(v), α(w))]. THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 29 and ξ =Pn i=1 ciδei, then n deδej(cid:17)E n n n n Xj=1 cjδejE(cid:17) ciδei, hα(ξ), α(ξ)i =Dα(cid:16) ciδei(cid:17), α(cid:16) Xi=1 = α(cid:16)D Xi=1 Xj=1 = α(cid:16) ci2δw(cid:17) Xi=1 Xi=1 ci2δα(w) = hξ, ξi . = n Proposition 5.1. Let (X,C) be the graph correspondence of (E, V ) and suppose (ρ, t, u,H) is a completely contractive representation of the dynamical system ((X,C),G, α). There exists a dilation to a completely contractive representation (ρ1, t1, u1,H1) such that for every v ∈ Vfin ρ(δv) = Xe∈r−1(v) t1(δe)t1(δe)∗ Proof. By [28, Lemma 3.5] because (ρ, t,H) is completely contractive then for v ∈ VK we have the matrix inequality [t(e)∗t(f )]e,f ∈r−1(v) ≤ [ρ(he, fi]e,f ∈r−1(v) = ⊕e∈r−1(v)ρ(δs(e)) and so [t(e) : e ∈ r−1(v)] is a row contraction. Hence, and so we can define ∆v := t(δe)t(δe)∗ ρ(δv) ≥ Xe∈r−1(v) pr−1(v)(cid:16)ρ(δv) − Xe∈r−1(v) 1 t(δe)t(δe)∗(cid:17)1/2 . Let Hw = ρ(δw)H and so we can assume H = ⊕w∈V Hw. For each pair (v′, w′) ∈ K let Hv′,w′ := Hv′ ⊗ [E(v′, w′)]. Then for each w ∈ V we define and w ≡ M(v′,w)∈K H+ Hv′,w = M(v′,w)∈K Hw,1 = Hw ⊕ H+ w. Hv′ ⊗ [E(v′, w)]. 30 E.G. KATSOULIS AND C. RAMSEY We combine these to define Hence H1 = ⊕w∈V Hw,1. ρ1(δw) = IHw,1, w ∈ V, extends to a ∗-homomorphism of C that dilates ρ. We also have a continuous unitary representation u1 : G → B(H1) dilating u : G → B(H) and defined as follows. Given g ∈ G and h ∈ H, we let u1(g)h = u(g)h. Otherwise, on each H+ w the operator u1(g) is defined by H+ w ⊇ Hv′,w ∋ h ⊗ ξ 7−→ ug(h) ⊗ αg(ξ) ∈ Hαg (v′),αg(w) ⊆ H+ α(w). It is easy to see that u1 : G → B(H1) is a continuous unitary represen- tation dilating u. We are ready to dilate t : X → B(H). If r(e) /∈ Vfin, then we let t1(e) = t(e). Otherwise, if e ∈ E with s(e) = w and r(e) = v ∈ Vfin, then t1(e) ∈ B(H1) has cokernel contained in Hw ⊕ (. . . 0 ⊕ 0 ⊕ Hv,w ⊕ 0 . . . ) ⊆ Hw ⊕ H+ w ≡ Hw,1 range contained in Hv ⊕ 0 ⊆ Hv ⊕ H+ v and it is given by t1(δe) = [t(δe) ∆vτ (e)] ∈ B(Hw ⊕ Hv,w,Hv), where τ (e) : Hv,w −→ Hv ; h ⊗ ξ 7−→ he, ξi h. (In general, for ζ ∈ [E(v, w)], τ (ζ) will be given by τ (ζ)(h ⊗ ξ) = hζ, ξi h.) It is easy to see that τ (e)τ (e)∗ = IHv and so, Xe∈r−1(v) t1(δe)t1(δe)∗ = Xe∈r−1(v) =(cid:16) Xe∈r−1(v) = ρ(δv) t(δe)t(δe)∗ + ∆vτ (e)τ (e)∗∆v t(δe)t(δe)∗(cid:17) + r−1(v)∆2 v This establishes one of the main conclusions of this proposition and gives that [t1(e)∗t1(f )]e,f ∈r−1(v) ≤ [ρ1(he, fi)]e,f ∈r−1(v) as in the start of the proof. Thus, by [28, Lemma 3.5] again, (ρ1, t1,H1) is a completely contractive representation of (X,C). THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 31 Lastly we must establish the covariance relations. For any g ∈ G recall that αg acts as a unitary between [E(v, w)] = [{e1, e2, . . . , en}] and [E(αg(v), αg(w))] = [{f1, f2, . . . , fn}]. This implies that n n n n t(αg(δei))t(αg(δei))∗ = Xi=1 = = and so n Xk=1 (αg)k,iδfk(cid:1)∗ Xi=1 Xj=1 t(cid:0) (αg)j,iδfj(cid:1)t(cid:0) Xj,k=1 n (αg)j,i(αg)k,i! t(δfj )t(δfk )∗ Xi=1 Xj=1 t(δfj )t(δfj )∗ n u(g)(ρ(δv) − Xe∈r−1(v) t(δe)t(δe)∗) t(αg(δe))t(αg(δe))∗(cid:1)u(g) =(cid:0)ρ(δαg (v)) − Xe∈r−1(v) =(cid:16)ρ(δαg (v)) − Xw∈s(r−1(v)) Xe∈E(v,w) =(cid:16)ρ(δαg (v)) − Xw∈s(r−1(αg (v))) Xf ∈E(αg (v),w) t(δf )t(δf )∗(cid:17)u(g). =(cid:16)ρ(δαg (v)) − Xf ∈r−1(αg (v)) t(αg(δe))t(αg(δe))∗(cid:17)u(g) t(δf )t(δf )∗(cid:17)u(g) By a standard functional analysis trick, u(g)∆v = ∆αg (v)u(g). Furthermore, if h ⊗ ξ ∈ Hv,w, then τ (αg(e))u1(g)(h ⊗ ξ) = τ (αg(e))(u(g)h ⊗ αg(ξ) = hαg(e), αg(ξ)i u(g)h = he, ξi u(g)h = u(g)τ (e)(h ⊗ ξ). Hence, w (cid:1) t1(αg(δe))u1(g) =(cid:2)t(αg(δe)) ∆αg (v)τ (αg(e))(cid:3)(cid:0)u1(g) Hw⊕H+ =(cid:2)u(g)t(δe) ∆αg(v)u(g)τ (e)(cid:3) = [u(g)t(δe) u(g)∆vu(g)τ (e)] = u1(g)t1(δe). 32 E.G. KATSOULIS AND C. RAMSEY It is also immediate that u1(g)ρ1(δv) = ρ1(δαg(v))u1(g). Therefore, (ρ1, t1, u1,H1) is a completely contractive representation of ((X,C),G, α). Theorem 5.2. Let (X,C) be the graph correspondence of (E, V ). Every completely contractive representation of ((X,C),G, α) can be dilated to a Cuntz-Pimsner representation. Proof. Let (ρ0, t0, u0,H0) be a completely contractive representation of ((X,C),G, α). Recursively use the previous proposition to gener- ate a sequence of completely contractive representations (ρn, tn, un,Hn) such that for each n ≥ 1, Hn−1 ⊂ Hn, (ρn, tn, un,Hn) is a dilation of (ρn−1, tn−1, un−1,Hn−1) and for every v ∈ V such that 1 ≤ r−1(v) < ∞ we have Thus, define H′ = ∪∞ Xe∈r−1(v) n=0Hn and tn(δe)tn(δe)∗ = ρn−1(δv). ρ′(c)Hn = ρn(c), t′(ξ)Hn = tn(ξ), and u′(g)Hn = un(g). Hence, (ρ′, t′, u′,H′) is a dilation (ρ0, t0, u0,H0) to a completely con- tractive representation such that for every v ∈ Vfin n→∞ Xe∈r−1(v) n→∞ Xe∈r−1(v) t′(δe)t′(δe)∗ = sot − lim = sot − lim Xe∈r−1(v) t′(δe)IHnt′(δe)∗ tn(δe)tn(δe)∗ = sot − lim = ρ′(δv). n→∞ ρn−1(δv) According to [28, Definition 5.3] the representation (ρ′, t′,H′) is JX- coisometric, i.e., Cuntz-Pimsner in the sense of Katsura but without being isometric. Lastly, by Theorem 4.6 there is a unique minimal isometric coex- tension of (ρ′, t′, u′,H′) to (ρ′′, t′′, u′′,H′′). By [28, Corollary 5.21] this coextension process preserves the property of being JX-coisometric. Therefore, (ρ′′, t′′, u′′,H′′) is an isometric and JX-coisometric dilation of (ρ0, t0, u0,H0), that is, a Cuntz-Pimsner representation of ((X,C),G, α). Corollary 5.3. Let (X,C) be the graph correspondence of a directed graph (E, V ) and ((X,C),G, α) a C∗-correspondence dynamical system, THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 33 that is, α is a generalized gauge action. Then all relative crossed prod- ucts for (T + X ,G, α) are canonically isomorphic via completely isometric maps. Proof. This is an immediate consequence of Theorem 5.2 and Theo- rem 2.4. With an added assumption this gives a positive solution to the Hao- Ng isomorphism problem. Corollary 5.4. Let (X,C) be the graph correspondence of a row-finite directed graph (E, V ) and ((X,C),G, α) a C∗-correspondence dynamical system. Then OX⋊αG ≃ OX ⋊α G. Proof. By the description of JX mentioned at the start of this section and Theorem 3.1, the graph correspondence of a row-finite directed graph (E, V ) is hyperrigid. The conclusion follows from Theorem 4.9. Acknowledgement. Part of this research was carried out in the summer of 2018 while EK was visiting Chongqing Normal University, Chongqing University, the Chinese Academy of Science in Beijing and the Chern Institute of Mathematics. EK would like to thank his hosts Professors Liming Ge, Hanfeng Li, Chi-Keung Ng, Wenhua Qian, Wenmimg Wu and Wei Yuan for the stimulating conversations and their hospitality during his stay at their institutions. References [1] B. Abadie, Takai duality for crossed products by Hilbert C∗-bimodules, J. Op- erator Theory 64 (2010), 19 -- 34. [2] W. Arveson, The noncommutative Choquet boundary II: hyperrigidity, Israel J. Math. 184 (2011), 349 -- 385. [3] E. Bedos, S. Kaliszewski, J. Quigg and D. Robertson, A new look at crossed product correspondences and associated C∗ -- algebras, J. Math. Anal. Appl. 426 (2015), 1080 -- 1098. [4] E. Bedos, S. Kaliszewski, J. Quigg and J. Spielberg, On finitely aligned left cancellative small categories, Zappa-Sz´ep products and Exel-Pardo algebras, Thy. App. Categ. 33 (2018), 1346 -- 1406. [5] D. Blecher and C. Le Merdy, Operator algebras and their modules-an operator space approach, London Mathematical Society Monographs New Series 30, Oxford University Press, 2004. [6] N. Brown and N. Ozawa, C∗ -- algebras and finite-dimensional approximations, Graduate Studies in Mathematics 88, American Mathematical Society, Prov- idence, RI, 2008. xvi+509 pp. 34 E.G. KATSOULIS AND C. RAMSEY [7] V. Deaconu, Group actions on graphs and C∗-correspondences, Houston J. Math. 44 (2018), 147 -- 168. [8] V. Deaconu, A. Kumjian and J. Quigg, Group actions on topological graphs, Ergodic Theory Dynam. Systems 32 (2012),1527 -- 1566. [9] B. Duncan, Certain free products of graph operator algebras, J. Math. Anal. Appl. 364 (2010), 534 -- 543. [10] R. Exel, Amenability for Fell bundles, J. Reine Angew. Math. 492 (1997), 41 -- 73. [11] N. Fowler, Discrete product systems of Hilbert bimodules, Pacific J. Math. 204 (2002), 335 -- 375. [12] E. Gootman and A. Lazar, Crossed products of type I AF C∗-algebras by abelian groups, Isr. J. Math 56 (1986), 267 -- 279. [13] M. Hamana, Injective envelopes of operator systems, Publ. Res. Inst. Math. Sci. 15 (1979), 773 -- 785. [14] G. Hao and C-K. Ng, Crossed products of C∗-correspondences by amenable group actions, J. Math. Anal. Appl. 345 (2008), 702 -- 707. [15] S. Harris and S. Kim, Crossed products of operator systems, J. Funct. Anal. 276 (2018), 2156 -- 2193. [16] E.T.A. Kakariadis, The Dirichlet property for tensor algebras, Bull. Lond. Math. Soc. 45 (2013), 1119 -- 1130. [17] S. Kaliszewski, J. Quigg and D. Robertson, Functoriality of Cuntz-Pimsner correspondence maps, J. Math. Anal. Appl. 405 (2013), 1 -- 11. [18] S. Kaliszewski, J. Quigg and D. Robertson, Coactions on Cuntz-Pimsner al- gebras, Math. Scand. 116 (2015), 222 -- 249. [19] E. Katsoulis, Non-selfadjoint operator algebras: dynamics, classification and C ∗-envelopes, Recent advances in operator theory and operator algebras, 27 -- 81, CRC Press, Boca Raton, FL, 2018. [20] E. Katsoulis, C∗-envelopes and the Hao-Ng isomorphism for discrete groups, Inter. Math. Res. Not. (2017), 5751 -- 5768. [21] E. Katsoulis and D. Kribs, Tensor algebras of C ∗-correspondences and their C∗-envelopes, J. Funct. Anal. 234 (2006), 226 -- 233. [22] E. Katsoulis and C. Ramsey, Crossed products of operator algebras, Memoirs Amer. Math. Soc 258 (2019), no. 1240, vii+85 pp. [23] E. Katsoulis and C. Ramsey, Crossed products of operator algebras: applica- tions of Takai duality, J. Funct. Anal. 275 (2018), 1173 -- 1207. [24] D.W. Kim, Coactions of Hopf C∗-algebras on Cuntz-Pimsner algebras, arXiv: 1407.6106 [math.OA]. [25] D. Larson and B. Solel, Structured triangular limit algebras Proc. London Math. Soc. 75 (1997), 177 -- 193. [26] M. McAsey and P. Muhly, Representations of nonselfadjoint crossed products, Proc. London Math. Soc. 47 (1983), 128 -- 144. [27] A. Morgan, Cuntz-Pimsner algebras and twisted tensor products, arXiv:1601.07826. [28] P. Muhly and B. Solel, Tensor algebras over C ∗-correspondences: representa- tions, dilations, and C ∗-envelopes, J. Funct. Anal. 158 (1998), 389 -- 457. [29] V. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge Studies in Advanced Mathematics 78, Cambridge University Press, 2002. MR1976867 THE NON-SELFADJOINT APPROACH TO THE HAO-NG ISOMORPHISM 35 [30] G. Popescu, Isometric dilations for infinite sequences on noncommuting oper- ators, Trans. Amer. Math. Soc. 316 (1989), 523 -- 536. [31] I. Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics 103, American Mathematical Society, Providence, RI, 2005. [32] J. Schaeffer, On unitary dilations of contractions, Proc. Amer. Math. Soc. 127 (1955), 322. [33] C. Schafhauser, Cuntz-Pimsner algebras, crossed products and K-theory, J. Funct. Anal. 269 (2015), 2927 -- 2946. [34] D. Williams, Crossed products of C∗-algebras, Mathematical Surveys and Monographs, Vol. 134, American Mathematical Society, 2007. Department of Mathematics, East Carolina University, Greenville, NC 27858, USA E-mail address: [email protected] Department of Mathematics and Statistics, MacEwan University, Edmonton, AB, Canada E-mail address: [email protected]
1704.06845	1	1704	2017-04-22T20:35:49	Positive definite functions on Coxeter groups with applications to operator spaces and noncommutative probability	[ "math.OA", "math.CO", "math.GR" ]	A new class of positive definite functions related to colour-length function on arbitrary Coxeter group is introduced. Extensions of positive definite functions, called the Riesz-Coxeter product, from the Riesz product on the Rademacher (Abelian Coxeter) group to arbitrary Coxeter group is obtained. Applications to harmonic analysis, operator spaces and noncommutative probability is presented. Characterization of radial and colour-radial functions on dihedral groups and infinite permutation group are shown.	math.OA	math	Positive definite functions on Coxeter groups with applications to operator spaces and noncommutative probability by Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski Uffe Haagerup ( -- ) in Memoriam Abstract. A new class of positive definite functions related to colour-length function on arbitrary Coxeter group is introduced. Extensions of positive definite functions, called the Riesz-Coxeter product, from the Riesz product on the Rademacher (Abelian Cox- eter) group to arbitrary Coxeter group is obtained. Applications to harmonic analysis, operator spaces and noncommutative prob- ability is presented. Characterization of radial and colour-radial functions on dihedral groups and infinite permutation group are shown. Introduction In  Uffe Haagerup in his seminal paper [haa], essentially proved the positive definitness, for  ≤ q ≤ , of the function Pq(x) = qx = exp(−tx), where · is the word lenght on a free Coxeter group W = Z/∗···∗Z/. From this he deduced also Khinchine type inequal- ities. He has shown that the regular C∗-algebra of W has bounded approximation property and later [dch] the completely bounded approximation property (cbap). These results of Uffe Haagerup have had significant impact on harmonic analysis on free groups and, more generally, on Coxeter groups; they also influenced free proba- bility theory and other noncommutative probability theories. In the paper [bjs] it was shown that the function Pq(x) = qx is positive definite for q ∈ [−, ] and all Coxeter groups, where the length · is the natural word length function on a Coxeter group with repect to the set of its Coxeter generators. This fact implies that infinite Coxeter groups have the Haagerup property and do not have Kazhdan's propery (T).  Mathematics Subject Classification. Primary f, a, l, Secondary a, a. Key words and phrases. Coxeter group, positive definite functions, operator spaces, Sidon sets, Khinchine inequality, length function, de Finetti theorem.   Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski Later, Januszkiewicz [jan] and Fendler [fenb] showed, in the spirit of Haagerup proof, that w 7→ zw is a coefficient of a uniformely bounded Hilbert representation of W for all z ∈ C such that z < . As shown in a very short paper of Valette [val], this implies cbap. See the book [bo] for futher extension of Uffe Haagerup results for a big class of groups. In the paper [bs] Bo zejko and Speicher considered the free prod- uct (convolution) of classic normal distribution N(, ) and the new length function on the permutation group Sn (i.e. the Coxeter group of type a) was introduced, which we shall call the colour-length func- tion k·k. It is defined as follows: for w ∈ Sn in the minimal (reduced) representations w = s . . . sk, where each sj belong to the set S of trans- posions of the form (i, i + ), we put kwk = #{s, s, . . . , sk}. For our study one of the most important results of this paper is that the function called Riesz-Coxeter product Rq defined on all Coxeter groups (W, S) as Rq(s) = qs, for s ∈ S, and Rq(xy) = Rq(x)Rq(y), if kxyk = kxk +kyk is positive definite for  ≤ qs ≤ . This implies, in particular, that in an arbitrary Coxeter group the set of its Coxeter generators is a weak Sidon set and also it is completely bounded Λcb p -set, see Theorems . and .. Equivalently, the span of the linear operators {λ(s)s ∈ S} in the noncommutative Lp-space Lp(W) is completely boundedly isomorphic to row and column op- erator Hilbert space (see Theorem .). Another interesting connection between the two length functions · and k · k appeared in [bs] in the formula for the moments of free additive convolution power of the Bernoulli law µ− = (δ− + δ)/ (cf. Corollary  in cited paper): (−)πq−kπk, mn(cid:16)µ ⊞q −(cid:17) = qnXπ∈P(n) for q ∈ N. (See also Section  of the present paper.) Also, in [bbls] the colour-length function on the permutation group Sn was studied. Some of its extensions to pairpartitions appeared in the presentation of the proof that classical normal law N(, ) is free infinitely divisible under free additive convolution ⊞. Since we have recent extensions of the free probability (which is re- lated to type a Coxeter groups) to the free probability of type b Cox- eter groups (see [beh]), it seems to be interesting to determine the role of the colour-length functions for the Coxeter groups of type b and d. Positive definite functions on Coxeter groups. . .  The plan of the paper is as follows. In Section  we recall definitions of Coxeter groups and of the length and the colour-length functions. In Section  we recall the definition of positive definite funcios and discuss various classes of those, namely radial, colour-radial, and colour-dependant. In Section  we discuss Abelian Coxeter groups. In Section  we show the following formula characterizing the ra- dial normalised positive definite functions on these Coxeter groups which contain the infinite Rademacher group L∞ i= Z/ as a para- bolic subgroup (these include the infinite permutation group S∞): every radial positive definite function ϕ is of the form ϕ(w) =Z  − qw µ(dq) for a probability measure µ. That characterisation is a variation on the classical de Finetti theo- rem. A noncommutative version was shown by Kostler ans Speicher [ks] (see also [leh]). We also show in Theorem ., that the function exp(−twp) is posi- tive definite for all t ≥  if and only if p ∈ [, ]. In Section  we give a short proof of the equivalence of the two known results concerning positive definite functions on finite Cox- eter groups. In Section  we present the main properties of the colour-dependent positive definite functions on Coxeter groups, in particular we show in Proposition .. that on S∞ and some other Coxeter groups, the function w 7→ rkwk is positive definite if and only if r ∈ [, ]. The Section  gives characterization of all colour-length functions on finite and infinite dihedral groups Dm, for m = , , . . . ,∞. In Section  we prove that the set S of Coxeter generators is a weak Sidon set in arbitrary Coxeter groups (W, S) with constant  and that it is also a completely bounded Λ(p) set with contants as C√p, for p > . In Section  we prove for arbitrary finitely generated Coxeter group an identity involving both lengths · and k · k (see Proposition .). We apply it to give a proof of Corollary  from [bs], (see Equation (.)) where the proof, involving probabilistic considerations, was not presented in [bs].  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski . Coxeter groups In this part we recall the basic facts regarding Coxeter groups and introduce notation which will be used throughout the rest of the paper. For more details we refer to [bou, hum]. A group W is called a Coxeter group if it admits the following pre- sentation: W =(cid:28)S(cid:12)(cid:12)(cid:12)(cid:12)n(ss)m(s,s) =  : s, s ∈ S, m(s, s) , ∞o(cid:29) , where S ⊂ W is a set and m is a function m : S×S → {, , , . . . ,∞} such that m(s, s) = m(s, s) for all s, s ∈ S and m(s, s) =  if and only if s = s. The pair (W, S) is called a Coxeter system. In particular, every generator s ∈ S has order two and every element w ∈ W can be represented as (.) w = ss . . . sm for some s, s, . . . , sm ∈ S. If the sequence (s, . . . , sm) ∈ Sm is chosen in such a way that m is minimal then we write w = m and call it the length of w. In such a case the right hand side of (.) is called a reduced representation or reduced word of w. This is not unique in general, but the set of involved generators is unique [bou, Ch. iv, §, Prop. ], i.e. if w = ss . . . sm = tt . . . tm are two reduced rep- resentations of w ∈ W then {s, s, . . . , sm} = {t, t, . . . , tm}. This set {s, s, . . . , sm} ⊆ S will be denoted Sw and called the colour of w. Given a subset T ⊂ S by WT we denote the subgroup generated by T and call it the parabolic subgroup associated with T. To see that Sw is independent of the reduced representation of w notice that (.) s ∈ Sw ⇐⇒ w < WSr{s}. We define the colour-length of w putting kwk = #Sw (the cardinality of Sw). Both lengths satisfy the triangle inequality and we have kwk ≤ w. In the case of the permutation group the colour-length has the fol- lowing pictorial interpretation. If σ is a permutation in Sn+ then kσk equals n minus the number of connected components of the diagram representing σ. Notice, that σ equals to the number of crossings in the diagram (the number of pairs of chords that cross). σ σ kσk e   Positive definite functions on Coxeter groups. . .  () ()() ()()()       It would be convenient to define (.) then, clearly, kwk =Ps∈Skwks. kwks : =( if s < Sw,  if s ∈ Sw, . Positive defined functions Xx,y∈Γ A complex function ϕ on a group Γ is called positive definite if we have ϕ(y−x)α(x)α(y) ≥  for every finitely supported function α : Γ → C. A positive definite ϕ function is Hermitian and satisfies ϕ(x) ≤ ϕ(e) for all x ∈ Γ. Usually it is assumed, that ϕ is normalised, i.e. that ϕ(e) = . In this and the following sections we discuss the radial functions on Coxeter groups. These are functions which depend on w rather then on w. We call a function ϕ on (W, S) colour-dependent if ϕ(w) depends only on Sw. We call it colour-radial if it depends only on kwk. An Abelian Coxeter group generated by S is isomorphic to the direct product ⊕s∈SZ/. On these groups the lengths · and k · k coincide and all functions are colour dependent. s The main example of a positive definite function will be the Riesz -- Coxeter function. Given a sequence q = (qs)s∈S we define Rq(w) = Qs∈S qkwks =Qs∈Sw qs. We will abuse notation and denote by Rq also the associated operatorPw∈W Rq(w)w. That is Rq =  +Xs∈S qsqsw + Xw:Sw={s,s,s} qss + Xw:Sw ={s,s} qsqsqsw + . . .  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski In the case all qs = q we get Rq =P qkwkw. This generalises the classical case of Rademacher -- Walsh functions in the Rademacher group Radn. If we denote the generator of the i-th factor Z/ of the latter by the symbol ri then, by definition, r  =  and ri rj = rj ri and Rq = nYi= ( + qiri ). . Rademacher groups In this section we are going to study positive definite radial func- tions on the Abelian Coxeter groups, (W, S) = RadS. Since positive definiteness is tested on functions with finite support, we can as- sume that S is countable. If #S = n we will write Radn instead of RadS. Given n ∈ N ∪ {∞}, we denote by Pn the class of all func- tions f : {, , . . . , n} → R for n finite and f : N → R if n = ∞ such that ϕ(w) = f (w) is a normalised positive definite on Radn. The following observation is straightforward. Proposition .. Assume that  ≤ m < n ≤ ∞ and f ∈ Pn. Then the restriction of f to {, . . . , m} belongs to Pm. A fuction f belogs to P∞ if and only if all its restrictions to {, . . . , m} for any m ∈ N belong to Pm. Theorem .. Assume n is finite. The set Pn form a simplex whose ver- l−i(cid:1), where  ≤ l ≤ i(cid:1)(cid:0)n−k tices (extreme points) are f n n. Equivalently, every normalised radial positive definite function on the group Radn is of the form i=(−)i(cid:0)k l(cid:1)−Pl l (k) =(cid:0)n nXl= λl f n ϕ(x) = l (x), l= λl = . l= is unique and satisfies where the sequence of nonnegative numbers (λl)n Pn Proof. We can indentify the dual [Radn group of Radn with Radn via the paring (x, y) = (−)Pn i= xi yi . By Bochner's theorem every nor- ϕ(x) =Z malised positive definite function ϕ on Radn is of the form (x, y) µ(dy), for some probability measure µ. Clearly, such a function is radial if and only if µ is invariant under the action of Sn. dRad∞ Positive definite functions on Coxeter groups. . .  Among such measures extreme ones are measures µl for  ≤ l ≤ n, where µl is equally distributed among elements of length l. More- over, ϕ(x) =Z (x, y) µl(dy) = f n l (x) dRad∞ (cid:3) as claimed. The following theorem is a version of the classical de Finetti Theo- rem (see [fel, p. ]) for the infinite Rademacher group. Theorem .. Assume that ϕ is a radial function on the Rademacher group Rad∞. Then ϕ is a normalised positive definite if and only if there exists a probability measure µ on [−, ] such that ϕ(x) =Z  − qxµ(dq). This measure µ is unique. Proof. Since the function qx is normalised positive definite for q ∈ [−, ], the "if" implication is obvious. Assume that ϕ is normlised positive definite. The group Rad∞ is dis- crete and Abelian and its dual is the compact group [Rad∞ =Q∞i= Z/. By Bochner's theorem, there exists a probability measure η on dRad∞ such that (x, y) dη(y), ϕ(x) =Z dRad∞ where for x = (x, x, . . .) ∈ Rad∞, y = (y, y, . . .) ∈ dRad∞ we put (x, y) = (−)P∞i= xi yi . The radiality of ϕ is equivalent to the fact that for every permutation σ ∈ S∞ we have ϕ(x) = ϕ(σ(x)), where σ(x) = (xσ(), xσ(), . . .). This, in turn, implies that η is σ-invariant for every σ ∈ S∞, i.e. we have η(A) = η(σ(A)) for every Borel subset A ⊂ dRad∞. Cn(ǫ) = {y ∈ dRad∞yi = ǫi :  ≤ i ≤ n}, i= ∈ {, }n we define Cn(ǫ) ⊆ dRad∞ by in particular C(∅) = dRad∞. Then we have η(Cn(ǫ)) = η(Cn(ǫ′)) if ǫ′i = ǫσ(i) for some σ ∈ Sn and every  ≤ i ≤ n. For ε ∈ R we put For a sequence ǫ = (ǫi)n εn = ε, ε, . . . , ε {z }n and an = η(Cn(n)). Moreover, for n, k ≥  we define the difference operators ∆kan by induction: ∆an = an and ∆k+an = ∆kan+ − ∆kan. We claim that (.) (−)k∆kan = η(cid:16)Cn+k(cid:16)nk(cid:17)(cid:17) .  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski Denoting the right hand side of (.) by c(n, k) we note that c(n, ) = an and Cn+k+(cid:16)nk(cid:17) ∪ Cn+k+(cid:16)nk(cid:17) = Cn+k(cid:16)nk(cid:17) , is a disjoint union. This implies This formula, by induction on k, leads to (.). c(n, k + ) = c(n, k)− c(n + , k). From (.) we see that the sequence (an) is completely monotone, i.e. that (−)k∆kan ≥  for all n, k ≥ . By the celebrated theorem of Haus- dorff (see [hau, Satze ii und iii]), there exists a unique probability measure ρ on [, ] such that (.) (−)k∆kan =Z   un(− u)k dρ(u). (Note that Equation (.) for arbitrary k ≥  follows from the case k = .) For x = (n∞) ∈ Rad∞ so that x = n, we have ϕ(x) =Z (x, y) dη(y) =Z dRad∞ = Xǫ∈{,}n (−)Pn k!(−)kZ  n nXk= =Z  qn dµ(q), =  dRad∞ i= ǫi η(Cn(ǫ)) = i= yi dη(y) (−)Pn k!(−)kη(cid:16)Cn(cid:16)kn−k(cid:17)(cid:17) n nXk= (− u)n dρ(u) uk(− u)n−k dρ(u) =Z   − where µ is defined by µ(A) = ρ(cid:16)   +   A(cid:17) for a Borel set A ⊆ [−, ]. (cid:3) . Remarks on radial positive definite functions on some infin- itely generated Coxeter groups In this Section we extend the last theorem of the previous section to a certain class of Coxeter groups. Theorem .. Assume that (W, S) is a Coxeter system and that there is an infinite subset S ⊆ S such that st = ts for s, t ∈ S. Assume that ϕ is a radial function on W with ϕ(e) = . Then ϕ is positive definite if and only if there exists a probability measure µ on [−, ] such that ϕ(σ) =Z  − qσµ(dq). This measure µ is unique. Positive definite functions on Coxeter groups. . .  Proof. It is sufficient to note that the group generated by S is a par- abolic subgroup isomorphic with Rad∞. (cid:3) Example. For W = S∞ we have S = {(n, n + ) : n ∈ N}. Then we can take S = {(n − , n) : n ∈ N}. Similar S can be found in infinitely generated groups of type b and d. Problem .. When − ≤ q ≤ , q ,  is the positive definite function qx on S∞ an extreme point in the set of normalised positive definite functions? Theorem .. The function ϕp(σ) = e−tσp and only if  ≤ p ≤ . Proof. A contrario. Assume that for some p >  and t >  the function ψp(σ) = e−tσp For q = e−t, choosing σ such that σ = n we have q for some probability measure µ on [−, ]. Since(cid:16)R  tends to max{qq ∈ supp µ} while(cid:18)q =R  qn dµ(q)(cid:17)/n → , we conclude that is positive definite on S∞ if is positive definite on S∞. µ is the Dirac measure at , which is a contradiction. The "if" part is standard. We need to show that f (x) = e−txp is the Laplace transform of some probability measure supported on [,∞), so f is a convex combination of functions of the form e−sx. By characterisation of Laplace transforms (see [hau, Satz iii]) this is equivalent to complete monotonicity, that is (−)nf (n) >  for all n = , , . . . . And indeed, by induction, (−)nf (n) is a positive linear combination of positive functions of the form xpj−nf (x) for  ≤ j ≤ n. (cid:3) (n)p  (cid:19)/n qn dµ(q) − (n)p  − The measures with Laplace transforms e−txp for t ≥  and  ≤ p ≤  are studied in detail in [yos, Ch. ix.] (see Propositions  and  there). Let us note that for such groups like Zk or Rk with the Euclidean distance d the functions exp(−tdp) are positive definite for all t ≥  and  ≤ p ≤  (the case p =  corresponds to the Gaussian Law). . The longest element If a Coxeter group W is finite, then it contains the unique element ω◦ which has the maximal length with respect to ·.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski From the definition it is clear, that a function ϕ on a group Γ with values in the field of complex numbers C is positive definite if and onlyPg∈Γ ϕ(g)g is a nonnegative (bounded if the group is finite) op- erator on ℓΓ. (We will identify g ∈ Γ with λ(g) ∈ B(ℓΓ), where λ is the left regular representation, for short.) Let W be a finite Coxeter group. The following two statements are well known. (a) The function qw is positive definite for any  ≤ q ≤ . (b) The function ∆(w) = ω◦/−w is positive definite. The first one was proven in [bjs] (even for infinite Coxeter groups and also for − ≤ q ≤ ) while the second -- in [bs, Proposition ]. Here we give a short direct prove of the following. Proposition .. The above statements (a) and (b) are equivalent. Proof. Let q = e−t (with t ≥ , as we assume q ≤ ). The case (a) is equivalent to Φt =Pw∈W et∆(w)w = etω◦/Pw∈W qww being nonnega- tive. Assume (a). Recall first, that ω◦w = ω◦ − w = wω◦. Therefore ω◦/ − ω◦w = −(ω◦/ − w), ie. ∆(ω◦w) = −∆(w) and similarly, ∆(wω◦) = −∆(w). The equality ∆(ω◦w) = ∆(wω◦) implies that ω◦ (and thus Q = ( − ω◦)/) commutes with ∆ (and thus Φt). Since Q = Q is nonnegative we conclude that t sinh(t∆(w)) et(ω◦/−w) − et(ω◦/−wω◦) t−ΦtQ = Xw∈W is nonnegative. Therefore, taking the limit as t → , we obtain that Pw∈W ∆(w)w is nonnegative. Thus (b). Assuming (b) and using the Schur lemma, which says that the (point- wise) product of positive definite functions is positive definite, we get that w = Xw∈W w. t Φt = Xw∈W et∆(w)w =Xn≥ is nonnegative. Thus (a). tn n!Xw∈W ∆(w)nw (cid:3) . Colour-dependent positive definite functions on Coxeter groups The question which colour-dependant or colour-radial functions are positive functions on Coxeter groups is wide open. In this section Positive definite functions on Coxeter groups. . .  we provide some sufficient conditions. In the next section we will examine the dihedral groups in full details. Lemma .. Let H be a subgroup of a group Γ of index d. Then the func- tion ϕr defined by ϕr(x) =  if x ∈ H and ϕr(x) = r otherwise is positive definite on Γ if and only if r ∈ [−/(d − ), ], with natural convention that if d = ∞ then −/(d − ) = . Note, that if H = {} then d = Γ. Proof. First assume that d is finite and let us enumerate the left cosets: {g H : g ∈ Γ} = {H, H, . . . , Hd}. Note, that for x ∈ Hi, y ∈ Hj we have y−x ∈ H if and only if i = j. Therefore, for r = −/(d − ) and for a finitely supported complex function f on Γ we have − ds −  ≤ qs ≤ , where ds denotes the index if the parabolic subgroup generated by S r{s} in W: ds = [W : WSr{s}]. Then the Riesz -- Coxeter Rq is positive definite on W. Proof. From Lemma . the function w 7→ qkwks is positive definite for s ∈ S and −/(ds − ) ≤ qs ≤ . Since the pointwise product of positive definite functions is positive definite, the statement holds. (cid:3) s (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xx∈Hi  , f (y)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕr(y−x)f (x)f (y) = Xx,y∈Γ  d −  X≤i<j≤d f (x)−Xy∈Hj which proves that ϕr is positive definite. For r ∈ [−/(d − ), ] the function ϕr is positive definite as a convex combination of ϕr and the constant function ϕ. On the other hand, if we choose xi ∈ Hi for each i ≤ d and define f as the characteristic function of the set {x, . . . , xd} then ϕr(y−x)f (x)f (y) = d + (d  − d)r, (.) Xx,y∈Γ which proves that r ≥ −/(d − ) is a necessary condition for positive definiteness of ϕr. If d = ∞ then r =  and the function ϕ is positive definite as the characteristic function of the subgroup H. For "only if" part we chose an arbitrarily long sequence x, . . . , xd′ of elements from dif- ferent left cosets and use (.) with d′ instead of d. (cid:3) Theorem .. Assume that for every s ∈ S we are given a number qs,  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski Example. Take W = Sn, the permutation group on the set {, , . . . , n}. It is generated by the transpositions S = {si = (i, i + ),  ≤ i ≤ n − }. For  ≤ i ≤ n −  the parabolic subgroup generated by S r {si} is i(cid:1). isomorphic with Si− × Sn−i−, so its index is i(cid:0)n It would be interesting to determine for which r the function w 7→ rkwk is positive definite. By Proposition . this holds for r ∈ [−/(d − ), ], where d is the maximal index of the parabolic subgroups of the form WSr{s}. We note a necessary condition. Proposition .. Assume that we have distinct generators s, s, . . . , sn ∈ S such that ssk , sks (i.e. m(s, sk) > ) for  ≤ k ≤ n. If the function w 7→ rkwk is positive definite on W, then − /(n − ) ≤ r ≤ . If there is an element s ∈ S for which there are infinitely many s ∈ S such that ss , ss then rkwk is positive definite on W if and only if  ≤ r ≤ . Proof. Consider elements wk = ssks. Note, that for k , l we have kw− l wkk = . If ϕr is positive definite on W then we have ϕr(x− l xk) = n + (n − n)r,  ≤ nXk,l= which implies r ≥ − /(n − ). Corollary .. The function w 7→ qkwk on S∞ is positive definite if and only if  ≤ q ≤ . Problem .. Thus, it is valid to ask the following. Is it true that ev- ery normalised positive definite colour-lenght-radial function φ : S∞ → R is of the form φ(σ) =R  on [, ]? qkσk dµ(q) for some probability measure µ (cid:3)  . Dihedral groups In this part we are going to examine the class of colour-dependent positive definite functions on the case the simplest nontrivial Cox- eter groups. Assume that W = Dn = hs, t(st)ni (i.e. the group of sym- metries of a regular n-gon), and define a colour-dependent function on W: (.) φ(w) =  if w = e, p if w = s, if w = t, q otherwise. r  If p = q then φ is colour radial. We are going to determine for which parameters p, q, r the function φ is positive definite on W. It is easy Positive definite functions on Coxeter groups. . .  to observe necessary conditions: p, q, r ∈ [−, ]. Moreover, since hsti is a cyclic subgroup of order n, Lemma ., implies a necessary con- dition: −/(n − ) ≤ r ≤ . Finite dihedral groups. Assume that W is a finite dihedral group, W = Dn, so that (st)n = . We will use the following version of Bochner's theorem: A function f on a compact group G is positive definite if and only if its Fourier transform: is a positive operator for every π ∈bG, wherebG denotes the dual ob- ject of G, i.e. the family of all equivalency classes of unitary irre- ducible representations of G, see [sim]. Then we have bf (π) =ZG f (x)π(x−)dx f (x) =Xπ∈bG dπ trhbf (π)π(x)i . nXg∈G bφ(π) = φ(g)π(g−).  Therefore, for every irreducible representation π of Dn we are going to find We will identify s with (,−) and t with (,−). If n is odd then Dn possesses two characters: χ+,+ such that χ+,+(w) =  for every w ∈ Dn and χ−,− such that χ−,−(s) = χ−,−(t) = −. If n is even then we have two additional characters χ+,− and χ−,+ such that χ+,−(s) = χ−,+(t) =  and χ+,−(t) = χ−,+(s) = −. It is easy to check that which gives and, for n even, which implies nbφ(χ+,+) =  + p + q + (n − )r, nbφ(χ−,−) = − p − q + r, −− (n − )r ≤ p + q ≤  + r nbφ(χ+,−) =  + p − q − r, nbφ(χ−,+) = − p + q − r, p − q ≤ − r. Ua(k, ) = eπika/n Ua(k,−) = e−πika/n    e−πika/n! , ! , eπika/n  We have also the family of two dimensional representations Ua:  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski  k. Then for the function given by (.) we have where a = , , . . . ,j n− = p − r + (q − r)e−πia/n nbφ(Ua) = (− r)Id + (p − r)Ua(,−) + (q − r)Ua(,−) ! . p − r + (q − r)eπia/n − r − r This matrix is positive definite if and only if r ≤  and Therefore we have (cid:12)(cid:12)(cid:12)p − r + (q − r)eπia/n(cid:12)(cid:12)(cid:12) ≤ − r. Proposition .. The function φ given by (.) is positive definite on Dn if and only if  + p + q + (n − )r ≥ , − p − q + r ≥  (plus whenever n is even) and  + p − q − r ≥ , − p + q − r ≥  (cid:12)(cid:12)(cid:12)p − r + (q − r)eπia/n(cid:12)(cid:12)(cid:12) ≤ − r. for a = , , . . . ,j n−  k. Let us confine ourselves to colour-radial functions. Corollary .. Assuming that p = q, the function φ defined by (.) is positive definite on W = Dn if and only if i.e. if and only if the point (p, r) belongs to the triangle whose vertices are max(−p −  n −  , p − ) ≤ r ≤  + (n − ) cos(π/n)! , − cos(π/n) − n − cos(π/n)  + (n − ) cos(π/n) ,  + p cos(π/n)  +  cos(π/n) , (cid:18) n −  n −  , − n − (cid:19) , (, ). Proof. For p = q the conditions from Proposition . reduce to p −  ≤ r, −− p ≤ (n − )r, and  cos(π/n)p − r ≤ − r. It is sufficient to note that p−  ≤ r implies  cos(π/n)(p− r) ≤ − r for p ≤ . Example. For D we have the positive definiteness of φ is equivalent to (cid:3) − +p + q ≤ r ≤ −p − q, which means that the set of all possible (p, q, r) forms a tetrahedron with vertices (−, ,−), (,−,−), (−,−, ), (, , ). For p = q the condition reduces to p−  ≤ r ≤ . In the case of D Proposition . leads to the following conditions: Positive definite functions on Coxeter groups. . .  − p − q + r ≥ ,  + p + q + r ≥ , − r ≥pp + q + r − pq − pr − qr, which can be expressed as max(−− p − q  , p + q − ) ≤ r ≤ − p − q + pq − p − q . The infinite dihedral group. Here we are going to study W = D∞. Proposition .. The function φ given by (.) is positive definite on W = D∞ if and only if  ≤ r and p − r +q − r ≤ − r, i.e.  + p + q (.) max{, p + q − } ≤ r ≤ min(−p − q, ) .  Proof. First we note that the set of (p, q, r) ∈ R satisfying (.) con- stitutes a pyramid which is the convex hull of the points (±, , ), (,±, ) and (, , ) (apex). For these particular parameters it is easy to see that φ is positive definite: (, , ) corresponds to the constant function , (, , ) to the characteristic function of the sub- group hsi = {, s}, and (−, , ) to the character χ−,− times the char- acteristic function of hsi. Similarly for (,±, ). This, by convexity, proves that (.) is a sufficient condition. On the other hand, we know already that r ≥  is a necessary con- dition. Let us fix n and define W+(n) = {x ∈ W : sx < x ≤ n}, W−(n) = {x ∈ W : tx < x ≤ n} and f (x) =(± if x ∈ W±(n), otherwise.  For x, y ∈ W+(n) we have Sy−x = ∅ in n cases (namely, if x = y) Sy−x = {s} in n −  cases (namely if x = k, y = k +  or vice- versa, k = , . . . , n − ) Sy−x = {t} in n cases (namely if x = k, y = k −  or vice-versa, k = , . . . , n) and Sy−x = {s, t} in all the other (n − )(n − ) cases. Similarly, for x, y ∈ W−(n) we have Sy−x = ∅ in n cases, Sy−x = {s} in n cases, Sy−x = {t} in n −  cases and Sy−x = {s, t} in (n − )(n − ) cases. If x ∈ W+(n), y ∈ W−(n) or vice- versa then Sy−x = {s, t}. Summing up, we get Xx,y∈W φ(y−x)fn(x)fn(y) = n + (n − )p + (n − )q + (n − )(n − )r − nr = n + (n − )p + (n − )q − (n − )r.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski    +(cid:18)− Therefore for every n ∈ N we have a necessary condition n(cid:19) r ≥ . n(cid:19) p +(cid:18)− Letting n → ∞ we get  + p + q ≥ r. Put xk = stst . . ., xk = k. Fix n and define n(cid:19) q −(cid:18)−  g(x) =(χ−,+(x)  if x = xk for  ≤ k ≤ n, otherwise, where, as before, χ−,+ is the character on W for which χ−,+(s) = −, χ−,+(t) = . Then φ(y−x)g(x)g(y) = Xx,y∈W nXk,l= φ(x− l xk)g(xk)g(xl). Denote ck,l = φ(x− l xk)g(xk)g(xl). Then we have ck,k = ,  ≤ k ≤ n, ck,k− = q if k is even, ck,k− = −p if k is odd,  ≤ k ≤ n and ck,l = cl,k for all  ≤ k, l ≤ n. If  ≤ k, l ≤ n and k − l ≥  then ck,l = (−)j r, where j is the total number of s appearing in xk and xl. Now it is not difficult to check that nXl= ck,l =( + q − r − p + q − r if k =  or k = n, if  < k < n, which implies Xx,y∈W φ(y−x)g(x)g(y) = n − (n − )p + nq − (n + )r and leads to necessary condition r ≤ − p +q. In a similar manner we get r ≤  + p − q. Finally, define a function h similarly like g, but now we use the char- acter χ−,−: h(x) =(χ−,−(x) = (−)k if x = xk for  ≤ k ≤ n, otherwise.  Putting dk,l = φ(x− l xk)h(xk)h(xl) we have dk,k = , dk,k− = −p if  ≤ k ≤ n is even and dk,k− = −q if k is odd. Moreover, if k − l ≥ ,  ≤ k, l ≤ n then dk,l = (−)k+l r. Now one can check that if k =  or k = n, if  < k < n, dk,l =(− q nXl= − p − q + r which yields − p − q + r ≥  and completes the proof that the condi- tions (.) are necessary. (cid:3) . Weak Sidon sets and operator Khinchin inequality Positive definite functions on Coxeter groups. . .  The aim of this section is to show that the set of Coxeter generators S in an arbitrary Coxeter group W is a weak Sidon set, ie. an interpo- lation set for the Fourier -- Stieltjes algebra B(W). Given a group Γ, the Fourier -- Stieltjes algebra consists of linear com- binations of positive definite functions on Γ, ie. every element of B(Γ) is of the form f = ϕ − ϕ + i(ϕ − ϕ) for some positive definite functions ϕi ( ≤ i ≤ ) on Γ. The norm on B(Γ) is defined as kf kB(Γ) = infnXϕi(e)where f decomposes as aboveo Theorem .. The set of Coxeter generators S in an arbitrary Coxeter for every bounded function f : S → group W is a weak Sidon set, ie. [−, ] there exists positive definite functions ϕ±, such that f (s) = ϕ+(s)− ϕ−(s) for any s ∈ S. One can take ϕ± = Rq± for a suitable choice of q±. Moreover kϕ+ − ϕ−kB(W) ≤  Proof. Put S±(f ) = {s ∈ S± f (s) > }. Set q±s =(±f (s)  for s ∈ S±(f ), otherwise. Then f (s) = Rq+(s) − Rq−(s) as claimed. The rest of the statement hold as the Riesz-Coxeter function at the identity element equals to one. (cid:3) Given a matrix A ∈ Mn(C) and p ≥  the Schatten p-class norm kAkSp is defined as kAkp Sp finite sum f =P cgλ(g) ∈ C[Γ] we define noncommutative Lp-norm Let λ denote the left regular representation of a group Γ. Given a = (trA)/p, where A = (A∗A)/. Lp(Γ) =(cid:16)τ(cid:16)(f ∗ ∗ f )/(cid:17)(cid:17)/p kf kp where τ(f ) = ce is the von Neumann trace and Lp(Γ) is a completion of C[Γ] with respect to the above norm. We recall, that a scalar-valued map ϕ on a group Γ is called a com- pletely bounded Fourier multiplier on Lp(Γ) if the associated operator Mϕ(λ(g)) = ϕ(g)λ(g), g ∈ Γ extends to a completely bounded operator on Lp(Γ). We let Mcb(Lp(Γ)) to be an algebra of completely bounded Fourier multipliers equipped with the norm kϕkMcb(Lp(Γ)) = kMϕ ⊗ idS p k.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski Following Pisier [pis], for as ∈ Mn(C), where s ∈ S, we define For a set E ∈ Γ we define the completely bounded constant Λcb infimum of C such that p (E) as ,(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (a∗sas)/(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)S p Xs∈S .  ≤ Ck(as)s∈SkR∩ C k(as)s∈SkR∩ C = max Xs∈S (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k(as)s∈SkR∩ C ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xs∈S (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (asa∗s)/(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)S p Xs∈S as ⊗ λ(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(W) as ⊗ λ(s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp(W) for all matrices as ∈ Mn(C) and n ∈ N. Theorem .. If as ∈ Mn(C), then for all p ≥  and any Coxeter system (W, S) we have ≤ A′√pk(as)s∈SkR∩ C . Proof. It was shown by Harcharras [lp, Prop. .] that Λcb p (E) if finite if and only if E is an interpolation set for Mcb(Lp(Γ)), i.e. every bounded function on E can be extended to a multiplier, and Λcb p (E) ≤ Λcb p (R)µcb p (E), where R is the generating set in the Rademacher group Rad∞ and µcb p (E) is the interpolation constant. As shown by Buchholz [buc, Thm. ] for p = n, and S te stan- dart generating set in Rad∞, Λcb absolute A. This was extended by Pisier [pis, Thm. ..] for any p ≥ , i.e n(R) = ((n − )!!)/n ≤ A√p for some p (R) ≤ A′√p, Λcb for an absolute constant A′. We have shown in Theorem . that in an arbitrary Coxeter group W its Coxeter generating set S is a weak Sidon set, i.e. it is interpolation set for the Fourier -- Stieltjes algebra B(W). Since for p ≥ , B(Γ) is a subalgebra of Mcb(Lp(Γ)) and kϕkMcb(Lp(Γ)) ≤ kϕkB(Γ), p (S) ≤ . Thus Λcb we see that µcb of the right inequality. The left inequality holds for any group Γ and any S ⊂ Γ (see [lp]). p (S) ≤ A′√p. This finishes the proof (cid:3) Positive definite functions on Coxeter groups. . .  Remark .. Fendler [fena] has shown that if for all s, t ∈ S, s , t, we have ms,t ≥ , then Λcb p (S) ≤ √. See also [bo z] and [buc] for related results in the case of free Coxeter groups. Also Haagerup and Pisier have shown that Λcb ∞ (S) = , where Λcb p (E) [hp]. See the paper of Haagerup [haa] where the best constant was calculated for the set of Cox- eter generators of the Rademacher group in case when as are scalars. ∞ (E) = supp≥ Λcb . Chromatic length function for Coxeter groups and pairparti- tions Let [n] = {, . . . , n}. Let [n] denote the set of subsets of [n]. By a partition of [n] we mean π ⊂ s[n] such that S π = [n] and if π′, π′′ ∈ π then π′ = π′′ or π′ ∩ π′′ = ∅. We say, that partition is a coarsening of a partition π if for any π′ ∈ π there exists ′ ∈ such that π′ ⊂ ′. A partition is called crossing if there exist  ≤ a < b < c < d ≤ n and π, π ∈ π with a, c ∈ π , π ∋ b, d; otherwise it is called noncrossing. For any partition π there exists th the smallest noncrossing coars- ening Φ(π) of π (ie. if is a noncrosing coarsening of π then it is a coarsening of Φ(π)). We define kπk = n − #Φ(π). The notion for the map Φ was introduced in [by]. We say that π is a pairpartition if every member of π has cardinality two. The set of pairpartitions of [n] is denoted by P(n). Given π ∈ P(n) we write π to denote the number of ordered quadruples  ≤ a < b < c < d ≤ n such that both {a, c} and {b, d} belong to π. Note, that π =  precisely when π is noncrossing. The set of noncrossing pairpartitions is denoted N C(n). Given a noncrossing pairpartition we call {b, c} ∈ an inner block if there exists {a, d} ∈ with a < b < c < d. The number of inner blocks of we denote as inn(). In [bs, Cor. ] Bo zejko and Speicher observed the following iden- tity. (.) Let fn(q) =P∈N C(n)(− q)inn(). It is elementary to derive (− q)inn(). Xπ∈P(n) (−)πqkπk = X∈N C(n) n +  (cid:12)(cid:12)(cid:12)(cid:12) q(cid:19) , fn(q) = Cn F(cid:18) n, − n  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski j= tn n(cid:1)−(cid:0) n j qj, then the triangle (tn n−(cid:1) = #N C(n) denote the n-th Catalan number where Cn =(cid:0)n and F is the classical hypergeometric funcion. If we write f (q) = Pn− j )≤j<n appears in [slo] as "sequence" a). Since we are not going to use this formula, we leave it as an exercise to the reader. For the expansion of fn( − t) and the Delanoy triangle appearing there the reader may consult [bw, Prop .]. In what follows, we prove a result about an arbitrary finitely gen- erated Coxeter group, which for permutation groups implies the above one. Given a permutation σ ∈ Sn, we construct a pairparti- tion σ = {{i, n +  − σ(i)} ≤ i ≤ n}. Note, that σ is equal to the length of σ with respect to the Coxeter generators (, ), . . . , (n− , n) of Sn. Therefore, denoting by w the Coxeter length of an element w of some Coxeter group W will not lead into any confusion. It is also clear that, with respect to the identification of permutations with a subset of pairpartitions, the two definitions of k · k agree (see Equations (.) and (.)). By W(t) we denote a growth series of a finitely generated Coxeter (Note, that the coefficient at t equals to #S. This explains why here and in the rest of this section we consider only finitely generated Coxeter groups. We will not repeat this assumption for short.) More- group W (length function). That is, a power series W(t) =Pw∈W tw. over, for X ⊂ W we write X(t) =Pw∈X tw. Let us define a multivariable formal power series (chromatic length function. For any X ⊂ W define X(t, q) =Xw∈X twYs∈Sw qs. In particular X(t) = X(t, ), where  = ()s∈S. Proposition .. The polynomial (or formal power series, if W is infi- nite) W(t, q) satisfies W(t, q) =XT⊂S WT(t)Yr∈T rs Ys∈SrT (− qs). Proof. Let W◦R denote the set of all elements of WR not contained in any proper parabolic subgroup of WR, ie. W◦R : = WR −ST R WT. Then, by inclusion-exclusion principle, W◦R(t) =PT⊂R(−)#(R−T)WT(t). Positive definite functions on Coxeter groups. . .  Therefore, qr twYs∈Sw W(t, q) = Xw∈W W◦R(t)Yr∈R qs =XR⊂S WT(t)(−)#(R−T)Yr∈R =XR⊂SXT⊂R WT(t)Yr∈T =XT⊂S qr XT⊂R⊂S Ys∈RrT qi Ys∈SrT WT(t)Yr∈T =XT⊂S (− qs). qr (−qs) (cid:3) Corollary .. If W is a finite Coxeter group then (.) W(−, q) =Ys∈S (− qs). Proof. Choose s ∈ T and put W{s}T = {w ∈ W : w < ws}. Clearly, WT = W{s}T WT therefore WT(t) = W{s}T (t)W{s}(t). Since W is a finite group, W{s}T is a polynomial. Thus WT(−) =  if T is nonempty (and W∅(−) = ). In order to prove Equation (.) we define the Wick map P(n) ∋ π 7→ :π: ∈ N C(n) (related to the normal order in quantum field theory). Given a pairpartition π we define :π: by repetitive resolving crossings. That is, we replace repetitively every crossing pair {a, c} and {b, d} with a < b < c < d by {a, d} and {b, c}. In order to see that the result is independent of the order of resolution we describe :π: in an equivalent way. (cid:3) Let Φ(π) be the smallest noncrossing coarsening of π. For each block β of Φ(π) define β+ = {y(∃x) x ∈ β, y > x, {x, y} ∈ π} and β− = {x(∃y) y ∈ β, y > x, {x, y} ∈ π}. Order β+ = {y, . . . , yk} in in- creasing way and β− = {x, . . . , xk} in decreasing way. Then all pairs {xi, yi} will be parts of :π:. Equation (.) will follow from a more refined statement. Proposition .. For every ∈ N C(n) (.) (−)πqkπk = (− q)inn(). Xπ∈P(n) :π:= Proof. Let us first consider the case of =  = {(i, n + − i) ≤ i ≤ n}. Clearly, {π:π: = } = {σσ ∈ Sn}. And Equation (.) is equivalent to Equation (.) (as all qs are set to q) for W = Sn.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski π :π: Φ(π) Figure. Examples of π, :π:, and Φ(π). In a general case observe, that Φ(π) is a coarsening of = :π:. Yet, not every coarsening may appear. The obvious condition is that for each block β of :π: the pair {min β, max β} belong to . For the purpose of this proof we will call such a coarsening admissible. Its clear, that abmissible coarsenings ρ are in one to one correspondence with subsets of containing all outer (not inner) parts of of the form {{min ρ′, max ρ′}ρ′ ∈ ρ}. Let us refine Equation (.) further. For every ∈ N C(n) and any admissible coarsening η of we have Xπ∈P(n) :π:=, Φ(π)=ρ (.) (−)π = (−)#ρ. Equation (.) follows from (.) by multiplying by qn−#η and sum- ming over all admissible coarsenings η of . Equation (.) is again equivalent to to Equation (.) (for all aper- mutation groups and all qs set to q) as both sides factor as a product over blocks of ρ. (cid:3) Question .. We have proven Equation . with the help of an em- bedding Sn ∋ σ 7→ σ ∈ N C(n) (or several such embeddings, one for each outer block of ). Corollary . holds for any Coxeter group. Is there a corresponding formula concerning some generalization of pairpartitions? In the proof of Proposition . we have not assumed that W was finite. Let us finish this section with a discussion of infinite Coxeter groups. Recall, that − does not lie in the radius of convergence on W(t) if W is not finite. Nevertheless, W(t) represents a rational function as follows from the following result. Positive definite functions on Coxeter groups. . .  Proposition .. ([ste],[ser, Prop. ]) Let (W, S) be an an infi- nite Coxeter system. Then (.)  W(t) =XT∈F (−)#T WT(/t) . Where F denote the family of subsets T ⊂ S, such that the group WT gen- erated by T is finite. In particular, W(t) is a series of a rational function (i.e. a quotient of polynomials). One may ask a question what is the class of (infinite) Coxeter groups such that WT(−) =  for any nonempty subset T of generators. A naıve argument that W(t) = W{s}(t)W{s}(t) = ( + t)W{s}(t) shows, that the question if W(−) ,  is equivalent to whether W{s}(t) can have a pole at t = −. On the other hand note, that if W is of type a, ie. W is given by a presentation hsi :  ≤ i ≤ s i , (si sj) :  ≤ i < j ≤ i then, by Equation (.), W(t) = +t+t More generally, it is known ([bou]) that in each coset of WT there exists the unique shortest element. Let WT denote the set of those shortest representatives. Moreover if w = wTwT with wT ∈ WT and wT ∈ WT then w = wT + wT. Therefore WT(t)WT(t) = W(t). In par- ticular, WT(t) represents a rational function, and it is legitimate to ask about the value of WT(−). In the case of finite Coxeter group W, Eng [eng] observed that (−t) and W(−) = /. WT(−) = #nw ∈ WTwww ∈ WTo , where w is the longest element in W. (Eng's proof was case-by-case. Later, a general classification-free proof of Eng's theorem was given in [rsw]). Subsequently, Reiner [rei] has shown that if W is crystallographic (ie. the Weyl group in a compact Lie group G), then both sides of the above equality compute the signature of the corresponding flag variety G/ QT, where QT is a parabolic subgroup associated to T. What is the meaning of W(−) or WT(−) for infinite W? We do not know if it possible for W(−) to be negative. If one takes W = hsi :  ≤ i ≤ s i , (si sj) :  ≤ i < j ≤ i. Then, by Equation (.), W(t) = (+t)(+t+t) t−t−t+ and Wsi (−) = − /.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski References [bbls] Serban T. Belinschi, Marek Bo zejko, Franz Lehner, and Roland Speicher. The normal distribution is ⊞-infinitely divisible. Adv. Math., (): -- , . [beh] Marek Bo zejko, Wiktor Ejsmont, and Takahiro Hasebe. Fock J. Funct. Anal., type B. space associated to Coxeter groups of (): -- , . [bjs] M. Bo zejko, T. Januszkiewicz, and R. J. Spatzier. Infinite Coxeter groups do not have Kazhdan's property. J. Operator Theory, (): -- , . Nathanial P. Brown and Narutaka Ozawa. C∗-algebras and finite- dimensional approximations, volume  of Graduate Studies in Mathemat- ics. American Mathematical Society, Providence, RI, . [bo] [bou] N. Bourbaki. ´El´ements de math´ematique. Fasc. XXXIV. Groupes et alg`ebres de Lie. Chapitre IV: Groupes de Coxeter et syst`emes de Tits. Chapitre V: Groupes engendr´es par des r´eflexions. Chapitre VI: syst`emes de racines. Actualit´es Scientifiques et Industrielles, No. . Hermann, Paris, . [bo z] Marek Bo zejko. On Λ(p) sets with minimal constant in discrete non- commutative groups. Proc. Amer. Math. Soc., : -- , . [bs] Marek Bo zejko and Roland Speicher. Interpolations between bosonic and fermionic relations given by generalized Brownian mo- tions. Math. Z., (): -- , . [bs] Marek Bo zejko and Ryszard Szwarc. Algebraic length and Poincar´e series on reflection groups with applications to representa- tions theory. In Asymptotic combinatorics with applications to mathemati- cal physics (St. Petersburg, ), volume  of Lecture Notes in Math., pages  -- . Springer, Berlin, . [buc] Artur Buchholz. Norm of convolution by operator-valued functions on free groups. Proc. Amer. Math. Soc., (): -- , . [buc] Artur Buchholz. Optimal constants in Khintchine type inequalities for fermions, Rademachers and q-Gaussian operators. Bull. Pol. Acad. Sci. Math., (): -- , . [bw] Marek Bo zejko and Janusz Wysocza ´nski. Remarks on t- transformations of measures and convolutions. Ann. Inst. H. Poincar´e Probab. Statist., (): -- , . [by] Marek Bo zejko and Hiroaki Yoshida. Generalized q-deformed Gaussian random variables. In Quantum probability, volume  of Ba- nach Center Publ., pages  -- . Polish Acad. Sci. Inst. Math., Warsaw, . Jean De Canni `ere and Uffe Haagerup. Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups. Amer. J. Math., (): -- , . [dch] [eng] Oliver D. Eng. Quotients of Poincar´e polynomials evaluated at −. J. [fel] William Feller. An introduction to probability theory and its applica- Algebraic Combin., (): -- , . tions. Vol. II. Second edition. John Wiley & Sons Inc., New York, . [fena] Gero Fendler. A note on L-sets. Colloq. Math., (): -- , . [fenb] Gero Fendler. Weak amenability of Coxeter groups. arXiv:math/ , . [haa] Uffe Haagerup. The best constants in the Khintchine inequality. Stu- dia Math., (): --  (), . Positive definite functions on Coxeter groups. . .  [haa] Uffe Haagerup. An example of a nonnuclear C∗-algebra, which has the metric approximation property. Invent. Math., (): -- , /. [hau] Felix Hausdorff. Summationsmethoden und Momentfolgen. II. [hp] Math. Z., (-): -- , . Uffe Haagerup and Gilles Pisier. Bounded linear operators be- tween C∗-algebras. Duke Math. J., (): -- , . [hum] James E. Humphreys. Reflection groups and Coxeter groups, volume  of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, . [jan] Tadeusz Januszkiewicz. For Coxeter groups zg is a coefficient of a [ks] uniformly bounded representation. Fund. Math., (): -- , . Claus K ostler and Roland Speicher. A noncommutative de Finetti theorem: invariance under quantum permutations is equivalent to free- ness with amalgamation. Comm. Math. Phys., (): -- , . [lp] [leh] Franz Lehner. Cumulants in noncommutative probability theory. I. Noncommutative exchangeability systems. Math. Z., (): -- , . Franc¸ oise Lust-Piquard. In´egalit´es de Khintchine dans Cp ( < p < ∞). C. R. Acad. Sci. Paris S´er. I Math., (): -- , . Gilles Pisier. Introduction to operator space theory, volume  of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, . [pis] [rei] Victor Reiner. Note on a theorem of Eng. Ann. Comb., (): -- , . [rsw] V. Reiner, D. Stanton, and D. White. The cyclic sieving phenome- [ser] [sim] non. J. Combin. Theory Ser. A, (): -- , . Jean-Pierre Serre. Cohomologie des groupes discrets. In Prospects in mathematics (Proc. Sympos., Princeton Univ., Princeton, N.J., ), pages  -- . Ann. of Math. Studies, No. . Princeton Univ. Press, Princeton, N.J., . Barry Simon. Representations of finite and compact groups, volume  of Graduate Studies in Mathematics. American Mathematical Society, Prov- idence, RI, . [slo] N. J. A. Sloane. The On-Line Encyclopedia of Integer Sequences. A, Jul . [ste] Robert Steinberg. Endomorphisms of linear algebraic groups. Memoirs of the American Mathematical Society, No. . American Mathematical Society, Providence, R.I., . [val] Alain Valette. Weak amenability of right-angled Coxeter groups. Proc. Amer. Math. Soc., (): -- , . [yos] K osaku Yosida. Functional analysis, volume  of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin-New York, sixth edition, . Acknowledgemens Marek Bo zejko was supported by ncn maestro grant dec-// a/st/. Marek Bo zejko and Wojciech Młotkowski were sup- ported by ncn grant //b/st/.  Marek Bo zejko, ´Swiatosław R. Gal, and Wojciech Młotkowski The authors would like to thank Ryszard Szwarc and Janusz Wysocza- ski for many discussions and help with the preparation of this paper. The first author would like to thank professor F. Gotze for his kind invitation to sfb and his hospitality in Bielefeld in . M. Bo zejko -- Polska Akademia Nauk, ul. Kopernika , - Wrocław E-mail address: [email protected] ´S. R. Gal -- Uniwersytet Wrocławski, pl. Grunwaldzki /, - Wrocław E-mail address: [email protected] W. Młotkowski -- Uniwersytet Wrocławski, pl. Grunwaldzki /, - Wrocław E-mail address: [email protected]
1704.04403	1	1704	2017-04-14T12:00:36	Morita embeddings for dual operator algebras and dual operator spaces	[ "math.OA" ]	We define a relation < for dual operator algebras. We say that B < A if there exists a projection p in A such that B and pAp are Morita equivalent in our sense. We show that < is transitive, and we investigate the following question: If A < B and B < A, then is it true that A and B are stably isomorphic? We propose an analogous relation < for dual operator spaces, and we present some properties of < in this case.	math.OA	math	MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES G. K. ELEFTHERAKIS Abstract. We define a relation ⊂∆ for dual operator algebras. We say that B ⊂∆ A if there exists a projection p ∈ A such that B and pAp are Morita equivalent in our sense. We show that ⊂∆ is transitive, and we investigate the following question: If A ⊂∆ B and B ⊂∆ A, then is it true that A and B are stably isomorphic? We propose an analogous relation ⊂∆ for dual operator spaces, and we present some properties of ⊂∆ in this case. 1. Introduction An operator space X is said to be a dual operator space if X is completely isometrically isomorphic to the operator space dual Y ∗ of an operator space Y. If, in addition, X is an operator algebra, then we call it a dual operator algebra. For example, Von Neumann algebras and nest algebras are dual operator algebras. Blecher, Muhly and Paulsen introduced the notion of the Morita equivalence of non-self-adjoint operator algebras [4]. Subsequently, Blecher and Kashyap developed a parallel theory for dual operator algebras [1], [14]. At the same time, the author of the present article proposed a different notion of Morita equivalence for dual operator algebras, called ∆- equivalence. Two unital dual operator algebras A and B are ∆-equivalent if there exist faithful normal representations α : A → α(A), β : B → β(B) and a ternary ring of operators M (i.e., a space satisfying M M ∗M ⊆ M) such that α(A) = [M ∗β(B)M]−w∗ and β(B) = [M α(A)M ∗]−w∗ [9]. In this case, we write A ∼∆ B. An important property is that two algebras are ∆-equivalent if and only if they are stably isomorphic, as was proved by Paulsen and the present author in [12]. Subsequently, Paulsen, Todorov and the present author defined a Morita-type equivalence ∼∆ for dual operator spaces [13]. This equivalence also has the property of being equivalent with the notion of a stable isomorphism. Key words and phrases. Dual operator algebras, Dual operator spaces, TRO, Stable isomorphism, Morita equivalence. 2010 Mathematics Subject Classification. 47L05 (primary), 47L25, 47L35, 46L10, 16D90 (secondary). 1 2 G. K. ELEFTHERAKIS In this paper, we define a weaker relation between dual operator algebras. We say that the dual operator algebra B ∆-embeds into the dual operator algebra A if there exists a projection p ∈ A such that B ∼∆ pAp. In this case, we write B ⊂∆ A. We investigate the relation ⊂∆ between unital dual operator algebras, and we prove that it is a transitive relation. In the case of von Neumann algebras, it is a partial order relation. This means that it has the additional property that if A, B are von Neumann algebras and A ⊂∆ B, B ⊂∆ A, then A ∼∆ B. We present a counterexample to demonstrate In that this does not always hold in the case of non-self-adjoint algebras. Section 2, we also present a characterisation of the relation ⊂∆ in the terms of reflexive lattices. In Section 3, we present an analogous theory defining the relation ⊂∆ for dual operator spaces. In this case, if X, Y are dual operator spaces such that Y ⊂∆ X, then there exist projections p and q such that pX ⊆ X, Xq ⊆ X and Y ∼∆ pXq. We also define a weaker relation ⊂cb∆ . We say that Y ⊂cb∆ X if there exist w∗-continuous completely bounded isomorphisms φ : X → φ(X), ψ : Y → ψ(Y ) such that φ(X) ⊂∆ ψ(Y ). We present a theorem describing ⊂cb∆ in the terms of stable isomorphisms ( Theorem 3.11), and we investigate the problem of whether ⊂∆ is a transitive relation for dual operator spaces ( Theorem 3.14). In the following, we briefly describe the notions used in this paper. We refer the reader to the books [3], [6], [7], [15] and [16] for further details. If V is a linear space and S ⊆ V, then by [S] we denote the linear span of S. If H, K are Hilbert spaces, then we write B(H, K) for the space of bounded operators from H to K. We denote B(H, H) as B(H). If L is a subset of B(H), then we write L′ for the commutant of L, and L′′ for (L′)′. If A is an operator algebra, then by ∆(A) we denote its diagonal A∩ A∗. A ternary ring of operators M, referred to as a TRO from this point, is a subspace of some B(H, K) satisfying the following: m1, m2, m3 ∈ M ⇒ m1m∗ 2m3 ∈ M. It is well known that in the case that M is norm closed, it is equal to [M M ∗M]−k·k. If X is a dual operator space and I is a cardinal, then we write MI(X) for the set of I × I matrices whose finite submatrices have uni- formly bounded norm. We underline that MI(X) is also a dual operator space, and it is completely isometrically and w∗-homeomorphically isomor- phic with X ¯⊗B(l2(I)). Here, ¯⊗ denotes the normal spatial tensor product. We say that two dual operator spaces X and Y are stably isomorphic if there exists a cardinal I and a w∗-continuous completely isometric map from MI (X) onto MI (Y ). If L ⊆ B(H) is a lattice, then we write Alg(L) for the algebra MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES3 of operators x ∈ B(H) satisfying If A ⊆ B(H) is an algebra, then we write Lat(A) for the lattice of projections l ∈ B(H) satisfying A lattice L is called reflexive if l⊥xl = 0, ∀ l ∈ L. l⊥xl = 0, ∀ x ∈ A. L = Lat(Alg(L)). A reflexive algebra is an algebra of the form Alg(L), for some lattice L. An important example of a class of reflexive lattices is given by nests. A nest N ⊆ B(H) is a totally ordered set of projections containing the zero and identity operators, which is closed under arbitrary suprema and infima. The corresponding algebra Alg(N ) is called a nest algebra. If A ⊆ B(H) is a w∗- closed algebra and I is cardinal, then we write AI for the algebra of operators x ∈ B(H ⊗ l2(I)) satisfying x((ξi)i∈I) = (a(ξi))i∈I , ∀ (ξi)i∈I ∈ H ⊗ l2(I) for some a ∈ A. 2. Morita embeddings for dual operator algebras We consider the following known theorem concerning von Neumann alge- bras. Theorem 2.1. Let A, B be von Neumann algebras. Then, the following are equivalent: (i) There exist w∗-continuous, one-to-one, ∗-homomorphisms α : A → B(H), β : B → B(K), where H, K are Hilbert spaces such that the commutants α(A)′, β(B)′ are ∗- isomorphic. (ii) The algebras A, B are weakly Morita equivalent in the sense of Rieffel. (iii) There exists a cardinal I and a ∗-isomorphism from MI (A) onto MI(B). Definition 2.1. [8] Let A, B be w∗-closed algebras acting on the Hilbert spaces H and K, respectively. We call these weakly TRO-equivalent if there exists a TRO M ⊆ B(H, K) such that A = [M ∗BM]−w∗ , B = [M AM ∗]−w∗ . In this case, we write A ∼T RO B. The following defines our notion of weak Morita equivalence for dual oper- ator algebras. 4 G. K. ELEFTHERAKIS Definition 2.2. [9] Let A, B be dual operator algebras. We call these weakly ∆-equivalent if there exist w∗-continuous completely isometric homomorphisms α and β, respectively, such that α(A) ∼T RO β(B). In this case, we write A ∼∆ B. The following theorem is a generalisation of Theorem 2.1 to the setting of unital dual operator algebras: Theorem 2.2. Let A, B be unital dual operator algebras. Then, the following statements are equivalent: (i) There exist reflexive lattices L1 and L2, w∗-continuous completely iso- β : B → Alg(L2), and a metric onto homomorphisms α : A → Alg(L1), ∗-isomorphism θ : ∆(A)′ = L′′ (ii) The algebras A, B are weakly ∆-equivalent. (iii) There exists a cardinal I and a w∗-continuous completely isometric 1 → ∆(B)′ = L′′ 2 such that θ(L1) = L2. homomorphism from MI(A) onto MI(B). The previous theorem has been proved in various papers. In fact, if (i) holds, then by Theorem 3.3 in [8] Alg(L1) ∼T RO Alg(L2), and thus A ∼∆ B. Conversely, if (ii) holds, then by Theorems 2.7 and 2.8 in [10], by choosing a w∗-continuous completely isometric homomorphism α : A → α(A) with reflexive range, there exists a w∗-continuous completely isometric homomor- phism β : B → β(B), also with a reflexive range, such that α(A) ∼T RO β(B). Thus, by Theorem 3.3 in [8], (i) holds. The equivalence of (ii) and (iii) con- stitutes the main result of [12]. Remark 2.3. In the remainder of this section, if A is a unital dual operator algebra and p ∈ A is a projection, then pAp is also a dual operator algebra with unit p. If A is a w∗-closed unital algebra acting on the Hilbert space H and p ∈ A is a projection, then we identify pAp with the algebra pAp(H) ⊆ B(p(H)). 2.1. TRO-embeddings for dual operator algebras. Definition 2.3. Let A, B be unital w∗-closed algebras acting on the Hilbert spaces H and K, respectively. We say that B weakly TRO-embeds into A if there exists a projection p ∈ A such that B ∼T RO pAp. In this case, we write B ⊂T RO A. Remark 2.4. The above definition is equivalent to the following statements. Let A, B be unital w∗-closed algebras acting on the Hilbert spaces H and K, respectively. Then, (i) The algebra B weakly TRO-embeds into A if and only if there exists a TRO M ⊆ B(H, K) such that B = [M ∗AM]−w∗ , M BM ∗ ⊆ A. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES5 (ii) The algebra B weakly TRO-embeds into A if and only if there exists a TRO M ⊆ B(H, K) such that B = [M ∗AM]−w∗, M M ∗ ⊆ A. Remark 2.5. If A is a unital w∗-closed algebra and p ∈ A is a projection, then pAp ⊂T RO A. For the proof, we can take the linear span of the element p as a TRO. Proposition 2.6. Suppose that A, B, C are unital w∗-closed algebras acting on the Hilbert spaces H, K, and L, respectively. If C ⊂T RO B and B ⊂T RO A, then C ⊂T RO A. Proof. We may assume that there exist projections p ∈ ∆(B), q ∈ ∆(A) such that C ∼T RO pBp, B ∼T RO qAq. By Proposition 2.8 in [8], there exists a TRO M such that qAq = [M ∗BM]−w∗ , , q∆(A)q = [M ∗M]−w∗ B = [M qAqM ∗]−w∗ ∆(B) = [M M ∗]−w∗ Define N = pM. Then, we have that , . (pM)(pM)∗(pM) = pM M ∗pM ⊆ pM M ∗∆(B)M ⊆ p∆(B)M ⊆ pM. Thus, N is a TRO. Then, we have that pBp = [pM qAqM ∗p]−w∗ = [N qAqN ∗]−w∗ = [N(N ∗N qAqN ∗N)N ∗]−w∗ , and therefore Thus, [N ∗pBpN]−w∗ = [N ∗N qAqN ∗N]−w∗ . Therefore, pBp ∼T RO [N ∗N qAqN ∗N]−w∗ C ∼T RO [N ∗N qAqN ∗N]−w∗ . We may assume that there exists a TRO L such that . , = [L∗CL]−w∗ . (2.1) [N ∗N qAqN ∗N]−w∗ C = [LN ∗N qAqN ∗N L∗]−w∗ We make the following observations: (2.2) [N q∆(A)qN ∗]−w∗ = [pM M ∗p]−w∗ Furthermore, N ∗N = M ∗pM ⊆ [M ∗M]−w∗ = q∆(A)q. Thus, ) ⊆ q∆(A)q. (2.3) Define D = [LN ∗N q]−w∗. We shall prove that D is a TRO. We have that L∗L ⊆ ∆([N ∗N qAqN ∗N]−w∗ = [pM M ∗M M ∗p]−w∗ = [N N ∗]−w∗ . (LN ∗N q)(qN ∗N L∗)(LN ∗N q) = LN ∗N qN ∗(N L∗LN ∗)N q. By (2.3), it holds that N L∗LN ∗ ⊆ N q∆(A)qN ∗ ⊆ N N ∗. 6 Thus, G. K. ELEFTHERAKIS By (2.2), we have that N qN ∗ ⊆ N N ∗. Thus, DD∗D ⊆ [LN ∗N qN ∗N N ∗N q]−w∗ . DD∗D ⊆ [LN ∗N N ∗N N ∗N q]−w∗ = D. Thus, D is a TRO. By (2.1), we have that Furthermore, By (2.3), we have that C = [DAD∗]−w∗ . D∗D = qN ∗N L∗LN ∗N q. In addition, by (2.2) we have that D∗D ⊆ qN ∗N q∆(A)qN ∗N q. D∗D ⊆ qN ∗N N ∗N q ⊆ qN ∗N q ⊆ q∆(A)q ⊆ A. Thus, by Remark 2.4 (ii), we have that C ⊂T RO A. Remark 2.7. In light of the above proposition, one could expect that the rela- tion ⊂T RO is a partial order relation in the class of unital w∗-closed operator algebras, if we identify those algebras that are TRO-equivalent. This means that ⊂T RO has the additional property that (cid:3) A ⊂T RO B, B ⊂T RO A ⇒ A ∼T RO B. This is true in the case of von Neumann algebras, as we will show in Section 1.3. However, it fails in the case of non-self-adjoint algebras, as we prove in Section 1.4. The following Lemma will be useful. Lemma 2.8. Let B be a w∗-closed unital operator algebra acting on the Hilbert space H, and let q ∈ B be a projection. If p is the projection onto ∆(B)(q(H)), then p is a central projection for the algebra ∆(B), and qBq ∼T RO pBp. Proof. Clearly, p is a central projection for ∆(B). We consider the TRO M = ∆(B)q ⊆ B(q(H), p(H)). M(q(H)) = p(H), M ∗(p(H)) = q(H), We have that and Then, Proposition 2.1 in [8] implies that M ∗pBpM ⊆ qBq, M qBqM ∗ ⊆ pBp. [M ∗pBpM]−w∗ = qBq, [M qBqM ∗]−w∗ = pBp. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES7 (cid:3) 2.2. ∆-embeddings for dual operator algebras. Definition 2.4. Let A, B be dual operator algebras. We say that B weakly ∆-embeds into A if there exist w∗-continuous completely isometric homomor- phisms α : A → α(A), β : B → β(B) such that β(B) ⊂T RO α(A). In this case, we write B ⊂∆ A. The following theorem is a generalisation of Theorem 2.2. Theorem 2.9. Let A, B be unital dual operator algebras. Then, the following are equivalent: (i) There exist reflexive lattices L1,L2 acting on the Hilbert spaces H and K, respectively, w∗-continuous completely isometric onto homomorphisms α : A → Alg(L1), β : B → Alg(L2), and an onto w∗-continuous ∗-homomorphism θ : ∆(A)′ = L′′ 1 → ∆(B)′ = L′′ 2 such that θ(L1) = L2. (ii) B ⊂∆ A. (iii) There exists a cardinal I, a projection q ∈ A, and a w∗-continuous completely isometric homomorphism from MI(B) onto MI(qAq). Proof. The equivalence of (ii) and (iii) is a consequence of Definition 2.4 and Theorem 2.2. It suffices to prove that (i) ⇒ (ii) Define the lattice and the spaces Alg(L2) ⊂T RO AlgL1. L = {(cid:18) l 0 0 θ(l) (cid:19) : l ∈ L1} U = {x : l⊥xθ(l) = 0 ∀ l ∈ L1}, V = {y : θ(l)⊥yl = 0 ∀ l ∈ L1}. We can easily prove that Because the map V Alg(L) =(cid:18) Alg(L1) U 1 → L′′, a →(cid:18) a 0 Alg(L2) (cid:19) . 0 θ(a) (cid:19) ρ : L′′ 8 G. K. ELEFTHERAKIS is a ∗-isomorphism such that θ(L1) = L, from Theorem 3.3 in [8] we have that Define the TRO Now, observe that and Thus, Alg(L1) ∼T RO Alg(L). M = (0 C). MAlg(L)M ∗ = Alg(L2) M ∗M ⊆ Alg(L). Then, Proposition 2.6 implies that Alg(L2) ⊂T RO AlgL. Alg(L2) ⊂T RO AlgL1. (iii) ⇒ (i) Suppose that MI (B) and MI(qAq) are completely isometrically and w∗- homeomorphically isomorphic. Then, B ∼∆ qAq. Every unital dual operator algebra has a w∗-completely isometric representation whose image is reflexive. Thus, we may assume that α : A → B(H) is a w∗-continuous completely isometric homomorphism such that α(A) = Alg(L1) for a reflexive lattice L1. If p = α(q), then α(qAq) = Alg(L1p(H)). Theorem 2.7 in [10] implies that there exists a reflexive lattice L2 and a w∗-continuous completely isometric homomorphism β : B → Alg(L2) such that Alg(L2) ∼T RO Alg(L1p(H)). Thus, by Theorem 3.3 in [8] there exists a ∗-isomorphism ρ : (L1p(H))′′ → L′′ such that ρ(L1p(H)) = L2. If τ : L′′ 1p(H) : x → xp(H), then we write θ = ρ ◦ τ. This is the required map. Theorem 2.10. Let A, B, C be unital dual operator algebras such that 1 → L′′ 2 (cid:3) C ⊂∆ B, B ⊂∆ A. Then, C ⊂∆ A. Proof. We may assume that there exist projections p ∈ B, q ∈ A such that C ∼T RO pBp, β(B) ∼T RO qAq, where β : B → β(B) is a w∗-continuous completely isometric homomor- phism. From Theorem 2.7 in [10], we know that for the representation MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES9 βpBp : pBp → β(p)β(B)β(p) there exists a w∗-continuous completely iso- metric homomorphism γ : C → γ(C) such that By Proposition 2.6, we have that γ(C) ∼T RO β(p)β(B)β(p). β(p)β(B)β(p) ⊂T RO qAq. (cid:3) Because qAq ⊂T RO A, we have that γ(C) ⊂T RO A, and thus C ⊂∆ A. Remark 2.11. In view of Theorem 2.10, one should expect that weak ∆- embedding is a partial order relation in the class of unital dual operator al- gebras if we identify those unital dual operator algebras that are weakly ∆- equivalent. Thus, one should expect that A ⊂∆ B, B ⊂∆ A ⇒ A ∼∆ B. In Section 1.4 we shall see that this is not true. However, in the case of von Neumann algebras, this is indeed true. For further details, see Section 2.3 below. Example 2.12. Let A = Alg(N1), B = Alg(N2), where N1 is a continuous nest, and N2 is a nest with at least one atom. We shall prove that it is impossible that B ⊂∆ A. Proof. Suppose on the contrary that B ⊂∆ A. Thus, there exists a projection p ∈ ∆(A) such that B ∼∆ pAp. Because B and pAp are nest algebras, it follows from Theorem 3.2 in [10] that B ∼T RO pAp. Thus, by Theorem 3.3 in [8] there exists a homeomorphism θ : N2 → N1p. This is impossible, because N2 contains an atom, and N1p is a continuous nest. (cid:3) 2.3. The case of von Neumann algebras. Lemma 2.13. Let A be a von Neumann algebra, and let p, q be central pro- jections of A such that p ≤ q and A ∼T RO Ap. Then, A ∼T RO Aq. Proof. By Theorem 3.3 in [8], there exists a ∗-isomorphism θ : A′ → A′p. We need to prove that there exists a ∗-isomorphism ρ : A′ → A′q. Suppose that e0 = IdA, e1 = q, e2 = p, en = θ(en−2), n = 2, 3, ... Clearly (en)n is a decreasing sequence of central projections. Observe that Thus, the map ρ : A′ → A′q sending ⊕(e2n − e2n+1) ⊕ (e2n+1 − e2n+2)! ⊕ ∧nen. e0 = ∞ Xn=0 ⊕a(e2n − e2n+1) ⊕ a(e2n+1 − e2n+2)! ⊕ (a ∧n en) a = ∞ Xn=0 10 to G. K. ELEFTHERAKIS ρ(a) = ∞ Xn=0 is a ∗-isomorphism . ⊕θ(a)(e2n+2 − e2n+3) ⊕ a(e2n+1 − e2n+2)! ⊕ (a ∧n en) (cid:3) The above Lemma is based on the fact if p, q are central projections of the von Neumann algebra A such that p ≤ q and A ∼= Ap, then A ∼= Aq, where ∼= is the ∗-isomorphism. We acknowledge that this was known to the authors of [5] (see the proof of Lemma 6.2.3). In this Lemma, an alternative proof to ours was provided. Theorem 2.14. The relation ⊂T RO is a partial order relation for von Neu- mann algebras, if we identify those von Neumann algebras that are TRO- equivalent. Proof. Let A, B be von Neumann algebras. It suffices to prove the implication that A ⊂T RO B, B ⊂T RO A ⇒ A ∼T RO B. Let q0 ∈ B, p0 ∈ A be projections such that A ∼T RO q0Bq0, B ∼T RO p0Ap0. By Lemma 2.8, there exist central projections q ∈ B, p ∈ A such that A ∼T RO Bq, B ∼T RO Ap. Thus, there exist ∗-isomorphisms θ : A′ → B′q, ρ : B′ → A′p. We can easily see that there exists a central projection p ∈ A such that p ≤ p and ρ(B′q) = A′ p. Therefore, we obtain a ∗-isomorphism from A′ onto A′ p. Because p ≤ p and p, p are central, Lemma 2.13 implies that there exists a ∗-isomorphism from A′ onto A′p. Thus, A ∼T RO Ap, which implies that A ∼T RO B. Theorem 2.15. Let A, B be von Neumann algebras. Then, the following are equivalent: (cid:3) continuous onto ∗-homomorphism θ : α(A)′ → β(B)′. (i) There exist ∗-isomorphisms α : A → α(A), β : B → β(B) and a w∗- (ii) B ⊂∆ A. (iii) There exists a cardinal I and a w∗-continuous onto ∗-homomorphism ρ : MI (A) → MI(B). MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES11 Proof. The equivalence of (i) and (ii) is a consequence of Theorem 2.9. (ii) ⇒ (iii) Suppose that B ⊂∆ A. By Theorem 2.9, we may assume that there exists a w∗- continuous onto ∗-homomorphism θ : A′ → B′. Suppose that A′p⊥ = Kerθ for a projection p in the centre of A. We also assume that A ⊆ B(H). Then, the map is a ∗-isomorphism. Because A′p(H) → B′ : ap(H) → θ(a) Theorem 2.1 implies that there exists a cardinal I and a w∗-continuous onto ∗-isomorphism (Ap(H))′ = A′p(H), MI (pAp) → MI (B). Suppose that ρ : MI (A) → MI(B) is a w∗-continuous onto ∗-homomorphism. Let p be a projection in the centre of MI(A) such that (iii) ⇒ (ii) MI(A)p⊥ = Kerρ. Because Z(MI(A)) = Z(A)I , where Z(A) (resp. Z(MI(A))) is the centre of A (resp. MI(A)), we may assume that p = qI for q ∈ Z(A). Thus, the map MI(Aq) → MI(B) : (ai,jq) → ρ((ai,j)) (cid:3) is a ∗-isomorphism. Then, Theorem 2.9 implies that B ⊂∆ A. Theorem 2.16. The weak ∆-embedding is a partial order relation in the class of von Neumann algebras, if we identify those von Neumann algebras that are weakly Morita equivalent in the sense of Rieffel. Proof. Claim: Let A be a von Neumann algebra, and let r ∈ A be a projec- tion such that A ∼∆ rAr. If q is a projection in A such that r ≤ q, then it also holds that A ∼∆ qAq. Proof of the claim: There exists a cardinal I such that the algebras MI(A) and MI(rAr) are ∗- isomorphic. We suppose that M = MI (A), N = MI(qAq). Then, we have that M ∼= MI (rAr) = rI N rI and N ∼= qI M qI . By Lemma 6.2.3 in [5], the von Neumann algebras M and N are stably isomorphic. Thus, M ∼∆ N. However, A ∼T RO M, qAq ∼T RO N. Therefore, A ∼∆ qAq, and the proof of the claim is complete. To prove the theorem, it suffices to prove that if A, B are von Neumann algebras such that A ⊂∆ B, B ⊂∆ A, then A ∼∆ B. We may assume that 12 G. K. ELEFTHERAKIS there exist projections p ∈ B, q ∈ A and w∗-continuous completely isometric homomorphisms α : A → α(A), β : B → β(B), such that α(A) ∼T RO pBp, β(B) ∼T RO qAq. For the representation βpBp : pBp → β(p)β(B)β(p), there exists a w∗- continuous one-to-one ∗-homomorphism γ : α(A) → γ(α(A)) such that γ(α(A)) ∼T RO β(p)β(B)β(p). By Proposition 2.6, we have that β(p)β(B)β(p) ⊂T RO qAq. Therefore, there exists a projection r ≤ q such that γ(α(A)) ∼T RO rAr ⇒ A ∼∆ rAr. The claim implies that A ∼∆ qAq. However, qAq ∼∆ B. Thus A ∼∆ B. Thus, the proof is complete. (cid:3) Corollary 2.17. Let A, B be von Neumann algebras, I, J be cardinals, and θ : MI (A) → MI(B), ρ : MJ (B) → MJ (A) be onto w∗-continuous homomorphisms. Then, A ∼∆ B. Proof. By Theorem 2.15, A ⊂∆ B and B ⊂∆ A. The conclusion then follows from the above theorem. Corollary 2.18. Let A, B be unital dual operator algebras such that A ⊂∆ B and B ⊂∆ A. Then, ∆(A) ∼∆ ∆(B). Proof. We can easily see that ∆(A) ⊂∆ ∆(B) and ∆(B) ⊂∆ ∆(A). Now, we can apply the above theorem. (cid:3) (cid:3) Example 2.19. Let A be a factor, and B be a unital dual operator algebra such that B ⊂∆ A. Then, B is a von Neumann algebra, and B ∼∆ A. Proof. There exist a ∗-isomorphism α : A → α(A), a w∗- continuous com- pletely isometric homomorphism β : B → β(B), and a TRO M such that if p is the projection onto [M M ∗]−w∗, then β(B) = [M ∗α(A)M]−w∗ , pα(A)p = [M β(B)M ∗]−w∗ . Define N = [α(A)pM]−w∗ . Because it follows that M M ∗pα(A) ⊆ pα(A) ⊆ α(A), pM M ∗p ⊆ α(A) ⇒ [α(A)pM M ∗pα(A)]−w∗ = [N N ∗]−w∗ MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES13 is an ideal of α(A). However, α(A) is a factor, and thus α(A) = [N N ∗]−w∗. On the other hand, [N ∗N]−w∗ = [M ∗pα(A)α(A)pM]−w∗ = [M ∗α(A)M]−w∗ = β(B). Thus, A and B are weakly Morita equivalent in the sense of Rieffel. However, in the case of von Neumann algebras, Rieffel's Morita equivalence is the same as ∆-equivalence. (cid:3) 2.4. A counterexample in non-self-adjoint operator algebras. Despite the situation for von Neumann algebras, we shall prove that if A, B are uni- tal non-self-adjoint dual operator algebras, it does not always hold that the implication (2.4) A ⊂∆ B, B ⊂∆ A ⇒ A ∼∆ B. By Theorem 3.12 in [10], if A, B are nest algebras, then Because for every nest algebra B and every projection p ∈ B the algebra pBp is a nest algebra, we can conclude that A ∼∆ B ⇔ A ∼T RO B. A ⊂∆ B ⇔ A ⊂T RO B. Thus, in order to prove that (2.4) does not hold, it suffices to find nest algebras A and B such that A ⊂T RO B, B ⊂T RO A and A is not TRO-equivalent to B. Let m be the Lebesgue measure on the Borel sets of the interval [0, 1]. Suppose that Q is the set of rationals, and Q+ (resp. Q−) is the projection onto l2(Q∩ [0, t]) (resp. l2(Q∩ [0, t)) ). Furthermore, let Nt be the projection onto L2([0, t], m) for 0 ≤ t ≤ 1. We define the nest N = {Q+ t ⊕ Nt, Q− t ⊕ Nt, 0 ≤ t ≤ 1}. By A = Alg(N ), we denote the corresponding nest algebra acting on the Hilbert space H = l2(Q ∩ [0, 1]) ⊕ L2([0, 1], m). The above nest appeared in Example 7.18 in [6]. Suppose that f (t) = 1 [0, 1], and define 2 , t ∈ M = {Q+ We can define a unitary f (t) ⊕ Nf (t), Q− f (t) ⊕ Nf (t), u2 : L2([0, 1]) → L2([0, 1 2 0 ≤ t ≤ 1}. ]) such that u2(χΩ) = √2χf (Ω) where χΩ is the characteristic function of the Borel set Ω. This unitary maps Nt onto Nf (t) in the sense that u2Ntu∗ 2 = Nf (t), 0 ≤ t ≤ 1. 14 G. K. ELEFTHERAKIS Furthermore, the map t , Q− {Q+ t onto Qj t : 0 ≤ t ≤ 1} −→ {Q+ f (t), Q− f (t) : 0 ≤ t ≤ 1} sending Qj f (t) for j = +,− is a nest isomorphism. Because these nests are multiplicity free (they generate a maximal abelian self-adjoint algebra, referred to as an MASA from this point) and totally atomic, the above map extends as a ∗-isomorphism between the corresponding MASAs. Thus, there exists a unitary such that u1 : l2(Q ∩ [0, 1]) → l2(Q ∩ [0, 1 2 ]) u2Q+ t u∗ 2 = Q+ f (t), u2Q− t u∗ 2 = Q− f (t), 0 ≤ t ≤ 1. Therefore, the unitary u = u1 ⊕ u2, implies a unitary equivalence between N and M. Let s be the projection s : l2(Q ∩ [0, 1]) → l2(Q ∩ [0, 1 2 ]) and r be the projection r : L2([0, 1], m) → L2([0, 1 2 ], m). If p = s⊕ r, then p ∈ A and pAp = Alg(M). By the above arguments, A and pAp are unitarily equivalent, and thus they are TRO-equivalent. Suppose that q0 is the projection 0 IL2([0,1])−r (cid:19) ∈ A q0 =(cid:18) 0 0 and q = p + q0. Because p ≤ q ≤ IdA, we have that pAp ⊂T RO qAq ⊂T RO A. However, A ∼T RO pAp. This implies that A ⊂T RO qAq. Thus, if (2.4) holds, then we should have that A ∼T RO qAq. Suppose that L is the nest Lat(qAq). By Theorem 3.3 in [8], there exists a ∗-isomorphism θ : ∆(A)′ → (∆(A)q(H))′ such that θ(N ) = L. However, the algebras ∆(A), ∆(A)q(H) are MASAs. Therefore, there exists a unitary w : q(H) → H such that θ(x) = w∗xw, ∀x ∈ ∆(A) = ∆(A)′. We have that A = wqAqw∗. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES15 We can easily see that L = L1 ∪ L2, where t ⊕ Nt, t ⊕ Nt, Q− L1 = {Q+ 0 ≤ t ≤ 1 2} and L2 = {Q+ 2 ⊕ Nt, 1 1 2 ≤ t ≤ 1}. Observe that L1 ≤ L0 ≤ L2 for all Li ∈ Li, i = 1, 2 where L0 = Q+ 2 ⊕ N 1 1 2 . If M0 = wL0w∗, N1 = wL1w∗, N2 = wL2w∗, then M1 ≤ M0 ≤ M2 for all Mi ∈ Ni, i = 1, 2. Suppose that M0 = Q+ N2 = {Q+ t ⊕ Nt, Q− t ⊕ Nt, If L2 = {(Q+ 1 2 ⊕ Nt) − L0 : we can consider L2 to be a nest acting on L2([ 1 t0 ⊕ Nt0. Then, t0 ≤ t ≤ 1}. 1 2 ≤ t ≤ 1}, 2 , 1], m). Furthermore, if N2 = {(Q+ t ⊕ Nt) − M0, (Q− t ⊕ Nt) − M0, t0 ≤ t ≤ 1}, then L2 and N2 are isomorphic nests. However, this is impossible, because L2 is a continuous nest and N2 is a nest with atoms. This contradiction shows that A and qAq are not TRO-equivalent. Remark 2.20. Let A and q be as above. As we have seen, qAq ⊂∆ A, A ⊂∆ qAq but A and qAq are not ∆-equivalent. We can prove further that they are not Morita equivalent even in the sense of Blecher and Kashyap [1], [14]. If they were, then by [11] N and L would be isomorphic as nests. However, we can see that this is impossible by applying the same arguments as above. 3. Morita embeddings for dual operator spaces Definition 2.1 can be adapted to the setting of dual operator spaces as follows. Definition 3.1. [13] Let H1, H2, K1, K2 be Hilbert spaces, and let X ⊆ B(H1, H2), Y ⊆ B(K1, K2) be w∗-closed spaces. We call these weakly TRO-equivalent if there exist TROs Mi ⊆ B(Hi, Ki), i = 1, 2 such that X = [M ∗ 2 Y M1]−w∗ , Y = [M2XM ∗ 1 ]−w∗ . In this case, we write X ∼T RO Y. 16 G. K. ELEFTHERAKIS Remark 3.1. If W1, W2 are Hilbert spaces and Z is a subspace of B(W1, W2), then we call it nondegenerate if Z(W1) = W2, Z ∗(W2) = W1. If H1, H2, K1, K2, X, Y are as in the above definition, p2 (resp. q2) is the projection onto X(H1) (resp. Y (K1)), and p1 (resp. q1) is the projection onto X ∗(H2) (resp. Y ∗(K2)), then the spaces p2Xp1(H1), q2Y q1(K1) are nondegenerate, and also weakly TRO-equivalent. This can be concluded from Proposition 2.2 in [13]. The following defines our notion of weak Morita equivalence for dual oper- ator spaces. Definition 3.2. [13] Let X, Y be dual operator spaces. We call these weakly ∆-equivalent if there exist w∗-continuous completely isometric maps φ, ψ, re- spectively, such that φ(X) ∼T RO ψ(Y ). In this case, we write X ∼∆ Y. The following theorem constitutes the main result of [13]. Theorem 3.2. Let X, Y be dual operator spaces. Then, the following are equivalent: (i) X ∼∆ Y. (ii) There exists a cardinal I and a w∗-continuous completely isometric map from MI(X) onto MI(Y ). Remark 3.3. Throughout Section 3, we shall employ the following notation. If H1, H2 are Hilbert spaces, X ⊆ B(H1, H2) is a w∗-closed subspace, and q ∈ B(H1), p ∈ B(H2) are projections such that pX ⊆ X, Xq ⊆ X, then by pXq we denote the space {pxq : x ∈ X} ⊆ B(H1, H2). This space is w∗-closed and completely isometrically and w∗-homeomorphically isomorphic with the space pXq(H1) ⊆ B(q(H1), p(H2)). 3.1. TRO-embeddings for dual operator spaces. Definition 3.3. Let H1, H2, K1, K2 be Hilbert spaces, and let X ⊆ B(H1, H2), Y ⊆ B(K1, K2) be w∗-closed spaces. We say that Y weakly TRO embeds into X if there exist TROs M1 ⊆ B(H1, K1) and M2 ⊆ B(H2, K2) such that Y = [M2XM ∗ 1 M1 ⊆ X. In this case, we write Y ⊂T RO X. Remark 3.4. We can easily see that if Y ⊂T RO X, then there exist projec- tions p, q such that pX ⊆ X, Xq ⊆ X and Y ∼T RO pXq. Examples 3.5. (i) If X ∼T RO Y , then clearly X ⊂T RO Y and Y ⊂T RO X. 2 Y M1 ⊆ X and M ∗ 2 M2X ⊆ X, XM ∗ 1 ]−w∗ , M ∗ (ii) If Ki, Wi, i = 1, 2 are Hilbert spaces, are w∗-closed spaces, and Y ⊆ B(K1, K2), Z ⊆ B(W1, W2) X = Y ⊕ Z ⊆ B(K1 ⊕ W1, K2 ⊕ W2), MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES17 then Y ⊂T RO X. For the proof, we apply the TROs M1 = (CIK1, 0), M2 = (CIK2, 0). (iii) If X ⊆ B(H1, H2) is a w∗-closed operator space and p ∈ B(H2), q ∈ B(H1) are projections such that pX ⊆ X, Xq ⊆ X, then pXq ⊂T RO X. (iv) A generalisation of W ∗-modules over von Neumann algebras is given by the projectively w∗-rigged modules over unital dual operator algebras. See [2] for more details. Given a unital dual operator algebra A, a projectively w∗-rigged module over A is a dual operator space Z that is completely isomet- rically and w∗-homeomorphically isomorphic to a space Y = [M A]−w∗ , where M is a TRO satisfying M ∗M ⊆ A. Observe that Y = [M AC]−w∗, M ∗Y C ⊆ A and M ∗M A ⊆ A, AC ⊆ A. Thus, Y ⊂T RO A. Therefore, for every pro- jectively w∗-rigged module Z over a unital dual operator algebra A, we have that Z ⊂∆ A. Here, ⊂∆ is the relation defined in Definition 3.4 below. Proposition 3.6. Let X, Y, Z be w∗-closed operator spaces. If Z ⊂T RO Y, Y ⊂T RO X, then there exist projections p, q such that pX ⊆ X, Xq ⊆ X and Z ⊂T RO pXq. Proof. There exist projections p, q, r, s such that pX ⊆ X, Xq ⊆ X, rY ⊆ Y, Y s ⊆ Y and TROs Mi, Ni, i = 1, 2 such that 1 ]−w∗ 1 ]−w∗ Y = [M2pXqM ∗ Z = [N2rY sN ∗ , pXq = [M ∗ , rY s = [N ∗ , 2 Y M1]−w∗ 2 ZN1]−w∗ . We may assume that M2p = M2, M1q = M1, N2r = N2, N1s = N1. Suppose that Di is the W ∗-algebra generated by the set Define {MiM ∗ i } ∪ {N ∗ i Ni}, i = 1, 2. Because M1M ∗ 1 ⊆ D1, N ∗ Li = [NiDiMi]−w∗ , i = 1, 2. 1 N1 ⊆ D1, it follows that N1D1M1M ∗ 1 D1N ∗ 1 N1D1M1 ⊆ N1D1M1, and thus Therefore, L1, and similarly L2, are TROs. Now, we have that L1L∗ 1L1 ⊆ L1. [L2pXqL∗ 1]−w∗ = [N2D2M2pXqM ∗ 1 D1N ∗ 1 ]−w∗ = [N2D2Y D1N ∗ 1 ]−w∗ . 18 Because G. K. ELEFTHERAKIS [M2M ∗ 2 Y ]−w∗ = Y = [Y M1M ∗ 1 ]−w∗ , N ∗ 2 N2Y ⊆ Y, Y N ∗ 1 N1 ⊆ Y, we have that D2Y = Y = Y D1. Thus, = [N2Y N ∗ [L2pXqL∗ (3.1) 1]−w∗ 1 ]−w∗ Furthermore, 2ZL1 ⊆ [M ∗ L∗ 2 D2N ∗ [M ∗ 2 ZN1D1M1]−w∗ 2 D2Y D1M1]−w∗ = [M ∗ 2 Y M1]−w∗ . = [N2rY sN ∗ 1 ]−w∗ = Z. = [M ∗ 2 D2rY sD1M1]−w∗ ⊆ Thus, (3.2) On the other hand, L∗ 2ZL1 ⊆ pXq. L∗ 2L2pXq ⊆ [M ∗ [M ∗ 2 D2N ∗ 2 D2M2M ∗ 2 N2D2M2pXq]−w∗ ⊆ [M ∗ 2 D2M2pXq]−w∗ = 2 Y M1]−w∗ ⊆ [M ∗ 2 Y M1]−w∗ . 2L2pXq ⊆ pXq, and similarly pXqL∗ Thus, L∗ tions (3.1) and (3.2) imply that Z ⊂T RO pXq. Remark 3.7. From the above proof, we isolate the fact that if Z ⊂T RO Y and Y ∼T RO X, then Z ⊂T RO X. 3.2. ∆-embeddings for dual operator spaces. 1L1 ⊆ pXq. Therefore, the rela- (cid:3) Definition 3.4. Let X, Y be dual operator spaces. We say that Y weakly ∆-embeds into X if there exist w∗-continuous completely isometric maps φ : X → φ(X), ψ : Y → ψ(Y ) such that ψ(Y ) ⊂T RO φ(X). In this case, we write Y ⊂∆ X. Definition 3.5. Let X, Y be dual operator spaces. A map φ : X → Y that is one-to-one, w∗-continuous, and completely bounded with a completely bounded inverse is called a w∗-c.b. isomorphism, and the spaces X, Y are called w∗-c.b. isomorphic. Under the above assumptions, the map φ−1 is also w∗-continuous. Definition 3.6. Let X, Y be dual operator spaces. We call these c.b. ∆- equivalent if there exist w∗- c.b. isomorphisms φ : X → φ(X), ψ : Y → ψ(Y ) such that φ(X) ∼T RO ψ(Y ). In this case, we write X ∼cb∆ Y. Definition 3.7. Let X, Y be dual operator spaces. We say that Y c.b. ∆- embeds into X if there exist w∗-c.b. isomorphisms φ : X → φ(X), ψ : Y → ψ(Y ) such that ψ(Y ) ⊂T RO φ(X). In this case, we write Y ⊂cb∆ X. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES19 Remark 3.8. Observe the following: (i) X ∼∆ Y ⇒ X ∼cb∆ Y (ii) X ⊂∆ Y ⇒ X ⊂cb∆ Y In what follows, if X is a dual operator space, then Ml(X) (resp. Mr(X)) denotes the algebra of left (resp. right) multipliers of X. In this case, Al(X) = ∆(Ml(X)), (resp. Ar(X) = ∆(Mr(X))) is a von Neumann algebra [3]. Lemma 3.9. Suppose that Z, Y are w∗-closed operator spaces satisfying Z ∼T RO Y, H1, H2 are Hilbert spaces such that Al(Y ) ⊆ B(H2), Ar(Y ) ⊆ B(H1), and ψ : Y → B(H1, H2) is a w∗-continuous complete isometry such that Al(Y )ψ(Y )Ar(Y ) ⊆ ψ(Y ). Then, there exists a w∗-continuous complete isometry ζ : Z → ζ(Z) such that ζ(Z) ∼T RO ψ(Y ). Proof. Assume that M1, M2 are TROs such that Z = [M ∗ 2 Y M1]−w∗ , Y = [M2ZM ∗ 1 ]−w∗ . By Remark 3.1, we may assume that Z and Y are nondegenerate spaces. We denote A = [M ∗ 2 M2]−w∗ , B = [M ∗ 1 M1]−w∗ , C = [M2M ∗ 2 ]−w∗ , D = [M1M ∗ 1 ]−w∗ . The algebras Ω(Z) =(cid:18) A Z 0 B (cid:19) , Ω(Y ) =(cid:18) C Y 0 D (cid:19) are weakly TRO-equivalent as algebras. Indeed, Ω(Z) = [M ∗Ω(Y )M]−w∗ , Ω(Y ) = [MΩ(Z)M ∗]−w∗ , where M is the TRO M2 ⊕ M1. If c ∈ C, then define γ(c) : ψ(Y ) → ψ(Y ), γ(c)ψ(y) = ψ(cy). We can easily see that γ(c) ∈ Al(Y ) and kγ(c)k ≤ 1. Thus, γ : C → Al(Y ) is a contractive homomorphism and hence a ∗-homomorphism. If γ(c) = 0, then cY = 0. Because Y is nondegenerate, we conclude that c = 0. Thus, γ is a one-to-one ∗-homomorphism. Similarly, there exists a one-to-one ∗- homomorphism δ : D → Ar(Y ) such that ψ(y)δ(d) = ψ(yd), ∀ y. The map π : Ω(Y ) → π(Ω(Y )), given by π(cid:18)(cid:18) c y 0 d (cid:19)(cid:19) =(cid:18) γ(c) ψ(y) δ(d) (cid:19) , is a w∗-continuous completely isometric homomorphism. This can be shown by applying 3.6.1 in [3]. By Theorem 2.7 in [10], there exists a w∗-continuous 0 20 G. K. ELEFTHERAKIS completely isometric homomorphism ρ : Ω(Z) → ρ(Ω(Z)) and a TRO N such that ρ(Ω(Z)) = [N ∗π(Ω(Y ))N]w∗ , π(Ω(Y )) = [N ρ(Ω(Z))N ∗]w∗ . As in the discussion concerning the map Φ below Theorem 2.5 in [13], the map ρ is given by ρ(cid:18)(cid:18) a z 0 b (cid:19)(cid:19) =(cid:18) α(a) ζ(z) β(b) (cid:19) , 0 where α : A → α(A), ζ : Z → ζ(Z), β : B → β(B) are completely isometric maps. By Lemma 2.8 in [13], the TRO N is of the form N = N2 ⊕ N1. Thus, ζ(Z) = [N ∗ 2 ψ(Y )N1]−w∗ , ψ(Y ) = [N2ζ(Z)N ∗ 1 ]−w∗ . (cid:3) Lemma 3.10. Let Z, Ω, X be dual operator spaces. We assume that Z ∼T RO Ω, and that ψ0 : Ω → ψ0(Ω) is a w∗-continuous complete isometry such that ψ0(Ω) ⊂T RO X. Then, there exists a w∗-c.b. isomorphism φ : X → φ(X) and a w∗-continuous complete isometry ζ : Z → ζ(Z) such that ζ(Z) ⊂T RO φ(X). Thus, Z ⊂cb∆ X. Proof. Suppose that Al(Ω) ⊆ B(H2), Ar(Ω) ⊆ B(H1) and ψ : Ω → B(H1, H2) is a w∗-continuous complete isometry such that Al(Ω)ψ(Ω)Ar(Ω) ⊆ ψ(Ω). By Lemma 3.9, there exists a w∗-continuous complete isometry ζ : Z → ζ(Z) such that We assume that p, q are projections such that pX ⊆ X, Xq ⊆ X and ζ(Z) ∼T RO ψ(Ω). ψ0(Ω) ∼T RO pXq. Again by Lemma 3.9, there exists a w∗-continuous complete isometry φ : pXq → φ(pXq) such that ψ(Ω) ∼T RO φ(pXq). Define xq⊥   , for all x ∈ X. Observe that φ is a w∗-continuous completely bounded and one-to-one map. If φ∞ is the ∞×∞ amplification of φ, then φ∞ has a closed range. Thus, by the open mapping theorem, φ∞ has a bounded inverse. φ(x) =  p⊥xq 0 φ(pxq) 0 0 0 0 0 MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES21 Therefore, φ−1 is completely bounded. We have that ζ(Z) ∼T RO φ(pXq). Thus, there exist TROs M1, M2 such that ζ(Z) = [M2φ(pXq)M ∗ 1 ]−w∗ , φ(pXq) = [M ∗ 2 ζ(Z)M1]−w∗ . Define the TROs We can see that and Ni =(cid:0) Mi 0 0 (cid:1) , i = 1, 2. [N2 φ(X)N ∗ 1 ]−w∗ = ζ(Z), N ∗ 2 ζ(Z)N1 ⊆ φ(X) 1 N1 ⊆ φ(X). 2 N2 φ(X) ⊆ φ(X), φ(X)N ∗ N ∗ Thus, ζ(Z) ⊂T RO φ(X). Theorem 3.11. Let X, Y be dual operator spaces. Then, the following are equivalent. (cid:3) (i) Y ⊂cb∆ X. (ii) There exist w∗-c.b. φ : X → φ(X); projections p, q such that pφ(X) ⊆ φ(X), φ(X)q ⊆ φ(X); a cardinal I; and a completely isometric w∗-continuous onto map isomorphisms ψ : Y → ψ(Y ), Proof. π : MI (ψ(Y )) → MI(pφ(X)q). By definition, there exist w∗-c.b. isomorphisms ψ : Y → ψ(Y ), φ : X → φ(X) such that (i) ⇒ (ii) ψ(Y ) ⊂T RO φ(X). There exist projections p, q such that pφ(X) ⊆ φ(X), φ(X)q ⊆ φ(X) and ψ(Y ) ∼T RO pφ(X)q. By Theorem 3.2, there exists a cardinal I and a completely isometric w∗-continuous onto map π : MI (ψ(Y )) → MI(pφ(X)q). Define (ii) ⇒ (i) pφ(x)q 0 0 0 p⊥φ(x)q 0 φ(x) =  0 0 φ(x)q⊥   , for all x ∈ X. As in the proof of Lemma 3.10, we can see that φ is a w∗-c.b. isomorphism. By Example 3.5 (ii), we have that φ(X). pφ(X)q ⊂T RO 22 G. K. ELEFTHERAKIS By Theorem 3.2, it holds that ψ(Y ) ∼∆ pφ(X)q. Thus, there exist completely isometric w∗-continuous maps µ : ψ(Y ) → µ(ψ(Y )), χ : pφ(X)q → χ(pφ(X)q) such that µ(ψ(Y )) ∼T RO χ(pφ(X)q). Now, apply Lemma 3.10 for Z = µ(ψ(Y )), Ω = χ(pφ(X)q), ψ0 = χ−1 : Ω → pφ(X)q. We have that We conclude that ψ0(Ω) ⊂T RO φ(X). µ(ψ(Y )) ⊂cb∆ φ(X) ⇒ Y ⊂cb∆ X. (cid:3) Lemma 3.12. Suppose that Z ⊆ B(W1, W2), Y ⊆ B(K1, K2) are w∗-closed spaces such that Z ⊂T RO Y. We also assume that p2 is the projection onto Y (K1), and p1 is the projection onto Y ∗(K2). Thus, Y0 = p2Y p1(K1) is a nondegenerate space into B(p1(K1), p2(K2)). We shall prove that Z ⊂T RO Y0. Proof. By definition, there exist TROs Ni ⊆ B(Wi, Ki), i = 1, 2 such that Z = [N ∗ 2 Y N1]−w∗ , N2ZN ∗ 1 ⊆ Y, N2N ∗ 2 Y ⊆ Y, Y N1N ∗ 1 ⊆ Y. Define Mi = piNi ⊆ B(Wi, pi(Ki)), i = 1, 2. We have that p2y = y ⇒ p2m1m∗ 2y = m1m∗ 2y, ∀ y ∈ Y, m1, m2 ∈ N2. Thus, p2N2N ∗ 2 p2 = N2N ∗ 2 p2 ⇒ p2N2N ∗ 2 p2 = p2N2N ∗ 2 . The above relation implies that 2 M2 = p2N2N ∗ M2M ∗ 2 p2N2 = p2N2N ∗ 2 N2 ⊆ p2N2 = M2. Therefore, M2 is a TRO. Similarly, M1 is also a TRO. Now, we have that [M ∗ 2 Y0M1]−w∗ = [N ∗ 2 p2Y p1N1]−w∗ = [N ∗ 2 Y N1]−w∗ = Z and Furthermore, M2ZM ∗ 1 = p2N2ZN ∗ 1 p1 ⊆ p2Y p1 = Y0. M2M ∗ 2 Y0 = p2N2N ∗ 2 p2Y p1 = p2N2N ∗ 2 Y p1 ⊆ p2Y p1 = Y0. Therefore, Y0 is a nondegenerate subspace of B(p1(K1), p2(K2)), and Z ⊂T RO Y0. (cid:3) The Lemma below is weaker than Lemma 3.9. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES23 Lemma 3.13. Suppose that Z, Y are w∗-closed operator spaces satisfying Z ⊂T RO Y , H1, H2 are Hilbert spaces such that Al(Y ) ⊆ B(H2), Ar(Y ) ⊆ B(H1), and ψ : Y → B(H1, H2) is a w∗-continuous complete isometry such that Al(Y )ψ(Y )Ar(Y ) ⊆ ψ(Y ). Then, there exists a w∗-continuous complete isometry ζ : Z → ζ(Z) such that ζ(Z) ⊂T RO ψ(Y ). Proof. Assume that M1, M2 are TROs such that 1 ⊆ Y, M2M ∗ 2 Y ⊆ Y, Y M1M ∗ 2 Y M1]−w∗ , M2ZM ∗ 1 ⊆ Y. Z = [M ∗ By Lemma 3.12, we may assume that Y is nondegenerate. Suppose that p is the identity of [M2M ∗ . Then, we have that pY ⊆ Y, Y q ⊆ Y. We denote A, B, C, D, Ω(Z) as in Lemma 3.9, and 2 ]−w∗ and q is the identity of [M1M ∗ 1 ]−w∗ If c ∈ C, then define map Ω(pY q) =(cid:18) C pY q 0 D (cid:19) . ψ(Y ) → ψ(Y ) : ψ(y) → ψ(cpy). Clearly, this map belongs to Al(Y ), and thus there exists γ(c) ∈ Al(Y ) satis- fying γ(c)ψ(y) = ψ(cpy) ∀ y ∈ Y. Note that we can define a ∗-homomorphism γ : C → Al(Y ). If γ(c) = 0, then cy = 0 for all y ∈ Y , and thus because Y is nondegenerate, it follows that c = 0. Therefore, γ is one-to-one. Similarly, there exists a one-to-one ∗-homomorphism such that δ : D → Ar(Y ) ψ(y)δ(d) = ψ(yqd) ∀ y ∈ Y. We can conclude that there exist projections p ∈ Al(Y ), q ∈ Ar(Y ) such that ψ(py) = pψ(y), ψ(yq) = ψ(y)q, ∀ y ∈ Y. The map π : Ω(pY q) → π(Ω(pY q)), given by π(cid:18)(cid:18) c pyq 0 d (cid:19)(cid:19) =(cid:18) γ(c) ψ(pyq) δ(d) 0 (cid:19) , is a w∗-continuous completely isometric homomorphism. Because Ω(Z) ∼T RO Ω(pY q), 24 G. K. ELEFTHERAKIS as in Lemma 3.9 we can find a w∗-continuous completely isometric map ζ : Z → ζ(Z) and TROs N1, N2 such that ζ(Z) = [N2ψ(pY q)N ∗ γ(C) = [N ∗ 1 ]−w∗ 2 N2]−w∗ , , ψ(pY q) = [N ∗ 2 ζ(Z)N1]−w∗ ⊆ ψ(Y ), 1 N1]−w∗ . δ(D) = [N ∗ Because N ∗ 2 N2ψ(Y ) ⊆ γ(C)ψ(Y ) = ψ(CY ) ⊆ ψ(Y ), 1 N1 ⊆ ψ(Y ), we have that ζ(Z) ⊂T RO ψ(Y ). and similarly ψ(Y )N ∗ Theorem 3.14. Let X, Y, Z be dual operator spaces such that Z ⊂∆ Y, Y ⊂∆ X. Then: (i) There exist a w∗ continuous complete isometry χ : X → χ(X) and pro- jections p, q such that pχ(X) ⊆ χ(X), χ(X)q ⊆ χ(X) and Z ⊂∆ pχ(X)q. (ii) Z ⊂cb∆ X. (cid:3) Proof. Suppose that H1, H2 are Hilbert spaces such that Al(Y ) ⊆ B(H2), Ar(Y ) ⊆ B(H1) and ψ : Y → B(H1, H2) is a w∗-continuous complete isometry such that By Lemma 3.13, there exists a w∗-continuous complete isometry ζ : Z → ζ(Z) such that Al(Y )ψ(Y )Ar(Y ) ⊆ ψ(Y ). Because Y ⊂∆ X, there exist a w∗-continuous complete isometry χ : X → χ(X) and projections p, q such that pχ(X) ⊆ χ(X), χ(X)q ⊆ χ(X) and ζ(Z) ⊂T RO ψ(Y ). Y ∼∆ pχ(X)q. By Lemma 3.9, there exists a w∗-continuous complete isometry such that φ : pχ(X)q → φ(pχ(X)q) Thus, by Remark 3.7 it holds that ψ(Y ) ∼T RO φ(pχ(X)q). There exist projections r, s and TROs M1, M2 such that ζ(Z) ⊂T RO φ(pχ(X)q) ⇒ Z ⊂∆ pχ(X)q. rφ(pχ(X)q) ⊆ φ(pχ(X)q), φ(pχ(X)q)s ⊆ φ(pχ(X)q) and ζ(Z) = [M2rφ(pχ(X)q)sM ∗ 1 ]−w∗ , rφ(pχ(X)q)s = [M ∗ 2 ζ(Z)M1]−w∗ . For every x ∈ X, define φ(x) = rφ(pχ(x)q)s ⊕ r⊥φ(pχ(x)q)s ⊕ φ(pχ(x)q)s⊥ ⊕ p⊥χ(x)q ⊕ χ(x)q⊥. MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES25 As in the proof of Lemma 3.10, we can see that φ is a w∗-c.b. isomorphism from X onto φ(X). Define the TROs Ni = (Mi 0 0 0 0), i = 1, 2. We can see that [N2 φ(X)N ∗ 1 ]−w∗ = ζ(Z), N ∗ 2 N2 φ(X) ⊆ φ(X), φ(X)N ∗ N ∗ 2 ζ(Z)N1 ⊆ φ(X) 1 N1 ⊆ φ(X). and Thus, ζ(Z) ⊂T RO φ(X) ⇒ Z ⊂cb∆ X. (cid:3) Example 3.15. Let Y be a dual operator space, and let H be a Hilbert space such that Y ⊂∆ B(H). Then, Y ∼∆ B(H). Proof. There exist w∗-continuous completely isometric maps ψ : Y → ψ(Y ) and projections p, q such that pφ(B(H)) ⊆ φ(B(H)), φ(B(H))q ⊆ φ(B(H)) and ψ(Y ) ∼T RO pφ(B(H))q. We define the map α : B(H) → B(H) given by α(x) = φ−1(pφ(x)). This is a multiplier of B(H), and also a projection. Because Al(B(H)) = B(H), there exists a projection p ∈ B(H) such that φ−1(pφ(x)) = px ⇒ φ(px) = pφ(x) ∀ x ∈ B(H). Similarly, there exists a projection q ∈ B(H) such that φ(xq) = φ(x)q ∀ x ∈ B(H). We have that Because φ−1(pφ(B(H))q) = pB(H)q. Al(pB(H)q) = B(p(H)), Ar(pB(H)q) = B(q(H)), it follows from Lemma 3.9 that there exists a w∗-continuous completely iso- metric map ζ : ψ(Y ) → ζ(ψ(Y )) such that ζ(ψ(Y )) ∼T RO pB(H)q. Thus, there exist TROs M1, M2 such that ζ(ψ(Y )) = [M2 pB(H)qM ∗ 1 ]−w∗ . Define N ∗ 2 = [M2 pB(H)]w∗ , N1 = [B(H)qM ∗ 1 ]−w∗ . 26 G. K. ELEFTHERAKIS Then, we have that [N2N ∗ 2 N2]w∗ = [B(H)pM ∗ 2 M2 pB(H)pM ∗ 2 ]w∗ = [B(H)pM ∗ 2 ]−w∗ = N2. Thus, N2 is a TRO. Similarly, N1 is a TRO. Now, ζ(ψ(Y )) = [N ∗ 2 B(H)N1]−w∗ and Thus, [N2B(H)N ∗ 1 ]−w∗ = [B(H)pM ∗ 2 B(H)M1 qB(H)]−w∗ = B(H). ζ(ψ(Y )) ∼T RO B(H) ⇒ Y ∼∆ B(H). (cid:3) Example 3.16. Let H be a Hilbert space, Φ : B(H) → B(H) be a w∗- continuous completely bounded idempotent map, and Y = RanΦ, Z = RanΦ⊥. Let φ : B(H) → Y ⊕ Z be the map given by φ(x) = Φ(x) ⊕ Φ⊥(x). We can easily prove that φ is a w∗-c.b. isomorphism onto Y ⊕ Z. By Example 3.5 (ii), we have that Y ⊂T RO φ(B(H)), and thus Y ⊂cb∆ B(H). Example 3.17. Let H be a Hilbert space, and let e, f ∈ B(H) be nontrivial projections. Define Φ : B(H) → B(H), Φ(x) = exf + exf ⊥ + e⊥xf ⊥, and denote Y = RanΦ. By Example 3.16, we have that Y ⊂cb∆ B(H). If Y weakly ∆-embeds into B(H), then by Example 3.15 we should have that Y ∼∆ B(H). However, this contradicts the fact that B(H) is a self-adjoint algebra and Y is a non-self-adjoint algebra. Thus, the relation Y ⊂cb∆ X does not always imply that Y ⊂∆ X holds. Acknowledgement: I would like to express appreciation to Dr Evgenios Kakariadis for his helpful comments and suggestions during the preparation of this work. References [1] D. P. Blecher and U. Kashyap, Morita equivalence of dual operator algebras, J. Pure Appl. Algebra, 212 (2008), 2401-2412 [2] D. P. Blecher and J. E. Kraus, On a generalization of W*-modules, Banach Center Publ. 91, (2010), 77-86 [3] D. P. Blecher and C. Le Merdy, Operator Algebras and Their Modules -- An Operator Space Approach, Oxford University Press, 2004 [4] D. P. Blecher, P. S. Muhly, and V. I. Paulsen, Categories of operator modules -- -Morita equivalence and projective modules, Memoirs of the A.M.S. 143 (2000) no. 681 MORITA EMBEDDINGS FOR DUAL OPERATOR ALGEBRAS AND DUAL OPERATOR SPACES27 [5] D. P. Blecher and V. Zarikian, The calculus of one sided M-ideals and multipliers in operator spaces, Memoirs of the A.M.S. 179 (842), 2003 [6] K. R. Davidson, Nest Algebras, Longman Scientific & Technical, Harlow, 1988 [7] E. Effros and Z.-J. Ruan, Operator Spaces, Oxford University Press, 2000 [8] G.K. Eleftherakis, TRO equivalent algebras, Houston J. of Mathematics, 38:1 (2012), 153-175 [9] G.K. Eleftherakis, A Morita type equivalence for dual operator algebras, J. Pure Appl. Algebra, 212:5 (2008), 1060-1071 [10] G.K. Eleftherakis, Morita type equivalences and reflexive algebras, J. Operator Theory, 64 (2010) no 1, 3-17 [11] G.K. Eleftherakis, Morita equivalence of nest algebras, Math. Scand., 113 (2013), no 1, 83-107 [12] G.K. Eleftherakis, V. I. Paulsen, Stably isomorphic dual operator algebras, Math. Ann., 341:1 (2008), 99-112 [13] G. K. Eleftherakis, V. I. Paulsen, and I. G. Todorov, Stable isomorphism of dual operator spaces, J. Funct. Anal. 258 (2010), 260 -- 278 [14] U. Kashyap, A Morita theorem for dual operator algebras, J. Funct. Analysis, 256 (2009) 3545-3567 [15] V. I. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge University Press, 2002 [16] G. Pisier, Introduction to Operator Space Theory, Cambridge University Press, 2000 G. K. Eleftherakis, University of Patras, Faculty of Sciences, Depart- ment of Mathematics, 265 00 Patras Greece E-mail address: [email protected]
1812.07194	1	1812	2018-12-18T06:31:24	Quotients of \'etale groupoids and the abelianizations of groupoid C*-algebras	[ "math.OA" ]	In this paper, we introduce quotients of \'etale groupoids. Using the notion of quotients, we describe the abelianizations of groupoid C-algebras. As another application, we obtain a simple proof that effectiveness of an \'etale groupoid is implied by the full uniqueness property of its groupoid C-algebra.	math.OA	math	QUOTIENTS OF ´ETALE GROUPOIDS AND THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS FUYUTA KOMURA Abstract. In this paper, we introduce quotients of ´etale groupoids. Using the notion of quotients, we describe the abelianizations of groupoid C-algebras. As another application, we obtain a simple proof that effec- tiveness of an ´etale groupoid is implied by the full uniqueness property of its groupoid C-algebra. 0. Introduction The study of C-algebras associated to ´etale groupoids, groupoid C- algebras, was initiated by Renault in [7]. Since then, many researchers have studied the relationship between ´etale groupoids and groupoid C- algebras. In the previous studies, there are many results for C-algebras associated to Hausdorff ´etale groupoids. For a non-Hausdorff ´etale groupoid, its C-algebra seems not to be studied sufficiently. However, non-Hausdorff groupoids naturally arise as mentioned in [3], [5] and so on. Exel pointed out that some results known for Hausdorff ´etale groupoids do not necessarily hold for non-Hausdorff groupoids in [4]. In [2], the authors treat simplicity of groupoid C-algebras associated to non-Hausdorff ´etale groupoids. In this paper, we calculate the abelianization of a groupoid C-algebra. For a discrete group Γ, the abelianization C∗(Γ)ab of its group C-algebra C∗(Γ) is isomorphic to C∗(Γab), where Γab is the abelianization of Γ. Fur- thermore, C∗(Γab) is isomorphic to C((cid:100)Γab), where (cid:100)Γab is the Pontryagin (cid:100)Gab so that C∗(G)ab (cid:39) C∗(Gab) (cid:39) C0((cid:100)Gab) holds. In order to construct dual of Γab. It is natural to consider an ´etale groupoid analogy. For an ´etale groupoid G, we construct an ´etale groupoid Gab and a topological groupoid Gab, we introduce the notion of quotient ´etale groupoids. A quotient ´etale groupoid often becomes non-Hausdorff even if the original ´etale groupoid is Hausdorff. Therefore, we treat not necessarily Hausdorff ´etale groupoids and their C-algebras, which are defined by Connes [3]. As a byproduct, we obtain a simple proof that effectiveness of an ´etale groupoid is implied by the full uniqueness property of its groupoid C-algebra (see Corollary 2.2.5). We remark that this result has been shown in [1] for Hausdorff ´etale groupoids in a different way and the proof in [1] seems to work for non-Hausdorff ´etale groupoids. 1 2 FUYUTA KOMURA This paper is organized as follows. In Section 1, we recall definitions and basic facts about not necessarily Hausdorff ´etale groupoids and their C-algebras. In Section 2, we introduce the notion of quotient ´etale groupoids and show some applications. Using quotients, we obtain a simple proof that full uniqueness property of a groupoid C-algebra induces effectiveness of an ´etale groupoid. In Section 3, for an ´etale groupoid G, we construct an ´etale abelian group bundle Gab through quotients. Finally, we show that the abelianization of a groupoid C-algebra is isomorphic to the C-algebra associated to Gab. Since the abelianization of a C-algebra is commutative, it is isomorphic to C0(X) for some locally compact Hausdorff space X by the Gelfand-Naimark theorem. We show that the abelianization of a groupoid C-algebra C∗(G) is isomorphic to C0((cid:100)Gab), where (cid:100)Gab is introduced in this paper. We obtain some results by using quotients of ´etale groupoids, which are not necessarily Hausdorff. We expect that this paper stimulates the study of non-Hausdorff ´etale groupoids. Acknowledgement. The author would like to thank his supervisor, Prof. Takeshi Katsura, for fruitful discussions and guidance. 1. ´Etale groupoids and groupoid C-algebras In this section, we recall the notions of ´etale groupoids and groupoid C-algebras. We refer to [7], [6] and [9] for details. 1.1. ´Etale groupoids. A groupoid is a set G together with a distinguished subset G(0) ⊂ G, source and range maps s, r : G → G(0) and a multiplication G(2) ··= {(α, β) ∈ G × G s(α) = r(β)} (cid:51) (α, β) (cid:55)→ αβ ∈ G such that (1) for all x ∈ G(0), s(x) = x and r(x) = x hold, (2) for all α ∈ G, αs(α) = r(α)α = α holds, (3) for all (α, β) ∈ G(2), s(αβ) = s(β) and r(αβ) = r(α) hold, (4) if (α, β), (β, γ) ∈ G(2), we have (αβ)γ = α(βγ), (5) every γ ∈ G, there exists γ(cid:48) ∈ G which satisfies (γ(cid:48), γ), (γ, γ(cid:48)) ∈ G(2) and s(γ) = γ(cid:48)γ and r(γ) = γγ(cid:48). Since the element γ(cid:48) in (5) is uniquely determined by γ, γ(cid:48) is called the inverse of γ and denoted by γ−1. We call G(0) the unit space of G. A subgroupoid of G is a subset of G which is closed under the inversion and multiplication. For U ⊂ G(0), we define GU ··= s−1(U ) and GU ··= r−1(U ). We define also Gx ··= G{x} and Gx ··= G{x} for x ∈ G(0). The isotropy bundle of G is denoted by Iso(G) ··= {γ ∈ G s(γ) = r(γ)}. If G satisfies G = Iso(G), G is called a group bundle over G(0). A group bundle G is said to be abelian if Gx is an abelian group for all x ∈ G(0). THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 3 A topological groupoid is a groupoid equipped with a topology where the multiplication and the inverse are continuous. Note that the source map and range map of a topological groupoid are continuous. Definition 1.1.1. A topological groupoid G is said to be ´etale if (1) the unit space G(0) ⊂ G is a locally compact Hausdorff space with respect to the relative topology of G, (2) the source map s : G → G(0) is locally homeomorphic (i.e. for all α ∈ G, there exists an open neighborhood U ⊂ G of α such that s(U ) ⊂ G(0) is open and sU is a homeomorphism onto s(U )). An ´etale topological groupoid is called an ´etale groupoid in short. In this paper, we assume that the unit space of an ´etale groupoid is a locally compact Hausdorff space. We do not assume that an ´etale groupoid is a Hausdorff space as a topological space. Note that a local homeomorphism is an open map. If s is locally homeo- morphic, then r is also locally homeomorphic since r(γ) = s(γ−1) holds for all γ ∈ G. By Definition 1.1.1, the family of all locally compact Hausdorff open subsets of G is an open basis for the topology of G. An ´etale groupoid G is said to be effective if G(0) = Iso(G)◦, where Iso(G)◦ denotes the interior of Iso(G). In some papers, the condition that the source map s : G → G(0) is locally homeomorphic in Definition 1.1.1 is replaced by the condition that the source map s : G → G is locally homeomorphic. As in Proposition 1.1.2, these definitions are equivalent. Proposition 1.1.2. Let G be an ´etale groupoid. The unit space G(0) is an open subset of G. Furthermore, the source and range maps s, r are local homeomorphisms as maps from G to G. Proof. First, we show that G(0) is an open subsets of G. Take x ∈ G(0) arbitrarily. Then, there exists an open subset U ⊂ G with x ∈ U such that s(U ) is an open subsets of G(0) and sU is a homeomorphism onto s(U ). Define V ··= U ∩ s−1(G(0) ∩ U ). Observe that V is an open subset of G with x ∈ V . Moreover, observe that V ⊂ G(0). Indeed, for every γ ∈ V , we have s(γ) ∈ U and s(γ) = s(s(γ)). Since sU is injective, it follows that γ = s(γ) ∈ G(0). Therefore, we have shown V ⊂ G(0) and this implies that G(0) is an open subset of G. Now, the second assertion can be easily (cid:3) checked. Definition 1.1.3. Let G be an ´etale groupoid. A subset U ⊂ G is called a bisection if both sU and rU are injective. For an ´etale groupoid G, an open bisection of G is a locally compact Hausdorff space because it is homeomorphic to open subset of G(0) and we assume that G(0) is locally compact Hausdorff. Note that the set of all open bisections composes an open basis of G. 4 FUYUTA KOMURA Example 1.1.4. Every locally compact Hausdorff space is regarded as a Hausdorff ´etale groupoid whose unit space coincides with the whole space. Example 1.1.5. Every topological group is regarded as a topological groupoid whose unit space is a singleton. A topological group is discrete if and only if it is ´etale as a topological groupoid. Example 1.1.6. Let X be a locally compact Hausdorff space and Γ be a discrete group. We denote by Aut(X) the set of all homeomorphisms on X, which is a group under the composition. An action of Γ on X is a group homomorphism α : Γ (cid:51) s (cid:55)→ αs ∈ Aut(X) and written α : Γ (cid:121) X. Now, we construct the groupoid associated to an action α : Γ (cid:121) X. Define Γ (cid:110)α X ··= Γ × X as a topological space. The unit space of Γ (cid:110)α X is X, which is identified with the subset of Γ (cid:110)α X via an inclusion X (cid:51) x (cid:55)→ (e, x) ∈ Γ (cid:110)α X. The source map and range map are defined by s((t, x)) = x and r((t, x)) = αt(x) respectively for (t, x) ∈ Γ (cid:110)α X. For a pair (t1, y), (t2, x) ∈ Γ (cid:110)α X with y = αt2(x), their multiplication is defined by (t1, y) · (t2, x) ··= (t1t2, x). An inverse is given by (t, x)−1 = (t−1, αt(x)). Then, Γ (cid:110)α X is a Hausdorff ´etale groupoid. Proposition 1.1.7 ([6, Proposition 2.2.4]). Let G be an ´etale groupoid and U, V ⊂ G be open sets. Then, a set U V ··= {αβ ∈ G α ∈ U, β ∈ V, s(α) = r(β)} ⊂ G is an open set. Furthermore, if U, V ⊂ G are open bisections, U V is also an open bisection. Proof. Take γ ∈ U V and an open bisection W ⊂ G with γ ∈ W . Then, there exist α ∈ U and β ∈ V such that γ = αβ. By the continuity of the multiplication of G, there exist open bisections U1, V1 ⊂ G with α ∈ U1 ⊂ U, β ∈ V1 ⊂ V and U1V1 ⊂ W . Note that r(U1V1) = r(U1 ∩ s−1(r(V1))) ⊂ G(0) is an open subset. Therefore, U1V1 = r−1(r(U1V1)) ∩ W ⊂ G is open. Now, we have γ ∈ U1V1 ⊂ U V , so U V ⊂ G is an open subset. Assume that U, V ⊂ G are open bisections. One can show that sU V and rU V are injective. Therefore, U V is an open bisection. (cid:3) Definition 1.1.8. Let G be a groupoid. A subset F ⊂ G(0) is said to be invariant if s(γ) ∈ F implies r(γ) ∈ F for all γ. A point x ∈ G(0) is called a fixed point if {x} ⊂ G(0) is invariant. Note that a set F ⊂ G(0) is invariant if and only if G(0) \ F is invariant. If F ⊂ G(0) is invariant, then GF = GF ∩ GF ⊂ G is a subgroupoid whose unit space is F . Proposition 1.1.9. Let G be an ´etale groupoid. Then, the set of all fixed points F ⊂ G(0) is a closed subset. Proof. We show that G(0) \ F ⊂ G(0) is an open set. Take x ∈ G(0) \ F . Then, there exists γ ∈ G such that x = s(γ) and x (cid:54)= r(γ). Take an open bisection U which contains γ. Let SU : s(U ) → r(U ) denote a homeomorphism defined by SU (s(α)) = r(α) for each α ∈ U . Since G(0) THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 5 is Hausdorff, there exist open sets U1, V1 ⊂ G(0) such that s(γ) ∈ U1 , r(γ) ∈ V1 and U1 ∩ V1 = ∅. By the continuity of SU , there exists an open set U2 ⊂ U such that γ ∈ U2 and SU (U2) ⊂ V1. Now, one can see U2 ⊂ G(0)\ F . Therefore, G(0) \ F ⊂ G(0) is an open set. (cid:3) We will use the next proposition for the set of all fixed points. Proposition 1.1.10. Let G be an ´etale groupoid and U, F ⊂ G(0) be an invariant open and closed subset respectively. Then, GU ⊂ G is an open subgroupoid of G and an ´etale groupoid in the relative topology. Similarly, GF ⊂ G is a closed subgroupoid of G and an ´etale groupoid in the relative topology. Proof. Observe that U and F are locally compact Hausdorff spaces in the relative topology of G(0). Now, it is clear that GU and GF are ´etale (cid:3) groupoids. In particular, if x ∈ G(0) is a fixed point, then Gx ⊂ G is a discrete group. 1.2. ´Etale groupoid C-algebras. Following Connes's idea in [3], we as- sociate a C-algebra to an ´etale groupoid which is not necessarily Hausdorff. Let G be an ´etale groupoid. For an open Hausdorff subset U ⊂ G, we denote the set of all continuous functions with compact support on U by Cc(U ). We regard an element in Cc(U ) as an element in Funct(G), the vector space of all complex valued functions on G, by defining it to be 0 U Cc(U ) ⊂ Funct(G), where the union is taken over all open Hausdorff subsets U ⊂ G. If G is Hausdorff, then C(G) coincides with Cc(G). If G is not Hausdorff, an element in C(G) may not be continuous. Proposition 1.2.1. Let G be an ´etale groupoid. Take an open basis {Ui}i∈I of G consisting of open Hausdorff subsets. Then, C(G) is the linear span of U Cc(U ), where the i∈I Cc(Ui). In particular, C(G) is the linear span of(cid:83) outside of U . We define C(G) ··= span(cid:83) (cid:83) union is taken over all open bisections of G. (cid:3) Proof. This follows from a partition of unity argument. Definition 1.2.2. Let G be an ´etale groupoid. Recall that C(G) is equipped with a structure of C-vector space by pointwise addition and scalar multi- plication. The multiplication f ∗ g ∈ C(G) and involution f∗ ∈ C(G) of f, g ∈ C(G) are defined by (cid:88) β∈Gs(γ) f ∗ g(γ) = f (γβ−1)g(β), f∗(γ) = f (γ−1). Then, C(G) is a -algebra under these operations. One can see that Cc(G(0)) ⊂ C(G) is a -subalgebra. Lemma 1.2.3. Let G be an ´etale groupoid and f ∈ C(G). Then, there exists Cf ≥ 0 such that (cid:107)ρ(f )(cid:107) ≤ Cf for all Hilbert spaces H and - homomorphisms ρ : C(G) → B(H). 6 FUYUTA KOMURA Proof. We may assume that f ∈ Cc(U ) for some open bisection U ⊂ G. One can see that f∗∗f ∈ Cc(G(0)). Since Cc(G(0)) is a union of commutative C-algebras, we have (cid:107)ρ(h)(cid:107) ≤ supx∈G(0)h(x) for all h ∈ Cc(G(0)). Then, we obtain (cid:107)ρ(f )(cid:107)2 = (cid:107)ρ(f∗ ∗ f )(cid:107) ≤ supx∈G(0)f∗ ∗ f (x) < ∞. (cid:3) The universal norm of f ∈ C(G) is defined by (cid:107)f(cid:107) ··= sup{(cid:107)ρ(f )(cid:107) ρ : C(G) → B(H) is a -representation}. is spanned by(cid:83) By Lemma 1.2.3, the universal norm takes values in [0,∞). Since the left regular representations of C(G) induces a faithful -representation of C(G), the universal norm becomes a C-norm (see [2, Section 4]). The completion of C(G) by universal norm is denoted by C∗(G). We shall remark that every -representation of C(G) induces the -representation of C∗(G). Note that the inclusion Cc(G(0)) ⊂ C(G) extends to C0(G(0)) ⊂ C∗(G). Proposition 1.2.4. Let G be an ´etale groupoid and F ⊂ G(0) be a closed invariant set. Then, the restriction C(G) (cid:51) f (cid:55)→ fGF ∈ C(GF ) extends to the surjective -homomorphism C∗(G) → C∗(GF ). Proof. First, we check that fGF ∈ C(GF ) for all f ∈ C(G). We may assume that f ∈ Cc(U ) for some open Hausdorff subset U ⊂ G, since C(G) U Cc(U ), where the union is taken over all open Hausdorff subsets U ⊂ G. Defining V ··= GF ∩ U , V is a Hausdorff open subset of GF . Then fGF is contained in Cc(V ) ⊂ C(GF ). Direct calculations imply that the restriction C(G) (cid:51) f (cid:55)→ fGF ∈ C(GF ) is a -homomorphism. Next, we show that the restriction C(G) (cid:51) f (cid:55)→ fGF ∈ C(GF ) is surjec- tive. Note that {GF ∩ U U ⊂ G is an open Hausdorff subset} is an open basis of GF . Take an open Hausdorff subset U ⊂ GF and f ∈ Cc(GF ∩ U ) arbitrarily. Put V ··= GF ∩ U . Since V ⊂ U is a closed subset of U and f ∈ Cc(V ), there exists f ∈ Cc(U ) such that fV = f by the Tietze exten- sion theorem. Now, we obtain f ∈ C(G) such that fGF = f . By Propo- U Cc(GF ∩ U ), where the union is taken over all open Hausdorff subsets U ⊂ G. Therefore, the restriction C(G) (cid:51) f (cid:55)→ fGF ∈ C(GF ) is surjective. By the universality of C∗(G), the restriction C(G) (cid:51) f (cid:55)→ fGF ∈ C(GF ) extends to the -homomorphism C∗(G) → C∗(GF ). Since the image of C∗(G) is dense in C∗(GF ), C∗(G) → C∗(GF ) is surjective. (cid:3) 2. Quotients of ´etale groupoids sition 1.2.1, C(GF ) is the linear span of (cid:83) In this section, we introduce the notion of quotient ´etale groupoids. 2.1. Quotients of ´etale groupoids. Definition 2.1.1. Let G be a groupoid. A subgroupoid H ⊂ G is said to be normal if (1) G(0) ⊂ H ⊂ Iso(G) holds and THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 7 (2) αHα−1 ⊂ H holds for all α ∈ G. Definition 2.1.2. Let G be a groupoid and H ⊂ G be a normal sub- groupoid. Then, we define an equivalence relation ∼ on G by declaring that α ∼ β if s(α) = s(β) and αβ−1 ∈ H. We denote the quotient set G/∼ by G/H. Lemma 2.1.3. Let G be a groupoid and H ⊂ G be a normal subgroupoid. Suppose that α, α(cid:48) ∈ G satisfy α ∼ α(cid:48). Then, we have s(α) = s(α(cid:48)) and r(α) = r(α(cid:48)). Proof. It follows that s(α) = s(α(cid:48)) from the definition of α ∼ α(cid:48). Since αα(cid:48)−1 ∈ H ⊂ Iso(G), we have r(α) = r(αα(cid:48)−1) = s(αα(cid:48)−1) = r(α(cid:48)). (cid:3) Lemma 2.1.4. Let G be a groupoid and H ⊂ G be a normal subgroupoid. Suppose that α, α(cid:48), β, β(cid:48) ∈ G satisfy α ∼ α(cid:48), β ∼ β(cid:48), s(α) = r(β). Then, we have s(α(cid:48)) = r(β(cid:48)) and αβ ∼ α(cid:48)β(cid:48). Proof. By Lemma 2.1.3, we have s(α) = s(α(cid:48)) and r(β) = r(β(cid:48)). Using s(α) = r(β), we obtain s(α(cid:48)) = r(β(cid:48)). s(αβ) = s(β) = s(β(cid:48)) = s(α(cid:48)β(cid:48)) and The last assertion follows from a simple calculation. Indeed, we have αβ(α(cid:48)β(cid:48))−1 = αββ(cid:48)−1α(cid:48)−1 = (αββ(cid:48)−1α−1)(αα(cid:48)−1) ∈ H. Note that αββ(cid:48)−1α−1 ∈ H, since H is normal. (cid:3) Definition 2.1.5. Let G be a groupoid, H ⊂ G be a normal subgroupoid and q : G → G/H be the quotient map. A groupoid structure of G/H is defined as the following; • a unit space (G/H)(0) is q(G(0)), which can be identified with G(0) via an injection qG(0), • source and range maps s, r : G/H → G(0) are defined by s(q(γ)) ··= q(s(γ)), r(q(γ)) ··= q(r(γ)) for γ ∈ G, • a multiplication of G/H is defined by q(α)q(β) ··= q(αβ) for α, β ∈ G with s(α) = r(β). One can see that the inverse map of G/H satisfies q(γ)−1 = q(γ−1) for γ ∈ G. Then, G/H is a groupoid under these operations. Remark 2.1.6. The operations of G/H are well-defined by Lemma 2.1.3 and Lemma 2.1.4. If G is a topological groupoid, then we consider the quotient topology as a topology of G/H. Lemma 2.1.7. Let G be an ´etale groupoid and H ⊂ G be an open nor- mal subgroupoid. Then, the quotient map q : G → G/H is an open map. Furthermore, q is a local homeomorphism. Proof. Let U ⊂ G be an open subset. Then, q−1(q(U )) = U H is an open subset of G by Proposition 1.1.7. Hence, q(U ) ⊂ G/H is an open subset by the definition of the quotient topology. 8 FUYUTA KOMURA Next, we show that the quotient map q : G → G/H is a local homeomor- phism. Fix a γ ∈ G. Then, take an open bisection U ⊂ G with γ ∈ U . One can see that qU is injective. Since q is an open map, qU is a homeomorphism onto an open subset q(U ) ⊂ G. Hence, q is a local homeomorphism. (cid:3) Observe that qG(0) : G(0) → (G/H)(0) is homeomorphic. Proposition 2.1.8. Let G be an ´etale groupoid and H ⊂ G be an open normal subgroupoid. Then, G/H is an ´etale groupoid. Proof. First, we show the continuity of the inverse G/H (cid:51) δ (cid:55)→ δ−1 ∈ G/H. One can see that a map G (cid:51) γ (cid:55)→ q(γ)−1 ∈ G/H is continuous, since the following diagram is commutative; G q inverse G q G/H inverse G/H. Next, we show that the multiplication of G/H is continuous. Take (q(α), q(β)) ∈ By the definition of the quotient topology, the inverse of G/H is continuous. (G/H)(2) and an open set U ⊂ G/H such that q(α)q(β) ∈ U . Since αβ ∈ q−1(U ) and q−1(U ) ⊂ G is open, there exist open sets V1, V2 ⊂ G such that α ∈ V1, β ∈ V2 and V1V2 ⊂ q−1(U ). Subsets V1, V2 ⊂ G are open, so q(V1), q(V2) ⊂ G/H are open. One can see that q(α) ∈ q(V1), q(β) ∈ q(V2) and q(V1)q(V2) = q(V1V2) ⊂ U . Therefore, the multiplication of G/H is continuous. Finally, we show that G/H is ´etale. Since the restriction qG(0) gives a homeomorphism from G(0) to (G/H)(0), (G/H)(0) is a locally compact Hausdorff space. One can see that the source map s : G/H → (G/H)(0) is a local homeomorphism, since we have Lemma 2.1.7 and the following diagram is commutative for every open bisection U ⊂ G; U s s(U ) q q q(U ) s s(q(U )). (cid:3) Therefore, G/H is an ´etale groupoid. Definition 2.1.9. Let G1 and G2 be groupoids. A map Φ : G1 → G2 is called a groupoid homomorphism if (Φ(α), Φ(β)) ∈ G(2) and Φ(αβ) = Φ(α)Φ(β) hold for all (α, β) ∈ G(2) 1 . 2 Now, we obtain the next theorem by Lemma 2.1.7 and Proposition 2.1.8, Theorem 2.1.10. Let G be an ´etale groupoid and H ⊂ G be an open normal subgroupoid. Then, the sequence of the ´etale groupoids THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 9 inclusion q G/H H G is exact, that is, q−1((G/H)(0)) = H. Proposition 2.1.11. Let G be an ´etale groupoid and H ⊂ G be an open normal subgroupoid. Then, G/H is Hausdorff if and only if H ⊂ G is closed. Proof. Recall that an ´etale groupoid G is Hausdorff if and only if its unit space G(0) is a closed subset of G (for example, see [9, Lemma 2.3.2]). If G/H is Hausdorff, (G/H)(0) ⊂ G/H is closed. Hence, H = q−1((G/H)(0)) is a closed subset of G. Suppose that H ⊂ G is closed. Since q is an open map, (G/H) \ (G/H)(0) = q(G \ H) ⊂ G/H is open. Hence, (G/H)(0) ⊂ G/H is closed, (cid:3) which implies that G/H is Hausdorff. Proposition 2.1.12. Let G be an ´etale groupoid. Then, the interior of isotropy Iso(G)◦ ⊂ Iso(G) is a normal subgroupoid. Proof. We show that Iso(G)◦ is normal. By Proposition 1.1.2, G(0) is contained in Iso(G)◦. Take α ∈ G and γ ∈ Iso(G)◦ with s(α) = r(γ). There exist open bisections U, V ⊂ G with α ∈ U and γ ∈ V ⊂ Iso(G). Then, by Proposition 1.1.7, U V U−1 ⊂ G is an open subset which contains αγα−1. Since U is bisection and V ⊂ Iso(G), we have U V U−1 ⊂ Iso(G). Therefore, αγα−1 ∈ Iso(G)◦ and Iso(G)◦ is an open normal subgroupoid. (cid:3) An ´etale groupoid G/ Iso(G)◦, which is a special case of quotient groupoids, coincides with a groupoid of germs of the canonical action (see [8, Section 3]). One can see that G/ Iso(G)◦ is effective. 2.2. -homomorphisms induced by quotients of ´etale groupoids. For an ´etale groupoid G and an open normal subgroupoid H ⊂ G, we have obtained the quotient ´etale groupoid G/H. Next, we see that the quotient map q : G → G/H induces a -homomorphism C∗(G) → C∗(G/H). For f ∈ C(G), we define f : G/H → C by (cid:88) f (γ) ··= f (α) q(α)=γ for γ ∈ G/H. Then, the following proposition holds. Proposition 2.2.1. Let G be an ´etale groupoid and H ⊂ G be an open normal subgroupoid. Then, C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) is a surjective - homomorphism. Proof. First, we show f ∈ C(G/H). We may assume that there exists an open bisection U ⊂ G such that fU ∈ Cc(U ) and fG\U = 0. Then, q(U ) ⊂ G/H is an open bisection and fq(U ) = f ◦ (qU )−1 ∈ Cc(q(U )), since qU is a homeomorphism onto the image. Moreover, one can see that f(G/H)\q(U ) = 0. Hence, f ∈ Cc(q(U )) ⊂ C(G/H). 10 FUYUTA KOMURA We show that C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) is a -homomorphism. We only check that C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) preserves the multiplications, since it is easy to check that this map is linear and preserves the involutions. For all f, g ∈ C(G) and γ(cid:48) ∈ G/H, we have (cid:88) (cid:88) (cid:88) q(γ)=γ(cid:48) α(cid:48)β(cid:48)=γ(cid:48) (cid:93)f ∗ g(γ(cid:48)) = f ∗ g(γ(cid:48)) = = (cid:88) (cid:88) (cid:88) αβ=γ q(γ)=γ(cid:48) (cid:88) f ∗ g(γ) = f (α(cid:48))g(β(cid:48)) = (cid:88) (cid:88) q(αβ)=γ(cid:48) f (α)g(β) α(cid:48)β(cid:48)=γ(cid:48) q(α)=α(cid:48) q(β)=β(cid:48) f (α)g(β) = f (α)g(β), f (α)g(β). q(αβ)=γ(cid:48) Hence, C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) is a surjective -homomorphism. Finally, we show that C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) is surjective. Note that {q(U ) ⊂ G/H U ⊂ G is an open bisection} is an open basis of G. Let U ⊂ G be an open bisection and f ∈ Cc(q(U )). One can see that Then, we have (cid:101)g = f . By Proposition 1.2.4, C(G/H) is the linear span qU is a homeomorphism onto its image. Define g ··= f ◦ qU ∈ Cc(U ). of (cid:83) U Cc(q(U )), where the union is taken over all open bisections U ⊂ G. (cid:3) By the universality of C∗(G), the surjective -homomorphism C(G) (cid:51) f (cid:55)→ f ∈ C(G/H) induces the surjective -homomorphism Q : C∗(G) → C∗(G/H). Proposition 2.2.2. Let Q : C∗(G) → C∗(G/H) be the -homomorphism as above. Then, ker Q ∩ C0(G(0)) = {0} holds. Proof. Since the universal norm of a function in Cc(G(0)) coincides with the supremum norm, QCc(G(0)) is isometric. Therefore, QC0(G(0)) is isometric and ker Q ∩ C0(G(0)) = {0}. (cid:3) Definition 2.2.3. Let G be an ´etale groupoid. We say that G has the full uniqueness property if every non-zero ideal I ⊂ C∗(G) satisfy I∩C0(G(0)) (cid:54)= {0}. The full uniqueness property of G is equivalent to the condition that a -representation π : C∗(G) → B(H) is injective if and only if πC0(G(0)) is injective. Lemma 2.2.4. Let G be an ´etale groupoid and H ⊂ G be an open normal subgroupoid. Then, the -homomorphism Q : C∗(G) → C∗(G/H) induced by Proposition 2.2.1 is injective if and only if H = G(0). Proof. It is clear that the -homomorphism Q : C∗(G) → C∗(G/H) is injective if H = G(0). Suppose that G(0) (cid:40) H and take γ0 ∈ H \ G(0). There exists an open bisection U ⊂ G with γ0 ∈ U ⊂ H. By the Urysohn lemma, THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 11 there exists f1 ∈ Cc(U ) with f1(γ0) = 1. Define f2 ∈ Cc(G(0)) by f2(γ) = f1 ◦ (sU )−1(γ) 0 (γ ∈ s(U )) (γ ∈ G(0) \ s(U )). (cid:40) We have f ··= f1 − f2 (cid:54)= 0, since f (γ0) = 1. One can see that Q(f ) = 0, (cid:3) which means that Q is not injective. Recall that an ´etale groupoid G is said to be effective if G(0) = Iso(G)◦. Corollary 2.2.5 (cf. [1, Proposition 5.5]). Let G be an ´etale groupoid. If G has the full uniqueness property, then G is effective. Proof. By Proposition 2.1.12, Iso(G)◦ is a normal subgroupoid of G. Let- ting Q : C∗(G) → C∗(G/ Iso(G)◦) be the -homomorphism induced by Propo- sition 2.2.1, we have ker Q ∩ C0(G(0)) = {0}. The full uniqueness property implies that Q : C∗(G) → C∗(G/ Iso(G)◦) is injective. Therefore, we obtain Iso(G)◦ = G(0) by Lemma 2.2.4. (cid:3) Remark 2.2.6. It had been known that Corollary 2.2.5 holds for Hausdorff ´etale groupoids as proved in [1, Proposition 5.5]. In [1, Proposition 5.5], the authors use the augmentation representations, which seems to work for non-Hausdorff ´etale groupoids. As shown in Proposition 2.2.1, the quotient map G → G/ Iso(G)◦ of ´etale groupoids induces the -homomorphism C∗(G) → C∗(G/ Iso(G)◦). Using this -homomorphism, we obtain the proof of Corollary 2.2.5, which seems to be more direct than the one in [1, Proposition 5.5]. We shall remark that G/ Iso(G)◦ coincides with the groupoid of germs of the canonical action in [8, Proposition 3.2]. The converse of Corollary 2.2.5 does not hold for non-Hausdorff ´etale groupoids. Indeed, Exel showed that there exists an effective ´etale groupoid G which does not have the full uniqueness property in [4] (cf. Example 2.2.7). Example 2.2.7 ([4, Section 2]). Let X ··= ([−1, 1]×{0})∪ ({0}× [−1, 1]) ⊂ R2 and K ··= {e, s, t, st} be the Klein group, which is isomorphic to Z/2Z ⊕ Z/2Z. We define an action σ of K on X by fot (x, y) ∈ X. σs((x, y)) = (−x, y), σt((x, y)) = (x,−y), σst((x, y)) = (−x,−y) Consider the transformation groupoid G ··= K (cid:110)σ X (see Example 1.1.6). One can see that Iso(G) = G(0) ∪ {(s, (0, y)) ∈ G y ∈ [−1, 1]} ∪ {(t, (x, 0)) ∈ G x ∈ [−1, 1]} ∪ {(st, (0, 0))}. Moreover, we have Iso(G)◦ = Iso(G)\{(s, (0, 0)), (t, (0, 0)), (st, (0, 0))}. Since Iso(G)◦ is not closed in G (for example, (s, (0, 0)) ∈ Iso(G)◦ \ Iso(G)◦), the quotient ´etale groupoid G/ Iso(G)◦ is not Hausdorff by Proposition 2.1.11. In [4], Exel shows that G/ Iso(G)◦ does not have the full uniqueness property, although it is effective. 12 FUYUTA KOMURA (0,1) (−1,0) (1,0) (0,−1) Figure 1. Picture of X in Example 2.2.7 Based on the work in [2], we shall give a necessary and sufficient condition of the full uniqueness property. Let G be an ´etale groupoid. We denote the left representation at x ∈ G(0) by λx : C(G) → B((cid:96)2(Gx)). By the universality of C∗(G), the left representation extends to the -representation λx : C∗(G) → B((cid:96)2(Gx)). Following [2], we say that an element a ∈ C∗(G) is singular if the interior of {γ ∈ G (cid:104)δγλs(γ)(a)δs(γ)(cid:105) (cid:54)= 0} is empty, where δγ ∈ (cid:96)2(Gx) denotes the delta function at γ ∈ Gx. We denote the reduced groupoid C-algebra of G by C∗ λ(G). We denote the canonical surjective -homomorphism by Q : C∗(G) → C∗ λ(G). In [2], the notion of a singular element is defined for elements in C∗ λ(G) in the same way as elements in C∗(G). Since an element in ker Q is singular, C∗(G) has no nonzero singular elements if and only if C∗ λ(G) has no nonzero singular elements and C∗(G) (cid:39) C∗ λ(G) via the canonical -homomorphism Q. Proposition 2.2.8. Let G be a second countable ´etale groupoid. Then, G has the full uniqueness property if and only if • G is effective and • C∗(G) has no nonzero singular elements. 3.6]. Therefore, letting π ··=(cid:76) Proof. Assume that G has the full uniqueness property. Corollary 2.2.5 implies that G is effective. We show that C∗(G) has no nonzero singular elements. Let a ∈ C∗(G) be a singular element. We define S ··= {x ∈ G(0) Gx ∩ Gx = {x}}, which is a dense subset of G(0) by [8, Proposition x∈S λx, π is injective on C0(G(0)). The full uniqueness property implies that π is injective. Since λx(a) = λx(Q(a)) holds for all x ∈ G(0), we have s({γ ∈ G (cid:104)δγλs(γ)(a)δs(γ)(cid:105) (cid:54)= 0}) ⊂ G(0) \ S by [2, Lemma 4.2]. Using this fact, one can see that π(a) = 0 if a ∈ C∗(G) is singular. Hence, C∗(G) has no nonzero singular element. Next, we show the converse. Note that Q is an isomorphism by the assumpution that C∗(G) has no nonzero singular elements. Now, the full (cid:3) uniqueness property of G follows from [2, Theorem 4.4]. THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 13 3. The abelianizations of ´etale groupoid C-algebras In this section, we calculate the abelianizations of ´etale groupoid C- algebras. First, recall the abelianizations of C-algebras. For a C-algebra A, its abelianization is defined by Aab = A/I, where I ⊂ A is the closed two-sided ideal generated by {xy − yx ∈ A x, y ∈ A}. The abelianization Aab is a commutative C-algebra with the following universality; for all commutative C-algebra B and -homomorphism π : A → B, there exists the unique -homomorphism π : Aab → B such that π◦q = π, where q : A → Aab denotes the quotient map. 3.1. One dimensional representations of a groupoid C-algebra. For a C-algebra A, we denote the set of all one dimensional nondegen- erate representations of A by ∆(A). Namely, ∆(A) is the set of all nonzero -homomorphisms from A to C. We suppose that ∆(A) is equipped with the pointwise convergence topology. If A is commutative, ∆(A) is known as the Gelfand spectrum of A. First, we calculate ∆(C∗(G)). Let G be an ´etale groupoid and x ∈ G(0) be a fixed point of G. Note that Gx is a discrete group. We denote the surjection in Proposition 1.2.4 by Qx : C∗(G) → C∗(Gx) temporarily. Also, we denote the circle group by T ··= {z ∈ C z = 1}. For a group homomorphism χ : Gx → T, a χ(γ)f (γ) ∈ C is a -homomorphism. This - homomorphism extends to the -homomorphism C∗(Gx) → C, which we also denote by χ : C∗(Gx) → C. Definition 3.1.1. Let G be an ´etale groupoid, x ∈ G(0) be a fixed point and χ : Gx → T be a group homomorphism. Then, we define a -homomorphism ϕx,χ : C∗(G) → C by ϕx,χ ··= χ ◦ Qx. map Cc(Gx) (cid:51) f (cid:55)→ (cid:80) γ∈Gx We will show that all elements of ∆(C∗(G)) are obtained by this form (Theorem 3.1.8). Proposition 3.1.2. Let G be an ´etale groupoid and ϕ ∈ ∆(C∗(G)). Then, there uniquely exists xϕ ∈ G(0) which satisfies ϕ(f ) = f (xϕ) for all f ∈ C0(G(0)). Proof. We have ϕC0(G0) (cid:54)= 0 since C0(G(0)) has an approximate identity of C∗(G). Therefore, ϕC0(G(0)) belongs to ∆(C0(G(0))). Now, the existence and uniqueness of xϕ ∈ G(0) follow from the Gelfand-Naimark theorem. (cid:3) Proposition 3.1.3. Let G be an ´etale groupoid and ϕ ∈ ∆(C∗(G)). Then, xϕ ∈ G(0) is a fixed point. Proof. Assume that γ ∈ G satisfies s(γ) = xϕ. We will show r(γ) = xϕ. There exists an open bisection U ⊂ G with γ ∈ U . Take nγ ∈ Cc(U ) γ ∗ nγ ∈ Cc(G(0)) and which satisfies nγ(γ) = 1. Note that we have n∗ γ ∗ nγ(xϕ) = nγ(γ)2 = 1. Fix f ∈ Cc(G(0)) arbitrarily. n∗ 14 FUYUTA KOMURA γ∗f ∗nγ) = ϕ(n∗ γ)ϕ(f )ϕ(nγ) = ϕ(n∗ Direct calculations show that n∗ γ ∗ f ∗ nγ(xϕ) = nγ(γ)f (r(γ))nγ(γ) = γ ∗ f ∗ nγ ∈ Cc(G(0)). Then, f (r(γ)). On the other hand, one can see that n∗ we have γ∗nγ)ϕ(f ) = f (xϕ). γ∗f ∗nγ(xϕ) = ϕ(n∗ n∗ Therefore, f (r(γ)) = f (xϕ) holds for all f ∈ Cc(G(0)), which implies r(γ) = xϕ. Hence, xϕ ∈ G(0) is a fixed point of G. (cid:3) Proposition 3.1.4. Let G be an ´etale groupoid, ϕ ∈ ∆(C∗(G)) and γ ∈ Gxϕ. Take an open bisection Uγ ⊂ G with γ ∈ Uγ and fγ ∈ Cc(Uγ) with fγ(γ) = 1. Then, ϕ(fγ) is independent of the choice of Uγ and fγ. Moreover, we have ϕ(fγ) ∈ T. Proof. First, we show ϕ(fγ) ∈ T. Since f∗ γ ∗ fγ ∈ C0(G(0)), we have ϕ(fγ)2 = ϕ(f∗ γ ∗ fγ) = f∗ γ ∗ fγ(xϕ) = fγ(γ)2 = 1. Therefore, ϕ(fγ) ∈ T. Second, we show that ϕ(fγ) is independent of the choice of Uγ and fγ. Assume that fγ ∈ Cc(Uγ) and gγ ∈ Cc(Vγ) satisfies fγ(γ) = gγ(γ) = 1, where Uγ and Vγ ⊂ G are open bisections. Find a function h ∈ Cc(s(Uγ ∩ Vγ)) ⊂ Cc(G(0)) such that h(s(γ)) = 1. Recall that s(γ) = r(γ) = xϕ since xϕ is a fixed point. Also, note that ϕ(h) = h(xϕ) = 1. Putting (cid:101)fγ ··= fγ ∗ h and (cid:101)gγ = gγ ∗ h, we have that (cid:101)fγ and (cid:101)gγ are contained in Cc(Uγ ∩ Vγ). Then, it follows that (cid:101)fγ ∗ ∗ (cid:101)gγ ∈ C0(G(0)) and ϕ(fγ)ϕ(gγ) = ϕ(h)ϕ(fγ)ϕ(gγ)ϕ(h) = ϕ((cid:101)fγ ∗ ∗ (cid:101)gγ) ∗ ∗ (cid:101)gγ(xϕ) = h(r(γ))fγ(γ)gγ(γ)h(s(γ)) = 1. Now, we have ϕ(fγ) = ϕ(gγ) since ϕ(fγ) ∈ T. (cid:3) Proposition 3.1.5. Let G be an ´etale groupoid and ϕ ∈ ∆(C∗(G)). We define χϕ : Gxϕ → T by χϕ(γ) ··= ϕ(fγ), where γ ∈ Gxϕ and fγ ∈ C(G) is a function as in Proposition 3.1.4. Then, χϕ : Gxϕ → T is a group homomor- phism. Proof. Take α, β ∈ Gxϕ. We show χϕ(α)χϕ(β) = χϕ(αβ). Take fα, fβ ∈ C(G) as in Proposition 3.1.4. It follows that fα ∗ fβ ∈ Cc(U ) for some open bisection U ⊂ G and fα ∗ fβ(αβ) = 1. Hence, we have = (cid:101)fγ χϕ(αβ) = ϕ(fα ∗ fβ) = ϕ(fα)ϕ(fβ) = χϕ(α)χϕ(β) by the definition of χϕ. Proposition 3.1.6. Let G be an ´etale groupoid. Then, we have ϕ = ϕxϕ,χϕ for all ϕ ∈ ∆(C∗(G)). Proof. Take f ∈ Cc(U ), where U ⊂ G is an open bisection. It suffices to show ϕ(f ) = ϕxϕ,χϕ(f ), since C∗(G) is generated by such functions. If Gxϕ ∩ f−1(C \ {0}) = ∅, then we have Note that f∗ ∗ f ∈ Cc(G(0)). (cid:3) THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 15 0 = f∗ ∗ f (xϕ) = ϕ(f )2. Since the restriction of fGxϕ is zero, it follows that ϕxϕ,χϕ(f ) = 0 = ϕ(f ). If Gxϕ ∩ f−1(C \ {0}) (cid:54)= ∅, Gxϕ ∩ f−1(C \ {0}) is a singleton because f is supported on an open bisection. Let γ ∈ Gxϕ ∩ f−1(C \ {0}) be the unique element of Gxϕ ∩ f−1(C \ {0}). Observe that F ··= f /f (γ) ∈ Cc(U ) satisfies F (γ) = 1. Now, we have ϕxϕ,χϕ(f ) = f (γ)χϕ(γ) = f (γ)ϕ(F ) = ϕ(f ). (cid:3) Hence, we have ϕxϕ,χϕ = ϕ. Proposition 3.1.7. Let G be an ´etale groupoid, x ∈ G(0) be a fixed point and χ : Gx → T be a group homomorphism. Then, x = xϕx,χ and χ = χϕx,χ. Proof. First, we show x = xϕx,χ. Take f ∈ Cc(G(0)) arbitrarily. Then, we have f (xϕx,χ) = ϕx,χ(f ) = f (x)χ(x) = f (x). Hence, it follows x = xϕx. Next, we show χ = χϕx,χ. Take γ ∈ Gx arbitrarily. There exist an open bisection U ⊂ G with γ ∈ U and f ∈ Cc(U ) with f (γ) = 1. Then, we have χϕx,χ(γ) = ϕx,χ(f ) = f (γ)χ(γ) = χ(γ). Hence, we have shown x = xϕx,χ and χ = χϕx,χ. (cid:3) Combining Proposition 3.1.6 and Proposition 3.1.7, we obtain the next theorem. Theorem 3.1.8. Let G be an ´etale groupoid. Define a set D ··= {(x, χ) x ∈ G(0) is a fixed point and χ : Gx → T is a group homomorphism}. Then, a map is bijective. D (cid:51) (x, χ) −→ ϕx,χ ∈ ∆(C∗(G)) tator subgroupoid of G by [G, G] ··=(cid:83) 3.2. Construction of an ´etale abelian group bundle Gab. For an ´etale groupoid G, we construct an ´etale abelian group bundle Gab so that C∗(G)ab (cid:39) C∗(Gab) holds. Proposition 3.2.1. Let G be an ´etale group bundle. We define the commu- x∈G(0)[Gx, Gx], where [Gx, Gx] is the commutator subgroup of Gx. Then, [G, G] is an open normal subgroupoid of G. Proof. It is obvious that [G, G] ⊂ G is a normal subgroupoid. We show that [G, G] ⊂ G is open. Take γ ∈ [G, G]. By the definition of the commu- tator subgroup, there exists {αj}k j=1 ⊂ Gs(γ) such that γ = α1β1α−1 1 β−1 ··· αkβkα−1 k β−1 k . j=1,{βj}k 2 β−1 1 α2β2α−1 2 16 FUYUTA KOMURA 1 V −1 1 V −1 1 Take open bisections Uj, Vj ⊂ G such that αj ∈ Uj and βj ∈ Vj for all j = 1, 2, . . . , k. We show that U1V1U−1 1 ⊂ [G, G], where we define U−1 ··= {γ−1 γ ∈ U} for U ⊂ G. Fix γ(cid:48) ∈ U1V1U−1 . Then, there exist α, α(cid:48) ∈ U1 and β, β(cid:48) ∈ V1 which satisfy γ = αβα(cid:48)−1β(cid:48)−1. Since G is a group bundle, we have s(α) = s(α(cid:48)) = s(β) = s(β(cid:48)). We obtain α = α(cid:48) and β = β(cid:48) because U1 and V1 are bijections. Therefore, γ(cid:48) = αβα−1β−1 ∈ [G, G]. Similarly, one can show that U1V1U−1 k ⊂ [G, G]. 1 V −1 is an open set and this contains γ. Hence, [G, G] ⊂ G is an open normal sub- (cid:3) groupoid. By Proposition 1.1.7, U1V1U−1 ··· UkVkU−1 k V −1 ··· UkVkU−1 1 V −1 1 U2V2U−1 1 U2V2U−1 2 V −1 k V −1 2 V −1 2 2 k Let G be an ´etale groupoid. Recall that the set of all fixed points F ⊂ G(0) is a closed subset of G(0) (Proposition 1.1.9). We define Gfix ··= GF , which is an ´etale groupoid from Proposition 1.1.10. Since we have Gfix = Iso(Gfix), Gfix is an ´etale group bundle. Definition 3.2.2. Let G be an ´etale groupoid. We define the abelianization of G by Gab ··= Gfix/[Gfix, Gfix]. Let G be an ´etale groupoid. Then, we have a -homomorphism C∗(G) → C∗(Gfix) induced by the restriction (Proposition 1.2.4). Composing with the -homomorphism C∗(Gfix) → C∗(Gab) in Proposition 2.2.1, we obtain a -homomorphism π : C∗(G) → C∗(Gab). Note that C∗(G) is commutative if and only if G is an ´etale abelian group bundle. In particular, C∗(Gab) is commutative. Lemma 3.2.3. Let G be an ´etale groupoid. Then, a map Φ : ∆(C∗(Gab)) (cid:51) χ (cid:55)→ χ ◦ π ∈ ∆(C∗(G)) is bijective. Proof. Surjectivity of π implies that Φ is injective. We show that Φ is surjective. Take ϕ ∈ ∆(C∗(G)). Then, we have the fixed point xϕ ∈ G(0) and the group homomorphism χϕ which makes the following diagram com- mutative; C∗(G) ϕ C q C∗(Gxϕ), χϕ where q : C∗(G) → C∗(Gxϕ) is the -homomorphism obtained in Proposition 1.2.4. ··= (Gxϕ)ab = (Gab)xϕ, we obtain the group xϕ → T which makes the following diagram commu- By the universality of Gab xϕ homomorphism ¯χϕ : Gab tative; THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS 17 C∗(Gxϕ) χϕ C q(cid:48) C∗(Gab xϕ), ¯χϕ where q(cid:48) : C∗(Gxϕ) → C∗(Gab xϕ) denotes the -homomorphism induced by the quotient map Gxϕ → Gab xϕ. Let res : C∗(Gab) → C∗(Gab the restriction C(Gab) → C(Gab following commutative diagram; xϕ) denote the -homomorphism obtained by xϕ) (see Proposition 1.2.4). Now, we have the ϕ C∗(G) q C∗(Gxϕ) χϕ C π q(cid:48) ¯χϕ C∗(Gab) res C∗(Gab xϕ). In particular, we have ϕ = ( ¯χϕ ◦ res)◦ π and ¯χϕ ◦ res ∈ ∆(C∗(Gab)). Hence, (cid:3) Φ is surjective. Now, we are ready to calculate the abelianization of C∗(G). Theorem 3.2.4. Let G be an ´etale groupoid. Then, C∗(G)ab is isomorphic to C∗(Gab) via the unique isomorphism ¯π which makes the following diagram commutative; C∗(G) π C∗(Gab) Q C∗(G)ab, ¯π where Q : C∗(G) → C∗(G)ab denotes the quotient map. Proof. By the universality of C∗(G)ab, we obtain a -homomorphism which makes the following diagram commutative; C∗(G) π C∗(Gab) Q C∗(G)ab. ¯π It is clear that ¯π is surjective. We show that ¯π is injective. Suppose that a ∈ C∗(G) satisfies π(a) = 0. It suffices to show Q(a) = 0, which is equivalent to ¯ϕ(Q(a)) = 0 for all ¯ϕ ∈ ∆(C∗(G)ab) since C∗(G)ab is commutative. Take 18 FUYUTA KOMURA ¯ϕ ∈ ∆(C∗(G)ab) and define ϕ ··= ¯ϕ◦ Q. Then, by Lemma 3.2.3, there exists ϕ ∈ ∆(C∗(Gab)) which makes the following diagram commutative; C∗(G) π C∗(Gab) ϕ ϕ C. Now, we have the following commutative diagram; C∗(Gab) C∗(G) π Q C∗(G)ab ϕ ¯ϕ ϕ C. Hence, we have ¯ϕ(Q(a)) = ϕ(π(a)) = 0. (cid:3) 3.3. Duals of ´etale abelian group bundles. Let G be an ´etale groupoid. Since the abelianization of C∗(G) is a commutative C-algebra, C∗(G)ab is isomorphic to C0(∆(C∗(G)ab)) via the Gelfand transformation. In this section, we calculate the Gelfand spectrum ∆(C∗(G)ab). the set of all group homomorphisms from Γ to T, which is denoted by (cid:98)Γ. Then,(cid:98)Γ is an abelian group with respect to the pointwise multiplication. It is known that(cid:98)Γ is a compact abelian topological group with respect to the For a discrete abelian group Γ, its Pontryagin dual group is defined as topology of pointwise convergence. Proposition 3.3.1. Let Γ be a discrete group and Q : C∗(Γ) → C∗(Γab) be the -homomorphism induced by the quotient map Γ → Γab. Then, a map Φ :(cid:100)Γab (cid:51) χ (cid:55)→ χ ◦ Q ∈ ∆(C∗(Γ)) is a homeomorphism. Hence, C∗(Γ)ab is isomorphic to C((cid:100)Γab). Proof. This follows from the universality of Γab and C∗(Γ). (cid:3) As seen in the previous proposition, the key to calculate ∆(C∗(G)) is the Pontryagin dual. Definition 3.3.2. Let G be an ´etale abelian group bundle. We define a group bundle (cid:98)G ··= {(χ, x) x ∈ G(0), χ ∈ (cid:99)Gx} over G(0). Note that (cid:98)G is a group bundle such that (cid:98)Gx = (cid:99)Gx × {x}((cid:39) (cid:99)Gx) for every Let G be an ´etale abelian group bundle and (χ, x) ∈ (cid:98)G. Recall that we C(G), we define evf : (cid:98)G → C by evf ((χ, x)) = ϕx,χ(f ), where (χ, x) ∈ (cid:98)G. We x ∈ G(0). obtain the -homomorphism ϕx,χ ∈ ∆(C∗(G)) as in Definition 3.1.1. Definition 3.3.3. Let G be an ´etale abelian group bundle. For each f ∈ 19 THE ABELIANIZATIONS OF GROUPOID C-ALGEBRAS define a topology of (cid:98)G as the weakest topology in which evf is continuous Ψ : ∆(C∗(G)) (cid:51) ϕ (cid:55)→ (χϕ, xϕ) ∈ (cid:98)G for all f ∈ C(G). Proposition 3.3.4. Let G be an ´etale abelian group bundle. Then, the map1) is a homeomorphism. Hence, C∗(G) is isomorphic to C0((cid:98)G) by Ψ−1((χ, x)) = ϕx,χ for each (χ, x) ∈ (cid:98)G. For each f ∈ C(G), a map Proof. Proposition 3.1.8 states that Ψ is a bijection and Ψ−1 is given ∆(C∗(G)) (cid:51) ϕ (cid:55)→ evf ((χϕ, xϕ)) = ϕ(f ) ∈ C is continuous. This means that Ψ is continuous. The continuity of Ψ−1 follows from approximation (cid:3) arguments. Therefore, Ψ is a homeomorphism. compact if and only if G(0) is compact. Let G be an ´etale groupoid. Recall that Gab is an ´etale abelian group bundle. Corollary 3.3.5. Let G be an ´etale groupoid. Then, C∗(G)ab is isomorphic Proof. Recall that C∗(G)ab is isomorphic to C∗(Gab) by Theorem 3.2.4. Since Gab is an ´etale abelian group bundle, Proposition 3.3.4 implies that (cid:3) to C0((cid:100)Gab). C∗(Gab) is isomorphic to C0((cid:100)Gab). Proposition 3.3.6. Let G be an ´etale abelian group bundle. Then, (cid:98)G is a locally compact Hausdorff topological group bundle. Furthermore, (cid:98)G is Proof. It is clear that (cid:98)G is locally compact Hausdorff, since (cid:98)G is homeo- take f ∈ C(G) arbitrarily. Then, the map (cid:98)G(2) (cid:51) (χ1, χ2) (cid:55)→ evf (χ1χ2) = evf (χ1) evf (χ2) ∈ C is continuous. Therefore, the multiplication of (cid:98)G(2) is continuous. Similarly, one can show that the inverse is continuous. Hence, (cid:98)G follows from the fact that G(0) is compact if and only if C∗(G) (cid:39) C0((cid:98)G) is morphic to ∆(C∗(G)). In order to show the continuity of the operations, is a locally compact Hausdorff topological group bundle. The last assertion (cid:3) unital. Example 3.3.7. Let S3 = (cid:104)s, t s3 = t2 = e, st = ts2(cid:105) = {e, s, s2, t, ts, ts2} be the symmetric group of degree 3 and A3 ··= {e, s, s2} ⊂ S3 be the sub- group of even permutations. Let G ··= S3 × [0, 1] \ {(t, 1) t (cid:54)∈ A3} be an ´etale group bundle over [0, 1]. Then, G can be drawn as follows; (ts2,0) (s2,0) (ts,0) (s,0) (t,0) (e,0) (s2,1) (s,1) (e,1) 1)See Proposition 3.1.2 and 3.1.5 for the definition of xϕ and χϕ. 20 FUYUTA KOMURA One can see that [G, G] ⊂ G is not closed. By Proposition 2.1.11, Gab = G/[G, G] is not Hausdorff. Indeed, letting q : G → Gab denote the quotient map, Gab looks like the following; q((t,0)) The dual (cid:100)Gab of Gab can be drawn as follows; q((e,0)) q((e,1)),q((s,1)),q((s2,1)) Note that (cid:100)Gab is not ´etale. References [1] J. Brown, L. O. Clark, C. Farthing, and A. Sims. Simplicity of algebras associated to ´etale groupoids. Semigroup Forum, 88(2):433 -- 452, 2014. [2] L. O. Clark, R. Exel, E. Pardo, A. Sims, and C Starling. Simplicity of algebras asso- ciated to non-Hausdorff groupoids. arXiv:1806.04362, 2018. [3] A. Connes. A survey of foliations and operator algebras. In Proc. Sympos. Pure, volume 38, pages 521 -- 628, 1982. [4] R. Exel. Non-Hausdorff ´etale groupoids. Proceedings of the American Mathematical Society, 139(3):897 -- 907, 2011. [5] R. Exel and E. Pardo. The tight groupoid of an inverse semigroup. Semigroup Forum, 92(1):274 -- 303, 2016. [6] A. Paterson. Groupoids, Inverse Semigroups, and their Operator Algebras. Progress in Mathematics. Birkhauser Boston, 2012. [7] J. Renault. A Groupoid Approach to C-Algebras. Lecture Notes in Mathematics. Springer-Verlag, 1980. [8] J. Renault. Cartan subalgebras in C-algebras. Irish Math. Soc. Bull., 61:29 -- 63, 2008. [9] A. Sims. Hausdorff ´etale groupoids and their C-algebras. arXiv:1710.10897v1, 2017. Department of Mathematics, Faculty of Science and Technology, Keio Uni- versity 3141 Hiyoshi, Kohoku-ku, Yokohama, 2238522, Japan E-mail address: [email protected]
0905.0478	3	0905	2010-04-02T14:52:37	Leavitt path algebras with coefficients in a commutative ring	[ "math.OA", "math.RA" ]	Given a directed graph E we describe a method for constructing a Leavitt path algebra $L_R(E)$ whose coefficients are in a commutative unital ring R. We prove versions of the Graded Uniqueness Theorem and Cuntz-Krieger Uniqueness Theorem for these Leavitt path algebras, giving proofs that both generalize and simplify the classical results for Leavitt path algebras over fields. We also analyze the ideal structure of $L_R(E)$, and we prove that if $K$ is a field, then $L_K(E) \cong K \otimes_\Z L_\Z(E)$.	math.OA	math	LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING MARK TOMFORDE Abstract. Given a directed graph E we describe a method for con- structing a Leavitt path algebra LR(E) whose coefficients are in a com- mutative unital ring R. We prove versions of the Graded Unique- ness Theorem and Cuntz-Krieger Uniqueness Theorem for these Leav- itt path algebras, giving proofs that both generalize and simplify the classical results for Leavitt path algebras over fields. We also analyze the ideal structure of LR(E), and we prove that if K is a field, then LK (E) ∼= K ⊗Z LZ(E). 1. Introduction In [1] the authors introduced a class of algebras over fields, which they constructed from directed graphs and called Leavitt path algebras. (The definition in [1] was given for row-finite directed graphs, but the authors later extended the definition in [2] to all directed graphs.) These Leavitt path algebras generalize the Leavitt algebras L(1, n) of [21], and also contain many other interesting classes of algebras over fields. In addition, Leavitt path algebras are intimately related to graph C ∗-algebras (see [23]), and for any graph E the Leavitt path algebra LC(E) is ∗-isomorphic to a dense ∗-subalgebra of the graph C ∗-algebra C ∗(E) [26, Theorem 7.3]. In this paper we generalize the construction of Leavitt path algebras by replacing the field K with a commutative unital ring R. We use the no- tation LR(E) for our Leavitt path algebra, and prove that it is a Z-graded R-algebra with characteristic equal to the characteristic of R. We also prove versions of the Graded Uniqueness Theorem and the Cuntz-Krieger Unique- ness Theorem, which are fundamental to the study of Leavitt path algebras. The Graded Uniqueness Theorem for Leavitt path algebras over a field says that a graded homomorphism φ : LK (E) → A is injective if φ(v) 6= 0 for all v ∈ E0. For Leavitt path algebras over rings we need slightly different hypotheses: We prove that a graded homomorphism φ : LR(E) → A is injective if φ(rv) 6= 0 for all v ∈ E0 and for all r ∈ R \ {0}. Similarly, the Cuntz-Krieger Uniqueness Theorem for Leavitt path algebras over a field says that if every cycle in E has an exit, then a homomorphism φ : Date: May 4, 2009. 2000 Mathematics Subject Classification. 16W50, 46L55. Key words and phrases. graph algebras, rings, R-algebras, C ∗-algebras. The author was supported by NSA Grant H98230-09-1-0036. 1 2 MARK TOMFORDE LK(E) → A is injective if φ(v) 6= 0 for all v ∈ E0. Again, our hypotheses for Leavitt path algebras over rings are slightly different: We prove that if every cycle in E has an exit, then a homomorphism φ : LK(E) → A is injective if φ(rv) 6= 0 for all v ∈ E0 and for all r ∈ R \ {0}. Our proofs of the Uniqueness Theorems use techniques that are different from those that have been used in the proofs for Leavitt path algebras over fields. Consequently, this paper gives new proofs of each of the Uniqueness Theorems in the case that R = K is a field. One of the main points of this article is that our proofs of the Uniqueness Theorems are shorter than those in the existing literature. After proving our Uniqueness Theorems we continue by analyzing the ideal structure of LR(E). For ease and clarity as we analyze ideals, we re- strict our attention to the case when the graph E is row-finite. Because of the hypothesis φ(rv) 6= 0 for all v ∈ E0 and for all r ∈ R \ {0}, the Unique- ness Theorems only allow us to analyze what we call basic ideals: an ideal I of LR(E) is basic if rv ∈ I for r ∈ R \ {0} implies that v ∈ I. In analogy with Leavitt path algebras over fields, we prove in Theorem 7.9 that the map H 7→ IH is a lattice isomorphism from the saturated hereditary subsets of E onto the graded basic ideals of LR(E). We also prove in Theorem 7.17 that all basic ideals in LR(E) are graded if and only if E satisfies Condi- tion (K). Finally, in Theorem 7.20 and Proposition 7.22 we derive conditions for LR(E) to have no nontrivial proper basic ideals. These results are sim- ilar to the classification of gauge-invariant ideals of graph C ∗-algebras and Cuntz-Krieger C ∗-algebras, and we use similar techniques in this paper. We refer the reader to Remark 7.23 for references to the corresponding results for Cuntz-Krieger algebras, graph C ∗-algebras, and Leavitt path algebras over fields. In the final section, we discuss extending the coefficients of a Leavitt path algebra by tensoring with a commutative unital ring. In particular, we show that if K is a field, then LK(E) ∼= K ⊗Z LZ(E); and if K is a field of characteristic p, then LK(E) ∼= K ⊗Zp LZp(E). This allows us to relate properties of LZ(E) and LZp(E) to properties of LK(E). This paper is organized as follows: After some preliminaries in §2, we continue in §3 by constructing the Leavitt path algebra over a commutative until ring, and prove that LR(E) exists and has the appropriate universal property. In §4 we establish some basic properties of LR(E). In §5 we prove the Graded Uniqueness Theorem for LR(E), and in §6 we prove the Cuntz- Krieger Uniqueness Theorem for LR(E). In §7 we analyze the ideal structure of LR(E). Finally, in §8 we discuss extending the coefficients of a Leavitt path algebra by taking tensor products. We conclude with a discussion of the significance of the rings LZ(E) and LZn(E). LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 3 2. Preliminaries When we refer to a graph in this paper, we shall always mean a directed graph E := (E0, E1, r, s) consisting of a countable set of vertices E0, a countable set of edges E1, and maps r : E1 → E0 and s : E1 → E0 identifying the range and source of each edge. Definition 2.1. Let E := (E0, E1, r, s) be a graph. We say that a vertex v ∈ E0 is a sink if s−1(v) = ∅, and we say that a vertex v ∈ E0 is an infinite emitter if s−1(v) = ∞. A singular vertex is a vertex that is either a sink or an infinite emitter, and we denote the set of singular vertices by E0 sing, and refer to the elements of E0 reg as regular vertices; i.e., a vertex v ∈ E0 is a regular vertex if and only if 0 < s−1(v) < ∞. sing. We also let E0 reg := E0 \ E0 in E0 to be paths of length zero. We also let E∗ := S∞ Definition 2.2. If E is a graph, a path is a sequence α := e1e2 . . . en of edges with r(ei) = s(ei+1) for 1 ≤ i ≤ n−1. We say the path α has length α := n, and we let En denote the set of paths of length n. We consider the vertices n=0 En denote the paths of finite length, and we extend the maps r and s to E∗ as follows: For α := e1e2 . . . en ∈ En, we set r(α) = r(en) and s(α) = s(e1). A cycle in E is a path α ∈ E∗ \ E0 with s(α) = r(α). If α := e1 . . . en, then an exit for α is an edge f ∈ E1 such that s(f ) = s(ei) but f 6= ei for some 1 ≤ i ≤ n. We say that a graph E satisfies Condition (L) if every cycle in E contains an exit. Definition 2.3. We let (E1)∗ denote the set of formal symbols {e∗ : e ∈ E1}, and for α = e1 . . . en ∈ En we define α∗ := e∗ 1. We also define v∗ = v for all v ∈ E0. We call the elements of E1 real edges and the elements of (E1)∗ ghost edges. n−1 . . . e∗ ne∗ Definition 2.4. Let E be a directed graph and let R be a ring. A collection {v, e, e∗ : v ∈ E0, e ∈ E1} ⊆ R is a Leavitt E-family in R if {v : v ∈ E0} consists of pairwise orthogonal idempotents and the following conditions are satisfied: (1) s(e)e = er(e) = e for all e ∈ E1 (2) r(e)e∗ = e∗s(e) = e∗ for all e ∈ E1 (3) e∗f = δe,f r(e) for all e, f ∈ E1 (4) v = X{e∈E 1:s(e)=v} ee∗ whenever v ∈ E0 reg. Definition 2.5. Let E be a directed graph, and let K be a field. The Leavitt path algebra of E with coefficients in K, denoted LK(E), is the universal K-algebra generated by a Leavitt E-family (see Definition 2.4). Note that LK(E) is universal for Leavitt E-families in K-algebras; i.e., if A is a K-algebra and {av, be, ce∗ : v ∈ E0, e ∈ E1} is a Leavitt E-family in A, then there exists a K-algebra homomorphism φ : LK(E) → A such 4 MARK TOMFORDE that φ(v) = av, φ(e) = be, and φ(e∗) = ce∗ for all v ∈ E0 and e ∈ E1. It is shown in [1, §1] and [2, §1] that for any graph E the generators {v, e, e∗ : v ∈ E0, e ∈ E1} of LK(E) are all nonzero. In any algebra generated by a Leavitt E-family {v, e, e∗ : v ∈ E0, e ∈ E1}, we see that (2.1) (αβ∗)(γδ∗) =  αγ′δ∗ αδ∗ αβ′∗δ∗ 0 if γ = βγ′ if β = γ if β = γβ′ otherwise. 2.1. Algebras over commutative rings. If R is a commutative ring with unit 1, then an R-algebra is an abelian group A that has the structure of both a ring and a (left) R-module in such a way that (1) r · (xy) = (r · x)y = x(r · y) for all r ∈ R and x, y ∈ A; and (2) 1 · x = x for all x ∈ A. Note that as a ring, A is not necessarily commutative and A does not neces- sarily contain a unit. By a homomorphism between R-algebras we mean an R-linear ring homomorphism. If A and B are R-algebras, we let HomR(A, B) denote the collection of R-linear ring homomorphisms from A to B. We ob- serve that for any R-algebra A, the endomorphism ring HomR(A, A) is an R-algebra in the obvious way. If R is a commutative ring, the characteristic of R, denoted char(R), is defined to be the smallest positive integer n such that nr = 0 for all r ∈ R, if such an n exists, and 0 otherwise. It is a fact that if K is a field, then char K is either equal to 0 or a prime p. Any ring R may be viewed as a Z-algebra in the natural way, and if R has characteristic n, then R may also be viewed as a Zn-algebra. Furthermore, if A is an R-algebra and X ⊆ A, then we define spanR X :=( n Xi=1 rixi : ri ∈ R and xi ∈ X for all 1 ≤ i ≤ n) to be the R-submodule of A generated by the set X. 3. Constructing Leavitt path algebras with coefficients in a commutative ring with unit. In this section we wish to extend the definition of a Leavitt path algebra to allow for coefficients in an arbitrary commutative ring with unit. Definition 3.1. Let E be a directed graph, and let R be a commutative ring with unit. The Leavitt path algebra with coefficients in R, denoted LR(E), is the universal R-algebra generated by a Leavitt E-family (see Definition 2.4). Note that LR(E) is universal for Leavitt E-families in R-algebras; i.e., if A is a R-algebra and {av, be, ce∗ : v ∈ E0, e ∈ E1} is a Leavitt E-family in LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 5 A, then there exists a R-algebra homomorphism φ : LR(E) → A such that φ(v) = av, φ(e) = be, and φ(e∗) = ce∗ for all v ∈ E0 and e ∈ E1. Recall that any ring is a Z-algebra and any ring of characteristic n is a Zn-algebra. This motivates the following definitions. Definition 3.2. If E is a graph, the Leavitt path ring of characteristic 0 is the ring LZ(E), and for each n ∈ N the Leavitt path ring of characteristic n is the ring LZn(E). Remark 3.3. In the next proposition we show that the elements of {v, e, e∗ : v ∈ E0, e ∈ E1} are all nonzero, and that rv 6= 0 for all v ∈ E0 and all r ∈ R \ {0}. In Proposition 4.9, we are able to prove a stronger result: The set of paths E∗ in LR(E) is linearly independent over R, and the set of ghost paths {α∗ : α ∈ E∗} in LR(E) is linearly independent over R. The construction in the next proposition is an R-algebra version of a sim- ilar construction that has been done for graph C ∗-algebras (see [19, Theo- rem 1.2]) and for Leavitt path algebras over fields (see [14, Lemma 1.5]). Proposition 3.4. If E is a graph and R is a commutative ring with unit, then the Leavitt path algebra LR(E) has the property that the elements of the set {v, e, e∗ : v ∈ E0, e ∈ E1} are all nonzero. Moreover, LR(E) = spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β)} and rv 6= 0 for all v ∈ E0 and all r ∈ R \ {0}. Proof. The fact that e∗f = δe,f r(e) allows us to write any word in the generators {v, e, e∗ : v ∈ E0, e ∈ E1} as αβ∗ with α, β ∈ E∗. It follows that LR(E) = spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β)}. To see that the elements of the set {v, e, e∗ : v ∈ E0, e ∈ E1} ⊆ LR(E) are all nonzero, it suffices (due to the universal property) to construct an R-algebra generated by nonzero elements satisfying the relations described in Definition 3.1. Define Z := R ⊕ R ⊕ . . . to be the direct sum of countably many copies of R. For each e ∈ E1 let Ae := Z, and for each v ∈ E0 let Av := Ae Ms(e)=v Z ⊕ Ms(e)=v Z if 0 < s−1(v) < ∞ Ae if s−1(v) = ∞ if s−1(v) = 0.   Note that the Av's and Ae's are all mutually isomorphic since each is the direct sum of countably many copies of R. Let A := Lv∈E 0 Av. For each v ∈ E0 define Tv : Av → Av to be the identity map, and extend to a homomorphism Tv : A → A by defining Tv to be zero on A ⊖ Av. Also, for each e ∈ E1 choose an isomorphism Te : Ar(e) → Ae ⊆ As(e) and extend to a homomorphism Te : A → A by defining Te to be zero on A ⊖ Ae. Finally, we define Te∗ : A → A by taking the isomorphism T −1 : Ae ⊆ e 6 MARK TOMFORDE As(e) → Ar(e) and extending to obtain a homomorphism Te∗ : A → A by defining Te∗ to be zero on A ⊖ Ae. Let A be the subalgebra of HomR(A, A) generated by {Tv, Te, Te∗ : v ∈ E0, e ∈ E1}. One can check that {Tv, Te, Te∗ : v ∈ E0, e ∈ E1} is a collection of nonzero elements satisfying the relations described in Definition 3.1. Thus the subalgebra of HomR(A, A) generated by {Tv, Te, Te∗ : v ∈ E0, e ∈ E1} is the desired R-algebra. Finally, we note that for any v we have Av = R ⊕ M for some R-module M . Thus for any r ∈ R \ {0}, using the fact that R is unital we have rTv(1, 0) = Tv(r, 0) = (r, 0) 6= 0. Hence rTv 6= 0. The universal property of LR(E) then implies that rv 6= 0 for any v ∈ E0 and any r ∈ R \ {0}. (cid:3) Corollary 3.5. Let E be a graph and let R be a commutative ring with unit. Then char LR(E) = char R. Remark 3.6 (A realization of LR(E)). Suppose E is a graph and R is a commutative ring with unit. The path algebra of E with coefficients in R is the R-algebra generated by paths with the operation of path concatena- tion. (Here vertices are considered as paths of length zero.) In other words, n=0 En with the AR(E) is the free R-algebra generated by the paths E∗ =S∞ following relations: (i) vw = δv,wv for all v, w ∈ E0 (ii) e = er(e) = s(e)e for all e ∈ E1. If E = (E0, E1, r, s) is a graph, we let E be the graph with vertex set E0 := E0, edge set E1 := {e, e∗ : e ∈ E1}, and maps r and s extended to E1 by r(e∗) := s(e) and s(e∗) = r(e) for all e ∈ E1. We see that LR(E) may be realized as the quotient AR( E)/I, where AR( E) is the path algebra of E with coefficients in R, and I is the ideal of AR( E) generated by the elements (3.1) (cid:8)e∗f − δe,f r(e) : e, f ∈ E1(cid:9) ∪(cid:8)v − Xs(e)=v ee∗ : v ∈ E0 reg(cid:9). 4. Properties of Leavitt Path Algebras 4.1. Involution and selfadjoint ideals. As we have seen, any element k where αk, βk ∈ E∗ with k=1 rkαkβ∗ x ∈ LR(E) may be written x = PN r(αk) = r(βk) and rk ∈ R for 1 ≤ k ≤ N . Remark 4.1. If E is a graph, R is a commutative ring with unit, and LR(E) is the associated Leavitt path algebra, we may define an R-linear involution k=1 rkβkα∗ k. Note that this operation is R-linear, involutive ((x∗)∗ = x), and antimulti- plicative ((xy)∗ = y∗x∗). x 7→ x∗ on LR(E) as follows: If x =PN k, then x∗ =PN k=1 rkαkβ∗ Definition 4.2. If LR(E) is the Leavitt path algebra of a graph E with coefficients in R, an ideal I of LR(E) is selfadjoint if I ∗ = I. LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 7 4.2. Enough idempotents and local units. A ring R has enough idempo- tents if there exists a collection of pairwise orthogonal idempotents {eα}α∈Λ such that R = Lα∈Λ eαR = Lα∈Λ Reα. A set of local units for a ring R is a set Λ ⊆ R of commuting idempotents with the property that for any x ∈ R there exists t ∈ Λ such that tx = xt = x. If E is a graph, R is a commutative ring with unit, and LR(E) is the associated Leavitt path algebra, then LR(E) = Mv∈E 0 vLR(E) = Mv∈E 0 LR(E)v so LR(E) is a ring with enough idempotents. Furthermore, if E0 is finite, k=1 vk, then {tn}n∈N is a set of local units for LR(E). Definition 4.3. A ring R is idempotent if R2 = R; that is, if every x ∈ R not have a unit, but if we list the vertices of E as E0 = {v1, v2, . . .} and set then 1 = Pv∈E 0 v is a unit for LR(E). If E0 is infinite, then LR(E) does tn :=Pn can be written as x =Pn Remark 4.4. We see that if R is a ring with a set of local units, then R is idempotent: If x ∈ R, then there exists an idempotent t ∈ R with x = tx. Consequently, the Leavitt path algebra LR(E) is an idempotent ring. k=1 akbk for a1, . . . an, b1, . . . , bn ∈ R. 4.3. Z-graded rings. We show that all Leavitt path algebras have a natural Z-grading. Definition 4.5. If R is a ring, we say R is Z-graded if there is a a collection of additive subgroups {Rk}k∈Z of R with the following two properties: (1) R =Lk∈Z Rk (2) RjRj ⊆ Rj+k for all j, k ∈ Z. The subgroup Rk is called the homogeneous component of R of degree k. Definition 4.6. If R is a graded ring, then an ideal I of R is a Z-graded ideal if I = Lk∈Z(I ∩ Rk). If φ : R → S is a ring homomorphism between Z-graded rings, then φ is a graded ring homomorphism if φ(Rk) ⊆ Sk for all n ∈ Z. Note that the kernel of a Z-graded homomorphism is a Z-graded ideal. Also, if I is a Z-graded ideal in a Z-graded ring R, then the quotient R/I admits a natural Z-grading and the quotient map R → R/I is a Z-graded homomorphism. In this paper we will be concerned only with Z-gradings, and hence we will often omit the prefix Z and simply refer to rings, ideals, homomorphisms, etc. as graded. Proposition 4.7. If E is a graph and R is a commutative ring with unit, then we may define a Z-grading on the associated Leavitt path algebra LR(E) by setting LR(E)k :=( N Xi=1 riαiβ∗ i : αi, βi ∈ E∗, ri ∈ R, and αi − βi = k for all i) . 8 MARK TOMFORDE Proof. Let A be the free R-algebra generated by E0 ∪ E1 ∪ (E1)∗. Then A has a unique Z-grading for which the elements of E0, E1, and (E1)∗ have degrees 0, 1, and −1, respectively. Let I be the ideal in A generated by elements of the following type: • vw − δv,wv for v, w ∈ E0 • e − er(e) for e ∈ E1 • e − s(e)e for e ∈ E1 • e∗f − δe,f r(e) for e, f ∈ E1 • v −Ps(e)=v ee∗ for v ∈ E0 reg. Since the elements generating I are all homogeneous of degree zero, it follows that I is a graded ideal. Furthermore, we see that A/I ∼= LR(E), so that LR(E) is graded with the homogeneous elements of degree k equal to the set of R-linear combinations of elements of the form αβ∗ with α − β = k. (cid:3) Definition 4.8. If x ∈ LR(E), we say that x is a polynomial in real edges if i=1 riαi for ri ∈ R \ {0} and αi ∈ E∗. In this case we also define the x =Pn degree of x to be deg x = max{αi : 1 ≤ i ≤ n}. Note that deg x is independent of how x is written. Proposition 4.9. Let E be a graph and let R be a commutative ring with unit. The set of paths E∗ in LR(E) is linearly independent over R. Likewise, the set of ghost paths {α∗ : α ∈ E∗} in LR(E) is linearly independent over R. Proof. Suppose that α1, . . . , αn ∈ E∗, andPn i=1 riαi = 0 for some r1, . . . , rn ∈ R. Using the Z-grading on LR(E) we may, without loss of generality, assume that all the αi's have the same length. Then for any 1 ≤ j ≤ n we have rj(αj) = α∗ i=1 riαi) = 0. Proposition 3.4 implies that ri = 0. It follows that {α1, . . . , αn} is linearly independent over R. A similar argument works for ghost paths. (cid:3) j (Pn j αj = α∗ 4.4. Morita equivalence. Throughout this paper we will need to discuss Morita equivalence for rings that do not necessarily have an identity element. We recall the necessary definitions and results here. Definition 4.10. If R is a ring, we say that a left R-module M is unital if RM = M . We also say that M is nondegenerate if for all m ∈ M we have that Rm = 0 implies that m = 0. We let R-MOD denote the full subcategory of the category of all R-modules whose objects are unital nondegenerate R-modules. (Note that if R is unital, R-MOD is the usual category of R-modules.) When R and S are rings, and RMS is a bimodule, we say M is unital if RM = M and M S = M . Definition 4.11. Let R and S be idempotent rings. A (surjective) Morita context (R, S, M, N, ψ, φ) between R and S consists of unital bimodules RMS LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 9 and SNR, a surjective R-module homomorphism ψ : M ⊗S N → R, and a surjective S-module homomorphism φ : N ⊗R M → S satisfying φ(n ⊗ m)n′ = nψ(m ⊗ n′) and m′φ(n ⊗ m) = ψ(m′ ⊗ n)m for every m, m′ ∈ M and n, n′ ∈ N . We say that R and S are Morita equivalent in the case that there exists a Morita context. It is proven in [13, Proposition 2.5] and [13, Proposition 2.7] that R-MOD and S-MOD are equivalent categories if and only if there exists a Morita context (R, S, M, N, ψ, φ). In addition, the following result is obtained in [13]. Proposition 4.12. [13, Proposition 3.5] Let R and S be Morita equivalent idempotent rings, and let (R, S, M, N, ψ, φ) be a Morita context. If LR := {I ⊆ R : I is an ideal and RIR = I} and LS := {I ⊆ S : I is an ideal and SIS = I}, then there is a lattice isomorphism from LR onto LS given by I 7→ φ(N I, M ) with inverse given by I 7→ ψ(M I, N ). Remark 4.13. Note that when R is a ring with a set of local units, LR is the lattice of ideals of R. Thus if each of R and S is a ring with a set of local units, and if R and S are Morita equivalent, then the lattice of ideals of R is isomorphic to the lattice of ideals of S. Recall that in rings the property of being a ring ideal is not transitive; i.e., if R is a ring, I is an ideal of R, and J is an ideal of I, then it is not necessarily true that J is an ideal of R. Despite this fact, there is a special case when the implication does hold, and this will be of use to us. Lemma 4.14. Let R be a ring and let I be an ideal of R with the property that I has a set of local units. If J is an ideal of I, then J is an ideal of R. Proof. Let r ∈ R and x ∈ J. Since I has a set of local units, there exists t ∈ I with tx = x. Because I is an ideal, we have that rt ∈ I. Hence rx = r(tx) = (rt)x ∈ J. A similar argument shows that xr ∈ I. (cid:3) 5. The Graded Uniqueness Theorem Lemma 5.1. Let I be a graded ideal of LR(E). Then I is generated as an ideal by the set I0 := I ∩ LR(E)0. Proof. Let k > 0. Given x ∈ Ik := I ∩LR(E)k, we may write x =Pn i=1 αixi, where xi ∈ LR(E)0 and αi ∈ Ek for all 1 ≤ i ≤ n, and αi 6= αj for i 6= j. Then for any 1 ≤ j ≤ n we have xj = α∗ j n Xi=1 αixi! = α∗ j x ∈ I. 10 MARK TOMFORDE Thus xj ∈ I0 and Ik = LR(E)kI0. Similarly, I−k = I0LR(E)−k. Since I is a (cid:3) graded ideal, I =Lk∈Z Ik, and I is generated as an ideal by I0. Lemma 5.2. Let E be a graph, and let R be a commutative ring with unit. If x ∈ LR(E)0 and x 6= 0, then there exists α, β ∈ E∗ such that α∗xβ = rv for some v ∈ E0 and some r ∈ R \ {0}. Proof. Define GN := spanR{αβ∗ : α, β ∈ Em for 1 ≤ m ≤ N }. Then N =0 GN . We will prove by induction on N that if x ∈ GN and x 6= 0, then there exists α, β ∈ E∗ such that α∗xβ = rv for some v ∈ E0 i=1 rivi for vi ∈ E0 and nonzero ri ∈ R with vi 6= vj for i 6= j. If we let α = β = v1, then α∗xβ = r1v1. LR(E)0 = S∞ and some r ∈ R \ {0}. In the base case we have N = 0, and x = Pn In the inductive step, we assume that for all nonzero y ∈ GN −1 there exists α′, β′ ∈ E∗ such that (α′)∗yβ′ = rv for some v ∈ E0 and some r ∈ R \ {0}. Suppose that x ∈ GN and x 6= 0. Then we can write x = M Xi=1 riαiβ∗ i + sjvj, P Xj=1 for α, β ∈ E∗ with αi = βi ≥ 1, vj ∈ E0 with vj 6= vj ′ for j 6= j′, and ri, sj ∈ R \ {0}. If any vj is a sink, we may let α = β = vj, and then α∗xβ = sjvj. If any vj is an infinite emitter, then we may choose an edge e ∈ E1 with s(e) = vj and e not equal to any edge appearing in any of the αi's. If we let α = β = e, then α∗xβ = e∗sjvje = sjr(e). The only other case to consider is when every vj is a regular vertex (i.e., neither a sink nor ee∗ to write x as a linear combination of elements γδ∗ where γ, δ ∈ E∗ with γ = δ ≥ 1. By regrouping the elements in this linear combination, we may write an infinite emitter). In this case we may use the relation vj = Ps(e)=vj P Q x = Xi=1 Xj=1 eixi,jf ∗ j j 6= 0 for all i, j. Since e1x1,1f ∗ 6= ei′ for i 6= i′ and fj 6= fj ′ for j 6= j′; and where ei, fi ∈ E1 with ei xi,j ∈ GN −1 with eixi,jf ∗ 1 6= 0, it follows that r(e1)x1,1r(f1) 6= 0. Because r(e1)x1,1r(f1) 6= 0 and r(e1)x1,1r(f1) ∈ GN −1, the inductive hypothesis implies that there exists α′, β′ ∈ E∗ such that (α′)∗r(e1)x1,1r(f1)β′ = rv for some v ∈ E0 and some r ∈ R \ {0}. If we let α := e1α′ and β := f1β′, then α∗xβ = (α′)∗e∗ 1xf1β′ = (α′)∗e∗ 1 f1β′ = (α′)∗r(e1)x1,1r(f1)β′ = rv. The Principle of Mathematical Induction shows that the claim holds for all N , and hence the lemma holds for all nonzero x in LR(E)0. (cid:3) 1e1x1,1f ∗ Theorem 5.3 (Graded Uniqueness Theorem). Let E be a graph, and let R be a commutative ring with unit. If S is a graded ring and φ : LR(E) → S is LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 11 a graded ring homomorphism with the property that φ(rv) 6= 0 for all v ∈ E0 and for all r ∈ R \ {0}, then φ is injective. Proof. Suppose that x ∈ LR(E)0 ∩ ker φ. If x is nonzero, then by Lemma 5.2 there exists α, β ∈ E∗ such that α∗xβ = rv for some v ∈ E0 and some r ∈ R \ {0}. But then φ(rv) = φ(α∗xβ) = φ(α∗)φ(x)φ(β) = 0, which is a contradiction. Hence x = 0, and LR(E)0 ∩ ker φ = {0}. Since φ is a graded ring homomorphism, ker φ is a graded ideal of LR(E). It follows from Lemma 5.1 that ker φ is generated as an ideal by LR(E)0 ∩ ker φ = {0}. Thus ker φ = {0}, and φ is injective. (cid:3) Corollary 5.4. Let E be a graph, and let K be a field. If S is a graded ring and φ : LK(E) → S is a graded ring homomorphism with the property that φ(v) 6= 0 for all v ∈ E0, then φ is injective. Remark 5.5. The Graded Uniqueness Theorem for Leavitt path algebras may be thought of as an analogue of the Gauge-Invariant Uniqueness The- orem for graph C ∗-algebras, with the grading playing the role of the gauge action. In [8] Cuntz and Krieger showed that if A is a finite {0, 1}-matrix satisfy- ing Condition (I), then there is a unique C ∗-algebra generated by a nonzero Cuntz-Krieger A-family, which they denote by OA. Universal Cuntz-Krieger algebras of finite {0, 1}-matrices were introduced in [15], and a Gauge- Invariant Uniqueness Theorem for these algebras was proven in [15, The- orem 2.3]. A Gauge-Invariant Uniqueness Theorem for C ∗-algebras of row- finite graphs was obtained in [5, Theorem 2.1], and this was extended to C ∗-algebras of non-row-finite graphs in [6, Theorem 2.1]. Furthermore, the Gauge-Invariant Uniqueness Theorem was generalized to Cuntz-Krieger al- gebras of infinite matrices in [24, Theorem 2.7] and to Cuntz-Pimnser alge- bras in [11, Theorem 4.1] and [17, Theorem 6.4]. In the Leavitt path algebra setting, the gauge action is replaced by a Z- grading -- in fact, if one views the Leavitt path algebra LC(E) as a dense ∗-subalgebra of the graph C ∗-algebra C ∗(E), then the gauge action on C ∗(E) induces a Z-grading on LC(E) (see the proof of [26, Theorem 7.3] for details). In [3, Theorem 5.1], Ara, Moreno, and Pardo proved the Graded Uniqueness Theorem for LK(E), where K is a field and E is a row-finite graph. A proof of the Graded Uniqueness Theorem for LK(E), where K is a field and E is an arbitrary graph, was given by the author in [26, Theorem 4.8]. The proof in Theorem 5.3 uses different techniques than [3] or [26]. 6. The Cuntz-Krieger Uniqueness Theorem Recall that a graph E is said to satisfy Condition (L) if every cycle in E has an exit. (See Definition 2.2 for more details.) Lemma 6.1. Suppose E is a graph satisfying Condition (L). If F is a finite subset of E∗ \ E0 and v ∈ E0, then there exists a path α ∈ E∗ such that s(α) = v and for every µ ∈ F we have α∗µα = 0. 12 MARK TOMFORDE Proof. Given v ∈ E0 and a finite subset F ⊆ E∗, consider two cases. Case I: There is a path from v to a sink in E. In this case, let α be a path with s(α) = v and r(α) a sink. For any µ ∈ F , we see that α∗µα is nonzero if and only if there exists ν ∈ E∗ \ E0 such that µα = αν, which is impossible since r(α) is a sink. Thus α∗µα = 0. Case II: There is no path from v to a sink in E. Let M = max{µ : µ ∈ F } + 1. If there is a path α = α1 . . . αM ∈ EM with s(α) = v and no repeated vertices, then for any µ ∈ F we see that α∗µα is nonzero if and only if there exists ν ∈ E∗ \ E0 such that µα = αν, which is impossible since this would imply that s(α1) = s(αj) for some j ≥ 2 contradicting that α has no repeated vertices. Thus α∗µα = 0. Otherwise, every path EM with s(λ) = v has repeated vertices, and there exists a path from v to the base point of a cycle in E. Choose a path τ of minimal length such that s(τ ) = v and r(τ ) is the base point of a cycle. Choose a cycle β of minimal length based at r(τ ). Let f be an exit for β, and let β′ be the segment of β from r(τ ) to s(f ). By the minimality of τ , the edge f is not equal to any of the edges in the path τ . Likewise, by the minimality of β, the edge f is not equal to any of the edges on the cycle β or the path β′. Thus the path α := τ ββ . . . ββ′f has the property that f is not equal to any edge αi for 1 ≤ i ≤ α − 1. By choosing sufficiently many repetitions of the cycle β we can ensure that α has length greater than or equal to M (to avoid the possibility that α ∈ F ). Then we have that α∗µα is nonzero if and only if there exists ν ∈ E∗ \ E0 such that µα = αν, which is impossible since this would imply that f = αj for some 1 ≤ j ≤ α − 1. Thus α∗µα = 0. (cid:3) Lemma 6.2. Let E be a graph satisfying Condition (L), and let R be a commutative ring with unit. If x ∈ LR(E) is a polynomial in only real edges and x 6= 0, then there exist paths α, β ∈ E∗ such that α∗xβ = rv for some v ∈ E0 and some r ∈ R \ {0}. Proof. We will prove by induction on N that if x ∈ LR(E) is a nonzero polynomial in only real edges with deg x = N , then there exist paths α, β ∈ E∗ such that α∗xβ = rv for some v ∈ E0 and some r ∈ R \ {0}. In the i=1 rivi for vi ∈ E0 and nonzero base case we have deg x = 0, so that x =PM ri ∈ R with vi 6= vj for i 6= j. If we let α = β = v1, then α∗xβ = r1v1. In the inductive step, we assume that our claim holds for all nonzero polynomials in real edges with degree N − 1 or less. Suppose x ∈ LR(E) is a nonzero polynomial in real edges with deg x = N . If x has no terms of degree 0, then we may write M Xi=1 with each xi a nonzero polynomial in real edges of degree N − 1 or less, and ei ∈ E1 with ei 6= ej for i 6= j. Then e∗ 1x = x1 is a nonzero polynomial of degree N − 1 or less, so by the inductive hypothesis there exists α′, β ∈ E∗ x = eixi LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 13 such that (α′)∗x1β = rv for some v ∈ E0 and r ∈ R\{0}. If we let α := e1α′, 1xβ = (α′)∗x1β = rv and the claim holds. On the other then α∗xβ = (α′)∗e∗ hand, if x does have a term of degree 0, then we may write M K x = riαi + sjvj Xi=1 Xj=1 where the αi's are paths of length 1 or greater, each ri, sj ∈ R \ {0}, and the vj's are vertices with vj 6= vj ′ for j 6= j′. Let F := {αi : 1 ≤ i ≤ M }. By Lemma 6.1 there exists α ∈ E∗ such that s(α) = v1 and for every αi we have α∗αiα = 0. If we let β = α, then we have M K α∗xβ = riα∗αiα + sjα∗vjα = s1α∗v1α = s1r(α). Xi=1 Xj=1 By the Principle of Mathematical Induction, we may conclude that the claim holds for all N , and hence the lemma holds for all nonzero polynomials in only real edges. (cid:3) Lemma 6.3. Let E be a graph and let R be a commutative ring with unit. Let x ∈ LR(E) and suppose that x is a polynomial in real edges with x 6= 0. If there exists v ∈ E0 with xv = x, then for any e ∈ E1 with s(e) = v we have xe 6= 0. Proof. Since LR(E) is graded with LR(E) = Lk∈Z LR(E)k, it suffices to case we may write x = PM prove the claim when x is homogeneous of degree k for some k ≥ 0. In this i=1 riαi with each ri ∈ R \ {0} and each αi ∈ Ek with αi 6= αi′ for i 6= i′. Since xv = x, we may also assume that r(αi) = v for all i. For any e ∈ E1 with s(e) = v we see that αie ∈ Ek+1. If xe = 0, then r1r(e) = e∗α∗ 1(r1α1e) = e∗α∗ 1 M Xi=1 riαi! e = e∗α∗ 1(xe) = 0, which contradicts Proposition 3.4. Hence xe 6= 0. (cid:3) Lemma 6.4. Let E be a graph and let R be a commutative ring with unit. If x ∈ LR(E) and x 6= 0 then there exists γ ∈ E∗ such that xγ 6= 0 and xγ is a polynomial in only real edges. Proof. Define riαiβ∗ i : ri ∈ R, αi, βi ∈ E∗, and βi ≤ N for 1 ≤ i ≤ M) . AN :=( M Xi=1 Then LR(E) =S∞ in only real edges. In the base case we have N = 0, and x = PM N =0 AN . We will prove by induction on N that if x ∈ AN and x 6= 0, then there exists γ ∈ E∗ such that xγ 6= 0 and xγ is a polynomial i=1 riαi so that x is a polynomial in real edges. Choose v ∈ E0 such that xv 6= 0. Then xv is a polynomial in only real edges, and the claim holds. 14 MARK TOMFORDE In the inductive step, we assume that for all nonzero x′ ∈ AN −1, there exists γ ∈ E∗ such that x′γ 6= 0 and x′γ is a polynomial in only real edges. i ∈ AN , we may choose v ∈ E0 such that i=1 riαiβ∗ Given an element x =PM xv 6= 0. By regrouping terms, we may write xv = P Xj=1 xje∗ j + y where the xj's are polynomials in which each term has N − 1 ghost edges or fewer (so that xj ∈ AN −1), each ej ∈ E1 with s(ej) = v and ej 6= ej ′ for j 6= j′, and y is a polynomial in only real edges with yv = y. If y = 0, then xve1 = x1 6= 0 and by the inductive hypothesis there exists γ′ such that x1γ′ is a nonzero polynomial in only real edges. If γ := e1γ′, then xγ = xve1γ′ = x1γ′ is a nonzero polynomial in only real edges. If y 6= 0, then we consider three possibilities for v. If v is a regular vertex, then v =Ps(e)=v ee∗ and xv =PP j=1 xje∗ j +Ps(e)=v yee∗ and by regrouping we are as in the situation described in the previous paragraph, so we may argue as done there. If v is a sink, then there are no edges whose source is v, so xv = y and we may choose γ := v, and the claim holds. If v is an infinite emitter, then we may choose e ∈ E1 with s(e) = v and e 6= ej for all j e + ye = ye. Since y is a nonzero polynomial in only real edges with yv = y, it follows from Lemma 6.3 that ye is a nonzero polynomial in only real edges. By the Principle of Mathematical Induction, we may conclude that the claim holds for all N , and hence the lemma holds for all nonzero x ∈ LR(E). (cid:3) 1 ≤ j ≤ P . If we let γ := e, then xγ = xe = xve = PP j=1 xje∗ Theorem 6.5 (Cuntz-Krieger Uniqueness Theorem). Let E be a graph sat- isfying Condition (L), and let R be a commutative ring with unit. If S is a ring and φ : LR(E) → S is a ring homomorphism with the property that φ(rv) 6= 0 for all v ∈ E0 and for all r ∈ R \ {0}, then φ is injective. Proof. Suppose x ∈ ker φ and x 6= 0. By Lemma 6.4 there exists γ ∈ E∗ such that xγ is a nonzero polynomial in all real edges. Consequently, Lemma 6.2 implies that there exists α, β ∈ E∗ such that α∗xγβ = rv for some v ∈ E0 and some r ∈ R \ {0}. Then φ(rv) = φ(α∗)φ(x)φ(γβ) = 0, which is a contradiction. Hence ker φ = {0}, and φ is injective. (cid:3) Corollary 6.6. Let E be a graph satisfying Condition (L), and let K be a field. If S is a ring and φ : LK(E) → S is a ring homomorphism with the property that φ(v) 6= 0 for all v ∈ E0, then φ is injective. Remark 6.7. The Cuntz-Krieger Uniqueness Theorem has a long history in the C ∗-algebra setting, and the Cuntz-Krieger Uniqueness Theorem for graph C ∗-algebras may be viewed as a vast generalization of Coburn's The- orem [22, Theorem 3.5.18]. LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 15 The first Cuntz-Krieger Uniqueness Theorem was proven by Cuntz and Krieger in [8, Theorem 2.13], where they showed that if A is a finite {0, 1}- matrix satisfying Condition (L) then any two Cuntz-Krieger A-families com- posed of nonzero partial isometries generate isomorphic C ∗-algebras. This was generalized to C ∗-algebras of locally finite graphs in [19, Theorem 3.7] using groupoid techniques, and [19] is also where Condition (L) was first introduced. In [5, Theorem 3.1] a Cuntz-Krieger Theorem was proven for C ∗-algebras of row-finite graphs, and the proof avoided groupoid methods in favor of a direct analysis of the AF-core. A Cuntz-Krieger Uniqueness Theorem for C ∗-algebras of non-row-finite graphs was obtained in [24, The- orem 1.5] by realizing the graph C ∗-algebra as an increasing union of C ∗- algebras of finite graphs. When Cuntz and Krieger's original theorem [8, Theorem 2.13] is translated into a theorem about graphs, one obtains the Cuntz-Krieger Uniqueness Theorem for C ∗-algebras of finite graphs with no sinks and, moreover, Condition (I) is equivalent to Conidition (L) for these graphs. Additionally, the Cuntz-Krieger Uniqueness Theorem was extended to Cuntz-Krieger algebras of infinite matrices [10, Theorem 13.1]. Although there is a Cuntz-Krieger Uniqueness Theorem for Leavitt path algebras, this result is independent of the graph C ∗-algebra result -- neither the Cuntz-Krieger Uniqueness Theorem for Leavitt path algebras nor the Cuntz-Krieger Uniqueness Theorem for graph C ∗-algebras may be used to obtain the other. In [1, Corollary 3.3], Abrams and Aranda-Pino first proved a weak version of the Cuntz-Krieger Uniqueness Theorem for LK(E), where K is a field and E is a row-finite graph. Later, the author proved a lemma (see [26, Lemma 6.5]) that, with [1, Corollary 3.3], gives a full Cuntz-Krieger Uniqueness Theorem for LK(E) when E is a row-finite graph. A proof of the Cuntz-Krieger Uniqueness Theorem for LK (E), where K is a field and E is an arbitrary graph, was given by the author in [26, Theorem 6.8]. The proof in [26] uses the process of desingularization [26, Lemma 6.7] to show that the Cuntz-Krieger Uniqueness Theorem in the row-finite case implies the Cuntz-Krieger Uniqueness Theorem for arbitrary graphs. The proof in Theorem 6.5 uses different techniques than [1] or [26], and does not require one to consider the row-finite case first. 7. Ideals in Leavitt path algebras In this section we analyze ideals in LR(E). We will see that for ideals of LR(E) we will not only be concerned with which vertices are in the ideal, but also which multiples of the vertices are in the ideal. To motivate the results in this section, we start with an example. Example 7.1. Let E be the graph with two vertices and no edges, and let R = Z. Then LZ(E) ∼= Z ⊕ Z. If we consider the ideals of LZ(E), we see that they are of the form nZ ⊕ mZ for n, m ∈ {0, 1, 2, . . . , ∞}. We would like to consider the ideals that are reflected in the structure of the graph -- in particular, those ideals that are generated by vertices of the 16 MARK TOMFORDE graph. However, if we list the vertices of E as E0 = {v, w}, then there are four subsets of vertices, ∅, {v}, {w}, {v, w}, and the ideals generated by these sets are 0, Z⊕0, 0⊕Z, Z⊕Z. These are the only ideals generated by subsets of vertices, and each of them has the property that if a nonzero multiple of a vertex in in the ideal, then that vertex is in the ideal. Consequently, it is only these kind of ideals that will be determined by subsets of vertices in the graph. This motivates the following definition. Definition 7.2. Let R be a commutative ring with unit, and let E be a graph. If I is an ideal in LR(E), we say that I is basic if whenever r ∈ R \ {0} and v ∈ E0, we have rv ∈ I implies v ∈ I. Remark 7.3. Observe that if K is a field, then every ideal in LK(E) is basic. In this section we show that saturated hereditary subsets of vertices cor- respond to graded basic ideals. Throughout this section we restrict our attention to the case of row-finite graphs in order to avoid many of the complications that arise in the non-row-finite case. Our hope is that this will make our investigations easier for the reader to follow. Despite this, most of the results in this section do generalize to the non-row-finite setting, provided one uses admissible pairs in place of saturated hereditary subsets. Definition 7.4. Let E be a graph. A subset H ⊆ E0 is hereditary if for all e ∈ E0 and s(e) ∈ H imply that r(e) ∈ H. A hereditary subset H is saturated if whenever v ∈ E0 reg then r(s−1(v)) ⊆ H implies that v ∈ H. For any hereditary set X, we define the saturation X to be the smallest saturated hereditary subset of E0 containing X. Observe that intersections of saturated hereditary subsets are saturated hereditary. Also, unions of saturated hereditary subsets are hereditary, but not necessarily saturated. In any R-algebra A, the ideals of A are partially ordered by inclusion and form a lattice under the operations I ∧ J := I ∩ J and I ∨ J := I + J. (Note that I + J is the smallest ideal containing I ∪ J.) This lattice has a maximum element A and a minimum element {0}. Likewise, for any graph E = (E0, E1, r, s), the saturated hereditary sub- sets of E0 are partially ordered by inclusion and form a lattice under the operations H1 ∧ H2 := H1 ∩ H2 and H1 ∨ H2 := H1 ∪ H2. This lattice has a maximum element E0 and a minimum element ∅. Definition 7.5. Let E = (E0, E1, r, s) be a graph and H ⊆ E0 be a saturated hereditary subset. We define (E\H) to be the graph with (E\H)0 := E0\H, (E \ H)1 := E1 \ r−1(H), and r(E\H) and s(E\H) are obtained by restricting r and s to (E \ H)1. We also define EH to be the graph with E0 H := H, E1 H := s−1(H), and rEH and sEH are obtained by restricting r and s to E1 H. Lemma 7.6. Let E be a graph, and let R be a commutative ring with unit. If I is an ideal of LR(E), then the set HI := {v : v ∈ I} is a saturated hereditary subset. LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 17 Proof. If e ∈ E1 and s(e) ∈ H, then s(e) ∈ I so r(e) = e∗e = e∗s(e)e ∈ I and r(e) ∈ H. Thus H is hereditary. If v ∈ E0 reg and r(s−1(v)) ⊆ H, then for each e ∈ s−1(v) we have r(e) ∈ H and r(e) ∈ I so ee∗ = er(e)e∗ ∈ I. Thus v = Ps(e)=v ee∗ ∈ I, and v ∈ H. Hence H is saturated. (cid:3) Proposition 7.7. Let E be a graph, and let R be a commutative ring with unit. If H is a saturated hereditary subset of E0, and IH is the two-sided ideal in LR(E) generated by {v : v ∈ H}, then IH = spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β) ∈ H}, IH is a graded basic ideal, and {v ∈ E0 : v ∈ IH} = H. Moreover, IH is a selfadjoint ideal that is also an idempotent ring. Proof. We first observe that the multiplication rules of (2.1) imply that spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β) ∈ H} is a two-sided ideal containing H. It follows that IH ⊆ spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β) ∈ H}. Furthermore, if v ∈ H, then for any α, β ∈ E∗ with r(α) = r(β) = v, the element αvβ∗ = αβ∗ is in any ideal containing v. Hence IH = spanR{αβ∗ : α, β ∈ E∗ and r(α) = r(β) ∈ H}. To see that IH is graded it suffices to notice that αβ∗ is homogeneous of degree α − β. In addition, we see IH is selfadjoint because (αβ∗) = βα∗. Next we show that IH is a basic ideal. Let v ∈ E0 and suppose that rv ∈ IH for some r ∈ R \ {0}. Let E \ H be the graph of Definition 7.5. Then the vertices, edges, and ghost edges of E \ H, which generate LR(E \ H), may be extended to a Leavitt E-family by simply defining elements to be zero if v ∈ H or r(e) ∈ H. By the universal property of LR(E), we obtain an R-algebra homomorphism φ : LR(E) → LR(E \ H) with φ(v) =(v 0 if v ∈ E0 \ H if v ∈ H φ(e) =(e if r(e) ∈ E0 \ H 0 if r(e) ∈ H and φ(e∗) =(e∗ 0 if r(e) ∈ E0 \ H if r(e) ∈ H. Thus ker φ is a two-sided ideal of LR(E) containing H, and it follows that IH ⊆ ker φ. Hence rφ(v) = φ(rv) = 0, and since v is a vertex in E0, either φ(v) = v or φ(v) = 0. But Proposition 3.4 implies that in LR(E \ H) we have rv 6= 0 for all v ∈ (E \ H)0 and all r ∈ R \ {0}. Thus φ(v) = 0 and v ∈ H. Hence v ∈ IH, and IH is a basic ideal. We next show that the set {v ∈ E0 : v ∈ IH} is precisely H. To begin, we trivially have H ⊆ {v ∈ E0 : v ∈ IH}. For the reverse inclusion we use the fact that IH ⊆ ker φ to conclude that v /∈ H implies that φ(v) 6= 0 so that v /∈ ker φ and v /∈ IH . Hence {v ∈ E0 : v ∈ IH} = H. Finally we show that IH is an idempotent ring. Any x ∈ IH has the i with r(αi) = r(βi) ∈ H. For each i, define vi := i=1 riαiβ∗ form x = PN 18 MARK TOMFORDE r(αi) = r(βi). Then riαiβ∗ viβ∗ for a1, . . . , aN , b1, . . . , bN ∈ IH. Thus IH is an idempotent ring. i ), and since riαivi ∈ IH and i ∈ IH, we see that any x ∈ IH may be written as x = a1b1 + . . . + aN bN (cid:3) i = (riαivi)(viβ∗ Lemma 7.8. Let E be a graph, and let R be a commutative ring with unit. If X is a hereditary subset of E0, and IX is the two-sided ideal in LR(E) generated by {v : v ∈ X}, then IX = IX. In particular, IX is a graded basic ideal that is also an idempotent ring. Proof. Since X ⊆ X, we have IX ⊆ IX . Conversely, if we let H := {v ∈ E0 : v ∈ IX}, then it follows from Lemma 7.6 that H is a saturated hereditary subset containing X. Thus X ⊆ H, and v ∈ X implies v ∈ IX . Hence IX ⊆ IX. (cid:3) Theorem 7.9. Let E be a graph, and let R be a commutative ring with unit. Using the notation of Definition 7.5 and Proposition 7.7, we have the following: (1) The map H 7→ IH is a lattice isomorphism from the lattice of satu- rated hereditary subsets of E0 onto the lattice of graded basic ideals of LR(E). In particular, the graded basic ideals of LR(E) form a lattice with IH1 ∧ IH2 = IH1∩H2 and IH1 ∨ IH2 = IH1∪H2 for any saturated hereditary subsets H1 and H2. (2) For any saturated hereditary subset H we have that LR(E)/IH is canonically isomorphic to LR(E \ H). (3) For any hereditary subset X the ideal IX and the Leavitt path algebra LR(EX ) are Morita equivalent as rings. Proof. We shall first prove (2), then (1), and then (3). Proof of (2): We shall show that LR(E)/IH ∼= LR(E \ H). Let {v : v ∈ E0}∪{e, e∗ :∈ E1} be the generators for LR(E). Then {v +IH : v ∈ E \H}∪ e + IH, e∗ + IH : r(e) /∈ H} is a collection of elements satisfying the Leav- itt path algebra relations for EH and generating LR(E)/IH . Hence there exists a surjective R-algebra homomorphism φ : LR(EH ) → LR(E)/IH . Proposition 7.7 shows that IH is a graded ideal, and hence φ is a graded homomorphism. Furthermore, if v ∈ E0 H, then v /∈ H and the previous paragraph implies that v /∈ IH. Since Proposition 7.7 shows that IH is a basic ideal, for all v ∈ E0 H and all r ∈ R \ {0}, we have φ(rv) = rv + IH 6= 0. It follows from the Graded Uniqueness Theorem 5.3 that φ is injective. Thus φ is an isomorphism and LR(E)/IH Proof of (1): We shall show that H 7→ IH is a lattice isomorphism. To see that this map is surjective, let I be a graded basic ideal in LR(E), and set H := {v ∈ E0 : v ∈ I}. Since IH ⊆ I, we see that IH and I contain the same v's. Therefore, just as in the proof of Part (2), we see that LR(E)/IH ∼= LR(E \ H). LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 19 and LR(E)/I are generated by nonzero elements satisfying the Leavitt path algebra relations for E \ H. Since both IH and I are graded, both quotients are graded, and the quotient map π : LR(E)/IH → LR(E)/I is a graded homomorphism. Furthermore, since I and IH contain the same v's, and since I is a basic ideal, it follows that if v ∈ E0 \ H, then v /∈ IH and rv /∈ I for all r ∈ R \ {0}. Thus the Graded Uniqueness Theorem implies that the quotient map π : LR(E \ H) ∼= LR(E)/IH → LR(E)/I is injective. Hence I = IH. The fact that H 7→ IH is injective follows immediately from the fact that {v ∈ E0 : v ∈ IH} is precisely H, which was obtained in Proposition 7.7. Thus the correspondence H 7→ IH is bijective. Since H 7→ IH is a bijection that preserves inclusions, the map H 7→ IH is a poset isomorphism and hence automatically a lattice isomorphism Proof of (3): To see that IX is Morita equivalent to LR(EX ), list the elements of X = {v1, v2, . . .}, let Λ :=({1, 2, . . . , X} {1, 2, . . .} if X is finite if X is infinite, and let en :=Pn i=1 vi for n ∈ Λ. If we consider the elements {v : v ∈ H} and {e, e∗ : e ∈ E1 and s(e) ∈ H} in LR(E), we see that they are a Leavitt EX -family and thus there exists a homomorphism π : LR(EX ) → LR(E) taking the generators of LR(EX ) to these elements. Since this homomorphism is graded, Theorem 5.3 shows that π is injective. Hence we may identify LR(EX ) with the subalgebra spanR{αβ∗ : α, β ∈ E∗ X and r(α) = r(β) ∈ X} n=1 enLR(E)en. enLR(E), ψ, φ! In addition, of LR(E). With this identification, we see that LR(EX ) =P∞ Moreover, Lemma 7.7 shows that IX =P∞ Xn∈Λ LR(E)enLR(E),Xn∈Λ enLR(E)en,Xn∈Λ LR(E)en,Xn∈Λ n=1 LR(E)enLR(E). with ψ(m ⊗ n) = mn and φ(n ⊗ m) = nm is a (surjective) Morita con- text for the idempotent rings LR(EX) and IX . It then follows from [13, Proposition 2.5] and [13, Proposition 2.7] that LR(EX ) and IX are Morita equivalent. Corollary 7.10. Let E be a graph, and let R be a commutative ring with unit. Then every graded basic ideal of LR(E) is selfadjoint. Using the Cuntz-Krieger Uniqueness Theorem we can characterize those graphs whose associated Leavitt path algebras have the property that every basic ideal is a graded ideal. (cid:3) 20 MARK TOMFORDE Definition 7.11. We say that a closed path α = e1 . . . en ∈ En is simple if s(ei) 6= s(e1) for i = 2, 3, . . . , n. Definition 7.12. A graph E satisfies Condition (K) if every vertex in E0 is either the base of no closed path or the base of at least two simple closed paths. The following proposition is well known. It has been proven in [25, Propo- sition 1.17] and [4, Theorem 4.5(2),(3)]. Proposition 7.13. If E is a row-finite graph, then E satisfies Condition (K) if and only if for every saturated hereditary subset H, the graph E \ H of Definition 7.5 satisfies Condition (L). Lemma 7.14. If E is the graph consisting of a single simple closed path of length n; i.e., E0 = {v1, . . . , vn} E1 = {e1, . . . en} s(ei) = vi for 1 ≤ i ≤ n r(ei) = vi+1 for 1 ≤ i < n and r(en) = v1, an d R is a commutative ring with unit, then LR(E) ∼= Mn(R[x, x−1]). The proof of Lemma 7.14 is the same as the proof of [26, Lemma 6.12]. Lemma 7.15. Let R be a commutative ring with unit, let E be a row-finite graph, and let H be a saturated hereditary subset of E. Then the ideal IH in LR(E) is a ring with a set of local units. The proof of Lemma 7.15 is the same as the proof of [26, Lemma 6.14]. Lemma 7.16. Let R be a commutative ring with unit, and let E be a row- finite graph that contains a simple closed path with no exit. Then LR(E) contains an ideal that is basic but not graded. Proof. Let α := e1 . . . en be a simple closed path with no exits in E. If we let X := {s(ei)}n i=1, then since α has no exits, X is a hereditary subset of E0. By Theorem 7.9(3) LR(EX ) is Morita equivalent to the ideal IX in LR(E). However, EX is the graph which consists of a single closed path, and thus LR(EX ) ∼= Mn(R[x, x−1]) by Lemma 7.14. Theorem 7.9(1) implies that LR(E) ∼= Mn(R[x, x−1]) has no proper nontrivial graded ideals. Let I := hx + 1i be the ideal in R[x, x−1] generated by x + 1. Then any element of I has the form p(x)(x + 1) for some p(x) ∈ R[x, x−1] and hence has −1 as a root. It follows that for every r ∈ R \ {0} we have that r1 /∈ I. Since v = 1 in R[x, x−1], it follows that rv /∈ I for all r ∈ R \{0}. Thus I is a basic ideal. It follows that Mn(I) is a proper nontrivial ideal of Mn(R[x, x−1]), which is basic but not graded. Because the Morita context described in the proof of Theorem 7.9(3) gives a lattice isomorphism from ideals of LR(EX ) to ideals of IX that preserves the grading, we may conclude that IX contains an ideal that is basic but not graded. Since IX has a set of local units by Lemma 7.15, it follows from Lemma 4.14 that ideals of IX are ideals of LR(E). Hence LR(E) contains an ideal that is basic but not graded. (cid:3) LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 21 These results together with the Cuntz-Krieger Uniqueness Theorem give us the following theorem. Theorem 7.17. Let R be a commutative ring with unit. If E is a row- finite graph, then E satisfies Condition (K) if and only if every basic ideal in LR(E) is graded. Proof. Suppose that E satisfies Condition (K). If I is a basic ideal of LR(E), let H := {v : v ∈ I}. Then IH ⊆ I, and we have a canonical surjection q : LR(E)/IH → LR(E)/I. By Theorem 7.9(2) there exists a canonical isomorphism φ : LR(E \ H) → LR(E)/IH . Since I is basic, the composition q ◦ φ : LR(E \ H) → LR(E)/I has the property that (q ◦ φ)(rv) 6= 0 for all v ∈ E0 and r ∈ R \ {0}. Since E satisfies Condition (K), it follows from Proposition 7.13 that E \ H satisfies Condition (L). Hence we may apply Theorem 6.5 to conclude that q ◦ φ is injective. Since φ is an isomorphism, this implies that q is injective and I = IH. It then follows from Lemma 7.7 that I is graded. Conversely, suppose that E does not satisfy Condition (K). Then Proposi- tion 7.13 implies that there exists a saturated hereditary subset H such that E \ H does not satisfy Condition (L). Thus there exists a closed simple path with no exit in E\H, and by Lemma 7.16 the algebra LR(E\H) ∼= LR(E)/IH contains an ideal I that is basic and not graded. If we let q : LR(E) → LR(E)/IH be the quotient map, then q is graded and q−1(I) is an ideal of LR(E) that is basic but not graded. (cid:3) Corollary 7.18. If E is a row-finite graph that satisfies Condition (K), then the map H 7→ IH is a lattice isomorphism from the lattice of saturated hereditary subsets of E onto the lattice of basic ideals of LR(E). Definition 7.19. The Leavitt path algebra LR(E) is basically simple if the only basic ideals of LR(E) are {0} and LR(E). (Note that if R = K is a field, then LK(E) is basically simple if and only if LK(E) is simple.) Theorem 7.20. Let R be a commutative ring with unit, and let E be a row-finite graph. The Leavitt path algebra LR(E) is basically simple if and only if E satisfies both of the following conditions: (i) The only saturated hereditary subsets of E are ∅ and E0, and (ii) The graph E satisfies Condition (L). Proof. Suppose that LR(E) is basically simple. Then the only basic ideals of LR(E) are {0} and LR(E), both of which are graded. By Theorem 7.17 we have that E satisfies Condition (K). It then follows from Theorem 7.9(1) and the fact that LR(E) is basically simple, that the only saturated hereditary subsets of E are ∅ and E0. Hence (i) holds. In addition, since Condition (K) implies Condition (L) (cf. Proposition 7.13) we have that (ii) holds. Conversely, suppose that (i) and (ii) hold. We shall show that E satisfies Condition (K). Let v be a vertex and let α = e1 . . . en be a closed simple path based at v. By (ii) we know that α has an exit f ; i.e., there exists 22 MARK TOMFORDE f ∈ E1 with s(f ) = s(ei) and f 6= ei for some i. If we let H be the set of vertices in E0 such that there is no path from that vertex to v, then H is saturated hereditary. By (i) we must have either H = ∅ or H = E0. Since v /∈ H, we have H = ∅. Hence for every vertex in E0, there is a path from that vertex to v. Choose a path β ∈ E∗ from r(f ) to v of minimal length. Then e1 . . . ei−1f β is a simple closed path based at v that is distinct from α. Hence E satisfies Condition (K). It then follows from Theorem 7.9(1) and (i) that LR(E) is basically simple. (cid:3) Condition (i) and (ii) in the above theorem can be reformulated in a number of equivalent ways. The equivalence of the statements (2) -- (5) in Proposition 7.22 are elementary facts about directed graphs (cf. [25, Theo- rem 1.23] and [2, Proposition 3.2]). Definition 7.21. A graph E is cofinal if whenever e1e2e3 . . . is an infinite path in E and v ∈ E0, then there exists a finite path from v to s(ei) for some i ∈ N. Proposition 7.22. Let E be a row-finite graph, let R be a commutative ring with unit, and let LR(E) be the associated Leavitt path algebra. Then the following are equivalent. (1) LR(E) is basically simple. (2) E satisfies Condition (L), and the only saturated hereditary subsets of E0 are ∅ and E0. (3) E satisfies Condition (K), and the only saturated hereditary subsets of E0 are ∅ and E0. (4) E satisfies Condition (L), E is cofinal, and whenever v is a sink in E and w ∈ E0 there is a path from w to v. (5) E satisfies Condition (K), E is cofinal, and whenever v is a sink in E and w ∈ E0 there is a path from w to v. Remark 7.23. The techniques used in this section are similar to those used to analyze ideals of graph C ∗-algebras, which were inspired by the work of Cuntz and Krieger. In [8] Cuntz and Krieger showed that the Cuntz-Krieger algebra of an irreducible matrix satisfying Condition (I) is simple. In [7, Theorem 2.5] Cuntz showed that for the Cuntz-Krieger algebra of a matrix satisfying Condition (II) there is a bijective correspondence between the ideals of the Cuntz-Krieger algebra and the hereditary subsets of a certain finite partially ordered set associated with the matrix. Subsequently, it was shown in [15, Theorem 3.5] that the hereditary subsets of this partially ordered set correspond to the gauge-invariant ideals in any universal Cuntz- Krieger algebra of a finite {0, 1}-matrix. In [20, Theorem 6.6], the authors introduced Condition (K) for graphs, and showed that for a locally finite graph satisfying Condition (K) there is a bijective correspondence between ideals in the graph C ∗-algebra and saturated hereditary subsets of the graph. Their proof used groupoid tech- niques and relied on realizing the graph C ∗-algebra as the C ∗-algebra of a LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 23 groupoid. In [5, Theorem 4.1] it was shown that for C ∗-algebras of row-finite graphs there is a bijective correspondence between gauge-invariant ideals in the graph C ∗-algebra and saturated hereditary subsets of the graph, and in [5, Theorem 4.4] it is proven that when a graph satisfies Condition (K) all the ideals of the associated C ∗-algebra are gauge invariant. The techniques used in [5] avoided the use of groupoids, and instead used methods similar to those used by Cuntz in [7]. In [6, Theorem 3.6] and [9, Theorem 3.5] the analysis of ideals was extended to non-row-finite graphs, where new phenom- ena had to be accounted for, and it was shown that gauge-invariant ideals of the graph C ∗-algebra are in bijective correspondence with admissible pairs; i.e., pairs consisting of a saturated hereditary set and a subset of break- ing vertices for this saturated hereditary subset. Furthermore, these results have been generalized to Cuntz-Pimsner algebras, and it has been shown that the gauge-invariant ideals in a Cuntz-Pimsner algebra correspond to admissible pairs of ideals in the coefficient algebra of the Hilbert bimodule [18, Theorem 8.6]. In the past five years, methods similar to those in [8], [7], [15], [5], [6], and [9] have been used to analyze the ideal structure of Leavitt path algebras over fields. It has been shown in [1, Theorem 3.11] that LK(E) is simple if and only if E satisfies Condition (L) and the only saturated hereditary subsets of E are ∅ and E0. In [3, Theorem 5.3] it was shown that if E is a row-finite graph, then the graded ideals of LK(E) are in bijective correspondence with the saturated hereditary subsets of E. Furthermore, in [26, Theorem 5.7] it was shown that for a non-row-finite graph E the graded ideals of LK(E) are in bijective correspondence with the admissible pairs of E. Moreover, it was proven in [26, Theorem 6.16] that a graph E satisfies Condition (K) if and only if every ideal in the Leavitt path algebra LK(E) is graded. 8. Tensor products and changing coefficients If R is a commutative ring with unit and if A and B are R-algebras, then the tensor product A ⊗R B is an R-module that may also be given the structure of an R-algebra with a multiplication satisfying (a1 ⊗ b1)(a2 ⊗ b2) = a1a2 ⊗ b1b2. (See [16, Ch.IV, Theorem 7.4] for details on how this multiplication is obtained.) Furthermore, if R is a commutative ring with unit that contains a unital subring S, then we may view R as an S-algebra. If, in addition, A is an S-algebra, then R ⊗S A is an R-algebra with r1(r2 ⊗ a) = r1r2 ⊗ a. Theorem 8.1. Let R be an algebra over the commutative unital ring S, and let E be a graph. Then LR(E) ∼= R ⊗S LS(E) as R-algebras. Proof. One can verify that {1 ⊗ v : v ∈ E0} ∪ {1 ⊗ e, 1 ⊗ e∗ : e ∈ E1} 24 MARK TOMFORDE is a Leavitt E-family in the R-algebra R ⊗S LS(E), and hence there exists an R-algebra homomorphism φ : LR(E) → R ⊗S LS(E) with φ(v) = 1 ⊗ v, φ(e) = 1 ⊗ e, and φ(e∗) = 1 ⊗ e∗. Furthermore, LR(E) is an S-algebra that contains a Leavitt E-family {v : v ∈ E0} ∪ {e, e∗ : e ∈ E1}. Thus there exists an S-algebra homomorphism φ : LS(E) → LR(E) with φ(v) = v, φ(e) = e, and φ(e∗) = e∗. Furthermore, using the universal property of the tensor product, one can verify that there exists an R-module morphism ψ : R ⊗S LS(E) → LR(E) with ψ(r ⊗ x) = rφ(x). Finally, one can verify that ψ is an inverse for φ (simply check on generators), and hence φ is an R-algebra isomorphism. (cid:3) Corollary 8.2. Let E be a graph, and let K be a field. If we view K as a Z-module, then LK(E) ∼= K ⊗Z LZ(E). Furthermore, if K has characteristic p > 0, then we may view K as a Zp- module and LK(E) ∼= K ⊗Zp LZp(E). (Here LZ(E) denotes the Leavitt path ring of characteristic 0 associated to E, and LZp(E) denotes the Leavitt path ring of characteristic p associated to E, as described in Definition 3.2.) Let R be a commutative ring with unit that contains a unital subring S, and let E be a row-finite graph. For a saturated hereditary subset H of E, let I S H denote the ideal in LS(E) generated by {v : v ∈ H} and let I R H denote the ideal in LR(E) generated by {v : v ∈ H}. Theorem 7.9 shows that any graded basic ideal of LS(E) has the form I S H , and any graded basic ideal of LR(E) has the form I R H 7→ I R H. Thus the map I S H is a lattice isomorphism from the lattice of graded basic ideals of LS(E) onto the lattice of graded basic ideals of LR(E). If we use Theorem 8.1 to identify LR(E) with R ⊗S LR(E) via the isomorphism described in the proof, then I R H = R ⊗ I S H , and we see that I 7→ R ⊗ I is a map from ideals of LS(E) onto ideals of LR(E) that restricts to an isomorphism from graded basic ideals of LS(E) onto graded basic ideal of LR(E). In the special case that S = Z and R = K is a field (respectively, a field of characteristic p), we see that all ideals of LK(E) are basic, and hence the map I 7→ K ⊗ I is a lattice isomorphism from the lattice of graded basic ideals of LZ(E) (respectively, LZp(E)) onto the lattice of graded ideals of LK (E). This suggests that properties of graded ideals of LK(E) may derived from properties of graded basic ideals of LZ(E) and LZn(E). In the study of Leavitt path algebras over fields, it has frequently been found that properties of LK(E) depend only on properties of the graph E and are independent of the particular field K that is chosen. The fact that LK(E) ∼= K ⊗Z LZ(E) (and LK(E) ∼= K ⊗Zp LZp(E) if char K = p), suggests that properties of LK(E) may consequences of properties of the Leavitt path rings LZ(E) and LZp(E). One may speculate that this is the reason many properties of LK(E) are independent of K. LEAVITT PATH ALGEBRAS WITH COEFFICIENTS IN A COMMUTATIVE RING 25 References [1] G. Abrams and G. Aranda-Pino, The Leavitt path algebra of a graph, J. Algebra 293 (2005), 319 -- 334. [2] G. Abrams and G. Aranda-Pino, The Leavitt path algebras of arbitrary graphs, Hous- ton J. Math 34 (2008), 423 -- 442. [3] P. Ara, M. A. Moreno, and E. Pardo, Nonstable K-theory for graph algebras, Al- gebr. Represent. Theory 10 (2007), 157 -- 178. [4] G. Aranda-Pino, E. Pardo, and M. Siles-Molina, Exchange Leavitt path algebras and stable rank, J. Algebra 305 (2006), 912 -- 936. [5] T. Bates, D. Pask, I. Raeburn and W. Szyma´nski, The C ∗-algebras of row-finite graphs, New York J. Math. 6 (2000), 307 -- 324. [6] T. Bates, J. H. Hong, I. Raeburn, and W. Szyma´nski, The ideal structure of C ∗- algebras of infinite graphs, Illinois J. Math 46 (2002), 1159 -- 1176. [7] J. Cuntz, A class of C ∗-algebras and topological Markov chains II: reducible chains and the Ext-functor for C ∗-algebras, Invent. Math. 63 (1981), 25 -- 40. [8] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980), 251 -- 268. [9] D. Drinen and M. Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Mountain J. Math. 35 (2005), 105 -- 135. [10] R. Exel and M. Laca, Cuntz-Krieger algebras for infinite matrices, J. reine angew. Math. 512 (1999), 119 -- 172. [11] N. Fowler, P. Muhly, and I. Raeburn, Representations of Cuntz-Pimsner algebras, Indiana Univ. Math. J. 52 (2003), 569 -- 605. [12] K. Fuller, On rings whose left modules are direct sums of finitely generated modules, Proc. Amer. Math. Soc. 54 (1976), 39 -- 44. [13] J. L. Garc´ıa and J J. Sim´on, Morita equivalence for idempotent rings, J. Pure Appl. Algebra 76 (1991), 39 -- 56. [14] K. R. Goodearl, Leavitt path algebras and direct limits, preprint. [15] A. an Huef and I. Raeburn, The ideal structure of Cuntz-Krieger algebras, Ergodic Theory Dynam. Systems 17 (1997), 611 -- 624. [16] T. W. Hungerford, Algebra. Reprint of the 1974 original. Graduate Texts in Mathe- matics 73. Springer-Verlag, New York-Berlin, 1980. xxiii+502 pp. [17] T. Katsura, On C ∗-algebras associated with C ∗-correspondences, J. Funct. Anal. 217 (2004), 366 -- 401. [18] T. Katsura, Ideal structure of C ∗-algebras associated with C ∗-correspondences, Pacific J. Math. 230 (2007), 107 -- 145. [19] A. Kumjian, D. Pask, and I. Raeburn, Cuntz-Krieger algebras of directed graphs, Pacific J. Math. 184 (1998), 161 -- 174. [20] A. Kumjian, D. Pask, I. Raeburn, and J. Renault, Graphs, groupoids, and Cuntz- Krieger algebras, J. Funct. Anal. 144 (1997), 505 -- 541. [21] W. G. Leavitt, Modules without invariant basis number, Proc. Amer. Math. Soc. 8 (1957), 322 -- 328. [22] G. J. Murphy, C ∗-algebras and Operator Theory. Academic Press, Inc., Boston, MA, 1990. x+286 pp. [23] I. Raeburn, Graph algebras. CBMS Regional Conference Series in Mathematics, 103, Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. vi+113 pp. [24] I. Raeburn and W. Szyma´nski, Cuntz-Krieger algebras of infinite graphs and matrices, Trans. Amer. Math. Soc. 356 (2004), 39 -- 59. [25] M. Tomforde, Structure of graph C*-algebras and their generalizations, Chapter in the book "Graph Algebras: Bridging the gap between analysis and algebra", Eds. Gonzalo 26 MARK TOMFORDE Aranda Pino, Francesc Perera Dom`enech, and Mercedes Siles Molina, Servicio de Publicaciones de la Universidad de M´alaga, M´alaga, Spain, 2006. [26] M. Tomforde, Uniqueness theorems and ideal structure for Leavitt path algebras, J. Al- gebra 318 (2007), 270 -- 299. Department of Mathematics, University of Houston, Houston, TX 77204- 3008, USA E-mail address: [email protected]
1605.01202	1	1605	2016-05-04T10:02:12	A dynamical characterization of diagonal preserving $$-isomorphisms of graph $C^$-algebras	[ "math.OA" ]	We characterize when there exists a diagonal preserving $$-isomorphism between two graph $C^$-algebras in terms of the dynamics of the boundary path spaces. In particular, we refine the notion of "orbit equivalence" between the boundary path spaces of the directed graphs $E$ and $F$ and show that this is a necessary and sufficient condition for the existence of a diagonal preserving $$-isomorphism between the graph $C^$-algebras $C^(E)$ and $C^(F)$.	math.OA	math	A DYNAMICAL CHARACTERIZATION OF DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C∗-ALGEBRAS SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ Abstract. We characterize when there exists a diagonal preserving ∗-iso- morphism between two graph C ∗-algebras in terms of the dynamics of the boundary path spaces. In particular, we refine the notion of "orbit equivalence" between the boundary path spaces of the directed graphs E and F and show that this is a necessary and sufficient condition for the existence of a diagonal preserving ∗-isomorphism between the graph C ∗-algebras C ∗(E) and C ∗(F ). 1. Introduction The notion of continuous orbit equivalence, pioneered by Matsumoto ([Mat13]), has proven to be an extremely important vehicle for understanding the relationship between dynamical systems and the C∗-algebras that they define. Indeed, this concept was a key ingredient which allowed Matsumoto and Matui to prove that the stabilized Cuntz-Krieger algebras become complete invariants for flow equivalence of irreducible shifts of finite type when considered not just as C∗-algebras, but as C∗-algebras containing a canonical commutative subalgebra, the diagonal. The key result in [MM14] thus gave an extremely elegant answer to the question that has been left open since Rørdam as a key step in the proof of his seminal classification result [Rør95] showed that two such shift spaces can give the same Cuntz-Krieger algebra without being flow equivalent: The diagonal is precisely the extra structure which is needed for the C∗-algebra to remember its underlying shift space. The success of the approach of Matsumoto and Matui begs the question of whether or not similar results hold true for more general C∗-algebras such as non- simple Cuntz-Krieger algebras and (simple or non-simple) graph C∗-algebras, ob- jects which are currently (see [ERRS16]) giving way to classification in a way paral- lelling Rørdam's results. In a sweeping generalization of Matsumoto's fundamental result, Brownlowe, Carlsen and Whittaker in [BCW14] showed that continuous or- bit equivalence exactly translates to diagonal-preserving isomorphism of the graph C∗-algebras for any graph with the so-called Condition (L), and proved by example that this condition is necessary. In the paper at hand, we will study continuous orbit equivalences preserving even- tually periodic points and prove that they exactly correspond to diagonal-preserving isomorphism of graph C∗-algebras. This small adjustment of the notions studied in [MM14] and [BCW14] thus allow a complete understanding also when Condition (L) fails. In particular, we prove that the original notion of orbit equivalence cor- responds to diagonal-preserving isomorphism when the graphs are finite and have no sinks, a case prominently containing the classical Cuntz-Krieger case. Our method of proof involves reducing the general case to the Condition (L) case and hence to the main result of [BCW14] by an elaboration the concept of "plugging" and "unplugging" graphs introduced in [ERRS16]. Date: June 28, 2021. 2010 Mathematics Subject Classification. Primary: 46L55; Secondary: 46L35, 37B10. Key words and phrases. Orbit equivalence, diagonal preserving ∗-isomorphism. 1 2 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ After having circulated an early version of this paper, we were made aware that our main result had been simultaneously obtained by Carlsen and Winger ([CW16]) by completely different methods. 2. Preliminaries In this section, we provide the definitions of the objects considered in this paper. We start with some background on directed graphs, graph C∗-algebras and their diagonal subalgebras. The definitions of the boundary path space and the graph groupoid of a directed graph are also provided. 2.1. Graph C∗-algebras and the diagonal subalgebra. A directed graph E = (E0, E1, r, s) consists of sets E0 and E1 and functions, r, s : E1 → E0 called the range and source maps, respectively. The elements of E0 are called the vertices of E and the elements of E1 are called the edges of E. Assumption 2.1. Throughout the paper, unless stated otherwise, when we say a graph we mean a directed graph. Moreover, we will only consider graphs such that the set of vertices and the set of edges are countable sets. Let E be a graph. A path of length n in E is a finite sequence µ = e1e2 . . . en with ei ∈ E1 and r(ei) = s(ei+1) for all i = 1, . . . , n − 1. We will regard the vertices E0 of E as paths of length zero. Denote the set of paths of length n in E by En. Set E∗ =Sn∈N0 En and set E≥k =Sn≥k En. We extend the range and source maps to E∗ by r(v) = s(v) = v for v ∈ E0, and s(e1 · · · en) = s(e1) and r(e1 · · · en) = r(en). A loop in E is an edge e in E such that s(e) = r(e). A cycle in E is a path µ ∈ E≥1 such that s(µ) = r(µ). A cycle e1e2 · · · en in E is said to have an exit if there exists an f ∈ E1 such that s(ek) = s(f ) for some k = 1, 2, . . . , n with f 6= ek. A vertex-simple cycle in E is a cycle µ = e1e2 · · · en such that r(ei) 6= r(ej ) for all i 6= j. A return path in E is a cycle µ = e1e2 · · · en such that r(ei) 6= r(µ) for all i = 1, 2, . . . , n − 1. An infinite path in E is an infinite sequence (en)∞ n=1, denoted e1e2 · · · , such that ei ∈ E1 and r(ei) = s(ei+1) for all i. The set of infinite paths in E will be denoted by E∞. If e is a loop and n ∈ N, then en will denote the cycle of length n with edges equal to e and e∞ will denote the infinite path with edges equal to e. If µ is a cycle in E, then µ∞ denotes the infinite path µµµ · · · . Definition 2.2. A graph E is said to have Condition (L) if every cycle in E has an exit. Let V, W be subsets of E0, S a subset of E∗ ∪ E∞, and n ∈ N0. Define subsets of E∗ ∪ E∞, labelled V S, SW , and V SW , by V S = {µ ∈ S s(µ) ∈ V }, SW = {µ ∈ S r(µ) ∈ W }, and V SW = V S ∩ SW . Note that SW and V SW are subsets of E∗ since the range map is only defined on E∗. We will write vS if V = {v}. Similarly for SW and V SW . A vertex v ∈ E0 is called regular if 0 < vE1 < ∞. Denote the set of regular vertices in E by E0 reg. A singular vertex is a vertex v in E that is not regular. We denote the set of singular vertices by E0 sing. A vertex v in E is called a sink if vE1 = 0 and is called an infinite emitter if vE1 = ∞. Denote the set of sinks by E0 inf . Hence, E0 sing = E0 We call an infinite path e1e2 · · · in E a tail if s(ei)E1 = {ei} = E1s(ei+1), and non-wandering if s(ei)E1 = {ei}. If µ ∈ E∗ is a cycle with no exits, then µ∞ is non-wandering. Definition 2.3. Let E be a graph. A Cuntz-Krieger E-family in a C∗-algebra A sink and the set of infinite emitters by E0 consists of a set of mutually orthogonal projections (cid:8)Pv(cid:12)(cid:12) v ∈ E0(cid:9) ⊆ A and a set of partial isometries(cid:8)Se(cid:12)(cid:12) e ∈ E1(cid:9) ⊆ A satisfying sink ∪ E0 inf . The topology on ∂E is given as follows: For µ ∈ E∗, the cylinder set of µ is the set where µx is the concatenation of paths. For µ ∈ E∗ and a finite subset F of r(µ)E1, set ∂E = E∞ ⊔(cid:8)µ ∈ E∗(cid:12)(cid:12) r(µ) ∈ E0 sing(cid:9) . ZE(µ) = {µx ∈ ∂E x ∈ r(µ)∂E} , ZE(µ\F ) = ZE(µ)\ [e∈F ZE(µe)! . DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 3 (CK1) S∗ (CK2) S∗ (CK3) SeS∗ e Sf = 0 for all e, f ∈ E1 with e 6= f ; e Se = Pr(e) for all e ∈ E1; e ≤ Ps(e) for all e ∈ E1; and e for all v ∈ E0 (CK4) Pv =Pe∈vE1 SeS∗ sv = pv. Then the C∗-subalgebra span(cid:8)sµs∗ reg. The graph C∗-algebra C∗(E) is the universal C∗-algebra generated by a Cuntz- Krieger E-family. If µ = e1e2 · · · en ∈ E≥2, we set sµ = se1 se2 · · · sen and for v ∈ E0, we set diagonal subalgebra of C∗(E) and is denoted by D(E). Definition 2.4. Let E and F be graphs. A ∗-isomorphism Φ : C∗(E) → C∗(F ) is a diagonal preserving ∗-isomorphism if Φ(D(E)) = D(F ). µ(cid:12)(cid:12) µ ∈ E∗(cid:9) of C∗(E) is called the 2.2. Boundary path space and the graph groupoid of a graph. The defini- tions given in this section follows that of [BCW14, Section 2.2 and 2.3]. Definition 2.5. Let E be a graph. The boundary path space of E is the space (When there is no cause for confusion, we will at times omit the subscripts.) The topology of ∂E is the topology generated by BE =(cid:8)ZE(µ\F )(cid:12)(cid:12) µ ∈ E∗, F a finite subset of r(µ)E1(cid:9) . The boundary path space ∂E is a locally compact Hausdorff space with basis BE and every U ∈ BE is compact and open (see [Web14, Theorem 2.1 and Theo- rem 2.2]). The key relationship between ∂E and D(E) is the following theorem. Theorem 2.6 ([Web14, Theorem 3.7]). There exists a unique homeomorphism hE from ∂E to the spectrum of D(E) given by hE(x)(sµs∗ µ) =(1 0 if x ∈ Z(µ) otherwise. Lemma 2.7. Let E be a graph and let S be a subset of E0 (i.e., closed and open) subset of ∂E. Proof. First note that for all v ∈ E0 sink and for all µ ∈ E∗ and F finite subset of r(µ)E1, v ∈ ZE(µ\F ) if and only if v = µ and F = ∅. In particular, for all v ∈ E0 again that the only cylinder set that contains a sink v is {v}, we conclude that if x ∈ ∂E\S, then any cylinder set ZE(µ) containing x will satisfy ZE(µ) ⊆ ∂E\S. So, ∂E\S is open which implies that S is closed. (cid:3) sink, ZE(v) = {v}. Therefore, S = Sv∈S ZE(v), and hence open. Using sink. Then S is a clopen If x ∈ ∂E, then set x =(∞ if x ∈ E∞ ∂E since ∂E≥n is equal toSµ∈En Z(µ). n For n ∈ N0, set ∂E≥n = {x ∈ ∂E x ≥ n}. Note that ∂E≥n is an open subset of if x ∈ En for some n ∈ N0. 4 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ Definition 2.8. Let E be a graph. Define the shift map σE : ∂E≥1 → ∂E on E by σE(x) =(e2e3 · · · r(x) if x = e1e2 · · · ∈ ∂E≥2 if x ∈ ∂E ∩ E1. E : ∂E≥n → ∂E will be the n-fold composition of σE with itself and For n ≥ 1, σn σ0 E : ∂E → ∂E will be the identity map. One can check that for all n ∈ N0, σn E is continuous and moreover, σn E is a local homeomorphism. We now define the graph groupoid of a graph E. Definition 2.9. Let E be a graph. The graph groupoid GE is defined as follows: As a set, GE = {(x, m − n, y) x, y ∈ ∂E with x ≥ m, y ≥ n, and σm E (x) = σn E(y)} . The product is defined by (x, k, y)(w, l, z) = (x, k + l, z) if y = w and undefined otherwise, and the inverse of (x, k, y) is (y, −k, x). The set of units G(0) E of GE is {(x, 0, x) x ∈ ∂E}. Let m, n ∈ N0, U be an open subset of ∂E≥m such that σm EV is injective. Suppose σm E U is injective, and V E(V ). E (U ) = σn be an open subset of ∂E≥n such that σn Set ZE(U, m, n, V ) = {(x, m − n, y) ∈ GE x ∈ U, y ∈ V, σm E (x) = σn E(y)} . (When there is no cause for confusion, we will at times omit the subscripts.) Then GE is a locally compact, Hausdorff, étale topological groupoid with the topology generated by the basis consisting of sets ZE(U, m, n, V ). One checks that the map µ ∈ ∂E 7→ (µ, 0, µ) ∈ G(0) E . We will freely identify G(0) E is a homeomorphism from ∂E to G(0) E with ∂E using this map throughout the paper without further mention. Thus, we have range and source maps r, s : GE → ∂E defined by r((x, k, y)) = x and s((x, k, y)) = y. By [ADR00, Proposition 3.3.5 and 6.1.8], the reduced and universal C∗-algebra of GE are equal since GE is topologically amenable (see [Yee07, Proposition 6.2]). We denote this C∗-algebra by C∗(GE). Theorem 2.10 ([BCW14, Proposition 2.2]). If E is a graph, then there exists a unique ∗-isomorphism ΦE : C∗(E) → C∗(GE) such that ΦE(pv) = 1Z(v,v) for all v ∈ E0 and ΦE(se) = 1Z(e,r(e)) for all e ∈ E1, and such that ΦE(D(E)) = C0(G(0) E ). 3. Orbit equivalence preserving periodic points and pseudogroups We now define orbit equivalence between graphs E and F preserving periodic points. When E and F are graphs with finitely many vertices and no sinks, or when E and F are graphs satisfying Condition (L), then this notion of orbit equivalence coincides with the notion of orbit equivalence defined in [BCW14, Definiton 3.1]. 3.1. Orbit equivalence preserving periodic points. Let E be a graph, and let ∂Eiso denote the set of isolated points in ∂E. Definition 3.1. Let E be a graph. Then x ∈ ∂E is eventually periodic if x = µν∞ where µ ∈ E∗, ν is a cycle in E, and r(µ) = s(ν). Note that an eventually periodic point x ∈ ∂E is an isolated point if and only if x = µν∞ where µ ∈ E∗ and ν ∈ E∗ is a cycle with no exits satisfying r(µ) = s(ν). sink, and call x even- tually non-wandering if x = µy where µ ∈ E∗ and y ∈ r(µ)E∞ is non-wandering. We call x ∈ ∂E eventually a sink if x ∈ E∗ with r(x) ∈ E0 DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 5 The isolated points in ∂E are exactly the points that are eventually sinks or eventu- ally non-wandering. Clearly, the eventually periodic isolated points are exactly the eventually periodic eventually non-wandering points. We will refer to the eventually non-wandering points that are are not eventually periodic as eventually non-periodic non-wandering. Definition 3.2. Let E and F be graphs and let κ : ∂E → ∂F be a homeomor- phism. We say that κ is an orbit equivalence if there exist continuous functions l, m : ∂E≥1 → N0 and l′, m′ : ∂F ≥1 → N0 such that (1) σm(x) (2) σm′(y) F (κ(σE(x))) = σl(x) (κ−1(σF (y))) = σl′(y) F (κ(x)) for all x ∈ ∂E≥1; and E (κ−1(y)) for all y ∈ ∂F ≥1. E If such a κ exists, we say that E and F are orbit equivalent or there exists an orbit equivalence between E and F . If, in addition, κ satisfies (3) for all x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic, then we say that κ is an orbit equivalence preserving periodic points. If such a κ exists, then we say that there exists an orbit equivalence between E and F preserving periodic points. All eventually non-wandering points in ∂E are eventually periodic if the graph E has finitely many vertices. Hence if E and F are graphs with finitely many vertices and no sinks, an orbit equivalence between E and F will automatically preserve periodic points. Likewise, if E and F are graphs satisfying Condition (L), any orbit equivalence between E and F will preserve periodic points, as E and F contain no eventually periodic isolated points. In general, if E and F are orbit equivalent graphs, there may not exist an or- bit equivalence between E and F preserving periodic points, as all three types of isolated points may be interchanged by an orbit equivalence. Consider the graphs E : • and F : • . By [BCW14, Example 5.2], E and F are orbit equivalent but there is no orbit equivalence between E and F preserving periodic points, as the isolated point in ∂E is a sink while the isolated point in ∂F is periodic. In Example 3.3 we provide examples of orbit equivalences that interchange eventually periodic points with eventually non-periodic non-wandering points, and eventually sinks with eventually non-periodic non-wondering points. Example 3.3. Consider the graphs E, F and G: E : e2 e1 ... • • f F : e1 e2 • • ... G : ... • v e2 e1 D D 6 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ Then all points in ∂E are eventually periodic isolated points, all points in ∂F are eventually non-periodic non-wandering, and all points in ∂G are eventually sinks. We now show that E, F , and G are orbit equivalent. Define κ1 : ∂E → ∂F and κ2 : ∂F → ∂G by κ1(ei · · · e1f ∞) = ei+1ei+2 · · · for i ≥ 1 and κ1(f ∞) = e1e2 · · · , and κ2(eiei+1 · · · ) = ei−1 · · · e1 for i ≥ 2 and κ2(e1e2 · · · ) = v. Clearly, κ1 and κ2 are bijective, and they are continuous and open since ∂E, ∂F , and ∂G carry the discrete topologies. One readily checks that m1, l1 : ∂E≥1 → N0 defined by m1(f ∞) = 0, m1(ei · · · e1f ∞) = 1, and l1(x) = 0 for all x ∈ ∂E≥1 satisfies σm1(x) F for all x ∈ ∂E≥1, and that m′ satisfies (κ1(σE(x))) = σl1(x) 1 : ∂F ≥1 → N0 defined by m′ (κ1(x)) F 1, l′ 1(x) σm′ E (κ−1 1 (σF (x))) = σl′ 1(x) E (κ−1 1 (x)) 1(x) = 1 and l′ 1(x) = 0 for all x ∈ ∂F ≥1. So κ1 is an orbit equivalence, since the maps m1, l1, m′ 1 are automatically continuous. Similarly, m2, l2 : ∂F ≥1 → N0 defined by m2(x) = 1 and l2(x) = 0, and m′ 2(x) = 0, lets us conclude that κ2 is an orbit equivalence. 2 : ∂G≥1 → N0 defined by m′ 2(x) = 1 and l′ 1, l′ 2, l′ 3.2. The pseudogroup PE and the groupoid of germs of PE. We now recall the groupoid of germs defined in [Ren08, Section 3]. Let X be a topological space. A homeomorphism h : U → V where U and V are open subsets of X is called a partial homeomorphism. Under composition and inverse, the collection of partial homeomorphisms on X is an inverse semigroup. A pseudogroup on X is a family of partial homeomorphisms of X stable under composition and inverse. Let P be a pseudogroup on X. A partial homeomorphism h : U → V is said to locally belong to P if for all x ∈ U , there exists an open neighborhood W of x and there exists g ∈ P such that hW = gW . The pseudogroup P is ample if each partial homeomorphism h : U → V that locally belongs to P must also be element in P. Definition 3.4. Let P be a pseudogroup on the topological space X. The groupoid of germs of P is GP = {[x, h, y] h ∈ P, y ∈ dom(h), x = h(y)} where [x, h, y] = [x, g, y] if and only if there exists a neighborhood V of y in X such that hV = gV . The range and source maps are given by r([x, h, y]) = x and s([x, h, y]) = y. The partially defined product is [x, h, y][y, g, z] = [x, h ◦ g, z], undefined otherwise and the inverse [x, h, y]−1 = [y, h−1, x]. The groupoid GP is given the topology given by basic open sets U(U, h, V ) = {[x, g, y] ∈ GP x ∈ U, y ∈ V } where U and V are open subsets of X and h ∈ P. We recall how to construct a pseudogroup from an étale groupoid G. A subset A of a groupoid G is called a bisection if rA and sA are injective functions. Then the set of all open bisections S(G) forms an inverse semigroup with composition law and AB =nγγ′(cid:12)(cid:12)(cid:12) (γ, γ′) ∈ (A × B) ∩ G(2)o A−1 =(cid:8)γ−1(cid:12)(cid:12) γ ∈ A(cid:9) . DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 7 Let A be an open bisection. Then define αA : s(A) → r(A) by αA(s(γ)) = r(γ) for all γ ∈ A. One checks that αA is a homeomorphism. Then the pseudogroup on G(0) is P(G) = {αA A is an open bisection} . Assumption 3.5. Isomorphisms between topological groupoids are isomorphisms between groupoids that are also homeomorphisms. The following proposition follows from [Ren08, Proposition 3.6] and the proofs of [Ren08, Proposition 3.2 and Corollary 3.3]. Proposition 3.6. Let G be an étale groupoid. Define ϕG : G → GP(G) by ϕG(γ) = [r(γ), αA, s(γ)] where A is an open bisection containing γ. Then ϕG is a well-defined surjective ho- momorphism of groupoids. Moreover, if G is Hausdorff and topologically principal, then ϕG is an isomorphism. As an immediate consequence, we get the following corollary. Corollary 3.7. Let G and H be étale groupoids. Suppose there exists a homeomor- phism κ : G(0) → H(0) such that Then there exists an isomorphism ψκ : GP(G) → GP(H) defined by κ ◦ P(G) ◦ κ−1 :=(cid:8)κ ◦ g ◦ κ−1(cid:12)(cid:12) g ∈ P(G)(cid:9) = P(H). ψκ([x, g, y]) = [κ(x), κ ◦ g ◦ κ−1, κ(y)]. Consequently, if G and H are Hausdorff and topological principal, then ψκ induces an isomorphism eψκ : G → H such that eψκG(0) = κ. Of particular interest to us is the pseudogroup of the étale groupoid GE for a graph E. So, for a graph E, we denote P(GE) by PE and we call PE the pseudogroup of E. In [BCW14, Proposition 3.4], the authors prove that E and F are orbit equivalent if and only if the pseudogroups of E and F are isomorphic, i.e., there exists a homeomorphism κ : ∂E → ∂F such that κ ◦ PE ◦ κ−1 E =(cid:8)κ ◦ g ◦ κ−1(cid:12)(cid:12) g ∈ PE(cid:9) = PF . They actually prove a stronger statement in the sense that the orbit equivalence between E and F induces the isomorphism between the pseudogroups of E and F and vice versa. We record this in the following proposition. Proposition 3.8 ([BCW14, Proposition 3.4]). Let E and F be graphs and κ from ∂E to ∂F be a homeomorphism. Then κ is an orbit equivalence (preserving periodic points) if and only if κ ◦ PE ◦ κ−1 = PF (and for all x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic). Proposition 3.9. For all graphs E1, E2, and E3, if κ1 from ∂E1 to ∂E2 and κ2 from ∂E2 to ∂E3 are orbit equivalences (preserving periodic points), then κ2 ◦ κ1 from ∂E1 to ∂E3 is an orbit equivalence (preserving periodic points). Proof. Let E1, E2, and E3 be graphs. Suppose κ1 : ∂E1 → ∂E2 and κ2 : ∂E2 → ∂E3 are orbit equivalences. By Proposition 3.8, we have κ1 ◦ PE1 ◦ κ−1 1 = PE2 and κ2 ◦ PE2 ◦ κ−1 2 = PE3. It follows that (κ2 ◦ κ1) ◦ PE1 ◦ (κ2 ◦ κ1)−1 = PE3. Thus, by Proposition 3.8, κ2 ◦ κ1 : ∂E1 → ∂E3 is an orbit equivalence. Suppose κ1 and κ2 are orbit equivalences preserving periodic points. Then, as above (κ2 ◦ κ1) ◦ PE1 ◦ (κ2 ◦ κ1)−1 = PE3. Moreover, since κ1 and κ2 are orbit equivalences preserving periodic points, for all isolated points x in ∂E1, x is 8 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ eventually periodic if and only if κ1(x) is eventually periodic if and only if (κ2◦κ1)(x) is eventually periodic. (cid:3) We are now able to prove a stronger version of [BCW14, Theorem 5.1]. This version will be important for us in the proof of Theorem 5.3. Theorem 3.10 (cf. [BCW14, Theorem 5.1]). Let E and F be graphs satisfying Condition (L). Suppose κ : ∂E → ∂F is an orbit equivalence from E to F . Then there exists an isomorphism ϕ : GE → GF such that ϕ∂E = κ. Proof. By Proposition 3.8, κ ◦ PE ◦ κ−1 = PF . Since E and F satisfy Condi- tion (L), by [BCW14, Proposition 2.3], GE and GF are topologically principal. By Corollary 3.7, there exists an isomorphism ϕ : GE → GF such that ϕ∂E = κ. (cid:3) 3.3. The unplugged graph and orbit equivalence. For a graph E, let E0 cycle be the set of vertices of E that is on a vertex-simple cycle with no exits. Suppose E satisfies the property that if ν is a vertex-simple cycle with no exits, then ν is a loop. This entails that every vertex v ∈ E0 cycle supports a unique loop ev. Note that cycle and e ∈ E1 such that s(e) = v, then e = ev. Denote the set of all loops if v ∈ E0 based at a vertex in E0 cycle. Also note that if e, f ∈ E1 cycle with s(e) = s(f ) (equivalently, r(e) = r(f )), then e = f . cycle. Note that s(E1 cycle) = r(E1 cycle) = E0 cycle by E1 Let E be a graph such that all vertex-simple cycles with no exits are loops. Let the unplugged graph Eg of E be the graph defined by E0 g = E0 and E1 g = E1\E1 cycle with the range and source maps of Eg the restrictions of the range and source maps of E respectively. Proposition 3.11. Let E be a graph such that each vertex-simple cycle with no exits is a loop. Define κE : ∂Eg → ∂E by κE(x) =(xe∞ x, r(x), g with r(x) ∈ E0 cycle if x ∈ E∗ otherwise. Then κE is an orbit equivalence such that for each isolated point x ∈ ∂Eg, rEg(x) ∈ E0 cycle if and only if κE(x) is an isolated point in ∂E that is eventually periodic. Proof. A computation shows that κE is a bijection with κ−1 µe∞ all other x. Let µ = e1e2 · · · en ∈ E∗. Suppose µ ∈ E∗ r(µ) for some µ = e1 · · · en with r(µ) ∈ E0 g. Then cycle and en 6= er(µ), and κ−1 E (x) = µ when x = E (x) = x for κ−1 E (ZE(µ)) = {µx ∈ ∂Eg x ∈ r(µ)∂Eg} = ZEg (µ) which is open. Suppose µ /∈ E∗ and ej /∈ E1 ZE(µ) = ZE(e1 · · · ei0−1) = {e1 · · · ei0−1e∞ Hence, cycle cycle for all j < i0. Then ei0 = ek for all i0 ≤ k ≤ n. Therefore, i0 }, where e1 · · · ei0−1 = s(µ) if i0 = 1. g. Let i0 ∈ N with 1 ≤ i0 ≤ n such that ei0 ∈ E1 κ−1 E (ZE(µ)) = κ−1 E ({e1 · · · ei0−1e∞ i0 }) = {e1 · · · ei0−1} = ZEg(e1 · · · ei0−1) which is open. Let F be a finite nonempty subset of r(µ)E1. then F = {er(µ)} so ZE(µ\F ) = ∅, hence κ−1 If r(µ) ∈ E0 E (ZE(µ\F )) is trivially open. cycle If DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 9 r(µ) /∈ E0 cycle then µ ∈ E∗ g and F ⊆ r(µ)E1 g. Hence κ−1 E (ZE(µ\F )) = κ−1 E (ZE(µ))\ [e∈F = ZEg(µ)\ [e∈F E (ZE(µe))! κ−1 ZEg(µe)! = ZEg(µ\F ) which is open. We have just shown that κE is continuous. Let µ ∈ E∗ g. Then κE(ZEg(µ)) = ZE(µ). Let F be a finite subset of r(µ)E1 g. Then κE(ZEg(µe)) = ZE(µe) for all e ∈ F , hence κE(ZEg (µ\F )) = ZE(µ)\ [e∈F ZE(µe)! = ZE(µ\F ), which is open. Hence, κE is an open map. Therefore, κE is a homeomorphism. Define m : ∂E≥1 → N0 by m(y) =(0, if y ∈ κE(E0 cycle) 1, otherwise. cycle ⊆ (Eg)0 Since E0 homeomorphism, κE(E0 sink, by Lemma 2.7, E0 cycle is clopen in ∂Eg. Since κE is a cycle) is clopen in ∂E, so m is continuous. A computation shows that σE(κE(x)) = κE(σEg(x)) and σm(y) Eg (κ−1 E (y)) = κ−1 E (σE(y)). Therefore, κE is an orbit equivalence. The last part of the proposition follows immediately from the construction of κE. (cid:3) The next lemma shows that we can adjust an orbit equivalence so that sinks are sent to sinks. This will be used in the proof of Theorem 5.3 to construct a diagonal preserving ∗-isomorphism. Note that if µ ∈ E∗ such that r(µ) is a sink, then µ ∈ ∂Eiso. Lemma 3.12. Let E be a graph and let F be a subset of E≥1 such that (1) for all µ ∈ F , r(µ) ∈ E0 (2) for all µ, ν ∈ F , r(µ) = r(ν) if and only if µ = ν, and (3) F is a closed subset of ∂E. sink, Define κ : ∂E → ∂E by κ(x) = x r(x) µ if x /∈ F ∪ r(F ) if x ∈ F if x = r(µ) for some µ ∈ F . Then κ is an orbit equivalence. Proof. We must show that κ is a homeomorphism and there exist continuous func- tions l, m : ∂E≥1 → N0 and l′, m′ : ∂E≥1 → N0 such that σm(x) E (κ(σE(x))) = σl(x) E (κ(x)) and σm′(x) E (κ−1(σE(x))) = σl′(x) E (κ−1(x)) for all x ∈ ∂E≥1. A computation shows that κ ◦ κ = id. Hence, it is enough to show that κ is continuous and there exist l, m : ∂E≥1 → N0 such that σm(x) E (κ(σE(x))) = σl(x) E (κ(x)) for all x ∈ ∂E≥1. Since r(F ) ⊆ E0 sink, by Lemma 2.7, r(F ) is a clopen subset of ∂E. Since F ⊆ ∂Eiso, F is open in ∂E. By Assumption (3), F is closed in ∂E. As F , 10 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ r(F ) and thereby also ∂E\ (F ∪ r(F )) are clopen, it suffices to check for continuity on the three sets individually. Since F and r(F ) consist of isolated points and thereby carry the discrete subspace topology, κ is automatically continuous on F and r(F ). Since κ restricts to the identity on ∂E\ (F ∪ r(F )), κ is also continuous on ∂E\ (F ∪ r(F )). So κ is continuous and hence a homeomorphism. We now produce continuous functions l, m : ∂E≥1 → N0 such that σm(x) E (κ(σE(x))) = σl(x) E (κ(x)) for all x ∈ ∂E≥1. Note that F ⊆ E∗ (paths of finite length). For each v ∈ r(F ), we will denote the unique element in F with range v by µv. Note that if x ∈ σ−1 E (F ), then x ≥ 2 since σE(x) ∈ F and F ⊆ ∂E≥1. Define l, m : ∂E≥1 → N0 by and E (F ), 0 x 1 if x ∈ F , if x ∈ σ−1 otherwise, l(x) = m(x) = E (F )(cid:1), l is also continuous on ∂E≥1\(cid:0)F ∪ σ−1 µr(x) x − 1 if x ∈ F ∩ ∂E≥2, 0 if x ∈ E1r(F ), otherwise. We first show that l is continuous. Since F is clopen in ∂E≥1 and σE is continu- E (F ) consist E (F ). As l is constant ous, we have that σ−1 of isolated points, l is automatically continuous on F and σ−1 E (F ) is a clopen subset of ∂E≥1. Since F and σ−1 on ∂E≥1\(cid:0)F ∪ σ−1 on ∂E≥1. E (F )(cid:1) and thereby For continuity of m, we first note that F ∩ ∂E≥2 and E1r(F ) are both clopen in ∂E≥1 as F and ∂E≥2 are, and as E1r(F ) = σ−1 E (F ) ∩ E1 with ∂E≥1 ∩ E1 clopen in ∂E≥1. Since E1r(F ) and F ∩ ∂E≥2 consist of isolated points, they carry the discrete subspace topology, so m is continuous on E1r(F ) and F ∩ ∂E≥2. As m is constant on the complement ∂E≥1\(E1r(F ) ∪ (F ∩ ∂E≥2)) we conclude that m is continuous. Let x ∈ ∂E≥1. Suppose x ∈ E1r(F ). Then σm(x) E (κ(σE(x))) = σ µr(x) E (κ(r(x))) = σ µr(x) E (µr(x)) = r(x) and E (κ(x)) =(κ(x) σl(x) σE(κ(x)) if x ∈ F if x ∈ E1r(F )\F = r(x). Suppose x ∈ F ∩ ∂E≥2. Then σE(x) /∈ F since r(σE(x)) = r(x) with x ∈ F , and σE(x) /∈ r(F ) as σE(x) ∈ ∂E≥1, hence κ(σE(x)) = σE(x). Thus and σm(x) E (κ(σE(x))) = σx−1 E (σE(x)) = r(x) σl(x) E (κ(x)) = σ0 E(r(x)) = r(x). DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 11 Suppose x /∈ E1r(F ) and x /∈ F ∩ ∂E≥2. Then x /∈ F and σE(x) /∈ r(F ), so and κ(x) = x, hence if x ∈ σ−1 if x /∈ σ−1 E (F ) E (F ) if x ∈ σ−1 if x /∈ σ−1 E (F ) E (F ), if x ∈ σ−1 if x /∈ σ−1 E (F ) E (F ) if x ∈ σ−1 if x /∈ σ−1 E (F ) E (F ). σm(x) E (κ(σE(x))) = κ(σE(x)) σE(x) σE(x) =(r(σE (x)) =(r(x) E (κ(x))) =(σx =(r(x) E (x) σE(x) σE(x) σl(x) We have just shown that l, m : ∂E≥1 → N0 are continuous functions and σm(x) E (κ(σE(x))) = σl(x) E (κ(x)) for all x ∈ ∂E≥1. We conclude that κ is an orbit equivalence. (cid:3) 4. The extended Weyl groupoid In [BCW14], the authors prove that a diagonal preserving ∗-isomorphism be- tween C∗(E) and C∗(F ) implies that E and F are orbit equivalent. In this section, we point out that their arguments even prove that the existence of a diagonal pre- serving ∗-isomorphism between C∗(E) and C∗(F ) implies the existence of an orbit equivalence between E and F preserving periodic points. The arguments are actu- ally contained in [BCW14, Section 4]. For the convenience of the reader, we provide the arguments here. First we need to recall the extended Weyl groupoid of (C∗(E), D(E)) as defined in [BCW14, Section 4]. Definition 4.1. Let E be a graph. The normalizer of D(E) is defined to be the set N (D(E)) = {n ∈ C∗(E) ndn∗, n∗dn ∈ D(E) for all d ∈ D(E)} . By [Ren08, Lemma 4.6], for all n ∈ N (D(E)), nn∗ and n∗n are elements in D(E). Therefore, we may define for n ∈ N (D(E), the sets and dom(n) = {x ∈ ∂E hE(x)(n∗n) > 0} ran(n) = {x ∈ ∂E hE(x)(nn∗) > 0} . By [Ren08, Proposition 4.7], for each n ∈ N (D(E)), there exists a unique homeo- morphism αn : dom(n) → ran(n) such that for all d ∈ D(E), hE(x)(n∗dn) = hE(αn(x))(d)hE (x)(n∗n). For each x ∈ ∂Eiso, we let px denote the unique element in D(E) satisfying hE(y)(px) = 1 if y = x and zero otherwise, i.e., px is the unique element in D(E) corresponding to the characteristic function χ{x} ∈ C0(∂E) under the canonical ∗-isomorphisms D(E) ∼= C0(Spec(D(E))) ∼= C0(∂E). By [BCW14, Lemma 4.3], if x ∈ ∂Eiso, then pxD(E)px is either isomorphic to C (when x is not eventually periodic) or C(T) (when x is eventually periodic). 12 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ By [BCW14, Lemma 4.4], for each x ∈ ∂Eiso, n1, n2 ∈ N (D(E)) such that x ∈ dom(n1) ∩ dom(n2) and αn1 (x) = αn2(x), U(x,n1,n2) = (hE(x)(n∗ 1n1n∗ 2n2))−1/2 pxn∗ 1n2px is a unitary in pxC∗(E)px. We define an equivalence relation ∼ on {(n, x) n ∈ N (D(E)), x ∈ dom(n)} by (n1, x1) ∼ (n2, x2) if either (1) x1 = x2 ∈ ∂Eiso, αn1 (x1) = αn2 (x2), and [U(x1,n1,n2)] = 0 in K1(px1C∗(E)px1 ) (2) x1 = x2 /∈ ∂Eiso and there is an open set V such that x1 ∈ V ⊆ dom(n1) ∩ dom(n2) and αn1 (y) = αn2 (y) for all y ∈ V . It is shown in [BCW14, Proposition 4.6] that this relation is in fact an equivalence relation. Let G(C ∗(E),D(E)) be the collection of equivalence classes. Define a partially defined product by [(n1, x1)] · [(n2, x2)] = [(n1n2, x2)] if αn2 (x2) = x1 and undefined otherwise, define an inverse map by [(n, x)]−1 = [(n∗, αn(x))]. By [BCW14, Proposition 4.7 and Proposition 4.8], G(C ∗(E),D(E)) is a groupoid and is a topological groupoid with the topology generated by {{[(n, x)] x ∈ dom(n)} n ∈ N (D(E))} . Moreover, by [BCW14, Proposition 4.8 and its proof], the map ϕE from GE to G(C ∗(E),D(E)) defined by ϕE((x, k, y)) = [(sµs∗ ν, y)] morphism. where x = µz, y = νz, k = µ − ν for some µ, ν ∈ E∗ and z ∈ ∂E, and (cid:8)pv, se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) be a Cuntz-Krieger E-family generating C∗(E), is an iso- Proposition 4.2 (cf. [BCW14, Proposition 4.11]). Let E and F be graphs. Suppose there exists a diagonal preserving ∗-isomorphism Ψ : C∗(E) → C∗(F ). Then there exists an isomorphism β : GE → GF and a homeomorphism κ : ∂E → ∂F such that β((µ, 0, µ)) = (κ(µ), 0, κ(µ)) and for all x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic. Proof. Let(cid:8)pv, se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) be a Cuntz-Krieger E-family generating C∗(E). Since Ψ : C∗(E) → C∗(F ) is a ∗-isomorphism such that Ψ(D(E)) = D(F ), there exists a homeomorphism κ : ∂E → ∂F such that hE(x)(f ) = hF (κ(x))Ψ(f ) for all f ∈ D(E) and the map λ : G(C ∗(E),D(E)) → G(C ∗(F ),D(E)) given by λ([(n, x)]) = [(Ψ(x), κ(x))] is an isomorphism. Now, β = ϕ−1 F ◦ λ ◦ ϕE : GE → GF is an isomorphism. We claim that β((µ, 0, µ)) = ((κ(µ), 0, κ(µ)) for all µ ∈ ∂E. Let µ ∈ ∂E. Then ϕF (β((µ, 0, µ))) = λ([(sµs∗ µ, µ)]) = [(Ψ(sµs∗ µ), κ(µ))]. µ), κ(µ))]. Since ϕF ((ν, 0, ν)) = [(sν s∗ Since ϕF is an isomorphism, there exists (ν, 0, ν) ∈ G(0) [(Ψ(sµs∗ ϕ−1 F ([(Ψ(sµs∗ We will now show that for all x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic. To do this, we first show that for all x ∈ ∂Eiso, then F such that ϕF ((ν, 0, ν)) = ν, ν)], we have that ν = κ(µ). Hence, µ), κ(µ))]) = (κ(µ), 0, κ(µ)), thus proving the claim. DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 13 Ψ(px) = pκ(x). Let x ∈ ∂Eiso. Suppose y ∈ ∂F . Then hF (y)(Ψ(px)) = hE(κ−1(y))(px) 0 =(1 =(1 0 if κ−1(y) = x otherwise if κ(x) = y otherwise = hF (y)(pκ(x)). Therefore, by the uniqueness of pκ(x), we have that ψ(px) = pκ(x). Hence, proving the claim. Let x ∈ ∂Eiso. Then pκ(x)D(F )pκ(x) = Ψ(pxD(E)px), and hence pκ(x)D(F )pκ(x) only if pxD(E)px ∼= C(T) if and only if pκ(x)D(F )pκ(x) is eventually periodic. ∼= pxD(E)px. Therefore, x is eventually periodic if and ∼= C(T) if and only if κ(x) (cid:3) Theorem 4.3. Let E and F be graphs. Suppose there exists a diagonal preserving ∗-isomorphism Ψ : C∗(E) → C∗(F ). Then there exists an orbit equivalence between E and F preserving periodic points. Proof. By Proposition 4.2, there exist an isomorphism β : GE → GF and a homeo- morphism κ : ∂E → ∂F such that β((µ, 0, µ)) = (κ(µ), 0, κ(µ)) and for all x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic. One can check that κ ◦ PE ◦ κ−1 = PF . By Proposition 3.8, κ is an orbit equivalence between E and F preserving periodic points. (cid:3) 5. Main result Let E be a graph and let S be a subset of E0 sink. Define Ef,S to be the graph with vertices E0 f,S = E0 and edges E1 f,S = E1 ⊔ {e(v) v ∈ S} where the range and source maps extends the range and source maps of E respec- tively, and rEf,S (e(v)) = sEf,S (e(v)) = v for all v ∈ S. Proposition 5.1. Let E and F be graphs, S1 be a nonempty subset of E0 sink, S2 be a nonempty subset of F 0 sink. Suppose there exist a bijection w : S1 → S2 and a diagonal preserving ∗-isomorphism Φ : C∗(E) → C∗(F ) such that Φ(Pv) = Qw(v) for all Then there exists a diagonal preserving ∗-isomorphism Ψ : C∗(Ef,S1 ) → C∗(Ff,S2 ). v ∈ S1, where (cid:8)Pv, Se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) is a Cuntz-Krieger E-family generating C∗(E) and(cid:8)Qv, Te(cid:12)(cid:12) v ∈ F 0, e ∈ F(cid:9) is a Cuntz-Krieger F -family generating C∗(F ). Proof. Let (cid:8)pv, se(cid:12)(cid:12) v ∈ E0 f,S1(cid:9) be a Cuntz-Krieger Ef,S1 -family gen- erating C∗(Ef,S1 ) and (cid:8)qv, te(cid:12)(cid:12) v ∈ F 0 f,S2(cid:9) be a Cuntz-Krieger Ff,S2 - family generating C∗(Ff,S2 ). Clearly, f,S2 , e ∈ F 0 f,S1, e ∈ E1 (cid:8)pv, se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) is a Cuntz-Krieger E-family in C∗(Ef,S1 ). Therefore, there exists a ∗-homo- morphism Φ1 : C∗(E) → C∗(Ef,S1 ) such that Φ1(Pv) = pv and Φ1(Se) = se for all v ∈ E0 and e ∈ E1. For all µ ∈ E∗, Φ1(Sµ) = sµ. Hence, for µ a vertex- simple cycle in E with no exits, µ is a vertex-simple cycle in Ef,S1 with no exits, and so Φ1(Sµ) = sµ is a partial unitary with spectrum equal to T. And since Φ1(Pv) 6= 0 for all v ∈ E0, by [Szy02, Theorem 1.2], we have that Φ1 is injective. 14 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ Similarly, there exists an injective ∗-homomorphism Φ2 : C∗(F ) → C∗(Ff,S2 ) such that Φ2(Qv) = qv and Φ2(Te) = te for all v ∈ F 0 and e ∈ F 1. which we denote by D(A) and Set A = Φ1(C∗(E)) and B = Φ2(C∗(F )). Note that A is the C∗-subalgebra of C∗(Ef,S1 ) generated by (cid:8)pv, se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) and B is the C∗-subalgebra of C∗(Ff,S1 ) generated by(cid:8)qv, te(cid:12)(cid:12) v ∈ F 0, e ∈ F 1(cid:9). Moreover, µ(cid:12)(cid:12) µ ∈ E∗(cid:9) µ(cid:12)(cid:12) µ ∈ F ∗(cid:9) which we denote by D(B). Therefore, Φ induces a ∗-isomorphism eΦ : A → B such that eΦ(D(A)) = D(B) and eΦ(pv) = qw(v) for all v ∈ S1. Φ1(D(E)) = span(cid:8)sµs∗ Φ2(D(F )) = span(cid:8)tµt∗ f,S1 , set pv =eΦ(pv) and for e ∈ E1 For v ∈ E0 f,S1, set if e ∈ E1 if e = e(v) for some v ∈ S1. se =(eΦ(se) One can check that(cid:8)pv, se(cid:12)(cid:12) v ∈ E0 te(w(v)) f,S1 , e ∈ E1 f,S1(cid:9) is a Cuntz-Krieger Ef,S1-family in C∗(Ff,S2 ). Hence, there exists a ∗-homomorphism Ψ : C∗(Ef,S1 ) → C∗(Ff,S2) such that Ψ(pv) = pv and Ψ(se) = se for all v ∈ E0 f,S1 . We claim that Ψ is a ∗-isomorphism. f,S1 and e ∈ E1 First we show that Ψ is surjective. Let w ∈ F 0 f,S2 . Since F 0 f,S2 = F 0, we have that w ∈ F 0. Thus, qw is in the image of eΦ and hence qw is in the image of Ψ. f,S2. Suppose e ∈ F 1. Then te is in the image of eΦ which implies that Let e ∈ F 1 te is in the image of Ψ. Suppose e = e(z) for some z ∈ S2. Since w : S1 → S2 is a bijection, z = w(v) for some v ∈ S1. Hence, Ψ(se(v)) = te(w(v)) = te(z) = te. Thus, Ψ is surjective. To show that Ψ is injective, we will first show that for every vertex-simple cycle µ in Ef,S1 with no exits, Ψ(sµ) is a partial unitary with spectrum equal to T. Let µ be a vertex-simple cycle in Ef,S1 with no exits. Note from the construction of Ef,S1, µ is either a vertex-simple cycle in E with no exits or µ = e(v) for some v ∈ S1. Suppose µ is a vertex-simple cycle in E with no exits. Then Ψ(sµ) =eΦ(sµ), and since eΦ is a ∗-isomorphism, Ψ(sµ) =eΦ(sµ) is a partial unitary with spectrum equal to T. Suppose µ = e(v) for some v ∈ S1. Then Ψ(sµ) = te(w(v)) which is a partial unitary with spectrum equal to T. Since Ψ(pv) 6= 0 for all v ∈ E0 f,S1, by [Szy02, Theorem 1.2] Ψ is injective. We have just shown that Ψ is a ∗-isomorphism. We are left with showing that Ψ(D(Ef,S1 )) = D(Ff,S2 ). Noting that sµe(v)n s∗ µ for all v ∈ S1 and µ ∈ E∗ with r(µ) = v, we have that D(Ef,S1 ) = D(A). Similarly, D(Ff,S2 ) = D(B). (cid:3) It is now clear that Ψ(D(Ef,S1 )) = D(Ff,S2 ) since eΦ(D(A)) = D(B). To use the above result to prove that an orbit equivalence preserving periodic points implies diagonal preserving isomorphism, we must show that we may reduce the problem to graphs satisfying the property that all vertex-simple cycles with no exits are loops. These are the loops that we will unplug, and then plug again. µe(v)n = sµs∗ Proposition 5.2. Let E be a graph. Then there exists a graph F such that each vertex-simple cycle in F with no exits is a loop and there exists a diagonal preserving ∗-isomorphism Ψ : C∗(F ) → C∗(E). Proof. For each cycle µ, let Vµ be the vertices that support the cycle µ. If µ and ν are vertex-simple cycles in E with no exits, then Vµ = Vν if and only if Vµ ∩ Vν 6= ∅. Define a relation ≈ on the vertex-simple cycles in E with no exits by DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 15 µ ≈ ν if Vµ = Vν . Clearly, ≈ is an equivalence relation, and we may pick a set F of representatives of the equivalence classes. Then F ⊆ E∗ such that (a) For each µ ∈ F , µ is a vertex-simple cycle with no exits; (b) For each µ, ν ∈ F , Vµ ∩ Vν 6= ∅ if and only if µ = ν; and (c) For each vertex-simple cycle µ in E with no exits, there exists ν ∈ F such that Vµ = Vν . sF (ν) = sE(ν) for all ν ∈ F . It is clear from the construction of F , that each vertex-simple cycle in F with no exits is a loop. In fact, the vertex-simple cycles with no exits are {ν ν ∈ F}. Set S = (cid:8)e ∈ E1(cid:12)(cid:12) s(e) = s(ν) for some ν ∈ F(cid:9). Define F by F 0 = E0, F 1 = (cid:0)E1\S(cid:1) ⊔ {ν ν ∈ F}, and rF E1\S = rEE1\S, sF E1\S = sEE1\S, rF (ν) = Let(cid:8)pv, se(cid:12)(cid:12) v ∈ E0, e ∈ E1(cid:9) be a generating Cuntz-Krieger E-family in C∗(E) , and let(cid:8)qv, te(cid:12)(cid:12) v ∈ F 0, e ∈ F 1(cid:9) be a generating Cuntz-Krieger F -family in C∗(F ). shows that(cid:8)Qv, Te(cid:12)(cid:12) v ∈ F 0, e ∈ F 1(cid:9) is a Cuntz-Krieger F -family in C∗(E). Hence, We now define a Cuntz-Krieger F -family in C∗(E). For each v ∈ F 0, set Qv = pv and for each e ∈ E1\S, set Te = se. For ν ∈ F , set Tν = sν. A computation there exists a ∗-homomorphism Ψ : C∗(F ) → C∗(E) such that Ψ(qv) = Qv and Ψ(te) = Te for all v ∈ F 0, e ∈ F 1. Since the only vertex-simple cycles of F with no exits are the ν's and Ψ(tν ) = sν where ν is a vertex-simple cycle with no exits, Ψ(tν) is a partial unitary with spectrum equal to T. Since Ψ(qv) = pv 6= 0 for all v ∈ F 0, by [Szy02, Theorem 1.2], Ψ is injective. We now show that Ψ is surjective. Clearly, pv, se ∈ im Ψ, for all v ∈ E0 and e ∈ E1\S. For each ν ∈ F , s−1 E (sE(ν)) = {eν} for some eν ∈ E1 since ν is a vertex-simple cycle with no exits. So S = {eν ν ∈ F}. Let ν ∈ F . If ν = 1 then ν = eν so seν = Tν ∈ im Ψ. Assume ν ≥ 2. Then ν = eνµ where µ = e1 · · · eν−1 with each ei not an element of S. Hence, sei = Tei ∈ im Ψ for each i. Since ν is a vertex-simple cycle with no exits, we have that seis∗ ei = psE (ei). We now have seν = seν psE (e1) = seν se1 s∗ e1 e1 = seν se1 se2 s∗ e2 s∗ e1 = seν se1 psE (e2)s∗ ... = seν sµs∗ µ = sνs∗ µ = Ψ(tνtµ) ∈ ψ(C∗(F )). Therefore, Ψ is surjective, and hence Ψ is a ∗-isomorphism. It is clear that Ψ(D(F )) ⊆ D(E). Let µ ∈ E∗. Note that sµs∗ ν for a path ν ∈ E∗ where all edges of ν are not in S. Indeed, if µ = νe1 · · · en with e1 ∈ S, e2, . . . , en ∈ E1, and ν ∈ E∗ where all edges of ν are not in S, then s(ei)E1 = {ei} for all i since e1 ∈ S. So pr(ν) = se1···en s∗ ν. Therefore, Ψ(tνt∗ µ. Hence, Ψ(D(F )) ⊇ D(E), so Ψ is diagonal preserving. (cid:3) e1···en , hence sµs∗ ν) = sν s∗ ν = sµs∗ µ = sνs∗ µ = sν s∗ We provide an example to illustrate the construction in the proof of Proposi- tion 5.2. Let E denote the left-most graph below. Then E contains four vertex- simple cycles with no exits. Let F = {ν} with ν = e3e4e1e2. Then S = {e3} and 16 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ F is the right-most graph below. • • • e4 ? _❃❃❃❃❃❃❃❃ e3 E : • • F : • • • e4 ? ν • e1 ❃❃❃❃❃❃❃❃ e2 e1 ❃❃❃❃❃❃❃❃ e2 • . We are now ready to prove our main result. Theorem 5.3. Let E and E′ be graphs. Then the following are equivalent. (1) There exists a diagonal preserving ∗-isomorphism Ψ : C∗(E) → C∗(E′). (2) GE and GE ′ are isomorphic. (3) There exists a homeomorphism κ : ∂E → ∂E′ such that κ ◦ PE ◦ κ−1 = PE ′ and for each x ∈ ∂Eiso, x is eventually periodic if and only if κ(x) is eventually periodic. (4) There exists an orbit equivalence from E to E′ preserving periodic points. Proof. The equivalence of (1) and (2) is [BCW14, Theorem 5.1 (1) ⇐⇒ (2)]. The equivalence of (3) and (4) follows from Proposition 3.8. We are left to showing that (1) and (4) are equivalent. 1 such that all vertex-simple cycles in E1 and E′ By Theorem 4.3, we get that (1) implies (4). We now prove (4) implies (1). We will first show that we may assume that all vertex-simple cycles in E and E′ with no exits are loops. Indeed, by Proposition 5.2, there are graphs E1 and E′ 1 with no exits are loops, and diagonal preserving ∗-isomorphisms from C∗(E) to C∗(E1) and from C∗(E′) to C∗(E′ 1). By Theorem 4.3, there exists an orbit equivalence from E to E1 preserving periodic points and there exists an orbit equivalence from E′ to E′ 1 preserving periodic points. By Proposition 3.9, there exists an orbit equivalence from E1 to E′ 1 preserving periodic points if and only if there is an orbit equivalence from E to E′ preserving periodic points. And clearly, there is a diagonal preserving ∗-iso- 1) if and only if there is one from C∗(E) to C∗(E′). morphism from C∗(E1) to C∗(E′ This establishes the claim. Assume all vertex-simple cycles in E and E′ with no exits are loops, and that there exists an orbit equivalence β from E to E′ preserving periodic points. Let g → ∂E′ be the orbit equivalences provided in Propo- κE : ∂Eg → ∂E and κE ′ : ∂E′ sition 3.11. Then by Propostion 3.9, λ = κ−1 g is an orbit equiv- E ′ ◦β◦κE : ∂Eg → ∂E′ alence. Let V =nv ∈ E0 cycle(cid:12)(cid:12)(cid:12) λ(v) ∈ (∂E′ g)≥1o. Set F = λ(V ). By Lemma 2.7, V g. Let v ∈ E0 is closed in ∂Eg. Since λ is a homeomorphism, we have that F is closed in ∂E′ Hence, F satisfies (3) in Lemma 3.12. cycle. Then, κE(v) = e∞ v . Since β is an orbit equivalence preserving cycle and µ ∈ (E′)∗ with no edges equal to periodic points, there exist w ∈ (E′)0 g)∗ with ew and rE ′ (µ) = w such that β(e∞ rE ′ (µ) ∈ (E′)0 w ) = µ. Hence, g(F ) ⊆ (E′ sink, rE ′ so F satisfies (1) in Lemma 3.12. A similar argument using λ−1 shows that rEg(λ−1((E′)0 cycle. Therefore, λ(v) = κ−1 w . So, in particular, µ ∈ (E′ cycle. In particular, F ⊆ (E′ E ′ (β(κE(v))) = κ−1 E ′ (µe∞ cycle)) ⊆ (E′)0 g)≥1 with rE ′ v ) = µe∞ g(λ(E0 cycle)) ⊆ E0 cycle. g)0 ? _ ? $ $ DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 17 Let v ∈ (E′)0 cycle. Then λ−1(v) = µ ∈ E∗ g with rE(µ) = w ∈ E0 v . By [BCW14, Lemma 3.5], there exist n, m ∈ N0 such that β(µe∞ w ) = e∞ cycle and E ′ (β(µe∞ As σn edges in ν equal to ev. Hence λ(w) = ν with rE ′ w )) = e∞ v , β(e∞ g (ν) = v, so E ′ (β(σµ σn E (µe∞ w ) = νe∞ v E ′ (β(µe∞ w ))) = σn for some ν ∈ (E′)∗ with rE ′ (ν) = v and no w )). rE ′ g(λ(rEg (λ−1(v)))) = rE ′ g (λ(rEg (µ))) = rE ′ g(λ(w)) = v. Applying this to rEg (λ−1((E′)0 and conclude that (E′)0 cycle = rE ′ g (λ(E0 cycle)). cycle)) ⊆ E0 cycle we see that (E′)0 cycle ⊆ rE ′ g(λ(E0 cycle)) g(λ(v1)) = rE ′ w1 and β(e∞ We now show that F satisfies (2) in Lemma 3.12. Let v1, v2 ∈ E0 g(λ(v2)). We will show that λ(v1) = λ(v2). Then β(e∞ cycle and assume that rE ′ v1 ) = cycle, µi ∈ (E′)∗, no edges in µi are µ1e∞ g(λ(vi)) = wi. equal to ewi and rE ′ (µi) = wi. So, λ(vi) = µi, which implies that rE ′ So w1 = w2. Since β is an orbit equivalence, by [BCW14, Lemma 3.5], there exist n1, m1, n2, m2 ∈ N0 such that w2 where wi ∈ (E′)0 v2 ) = µ2e∞ E (β−1(σµi σni F (µie∞ w1))) = σmi E (β−1(µie∞ w1)) for all i ∈ {1, 2}. Thus, v1 = σn2 e∞ E (e∞ v1 ) = σn1+n2 E (β−1(e∞ w1 )) = σn1 E (e∞ v2 ) = e∞ v2 . g(λ(E0 g → ∂E′ Let κ : ∂E′ g (λ(v)) /∈ rE ′ g. Let v ∈ E0 cycle)) = (E′)0 cycle, we have that γ(E0 cycle. Either way, γ(v) = κ(λ(v)) = rE ′ cycle. If λ(v) ∈ F then κ(λ(v)) = rE ′ g (F ) as we saw above that rE ′ This implies that v1 = v2. We have just shown that F satisfies (2) in Lemma 3.12. g be the orbit equivalence given in Lemma 3.12 for the graph E′ g and the set F . Then γ = κ ◦ λ is an orbit equivalence from ∂Eg to ∂E′ g(λ(v)). If λ(v) /∈ F then g (λ(w)) im- g(λ(v)) = rE ′ λ(v) = rE ′ plies v = w for all w ∈ E0 g (λ(v)). Since sink. Since rE ′ Eg and E′ g are graphs satisfying Condition (L), by Theorem 3.10, there exists an isomorphism ϕ : GEg → GE ′ g such that ϕ∂Eg = γ. By [BCW14, Proposi- tion 2.2], this isomorphism of groupoids induces a ∗-isomorphism Φ : C∗(Eg) → C∗(E′ cycle, where (cid:8)pv, se(cid:12)(cid:12) v ∈ E0 (cid:8)qv, te(cid:12)(cid:12) v ∈ (E′ g(cid:9) is a Cuntz-Krieger Eg-family generating C∗(Eg) and g)1(cid:9) is a Cuntz-Krieger E′ g). ∼= E′, Proposition 5.1 implies that there Since (Eg)f,E0 exists a diagonal preserving ∗-isomorphism Ψ : C∗(E) → C∗(E′). Hence (4) implies (1). (cid:3) g) and Φ(pv) = qγ(v) for all v ∈ E0 g) such that Φ(D(Eg)) = D(E′ g-family generating C∗(E′ g, e ∈ E1 g)0, e ∈ (E′ cycle) = (E′)0 ∼= E and (E′ cycle ⊆ (E′ g)f,(E ′)0 g)0 cycle cycle We will denote the C∗-algebra of compact operators on ℓ2(N) by K and the maximal abelian subalgebra of K consisting of diagonal operators by C. For a commutative ring R with identity, we write M∞(R) for the ring of finitely supported, countably infinite square matrices over R and D∞(R) for the abelian subring of M∞(R) consisting of diagonal matrices. We write (C∗(E), D(E)) ∼= (C∗(F ), D(F )) if there exists a diagonal preserv- ing ∗-isomorphism Ψ from C∗(E) to C∗(F ), and write (C∗(E) ⊗ K, D(E) ⊗ C) ∼= (C∗(F ) ⊗ K, D(F ) ⊗ C) if there exists a ∗-isomorphism from Ψ from C∗(E) ⊗ K to C∗(F ) ⊗ K such that Ψ(D(E) ⊗ C) = D(F ) ⊗ C. Let R be a commutative ring with identity. If there is a ring ∗-isomorphism Ψ from LR(E) to LR(F ) such that Ψ(DR(E)) = DR(F ), then we write (LR(E), DR(E)) ∼= (LR(F ), DR(F )). 18 SARA E. ARKLINT, SØREN EILERS, AND EFREN RUIZ Similarly, if there is a ring ∗-isomorphism Ψ from LR(E) ⊗ M∞(R) to LR(F ) ⊗ M∞(R) such that Ψ(DR(E) ⊗ D∞(R)) = DR(F ) ⊗ D∞(R), then we write (LR(E) ⊗ M∞(R), DR(E) ⊗ D∞(R)) ∼= (LR(F ) ⊗ M∞(R), DR(F ) ⊗ D∞(R)). Let R be the full equivalence relation on N × N. We can regard R as a discrete principal groupoid with unit space N. Corollary 5.4. Let E and F be graphs, and let R be a commutative integral domain with 1. The following are equivalent: (1) (C∗(E) ⊗ K, D(E) ⊗ C) ∼= (C∗(F ) ⊗ K, D(F ) ⊗ C); (2) (LR(E)⊗ M∞(R), DR(E)⊗ D∞(R)) ∼= (LR(F )⊗ M∞(R), DR(F )⊗ D∞(R)); (3) (C∗(SE), D(SE)) ∼= (C∗(SF ), D(SF )); (4) (LR(SE), DR(SE)) ∼= (LR(SF ), DR(SF )); (5) GE × R ∼= GF × R; (6) GSE ∼= GSF ; (7) There exists an orbit equivalence from SE to SF preserving periodic points. Proof. By [CRS16, Theorem 4.2], (1) through (6) are equivalent. (6) ⇐⇒ (7) follows from Theorem 5.3 for the graphs SE and SF . (cid:3) We end by noting that the results above combine with [ERRS16] to completely re- solve the relationship between orbit and flow equivalence for countable shift spaces. Corollary 5.5. Let E and F be finite graphs with no sinks and sources such that the edge shift spaces XE and XF are countable sets. (1) If E and F are orbit equivalent, then XE and XF are flow equivalent. (2) SE and SF are orbit equivalent if and only if XE and XF are flow equiva- lent. Proof. Suppose E and F are orbit equivalent. Since E and F have no sinks, there exists an orbit equivalence from E to F preserving periodic points. Hence, by Theorem 5.3, C∗(E) ∼= C∗(F ). One easily sees by contradiction that the asserted countability translates to the condition that every vertex of E and F either supports exactly one return path or does not support a return path. Hence, [ERRS16, Theorem 7.1 (3) =⇒ (4)] applies, and thus the shift spaces XE and XF are flow equivalent, proving (1). For (2), we note that if SE and SF are orbit equivalent then by Corollary 5.4, C∗(E)⊗ K ∼= C∗(F )⊗ K, and the forward implication follows as above. In the other direction, suppose XE and XF are flow equivalent. By [ERRS16, Lemma 5.1], then E and F are Move equivalent. By [CRS16, Corollary 4.8], there exists a diagonal preserving isomorphism from C∗(E) ⊗ K to C∗(F ) ⊗ K. Hence, by Corollary 5.4, SE and SF are orbit equivalent. (cid:3) 6. Acknowledgements This work was partially supported by the Danish National Research Founda- tion through the Centre for Symmetry and Deformation (DNRF92), by VILLUM FONDEN through the network for Experimental Mathematics in Number Theory, Operator Algebras, and Topology, and by a grant from the Simons Foundation (# 279369 to Efren Ruiz). This work was completed while all three authors were attending the research program Classification of operator algebras: complexity, rigidity, and dynamics at the Mittag-Leffler Institute, January -- April 2016. We thank the institute and its staff for the excellent work conditions provided. The authors thank Aidan Sims for many helpful discussions. DIAGONAL PRESERVING ∗-ISOMORPHISMS OF GRAPH C ∗-ALGEBRAS 19 References [ADR00] C. Anantharaman-Delaroche and J. Renault, Amenable groupoids, Monographies de L'Enseignement Mathématique [Monographs of L'Enseignement Mathématique], vol. 36, L'Enseignement Mathématique, Geneva, 2000, With a foreword by Georges Skandalis and Appendix B by E. Germain. [BCW14] Nathan Brownlowe, Toke Meier Carlsen, and Michael F. Whittaker, Graph al- gebras and orbit equivalence, to appear in Ergod. Theory and Dynam. Systems, doi:10.1017/etds.2015.52 . [CRS16] Toke Meier Carlsen, Efren Ruiz, and Aidan Sims, Equivalence and stable isomorphism of groupoids, and diagonal-preserving stable isomorphisms of graph C ∗-algebras and Leavitt path algebras, eprint arXiv:1602.02602 (2016). Toke M. Carsen and Marius L. Winger, Orbit equivalence of graphs and isomorphism of graph groupoids, private communication, 2016. [CW16] [ERRS16] Søren Eilers, Gunnar Restorff, Efren Ruiz, and Adam P. W. Sørensen, Geometric classification of graph C ∗-algebras over finite graphs, eprint arXiv:1604.05439, 2016. [Mat13] Kengo Matsumoto, Classification of Cuntz-Krieger algebras by orbit equivalence of topological Markov shifts, Proc. Amer. Math. Soc. 141 (2013), 2329 -- 2342. [MM14] Kengo Matsumoto and Hiroki Matui, Continuous orbit equivalence of topological Markov shifts and Cuntz-Krieger algebras, Kyoto J. Math. 54 (2014), no. 4, 863 -- 877, doi:10.1215/21562261-2801849. Jean Renault, Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bull. (2008), no. 61, 29 -- 63. [Ren08] [Rør95] Mikael Rørdam, Classification of Cuntz-Krieger algebras, K-Theory 9 (1995), no. 1, 31 -- 58, doi:10.1007/BF00965458. [Szy02] Wojciech Szymański, General Cuntz-Krieger uniqueness theorem, Internat. J. Math. [Web14] [Yee07] 13 (2002), no. 5, 549 -- 555, doi:10.1142/S0129167X0200137X. Samuel B. G. Webster, The path space of a directed graph, Proc. Amer. Math. Soc. 142 (2014), no. 1, 213 -- 225, doi:10.1090/S0002-9939-2013-11755-7. Trent Yeend, Groupoid models for the C ∗-algebras of topological higher-rank graphs, J. Operator Theory 57 (2007), no. 1, 95 -- 120. Department of Mathematical Sciences, University of Copenhagen, Universitets- parken 5, DK-2100 Copenhagen, Denmark E-mail address: [email protected] Department of Mathematical Sciences, University of Copenhagen, Universitets- parken 5, DK-2100 Copenhagen, Denmark E-mail address: [email protected] Department of Mathematics, University of Hawaii, Hilo, 200 W. Kawili St., Hilo, Hawaii, 96720-4091 USA E-mail address: [email protected]
1106.3143	1	1106	2011-06-16T05:25:41	Ideals in Operator Space Projective Tensor Product of $C^*$-algebras	[ "math.OA" ]	For $C^$-algebras $A$ and $B$, we prove the slice map conjecture for ideals in the operator space projective tensor product $A \hat\otimes B$. As an application, a characterization of prime ideals in the Banach $\ast$-algebra $A\hat\otimes B$ is obtained. Further, we study the primitive ideals, modular ideals and the maximal modular ideals of $A\hat\otimes B$. It is also shown that the Banach $\ast$-algebra $A\hat\otimes B$ possesses Wiener property; and that, for a subhomogenous $C^$-algebra $A$, $A\hat\otimes B$ is symmetric.	math.OA	math	IDEALS IN OPERATOR SPACE PROJECTIVE TENSOR PRODUCT OF C ∗-ALGEBRAS RANJANA JAIN AND AJAY KUMAR Abstract. For C ∗-algebras A and B, we prove the slice map conjecture for ideals in the operator space projective tensor product A b⊗B. As an application, a characterization of prime ideals in the Banach ∗-algebra A b⊗B is obtained. Further, we study the primitive ideals, modular ideals and the maximal modular ideals of A b⊗B. It is also shown that the Banach ∗- algebra A b⊗B possesses Wiener property; and that, for a subhomogeneous C ∗-algebra A, A b⊗B is symmetric. 1. Introduction A systematic study of tensor products of subspaces and subalgebras of C ∗- algebras was initiated by Blecher and Paulsen [7], and Effros and Ruan [9, 10]. Analogous constructions to those of Banach spaces; for example, quotients, duals and tensor products were defined and studied. For a Hilbert space H, let B(H) denote the bounded operators on H. An operator space X on H is just a closed subspace of B(H). If E and F are operator spaces, then the operator space projective tensor product, denoted by Eb⊗F , is the completion of the algebraic tensor product E ⊗ F under the norm kuk∧ = inf{kαkkvkkwkkβk : u = α(v ⊗ w)β}, where the infimum runs over arbitrary decompositions with v ∈ Mp(E), w ∈ Mq(F ), α ∈ M1, pq, β ∈ Mpq, 1 with p, q ∈ N arbitrary; Mk, l being the space If E and F are C ∗-algebras, then Eb⊗F admits a of k × l matrices over C. Banach algebra with canonical isometric involution [16]. The main objective of this paper is to study the closed ∗-ideals of this Banach ∗-algebra. In Section 2, we study the slice map problem for ideals of Ab⊗B. Tomiyama [25] studied the slice maps on the tensor product of C ∗-algebras with respect to the 'min'-norm. Later, Wassermann [26] discussed the slice map problem in greater detail, which was then studied and used in different contexts - see, for instance, [2, 27]. It is interesting to know that the slice map property is not true for the 'min' norm for all C ∗-algebras. In fact, for the 'min' norm the slice map problem for ideals is equivalent to the problem of whether every tensor product A ⊗min B has Property F of Tomiyama [26, Remark 24]. In 1991, Smith [23] studied the slice map property for the Haagerup norm and proved that the slice map conjecture is true for all subspaces of B(H). We give 2000 Mathematics Subject Classification. 46L06,46L07,47L25. Key words and phrases. C ∗-algebras, Operator space projective tensor norm, Haagerup tensor product. 1 2 R. JAIN AND A. KUMAR an affirmative answer to the slice map conjecture for ideals with respect to the operator space projective tensor norm. The ideal structure for the Haagerup tensor product and the 'min' norm has been studied extensively in [1], [3] and [24]. In [16] and [14], the authors investigated some properties of the closed ideals of the projective tensor product In Section 3, we discuss a characterization of prime ideals, primitive ideals, and Ab⊗B, for example, sum of the product ideals, minimal and the maximal ideals. maximal modular ideals of the Banach ∗-algebra Ab⊗B. Finally, in Section 4, certain ∗-algebraic properties of Ab⊗B, namely, Wiener property and symmetry are studied. Throughout the paper, A and B will denote C ∗-algebras unless otherwise specified. Recall that the Haagerup norm of an element u in the algebraic tensor prod- uct A ⊗ B of two C ∗-algebras A and B is defined by kukh = inf{kΣi aia∗ i k1/2 kΣi b∗ i bik1/2 : u = Σn i=1ai ⊗ bi}. The Haagerup tensor product A ⊗h B is defined to be the completion of A ⊗ B in the norm k · kh. Also, the Banach space projective norm of u ∈ A ⊗ B is given by kukγ = inf{Σi kaikkbik : u = Σn i=1 ai ⊗ bi}. The norms k · kh, k · k∧ and k · kγ on the tensor product A ⊗ B of two C ∗-algebras A and B satisfy k · kh ≤ k · k∧ ≤ k · kγ. Necessary and sufficient conditions on A and B for the equivalence of these norms can be seen in [17]. 2. Slice Map Property for Ideals For each φ ∈ A∗, define a linear map Rφ : A ⊗ B → B by Rφ(Σn i=1 ai ⊗ bi) = Σn i=1 φ(ai)bi. Then, it can be easily seen that Rφ is well defined. Also, it is continuous with respect to the 'min'-norm [26] and hence for the larger operator space projective tensor norm with kRφk ≤ kφk; so, it can be extended to Ab⊗B as a bounded linear map and is known as the right slice map associated to φ. Similarly, one can define the left slice map Lψ for each ψ ∈ B∗. For a closed ideal J of B, Ab⊗J is a closed ideal of Ab⊗B [16] and clearly Rφ(x) ∈ J for all x ∈ Ab⊗J. We prove the converse of this statement which is known as the slice map problem for ideals. Lemma 2.1. The set {Rφ : φ ∈ A∗} is total on Ab⊗B, that is, if x ∈ Ab⊗B and Rφ(x) = 0 for all φ ∈ A∗, then x = 0. Proof. For φ ∈ A∗ and ψ ∈ B∗, consider φ ⊗ ψ : A ⊗ B → C given by (φ ⊗ ψ)(Σi ai ⊗ bi) = Σi φ(ai)ψ(bi). Note that, by the definition of the Banach space injective norm λ [24, page 188], we have Σi φ(ai)ψ(bi) ≤ kφkkψkkΣiai ⊗ bikλ. Thus φ ⊗ ψ is continuous with respect to larger norms, in particular, 'min'-norm and '∧'-norm; so, φ ⊗ ψ can be extended to continuous linear functionals on A ⊗min B and Ab⊗B. Let IDEALS IN O.S. PROJECTIVE TENSOR PRODUCT 3 us denote its extensions by φ ⊗min ψ and φb⊗ψ respectively. We claim that the set {φb⊗ψ : φ ∈ A∗, ψ ∈ B∗} is total on Ab⊗B. For this, consider an element x ∈ Ab⊗B such that (φb⊗ψ)(x) = 0, ∀φ ∈ A∗, ψ ∈ B∗. Observe that for the canonical map i : Ab⊗B → A ⊗min B, the maps φb⊗ψ and (φ ⊗min ψ) ◦ i both are continuous on Ab⊗B and agree on A ⊗ B, giving (φ ⊗min ψ)(i(x)) = 0 for all φ ∈ A∗, ψ ∈ B∗. Now, for faithful representations {πA, H} and {πB, K} of A and B respectively, for ξi ∈ H, ηi ∈ K, i = 1, 2; φ := hπA(·)ξ1, ξ2i ∈ A∗, ψ := hπB(·)η1, η2i ∈ B∗; so 0 = (φ ⊗min ψ)(i(x)) = h(πA ⊗ πB)(i(x))ξ1 ⊗ η1, ξ2 ⊗ η2i. This holds for all ξi ∈ H, ηi ∈ K; i = 1, 2 giving (πA ⊗ πB)(i(x)) = 0. Using the facts that πA ⊗ πB is faithful [24, Theorem IV.4.9], and that i is injective [13, Corollary 1] we obtain the claim. Finally, the relation hx, φb⊗ψi = hRφ(x), ψi = hLψ(x), φi, ∀x ∈ Ab⊗B, gives the required result. (cid:3) Recall that, for Banach spaces X and Y , a mapping θ : X → Y is said to be a quotient map if it maps the open unit ball of X onto that of Y [9]. Clearly, a quotient map is surjective, and for Banach space X and a closed subspace Y of X, the canonical quotient map π : X → X/Y is a quotient map in the above sense. Like in the case of Haagerup tensor product [1], the operator space projective tensor product of quotient maps behaves nicely. Although straight forward, we include a proof of the following for the sake of convenience: Lemma 2.2. Let I and J be closed ideals of the C ∗-algebras A and B, and π : A → A/I and ρ : B → B/J be the quotient maps. Then, (1) πb⊗ρ : Ab⊗B → (A/I)b⊗(B/J) is a quotient map with ker(πb⊗ρ) = Ab⊗J + Ib⊗B. (2) for a closed ideal K of Ab⊗B containing ker(πb⊗ρ), (πb⊗ρ)(K) is a closed ideal of (A/I)b⊗(B/J) with (πb⊗ρ)−1((πb⊗ρ)(K)) = K. Proof. (1) This follows directly from [14, Proposition 3.5]. (2) Consider an element (πb⊗ρ)(x) ∈ (A/I)b⊗(B/J) such that (πb⊗ρ)(x) ∈ cl((πb⊗ρ)(K)), where x ∈ Ab⊗B. Given an arbitrary ǫ > 0, there exists k ∈ K such that k(πb⊗ρ)(k − x)k(A/I) b⊗(B/J) < ǫ. Using part (1) above, there is an isomorphism between (Ab⊗B)/Z and (A/I)b⊗ (B/J), where Z = ker(πb⊗ρ). Therefore, k(k − x) + Zk(A b⊗B)/Z < cǫ, for some constant c. So, there exists some z ∈ Z ⊆ K with k(k + z) − xk(A b⊗B)/Z ≤ cǫ. Since K is closed and k + z ∈ K, we must have x ∈ K, 4 R. JAIN AND A. KUMAR which proves the claim. Finally, the equation in the statement is a routine verification. (cid:3) We are now prepared to present a proof of the slice map problem for ideals. Theorem 2.3. Let J be a closed ideal of B. Then Ab⊗J = {x ∈ Ab⊗B : Rφ(x) ∈ J for all φ ∈ A∗}. Lemma 2.2, corresponding to the quotient map π : B → B/J, we have another Proof. Consider an element x ∈ Ab⊗B such that Rφ(x) ∈ J for all φ ∈ A∗. From quotient map ib⊗π : Ab⊗B → Ab⊗(B/J) with ker(ib⊗π) = Ab⊗J, where 'i' is the identity map on A. Also observe that, by continuity and agreement on A ⊗ B, π ◦ Rφ = rφ ◦ (ib⊗π), where rφ : Ab⊗(B/J) → B/J is the right slice map. Using the fact that Rφ(x) ∈ J for all φ ∈ A∗, we obtain rφ(ib⊗π(x)) = 0 for all φ ∈ A∗. Thus, by Lemma 2.1, ib⊗π(x) = 0; so that x ∈ ker(ib⊗π) = Ab⊗J. The other containment is easy. (cid:3) We next give an application of Theorem 2.3 which will be used later to characterize the prime ideals. For the Haagerup norm such a result was proved for subspaces of B(H) in [23, Corollary 4.6]. Proposition 2.4. Let A1, A2 and B1, B2 be closed ideals of A and B, respec- tively. Then, (A1b⊗B1) ∩ (A2b⊗B2) = (A1 ∩ A2)b⊗(B1 ∩ B2). Proof. Since Aib⊗Bi, i = 1, 2 are closed ideals of Ab⊗B [16], it is easy to see that (A1 ∩ A2)b⊗(B1 ∩ B2) ⊆ (A1b⊗B1) ∩ (A2b⊗B2). For the other containment, consider an element v ∈ (A1b⊗B1)∩(A2b⊗B2). Then, Rφ(v) ∈ B1 ∩ B2 for all φ ∈ A∗; so, by Theorem 2.3, v ∈ Ab⊗(B1 ∩ B2). Next, consider any ψ ∈ (B1 ∩ B2)∗ and let ψ be an extension on B∗. Again, L ψ(v) ∈ (A1 ∩ A2) and Lψ(v) = L ψ(v); so that Lψ(v) ∈ (A1 ∩ A2). This is true for every ψ ∈ (B1 ∩ B2)∗; so, applying the slice map property once again for the left slice map, we obtain v ∈ (A1 ∩ A2)b⊗(B1 ∩ B2), which proves the claim. (cid:3) Using the slice map property for the right and the left slice maps, and the technique of extending linear functionals as done in Proposition 2.4, we can easily deduce the following: Corollary 2.5. For closed ideals I and J of A and B respectively, we have Ib⊗J = {x ∈ Ab⊗B : Rφ(x) ∈ J, Lψ(x) ∈ I; ∀ φ ∈ A∗, ∀ ψ ∈ B∗}. 3. Ideal Structure for Ab⊗B This section deals with the structure of prime ideals, primitive ideals and modular ideals of Ab⊗B which play an important role in determining the struc- ture of a Banach ∗-algebra. In a Banach algebra a proper closed ideal K is said to be prime if for any pair of closed ideals I and J satisfying IJ ⊆ K, either I ⊆ K or J ⊆ K. It is well known that a proper closed ideal K of a C ∗- algebra A is prime if and only if for any pair of closed ideals I and J satisfying IDEALS IN O.S. PROJECTIVE TENSOR PRODUCT 5 I ∩ J ⊆ K, either I ⊆ K or J ⊆ K. This property is also true for Ab⊗B as can be explicitly observed from the following result. The proof of the following result is largely inspired by [1]. Now consider the quotient maps π : A → A/E and ρ : B → B/F . Since Proof. Let K be a closed prime ideal. We can choose closed ideals E and F in Theorem 3.1. A closed ideal K in Ab⊗B is prime if and only if K = Ab⊗F + Eb⊗B for some prime ideals E and F in A and B respectively. A and B which are maximal with respect to the property Ab⊗F + Eb⊗B ⊆ K. ker(π ⊗ ρ) ⊆ K, by Lemma 2.2, (π ⊗ ρ)(K) is a closed ideal of A/Eb⊗B/F . We claim that (π ⊗ ρ)(K) = 0; this would imply K = Ab⊗F + Eb⊗B. If possible, let the ideal (π ⊗ ρ)(K) be non-zero. Then, it must contain a non-zero elementary tensor, say, π(a) ⊗ ρ(b), where a ⊗ b ∈ K [14, Proposition 3.7]. Let E0 and F0 be the closed ideals generated by a and b respectively. Then, the product ideal E0b⊗F0 is contained in K. Now, consider the product ideals M = Ab⊗(F + F0) and N = (E + E0)b⊗B. Using Proposition 2.4 and [14, Proposition 3.6], we have M N ⊆ M ∩ N = Eb⊗F + Eb⊗F0 + E0b⊗F + E0b⊗F0. It is clear that M N ⊆ K, so that either M ⊆ K or N ⊆ K. Using the maximality property of E and F , we have either E0 ⊆ E or F0 ⊆ F . Thus, either π(a) = 0 or ρ(b) = 0 contradicting the fact that (π ⊗ ρ)(a ⊗ b) 6= 0. Next we prove that E and F are prime ideals. Note that E and F both are proper ideals, K being proper. Let I ∩J ⊆ E for some closed ideals I and J of A. Then, (Ib⊗B)(J b⊗B) ⊆ (Ib⊗B)∩(J b⊗B) ⊆ K; so, either Ib⊗B ⊆ K or J b⊗B ⊆ K. Without loss of generality, let Ib⊗B ⊆ K. Consider any φ ∈ E⊥ ⊆ A∗ and 0 6= ψ ∈ F ⊥. Then, (φ ⊗ ψ)(K) = 0 which further gives (φ ⊗ ψ)(Ib⊗B) = 0. Since this is true for any φ ∈ E⊥, we must have I ⊆ E. Thus, E is prime and by a similar argument F is also prime. For the converse, let us assume that K = Ab⊗F + Eb⊗B for some prime ideals E and F in A and B respectively. Let IJ ⊆ K for some closed ideals I and J of Ab⊗B. Define the closed ideals M and N as M = cl(I + K) and N = cl(J + K). Then K ⊆ M, K ⊆ N and M N ⊆ K. We claim that either M = K or N = K, which further implies that either I ⊆ K or J ⊆ K. Suppose, on the contrary, that both the containments K ⊆ M and K ⊆ N are strict. We now claim that M contains a product ideal M1b⊗N1 which is not contained in K. As done previously, since K ( M, (π ⊗ ρ)(M ) is a non-zero closed ideal of A/Eb⊗B/F with (π⊗ρ)−1((π⊗ρ)(M )) = M . So, (π⊗ρ)(M ) contains a non-zero elementary tensor say π(a) ⊗ ρ(b). Define M1 and N1 to be the closed ideals generated by a and b. Then M1b⊗N1 is contained in M but not in K. Similarly, N contains a product ideal M2b⊗N2 which is not contained in K. By routine calculations, it is easily seen that M1M2b⊗N1N2 = cl((M1b⊗N1)(M2b⊗N2)) ⊆ cl(M N ) ⊆ K, which further gives π(M1M2) ⊗ ρ(N1N2) ⊆ (π ⊗ ρ)(M1M2b⊗N1N2) = {0}. 6 R. JAIN AND A. KUMAR So either M1M2 ⊆ ker π = E or N1N2 ⊆ ker ρ = F . Now, both E and F are prime, so at least one of the following containments must hold: M1 ⊆ E, M2 ⊆ E, N1 ⊆ F, N2 ⊆ F. In all these cases, either M1b⊗N1 or M2b⊗N2 is contained in K, which is a contradiction. Thus, K is prime. (cid:3) A closed ideal I of a Banach ∗-algebra E is said to be primitive if it is the kernel of an irreducible ∗-representation of E on some Hilbert space. The following gives a characterization of the primitive ideals of Ab⊗B. Theorem 3.2. For C ∗-algebras A and B, we have the following: Eb⊗B is also a primitive ideal of Ab⊗B. (1) If E and F are primitive ideals of A and B respectively, then Ab⊗F + (2) If K is a primitive ideal of Ab⊗B, then K = Ab⊗F + Eb⊗B for some (3) If A and B are separable, then K is primitive if and only if K = Ab⊗F + Eb⊗B for some primitive ideals E and F of A and B, respectively. prime ideals E and F of A and B, respectively. Proof. (1) Since E and F are primitive ideals, there exist irreducible ∗- repre- sentations π1 : A → B(H1) and π2 : B → B(H2) such that E = ker π1 and F = ker π2. Define π : A ⊗ B → B(H1 ⊗ H2) by π(a ⊗ b) = π1(a) ⊗ π2(b). Then, by the definition of min-norm [24], π is bounded with respect to the min-norm and hence the '∧' norm; so, π can be extended to Ab⊗B as a bounded ∗-representation. We first claim that π is irreducible, equivalently, π(Ab⊗B)′ = CI. Since π(Ab⊗B) ⊃ π1(A)⊗π2(B), we have π(Ab⊗B)′ ⊆ (π1(A)⊗π2(B))′, where ⊗ denotes the weak closure. Further, π1 and π2 being irreducible, π1(A) and π2(B) are non-degenerate ∗-subalgebras of B(H1) and B(H2), respectively; so that, by Double Commutant Theorem, π1(A) and π2(B) are weakly dense In particular, π1(A)⊗π2(B) = π1(A)′′ ⊗π2(B)′′; and, in π1(A)′′ and π2(B)′′. an appeal to Tomita's Commutation Theorem then yields (π1(A)⊗π2(B))′ = π1(A)′ ⊗π2(B)′ ⊆ CI, which shows that π is irreducible. are both contained in ker π; so that K ⊆ ker π. For the other containment, Next we claim that ker π = Ab⊗F + Eb⊗B = K(say). Clearly, Ab⊗F and Eb⊗B consider the quotient map θ : Ab⊗B → A/Eb⊗B/F with ker θ = K. Since, ker π contains ker θ, by Lemma 2.2, θ(ker π) is a closed ideal of A/Eb⊗B/F with θ−1(θ(ker π)) = ker π. If θ(ker π) 6= 0, then it must contain a non-zero elementary tensor say (a + E) ⊗ (b + F ) [14, Proposition 3.7]. Now a ⊗ b ∈ ker π implies π1(a) ⊗ π2(b) = 0, which further implies that either a ∈ E or b ∈ F , so that (a + E) ⊗ (b + F ) = 0, which is a contradiction. Thus, ker π ⊆ ker θ = K. (2) Let K = ker π for some irreducible ∗-representation π of Ab⊗B on H. By [24, Lemma IV.4.1], there exist commuting ∗-representations π1 : A → B(H) and π2 : B → B(H) such that π(a ⊗ b) = π1(a)π2(b), ∀ a ∈ A, b ∈ B. IDEALS IN O.S. PROJECTIVE TENSOR PRODUCT 7 Now, π(A ⊗ B) = π1(A)π2(B), so π(Ab⊗B) ⊆ cl(π1(A)π2(B)). Thus, we obtain ′ ′ ′ (π1(A)π2(B)) = cl(π1(A)π2(B)) ⊆ π(Ab⊗B) = CI. Also, note that π1 and π2 are both factor representations as for P = π1(A) and Q = π2(B) , we have ′′ ′′ ′ P ∩ P ′′ ′ ′ ′ ∩ π1(A) = π1(A) = (π1(A) ∪ π1(A)) ⊆ (π2(B) ∪ π1(A)) ⊆ {π1(A)π2(B)} = CI. ′ ′ (as π1(A) and π2(B) commute) Now, let E = ker π1 and F = ker π2. Then E and F , being kernels of factor representations, are both prime ideals [4, II.6.1.11]. Also, by the definition of ′′ π, Ab⊗F + Eb⊗B ⊆ K. For the reverse containment, consider a ⊗ b ∈ K. Then, we have π1(a)π2(b) = 0. Since π1(A) , using [24, Proposition IV.4.20], we see that either π1(a) = 0 or π2(b) = 0, i.e., a ⊗ b belongs to either Ab⊗F or Eb⊗B. In both cases, a ⊗ b ∈ Ab⊗F + Eb⊗B. Finally, exactly on the lines of (1), we conclude that K ⊆ Ab⊗F + Eb⊗B. (3) If A and B are separable, then every prime ideal is a primitive ideal. So, (cid:3) is a factor and π2(B) the result follows from parts (1) and (2). ⊆ π1(A) ′′ ′ In particular, among all the five proper closed ideals of B(H)b⊗B(H) - see [14, Theorem 3.12]- namely, {0}, B(H)b⊗K(H), K(H)b⊗B(H), B(H)b⊗K(H) + K(H)b⊗B(H) and K(H)b⊗K(H), the first four are prime as well primitive. We now discuss the modular ideals of Ab⊗B. In a Banach algebra A, an ideal I is said to be modular (or regular) if there exists an e ∈ A such that xe − x, ex − x ∈ I for all x ∈ A, or equivalently, if A/I is unital. It is clear that every proper ideal in a unital Banach algebra is modular. Also, {0} is modular if and only if A is unital. If I is a closed modular ideal of A, then the product ideal Ib⊗A need not be modular in Ab⊗A. This can be seen by considering A = C0(X), where X is a locally compact Hausdorff space (non-compact). A closed modular ideal of C0(X) is of the form I(E) = {f ∈ A : f (E) = 0}, where E is a compact subset of X [15]. So let us consider a closed modular ideal I = I(E) of A. Now note that Ib⊗A ⊆ Ab⊗A ⊆ A ⊗λ A = C0(X × X), we have the following result which characterizes the modular product ideals. where 'λ' is the Banach space injective tensor product. This shows that Ib⊗A ⊆ I(E × X). Thus, Ib⊗A is not modular, I(E × X) not being modular. In fact, Theorem 3.3. For closed modular ideals I and J of A and B respectively, Ib⊗J is modular in Ab⊗B if and only if both A and B are unital. Proof. If A and B are both unital, then so is Ab⊗B; so that every ideal is modular. Conversely, let Ib⊗J be a modular ideal. Since Ab⊗J and Ib⊗B both contain Ib⊗J, both are modular ideals of Ab⊗B. Using Lemma 2.2, we have an isomorphism between (Ab⊗B)/(Ab⊗J) and Ab⊗(B/J), and similarly between 8 R. JAIN AND A. KUMAR (cid:3) which further show that A and B are both unital [20, Theorem 1]. (Ab⊗B)/(Ib⊗B) and (A/I)b⊗B. Therefore, Ab⊗(B/J) and (A/I)b⊗B are unital In particular, K(H)b⊗K(H) is a closed modular ideal of B(H)b⊗B(H), but it is not modular in B(H)b⊗K(H). However, the maximal modular ideals behave well in Ab⊗B as can be seen in the following result: Theorem 3.4. A closed ideal K of Ab⊗B is maximal modular if and only if K = Ab⊗N + M b⊗B for some maximal modular ideals M and N of A and B, Proof. Let K be a maximal modular ideal of Ab⊗B. Since every maximal mod- ular ideal is also a maximal ideal, K is of the form K = Ab⊗N + M b⊗B for Now (Ab⊗B)/K is unital and is isomorphic to A/M b⊗B/N , by Lemma 2.2 ; some maximal ideals M and N of A and B respectively [14, Theorem 3.11]. respectively. therefore, the latter space is unital. But this implies that A/M and B/N are both unital [20, Theorem 1]. Thus, M and N are also modular ideals of A and B respectively. modular ideals of A and B respectively. Then, M and N being maximal, by For the converse, let K = Ab⊗N + M b⊗B, where M and N are maximal [14, Theorem 3.11], K is also a maximal ideal. Also, the facts that (Ab⊗B)/K and A/M b⊗B/N are isomorphic, and A/M and that B/N are both unital, together imply that Ab⊗B/K is unital, so that K is modular. (cid:3) 4. Wiener Property and Symmetry A Banach ∗-algebra is said to have Wiener property if every proper closed two-sided ideal is annihilated by an irreducible ∗-representation [22]. Wiener property for group algebras and the weighted group algebras has been studied in [12, 21] and others. It is well known that every C ∗-algebra has Wiener property. Theorem 4.1. The Banach ∗-algebra Ab⊗B has Wiener property. Proof. Consider a proper closed two-sided ideal J of Ab⊗B. Let Jmin denote the closure of i(J) in A ⊗min B, where i : Ab⊗B → A ⊗min B is the canonical homomorphism. By [16, Theorem 6], Jmin is also a proper closed two-sided ideal of the C ∗-algebra A ⊗min B, and so it is annihilated by an irreducible ∗-representation π : A ⊗min B → B(H). Note that the isometry of involution gives i is ∗-preserving, so that we have a ∗-representation π := π ◦ i of Ab⊗B on H. Using injectivity of i [13], we have π(J) = {0}. Also, the relation π(A ⊗ B) = π(A ⊗ B) gives π(Ab⊗B)′ ⊆ π(A ⊗ B)′ = π(A ⊗min B)′ = CI, where the equality between the middle expressions follows from the norm den- sity of π(A ⊗ B) in π(A ⊗min B). This further implies that π is irreducible; hence, Ab⊗B has Wiener property. (cid:3) A Banach ∗-algebra is said to be symmetric if every element of the form x∗x has positive spectrum, or equivalently, every self adjoint element has a real spectrum [22, Theorem 10.4.17]. Symmetry in group algebras has been IDEALS IN O.S. PROJECTIVE TENSOR PRODUCT 9 investigated by various authors, see, for instance, [21, 19]. One can easily verify that a Banach ∗-algebra A is symmetric if and only if for every left modular ideal I of A with modular unit α, the set SI of Hermitian sesquilinear forms given by SI = {B : A × A → C Bα = B, B(I, A) = {0}, B(u, u) ≥ 0, B(uw, vw) = B(v∗uw, w), ∀u, v, w ∈ A} is non-trivial, where Bα(v, w) := B(vα, wα), ∀v, w ∈ A [21]. It is well known that every C ∗-algebra is symmetric [22]. For C ∗-algebras A and B, we do not know whether the Banach ∗-algebra Ab⊗B is symmetric or not, but if one of them is subhomogeneous, then we have an affirmative answer. Recall that a C ∗-algebra A is subhomogeneous if there exists a positive integer n such that each irreducible representation of A has dimension less than or equal to n. We first modify a result from [16] in terms of operator algebras. We say that a Banach algebra A is an operator algebra if there exists a Hilbert space H and a bicontinuous homomorphism of A into B(H). Proposition 4.2. If A and B are operator algebras, then Ab⊗B is a Banach algebra. If A and B both have isometric involutions then Ab⊗B is a Banach ∗-algebra. Proof. It is known that if A is an operator algebra then the multiplication operator m : A ⊗h A → A given by m(a ⊗ b) = ab is completely bounded [6, Theorem 1.3]. Using this result, we get the completely bounded operators mA : A ⊗h A → A and mB : B ⊗h B → B. Now consider the canonical map i : Ab⊗A → A ⊗h A, which is a completely A : Ab⊗A → A, contractive homomorphism. Then, the multiplication operator m′ which can be regarded as m′ A = mA ◦ i, is completely bounded. Similarly, B : Bb⊗B → B is also completely bounded. In the multiplication operator m′ particular, the operator B : (Ab⊗A)b⊗(Bb⊗B) → Ab⊗B A ⊗ m′ m′ is bounded. Using the commutativity of '∧', the operator m′ A ⊗ m′ B : (Ab⊗B)b⊗(Ab⊗B) → Ab⊗B follows as in [16]. is also bounded. Hence, Ab⊗B is a Banach algebra. The proof for involution Lemma 4.3. Let A and B be C ∗-algebras with either A or B finite-dimensional. Then Ab⊗B is a symmetric operator algebra. Proof. If A or B is finite dimensional, then clearly, Ab⊗B is ∗-isomorphic to A ⊗min B, which gives the required result. (cid:3) Lemma 4.4. If A is a commutative unital C ∗-algebra and B is a symmetric unital operator algebra with isometric involution, then Ab⊗B is symmetric. (cid:3) 10 R. JAIN AND A. KUMAR denote the set of maximal ideals of A, then it is in one-one correspondence with the space of non-zero ∗-homomorphisms of A. For M ∈ Φ(A), define Proof. Note that Ab⊗B is a Banach ∗-algebra by Proposition 4.2. Let Φ(A) hM : A ⊗ B → B by hM (P ai ⊗ bi) = P ai(M )bi. It is bounded with respect to '∧'-norm, so can be extended to Ab⊗B as a ∗-homomorphism. Then, by [18, Corollary 2], an element x of Ab⊗B is invertible if and only if hM (x) is invertible for each maximal ideal M of A. Thus, σ(x) = [ M ∈Φ(A) σ(hM (x)), where σ(x) denotes the spectrum of x in Ab⊗B. Now consider a self-adjoint element u in Ab⊗B. For any M ∈ Φ(A), hM being ∗-preserving, hM (u) is self-adjoint in B. But B is symmetric, so σ(u) = [ σ(hM (u)) ⊆ R. M ∈Φ(A) (cid:3) Hence, Ab⊗B is symmetric. Remark 4.5. Note that one can also prove the above lemma using an argument similar to that in [8, Corollary 3.3]. Theorem 4.6. If A is a subhomogeneous C ∗-algebra, then for any C ∗-algebra B, Ab⊗B is symmetric. Proof. Since Ab⊗B can be isometrically embedded in A∗∗b⊗B∗∗ as a closed ∗- subalgebra, it is sufficient to show that A∗∗b⊗B∗∗ is symmetric. Let A be n- subhomogeneous, then A∗∗ is a direct sum of type Im von Neumann algebras for m ≤ n [4, Theorem IV.1.4.6]. Also each type Im von Neumann algebra is isomorphic to Mm⊗C, where Mm is the set of m × m complex matrices and C is a commutative von Neumann algebra [4, III.1.5.12]. Thus, A∗∗b⊗B∗∗ is ∗- isomorphic (not necessarily isometrically) to a direct sum of some Mm(C)b⊗B∗∗. For each m, Mm(C) is isomorphic to Mmb⊗C; so, using the commutativity and associativity of the operator space projective norm, we get Mm(C)b⊗B∗∗ is ∗- isomorphic to Cb⊗(Mmb⊗B∗∗). Note that, Lemma 4.3 gives Mmb⊗B∗∗ is an 4.4, Mm(C)b⊗B∗∗ is symmetric. Hence, A∗∗b⊗B∗∗ is symmetric being the direct Remark 4.7. If A is commutative and B is any C ∗-algebra, then, by [8, Corol- lary 3.3], A ⊗γ B is symmetric. However, the symmetry of A ⊗γ B when A is subhomogeneous and B is any C ∗-algebra follows as in Theorem 4.6. operator algebra with an isometric involution and is symmetric; so, by Lemma sum of symmetric Banach ∗-algebras [22, Theorem 11.4.2] (cid:3) References [1] S. D. Allen, A.M. Sinclair and R. R. Smith, The ideal structure of the Haagerup tensor product of C ∗-algebras, J. Reine Angew. Math. 442(1993), 111-148. [2] R. J. Archbold, A counter example for commutation in tensor products of C ∗-algebras, Proc. Amer. Math Soc. 81(4)(1981), 562-564. [3] R. J. Archbold, E. Kaniuth, G. Schlichting and D. W. B. Somerset, Ideal space of the Haagerup tensor product of C ∗-algebras, Internat. J. Math. 8(1997), 1-29. IDEALS IN O.S. PROJECTIVE TENSOR PRODUCT 11 [4] B. Blackadar, Operator algebras- Theory of C ∗-algebras and von-Neumann algebras, Springer-Verlag, 2006. [5] D. P. Blecher, Tensor products which do not preserve operator algebras, Math. Proc. Camb. Phil. Soc. 108(1990), 395-403. [6] D. P. Blecher and C. LeMerdy, On quotients of function algebras and operator algebra structures on l p, J. Operator Theory 34(1995), 315-346. [7] D. P. Blecher and V. I. Paulsen, Tensor product of operator spaces, J. Funct. Anal. 99(1991), 262-292. [8] R. A. Bonic, Symmetry in group algebras of discrete groups, Pacific J. Math 11(1)(1961), 73-94. [9] E. G. Effros and Z. J. Ruan, On matricially normed spaces, Pacific J. Math. 132(2)(1988), 243-264. [10] E. G. Effros and Z. J. Ruan, A new approach to operator space, Canad. Math. Bull. 34(3)(1991), 329-337. [11] U. Haagerup, The Grothendieck inequality for bilinear forms on C ∗-algebras, Advances in Math. 56(1985), 93-116. [12] W. Hauenschild, E. Kaniuth and A. Kumar, Ideal structure of Beurling algebras on [F C]− groups, J. Funct. Anal. 51(1983), 213-228. [13] R. Jain and A. Kumar, Operator space tensor products of C ∗-algebras, Math. Zeit. 260(2008), 805-811. [14] R. Jain and A. Kumar, Operator space projective tensor product: Embedding into second dual and ideal structure, arXiv:1106.2644v1 [math.OA] 14 Jun 2011. [15] E. Kaniuth, A course in commutative Banach algebras, Springer-Verlag, 2009. [16] A. Kumar, Operator space projective tensor product of C ∗-algebras, Math. Zeit. 237(2001), 211-217. [17] A. Kumar and A. M. Sinclair, Equivalence of norms on operator space tensor products of C ∗-algebras, Trans. Amer. Math. Soc. 350(1998), 2033-2048. [18] A. Lebow, Maximal ideals in tensor products of Banach algebras, Bull. Amer. Math. Soc. 74(11)(1968), 1020-1022. [19] H. Leptin and D. Poguntke, Symmetry and non symmetry for locally compact groups, J. Funct. Anal. 33(1979), 119-134. [20] R. J. Loy, Identities in tensor product of Banach algebras, Bull. Aust. Math. Soc. 2(1970), 253-260. [21] J. Ludwig, A class of symmetric and a class of Wiener group algebras, J. Funct. Anal. 31(1979), 187-194. [22] T. W. Palmer, Banach algebras and the general theory of ∗-algebras II, Cambridge Uni- versity Press, 2001. [23] R. R. Smith, Completely bounded module maps and the Haagerup tensor product, J. Funct. Anal. 102(1991), 156-175. [24] M. Takesaki, Theory of operator algebras I, Springer-Verlag, 2000. [25] J. Tomiyama, Applications of Fubini type theorem to the tensor products of C ∗-algebras, Tohoku Math J. 19(2)(1967), 213-226. [26] S. Wassermann, The slice map problem for C ∗-algebras, Proc. London Math. Soc. 32(3)1976, 537-559. [27] S. Wassermann, On tensor products of certain group C ∗-algebras, J. Funct. Anal. 23(1976), 239-254. Department of Mathematics, Lady Shri Ram College for Women, New Delhi- 110024, India. E-mail address: ranjanaj [email protected] Department of Mathematics, University of Delhi, Delhi-110007, India. E-mail address: [email protected]
1811.00456	1	1811	2018-11-01T15:52:03	Higher variations for free L\'evy processes	[ "math.OA", "math.PR" ]	For a general free L\'evy process, we prove the existence of its higher variation processes as limits in distribution, and identify the limits in terms of the L\'evy-It\^o representation of the original process. For a general free compound Poisson process, this convergence holds almost uniformly, This implies joint convergence in distribution to a $k$-tuple of higher variation processes, and so the existence of $k$-fold stochastic integrals as almost uniform limits. If the existence of moments of all orders is assumed, the result holds for free additive (not necessarily stationary) processes and more general approximants. In the appendix we note relevant properties of symmetric polynomials in non-commuting variables.	math.OA	math	HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES MICHAEL ANSHELEVICH, ZHICHAO WANG ABSTRACT. For a general free L´evy process, we prove the existence of its higher variation processes as limits in distribution, and identify the limits in terms of the L´evy-Ito representation of the original process. For a general free compound Poisson process, this convergence holds almost uniformly, This implies joint convergence in distribution to a k-tuple of higher variation processes, and so the existence of k-fold stochastic integrals as almost uniform limits. If the existence of moments of all orders is assumed, the result holds for free additive (not necessarily stationary) processes and more general approximants. In the appendix we note relevant properties of symmetric polynomials in non-commuting variables. 8 1 0 2 v o N 1 ] . A O h t a m [ 1 v 6 5 4 0 0 . 1 1 8 1 : v i X r a 1. INTRODUCTION A free (additive) L´evy process (in law; we will typically omit this qualifier) is a family of self- adjoint random variables {X(t) : t > 0} affiliated to a non-commutative probability space (A, τ ) which starts at zero, has free, stationary increments, and is stochastically continuous: free, (a) X(0) = 0, (b) For all n ∈ N and t0 < t1 < . . . < tn, {X(t0), X(t1) − X(t0), . . . , X(tn) − X(tn−1)} are (c) The distribution of the increment X(t + h) − X(t) depends only on h (and will be denoted (d) For all ε > 0, limh→0 µh(x > ε) = 0. µh), The distributions of increments of a free L´evy process form a semigroup with respect to the additive free convolution ⊞, and so are ⊞-infinitely divisible. This implies that the Voiculescu transform of the distribution µt of X(t) has the form (1) ϕµt(z) = tη + t a z + tZR(cid:20) z2 z − x − z − x1[−1,1](x)(cid:21) dρ(x), where η ∈ R, a ∈ R+, and ρ is a L´evy measure. Barndorff-Nielsen and Thorbjørnsen proved that a free L´evy process has a free L´evy-Ito decomposition. Theorem 1 (Theorems 6.4, 6.5 in [BNT05]). Let {X(t) : t > 0} be a free L´evy process, with the generating triple (η, a, ρ) as above. Then, X(t) is equal in distribution to a sum of three freely Date: November 2, 2018. 2010 Mathematics Subject Classification. Primary 46L54; Secondary 60F05, 60G51. This work was supported in part by a Simons Foundation Collaboration Grant. Parts of this article form part of the second author's 2018 Master's thesis at Texas A&M University. 1 2 MICHAEL ANSHELEVICH, ZHICHAO WANG independent parts. In general, (2) X(t) d= ηt1A0 + √aS(t) + lim ǫց0(cid:16)Z(0,t]×{x>ǫ} In particular, whenR[−1,1] xρ(dx) is finite and η := η −R 1 X(t) d= ηt1A0 + √aS(t) +Z(0,t]×R xdM(t, x) −Z(0,t]×{ǫ<x61} (3) −1 xρ(dx), then xdM(t, x). x(Leb ⊗ ρ)(dt, dx)1A0(cid:17). Here, S(t) is the free Brownian motion (in some W ∗-probability space (A0, τ 0)) and M is a free Poisson random measure on the measure space (R+ × R,B(R+ × R), Leb ⊗ ρ) with values in (A0, τ 0). The limit is taken in probability. In the representation in the theorem above, define the k'th variation of the process by (4) X (k)(t) = atδk,21A +Z(0,t]×R xkdM(t, x). We will show that these objects are well defined, and again form a free L´evy process. Later in the article we will define the corresponding object when xk is replaced by a more general function p(x). Our first main result concerns convergence in distribution to a higher variation process. Theorem 2. For each N ∈ N, let {Xi,N : i ∈ N} be free, identically distributed, self-adjoint ran- dom variables affiliated to (A, τ ). Suppose that for t > 0, Then for each k, the limits being taken in distribution. lim N→∞ lim N→∞ Xi,N d= X(t). X k i,N d= X (k)(t), [N t]Xi=1 [N t]Xi=1 We next discuss joint convergence in distribution. In the non-commutative case, there is at this point no universally accepted definition of this notion. Recall the following. Definition 1. A family of self-adjoint operators (a1,N , . . . , ak,N ) affiliated to a non-commutative probability space (A, τ ) converges to (a1, . . . , ak) jointly in moments if for any non-commutative self-adjoint polynomial P (x1, . . . , xk), τ [P (a1,N , . . . , ak,N )] is well-defined and τ [P (a1,N , . . . , ak,N )] → τ [P (a1, . . . , ak)] The family converges jointly in distribution if for any P as above, P (a1,N , . . . , ak,N ) → P (a1, . . . , ak) in distribution (see [MS13] for a related notion). HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 3 Recall that convergence in distribution and convergence in moments coincide for bounded opera- tors, but in general neither implies the other. The next result applies to free additive processes whose increments are not necessarily stationary. Theorem 3. For each N ∈ N, let {Xi,N : i ∈ N} be free self-adjoint random variables affiliated to (A, τ ) all of whose moments are finite. Suppose that for t > 0, converges in moments to X(t) as N → ∞. Suppose in addition that Xi,N [N t]Xi=1 (5) NXi=1 τ [X k i,N ]2 → 0  [N t]Xi=1 Xi,N , [N t]Xi=1 X 2 i,N , . . . , [N t]Xi=1 as N → ∞, for all k. Then there exist free additive processes(cid:8)X (j)(t)(cid:9) such that we have joint convergence in moments X k i,N →(cid:0)X(t), X (2)(t), . . . , X (k)(t)(cid:1) i,N ] → 0 andPN i=1 τ [X 2 as N → ∞. Remark 1. For triangular arrays of centered random variables with finite variance, the standard condition for convergence is max16i6N τ [X 2 i,N ] 6 c < ∞, see for example Section 22 in [Lo`e77]. The assumption (5) is clearly significantly stronger. On the other hand, it is significantly weaker that assuming that all Xi,N are identically distributed. In the latter case, the result follows from the limit theorem 13.1 in [NS06], itself based on a result of Speicher [Spe90]. The second case where we can prove joint convergence is when individual convergence holds in probability. In the following theorem, we actually have almost uniform convergence. Recall that in the commutative case, by Egorov's theorem this mode of convergence corresponds to the conver- gence almost surely. Theorem 4. Let ρ be a finite probability measure, and X(t) =Z(0,t]×R x dM(t, x) the corresponding free compound Poisson process. Then for Xi,N = X( i N ) − X( i−1 N ), we have the limit being taken almost uniformly. lim N→∞ X k i,N = X (k)(t), [N t]Xi=1 We expect similar convergence, in probability, for general free L´evy processes. At this point we have the following partial result. 4 MICHAEL ANSHELEVICH, ZHICHAO WANG Theorem 5. Let {X(t) : t > 0} be a free L´evy process whose increments have symmetric distribu- tions, and Xi,N = X( i N ) − X( i−1 N ). Then lim N→∞ [N t]Xi=1 X 2 i,N = X (2)(t), the limit being taken in probability. Corollary 6. For a free compound Poisson process {X(t) : t > 0} and increments Xi,N as above, we have joint convergence in distribution  [N t]Xi=1 Xi,N , [N t]Xi=1 X 2 i,N , . . . , [N t]Xi=1 X k i,N →(cid:0)X(t), X (2)(t), . . . , X (k)(t)(cid:1) as N → ∞. Corollary 7. Let {Xi,N : 1 6 i 6 N, N ∈ N} be as in either Theorem 3 or Corollary 6. Then for t > 0, (6) lim N→∞ X16i(1),i(2),...,i(k)6[N t] i(1)6=i(2),i(2)6=i(3),...,i(k−1)6=i(k) Xi(1),N Xi(2),N . . . Xi(k),N = kXj=1 (−1)k−j Xm1,...,mj >1 m1+...+mj =k X (m1)(t) . . . X (mj )(t). Here under the assumptions of Theorem 3 the limit is in moments, while under the assumptions of Corollary 6 the limit is almost uniformly, and so also in distribution. It was shown in Proposition 1 of [Ans00] that for free L´evy processes with bounded, centered increments, the limits (in norm) of the left-hand side of (6) and of (7) X16i(1),i(2),...,i(k)6[N t] {i(1),i(2),...,i(k)}=k Xi(1),N Xi(2),N . . . Xi(k),N . coincide. These limits should be interpreted as the free stochastic integral Z[0,t]k dX(s1) . . . dX(sk). See the end of the introduction, and the appendix, for the explanation of why the expression (6) is more appropriate in the free case. Prior results. The initial motivation for our analysis was the article [AT86] by Avram and Taqqu. We briefly compare some of their results with ours; the reader should consult their article for more details. Let {X(t)} be a L´evy process, and define its higher variations pathwise using jumps. Note that such a definition is unavailable in the non-commutative case. On the other hand, while the HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 5 classical version of the representation (4) is surely known, we have not found it in the literature. Let {Xi,N : 1 6 i 6 N, N ∈ N} be a triangular array with i.i.d. rows, such that in distribution as N → ∞. Then a multivariate limit theorem implies that NXi=1 Xi,N → X(t) (8) [N t]Xi=1(cid:0)Xi,N , X 2 i,N , . . . , X k i,N(cid:1) →(cid:0)X(t), X (2)(t), . . . , X (k)(t)(cid:1) jointly in distribution. At this point, in the non-commutative case such a theorem is only available for convergence in moments. On the other hand, we actually prove Theorem 2 not just for powers but for polynomials, that is, linear combinations of powers. For commuting variables, convergence in distribution of linear combinations is equivalent to joint convergence in distribution (an easy exercise left to the reader). So the appropriate commutative analog of Theorem 2 also implies the joint convergence in (8). Next, recall that the elementary symmetric polynomial ek(x1, . . . , xN ) = X16i(1)<i(2)<...<i(k)6N xi(1)xi(2) . . . xi(k) is a polynomial Pk(p1, . . . , pk) in the power sum symmetric polynomials (the polynomial Pk can be written down explicitly). Consequently, X 16i(1)<i(2)<...<i(k)6[N t] converges in distribution as N → ∞. Its limit is naturally identified with the multiple integral xj i pj(x1, . . . , xN ) = NXi=1 Xi(1),N Xi(2),N . . . Xi(k),N = Pk [N t]Xi=1 Z06s1<s2<...<sk6t dX(s1) dX(s2) . . . dX(sk). Xi,N , [N t]Xi=1 X 2 i,N , . . . , [N t]Xi=1 X k i,N Note that as explained in the appendix, if the variables {xi} do not commute, ek is not a polynomial in the pj's. Its natural replacement in the non-commutative setting is ek(x1, . . . , xN ) = X16i(1),i(2),...,i(k)6N i(1)6=i(2),i(2)6=i(3),...,i(k−1)6=i(k) xi(1)xi(2) . . . xi(k) used in equation (6). Motivated by [RW97], the first author studied related objects in [Ans00], but only for the case of free L´evy processes with compactly supported distributions. We are not aware of other sources where these specific topics are studied in the free probability setting. See however the study of homogeneous sums in [DN14, Sim15]. 6 MICHAEL ANSHELEVICH, ZHICHAO WANG The article is organized as follows. After the introduction and background in Section 2, Section 3 treats, for general free L´evy processes, convergence in distribution to the higher variation pro- cesses, and their generalization from powers to more general continuous functions. The key result is Theorem 17. Section 4 treats joint convergence in moments for more general additive processes. Section 5 contains results about almost uniform convergence and convergence in probability, as well as an alternative definition of joint convergence in distribution for non-commuting variables. Finally, in the appendix we explain which symmetric polynomials in non-commuting variables can be expressed in terms of the basic power sum symmetric polynomials. Acknowledgements. The authors are grateful to Matthieu Josuat-Verg`es for the references in the Appendix. 2. BACKGROUND AND THE FREE POISSON RANDOM MEASURE 2.1. Unbounded Operators and Affiliated Operators. A W ∗-probability space is a pair (A, τ ), where A is a von Neumann algebra acting on a Hilbert space and τ is a faithful normal tracial state on A. Throughout most of the paper, we will work with possibly unbounded operators affiliated to A. A self-adjoint operator a is affiliated to A if all of its spectral projections are in A. Equivalently, for any bounded Borel function, f (a) ∈ A. We denote the collection of all self-adjoint operators affiliated to A by Asa. A general closed, densely defined operator a is affiliated to A if in its polar decomposition a = u a, we have u ∈ A and a ∈ Asa. The collection of all such operators is denoted by A. Murray and von Neumann [MVN36] proved that A is an algebra, that is, if a, b ∈ A, then a + b and ab are densely defined and closable, and their closures are in A. For a ∈ Asa, its distribution is the unique probability measure µa on R such that for any bounded Borel function, (9) τ [f (a)] =ZR f (x) dµa(x). Definition 2. ([BNT02]) Let (A, τ ) be a W ∗-probability space and (an)n∈N be a sequence of op- erators affiliated with A. We say that an → a in probability if an − a → 0 in distribution as n → ∞. Here, a := √a∗a, which is self-adjoint. When an and a are self-adjoint operators affiliated with A, an → a in probability if and only if an − a converges to zero in distribution, i.e. the distribution of an − a as a probability measure on R converges weakly to probability measure δ0. We list the following proposition for completeness. Compare with Proposition 2.18 in [BNT02] Proposition 8. The following are equivalent. (a) an → a in probability. (b) ∀ε > 0, the traces of the spectral projections τ [1(ε,∞)(an − a)] → 0. (c) Denote N (ε, δ) =nb ∈ A : ∃ projection p ∈ A such that τ [1 − p] < δ, bp ∈ A,kbpk < εo . Then ∀ε, δ > 0, for sufficiently large n, an − a ∈ N (ε, δ). This mode of convergence is also called convergence in measure. HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 7 2.2. Freely infinitely divisible distributions and limit theorems. As mentioned in the introduc- tion, a probability measure µ on R is ⊞-infinitely divisible if and only if its Voiculescu transform has a representation (10) ϕµ(z) = η + where η ∈ R, a ∈ R+, and ρ is a L´evy measure, that is, z − x − z − x1[−1,1](x)(cid:21) dρ(x), a z +ZR(cid:20) z2 ρ({0}) = 0 and ZR ϕµ(z) = γ +ZR min(1, x2) dρ(x) < ∞. 1 + xz z − x dσ(x). ϕµ also has an alternative representation (11) For future reference, we record the relation between the generating triple (a, η, ρ) and the generating pair (γ, σ) for the same measure µ: (12) and, conversely, (13) x2 1 + x2 ρ(dx) 1 σ(dx) = aδ0(dx) + γ = η −ZR x(cid:20)1[−1,1](x) − 1 + x2(cid:21) dρ(x)  a = σ({0}) η = γ +ZR\{0} ρ(dx) = 1 + x2 x2 1 + x2 x (cid:20)1[−1,1](x) − 1 1 + x2(cid:21) dσ(x) 1R\{0}(x)σ(dx).  The following fundamental limit theorem was proved by Bercovici and Pata in [BP99]. Theorem 9. For a sequence of probability measures {µn} and a strictly increasing sequence of positive integers (kn), the following assertions are equivalent: (a) the sequence of kn-fold free convolutions µ⊞kn n converges weakly to a probability measure µ; (b) there exist a finite positive Borel measure σ on R and a real number γ such that (14) (15) and kn x2 x2 + 1 dµn(x) w.→ dσ(x) lim n→∞ knZR x 1 + x2 dµn(x) = γ. The pair of parameters (γ, σ) comes from the Voiculescu transform (11) of µ. This also implies the ⊞-infinite divisibility of µ. 8 MICHAEL ANSHELEVICH, ZHICHAO WANG 2.3. Free Poisson Random Measures. Definition 3 (Free Poisson Random Measures). Let (Θ,E, ν) be a measure space and put E0 = {E ∈ E : ν(E) < ∞}. Let further (A, τ ) be a W ∗−probability space and let A+ denote the cone of positive operators in A. A free Poisson random measure on (Θ,E, ν) with values in (A, τ ) is a mapping M : E0 → A+ with the following properties: (a) the distribution of M(E) is a free Poisson distribution Poiss⊞(ν(E)); (b) for mutually disjoint sets A1, ..., An in E0, the random variables M(A1), M(A2), ..., M(An) are freely independent and M(∪n j=1 M(Aj). Here, the free Poisson distribution Poiss⊞(λ) is obtained by the limit in distribution of j=1Aj) =Pn (cid:18)(1 − )δ0 + λ N λ N δ1(cid:19)⊞N , as N → ∞ (see Lecture 12 in [NS06]). The existence of free Poisson random measures is proved by Barndorff-Nielsen and Thorbjørnsen in [BNT05]. For an alternative approach, see Remark 3 below. We next discuss integration with respect to a free Poisson random measure. Definition 4. Let s be a real-valued simple function in L1(Θ,E, ν) of the form s = Pr j=1 aj 1Ej , where aj ∈ R \ {0} and Ej are disjoint sets from E0. Then, we define the integral of s with respect to M as ZΘ rXj=1 sdM = ajM(Ej) ∈ A. ple function in L1(Θ,E, ν). Next, we can extend this integration to general functions in L1(Θ,E, ν). Lemma 10. [BNT05, Proposition 4.3] Let f be a real-valued function in L1(Θ,E, ν). Choose a sequence of real-valued simple functions (sn) in L1(Θ,E, ν) which satisfies the assumptions of converges in probability to a self-adjoint (possibly unbounded) operator affiliated with A. This operator is independent of the choice of approximating sequence (sn). We denote this operator by Because M(Ej) are positive in A, the elementRΘ sdM is self-adjoint in A, for any real-valued sim- the Dominated Convergence Theorem, such that sn(θ) → f (θ), for all θ ∈ Θ. Then, RΘ sndM RΘ f dM . The proof of the following lemma follows by the same techniques as Proposition 4.3 and Corol- lary 4.5 in [BNT05]. Lemma 11. Let f be a real-valued function in L1(Θ,E, ν). Choose a sequence of real-valued func- tions (fn) in L1(Θ,E, ν) which satisfies the assumptions of the Dominated Convergence Theorem, such that fn(θ) → f (θ), for all θ ∈ Θ. Then,RΘ fndM converges in probability toRΘ f dM . In fact, we only use a special measure space with a concrete intensity measure in our situation. Let D = R+ × R and B(D) be the set of all Borel subsets of D. In our case, (Θ,E, ν) = (D,B(D), Leb ⊗ ρ), HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 9 where ρ is a L´evy measure. The free Poisson random measure M that we will use is defined on (D,B(D), Leb ⊗ ρ) with values in a W ∗−probability space (A, τ ). Besides, the integration with respect to this free Poisson measure M we will use is also a special case. Lemma 12. Let ρ be a L´evy measure on the real line, and let M be a free Poisson random measure on (D,B(D), Leb ⊗ ρ) with values in the W ∗−probability space (A, τ ). Suppose that p(x) is any continuous function on R. (a) For any ǫ > 0 and 0 6 s < t < ∞, the integral Z(s,t]×{ǫ<x6n} p(x)M(dt, dx) converges in probability, as n → ∞, to some self-adjoint operator affiliated with A, which is denoted by p(x)M(dt, dx). Z(s,t]×{ǫ<x<∞} Z(s,t]×{x6n} (b) IfR[−1,1] p(x)ρ(dx) < ∞, then for any ǫ > 0 and 0 6 s < t < ∞, the integral p(x)M(dt, dx) converges in probability to some self-adjoint operator affiliated with A, as n → ∞. We denote it by Z(s,t]×R p(x)M(dt, dx). The statement of Lemma 12 is quite similar with Lemma 6.3 of [BNT05]. In the paper [BNT05], the authors only proved the situation when p(x) = x but their methods in Lemma 6.1 and Lemma 6.2 of [BNT05] still work well for Lemma 12. According to Lemma 6.3 of [BNT05], there are only two things for us to check. Since ρ is a L´evy measure, we have that Z(s,t]×{ǫ<x6n} p(x)Leb ⊗ ρ(du, dx) = (t − s)Z{ǫ<x6n} p(x)ρ(dx) < ∞. IfR[−1,1] p(x)ρ(dx) < ∞, we have that Z(s,t]×{x6n} p(x)Leb⊗ ρ(du, dx) = (t− s)(cid:20)Z{x61} p(x)ρ(dx) +Z{1<x6n} p(x)ρ(dx)(cid:21) < ∞. Thus, integralsR(s,t]×{ǫ<x6n} p(x)M(dt, dx) andR(s,t]×{x6n} p(x)M(dt, dx) are well-defined by Proposition 4.3 of [BNT05]. Then, we can copy the proof of Lemma 6.3 of [BNT05] and replace the function f (x) = x by arbitrary continuous function p(x) directly to prove Lemma 12. The idea for proving Lemma 6.3 is employing the Bercovici-Pata bijection to transform the statement into classical sense and then using Lebesgue's dominated convergence theorem. 10 MICHAEL ANSHELEVICH, ZHICHAO WANG 3. THE HIGHER VARIATIONS OF FREE L ´EVY PROCESSES Proposition 13. If there exist a finite Borel measure σ and a constant γ such that N x2 x2 + 1 dµN (x) w.→ dσ(x) lim N→∞ NZR x 1 + x2 dµN (x) = γ, (16) and (17) and w.→ µt, for any then there exists a family {µt}t>0 of probability measures on R such that µ⊞[N t] t ∈ [0,∞). Each µt is ⊞-infinitely divisible and its Voiculescu transform is ϕµt(z) = tγ + tRR Moreover, there exists a free L´evy process {X(t)}t>0 such that the distribution of each X(t) is µt, for all t > 0. 1+xz z−x dσ(x) = tϕµ(z), where µ := µ1 is the distribution of X(1). N Proof. By Theorem 9, we know that if there exist a finite Borel measure σ and a constant γ such that (16) and (17) hold, then µ⊞N N w.→ µ1. For any t ∈ [0,∞), we have that x2 dµN (x) w.→ tdσ(x) =: dσt(x) x2 + 1 [Nt] x [Nt]ZR lim N→∞ 1 + x2 dµN (x) = t lim 1 + x2 dµN (x) = tγ =: γt. w.→ µt. Accord- Therefore, for any t ∈ [0,∞), there exists a probability measure µt such that µ⊞[N t] ing to Theorem 9, for any t ∈ [0,∞), µt is ⊞-infinitely divisible since the Voiculescu transform of µt is N→∞ N x NZR ϕµt(z) = γt +ZR 1 + xz z − x dσt(x) = tϕµ(z), where µ := µ1. Therefore, ϕµt = ϕµt−s + ϕµs, when t > s > 0. In other words, µt = µt−s ⊞ µs. w.→ δ0, as t → 0. Then, by Remark 6.7 in Meanwhile, ϕµt → 0 when t → 0, which means µt [BNT05], we can conclude that there exists a free L´evy process {X(t)}t>0, which is a family of self-adjoint operators affiliated with some W ∗-probability space (A0, τ 0), such that the distribution of each X(t) is µt, for all t > 0. Lemma 14. Let (A, τ ) be a W ∗-probability space. Let a ∈ Asa with distribution µ, and p(x) be a continuous real-valued function. Then the distribution µ(p) of operator p(a) (obtained via continuous functional calculus) can be obtained by the following formula: (cid:3) ZR f (p(x))dµ(x) =ZR f (x)dµ(p)(x), for any bounded Borel function f : R → R. Proof. By definition of the distribution, for any bounded Borel function f : R → R, τ (f (a)) =ZR f (x)dµ(x). HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES Then, f ◦ p(x) is still a bounded Borel function. Thus, ZR f (p(x))dµ(x) = τ (f (p(a))) =ZR f (x)dµ(p)(x). 11 (cid:3) Generally, Lemma 14 shows how to change variables between different probability measures. Note the difference between the notation µ(p) in the preceding lemma and ρp in the following one. Lemma 15. Let p(x) be any real-valued continuous function such that p(0) = 0 and p′(0) exists. Suppose that M is a free Poisson random measure determined by a L´evy measure ρ on the Borel measure space (D,B(D), Leb ⊗ ρ) with values in some W ∗-probability space A. If ρp is another measure defined by (18) f (p(x))1R\{0}(p(x))dρ(x), ZR f (x)dρp(x) =ZR for any bounded Borel function f (x) on R, then ρp is a L´evy measure. The free Poisson random measure M (p) defined by ρp on (D,B(D), Leb ⊗ ρp) has the following relation with M : (19) for any t, ǫ > 0, and (20) Z(0,t]×{ǫ<x} Z(0,t]×R xdM (p)(t, x) d=Z(0,t]×{ǫ<p(x)} xdM (p)(t, x) d=Z(0,t]×R provided thatR[−1,1] xdρp(x) < ∞. p(x)dM(t, x), p(x)dM(t, x), ∀t > 0, Proof. Since p(0) = 0, there exists an ε > 0 such that p(x) 6 1 when x 6 ε. Since p′(0) exists, the function h(x) :=( p(x) x , x 6= 0 p′(0), x = 0, RR ZR is continuous on R. First, we show that ρp is a L´evy measure. If f (x) = 1{0}(x), then ρp({0}) = 1{0}(x)dρp(x) is zero by the definition (18). Next, if f (x) = min{1, x2}, then we can get the following conclusion: 1R\[−1,1](x)dρp(x) min{1, x2}dρp(x) =ZR =ZR 6Z{x∈R:ε<p(x)61} 6ZR\[−ε,ε] 1[−1,1](x)x2dρp(x) +ZR 1[−1,1]\{0}(p(x))(p(x))2dρ(x) +ZR p(x)2dρ(x) +Z ε −ε6x6εh(x)2Z 1 1dρ(x) + max −ε 1R\[−1,1](p(x))dρ(x) h(x)2x2dρ(x) +ZR\[−ε,ε] x2dρ(x) +ZR\[−ε,ε] −1 1dρ(x) 1dρ(x) < ∞. Therefore, ρp is a L´evy measure. 12 MICHAEL ANSHELEVICH, ZHICHAO WANG since Second, we show that the relation (20) holds. IfR[−1,1] xdρp(x) is finite, then so isR 1 −1 p(x)dρ(x) < ∞. 1[−1,1]\{0}(p(x)) · p(x)dρ(x) =Z 1 Z 1 −1 xdρp(x) =ZR −1 p(x)dρ(x), Thus, the right-hand side and left-hand side of (20) make sense by Lemma 12. According to Lemma 12, we only need to show that for all t > 0 and n ∈ N. For any N ∈ N, consider mutually disjoint intervals p(x)dM(t, x), Z(0,t]×{x:−n6x<n} EN N 2nm ,−n + 2n(m − 1) xdM (p)(t, x) d=Z(0,t]×{x:−n6p(x)<n} m =h−n + N (cid:17), (cid:19) 1 NXm=1(cid:18)−n + sN (x)dM (p)(t, x) →Z(0,t]×{x:−n6x<n} N m (x) EN where 1 6 m 6 N and m ∈ N. Then, the simple functions 2n(m − 1) sN (x) = converge to f (x) = x, for any x ∈ [−n, n) as N → ∞. Thus, Z(0,t]×{x:−n6x<n} in probability. Let xdM (p)(t, x) Then, ∪N m=1J N m = {x : −n 6 p(x) < n} and {J N m} are mutually disjoint. The simple functions J N m = {x : p(x) ∈ EN m}, (1 6 m 6 N, m ∈ N). for any x ∈ {x : −n 6 p(x) < n}, as N → ∞. Therefore, when N → ∞, in probability. We conclude that it suffices to show the equality in distribution m (x) → p(x) J N N (cid:19) 1 2n(m − 1) NXm=1(cid:18)−n + gN (x)dM(t, x) →Z(0,t]×{x:−n6p(x)<n} gN (x)dM(t, x) d=Z(0,t]×{−n6x<n} m . By Definition 4, we know that p(x)dM(t, x) sN (x)dM (p)(t, x). gN (x) = Z(0,t]×{x:−n6p(x)<n} Z(0,t]×{x:−n6p(x)<n} Z(0,t]×{x:−n6x<n} Z(0,t]×{x:−n6p(x)<n} m and K N Let F N m = (0, t] × EN m = (0, t] × J N sN (x)dM (p)(t, x) = and gN (x)dM(t, x) = NXm=1(cid:18)−n + NXm=1(cid:18)−n + 2n(m − 1) N 2n(m − 1) N m ), (cid:19) M (p)(F N (cid:19) M(K N m ). HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 13 By Definition 3, the distribution of M(K N Poiss⊞(tρp(EN m )). According to (18), we conclude that, when m 6= N 2 + 1, 0 /∈ EN m , so m ) is Poiss⊞(tρ(J N m )) and the distribution of M (p)(F N m ) is ρp(EN m ) =ZR =ZR =ZR 1 m (x)dρp(x) EN 1 m \{0}(p(x))dρ(x) EN 1 {x:p(x)∈EN m}(x)dρ(x) = ρ(J N m ). m )) = Poiss⊞(tρp(EN So, Poiss⊞(tρ(J N M(K N and n ∈ N, we can apply the same method and show that m )), m 6= N ) and M (p)(F N 1+ N 2 1+ N 2 2 + 1. Notice that the coefficients in front of ) are zero. Then, we get the final result (20). In general, for any t, ǫ > 0 Z(0,t]×{ǫ<x<n} xdM (p)(t, x) d=Z(0,t]×{ǫ<p(x)<n} p(x)dM(t, x), to prove equation (19). (cid:3) Lemma 16. Let p(x) be any real-valued continuous function such that p(0) = p′(0) = 0 and p′′(0) = 2c exists. Then whether or notR 1 Proof. Denote −1 xdρ(x) is finite,R 1 q(x) :=( p(x) x2 , x 6= 0 c, x = 0. −1 xdρp(x) is finite. Then q(x) is a continuous function. So we can check that 1[−1,1](x)xdρp(x) Z 1 −1 xdρp(x) =ZR =ZR =Z 1 6 kqkC([−1,1])Z 1 6 CZR −1 −1 1[−1,1]\{0}(p(x))p(x)dρ(x) 1[−1,1]\{0}(p(x))p(x)dρ(x) +ZR\[−1,1] x2dρ(x) +ZR\[−1,1] min{1, x2}dρ(x) < ∞. 1[−1,1]\{0}(p(x))p(x)dρ(x) 1{0<p(x)61}(x)p(x)dρ(x) (cid:3) Theorem 2 follows from the following more general result by taking p(x) = xk. Theorem 17. For each N ∈ N, let {Xi,N : i ∈ N} be free, identically distributed, self-adjoint random variables affiliated to (A, τ ). Suppose that for t > 0, [N t]Xr=1 XN,r d.→ X(t), 14 MICHAEL ANSHELEVICH, ZHICHAO WANG where X(t) is a free L´evy process. Let p(x) be any real-valued continuous function such that p(0) = 0, p′(0) = b, and p′′(0) = 2c. Then, there exists a L´evy process X p(t) such that (21) [N t]Xr=1 p (XN,r) d.→ X p(t). In addition, if X(t) has the L´evy-Ito decomposition (2) with the generating triple (η, a, ρ), then X p(t) has a representation in the form: (22) X p(t) d= bX(t) + act1A0 +Z(0,t]×R (p(x) − bx)dM(t, x), where M is a free Poisson random measure coming from the L´evy-Ito decomposition of X(t). This is the case whether or notR 1 −1 xdρ(x) is finite. In particular, if p′(0) = 0, X p(t) d= act1A0 +Z(0,t]×R p(x)dM(t, x). Proof. Let X(1) be generated by the pair (γ, σ). Let µN and µp p(XN,r) respectively. Recall that by Lemma 14, RR f (x)dµp real-valued and bounded Borel function f (x). Let N be the distributions of XN,r and N (x) = RR f (p(x))dµN (x), for any q(x) :=( p(x)−bx x2 c, x = 0. , x 6= 0 Then q(x) is a continuous function. Therefore, lim N→∞ = t lim N→∞ [Nt]ZR NhZR x bx N (x) = t lim N→∞ NZR 1 + x2 dµp 1 + x2 dµN (x) +ZR(cid:18) p(x) NZR 1 + x2 dµN (x), gp(x) x2 = tbγ + t lim N→∞ p(x) 1 + p(x)2 dµN (x) 1 + p(x)2 − bx 1 + x2(cid:19) dµN (x)i where gp(x) = p(x)+q(x)−b(b+xq(x))p(x) 1+p(x)2 ∈ Cb(R) and gp(0) = c. So γp is defined by 1 + x2 dµp where γ and σ are defined by (14) and (15). Define γp := lim N→∞ N (x) = bγ +ZR gp(x)dσ(x), x NZR h(x) :=( p(x) x , x 6= 0 b, x = 0. HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 15 Then, h(x) is a continuous function and xh(x) = p(x). For any f (x) ∈ Cb(R), [Nt]ZR f (x) x2 x2 + 1 dµp N (x) = [Nt]ZR = [Nt]ZR N→∞−→ tZR f (p(x)) p(x)2 p(x)2 + 1 dµN (x) f (p(x)) p(x)2 x2 x2 + 1 p(x)2 + 1 p(x)2 + h(x)2 p(x)2 + 1 f (p(x)) dσ(x). x2 + 1 x2 dµN (x) Let hp(x) := p(x)2+h(x)2 p(x)2+1 , which is a positive bounded Borel function on R. We denote by deσ(x) the measure hp(x)dσ(x). The measure dσp(x) is defined byRR f (x)dσp(x) =RR f (p(x))deσ(x), for any bounded Borel function f (x). Then, N x2 x2 + 1 dµp N (x) w.→ dσp(x), Borel measure. Thus, the conclusion (21) follows immediately from Theorem 9. By Proposition as N → ∞. Since σ is a finite positive Borel measure on R, we know that σp is also a finite positive 13, we know that {X p(t)}t>0 can be a free L´evy process affiliated with some W ∗-probability space. Denote the free generating triplet of X p(1) by (ap, ηp, ρp). Next, based on Theorem 1, L´evy process X p(t) has a decomposition in the form of (2) with free generating triplet (ap, ηp, ρp). Hence, to prove the representation (17) of X p(t), it is necessary to compute the free generating triplet (ap, ηp, ρp) in terms of free generating pair (γ, σ) or free 1{0}(x)dσp(x) = σ({0})h(0)2 = generating triplet (a, η, ρ) of X(t). Firstly, ap = σp({0}) = RR ab2. Secondly, for L´evy measure ρp and any bounded Borel function f (x), we have that ZR f (x)dρp(x) =ZR =ZR =ZR =ZR 1 + x2 f (x) 1R\{0}(x)dσp(x) x2 1 + (p(x))2 p(x)2 f (p(x)) 1R\{0}(p(x)) p(x)2 + h(x)2 1 + p(x)2 dσ(x) f (p(x))1R\{0}(p(x)) 1 + x2 x2 dσ(x) f (p(x))1R\{0}(p(x))dρ(x). Therefore, ρp is precisely the measure from Lemma 15, and in particular a L´evy measure. Thirdly, 1+x2 )dσp(x), and the corresponding relation 1+x2 x (1[−1,1](x) − 1 by the relation ηp = γp +RR\{0} 16 MICHAEL ANSHELEVICH, ZHICHAO WANG between η and γ, using also a = σ({0}) and relation (12), we can deduce that ηp = bγ +ZR +ZR = bγ −ZR\{0} 1{p(x)6=0}(x) p 1 1 + (p(x))2(cid:19) dσ(x) b x h2 + p2 1{p(x)=0}(x) 1{p(x)=0}(x)gp(x)dσ(x) 1{p(x)6=0}(x)gp(x)dσ(x) +ZR (cid:18)1{−16p(x)61}(x) − dσ(x) +ZR 1{−16p61}(x)(cid:19) dσ −Z{p(x)>ǫ} dσ(x) −Z{x>ǫ} 1{0<p(x)61}(x)p(x) − 1{0<x61}(x)bx(cid:19) dρ(x). 1{x=0}(x)cdσ(x) 1{ǫ<p(x)61}(x) h2 + p2 b x b x p + lim ǫց0(cid:20)Z{ǫ<p(x)}(cid:18)h2 + p2 p = bγ + ac + lim ǫց0(cid:20)ZR = bη + ac +(cid:18)ZR dσ(x)(cid:21) dσ(x)(cid:21) Note that for some ε > 0, p(x) 6 1 for x 6 ε. So ZR(cid:12)(cid:12)1{0<p(x)61}(x)p(x) − 1{0<x61}(x)bx(cid:12)(cid:12) dρ(x) =Z ε −ε q(x) x2 dρ(x) +ZR(cid:12)(cid:12)1{0<p(x)61,x>ε}(x)p(x) − 1{ε<x61}(x)bx(cid:12)(cid:12) dρ(x) −ε6x6εq(x)Z ε x2 dρ(x) + 2Z{x>ε} dρ(x) < ∞ 6 sup −ε since ρ is a L´evy measure, and so the expression above makes sense. Combine three results we got above and recall the general free L´evy-Ito decomposition of X p(t) with the free generating triplet (ap, ηp, ρp). Let M (p) be the free Poisson random measure on (D,B(D), Leb ⊗ ρp). Then, we can simplify the last part of L´evy-Ito decomposition of X p(t) with respect to the free Poisson random measure M (p): lim ǫց0(cid:20)Z(0,t]×{x>ǫ} ǫց0(cid:20)Z(0,t]×{p(x)>ǫ} ǫց0(cid:20)Z(0,t]×{p(x)>ǫ} xdM (p)(t, x) −Z(0,t]×{ǫ<x61} p(x)dM(t, x) − tZR p(x)dM(t, x) − tZR xLeb ⊗ ρp(dt, dx)1A0(cid:21) x1{ǫ<x61}(x)dρp(x)1A0(cid:21) p(x)1{ǫ<x61}(p(x))dρ(x)1A0(cid:21) = lim = lim . HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 17 Here, we employ Lemma 15, integration by substitution with respect to free Poisson random mea- sures and relation (13). Thus finally, + lim ǫց0(cid:18)ZR X p(t) d= t(cid:18)bη + ac + lim ǫց0(cid:20)Z(0,t]×{p(x)>ǫ} = bhηt1A0 + √aS(t) ǫց0(cid:16)Z(0,t]×{x>ǫ} + lim + act + lim 1{ε<p(x)61}(x)p(x) − 1{ε<x61}(x)bx(cid:19) dρ(x)(cid:19) 1A0 + √abS(t) p(x)dM(t, x) − tZR xdM(t, x) − tZ{ǫ<x61} 1{ǫ<p(x)61}(x)p(x)dρ(x)1A0(cid:21) x dρ(x)1A0(cid:17)i p(x)dM(t, x) − bZ(0,t]×{x>ǫ} xdM(t, x)(cid:17) ǫց0(cid:16)Z(0,t]×{p(x)>ǫ} ǫց0Z(0,t]×{x>ǫ} = bX(t) + act + lim (p(x) − bx)dM(t, x) p(x)(1{p(x)>ǫ} − 1{x>ǫ})dM(t, x). Here we used the fact that the distribution of S(t) is symmetric. Since (p(x) − bx)dM(t, x) =Z(0,t]×R xdM (p(x)−bx)(t, x) exists by Lemmas 16 and 12, and the functions (p(x) − bx)1x6ε have a uniform integrable bound and converge to zero pointwise as ε → 0, by Lemma 11 we have + lim ǫց0Z(0,t]×R Z(0,t]×R lim ǫց0Z(0,t]×{x>ǫ} =Z(0,t]×R =Z(0,t]×R (p(x) − bx)dM(t, x) (p(x) − bx)dM(t, x) − lim (p(x) − bx)dM(t, x) ǫց0Z(0,t]×{x≤ǫ} (p(x) − bx)dM(t, x). Finally, the functions p(x)(1{p(x)>ǫ} − 1{x>ǫ}) = −p(x)(1{p(x)≤ǫ} − 1{x≤ǫ}) also have a uniform integrable bound and converge to zero pointwise as ε → 0. Therefore by Lemma 11, lim ǫց0Z(0,t]×R p(x)(1{p(x)>ǫ} − 1{x>ǫ})dM(t, x) = 0. (cid:3) Remark 2. It is natural to consider, more generally, free additive (not necessarily) stationary pro- cesses approximated by free, non-identically distributed triangular arrays which are infinitesimal, that is, their distributions µi,N satisfy lim N→∞ max 16i6kN µi,N ({x > ε}) = 0 18 MICHAEL ANSHELEVICH, ZHICHAO WANG for every ε > 0. The following very simple example shows how without additional assumptions, the results immediately break down. Let Xi,N = 1 N + (−1)i 1 N α , i = 1, . . . , 2N. Then clearly the array {Xi,N} is infinitesimal, and limN→∞P[2N t] i=1 Xi,N = 2t. But X 2 i,N ∼ 2t N 2α−1 [2N t]Xi=1 diverges for α < 1 do not converge to it. Compare with the remarks on page 494 of [AT86]. 2 . So while the quadratic variation of a non-random process is zero, these sums 4. CONVERGENCE IN MOMENTS For a non-crossing partition π ∈ NC(n), denote Recall that the free cumulant functional is defined by Mob(π)τπ[a1, . . . , an], τ"Yi∈V ai# . τπ [a1, . . . , an] = YV ∈π R[a1, . . . , an] = Xπ∈NC(n) where Mob is the Mobius function on the lattice of non-crossing partitions. The key property of the free cumulant functional is that if a1, . . . , ak are free, then unless u(1) = . . . = u(n). R[au(1), . . . , au(n)] = 0 Proof of Theorem 3. Note first that by freeness and the free moment-cumulant formula, X u(1) i,N , . . . , R [N t]Xi=1 [N t]Xi=1 [N t]Xi=1(cid:16)R(cid:16)X u(1) [N t]Xi=1 Xπ∈NC(k) = = π6=1k i,N , . . . , X u(k) Mob(π)τπhX u(1) X u(k) X u(1)+...+u(k) i,N [N t]Xi=1 i,N  − τ i,N (cid:17) − τhX u(1)+...+u(k) i,N i . i,N , . . . , X u(k) i,N  i(cid:17) HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 19 The absolute value of this expression is bounded by i,N τhXPj∈V2 i,N u(j) 1/2 i2 π6=1k π6=1k u(j) i,N max i,N 1/2 [N t]Xi=1 i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) i2 Mob(π)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) τhXPj∈V u(j) [N t]Xi=1 YV ∈π Xπ∈NC(k) Mob(π) τhXPj∈V1 [N t]Xi=1 6 Xπ∈NC(k) (cid:12)(cid:12)(cid:12) , 16i6[N t](cid:12)(cid:12)(cid:12)XPj∈V u(j) × YV ∈π\{V1,V2} τ i,N [N t]Xi=1 τ i,N − Rk Xi,N → 0 [N t]Xi=1 [N t]Xi=1 Rk Xi,N → Rk(X(t)). [N t]Xi=1 lim N→∞ X k X k as N → ∞. Finally, by assumption which goes to zero as N → ∞, by assumption. So to prove joint convergence in moments, it suffices to show that the limit exists for each k. Indeed, applying the derivation above to the case u(1) = . . . = u(k) = 1, The statement about processes follows as in Proposition 13. (cid:3) 5. CONVERGENCE IN PROBABILITY Definition 5. an → a almost uniformly (a.u.) if for any δ > 0, there exists a projection p such that τ [1 − p] < δ and k(an − a)pk → 0. We now quote a result from [BV93]. Lemma 18 (Lemma 4.4). Let (A, τ ) be a W ∗-probability space, T1, T2, T ′ orthogonal projections. Suppose T ′ such that 2 ∈ A, and p1, p2 ∈ A j = Tjpj, for j = 1, 2. Then there exist projections p, q ∈ A 1, T ′ 1T ′ (a) (T1T2)p = (T ′ (b) (T1 + T2)q = (T ′ 2)q, and (c) τ [p], τ [q] > τ [p1] + τ [p2] − 1. 2)p, 1 + T ′ We do not have a reference for the following result (compare with [Pet84, Sau91]), and so provide a short proof. Lemma 19. Let an → a and bn → b a.u. Then an + bn → a + b and anbn → ab a.u. 20 MICHAEL ANSHELEVICH, ZHICHAO WANG Proof. The first statement is obvious. For the second, let ε > 0, and choose projections p1, p2, p3, p4 so that k(an − a)p1k → 0, k(bn − b)p2k → 0, p3 6 p1, kap3k < ∞, kbp4k < ∞ and τ [p1], τ [p2], τ [p3], τ [p4] > 1 − ε. By Lemma 18, we may choose projections p′, p′′ with τ [p′], τ [p′′] > 1 − 2ε so that (an − a)bp′ = (an − a)p1bp4p′, an(bn − b)p′′ = anp3(bn − b)p2p′′. Since p3 6 p1, also supn kanp3k < ∞. Finally taking p = p′ ∧ p′′, with τ [p] > 1 − 4ε, we get k(anbn − ab)pk 6 k(an − a)bpk + kan(bn − b)pk 6 k(an − a)p1kkbp4k + kanp3kk(bn − b)p2k → 0 as n → ∞. Remark 3. Let ρ be a probability measure on R. In the tracial non-commutative probability space C = L∞((0, 1] × R, Leb ⊗ ρ), consider the projections P (B) = χB for every Borel set B. Let s be a semicircular element free from C. Then according to [NS96], the family of operators M : B → sP (B)s satisfies all the properties of a free Poisson random measure in Definition 3. Next, let (cid:3) e(t) =ZR x P ((0, t] × dx), meaning that the spectral projections of et are {P ((0, t] × (−∞, x))}. Then {e(t) : t ∈ (0, 1]} is a process with orthogonal increments, and {se(t)s : t ∈ (0, 1]} is a free compound Poisson process. Note that se(t)s =ZR xs P ((0, t] × dx)s =Z(0,t]×R x dM(t, x). Proposition 20. Let Z1, . . . , Zk be bounded and centered, free from a stationary process {e(t)} with orthogonal increments. Then NXi=1 em0 i,N Z1em1 i,N Z2 . . . emk−1 i,N Zkemk i,N → 0 a.u. as N → ∞. Here we denote as usual ei,N = e( i Proof. Without loss of generality, assume that {e(t)} has the form in Remark 3. For arbitrary ε > 0, choose T so that for q = χ[0,1]×(−T,T ), we have τ (q) = ρ((−T, T )) > 1 − ε. Then {e′(t) = e(t)q} is a bounded process with orthogonal increments, and ke′(1)k = T . According to Lemma 18, there N ) − e( i−1 N ). j=0 mjε such that i,N Zkemk i,N p = (e′ i,N )m0Z1(e′ i,N )m1Z2 . . . (e′ i,N )mk−1Zk(e′ i,N )mkp On the other hand, according to Theorem 3 from [Ans00], em0 i,N Z1em1 i,N Z2 . . . emk−1 is a projection p with τ (p) > 1 −Pk (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) NXi=1 The result follows. i,N )m0Z1(e′ i,N )m1Z2 . . . (e′ (e′ i,N )mk−1Zk(e′ i,N )mk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 6 42k(maxkZjk)kT Pk j=0 mj N −k/2. (cid:3) in probability as N → ∞. Next, write X(t) = se(t)s as before. By the same reasoning as in Remark 3, Z(cid:0)0, [Nt] N (cid:3)×R Z(0,t]×R xk dM(t, x) →Z(0,t]×R xk dM(t, x) xk dM(t, x) = se(t)ks. HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 21 Proof of Theorem 4. By using the addition part of Lemma 19, we may assume that t ∈ (0, 1]. Note first that by Lemma 11, Therefore [N t]Xi=1 X k i,N −Z(cid:0)0, [Nt] N (cid:3)×R k−1Xj=1 Xm0,m1,...,mj >1 m0+m1+...+mj=k = s [N t]Xi=1 xk dM(t, x) = [N t]Xi=1 (sei,N s)k − sek i,N s [N t]Xi=1 em0 i,N (s2 − 1)em1 i,N (s2 − 1) . . . emj−1 i,N (s2 − 1)emj i,N s. Now note that τ (s2 − 1) = 0 and apply Proposition 20. (cid:3) Proof of Corollary 6. By Lemma 19, addition and multiplication are continuous with respect to the i,No converge to the topology of a.u. convergence. Thus by Theorem 4, polynomials innP[N t] corresponding polynomials in(cid:8)X (j)(t)(cid:9) a.u. Finally, convergence a.u. clearly implies convergence in probability, and by Proposition 2.19 in [BNT02] (see also Proposition 2.1 in [LP97]), conver- gence in probability implies convergence in distribution. (cid:3) i=1 X j Proof of Corollary 7. According to Corollary 25, X16i(1),i(2),...,i(k)6[N t] i(1)6=i(2),i(2)6=i(3),...,i(k−1)6=i(k) Xi(1),N Xi(2),N . . . Xi(k),N = kXj=1 (−1)k−j Xm1,...,mj >1 m1+...+mj =k  [N t]Xi(1)=1 X m1 i(1),N . . . [N t]Xi(j)=1 X mj i(j),N . Now apply either Theorem 3 or Corollary 6. (cid:3) See the second author's thesis for a direct proof. Remark 4. In the case of a process which is not necessarily centered, normalizing it so that τ [X(t)] = t, a more natural definition of an n-fold stochastic integral ψn, according to Theorem 4 of [Ans00], is ψn = Xψn−1 + nXj=2 (−1)j−1 n−jXk=0(cid:18)k + j − 2 j − 2 (cid:19)X (j)ψn−j−k. 22 MICHAEL ANSHELEVICH, ZHICHAO WANG The recursion Pn = NXi=1 xi! Pn−1 + nXj=2 (−1)j−1 n−jXk=0(cid:18)k + j − 2 j − 2 (cid:19) NXi=1 i! Pn−j−k. xj for polynomials Pn(x1, . . . , xN , t) can be solved explicitly, but we find the resulting formula com- plicated and not particularly illuminating, and omit it from the article. We can similarly upgrade various results proven in [Ans00] for bounded free L´evy processes and uniform limits to general free compound Poisson processes and almost uniform limits. This applies to Theorem 1 (stochastic measures corresponding to crossing partitions are zero), Proposition 1 (for a centered process, stochastic measures corresponding to partitions with inner singletons are zero) and its corollary on the equality of expressions (6) and (7), Remark 5. Let µ, ν be probability measures on R, such that µ = µa, ν = µb for free a, b ∈ ( Asa, τ ). The additive free convolution µ ⊞ ν is the distribution of a + b. If µ is supported on R+ (so that a is positive), the multiplicative free convolution µ ⊠ ν is the distribution of a1/2ba1/2, which we identify (since τ is a trace) with the distribution of ab. According to Proposition 3.5 in [BN08], we have the relation (23) (µ⊞t) ⊠ (ν ⊞t) = (µ ⊠ ν)⊞t ◦ D1/t, where D1/t is the dilation operator corresponding to multiplying the operator by t. Note that in the proposition, the relation is stated for t > 1, but the same argument shows that it holds whenever all the convolution powers on the left-hand side are defined and at least one of them is supported on R+. Proposition 21. LetnX (1) i,N : 1 6 i 6 N, N ∈ No∪nX (2) i,N : 1 6 i 6 N, N ∈ No ⊂(cid:16)( A, τ )sa(cid:17) be two triangular arrays with free, identically distributed rows, free from each other, the first of which consists of positive operators. Denote and suppose that X (j) i,N = X (j) N , j = 1, 2 NXi=1 lim N→∞ X (j) N = X (j), j = 1, 2 in distribution, for some(cid:8)X (1), X (2)(cid:9). Then as N → ∞, i,N(cid:17)1/2 i,N(cid:16)X (1) i,N(cid:17)1/2 NXi=1(cid:16)X (1) X (2) in distribution, and so also in probability. → 0 Proof. Using the identity from the preceding remark, µ (cid:16)X (1) i,N(cid:17)1/2 and so i,N(cid:16)X (1) X (2) i,N(cid:17)1/2 = (µ⊞(1/N ) X (1) N ) ⊠ (µ⊞(1/N ) X (2) N ) = (µX (1) N ⊠ µX (2) N )⊞(1/N ) ◦ DN µ i=1(cid:16)X (1) i,N(cid:17)1/2 PN i,N(cid:16)X (1) X (2) i,N(cid:17)1/2(z) = (µX (1) N ⊠ µX (2) N ) ◦ DN . HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 23 N ⊠ µX (2) As N → ∞, µX (1) weakly. Remark 6. Denote Cµ(z) = zϕµ(1/z) the free cumulant transform. A measure σ is free regular if N → µX (1) ⊠ µX (2) weakly, and so the distribution above converges to δ0 (cid:3) Cσ(z) = η′z +ZR(cid:18) 1 1 − zx − 1(cid:19) ν(dx) for some η′ > 0 and ν((−∞, 0]) = 0. By Proposition 6.2 in [AHS13], if µ is ⊞-infinitely divisible and symmetric, then Here µ2 = µ(x2) in our earlier notation, m is the standard free Poisson distribution, and σ is a free regular measure. Moreover by Theorem 11 from [PAS12], this is equivalent to µ2 = m ⊠ σ. Cµ(z) = Cσ(z2). (cid:0)(µ(cid:3)ν)⊞1/2(cid:1)2 = µ2 ⊠ ν2. Next, let µ, ν be probability measures on R, such that µ = µa, ν = µb for free a, b ∈ ( Asa, τ ). Denote by µ(cid:3)ν the distribution of the anti-commutator ab + ba. If µ, ν are both symmetric, it coincides with the distribution of the commutator i(ab − ba), and satisfies See [NS98], Lectures 15 and 19 in [NS06], and Corollary 6.5 in [AHS13]. We also note that if in Remark 5, µ is free regular, then by Theorem 4.2 in [AHS13], µ⊞t is the distribution of a positive operator for all t > 0. So if in addition ν is ⊞-infinitely divisible, the identity (23) holds for all such t. Proposition 22. LetnX (1) i,N : 1 6 i 6 N, N ∈ No ∪nX (2) i,N : 1 6 i 6 N, N ∈ No ⊂ (cid:16)( A, τ )sa(cid:17) be two triangular arrays with free, identically distributed rows, free from each other, all of whose distributions are symmetric. Denote X (j) i,N = X (j) N , j = 1, 2 N is ⊞-infinitely divisible and N = X (j), j = 1, 2 X (j) and suppose that the distribution of each X (j) in distribution, for some(cid:8)X (1), X (2)(cid:9). Then as N → ∞, NXi=1 lim N→∞ NXi=1(cid:16)X (1) NXi=1 i,N X (2) i,N + X (2) i,N X (1) i,N(cid:17) → 0 X (1) i,N X (2) i,N → 0 in distribution, and so also in probability, and (24) in probability. 24 MICHAEL ANSHELEVICH, ZHICHAO WANG Proof. Denote by µj,N the distribution of X (j) N . Using the preceding remark, we may write where σj,N is a free regular measure, such that µ2 j,N = m ⊠ σj,N , Cµj,N (z) = Cσj,N (z2). Note that Thus Cµ ⊞(1/N) j,N (z) = 1 N Cµj,N (z) = 1 N Cσj,N (z2) = Cσ ⊞(1/N) j,N (z2). Next, (cid:18)(cid:16)µ⊞(1/N ) Therefore 1,N (cid:3)µ⊞(1/N ) 2,N (cid:17)⊞(1/2)(cid:19)2 = m ⊠ σ⊞(1/N ) . j,N j,N (cid:16)µ⊞(1/N ) (cid:17)2 =(cid:16)µ⊞(1/N ) (cid:17)2 1,N ⊠(cid:16)µ⊞(1/N ) 2,N (cid:17)2 = m ⊠ σ⊞(1/N ) 1,N ⊠ m ⊠ σ⊞(1/N ) 2,N . C(cid:16)µ ⊞(1/N) 1,N (cid:3)µ ⊞(1/N) 2,N (cid:17)⊞(N/2)(z) = NCσ ⊞(1/N) 1,N ⊠m⊠σ ⊞(1/N) 2,N (z2). Applying the relation (23) twice and distributing the dilation, we get (cid:16)σ⊞(1/N ) 1,N ⊠ m ⊠ σ⊞(1/N ) 2,N (cid:17)⊞N =(cid:0)σ1,N ⊠ m⊞N ⊠ σ2,N(cid:1) ◦ DN 2 =(cid:0)m⊞N ◦ DN(cid:1) ⊠ ((σ1,N ⊠ σ2,N ) ◦ DN ) . Using the (noncommutative) law of large numbers, or by a direct calculation, m⊞N ◦ DN → δ1, so these measures converge to δ0 weakly. Therefore their free cumulant transforms converge to zero pointwise, which implies that Since the same convergence in probability holds for the commutators (cid:16)µ⊞(1/N ) i(cid:16)X (1) NXj=1 1,N (cid:3)µ⊞(1/N ) 2,N (cid:17)⊞(N/2) → δ0. j,N X (2) j,N − X (2) j,N X (1) j,N(cid:17) , it holds for their linear combination (24). Proof of Theorem 5. Let X(t) = ηt1A0 +√aS(t) + lim Fix α ∈ (0, 1). Denote X ′(t) =(cid:18)η −Z{α6x61} xdM(t, x)−Z(0,t]×{ǫ<x61} ǫց0(cid:16)Z(0,t]×{x>ǫ} x ρ(dx)(cid:19) t1A0 + √aS(t) (cid:3) x((Leb⊗ ρ)(dt, dx)1A0)(cid:17). + lim ǫց0(cid:16)Z(0,t]×{ε<x<α} xdM(t, x) −Z(0,t]×{ǫ<x<α} x((Leb ⊗ ρ)(dt, dx)1A0)(cid:17). HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 25 and X ′′(t) =Z(0,t]×{x>α} xdM(t, x). Note that {X ′′(t)} is an (unbounded) free compound Poisson process, X(t) = X ′(t) + X ′′(t), {X ′(t)} and {X ′′(t)} are free from each other, and all of their distributions are ⊞-infinitely divisible and symmetric. Then X 2 i,N = [N t]Xi=1 [N t]Xi=1(cid:0)X ′ i,N(cid:1)2 + [N t]Xi=1(cid:0)X ′′ i,N(cid:1)2 + [N t]Xi=1(cid:0)X ′ i,N X ′′ i,N + X ′′ i,N X ′ i,N(cid:1) . By Theorem 4, the second term converges almost uniformly to (X ′′)(2) (t). By Proposition 22, the third term converges to zero in probability. By Theorem 2, for fixed α, the first term converges in distribution to (X ′)(2) (t) = at1A +Z(0,t]×(−α,α) x2dM(t, x). and Finally, as α → 0, (X ′)(2) (t) → at1A in probability. Thus, given ε, δ > 0, we may choose α small i=1(cid:0)X ′ i,N(cid:1)2−at1A ∈ N (ε, δ) so that (X ′)(2) (t)−at1A ∈ N (ε, δ). Then for sufficiently large N ,P[N t] i,N − (X ′′)(2) (t) − at1A ∈ N (ε, δ). X 2 [N t]Xi=1 It remains to note that also X (2)(t) − (X ′′)(2) (t) − at1A = (X ′)(2) (t) − at1A ∈ N (ε, δ). (cid:3) We finish this section with another possible definition of joint convergence in distribution. As al- ready noted, for commuting variables, convergence in distribution of linear combinations is equiv- alent to joint convergence in distribution. As pointed out by ´Eduard Maurel-Segala and Maxime Fevrier, this is not the case for non-commuting variables. However the following matricial version is its natural replacement. By the well-known linearization trick [HT05] (see also Chapter 10 of [MS17]), it implied the definition in the introduction; we do not know if they are in general equiv- alent. We show that convergence in probability implies joint convergence in this possibly stronger sense as well. Definition 6. Let {xi,N : 1 6 i 6 k, N ∈ N} ∪ {xi : 1 6 i 6 k} ⊂ ( Asa, τ ). We say that (x1,N , . . . , xk,N ) → (x1, . . . xk) jointly in distribution if for any d and any Hermitian matrices A1, . . . , Ak ∈ Md(C), and any B ∈ Md(C) with ℑB > εI for some ε > 0, the Cauchy transforms (I ⊗ τ ) B ⊗ 1 − Ai ⊗ xi,N!−1 → (I ⊗ τ ) B ⊗ 1 − kXi=1 Ai ⊗ xi!−1 kXi=1 in norm in MN (C). Proposition 23. If for each i, xi,N → xi in probability, then (x1,N , . . . , xk,N ) → (x1, . . . xk) in the sense of Definition 6. By assumption and a short argument, for any ε, δ > 0 there is an n such that for N > n, there is a projection pN with τ [pN ] > 1 − δ and Ai ⊗ xi,N! B ⊗ 1 − Ai ⊗ xi!−1 . kXi=1 and in particular this operator is bounded. By the resolvent identity, reader's convenience. Note first that for X ∈ Md( Asa), kXi=1 (cid:13)(cid:13)(B ⊗ 1 − X)−1(cid:13)(cid:13) 6(cid:13)(cid:13)(ℑB)−1(cid:13)(cid:13) , Ai ⊗ xi,N!−1 B ⊗ 1 − − B ⊗ 1 − Ai ⊗ xi!−1 kXi=1 = B ⊗ 1 − Ai ⊗ xi,N!−1 kXi=1 kXi=1 kXi=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kXi=1 Ai ⊗ xi,N!−1 kXi=1 Ai ⊗ xi,N! (I ⊗ pN )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kXi=1 − B ⊗ 1 − Ai ⊗ xi − Ai ⊗ xi − kXi=1 Thus for some projection qN with the same property, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  B ⊗ 1 − < ε kAik . kXi=1 Ai ⊗ xi!−1 (I ⊗ qN )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kXi=1 6 ε(cid:13)(cid:13)(ℑB)−1(cid:13)(cid:13)2 Ai ⊗ xi!−1 → 0. kXi=1 kAik . (cid:3) 26 MICHAEL ANSHELEVICH, ZHICHAO WANG Proof. The argument in Proposition 2.19 in [BNT02] largely goes through; we outline it for the In particular, the same estimate holds on each matrix entry on the left-hand side. Applying the rest of the argument from Proposition 2.19 in [BNT02] entry-wise, it follows that (I ⊗ τ ) B ⊗ 1 − Ai ⊗ xi,N!−1 kXi=1 − B ⊗ 1 − APPENDIX A. SYMMETRIC POLYNOMIALS IN NON-COMMUTING VARIABLES Symmetric functions in non-commuting variables (not to be confused with non-commutative sym- metric functions) have been considered in [RS06, BRRZ08] and subsequent work. We need the following observation, whose explicit statement we could not find in the literature. Proposition 24. Let pk =PN i be the basic power sum symmetric polynomials. In the algebra of non-commutative polynomials Chx1, . . . , xNi, the subalgebra generated by {pk : k > 1} is the linear span of polynomials i=1 xk Pu(x) = NXi(1),i(2),...=1 neighbors distinct xu(1) i(1) xu(2) i(2) . . . HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 27 for all choices of u with coordinates u(i) > 1. Note that these polynomials are obviously linearly independent. In particular, the elementary symmetric functions ek = Xi(1)6=i(2)6=...6=i(r) i(1)+i(2)+...+i(r)=k xi(1)xi(2) . . . xi(r) are not in this subalgebra for k > 1. Proof. Clearly the algebra generated by all pk is the span of all Qu(x) = pu(1)(x)pu(2)(x) . . . = NXi(1),i(2),...=1 xu(1) i(1) xu(2) i(2) . . . , where the i(j) are not necessarily distinct. Denote by Int(n) the interval partitions of [n]. Then we may re-index these polynomials as Qπ(x) = NXi(1),i(2),...,i(r)=1 rYj=1 xVj i(j) = pVj(x) rYj=1 for π = {V1, . . . , Vr} ∈ Int(n) for some n. For u ∈ [N]r, denote ker(u) ∈ P(n) the partition such that u(i) = u(j) if and only if i, j lie in the same block of ker(u). Note that for V ∈ ker(u), the notation u(V ) is unambiguous. Also, for π ∈ P(n), let I(π) be the largest interval partition such that I(π) 6 π. Note that I(π) = τ if π > τ and if V, V ′ are neighboring blocks of τ , they lie in different blocks of π. Finally, for π = {V1, . . . , Vr} ∈ Int(n), denote Pπ(x) = neighbors distinct xVj i(j). rYj=1 NXi(1),i(2),...,i(r)=1 I(π)=τ Xi:ker i=πYV ∈π τ >σ Xπ∈P(n) i(V ) = Xτ ∈Int(n) i(V ) = Xτ ∈Int (n) Pτ (x). xV xV τ >σ xV i(V ) Then for σ ∈ Int(n), π>σ Xi:ker i=πYV ∈π Qσ(x) = Xπ∈P(n) = Xτ ∈Int(n) τ >σ NXi(1),i(2),...,i(τ )=1 neighbors distinct YV ∈τ Then by Mobius inversion on the lattice Int(n), the spans of {Qπ} and of {Pπ} are the same. (cid:3) Corollary 25. In the notation of the preceding proof, Pσ = Xπ∈Int(n) π>σ (−1)σ−πYV ∈π pV (x). In particular, NXi(1),i(2),...,i(n)=1 neighbors distinct nYj=1 xi(j) = Xπ∈Int(n) (−1)n−πYV ∈π pV (x). 28 MICHAEL ANSHELEVICH, ZHICHAO WANG Proof. The first statement follows by Mobius inversion, since the Mobius function on the lattice Int(n) is Mob(σ, π) = (−1)σ−π. The second statement follows from the fact that the left-hand side is P0n(x). (cid:3) REFERENCES []Ans00 Michael Anshelevich, Free stochastic measures via noncrossing partitions, Adv. Math. 155 (2000), no. 1, 154 -- 179. MR 1789851 (2001k:46102) []AHS13 Octavio Arizmendi, Takahiro Hasebe, and Noriyoshi Sakuma, On the law of free subordinators, ALEA Lat. []AT86 Am. J. Probab. Math. Stat. 10 (2013), no. 1, 271 -- 291. MR 3083927 Florin Avram and Murad S. Taqqu, Symmetric polynomials of random variables attracted to an infinitely divisible law, Probab. Theory Relat. Fields 71 (1986), no. 4, 491 -- 500. MR 833266 []BNT02 Ole E. Barndorff-Nielsen and Steen Thorbjørnsen, Self-decomposability and L´evy processes in free proba- bility, Bernoulli 8 (2002), no. 3, 323 -- 366. MR 1913111 []BNT05 , The L´evy-Ito decomposition in free probability, Probab. Theory Related Fields 131 (2005), no. 2, []BN08 []BP99 []BV93 197 -- 228. MR 2117952 Serban T. Belinschi and Alexandru Nica, On a remarkable semigroup of homomorphisms with respect to free multiplicative convolution, Indiana Univ. Math. J. 57 (2008), no. 4, 1679 -- 1713. MR 2440877 (2009f:46087) Hari Bercovici and Vittorino Pata, Stable laws and domains of attraction in free probability theory, Ann. of Math. (2) 149 (1999), no. 3, 1023 -- 1060, With an appendix by Philippe Biane. MR 2000i:46061 Hari Bercovici and Dan Voiculescu, Free convolution of measures with unbounded support, Indiana Univ. Math. J. 42 (1993), no. 3, 733 -- 773. MR 1254116 (95c:46109) []DN14 []BRRZ08 Nantel Bergeron, Christophe Reutenauer, Mercedes Rosas, and Mike Zabrocki, Invariants and coinvari- ants of the symmetric groups in noncommuting variables, Canad. J. Math. 60 (2008), no. 2, 266 -- 296. MR 2398749 Aur´elien Deya and Ivan Nourdin, Invariance principles for homogeneous sums of free random variables, Bernoulli 20 (2014), no. 2, 586 -- 603. MR 3178510 Uffe Haagerup and Steen Thorbjørnsen, A new application of random matrices: Ext(C ∗ group, Ann. of Math. (2) 162 (2005), no. 2, 711 -- 775. MR 2183281 (2009k:46121) J. Martin Lindsay and Vittorino Pata, Some weak laws of large numbers in noncommutative probability, Math. Z. 226 (1997), no. 4, 533 -- 543. MR 1484709 red(F2)) is not a []HT05 []LP97 []Lo`e77 Michel Lo`eve, Probability theory. I, fourth ed., Springer-Verlag, New York-Heidelberg, 1977, Graduate []MS13 []MS17 Texts in Mathematics, Vol. 45. MR 0651017 Tobias Mai and Roland Speicher, Operator-valued and multivariate free Berry-Esseen theorems, Limit the- orems in probability, statistics and number theory, Springer Proc. Math. Stat., vol. 42, Springer, Heidelberg, 2013, pp. 113 -- 140. MR 3079141 James A. Mingo and Roland Speicher, Free probability and random matrices, Fields Institute Monographs, vol. 35, Springer, New York; Fields Institute for Research in Mathematical Sciences, Toronto, ON, 2017. MR 3585560 []MVN36 F. J. Murray and J. Von Neumann, On rings of operators, Ann. of Math. (2) 37 (1936), no. 1, 116 -- 229. []Nel74 []NS96 []NS98 []NS06 MR 1503275 Edward Nelson, Notes on non-commutative integration, J. Functional Analysis 15 (1974), 103 -- 116. MR 0355628 Alexandru Nica and Roland Speicher, On the multiplication of free N -tuples of noncommutative random variables, Amer. J. Math. 118 (1996), no. 4, 799 -- 837. MR 98i:46069 , Commutators of free random variables, Duke Math. J. 92 (1998), no. 3, 553 -- 592. MR 1620518 (99d:46084) , Lectures on the combinatorics of free probability, London Mathematical Society Lecture Note Series, vol. 335, Cambridge University Press, Cambridge, 2006. MR 2266879 (2008k:46198) []PAS12 Victor P´erez-Abreu and Noriyoshi Sakuma, Free infinite divisibility of free multiplicative mixtures of the Wigner distribution, J. Theoret. Probab. 25 (2012), no. 1, 100 -- 121. MR 2886381 HIGHER VARIATIONS FOR FREE L ´EVY PROCESSES 29 []Pet84 []RS06 []RW97 []Sau91 []Sim15 []Spe90 D´enes Petz, Quantum ergodic theorems, Quantum probability and applications to the quantum theory of irreversible processes (Villa Mondragone, 1982), Lecture Notes in Math., vol. 1055, Springer, Berlin, 1984, pp. 289 -- 300. MR 782910 Mercedes H. Rosas and Bruce E. Sagan, Symmetric functions in noncommuting variables, Trans. Amer. Math. Soc. 358 (2006), no. 1, 215 -- 232. MR 2171230 Gian-Carlo Rota and Timothy C. Wallstrom, Stochastic integrals: a combinatorial approach, Ann. Probab. 25 (1997), no. 3, 1257 -- 1283. MR 98m:60081 Jean-Luc Sauvageot, A note on almost uniform convergence in von Neumann algebras, Quantum probability & related topics, QP-PQ, VI, World Sci. Publ., River Edge, NJ, 1991, pp. 385 -- 390. MR 1149839 Rosaria Simone, Universality of free homogeneous sums in every dimension, ALEA Lat. Am. J. Probab. Math. Stat. 12 (2015), no. 1, 213 -- 244. MR 3343483 Roland Speicher, A new example of "independence" and "white noise", Probab. Theory Related Fields 84 (1990), no. 2, 141 -- 159. MR 1030725 (90m:46116) DEPARTMENT OF MATHEMATICS, TEXAS A&M UNIVERSITY, COLLEGE STATION, TX 77843-3368 E-mail address: [email protected], [email protected]
1406.6160	2	1406	2015-04-26T15:03:40	Asymptotic structure of free Araki-Woods factors	[ "math.OA" ]	The purpose of this paper is to investigate the structure of Shlyakhtenko's free Araki-Woods factors using the framework of ultraproduct von Neumann algebras. We first prove that all the free Araki-Woods factors $\Gamma(H_{\mathbb R}, U_t)^{\prime \prime}$ are $\omega$-solid in the following sense: for every von Neumann subalgebra $Q \subset \Gamma(H_{\mathbb R}, U_t)^{\prime \prime}$ that is the range of a faithful normal conditional expectation and such that the relative commutant $Q' \cap M^\omega$ is diffuse, we have that $Q$ is amenable. Next, we prove that the continuous cores of the free Araki-Woods factors $\Gamma(H_{\mathbb R}, U_t)^{\prime \prime}$ associated with mixing orthogonal representations $U : \mathbb R \to \mathcal O(H_{\mathbb R})$ are $\omega$-solid type ${\rm II_\infty}$ factors. Finally, when the orthogonal representation $U : \mathbb R \to \mathcal O(H_{\mathbb R})$ is weakly mixing, we prove a dichotomy result for all the von Neumann subalgebras $Q \subset \Gamma(H_{\mathbb R}, U_t)^{\prime \prime}$ that are globally invariant under the modular automorphism group $(\sigma_t^{\varphi_U})$ of the free quasi-free state $\varphi_U$.	math.OA	math	Asymptotic structure of free Araki-Woods factors by Cyril Houdayer1 and Sven Raum2 Abstract The purpose of this paper is to investigate the structure of Shlyakhtenko's free Araki-Woods factors using the framework of ultraproduct von Neumann algebras. We first prove that all the free Araki-Woods factors Γ(HR, Ut)′′ are ω-solid in the following sense: for every von Neumann subalgebra Q ⊂ Γ(HR, Ut)′′ that is the range of a faithful normal conditional expectation and such that the relative commutant Q′ ∩ M ω is diffuse, we have that Q is amenable. Next, we prove that the continuous cores of the free Araki-Woods factors Γ(HR, Ut)′′ associated with mixing orthogonal representations U ∶ R → O(HR) are ω-solid type II∞ factors. Finally, when the orthogonal representation U ∶ R → O(HR) is weakly mixing, we prove a dichotomy result for all the von Neumann subalgebras Q ⊂ Γ(HR, Ut)′′ that are globally invariant under the modular automorphism group (σϕU t ) of the free quasi-free state ϕU . 1 Introduction and statement of the main results Free Araki-Woods factors were introduced by Shlyakhtenko in [Shl97]. In the context of Voiculescu's free probability theory, these factors can be regarded as the analogues of the hyperfinite factors coming from the canonical anticommutation relations (CAR) functor. Alternatively, they can also be regarded as the analogues of the free group factors in the setting of type III factors. Following [Shl97], to any orthogonal representation U ∶ R → O(HR) on a separable real Hilbert space, one associates a von Neumann algebra denoted by Γ(HR , Ut)′′, called the free Araki-Woods von Neumann algebra. The von Neumann algebra Γ(HR , Ut)′′ comes equipped with a unique free quasi-free state ϕU that is always normal and faithful (see Subsection 2.2 for a detailed construction). dim(HR)) when U ∶ R → O(HR) is the trivial representation and Γ(HR , Ut)′′ We have Γ(HR , id)′′ ≅ L(F is a full type III factor when U ∶ R → O(HR) is not the trivial representation. Free Araki-Woods factors were first studied using the framework of Voiculescu's free probability theory. A complete description of their type classification as well as fullness and computation of their Connes's τ and Sd invariants was obtained in [Shl97, Shl98, Shl99] (see also the survey [Vae06]). More recently, free Araki-Woods factors were studied using the framework of Popa's Deformation/Rigidity theory [Pop06a]. This new approach allowed to obtain various indecom- posability results in [Hou08] and complete metric approximation property and absence of Cartan subalgebra in [HR10]. Because of their rich structure, free Araki-Woods factors form one of the most prominent classes of type III factors. The purpose of this paper is to investigate the asymptotic structure of free Araki-Woods factors using the framework of ultraproduct von Neumann algebras. Before stating our main results, we first introduce some terminology. 1Research supported by the ANR Grant NEUMANN and JSPS Invitation Fellowship Program for Research in Japan FY2014 2Research supported by the ANR Grant NEUMANN 1 We will say that a von Neumann subalgebra Q ⊂ M is with expectation if there exists a faithful normal conditional expectation EQ ∶ M → Q. We will say that a diffuse von Neumann algebra M is solid if for every von Neumann subalgebra Q ⊂ M with expectation whose relative commutant Q′ ∩ M is diffuse, we have that Q is amenable [Oza04]. The first class of solid von Neumann algebras was discovered by Ozawa in [Oza04]. He showed that every Gromov-word hyperbolic group G gives rise to a solid von Neumann algebra L(G). More conceptually, Ozawa showed that every finite diffuse von Neumann algebra satisfying the Akemann-Ostrand property (abbreviated property (AO) hereafter, see Subsection 2.4) is solid. It was observed in [VV05] that in fact every diffuse von Neumann algebra satisfying property (AO) is solid. Let now ω ∈ β(N)∖N be a non-principal ultrafilter. We refer to Subsection 2.3 for the construction of the ultraproduct von Neumann algebra M ω. We will say that a von Neumann algebra M is ω-solid if for every von Neumann subalgebra Q ⊂ M with expectation whose relative commutant Q′ ∩ M ω is diffuse, we have that Q is amenable. Since M sits in M ω as a von Neumann subalgebra with expectation, any ω-solid von Neumann is obviously solid. As of today, the converse implication is an open problem3. Ozawa proved in [Oza10] that any finite diffuse von Neumann algebra satisfying property (AO) is ω-solid. Our first result generalises Ozawa's result [Oza10] to arbitrary diffuse von Neumann algebras with separable predual satisfying property (AO). Theorem A. Any von Neumann algebra with separable predual satisfying property (AO) is ω-solid. In particular, any free Araki-Woods factor Γ(HR , Ut)′′ associated with an orthogonal representation U ∶ R → O(HR) on a separable real Hilbert space is ω-solid. The proof of Theorem A combines Ozawa's original argument [Oza04] together with several tech- niques from [AH12] on the structure of ultraproduct von Neumann algebras. The proof of Theo- rem A is carried out in Section 3. The fact that all free Araki-Woods factors satisfy property (AO) was proven in [Hou07, Chapter 4]. It follows from Theorem A that any von Neumann subal- gebra with expectation and with property Gamma of any free Araki-Woods factor is necessarily amenable. We also show in Proposition 3.3 that for every ω-solid von Neumann algebra M and every von Neumann subalgebra Q ⊂ M with expectation and with no amenable direct summand, the relative commutant Q′ ∩M ω is necessarily discrete and hence equal to Q′ ∩M (see Theorem 2.3). An interesting motivation for studying ω-solidity in the setting of type III factors is the fact that Connes's τ-invariant [Con74] is computable for all the ω-solid type III1 factors that possess faithful normal states with non-amenable centralizer. More precisely, we show in Proposition 3.10 that for every ω-solid factor M and for every faithful normal state ϕ ∈ M∗ such that the centralizer M ϕ is a non-amenable II1 factor, Connes's invariant τ (M ) is the weakest topology on R that makes the map R → Aut(M ) ∶ t ↦ σϕ Extending [Hou08, Theorem 1.2], we next show that the continuous cores of the free Araki-Woods factors associated with mixing orthogonal representations U ∶ R → O(HR) are ω-solid and so are their finite corners. t continuous. Theorem B. Let U ∶ R → O(HR) be any orthogonal representation on a separable real Hilbert space that is the direct sum of a mixing representation and a representation of dimension less than or 3The proof of [Oza04, Proposition 7] requires N0 = M and only shows that any finite diffuse von Neumann algebra that is solid and that has property Gamma is amenable. 2 equal to 1. Let M = Γ(HR , Ut)′′ be the associated free Araki-Woods factor. Then its continuous core c(M ) is an ω-solid type II∞ factor. For the proof of Theorem B, we can no longer rely on property (AO). Instead, we work within the framework of Popa's Deformation/Rigidity theory [Pop06a] and we apply Popa's spectral gap rigidity [Pop06b] to the free malleable deformation of the free Araki-Woods factors arising from second quantisation (see Subsection 2.2 for details). The proof of Theorem B is carried out in Section 3. When dealing with weakly mixing orthogonal representations U ∶ R → O(HR), we obtain a dichotomy result for all von Neumann subalgebras of the free Araki-Woods factors Γ(HR , Ut)′′ that are globally invariant under the modular automorphism group of the free quasi-free state. This result constitutes a new feature in the structure theory of type III factors. Theorem C. Let U ∶ R → O(HR ) be any weakly mixing orthogonal representation on a separable real Hilbert space and (M, ϕ) = (Γ(HR , Ut)′′, ϕU ) the associated free Araki-Woods factor. Let Q ⊂ M be any von Neumann subalgebra that is globally invariant under the modular automorphism group t ) of the free quasi-free state ϕ. Then either Q = C1 or Q is a full non-amenable type III1 factor (σϕ such that Q′ ∩ M ω = C1. Theorem C shows in particular that any amenable von Neumann subalgebra of M that is globally invariant under the modular automorphism group (σϕ t ) is necessarily trivial. Note that if the orthogonal representation U ∶ R → O(HR) is not weakly mixing then the centralizer M ϕ is not trivial. This shows that the assumption of U ∶ R → O(HR) being weakly mixing is necessary in Theorem C. The proof of Theorem C is based on the recent work of the first named author [Hou12a, Hou12b, Hou14] and uses in a novel fashion Popa's asymptotic orthogonality property [Pop83] in the framework of ultraproduct von Neumann algebras. The proof of Theorem C is carried out in Section 4. Acknowledgments This paper was completed when the first named author was visiting the Research Institute for Mathematical Sciences (RIMS) in Kyoto during Summer 2014. He warmly thanks Narutaka Ozawa and the RIMS for their kind hospitality. The authors also thank Stefaan Vaes for useful remarks regarding a first draft of this manuscript. Finally, the authors thank the anonymous referees for carefully reading the paper and providing valuable comments. 2 Preliminaries For a von Neumann algebra M, we will denote by Z(M ) the centre of M, by U(M ) the group of unitaries in M and by Ball(M ) the unit ball of M with respect to the uniform norm ⋅ ∞. Let now M be any σ-finite von Neumann algebra and ϕ ∈ M∗ any faithful normal state. We denote by L2(M, ϕ) (or simply L2(M ) when no confusion is possible) the GNS L2-completion of 3 M with respect to the inner product defined by ⟨x, y⟩ϕ = ϕ(y∗x) for all x, y ∈ M. We denote by Λϕ ∶ M → L2(M)∶ x ↦ Λϕ(x) the canonical embedding and by Jϕ ∶ L2(M) → L2(M) the canonical conjugation. We have xΛϕ(y) = Λϕ(xy) for all x, y ∈ M. ϕ = ϕ(x∗x+ xx∗)1~2 for all x ∈ M. Recall that on Ball(M), We will write xϕ = ϕ(x∗x)1~2 and x# ϕ ) coincides with the strong (resp. ∗-strong) topology. When the topology given by ⋅ϕ (resp. ⋅# ϕ = τ is a faithful normal tracial state, we will simply write x2 = τ(x∗x)1~2 for all x ∈ M. We will say that a von Neumann algebra M is tracial if it is endowed with a faithful normal tracial state τ. 2.1 The continuous core of a σ-finite von Neumann algebra Let (M, ϕ) be any σ-finite von Neumann algebra endowed with a faithful normal state. We denote by (σϕ t ) the modular automorphism group with respect to the state ϕ. The centraliser M ϕ of the state ϕ is by definition the fixed point algebra of (M,(σϕ t )). The continuous core of M with respect to ϕ, denoted by cϕ(M), is the crossed product von Neumann algebra M⋊σϕ R. The natural inclusion πϕ ∶ M → cϕ(M) and the unitary representation λϕ ∶ R → cϕ(M) satisfy the covariance relation λϕ(s)πϕ(x)λϕ(s)∗ = πϕ(σϕ s(x)) for all x ∈ M and all s ∈ R . Because of Connes's Radon-Nikodym cocycle theorem [Con73, Théorème 1.2.1] (see also [Tak03, There is a unique faithful normal conditional expectation Eϕ ∶ cϕ(M) → Lϕ(R) satisfying Eϕ(xλϕ(s)) = ϕ(x)λϕ(s). The semifinite faithful normal trace f ↦ ∫R exp(−t)f(t) on L∞(R) gives rise to a semifinite faithful normal trace Trϕ on Lϕ(R) via the Fourier transform. The formula Trϕ = Trϕ○Eϕ extends it to a semifinite faithful normal trace on cϕ(M). Theorem VIII.3.3]), the semifinite von Neumann algebra cϕ(M) together with its trace Trϕ does M, there is a canonical surjective ∗-isomorphism Πψ,ϕ ∶ cϕ(M) → cψ(M) such that Πψ,ϕ○ πϕ = πψ and Trψ○Πψ,ϕ = Trϕ. Note however that Πψ,ϕ does not map the subalgebra Lϕ(R) ⊂ cϕ(M) onto the subalgebra Lψ(R) ⊂ cψ(M). not depend on the choice of ϕ in the following precise sense. If ψ is another faithful normal state on 2.2 Free Araki-Woods factors Let U ∶ R → O(HR) be any orthogonal representation on a separable real Hilbert space. Denote by H = HR⊗R C the complexified Hilbert space of HR and by U ∶ R → U(H) the corresponding unitary representation. Let A be the positive selfadjoint closed operator defined on H satisfying Ait = Ut for all t ∈ R. Then there is an isometric embedding of HR into H given by HR → H ∶ ξ ↦ ξ, 2 1+ A−11~2 whose image we denote by KR. One can check that KR∩iKR ={0} and KR+iKR is dense in H. We denote by J the canonical conjugation on H = HR ⊕ iHR and by I = J A−1~2. Then I is an invertible anti-linear closed operator on H satisfying I = I−1. Oserve that KR = {ξ ∈ dom(T) ∶ Iξ = ξ}. From now on, we will simply write I ∶ ξ + iη ↦ ξ + iη = ξ − iη for all ξ, η ∈ KR. 4 The full Fock space of H is given by F(H) = CΩ ⊕ ∞ࣷ n=1 H⊗n. We call the vector Ω ∈ F(H) the vacuum vector. For all ξ ∈ H, the left creation operator ℓ(ξ) ∈ B(F(H)) is given by the formulae ℓ(ξ)Ω = ξ and ℓ(ξ)(ξ1 ⊗ ⋯ ⊗ ξn) = ξ ⊗ ξ1 ⊗ ⋯ ⊗ ξn. Note that ℓ(ξ) = ξ and ℓ(ξ) is an isometry if ξ = 1. Put W(ξ) = ℓ(ξ) + ℓ(ξ)∗ for all ξ ∈ KR. Following [Shl97], we define the free Araki-Woods factor associated with U ∶ R → O(HR) by Γ(HR , Ut)′′ ={W(ξ) ξ ∈ KR}′′. t . In particular, we have σϕU The vector state ϕU(x) =⟨xΩ, Ω⟩ on Γ(HR , Ut)′′ is called the free quasi-free state. It is faithful and = Ad(F(Ut)) for all one can show that the modular automorphism group of ϕU is given by σϕU t ∈ R, where F(Ut) = 1 ⊕ ࣷn≥1 U⊗n The GNS-representation of Γ(HR , Ut)′′ with respect to ϕU is isomorphic with its representation on F(H) with cyclic vector Ω. It is easy to check that for all n ≥ 1 and all ξ1, . . . , ξn ∈ KR + iKR there is a unique element W(ξ1 ⊗ ⋯ ⊗ ξn) ∈ Γ(HR , Ut)′′ such that W(ξ1 ⊗ ⋯ ⊗ ξn)Ω = ξ1 ⊗ ⋯ ⊗ ξn. We have W(ξ) = ℓ(ξ) + ℓ(ξ)∗ for all ξ ∈ KR + iKR. The following proposition describes a Wick-type formula Proposition 2.1 ([Hou12a, Hou12b, HR10]). Let ξj, ηk ∈ KR + iKR, for j, k ≥ 1. The following statements are true: t (W(ξ)) = W(Utξ) for all ξ ∈ KR. for such elements. t (i) The Wick formula W(ξ1 ⊗ ⋯ ⊗ ξn) = (ii) The product W(ξ1 ⊗ ⋯ ⊗ ξr)W(η1 ⊗ ⋯ ⊗ ηs) equals n Q k=0 ℓ(ξ1)⋯ℓ(ξk)ℓ(ξk+1)∗⋯ℓ(ξn)∗ holds. W(ξ1 ⊗ ⋯ ⊗ ξr ⊗ η1 ⊗ ⋯ ⊗ ηs) +⟨ξr, η1⟩W(ξ1 ⊗ ⋯ ⊗ ξr−1)W(η2 ⊗ ⋯ ⊗ ηs) . (iii) We have W(ξ1 ⊗ ⋯ ⊗ ξn)∗ = W(ξn ⊗ ⋯ ⊗ ξ1). (iv) The linear span of {1, W(ξ1 ⊗ ⋯ ⊗ ξn) ∶ n ≥ 1, ξi ∈ KR + iKR} forms a unital σ-strongly dense ∗-subalgebra of Γ(HR , Ut)′′. Proof. The proof of (i) is borrowed from [HR10, Lemma 3.2]. We prove the formula by induction on n. For n ∈{0, 1}, we have W(Ω) = 1 and we already observed that W(ξi) = ℓ(ξi) + ℓ(ξi)∗. Next, for ξ0 ∈ KR + iKR, we have W(ξ0)W(ξ1 ⊗ ⋯ ⊗ ξn)Ω = W(ξ0)(ξ1 ⊗ ⋯ ⊗ ξn) =(ℓ(ξ0) + ℓ(ξ0)∗)ξ1 ⊗ ⋯ ⊗ ξn = ξ0 ⊗ ξ1 ⊗ ⋯ ⊗ ξn +⟨ξ0, ξ1⟩ ξ2 ⊗ ⋯ ⊗ ξn. 5 Q 0≤j≤r−1 ℓ(ξ1)⋯ℓ(ξj)ℓ(ξj+1)∗⋯ℓ(ξr)∗ℓ(η1)∗⋯ℓ(ηs)∗ ℓ(ξ1)⋯ℓ(ξr)ℓ(η1)⋯ℓ(ηk)ℓ(ηk+1)∗⋯ℓ(ηs)∗ + Q 0≤k≤s +⟨ξr, η1⟩ Q 0≤j≤r−1,1≤k≤s Therefore W(ξ1 ⊗ ⋯ ⊗ ξr)W(η1 ⊗ ⋯ ⊗ ηs) is equal to ℓ(ξ1)⋯ℓ(ξj)ℓ(ξj+1)∗⋯ℓ(ξr−1)∗ℓ(η2)⋯ℓ(ηk)ℓ(ηk+1)∗⋯ℓ(ηs)∗ . So, we obtain W(ξ0 ⊗ ⋯ ⊗ ξn) = W(ξ0)W(ξ1 ⊗ ⋯ ⊗ ξn) −⟨ξ0, ξ1⟩W(ξ2 ⊗ ⋯ ⊗ ξn) = ℓ(ξ0)∗W(ξ1 ⊗ ⋯ ⊗ ξn) −⟨ξ0, ξ1⟩W(ξ2 ⊗ ⋯ ⊗ ξn) + ℓ(ξ0)W(ξ1 ⊗ ⋯ ⊗ ξn). Using the assumption for n and n − 1 and the relation ℓ(ξ0)∗ℓ(ξ1) =⟨ξ0, ξ1⟩, we obtain ℓ(ξ0)∗W(ξ1 ⊗ ⋯ ⊗ ξn) =⟨ξ0, ξ1⟩W(ξ2 ⊗ ⋯ ⊗ ξn) + ℓ(ξ0)∗ℓ(ξ1)∗⋯ℓ(ξn)∗ . Since ℓ(ξ0)W(ξ1 ⊗ ⋯ ⊗ ξn) gives the last n + 1 terms in the Wick formula at order n + 1 and ℓ(ξ0)∗ℓ(ξ1)∗⋯ℓ(ξn)∗ gives the first term, we are done. We now prove (ii). By the Wick formula, we have that W(ξ1 ⊗ ⋯ ⊗ ξr)W(η1 ⊗ ⋯ ⊗ ηs) is equal to Q 0≤j≤r,0≤k≤s ℓ(ξ1)⋯ℓ(ξj)ℓ(ξj+1)∗⋯ℓ(ξr)∗ℓ(η1)⋯ℓ(ηk)ℓ(ηk+1)∗⋯ℓ(ηs)∗ . Recall that we have ℓ(ξr)∗ℓ(η1) =⟨ξr, η1⟩. Therefore the above sum equals W(ξ1 ⊗ ⋯ ⊗ ξr ⊗ η1 ⊗ ⋯ ⊗ ηs) +⟨ξr, η1⟩W(ξ1 ⊗ ⋯ ⊗ ξr−1)W(η2 ⊗ ⋯ ⊗ ηs). It is now clear that (i) ⇒ (iii). Moreover, (iv) follows from (iii) using an induction procedure. 2.3 Ultraproduct von Neumann algebras Fix a non-principal ultrafilter ω ∈ β(N) ∖ N. Let M be any σ-finite von Neumann algebra. Define Iω(M) ={(xn)n ∈ ℓ∞(N, M) xn → 0 ∗ -strongly as n → ω} Mω(M) ={(xn)n ∈ ℓ∞(N, M)(xn)n Iω(M) ⊂ Iω(M) and Iω(M)(xn)n ⊂ Iω(M)} . ideal. Following [Ocn85, Chapter 5], we define the ultraproduct von Neumann algebra M ω by The multiplier algebra Mω(M) is a C∗-algebra and Iω(M) ⊂ Mω(M) is a norm closed two-sided M ω = Mω(M)~Iω(M). We denote the image of (xn)n ∈ Mω(M) by (xn)ω ∈ M ω. For all x ∈ M, the constant sequence (x)n lies in the multiplier algebra Mω(M). We will identify M with (M + Iω(M))~Iω(M) and regard M ⊂ M ω as a von Neumann subalgebra. The map Eω ∶ M ω → M ∶ (xn)ω ↦ σ-weak limn→ω xn is a faithful normal conditional expectation. For every 6 well-known proposition. faithful normal state ϕ ∈ M∗, the formula ϕω = ϕ ○ Eω defines a faithful normal state on M ω. Observe that ϕω((xn)ω) = limn→ω ϕ(xn) for all (xn)ω ∈ M ω. Put H = L2(M, ϕ). The ultraproduct Hilbert space Hω is defined to be the quotient of ℓ∞(N, H) by the subspace consisting in sequences (ξn)n satisfying limn→ωξnH = 0. We denote the image of (ξn)n ∈ ℓ∞(N, H) by (ξn)ω ∈ Hω. The inner product space structure on the Hilbert space Hω is defined by ⟨(ξn)ω,(ηn)ω⟩Hω = limn→ω⟨ξn, ηn⟩H. The GNS Hilbert space L2(M ω, ϕω) can be embedded into Hω as a closed subspace by Λϕω((xn)ω) ↦(Λϕ(xn))ω. Put xϕ = ϕ(⋅ x) and ϕx = ϕ(x ⋅) for all x ∈ M and all ϕ ∈ M∗. We will be using the following Proposition 2.2. Let (M, ϕ) be any σ-finite von Neumann algebra endowed with a faithful normal (i) For every λ > 0 and every (xn)n ∈ ℓ∞(N, M) satisfying limn→ωxnϕ − λϕxn = 0, we have (xn)n ∈ Mω(M) and (xn)ωϕω = λϕω(xn)ω. (ii) For every projection e ∈ M ω, there exists a sequence of projections (en)n ∈ Mω(M) such that e = (en)ω. If M is moreover diffuse, the projections en ∈ M may be chosen such that ϕ(en) = ϕω(e) for all n ∈ N. Proof. (i) Let (xn)n ∈ ℓ∞(N, M) such that limn→ωxnϕ − λϕxn = 0. Let (bn)n ∈ Iω(M). We may assume that max{xn∞,bn∞ ∶ n ∈ N} ≤ 1. Using the Cauchy-Schwarz inequality, for all n ∈ N, we state. have (xnbn# ϕ)2 = ϕ(b∗n x∗nxnbn) + ϕ(xn bnb∗nx∗n) ≤bnϕx∗nxnbnϕ + λ−1(xnϕ − λϕxn)(bnb∗nx∗n) + λ−1ϕ(bn b∗nx∗nxn) ≤bnϕ + λ−1xnϕ − λϕxnbnb∗nx∗n∞ + λ−1b∗nϕb∗nx∗nxnϕ ≤bnϕ + λ−1xnϕ − λϕxn + λ−1b∗nϕ. Therefore, we obtain limn→ωxnbn# ϕ = 0 and so (xnbn)n ∈ Iω(M). Likewise, for all n ∈ N, we have (bnxn# ϕ)2 = ϕ(x∗n b∗nbnxn) + ϕ(bn xnx∗nb∗n) ≤(λx∗nϕ − ϕx∗n)(b∗nbnxn) + λϕ(b∗n bnxnx∗n) +b∗nϕxnx∗nb∗nϕ ≤λx∗nϕ − ϕx∗nb∗nbnxn∞ + λbnϕbnxnx∗nϕ +b∗nϕ ≤xnϕ − λϕxn + λbnϕ +b∗nϕ. For the first part of the proof of (ii), see the proof of [Hou14, Proposition 2.4 (3)]. It remains to prove ϕ = 0 and so (bnxn)n ∈ Iω(M). This shows that (xn)n ∈ Therefore, we obtain limn→ωbnxn# Mω(M). Moreover, (xn)ωϕω = λϕω(xn)ω by [AH12, Lemma 4.36]. the moreover part of (ii) when M is diffuse. Let p ∈ M ω be any projection and (pn)n ∈ Mω(M) a sequence of projections such that p = (pn)ω. Let n ≥ 1. Assume that ϕ(pn) ≥ ϕω(p). Since pnM pn is diffuse, we may choose a projection rn ∈ pnM pn such that ϕ(rn) = ϕω(p). Assume that 7 ϕ(pn) ≤ ϕω(p). Since (1−pn)M(1−pn) is diffuse, we may choose a projection sn ∈(1−pn)M(1−pn) such that ϕ(sn) = ϕω(p) − ϕ(pn). Put rn = pn + sn. ϕ = limn→ωϕ(pn − rn) = 0 and hence (pn − rn)n ∈ Iω(M). Thus, we have We obtain limn→ωpn − rn2 p =(rn)ω and ϕ(rn) = ϕω(p) for all n ∈ N. The next theorem will be very useful to prove Theorem A. It is a generalization of [Ioa12, Lemma 2.7] to arbitrary von Neumann algebras. Theorem 2.3. Let Q ⊂ M be an inclusion of von Neumann algebras with faithful normal conditional expectation EQ ∶ M → Q. Assume that Q has separable predual. Denote by z ∈ Z(Q′ ∩ M ω) the unique maximal central projection such that (Q′ ∩ M ω)z is discrete. Then • z ∈ Z(Q′ ∩ M ω) ∩ Z(Q′ ∩ M), • (Q′ ∩(M ω)ϕω)(1 − z) is diffuse for all faithful normal states ϕ ∈ M∗ such that ϕ ○ EQ = ϕ and • (Q′ ∩ M ω)z =(Q′ ∩ M)z. We start by proving the following two lemmas. Lemma 2.4. Let Q ⊂ M be an inclusion of von Neumann algebras with faithful normal conditional expectation EQ ∶ M → Q. Assume that Q has separable predual. Let ϕ ∈ M∗ be any faithful normal state such that ϕ ○ EQ = ϕ. Denote by e ∈ Z(Q′ ∩(M ω)ϕω) the unique maximal central projection such that (Q′ ∩(M ω)ϕω)e is discrete. Then • e ∈ Z(Q′ ∩(M ω)ϕω) ∩ Z(Q′ ∩ M ϕ) and • (Q′ ∩(M ω)ϕω)e =(Q′ ∩ M ϕ)e. tion 2.5]). Put Q = Q′ ∩(M ω)ϕω and denote by e ∈ Z(Q) the unique maximal central projection in Q such that Qe is discrete. Choose a sequence of projections (en)n ∈ Mω(M) such that e =(en)ω. Let a = σ-weak limn→ω en ∈ Z(Q′ ∩ M ϕ). Next, we construct by induction a sequence of projections (fm)m≥1 in Q such that ϕω(efi) = ϕ(a2), ϕω(efia) = ϕ(a3) and ϕω(efifj) = ϕω(efia), ∀1 ≤ i < j. Proof. The proof is a generalisation of [Ioa12, Lemma 2.7] (see also the proof of [Hou14, Proposi- (2.1) Indeed, assume that f1, . . . , fm ∈ Q have been constructed. For every 1 ≤ j ≤ m, choose a sequence of projections (fj,n)n ∈ Mω(M) such that fj = (fj,n)ω. Let (xi)i∈N be a ⋅# ϕ -dense sequence in Ball(Q). Since e = (en)ω ∈(M ω)ϕω , limn→ωenxi − xien# ϕ = 0 for all i ∈ N and en → a σ-weakly as n → ω, we can find an increasing sequence (kn)n in N such that for every n ≥ 1, we have (P1) ekn ϕ − ϕekn ≤ 1 n, (P2) ekn xi − xiekn# ϕ ≤ 1 n for all 1 ≤ i ≤ n, 8 n for all 1 ≤ j ≤ m. (P3) ϕ(enekn) − ϕ(ena) ≤ 1 n, (P4) ϕ(enekna) − ϕ(ena2) ≤ 1 n and (P5) ϕ(enfj,nekn) − ϕ(enfj,na) ≤ 1 Properties (P1) and (P2) together with Proposition 2.2 imply that (ekn)n ∈ Mω(M) and f = (ekn)ω ∈ Q′ ∩(M ω)ϕω. Property (P3) implies that ϕω(ef) = ϕω(ea) = ϕ(a2), Property (P4) implies that ϕω(ef a) = ϕω(ea2) = ϕ(a3) and Property (P5) implies that ϕω(efjf) = ϕω(efja) for all 1 ≤ j ≤ m. We can now put fm+1 = f. This finishes the proof of the induction. Define pm = fme which is a projection in Qe. We have ϕω(pj) = ϕ(a2) and ϕω(pjpm) = ϕ(a3) countable direct sum of finite dimensional factors, Ball(Qe) is ⋅ϕω -compact. Thus, we may choose a subsequence (pmk)k≥1 that is ⋅ϕω-convergent in Ball(Qe). By the Cauchy-Schwarz inequality, for all 1 ≤ j < m. Observe that since Qe is a discrete tracial von Neumann algebra and hence a for all 1 ≤ j < k, we have ϕω(pmj pmk) − ϕω(pmj) =ϕω(pmj(pmk − pmj)) ≤pmj − pmkϕω . Put ε = 1 2f −EeM e(f)ϕω Taking the limit as (j, k) → ∞ and using (2.1), we obtain ϕ(a2) = ϕ(a3) and so 0 ≤ a ≤ 1 is a ϕ = ϕ(en) + ϕ(a) − 2ϕ(ena) → 0 as n → ω and so projection in Q′ ∩ M ϕ. Thus we have en − a2 e =(en)ω = a ∈ Z(Q′ ∩(M ω)ϕω) ∩ Z(Q′ ∩ M ϕ). It remains to prove that (Q′ ∩ (M ω)ϕω)e = (Q′ ∩ M ϕ)e. Assume by contradiction that this is not the case and choose a nonzero projection f ∈ (Q′ ∩(M ω)ϕω)e such that f ∉ (Q′ ∩ M ϕ)e. Let ϕe be the faithful normal state on eM e defined by ϕe = ϕ(e⋅e)ϕ(e) by EeM e ∶ (eM e)ω → eM e the canonical faithful normal conditional expectation. Recall that e > 0. Moreover, for all y ∈ Ball(eM e), ϕe○EeM e = ϕω we have e > 0 . e . Since f ∉(Q′∩M ϕ)e, we have f −EeM e(f)ϕω e ≥f − EeM e(f)ϕω . Recall that e ∈ M ϕ. Denote f − yϕω n ≥ 1, we have e , limk→ωf1,kxi − xif1,k# ϕe = 0 for all i ∈ N and limk→ωf1,k − fj,nϕe =f − fj,nϕω e and f1 = f ∈ Qe. Next, we construct by induction a sequence of projections fm ∈ Qe such that fp − fqϕω e ≥ ε for all p, q ≥ 1 such that p ≠ q. Assume that f1, . . . , fm ∈ Qe have been constructed. For every 1 ≤ j ≤ m, choose a sequence of projections (fj,n)n ∈ Mω(eM e) such that fj = (fj,n)ω. Let (xi)i∈N be a ⋅# ϕe-dense sequence in Ball(Qe). Since f = f1 = (f1,n)ω ∈ ((eM e)ω)ϕω e ≥ 2ε for all 1 ≤ j ≤ m and all n ∈ N, we can find an increasing sequence (kn)n in N such that for every (P1) f1,knϕe − ϕef1,kn ≤ 1 n, (P2) f1,knxi − xif1,kn# ϕe ≤ 1 n for all 1 ≤ i ≤ n and (P3) f1,kn − fj,nϕe ≥ ε for all 1 ≤ j ≤ m. (f1,kn)n ∈ Mω(eM e) and (f1,kn)ω ∈ (Qe)′ ∩((eM e)ω)ϕω By the same reasoning as before, Properties (P1) and (P2) together with Proposition 2.2 imply that e = Qe. Moreover, Property (P3) implies 9 of the induction. that (f1,kn)ω − fjϕω So, we have constructed a sequence of projections fm ∈ Qe such that fp − fqϕω such that p ≠ q. This however contradicts the fact that Ball(Qe) is ⋅ϕω e ≥ ε for all 1 ≤ j ≤ m. We can now put fm+1 =(f1,kn)ω. This finishes the proof e ≥ ε for all p, q ≥ 1 e -compact and finishes the proof Lemma 2.4. Lemma 2.5. Let Q ⊂ M be an inclusion of von Neumann algebras with faithful normal conditional expectation EQ ∶ M → Q. Assume that Q has separable predual. Let ϕ ∈ M∗ be any faithful normal state such that ϕ ○ EQ = ϕ. If Q′ ∩(M ω)ϕω = C then Q′ ∩ M ω = C. Proof. The proof is a straightforward generalisation of [AH12, Theorem 5.2] and so we will only sketch it. Assume that Q′ ∩(M ω)ϕω = C. Since (Q′ ∩ M ω)ϕω = Q′ ∩(M ω)ϕω = C, [AH12, Lemma 5.3] shows that Q′ ∩ M ω = C or Q′ ∩ M ω is a type III1 factor. By contradiction, assume that Q′ ∩ M ω is a type III1 factor. Choose (ai)i∈N a ⋅# ϕ -dense sequence in Ball(Q). Proceeding as in the proof of [AH12, Theorem 5.2], for all n ∈ N and all i, j ∈ {1, 2}, we find elements f(n)ij conditions of [AH12, Theorem 5.2, Claim 1] with respect to the sequence (ai)i∈N in Ball(Q). As in [AH12, Theorem 5.2, Claim 2], we obtain that (f(n)ij )n ∈ Mω(M) for all i, j ∈ {1, 2}. Finally, we obtain a projection g11 ∈ Q′ ∩(M ω)ϕω such that ϕω(g11) ≠ 0, 1. This is a contradiction and finishes ∈ M that satisfy the the proof of Lemma 2.5. Proof of Theorem 2.3. Let Q ⊂ M be an inclusion of von Neumann algebras and fix a faithful normal state ϕ ∈ M∗ such that ϕ ○ EQ = ϕ. By Lemma 2.4, the unique maximal central projection e ∈ Z(Q′ ∩(M ω)ϕω) such that (Q′ ∩(M ω)ϕω)e is discrete satisfies e ∈ Z(Q′ ∩(M ω)ϕω)∩Z(Q′ ∩ M ϕ) and (Q′ ∩(M ω)ϕω)e =(Q′ ∩ M ϕ)e. Observe that the projection e may a priori depend on the state ϕ. However, we will prove that this is not the case and show that the projection e satisfies the conclusion of Theorem 2.3. Since (Q′ ∩(M ω)ϕω)e is discrete, choose a family (pi)i∈I of pairwise orthogonal minimal projections in (Q′ ∩(M ω)ϕω)e such that ∑i∈I pi = e. We have pi ∈(Q′ ∩ M ϕ)e and (Qpi)′ ∩((piM pi)ω)ϕω pi = Cpi where ϕpi = ϕ(pi⋅pi) ϕ(pi) . Lemma 2.5 applied to the inclusion Qpi ⊂ piM pi implies that pi(Q′ ∩ M ω)pi =(Qpi)′ ∩(piM pi)ω = Cpi and hence pi is a minimal projection in e(Q′ ∩ M ω)e. Since ∑i∈I pi = e, we have that e(Q′ ∩ M ω)e is discrete. Denote by z(e) the central support of the projection e in Q′ ∩ M ω. We obtain that (Q′ ∩ M ω)z(e) is still discrete. Since z(e) ∈ Z(Q′ ∩ M ω) and Q′ ∩ M ω is globally invariant under the modular automorphism group (σϕω t ), we have z(e) ∈ Z(Q′ ∩(M ω)ϕω) and (Q′ ∩(M ω)ϕω)z(e) is discrete. By Lemma 2.4 and since e ≤ z(e), we obtain e = z(e) ∈ Z(Q′ ∩ M ω) ∩ Z(Q′ ∩ M) and (Q′ ∩ M ω)e is discrete. Observe that (Q′ ∩ M ω)(1 − e) is diffuse since (Q′ ∩(M ω)ϕω)(1 − e) is a diffuse subalgebra with expectation. Therefore, e ∈ Z(Q′∩M ω)∩Z(Q′∩M) is the unique projection such that (Q′ ∩ M ω)e is discrete and (Q′ ∩ M ω)(1 − e) is diffuse and hence e does not depend on the choice of the faithful normal state ϕ ∈ M∗ satisfying ϕ = ϕ ○ EQ. Thus, the above proof shows that (Q′ ∩(M ω)ψω)(1 − e) is diffuse for all faithful normal states ψ ∈ M∗ satisfying ψ = ψ ○ EQ. 10 It remains to prove that (Q′ ∩M ω)e =(Q′ ∩M)e. Recall that there exists a family (pi)i∈I of pairwise orthogonal minimal projections in (Q′ ∩ M ω)e such that ∑i∈I pi = e and pi ∈(Q′ ∩ M ϕ)e for all i ∈ I. In order to show that(Q′∩M ω)e =(Q′∩M)e, it suffices to prove that eF(Q′∩M ω)eF = eF(Q′∩M)eF for all finite subsets F ⊂ I, with eF = ∑i∈F pi ∈(Q′ ∩ M ϕ)e. Assume by contradiction that (Q′ ∩M ω)e ≠(Q′ ∩M)e. Hence there exists a finite subset F ⊂ I such that eF(Q′ ∩ M ω)eF ≠ eF(Q′ ∩ M)eF . For notational convenience, put q = eF and Q = Q′ ∩ M ω. Observe that qQq is discrete and finite. Let ϕq be the faithful normal state on qM q defined by . Recall that q ∈ Q′∩M ϕ. Denote by EqM q ∶(qM q)ω → qM q the canonical faithful normal ϕq = ϕ(q⋅q)ϕ(q) q . Since qQq =(Qq)′ ∩(qM q)ω is discrete, finite conditional expectation. Recall that ϕq ○ EqM q = ϕω q restricted to qQq is diagonalizable. For every and hence of type I, the faithful normal state ϕω eigenvalue λ > 0, we will denote by Eλ =(xn)ω ∈ qQq(xn)ωϕω q = λϕω q(xn)ω the spectral subspace of (qQq, ϕω q) corresponding to the eigenvalue λ. Since qQq ≠ q(Q′ ∩ M)q, we may choose an eigenvalue λ > 0 and a nonzero element f ∈ Ball(Eλ) such that f ∉ q(Q′ ∩ M)q. Since f ∉ q(Q′ ∩ M)q, we have f − EqM q(f)ϕω f − yϕω q > 0. Moreover, for all y ∈ Ball(qM q), we have q ≥f − EqM q(f)ϕω q > 0 . Put ε = 1 2f − EqM q(f)ϕω ϕq -dense sequence in Ball(Qq). Since f = f1 = (f1,n)ω ∈ Eλ, limk→ωf1,kxi − xif1,k# q and f1 = f ∈ Ball(Eλ). Next, we construct by induction a sequence of elements fm ∈ Ball(Eλ) such that fm − fpϕω q ≥ ε for all m, p ≥ 1 such that m ≠ p. As- sume that f1, . . . , fm ∈ Ball(Eλ) have been constructed. For every 1 ≤ j ≤ m, choose a sequence (fj,n)n ∈ Mω(qM q) such that fj,n ∈ Ball(qM q) for all n ∈ N and fj = (fj,n)ω. Let (xi)i∈N be a ⋅# ϕq = 0 for all i ∈ N and limk→ωf1,k − fj,nϕq = f − fj,nϕω q ≥ 2ε for all 1 ≤ j ≤ m and all n ∈ N, we can find an increasing sequence (kn)n in N such that for every n ≥ 1, we have (P1) f1,knϕq − λϕqf1,kn ≤ 1 n, (P2) f1,knxi − xif1,kn# n for all 1 ≤ i ≤ n and ϕq ≤ 1 (P3) f1,kn − fj,nϕq ≥ ε for all 1 ≤ j ≤ m. (f1,kn)n ∈ Mω(qM q) and (f1,kn)ω ∈ Ball(Eλ). Moreover, Property (P3) implies that (f1,kn)ω − e ≥ ε for all 1 ≤ j ≤ m. We can now put fm+1 =(f1,kn)ω. This finishes the proof of the induction. fjϕω So, we have constructed a sequence of elements fm ∈ Ball(Eλ) such thatfm−fpϕω q ≥ ε for all m, p ≥ 1 such that m ≠ p. However, since qQq is discrete and finite and hence a countable direct sum of finite dimensional factors, Ball(Eλ) is ⋅ϕω q -compact and hence we have obtained a contradiction. By the same reasoning as before, Properties (P1) and (P2) together with Proposition 2.2 imply that This finishes the proof of Theorem 2.3. Following [Con74], we define the asymptotic centraliser Mω of the von Neumann algebra M by Mω =(xn)n ∈ ℓ∞(N, M) ∀ψ ∈ M∗, lim n→ωxnψ − ψxn = 0~Iω(M) . By [Con74, Proposition 2.8], we have Mω =(M ′ ∩ M ω)ϕω for every faithful normal state ϕ ∈ M∗. 11 Corollary 2.6. Let M be any factor with separable predual such that M ′ ∩ M ω ≠ C1. Then Mω is diffuse. Proof. Let ϕ ∈ M∗ be any faithful normal state. We have Mω =(M ′ ∩ M ω)ϕω = M ′ ∩(M ω)ϕω. Since Mω ≠ C1 and since M ′ ∩ M = C1, the projection z obtained in Theorem 2.3 satisfies z = 0 and hence Mω is diffuse. 2.4 The Akemann-Ostrand property (AO) from the work of The Akemann-Ostrand property for von Neumann algebras arises Akemann-Ostrand [AO75] and was introduced by Ozawa in [Oza04]. A von Neumann algebra M ⊂ B(H) has property (AO) if there are unital σ-weakly dense C∗-subalgebras B ⊂ M and C ⊂ M ′ such that B is locally reflexive and such that the map B ⊗alg C → B(H)~K(H) ∶ x ⊗ y ↦ xy is continuous with respect to the minimal tensor C∗-norm. We recall the following well-known result. For a proof, we refer the reader to [Hou07, Chapter 4]. Proposition 2.7. Any free Araki-Woods factor satisfies property (AO). 2.5 Intertwining-by-bimodules techniques Popa introduced his powerful intertwining-by-bimodule techniques in [Pop02, Pop03, Pop04]. We first recall the intertwining-by-bimodule criterion in the case of finite von Neumann algebras. Let (M, τ) be any tracial von Neumann algebra together with von Neumann subalgebras A ⊂ 1AM1A and B ⊂ 1BM1B. Denote by EB ∶ 1BM1B → B the unique trace preserving faithful normal conditional expectation. Then the following statements are equivalent (see [Pop03, Lemma 2.1 and Corollary 2.3]): • There is n ≥ 1, a nonzero partial isometry v ∈ M1,n(1AM1B) and a possibly non-unital normal • There is no net of unitaries (wi)i in U(A) such that EB(x∗wiy) → 0 ∗-strongly as i → ∞ for ∗-homomorphism π ∶ A → Mn(B) such that av = vπ(a) for all a ∈ A. all x, y ∈ 1AM1B. We will say that A embeds into B inside M and write A ⪯M B if one of the above equivalent conditions is satisfied. Let (M, Tr) be any semifinite von Neumann algebra endowed with a semifinite faithful normal trace. Let B ⊂ M be any von Neumann subalgebra such that TrB is semifinite. Denote by EB ∶ M → B the unique trace preserving faithful normal conditional expectation. Let p ∈ M be any nonzero finite trace projection and A ⊂ pMp any von Neumann subalgebra. Let q ∈ B be any nonzero finite trace projection. Observe that p ∨ q is a nonzero finite trace projection in M. We will say that A embeds into qBq inside M and write A ⪯M qBq if A ⪯ (p∨q)M(p∨q) qBq in the usual sense for finite von Neumann algebras. 12 We will need the following useful intertwining-by-bimodule criterion for semifinite von Neumann algebras (see [HR10, Lemma 2.2] or [BHR14, Lemma 2.3]). Lemma 2.8. Let (M, Tr) be any semifinite von Neumann algebra endowed with a semifinite faithful normal trace. Let B ⊂ M be any von Neumann subalgebra such that TrB is semifinite. Denote by EB ∶ M → B the unique trace preserving faithful normal conditional expectation. Let p ∈ M be any nonzero finite trace projection and A ⊂ pMp any von Neumann subalgebra. The following conditions are equivalent: (i) For every nonzero finite trace projection q ∈ B, we have A M qBq. (ii) There exists an increasing sequence of nonzero finite trace projections qn ∈ B such that qn → 1 strongly and A M qnBqn for all n ∈ N. (iii) There exists a net of unitaries (wi) in U(A) such that limkEB(x∗wiy)2,Tr = 0 for all x, y ∈ pM. 2.6 Deformation/Rigidity theory for free Araki-Woods factors We introduce the s-malleable deformation of free Araki-Woods factors. It is an analogue of the malleable deformations considered in [Pop01, IPP05]. Let U ∶ R → O(HR) be any orthogonal representation on a separable real Hilbert space. Denote by (M, ϕ) = (Γ(HR , Ut)′′, ϕU) the associated free Araki-Woods factor together with its free quasi-free state ϕ. Put M = cϕ(M) and simply denote by Tr the canonical semifinite faithful normal trace on M. Furthermore, we write (M ,ϕ) = (Γ(HR ⊕ HR , Ut ⊕ Ut)′′, ϕU ⊕U) and M = cϕ(M). By [Shl97], there are ∗-isomorphisms (M ,ϕ) ≅ (M, ϕ) ∗(M, ϕ) and M ≅ M ∗Lϕ(R) M. We will identify M with its first copy in M and M with its first copy in M. The orthogonal representation V ∶ R → O(HR ⊕ HR) given by 2 s) − sin( π 2 s) Vs =cos( π 2 s) sin( π cos( π 2 s) commutes with the orthogonal representation U ⊕ U ∶ R → O(HR ⊕ HR). Hence the associated transformation Γ(Vs) on the free Fock space of HR ⊕ HR induces a ∗-automorphism αs of M . It satisfies α1(x ∗ 1) = 1 ∗ x for all x ∈ M. Since Γ(Vs) fixes the vacuum vector, it preserves the free quasi-free state ϕ on M. Hence it induces a trace preserving ∗-automorphism of M that we still denote by αs. Likewise, the orthogonal transformation induces a trace preserving ∗-automorphism β of M which moreover satisfies β2 = id M, βM = idM and βαs = α−sβ for all s ∈ R. Therefore, the deformation (αs, β)s∈R is s-malleable in the sense of Popa [Pop03] and satisfies the following transversality property. Proposition 2.9 (See [HR10, Proposition 4.2] and [Pop06b, Lemma 2.1]). x − α2s(x)2 ≤√2αs(x) −(EM ○ αs)(x)2 for all x ∈ L2(M, Tr) and all s ∈ R. 0 1 0 −1 13 3 Proofs of Theorems A and B 3.1 Preliminaries on ω-solidity Definition 3.1. Let ω ∈ β(N) ∖ N be a non-principal ultrafilter. We will say that a diffuse von Neumann algebra M is ω-solid if for every von Neumann subalgebra Q ⊂ M with expectation whose relative commutant Q′ ∩ M ω is diffuse, we have that Q is amenable. We first show that ω-solidity is stable under amplifications. is amenable and so is Q. Proof. If M is a diffuse ω-solid von Neumann algebra, then pM p is ω-solid for every nonzero Proposition 3.2. Let M be any diffuse ω-solid von Neumann algebra. Then p(M ⊗ B(ℓ2))p is ω-solid for every nonzero projection p ∈ M ⊗ B(ℓ2). projection p ∈ M. Indeed, let Q ⊂ pM p be any von Neumann subalgebra such that Q′ ∩(pM p)ω is diffuse. Put Q = Q ⊕ C(1 − p). Then Q′ ∩ M ω ⊃ Q′ ∩(pM p)ω ⊕(1 − p)M ω(1 − p) is diffuse. Thus Q It remains to prove that if M is a diffuse ω-solid von Neumann algebra, then M ⊗ B(ℓ2) is ω-solid. We may assume that M is not amenable. Observe that if M is properly infinite, then M ⊗B(ℓ2) ≅ M. Since any von Neumann algebra is the direct sum of a finite von Neumann algebra and a properly infinite von Neumann algebra, after cutting down by a central projection, we may assume that M is a diffuse ω-solid finite von Neumann algebra. Since M is the direct sum of a diffuse amenable von Neumann algebra and at most countably many non-amenable ω-solid II1 factors, after cutting down by a central projection, we may further assume without loss of generality that M is a non-amenable ω-solid II1 factor. We first prove that M t is ω-solid for all t > 0. Using the first paragraph of the proof, it suffices to corner of M, we obtain that M is not ω-solid. This is a contradiction. prove that M ⊗ Mn(C) is ω-solid for all n ≥ 1. Let Q ⊂ M ⊗ Mn(C) be any von Neumann subalgebra such that Q′ ∩(M ⊗ Mn(C))ω is diffuse. Assume by contradiction that Q is not amenable. We may choose a projection q ∈ Q such that qQq is not amenable and (τ ⊗ Trn)(q) ≤ 1~n. Note that (qQq)′ ∩(q(M ⊗ Mn(C))q)ω = (Q′ ∩(M ⊗ Mn(C))ω)q is diffuse. Regarding q(M ⊗ Mn(C))q as a We now prove that M = M ⊗ B(ℓ2) is ω-solid. Let Q ⊂ M be any von Neumann subalgebra with faithful normal conditional expectation EQ ∶ M → Q and such that Q′ ∩ Mω is diffuse. Denote by Eω ∶ Mω → M the canonical faithful normal conditional expectation. Let ϕ ∈ M∗ be any faithful normal state such that ϕ○EQ = ϕ. By Theorem 2.3, the relative commutant Q′ ∩(Mω)ϕω is diffuse as well. Fix a tracial faithful normal semifinite weight Tr on M. By [AH12, Lemma 4.26], Trω = Tr ○Eω is a tracial faithful normal semifinite weight on Mω. Denote by T ∈ L1(M, Tr)+ the unique positive selfadjoint operator affiliated with M satisfying ϕ = Tr(T ⋅). By [Con73, Lemme 1.2.3 (b) and Lemme 1.4.4], we have ϕω = Trω(T ⋅). Q = Q ∨ B. Since ϕω = Trω(T ⋅), we have Mϕω = B ′ ∩ Mω and hence Q′ ∩ Mω = Q′ ∩ B ′ ∩ Mω = Q′ ∩(Mω)ϕω Denote by B ⊂ M the von Neumann subalgebra generated by all the spectral projections of T . Put 14 is diffuse. Observe that Q is globally invariant under the modular automorphism group (σϕ t ). Since T ∈ L1(M, Tr), we may choose a sequence of finite trace projections pk ∈ B ⊂ Q such that (pkQpk)′ ∩(pkMpk)ω =(Q′ ∩Mω)pk =(Q′ ∩(Mω)ϕω)pk is diffuse, we have that pkQpk is amenable. pk → 1 strongly. Since pkMpk is an ω-solid II1 factor by the first part of the proof and since Since amenability is stable under direct limits, we finally obtain that Q is amenable. Since Q ⊂ Q is a von Neumann subalgebra with expectation, Q is also amenable. This finishes the proof of Proposition 3.2. Next, we prove a useful characterisation of ω-solidity. Proposition 3.3. Let M be any von Neumann algebra with separable predual that has no amenable direct summand. The following conditions are equivalent. (i) For every von Neumann subalgebra Q ⊂ M with expectation, if Q′ ∩ M ω is diffuse then Q is amenable. (ii) For every von Neumann subalgebra Q ⊂ M with expectation that has no amenable direct summand, the relative commutant Q′ ∩ M ω is discrete. Proof. (i) ⇒ (ii). Let Q ⊂ M be any von Neumann subalgebra with expectation that has no amenable direct summand. By Theorem 2.3, there is a unique central projection z ∈ Z(Q′ ∩ M ω) ∩ Z(Q′ ∩ M) such that (Q′ ∩ M ω)z is diffuse and (Q′ ∩ M ω)(1 − z) is discrete. Put Q = Qz ⊕ C(1 − z). Since Q′ ∩ M ω ⊃(Q′ ∩ M ω)z ⊕(1 − z)M ω(1 − z) is diffuse, we have that Q is amenable. Thus z = 0 and Q′ ∩ M ω is discrete. (ii) ⇒ (i). Let Q ⊂ M be any von Neumann subalgebra with expectation such that Q′ ∩ M ω is diffuse. Denote by z ∈ Z(Q) the unique central projection such that Qz has no amenable direct summand and Q(1 − z) is amenable. Since (1 − z)M(1 − z) has no amenable direct summand, Q = Qz ⊕(1 − z)M(1 − z) has no amenable direct summand either. Then Q′ ∩ M ω is discrete and so is (Q′ ∩ M ω)z =(Q′ ∩ M ω)z. Thus z = 0 and Q is amenable. 3.2 Proof of Theorem A Proof of Theorem A. Let M be any von Neumann algebra with separable predual that satisfies property (AO). Denote by Eω ∶ M ω → M the canonical faithful normal conditional expectation. Let Q ⊂ M be any von Neumann subalgebra with faithful normal conditional expectation EQ ∶ M → Q and such that Q′ ∩ M ω is diffuse. Fix a faithful normal state ϕ ∈ M∗ satisfying ϕ ○ EQ = ϕ. By Theorem 2.3, Q′∩(M ω)ϕω is diffuse and hence there is a sequence of unitaries Uk in U(Q′∩(M ω)ϕω) such that Uk → 0 weakly as k → ∞. Choose a sequence (uk m ∈ Ball(M) ϕ -dense sequence in Ball(Q) and (ψj)j≥1 be a for all m ∈ N and Uk = (uk ⋅-dense sequence in M∗. There exists an increasing sequence(kn)n in N such that for every n ∈ N, we have limm→ωψj(ukn m) = (ψj ○ Eω)(Ukn) < 1 n for all 1 ≤ j ≤ n. Therefore, there exists an increasing sequence (mn)n in N such that for every n ∈ N, the element un = ukn m)m ∈ Mω(M) such that uk m)ω. Let (xi)i≥1 be a ⋅# mn ∈ Ball(M) satisfies 15 n for all 1 ≤ j ≤ n. (P1) unϕ − ϕun ≤ 1 n, nun# n# n and 1 − unu∗ (P2) 1 − u∗ ϕ ≤ 1 ϕ ≤ 1 n (P3) unxi − xiun# ϕ ≤ 1 n for all 1 ≤ i ≤ n and (P4) ψj(un) ≤ 1 Property (P1) and Proposition 2.2 show that (un)n ∈ Mω(M) and together with Properties (P2) and (P3) they show that U = (un)ω ∈ U(Q′ ∩(M ω)ϕω). Finally, Property (P4) shows that un → 0 Regard M ⊂ B(H) where the Hilbert space H is given by property (AO). Define the unital completely positive map Θ ∶ B(H) → B(H) by Θ(T) = σ-weak limn→ω unT u∗ n. Observe that Θ(x) = Eω(U xU ∗) ∈ M for all x ∈ M and hence ΘM is normal. Next, define the unital com- weakly as n → ∞. pletely positive maps Ψk = 1 k k Q j=1 Θ○j ∶ B(H) → B(H) and let Ψ be the pointwise σ-weak limit of (Ψk)k, as k → ω. Since Θ(M) ⊂ M, we have Ψk(M) ⊂ M for all k ≥ 1 and hence Ψ(M) ⊂ M. Note that since ΘM is normal, we have ΨM = Θ ○ ΨM . Since U ∈ U((M ω)ϕω), we also have ϕ ○ ΘM = ϕ. This implies that ϕ ○ ΨkM = ϕ for all k ≥ 1 and hence ϕ ○ ΨM = ϕ. Put Q ={U, U ∗}′ ∩ M. Observe that Q ⊂ Q ⊂ M and that Q is globally invariant under the modular t ). Let x ∈ M and put y = Ψ(x) ∈ M. We have y = Ψ(x) = Θ(Ψ(x)) = automorphism group (σϕ Θ(y) = Eω(U yU ∗). Since U ∈ U((M ω)ϕω) and since yϕω =yϕ, we obtain U yU ∗ϕω =yϕ and hence y − U yU ∗2 ϕω − 2 Re ϕω(U y∗U ∗y) ϕω =y2 = 2y2 = 2y2 = 2y2 ϕω +U yU ∗2 ϕ − 2 Re ϕ(Eω(U y∗U ∗y)) ϕ − 2 Re ϕ(Eω(U y∗U ∗)y) ϕ − 2 Re ϕ(y∗y) = 0 . Therefore, we have y = U yU ∗ and hence Ψ(x) = y ∈ Q. Combining this with the fact that Θ(x) = x for all x ∈ Q, we see that ΨM is a norm one projection onto Q. We already saw that ΨM is ϕ-preserving and hence we infer that ΨM = EQ ∶ M → Q is the unique ϕ-preserving conditional j=1 EQ(bi) ci. By definition of Ψ, we have Ψ(c) = c Define ΦQ ∶ M ⊗alg M ′ → B(H) ∶ ∑n for all c ∈ M ′. Therefore [Cho74, Theorem 3.1] implies that for all n ≥ 1, all bi ∈ M and all ci ∈ M ′, we have expectation from M onto Q. j=1 bi ⊗ ci ↦ ∑n Ψ( n Q j=1 bi ci) = n Q j=1 Ψ(bi) ci = n Q j=1 EQ(bi) ci = ΦQ( n bi ⊗ ci) . Q j=1 The fact that un → 0 σ-weakly as n → ω implies that Θ(T) = σ-weak limn→ω unT u∗ n = 0 for all T ∈ K(H). This shows that Ψ(T) = 0 for all T ∈ K(H). Hence K(H) ⊂ ker Ψ. Denote by 16 π ∶ B(H) → B(H)~K(H) the canonical quotient map. Then there exists a unital completely positive map Ψ ∶ B(H)~K(H) → B(H) such that Ψ = Ψ ○ π. By property (AO) of M, there is a unital σ-weakly dense locally reflexive C∗-subalgebra B ⊂ M and a unital σ-weakly dense C∗-subalgebra C ⊂ M ′ together with a ∗-homomorphism ν ∶ B ⊗alg C → B(H)~K(H) ∶ n Q j=1 that is continuous with respect to the minimal tensor norm on B ⊗alg C. Therefore ΦQ = Ψ ○ ν is continuous with respect to the minimal tensor norm on B ⊗alg C. Applying [Oza04, Lemma 5], we obtain that Q is amenable and so is Q. bi ⊗ ci ↦ π( n bi ci) Q j=1 3.3 Proof of Theorem B The next theorem is a generalisation of [Hou08, Theorem 3.4] regarding the position of the rel- ative commutant of large subalgebras of the continuous core of free Araki-Woods factors in the ultraproduct framework. state ϕ. Theorem 3.4. Let U ∶ R → O(HR) be any orthogonal representation on a separable real Hilbert space. Denote by (M, ϕ) =(Γ(HR , Ut)′′, ϕU) the corresponding free Araki-Woods factor together with its free quasi-free state and by M = cϕ(M) the continuous core associated with the free quasi-free Then for every nonzero finite trace projection p ∈ Lϕ(R) ⊂ M and every von Neumann subalgebra q ∈ Lϕ(R) such that Q ⊂ pMp that has no amenable direct summand, there exists a nonzero finite trace projection Q′ ∩ pMωp ⪯Mω Lϕ(R)ωq . Proof of Theorem 3.4. The proof is very much inspired by [Pet09, Theorems 4.3 and 4.5] (see also [Hou12b, Theorem D]). Let αt ∶ M → M be the trace preserving s-malleable deformation introduced in Subsection 2.6. Write M = M ∗Lϕ(R) α1(M). Observe that if (xn)n ∈ Iω(M), then also (αt(xn))n ∈ Iω(M) for all t ∈ R. It follows that (αt) extends to a one-parameter family of trace preserving ∗-automorphisms of the ultraproduct von Neumann algebra Mω that we denote t ). We emphasise however that t ↦ αt(x) need not be continuous when x ∈ Mω. by (αω Step 1: Uniform convergence in ⋅2 of (αω t ) on Ball(Q′ ∩ pMωp). Assume by contradiction that (αω t ) does not converge uniformly in ⋅2 on Ball(Q′ ∩ pMωp). Thus there exist c > 0, a sequence (tk)k of positive reals such that limk tk = 0 and a sequence (Xk)k in Ball(Q′ ∩ pMωp) such that Xk − αω 2tk(Xk)2 ≥ 2c for all k ∈ N. Write Xk =(xk,n)ω with xk,n ∈ Ball(pMp) satisfying limn→ωyxk,n − xk,ny2 = 0 and Xk − αω 2tk(Xk)2 = limn→ωxk,n − α2tk(xk,n)2 for all k ∈ N and all Denote by I the directed set of all pairs (F , ε) with ε > 0 and F ⊂ Ball(Q) finite subset. Let i = (F , ε) ∈ I. Choose k ∈ N large enough so that a − αtk(a)2 ≤ ε~3 for all a ∈ F. Then choose n ∈ N large enough so that xk,n − α2tk(xk,n)2 ≥ c and axk,n − xk,na2 ≤ ε~3 for all a ∈ F. y ∈ Q. 17 Put ξi = αtk(xk,n) − EpMp(αtk(xk,n)) ∈ L2(pMp) ⊖ L2(pMp). By Proposition 2.9, we have ξi2 ≥ 1√2xk,n − α2tk(xk,n)2 ≥ c√2 . For all x ∈ pMp, we have xξi2 =(1 − EpMp)(xαtk(xk,n))2 ≤xαtk(xk,n)2 ≤x2 . By Popa's spectral gap argument [Pop06b], for all a ∈ F, we have aξi − ξia2 =(1 − EpMp)(aαtk(xk,n) − αtk(xk,n)a)2 ≤aαtk(xk,n) − αtk(xk,n)a2 Lemma 5.1]), it follows that Q has an amenable direct summand by Connes' characterisation of ≤ 2a − αtk(a)2 +axk,n − xk,na2 ≤ ε. Hence ξi ∈ L2(pMp) ⊖ L2(pMp) is a net of vectors satisfying lim supixξi2 ≤x2 for all x ∈ pMp, and limiaξi − ξia2 = 0 for all a ∈ Q. Since the pMp-pMp-bimodule L2(pMp)⊖ lim inf iξi2 ≥ c√2 L2(pMp) is weakly contained in the coarse pMp-pMp-bimodule L2(pMp) ⊗ L2(pMp) (see [HR10, amenability [Con76]. This is a contradiction and hence (αω t ) does converge uniformly in ⋅2 on Ball(Q′ ∩ pMωp). We now proceed by contradiction and assume that Q′∩pMωp Mω Lϕ(R)ωq for every nonzero finite trace projection q ∈ Lϕ(R). By Lemma 2.8, there exists a net (Uk)k of unitaries in U(Q′ ∩ pMωp) such that limkELϕ(R)ω(X ∗UkY)2 = 0 for all X, Y ∈ pMω. Step 2: Uniform convergence in ⋅2 of (αt) on Ball(Q). Take ε > 0. Since (αω t ) converges uniformly in ⋅2 on Ball(Q′ ∩ pMωp), there is some t0 > 0 such that for all t ∈ [0, t0], we have t (X)−X2 < ε2~4 for all X ∈ Ball(Q′∩pMωp). We show that for all t ∈ [0, t0] and all x ∈ Ball(Q), αω we have αt(x) − x2 < ε. Take t ∈ [0, t0] and x ∈ Ball(Q). Let (yi)i be a ⋅ 2-dense sequence in Ball(pM). There is an increasing sequence (kn)n such that for every n ≥ 1, the unitary Ukn ∈ U(Q′ ∩ pMωp) satisfies m ∈ Ball(pMp) for ELϕ(R)ω(y∗ all m ∈ N. There exists an increasing sequence (mn)n in N such that for every n ≥ 1, the element vn = ukn i Ukn yj)2 < 1~n for all i, j ∈ {1, . . . , n}. Write Ukn = (ukn m )ω with ukn mn ∈ Ball(pMp) satisfies n − x2 ≤ 1~n, • vnxv∗ • ELϕ(R)(y∗ • αt(vn) − vn2 ≤ ε2~4. i vnyj)2 ≤ 1~n for all i, j ∈ {1, . . . , n}, and Since (yi)i is ⋅ 2-dense in Ball(pM), the second condition implies that ELϕ(R)(a∗vnb)2 → 0 for all a, b ∈ pM. Writing now δt(x) = αt(x) − EpMp(αt(x)) ∈ p Mp ⊖ pMp, we obtain δt(x)2 2 = ⟨δt(x), δt(x)⟩ ≤ ⟨δt(vnxv∗ ≤ ⟨vnδt(x)v∗ ≤ ⟨vnδt(x)v∗ n), δt(x)⟩ + vnxv∗ n, δt(x)⟩ + vnxv∗ n, δt(x)⟩ + 1~n + ε2~2 . n − x2 n − x2 + 2vn − αt(vn)2 18 Observe moreover that by Cauchy-Schwarz inequality, we have n, δt(x)⟩ = Tr(δt(x)∗vnδt(x)v∗ n) ⟨vnδt(x)v∗ = Tr(EpMp(δt(x)∗vnδt(x))v∗ n) ≤ EpMp(δt(x)∗vnδt(x))2. Since δt(x) ∈ p( M⊖M) and since limn ELϕ(R)(a∗vnb)2 = 0 for all a, b ∈ pM, by [BHR14, Theorem 2.5, Claim], it follows that limn EpMp(δt(x)∗vnδt(x))2 = 0 and hence limn ⟨vnδt(x)v∗ n, δt(x)⟩ = 0. Hence, the transversality property of Proposition 2.9 now yields x − α2t(x)2 ≤ √2δt(x)2 ≤ ε. Thus, (αt) converges uniformly in ⋅ 2 on Ball(Q). Step 3: Deducing a contradiction. Since (αt) converges uniformly in ⋅ 2 on Ball(Q), [HR10, Theorem 4.3] implies that there exists a nonzero finite trace projection r ∈ Lϕ(R) such that Q ⪯M Lϕ(R)r. Since Lϕ(R)r is amenable, it follows that Q has an amenable direct summand, con- tradicting our assumption that it does not. It follows that the assumption Q′ ∩pMωp Mω Lϕ(R)ωq for every nonzero finite trace projection q ∈ Lϕ(R) of the previous step is wrong. This finishes the proof of the theorem. Before we can proceed to the proof of Theorem B, we need a few basic results regarding mixing inclusions in semifinite amalgamated free products. Recall that an inclusion of tracial von Neumann algebras B ⊂ (M, τ) is mixing if for every uniformly bounded net (wk)k of elements in B that goes to 0 weakly, we have k EB(xwky)2 = 0, ∀x, y ∈ M ⊖ B . lim Let now (M, Tr) be any semifinite von Neumann algebra endowed with a semifinite faithful normal trace. Let B ⊂ M be any von Neumann subalgebra such that TrB is semifinite. Denote by EB ∶ M → B the unique trace preserving faithful normal conditional expectation. Definition 3.5. Keep the same notation. We will say that the inclusion B ⊂ M is mixing if for every nonzero finite trace projection q ∈ B and for every uniformly bounded net (wk)k in qBq that goes to 0 weakly, we have k EB(x∗wky)2 = 0, ∀x, y ∈ q(M ⊖ B) . lim We prove a useful characterisation of mixing inclusions of semifinite von Neumann algebras. Lemma 3.6. Keep the same notation. The following conditions are equivalent. (i) The inclusion B ⊂ M is mixing. (ii) For every nonzero finite trace projection q ∈ B, the inclusion of tracial von Neumann algebras qBq ⊂ qMq is mixing. (iii) There exists an increasing sequence of nonzero finite trace projections qn ∈ B such that the inclusion of tracial von Neumann algebras qnBqn ⊂ qnMqn is mixing for all n ∈ N. 19 Proof. (i) ⇒ (ii) ⇒ (iii) are obvious. For (iii) ⇒ (i), let q ∈ B be a nonzero finite trace projection, (wk)k a net of elements in Ball(qBq) that goes to 0 weakly and x, y ∈ Ball(M) ∩ q(M ⊖ B). Take ε > 0. Since Tr(q) < ∞ and qn → 1 strongly, there exists n ∈ N such that q − qnq2 + q − qqn2 + x∗q − qnx∗q2 + qy − qyqn2 ≤ ε (3.1) . 4 This implies in particular that for all k, we have (3.2) wk − qnwkqn2 ≤ wk − qnwk2 + qn(wk − wkqn)2 ≤ ε 4 . Since the inclusion of tracial von Neumann algebras qnBqn ⊂ qnMqn is mixing, since qnqyqn, qnx∗qqn ∈ qnMqn ⊖ qnBqn and since qnwkqn → 0 weakly as k → ∞, there exists k0 such that for all k ≥ k0, we have (3.3) EB(qnx∗qqn qnwkqn qnqyqn)2 = Tr(qn)1~2 EqnBqn(qnx∗qqn qnwkqn qnqyqn)2,τqnMqn ≤ ε 2 . Combining (3.1), (3.2) and (3.3), we obtain EB(x∗wky)2 ≤ EB(qnx∗q wk qyqn)2 + ε 4 ≤ EB(qnx∗q qnwkqn qyqn)2 + ≤ ε. ε 2 An interesting class of mixing inclusions of semifinite von Neumann algebras arises from modular automorphism groups. Proposition 3.7. Let (M, ϕ) be any von Neumann algebra together with a faithful normal state t ) is mixing, that is, for all x, y ∈ M, we have such that the modular automorphism group (σϕ t (x)y) = ϕ(x)ϕ(y). Denote by cϕ(M) the continuous core associated with ϕ. Then the limt→∞ ϕ(σϕ inclusion Lϕ(R) ⊂ cϕ(M) is mixing. Proof. By Fourier transform, identify Lϕ(R) with L∞(R). The proof of [Hou08, Theorem 3.7] shows that the inclusion of tracial von Neumann algebras Lϕ(R)q ⊂ qcϕ(M)q is mixing for all nonzero projections q corresponding to the bounded intervals of the form [−T, T] with T > 0. Then Lemma 3.6 shows that the inclusion Lϕ(R) ⊂ cϕ(M) is mixing. let B ⊂ Mi be an inclusion of von Neumann algebras with faithful normal For all i ∈ {1, 2}, conditional expectation Ei ∶ Mi → B. Assume that B is semifinite with faithful normal semifinite tracial weight Tr. Assume moreover that Tr ○Ei is still a semifinite trace on Mi. Consider the amalgamated free product von Neumann algebra (M, E) = (M1, E1) ∗B (M2, E2) and observe that Tr ○E is still a faithful normal semifinite trace on M (see [BHR14, Section 2.2]). We say in that case that M = M1 ∗B M2 is a semifinite amalgamated free product von Neumann algebra. We prove the analogue of [Hou12a, Proposition 4.7] in the setting of semifinite amalgamated free product von Neumann algebras. 20 Proposition 3.8. Let M = M1 ∗B M2 be a semifinite amalgamated free product von Neumann algebra. Assume that the inclusion B ⊂ M2 is mixing. Then the inclusion M1 ⊂ M is mixing. Proof. Denote by EM1 ∶ M → M1 the unique trace preserving faithful normal conditional ex- pectation. To prove that the inclusion M1 ⊂ M is mixing, using Kaplansky's density theo- rem and Lemma 3.6, it suffices to show that for all nonzero finite trace projections q ∈ B, all nets (wk)k of elements in Ball(qM1q) that go to 0 weakly and all elements x, y ∈ q(M ⊖ M1) of the form x = qx1⋯x2m+1 and y = qy1⋯y2n+1 with m, n ≥ 1, x1, x2m+1, y1, y2n+1 ∈ Ball(M1), x2, . . . , x2m, y2, . . . , y2n ∈ Ball(M2)∩(M2⊖B) and x3, . . . , x2m−1, y3, . . . , y2n−1 ∈ Ball(M1)∩(M1⊖B), we have k EM1(x∗wky)2 = 0 . lim Using the property of freeness with amalgamation over B, we have EM1(x∗wky) = EM1(x∗ = EM1(x∗ = EM1(x∗ 2m+1⋯x∗ 2m+1⋯x∗ 2m+1⋯x∗ 2 x∗ 2 EB(x∗ 3 EB(x∗ 1qwkqy1 y2⋯y2n+1) 1 wky1) y2⋯y2n+1) 2 EB(x∗ 1 wky1) y2) y3⋯y2n+1) . Take ε > 0. Since Tr(q) < +∞, we may choose a large enough finite trace projection p ∈ B such that qy1 − qy1p2 + x∗ 1 q − px∗ 1q2 ≤ ε . We infer that EB(x∗ 1wky1) − EB(px∗ EM1(x∗wky) − EM1(x∗ 1q wk qy1p)2 ≤ ε for all k and hence 2m+1⋯x∗ k lim sup 3 EB(x∗ Since the inclusion B ⊂ M2 is mixing, since (EB(px∗ weakly and since px2, py2 ∈ p(M2 ⊖ B), it follows that limk EB(x∗ and hence 2 EB(px∗ 1 q wk qy1p) y2) y3⋯y2n+1)2 ≤ ε . 1 q wk qy1p))k is a net in Ball(pBp) that goes to 0 1q wk qy1p) py2)2 = 0 2 p EB(px∗ 2m+1⋯x∗ 3 EB(x∗ lim k EM1(x∗ 1 q wk qy1p) y2) y3⋯y2n+1)2 = 0 . Since ε > 0 is arbitrary, we deduce that 2 EB(px∗ This implies that lim supk EM1(x∗wky)2 ≤ ε. limk EM1(x∗wky)2 = 0. Proof of Theorem B. By Proposition 3.2, it suffices to prove that finite corners of continuous cores of free Araki-Woods factors are ω-solid. Let U ∶ R → O(HR) be any orthogonal representation on a separable real Hilbert space that is the direct sum of a mixing representation and a representation of dimension less than or equal to 1. Denote by (M, ϕ) = (Γ(HR , Ut)′′, ϕU) the associated free Araki-Woods factor together with its free quasi-free state and M = cϕ(M) its continuous core with respect to the free quasi-free state ϕ. Observe that M is a type III1 factor and hence M is a type II∞ factor. Let p ∈ Lϕ(R) be any nonzero finite trace projection and Q ⊂ pMp any von Neumann subalgebra such that Q′ ∩(pMp)ω is diffuse. 21 Assume by contradiction that Q is not amenable. Let z ∈ Z(Q) be a nonzero central projection such that Qz has no amenable direct summand. Since pMp is a II1 factor and since Lϕ(R)p is diffuse, there exists u ∈ U(pMp) and q ∈ Lϕ(R)p such that uzu∗ = q. So up to conjugating by a unitary and taking a smaller projection in Lϕ(R)p, we may assume without loss of generality that Q ⊂ pMp has no amenable direct summand and that Q′ ∩ (pMp)ω is diffuse. By Theorem 3.4, we know that there exists a nonzero finite trace projection q ∈ Lϕ(R) such that Q′ ∩ (pMp)ω ⪯Mω (Lϕ(R)q)ω. Up to replacing q by p ∨ q ∈ Lϕ(R), we may assume that p ≤ q. If (Ut) is mixing, then [Hou08, Proposition 2.4] and Proposition 3.7 show that the inclusion Lϕ(R) ⊂ M is mixing. Applying [Ioa12, Lemma 9.5], we obtain that Q ⪯qMq Lϕ(R)q. If (Ut) is the direct sum of a mixing orthogonal representation with an orthogonal representation of dimension one, then M = N ∗ L(Z), where N is the Araki-Woods factor associated with the mixing part of (Ut). Writing N = cϕN (N), we obtain M ≅ N ∗Lϕ(R) (L(Z) ⊗ Lϕ(R)). Hence Proposition 3.8 shows that the inclusion L(Z) ⊗ Lϕ(R) ⊂ M is mixing. Moreover, we know that Q′ ∩ (pMp)ω ⪯ (qMq)ω (Lϕ(R)q)ω and hence Q′ ∩ (pMp)ω ⪯ (qMq)ω ((L(Z) ⊗ Lϕ(R))q)ω. Applying [Ioa12, Lemma 9.5], we obtain that Q ⪯qMq (L(Z) ⊗ Lϕ(R))q. However, in both cases, this contradicts the fact that Q has no amenable direct summand. 3.4 Computation of Connes's τ -invariant for ω-solid factors Let M be any von Neumann algebra with separable predual. We endow Aut(M) with the topology of pointwise convergence in M∗, that is, αn → id in Aut(M) as n → ∞ if and only if n→∞ϕ ○ αn − ϕ = 0 for all ϕ ∈ M∗ . lim Endowed with this topology, Aut(M) becomes a Polish group. Recall from [Con74] that when M is a factor, we have that M is full if and only if the subgroup Inn(M) of inner automorphisms is closed in Aut(M). Equivalently, we have Mω = C1 for some (or any) ω ∈ β(N) ∖ N. In that case, the quotient group Out(M) = Aut(M)~Inn(M) endowed with the quotient topology is a Polish group. We will denote by π ∶ Aut(M) → Out(M) the quotient homomorphism. By Connes's Radon-Nikodym cocycle theorem [Con73, Théorème 1.2.1] (see also [Tak03, Theorem VIII.3.3]), the homomorphism δ ∶ R → Out(M) ∶ t ↦ π(σϕ t ) is well-defined and does not depend on the choice of a particular state on M. Definition 3.9 ([Con74]). Let M be a full factor of type III1 with separable predual. We define τ(M) to be the weakest topology that makes the map δ ∶ R → Out(M) continuous. It is typically difficult to calculate Connes's τ-invariant for arbitrary type III1 factors. In the case of the free Araki-Woods factors M = Γ(HR , Ut)′′, using a 14ε-type argument, it is proven in [Shl97, Vae06] that τ(M) is the weakest topology that makes the map R → O(HR) ∶ t ↦ Ut strongly continuous. In the next proposition, we show that Connes's τ-invariant is computable for a fairly large class of ω-solid type III1 factors. Our proof no longer relies on a 14ε-type argument and works in great generality. 22 t continuous. Proposition 3.10. Let M be any ω-solid factor of type III1 with separable predual and ϕ ∈ M∗ any faithful normal state whose centralizer is a non-amenable II1 factor. Then M is a full factor and τ(M) is the weakest topology on R that makes the map R → Aut(M) ∶ t ↦ σϕ Proof. Let M be any ω-solid factor of type III1 with separable predual and ϕ ∈ M∗ a faithful normal state whose centralizer is a non-amenable II1 factor. Since M is a non-amenable ω-solid factor, M ′ ∩ M ω is discrete by Proposition 3.3 and hence M ′ ∩ M ω = C1 by Corollary 2.6. This implies that M is a full factor. We next have to show that if (tn)n is a sequence in R that converges to 0 with respect to τ(M), then σϕ By Theorem A and Proposition 3.3, the relative commutant (M ϕ)′ ∩ M ω is discrete. Applying Theorem 2.3, we have that (M ϕ)′ ∩ M ω = (M ϕ)′ ∩ M. Since ((M ϕ)′ ∩ M)ϕ = (M ϕ)′ ∩ M ϕ = C1, [AH12, Lemma 5.3] implies that (M ϕ)′ ∩ M = C1 or (M ϕ)′ ∩ M is a factor of type III1. Since (M ϕ)′ ∩ M ω = (M ϕ)′ ∩ M is discrete, we obtain that (M ϕ)′ ∩ M ω = C1. Observe that this implies that (M ϕ)′ ∩ M ω = C1 for all non-principal ultrafilter ω ∈ β(N) ∖ N. Now take a sequence (tn)n in R that converges to 0 with respect to τ(M). Then there is a tn → id in Aut(M). Fix ω ∈ β(N) ∖ N a sequence of unitaries (un)n in M such that (Ad un) ○ σϕ non-principal ultrafilter. As in the proof of [Ued11, Proposition 3.1], we have that (un)n ∈ Mω(M) and (un)ω ∈ (M ϕ)′ ∩ M ω. Indeed, for all n ∈ N, we have n = ϕ ○ (Ad un) − ϕ = ϕ ○ (Ad un) ○ σϕ tn = ϕ ○ (Ad un) ○ σϕ tn → id in Aut(M). nϕ − ϕu∗ u∗ tn − ϕ ○ σϕ tn − ϕ. nϕ − ϕu∗ tn − ϕ = 0 and hence tn − ϕ = 0, we have limn→ω ϕ ○ (Ad un) ○ σϕ Since limn→∞ ϕ ○ (Ad un) ○ σϕ n = 0. Therefore (un)n ∈ Mω(M) and (un)ω ∈ (M ω)ϕω by Proposition 2.2. limn→ω u∗ tn(x) → x strongly as n → ∞ for all x ∈ M ϕ. This implies that We moreover have (Ad un) ○ σϕ n − xϕ = 0 for all x ∈ M ϕ. Since (un)n ∈ Mω(M) and (un)ω ∈ (M ω)ϕω, we finally limn→ω unxu∗ obtain (un)ω ∈ (M ϕ)′ ∩ M ω. Since (M ϕ)′ ∩ M ω = C1, we have limn→ω un − ϕ(un)1ϕ = (un)ω − ϕω((un)ω)ϕω = 0. Since this is true for every ω ∈ β(N) ∖ N, we obtain limn→∞ un − ϕ(un)1ϕ = 0. Proceeding now exactly as in the proof of [Con74, Theorem 5.2], we conclude that σϕ tn Aut(M). → id in 4 Proof of Theorem C We first recall a basic fact on ε-orthogonality. Definition 4.1. Let H be a complex Hilbert space and ε ≥ 0. Two (not necessarily closed) subspaces K, L ⊂ H are called ε-orthogonal if ⟨ξ, η⟩ ≤ εξη for all ξ ∈ K and all η ∈ L. In that case, we will denote K ⊥ε L. Proposition 4.2 ([Hou12a, Proposition 2.3]). There is a continuous function δ ∶ [0, 1~2) → R≥0 satisfying δ(0) = 0 and the following property. If k ≥ 1 and 0 ≤ ε < 1~2 are such that δ○(k−1)(ε) < 1~2, 23 then for all projections pi ∈ B(H), i ∈ {1, . . . , 2k}, satisfying piH ⊥ε pjH for all i, j ∈ {1, . . . , 2k}, i ≠ j, we have 2k Q i=1piξ2 ≤ k−1 M j=0(1 + δ○j(ε))2P ξ2 , where P = ⋁2k i=1 pi is the projection onto the closed linear span span ⋃2k i=1 piH. The main result of this section is the following asymptotic orthogonality result in the framework of ultraproducts of free Araki-Woods factors and is inspired by [Pop83, Lemma 2.1]. Theorem 4.3. Let U ∶ R → O(HR) be any weakly mixing orthogonal representation on a separable real Hilbert space and (M, ϕ) = (Γ(HR , Ut)′′, ϕU) the associated free Araki-Woods factor. Then for all x, y ∈ (M ω)ϕω ⊖ C1 and all a, b ∈ M ⊖ C1, we have ϕω(b∗y∗ax) = 0. Proof. Let H = HR ⊕ iHR and denote by H = F(H) the full Fock space. We view KR + iKR ⊂ H as a dense subspace of H. Put κt = id ⊕ ࣷn≥1 U ⊗n ∈ U(H). For every x ∈ M, we have t t (x)Ω = κt(xΩ) . σϕ Since the linear span of 1 and of all the reduced words W(ξ1 ⊗⋯⊗ ξm) with m ≥ 1 and ξj ∈ KR +iKR is a unital σ-strongly dense ∗-subalgebra of M, it suffices to prove the result when a = W(ξ1⊗⋯⊗ξk) and b = W(η1 ⊗ ⋯ ⊗ ηℓ) are reduced words with ξ1, . . . , ξk, η1, . . . , ηℓ ∈ KR + iKR. Approximating ηj ∈ KR + iKR by 1[λ−1,λ](A)(ηj) ∈ KR + iKR for all 1 ≤ j ≤ ℓ and for λ > 1 sufficiently large, we may further assume that ηj = 1[λ−1,λ](A)(ηj) for all 1 ≤ j ≤ ℓ. It follows that the map R → KR + iKR ∶ t ↦ Utηj can be extended to an entire analytic function which takes values in KR + iKR for all 1 ≤ j ≤ ℓ. This implies that the map R → M ∶ t ↦ W(Utηℓ ⊗ ⋯ ⊗ Utη1) can be extended to an t (W(ηℓ ⊗ ⋯ ⊗ η1)) = W(Utηℓ ⊗ ⋯ ⊗ Utη1) for all t ∈ R, M-valued entire analytic function. Since σϕ we obtain that W(ηℓ ⊗ ⋯ ⊗ η1) is analytic for the modular automorphism group (σϕ t ) and we have σϕ z (W(ηℓ ⊗ ⋯ ⊗ η1)) = W(Aizηℓ ⊗ ⋯ ⊗ Aizη1) for all z ∈ C. From now on and for the rest of the proof, define L = span(ξk, ξk, η1, η1) ⊂ KR + iKR. We will use the following notation: • X1 ⊂ H is the closed subspace generated by the linear span of all the reduced words e1 ⊗⋯⊗en with n ≥ 1 and such that e1 ∈ L. • X2 ⊂ H is the closed subspace generated by the linear span of all the reduced words e1 ⊗⋯⊗en with n ≥ 1 and such that en ∈ L. • Y ⊂ H is the closed subspace generated by the linear span of all the reduced words e1 ⊗ ⋯⊗ en with n ≥ 1 and such that e1, en ∈ (KR + iKR) ∩ L⊥. Observe that we have CΩ ⊕ (X1 + X2) ⋅ϕ ⊕ Y = H . Claim 1. Let ε ≥ 0 and t ∈ R be such that Ut(L) ⊥ε~ dim(L) L. Then for all i ∈ {1, 2}, we have κt(Xi) ⊥ε Xi . 24 Choose an orthonormal basis (ζ1, . . . , ζdim(L)) of L. We first prove the claim for X1. We will identify X1 with L ⊗ H using the following unitary defined by V1 ∶ H ⊗ H ∋ ζ ⊗ (e1 ⊗ ⋯ ⊗ en) ↦ ζ ⊗ e1 ⊗ ⋯ ⊗ en ∈ H , for all n ≥ 1 and all ζ, e1, . . . , en ∈ H. Observe that κtV1 = V1(Ut ⊗ κt) for all t ∈ R. Let ξ, η ∈ X1 be such that ξ = ∑dim(L) ζj ⊗ νj with µi, νj ∈ H. Further observe that ξ2 = ∑dim(L) νj2. We have κtξ = ∑dim(L) µi2 and η2 = ∑dim(L) ζi ⊗ µi and η = ∑dim(L) Utζi ⊗ κtµi and hence j=1 i=1 j=1 i=1 i=1 ⟨κtξ, η⟩ ≤ dim(L) Q i,j=1 ⟨Utζi, ζj⟩µiνj . Since ⟨Utζi, ζj⟩ ≤ ε~ dim(L), we obtain ⟨κtξ, η⟩ ≤ εξη by the Cauchy-Schwarz inequality. Next, we prove the claim for X2. We identify X2 with H ⊗ L using the unitary defined by V2 ∶ H ⊗ H ∋ (e1 ⊗ ⋯ ⊗ en) ⊗ ζ ↦ e1 ⊗ ⋯ ⊗ en ⊗ ζ ∈ H , for all n ≥ 1 and all ζ, e1, . . . , en ∈ H. Observe that κtV2 = V2(κt ⊗ Ut) for all t ∈ R. Let ξ, η ∈ X2 be such that ξ = ∑dim(L) νj ⊗ ζj with µi, νj ∈ H. Further observe that ξ2 = ∑dim(L) νj2. We have κtξ = ∑dim(L) µi2 and η2 = ∑dim(L) µi ⊗ ζi and η = ∑dim(L) κtµi ⊗ Utζi and hence j=1 i=1 j=1 i=1 i=1 ⟨κtξ, η⟩ ≤ dim(L) Q i,j=1 ⟨Utζi, ζj⟩µiνj . Since ⟨Utζi, ζj⟩ ≤ ε~ dim(L), we obtain ⟨κtξ, η⟩ ≤ εξη by the Cauchy-Schwarz inequality. This finishes the proof of the claim. Claim 2. For every x = (xn)ω ∈ (M ω)ϕω, we have and n→ωPX1(xnΩ)ϕ = 0 lim n→ωPX2(xnΩ)ϕ = 0 . lim Let x ∈ (M ω)ϕω. We may assume that x ∈ Ball((M ω)ϕω ) and then choose a sequence (xn)n ∈ Mω(M) such that xn ∈ Ball(M) for all n ∈ N and x = (xn)ω. For all i ∈ {1, 2}, all t ∈ R and all n ∈ N, we have PXi(xnΩ)2 ϕ ϕ = κtPXi(xnΩ)2 ≤ 2κtPXi(xnΩ) − Pκt(Xi)(xnΩ)2 = 2Pκt(Xi)(κt(xnΩ) − xnΩ)2 ≤ 2σϕ t (xn) − xn2 ϕ + 2Pκt(Xi)(xnΩ)2 ϕ. ϕ + 2Pκt(Xi)(xnΩ)2 ϕ ϕ + 2Pκt(Xi)(xnΩ)2 ϕ Furthermore, [AH12, Theorem 4.1] says that for all t ∈ R (xn)ω = x = σϕω t (x) = (σϕ t (xn))ω 25 holds. This implies that limn→ω xn − σϕ Fix p ≥ 1. Choose ε > 0 very small according to Proposition 4.2 so that ∏p−1 j=0(1 + δ○j(ε))2 ≤ 2. Since U ∶ R → O(HR) is weakly mixing and since L is finite dimensional, with ε′ = ε~ dim(L), we can choose inductively t1, . . . , t2p ∈ R such that ϕ = 0 for all t ∈ R. t (xn)# Using Claim 1, this implies that Utj(L) ⊥ε′ Uti(L), ∀1 ≤ i < j ≤ 2p . κtj(X1) ⊥ε κti(X1) and κtj(X2) ⊥ε κti(X2), ∀1 ≤ i < j ≤ 2p . Thus, using the above inequalities and Proposition 4.2, we obtain ϕ = lim n→ω 2p Q ϕ lim n→ω ≤ lim n→ω 2pPXi(xnΩ)2 j=1κtj PXi(xnΩ)2 2p 2σϕ tj(xn) − xn2 Q j=1 4xn2 ϕ . ϕ ≤ 22−p for all p ≥ 1. Thus, we have limn→ω PXi(xnΩ)ϕ = 0. We conclude that limn→ω PXi(xnΩ)2 This finishes the proof of the claim. Claim 3. The subspaces W(ξ1 ⊗ ⋯ ⊗ ξk) Y and Jϕσϕ −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ Y are orthogonal in H. Let m, n ≥ 1 and e1, . . . , em, f1 . . . , fn ∈ KR + iKR. Assume moreover that e1, em, f1, f n ∈ L⊥ so that e1 ⊗ ⋯ ⊗ em ∈ Y and f1 ⊗ ⋯ ⊗ fn ∈ Y. Then by Proposition 2.1 (ii) and since ξk, η1 ∈ L, we have 2Pκtj (Xi)(xnΩ)2 ϕ + lim n→ω ≤ lim n→ω 2p Q j=1 ϕ W(ξ1 ⊗ ⋯ ⊗ ξk)(e1 ⊗ ⋯ ⊗ em) = W(ξ1 ⊗ ⋯ ⊗ ξk)W(e1 ⊗ ⋯ ⊗ em)Ω = W(ξ1 ⊗ ⋯ ⊗ ξk ⊗ e1 ⊗ ⋯ ⊗ em)Ω = ξ1 ⊗ ⋯ ⊗ ξk ⊗ e1 ⊗ ⋯ ⊗ em , Jϕσϕ −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ (f1 ⊗ ⋯ ⊗ fn) = W(f1 ⊗ ⋯ ⊗ fn)W(η1 ⊗ ⋯ ⊗ ηℓ)Ω = W(f1 ⊗ ⋯ ⊗ fn ⊗ η1 ⊗ ⋯ ⊗ ηℓ)Ω = f1 ⊗ ⋯ ⊗ fn ⊗ η1 ⊗ ⋯ ⊗ ηℓ. Since ⟨ξ1, f1⟩ = 0, we see that the vectors W(ξ1 ⊗ ⋯ ⊗ ξk)(e1 ⊗ ⋯ ⊗ em) and Jϕσϕ −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ (f1 ⊗ ⋯ ⊗ fn) are orthogonal in H. Finally, using the density of the linear span of the words e1 ⊗ ⋯ ⊗ em and f1 ⊗ ⋯ ⊗ fn in Y finishes the proof of the claim. We are now ready to finish the proof of Theorem 4.3. Let x, y ∈ (M ω)ϕω ⊖ C1. Using Claim 2 and the fact that limn→ω PC Ω(xnΩ)ϕ = 0, we have Λϕω(ax) = (W(ξ1 ⊗ ⋯ ⊗ ξk) xnΩ)ω = (W(ξ1 ⊗ ⋯ ⊗ ξk) PY(xnΩ))ω , Λϕω(yb) = (Jϕσϕ = (Jϕσϕ −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ ynΩ)ω −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ PY(ynΩ))ω . 26 By Claim 3, we know that W(ξ1 ⊗ ⋯ ⊗ ξk) PY(xnΩ) ⊥ Jϕσϕ −i~2(W(ηℓ ⊗ ⋯ ⊗ η1))Jϕ PY(ynΩ) , for all n ∈ N. Hence Λϕω(ax) ⊥ Λϕω(yb) in Hω, which implies that ϕω(b∗y∗ax) = 0. Theorem 4.4. Let U ∶ R → O(HR) be any weakly mixing orthogonal representation on a separable real Hilbert space and (M, ϕ) = (Γ(HR , Ut)′′, ϕU) the associated free Araki-Woods factor. Let Q ⊂ M be any von Neumann subalgebra such that Q′ ∩ (M ω)ϕω ≠ C1. Then Q = C1. Proof. Assume that Q′ ∩ (M ω)ϕω ≠ C1. Choose a projection e ∈ Q′ ∩ (M ω)ϕω such that e ∉ {0, 1}. Then choose a sequence of projections (en)n ∈ Mω(M) such that e = (en)ω and limn→ω σϕ t (en) − ϕ = 0 for all t ∈ R. Put a = σ-weak limn→ω en ∈ Q′ ∩ M ϕ. Since M ϕ = C1, we obtain a = ϕ(a)1. en# Since e ∉ {0, 1}, we have ϕ(a) ∉ {0, 1}. Let y ∈ Q ⊖ C1. By Theorem 4.3, we have (e − ϕ(a)1)y2 ϕω = ϕω(y∗(e − ϕ(a)1)∗(e − ϕ(a)1)y) = ϕω(y∗(e − ϕ(a)1)∗y(e − ϕ(a)1)) = 0 Moreover, ϕω = lim n→ω(en−ϕ(a)1)y2 ϕ = lim n→ω⟨(en−2ϕ(a)en+ϕ(a)21)yΩ, yΩ⟩ϕ = (ϕ(a)−ϕ(a)2)y2 ϕ . (e−ϕ(a)1)y2 Since ϕ(a) ∉ {0, 1}, it follows that y = 0 and hence Q = C1. Proof of Theorem C. Let Q ⊂ M be any von Neumann subalgebra that is globally invariant under the modular automorphism group (σϕ t ). There is a unique ϕ-preserving faithful normal conditional expectation EQ ∶ M → Q. Assume that Q′ ∩ M ω ≠ C1. Then we have Q′ ∩ (M ω)ϕω ≠ C1 by Lemma 2.5. Therefore, we obtain Q = C1 by Theorem 4.4. References [AH12] [AO75] H. Ando and U. Haagerup. Ultraproducts of von Neumann algebras. J. Funct. Anal. 266, 6842 -- 6913, 2014. C. A. Akemann and P. A. Ostrand. On a tensor product C∗ -algebra associated with the free group on two generators. J. Math. Soc. Japan 27, 589 -- 599, 1975. [BHR14] R. Boutonnet, C. Houdayer, and S. Raum. Amalgamated free product type III factors with at most one Cartan subalgebra. Compos. Math. 150 (1), 143 -- 174, 2014. [Cho74] M. D. Choi. A Schwarz inequality for positive linear maps on C∗-algebras. Illinois J. Math. 18, 565 -- 574, 1974. [Con73] A. Connes. Une classification des facteurs de type III. Ann. Sci. Éc. Norm. Supér. (4) 6, 133 -- 252, 1973. [Con74] A. Connes. Almost periodic states and factors of type III1. J. Funct. Anal. 16, 415 -- 445, 1974. [Con76] A. Connes. Classification of injective factors. Cases II1, II∞, IIIλ, λ ≠ 1. Ann. Math. (2) 74, 73 -- 115, 1976. 27 [Hou07] [Hou08] C. Houdayer. Sur la classification de certaines algèbres de von Neumann. PhD thesis, Université de Paris VII, 2007. C. Houdayer. Structural results for free Araki-Woods factors and their continuous cores. J. Inst. Math. Jussieu 9 (4), 741 -- 767, 2010. [Hou12a] C. Houdayer. A class of II1 factors with an exotic abelian maximal amenable subalgebra. Trans. Am. Math. Soc. 366, 3693 -- 3707, 2014. [Hou12b] C. Houdayer. Structure of II1 factors arising from free Bogoljubov actions of arbitrary groups. [Hou14] [HR10] [Ioa12] [IPP05] Adv. Math. 260, 414 -- 457, 2014. C. Houdayer. Gamma stability in free product von Neumann algebras. arXiv:1403.4098, to appear in Commun. Math. Phys. C. Houdayer and É. Ricard. Approximation properties and absence of Cartan subalgebra for free Araki-Woods factors. Adv. Math. 228 (2), 764 -- 802, 2011. A. Ioana. Cartan subalgebras of amalgamated free product II1 factors. arXiv:1207.0054, to appear in Ann. Sci. Éc. Norm. Supér. A. Ioana, J. Peterson, and S. Popa. Amalgamated free products of weakly rigid factors and calculation of their symmetry groups. Acta Math. 200 (1), 85 -- 153, 2008. [Ocn85] A. Ocneanu. Actions of discrete amenable groups on von Neumann algebras, volume 1138 of [Oza04] [Oza10] [Pet09] [Pop83] [Pop01] [Pop02] [Pop03] [Pop04] Lecture Notes in Mathematics. Berlin-Heidelberg-New York: Springer-Verlag, 1985. N. Ozawa. Solid von Neumann algebras. Acta Math. 192 (1), 111 -- 117, 2004. N. Ozawa. A comment on free group factors. In Noncommutative harmonic analysis with applica- tions to probability II, volume 89 of Banach Center Publ., pages 241 -- 245. Polish Acad. Sci. Inst. Math., Warsaw, 2010. J. Peterson. L2-rigidity in von Neumann algebras. Invent. Math. 175 (2), 417 -- 433, 2009. S. Popa. Maximal injective subalgebras in factors associated with free groups. Adv. Math. 50, 27 -- 48, 1983. S. Popa. Some rigidity results for non-commutative Bernoulli shifts. J. Funct. Anal. 230 (2), 273 -- 328, 2006. S. Popa. On a class of type II1 factors with Betti numbers invariants. Ann. Math. (2) 163 (3), 809 -- 899, 2006. S. Popa. Strong rigidity of II1 factors arising from malleable actions of w-rigid groups. I. Invent. Math. 165 (2), 369 -- 408, 2006. S. Popa. Strong rigidity of II1 factors arising from malleable actions of w-rigid groups. II. Invent. Math. 165 (2), 409 -- 451, 2006. [Pop06a] S. Popa. Deformation and rigidity for group actions and von Neumann algebras. In M. Sanz-Solé et al., editors, Proceedings of the international congress of mathematicians, Madrid, Spain, Au- gust 22 -- 30, 2006, volume I: Plenary lectures and ceremonies, pages 445 -- 477. Zürich: European Mathematical Society, 2007. [Pop06b] S. Popa. On the superrigidity of malleable actions with spectral gap. J. Am. Math. Soc. 21 (4), [Shl97] [Shl98] 981 -- 1000, 2008. D. Shlyakhtenko. Free quasi-free states. Pac. J. Math. 177 (2), 329 -- 368, 1997. D. Shlyakhtenko. Some applications of freeness with amalgamation. J. Reine Angew. Math. 500, 191 -- 212, 1998. D. Shlyakhtenko. A-valued semicircular systems. J. Funct. Anal. 166 (1), 1 -- 47, 1999. [Shl99] [Tak03] M. Takesaki. Theory of operator algebras III. Berlin-Heidelberg-New York: Springer-Verlag, 2003. [Ued11] [Vae06] Y. Ueda. On type III1 factors arising as free products. Math. Res. Lett. 18 (5), 909 -- 920, 2011. S. Vaes. États quasi-libres libres et facteurs de type III. In Séminaire Bourbaki. 2003/2004, volume 937 of Astérisque. Paris: Société Mathématique de France, 2007. 28 [VV05] S. Vaes and R. Vergnioux. The boundary of universal discrete quantum groups, exactness, and factoriality. Duke Math. J. 140 (1), 35 -- 84, 2007. Cyril Houdayer CNRS - Université Paris-Est - Marne-la-Vallée LAMA UMR 8050 77454 Marne-la-Vallée cedex 2 France [email protected] Sven Raum RIMS Kitashirakawa-oiwakecho 606-8502 Sakyo-ku, Kyoto Japan [email protected] 29
1410.2051	5	1410	2017-01-09T19:30:56	Inverse semigroup actions on groupoids	[ "math.OA" ]	We define inverse semigroup actions on topological groupoids by partial equivalences. From such actions, we construct saturated Fell bundles over inverse semigroups and non-Hausdorff \'etale groupoids. We interpret these as actions on C*-algebras by Hilbert bimodules and describe the section algebras of these Fell bundles. Our constructions give saturated Fell bundles over non-Hausdorff \'etale groupoids that model actions on locally Hausdorff spaces. We show that these Fell bundles are usually not Morita equivalent to an action by automorphisms. That is, the Packer-Raeburn Stabilisation Trick does not generalise to non-Hausdorff groupoids.	math.OA	math	INVERSE SEMIGROUP ACTIONS ON GROUPOIDS ALCIDES BUSS AND RALF MEYER Abstract. We define inverse semigroup actions on topological groupoids by partial equivalences. From such actions, we construct saturated Fell bundles over inverse semigroups and non-Hausdorff étale groupoids. We interpret these as actions on C ∗-algebras by Hilbert bimodules and describe the section alge- bras of these Fell bundles. Our constructions give saturated Fell bundles over non-Hausdorff étale groupoids that model actions on locally Hausdorff spaces. We show that these Fell bundles are usually not Morita equivalent to an action by automorphisms. That is, the Packer -- Raeburn Stabilisation Trick does not generalise to non- Hausdorff groupoids. Contents Introduction Inverse semigroup actions on groupoids Inverse semigroup actions on C∗-algebras 1. 2. Partial equivalences 3. 3.1. Compatibility with order and involution 3.2. Transformation groupoids 3.3. Examples: group actions and actions on spaces 3.4. Morita invariance of actions by partial equivalences 3.5. Local centralisers 3.6. Decomposing proper Lie groupoids 4. 5. Fell bundles from actions of inverse semigroups 5.1. A Haar system on the transformation groupoid 5.2. Construction of the Fell bundle 5.3. Another construction of the Fell bundle 6. Actions of inverse semigroups and groupoids 6.1. The motivating example 6.2. 7. Actions by automorphisms are not enough 8. A simple explicit example Appendix A. Preliminaries on topological groupoids A.1. Topological groupoids, principal bundles, and equivalences A.2. Basic actions versus free and proper actions A.3. Covering groupoids and equivalence Appendix B. Fields of Banach spaces over locally Hausdorff spaces B.1. Proof of Theorem 5.5 References Inverse semigroup models for étale groupoids 2 4 11 14 18 20 23 25 26 27 31 32 32 35 36 38 40 43 44 46 47 49 50 53 56 57 2010 Mathematics Subject Classification. 46L55, 20M18, 22A22. Key words and phrases. Inverse semigroups, groupoids, actions, partial equivalences, Fell bun- dles, stabilisation trick. Supported by CNPq/CsF (Brazil) and the German Research Foundation (Deutsche Forschungs- gemeinschaft (DFG)) through the grant "Actions of 2-groupoids on C*-algebras." 1 2 ALCIDES BUSS AND RALF MEYER 1. Introduction Two of the most obvious actions of a groupoid G are those by left and right If G is Hausdorff, they induce continuous translations on its arrow space G1. actions of G on the C∗-algebra C0(G1). What happens if G is non-Hausdorff? Let G be a non-Hausdorff, étale groupoid with Hausdorff, locally compact object space G0. Then G1 is locally Hausdorff, that is, it has an open covering U = (Ui)i∈I by Hausdorff subsets: we may choose Ui so that the range and source maps restrict to homeomorphisms from Ui onto open subsets of the Hausdorff space G0. The covering U yields an étale, locally compact, Hausdorff groupoid H with object space H 0 := Fi∈I Ui, arrow space H 1 := Fi,j∈I Ui ∩ Uj, range and source maps r(i, j, x) := (i, x) and s(i, j, x) := (j, x), and multiplication (i, j, x) · (j, k, x) = (i, k, x). The groupoid H is known as the Čech groupoid for the covering U. In noncommutative geometry, we view the groupoid C∗-algebra C∗(H) as the algebra of functions on the non-Hausdorff space G1. Is there some kind of action of G on C∗(H) that corresponds to the translation action of G on G1? There is no action of G on C∗(H) in the usual sense because there is no action of G on H by automorphisms. The problem is that arrows g ∈ G1 have many liftings (i, g) ∈ H 0. To let g ∈ G act on H, we must choose k ∈ I with gh ∈ Uk for h ∈ Uj with r(h) = s(g). It may, however, be impossible to choose k continuously when h varies in Uj. This article introduces actions by partial equivalences in order to make sense of the actions of G on H and C∗(H). At first, we replace G by its inverse semigroup of bisections S = Bis(G). This inverse semigroup cannot act on H by partial groupoid isomorphisms for the same reasons as above. It does, however, act on H by partial equivalences because the equivalence class of H is independent of the covering (see also [16, Lemma 4.1]); thus partial homeomorphisms on G1 lift to partial equivalences of H in a canonical way. We will see that an S-action by partial equivalences on a Čech groupoid for a locally Hausdorff space Z is equivalent to an S-action on Z by partial homeomorphisms. Let S act on a groupoid H by partial equivalences. Then we build a transforma- tion groupoid H ⋊ S. Special cases of this construction are the groupoid of germs for an action of S on a space by partial homeomorphisms, the semidirect product for a group(oid) action on another group(oid) by automorphisms, and the linking groupoid of a single Morita -- Rieffel equivalence. The original action is encoded in the transformation groupoid L := H ⋊ S and open subsets Lt ⊆ L with Lt · Lu = Ltu, L−1 t = Lt∗ , Lt ∩ Lu = [v≤t,u Lv, L1 = [t∈S Lt and H = L1. We call such a family of subsets an S-grading on L with unit fibre H. Any S-graded groupoid is a transformation groupoid for an essentially unique action of S by partial equivalences on its unit fibre. This is a very convenient characteri- sation of actions by partial equivalences. An action of an inverse semigroup S on H by partial equivalences cannot induce, in general, an action of S on C∗(H) by partial automorphisms in the usual sense (as defined by Sieben [33]). But we do get an action by partial Morita -- Rieffel equivalences, that is, by Hilbert bimodules. We show that actions of S by Hilbert bimodules are equivalent to (saturated) Fell bundles over S. Along the way, we also drastically simplify the definition of Fell bundles over inverse semigroups in [13]. Our approach clarifies in what sense a Fell bundle over an inverse semigroup is an "action" of the inverse semigroup on a C∗-algebra. In the end, we want an action of the groupoid G itself, not of the inverse semi- group Bis(G). For actions by automorphisms, Sieben and Quigg [28] characterise which actions of Bis(G) come from actions of G. We extend this characterisation to INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 3 Fell bundles: a Fell bundle over Bis(G) comes from a Fell bundle over G if and only if the restriction of the action to idempotents in Bis(G) commutes with suprema of arbitrarily large subsets. This criterion only works for Bis(G) itself. In practice, we may want to "model" G by a smaller inverse semigroup S such that G0 ⋊ S ∼= G. We characterise which Fell bundles over such S come from Fell bundles over G. In particular, our action of Bis(G) on C∗(H) for the Čech groupoid associated to G1 does come from an action of G, so we get Fell bundles over G that describe the left and the right translation actions on G1. For these Fell bundles, we show that the section C∗-algebras are Morita equivalent to C0(G0). More generally, for any principal G-bundle X → Z, the section algebra of the Fell bundle over G that describes the action of G on a Čech groupoid for X is Morita -- Rieffel equivalent to C0(Z), just as in the more classical Hausdorff case (see Proposition 6.5). For any action of an inverse semigroup S on a locally compact groupoid H by partial equivalences, we identify the section C∗-algebra of the resulting Fell bundle over S with the groupoid C∗-algebra of the transformation groupoid. In brief notation, This generalises the well-known isomorphism C∗(H) ⋊ S ∼= C∗(H ⋊ S). C0(X) ⋊ S ∼= C∗(X ⋊ S) for inverse semigroup actions on Hausdorff locally compact spaces by partial home- omorphisms. For a Hausdorff locally compact groupoid, any Fell bundle is equivalent to an ordinary action on a stabilisation (Packer -- Raeburn Stabilisation Trick, see also [6, Proposition 5.2]). In contrast, our Theorem 7.1 shows that a non-Hausdorff groupoid has no action by automorphisms that describes its translation action on G1. Thus we really need Fell bundles to treat these actions of a non-Hausdorff groupoid. Now we explain the results of the individual sections of the paper. In Section 2, we study partial equivalences between topological groupoids. We show, in particular, that the involution that exchanges the left and right actions on a partial equivalence behaves like the involution in an inverse semigroup. Section 3 introduces inverse semigroup actions by partial equivalences. We show that the rather simple-minded definition implies further structure, which is needed to construct the transformation groupoid. Once we know that actions by partial equivalences are essentially the same as S-graded groupoids, we treat many exam- ples. This includes actions on spaces and Čech groupoids; in particular, an S-action on a space by partial homeomorphisms induces an action by partial equivalences on any Čech groupoid for a covering of the space. We describe a group action by (par- tial) equivalences as a kind of extension by the group. We show that any (locally) proper Lie groupoid is a transformation groupoid for an inverse semigroup action on a very simple kind of groupoid: a disjoint union of transformation groupoids of the form V ⋊ K, where V is a vector space, K a compact Lie group, and the K-action on V is by an R-linear representation. This is meant as an example for gluing together groupoids along partial equivalences. In Section 4 we define inverse semigroup actions on C∗-algebras by Hilbert bimod- ules. The theory is parallel to that for actions on groupoids by partial equivalences because both cases have the same crucial algebraic features. We show that actions by Hilbert bimodules are equivalent to saturated Fell bundles. This simplifies the original definition of Fell bundles over inverse semigroups in [13]. In Section 5, we turn inverse semigroup actions on groupoids by partial equiva- lences into actions on groupoid C∗-algebras by Hilbert bimodules. We do this in 4 ALCIDES BUSS AND RALF MEYER two different (but equivalent) ways, by using transformation groupoids and abstract functorial properties of our constructions. The approach using transformation groupoids suggests that the section C∗-algebra of the resulting Fell bundle is simply the groupoid C∗-algebra of the transformation groupoid: C∗(H) ⋊ S ∼= C∗(H ⋊ S). We prove this, and a more general result for Fell bundles over H ⋊ S. In Section 6 we relate inverse semigroup actions to actions of corresponding étale groupoids. In particular, we characterise when an action of Bis(G) comes from an action of G. Finally, we can then treat our motivating example and turn a groupoid action on a locally Hausdorff space Z into a Fell bundle over the groupoid. We may also describe the section C∗-algebra in this case, which plays the role of the crossed product. If the action is free and proper, then the result is Morita -- Rieffel equivalent to C0(Z/G). We also define "proper actions" of inverse semigroups on groupoids. We show that a free and proper action can only occur on a groupoid that is equivalent to a locally Hausdorff and locally quasi-compact space. Section 7 shows that the translation action of a non-Hausdorff étale groupoid on its arrow space cannot be described by a groupoid action by automorphisms in the usual sense. Our previous theory shows, however, that we may describe such actions by groupoid Fell bundles. Thus the no-go theorem in Section 7 shows that the Packer -- Raeburn Stabilisation Trick fails for non-Hausdorff groupoids, so Fell bundles are really more general than ordinary actions in that case. In Section 8, we examine a very simple explicit example to illustrate the no-go theorem and to see how our main results avoid it. Appendix A deals with topological groupoids, their actions on spaces and equiv- alences between them. The main point is to define principal bundles and (Morita) equivalence for non-Hausdorff groupoids in such a way that the theory works just as well as in the Hausdorff case. Among others, we show that a non-Hausdorff space is equivalent to a Hausdorff, locally compact groupoid if and only if it is locally Hausdorff and locally quasi-compact, answering a question in [8]. Appendix B contains a general technical result about upper semicontinuous fields of Banach spaces over locally Hausdorff spaces and uses it to prove C∗(H) ⋊ S ∼= C∗(H ⋊S) for inverse semigroup actions on groupoids and a more general statement involving Fell bundles over H ⋊ S. 2. Partial equivalences In this section and the next one, we work in the category of topological spaces and continuous maps, without assuming spaces to be Hausdorff or locally com- pact. Appendix A shows how topological groupoids, their actions, principal bun- dles, and equivalences between them should be defined so that the theory goes through smoothly without extra assumptions on the underlying topological spaces. Our main applications deal with groupoids that have a Hausdorff, locally com- pact object space and a locally Hausdorff, locally quasi-compact arrow space. We care about actions of such groupoids G on locally Hausdorff spaces Z. It is very convenient to encode such an action by the transformation groupoid G ⋉ Z. Its object space Z is only locally Hausdorff. When we allow such topological groupoids, the usual definition of equivalence for topological groupoids breaks down because orbit spaces of proper actions are always Hausdorff, so the actions on an equivalence bispace cannot be proper unless the object spaces of the two groupoids are Haus- dorff. Jean-Louis Tu's definition in [36] works -- it is equivalent to what we do. But the theory becomes more elegant if we also drop the local compactness assumption and thus no longer use proper maps in our basic definitions. The replacement for free and proper actions are "basic" actions, which are characterised by the map G ×s,G0,r X → X × X, (g, x) 7→ (gx, x), INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 5 being a homeomorphism onto its image with the subspace topology from X × X. Readers already familiar with the usual theory of locally compact groupoids may read on and only turn to Appendix A in cases of doubt; they should note that range and source maps of groupoids are assumed to be open, whereas anchor maps of groupoid actions are not assumed open. Less experienced readers should read Appendix A first. Definition 2.1. Let G and H be topological groupoids. A partial equivalence from H to G is a topological space with anchor maps r : X → G0 and s: X → H 0 and multiplication maps G1 ×s,G0,r X → X and X ×s,H0,r H 1 → X, which we write multiplicatively, that satisfy the following conditions: (P1) s(g · x) = s(x), r(g · x) = r(g) for all g ∈ G1, x ∈ X with s(g) = r(x), and s(x · h) = s(h), r(x · h) = r(x) for all x ∈ X, h ∈ H 1 with s(x) = r(h); (P2) associativity: g1 · (g2 · x) = (g1 · g2) · x, g2 · (x · h1) = (g2 · x) · h1, x · (h1 · h2) = (x · h1) · h2 for all g1, g2 ∈ G1, x ∈ X, h1, h2 ∈ H 1 with s(g1) = r(g2), s(g2) = r(x), s(x) = r(h1), s(h1) = r(h2); (P3) the following two maps are homeomorphisms: G1 ×s,G0,r X → X ×s,H0,s X, X ×s,H0,r H 1 → X ×r,G0,r X, (g, x) 7→ (x, g · x), (x, h) 7→ (x, x · h); (P4) s and r are open. The first two conditions say that X is a G, H-bispace. The only difference be- tween a partial and a global equivalence is whether the anchor maps are assumed surjective or not: conditions (P1) -- (P4) are the same as conditions (E1) -- (E4) in Proposition A.5. We view a partial equivalence X from H to G as a generalised map from H to G. Indeed, there is a bicategory with partial equivalences as arrows H → G (Theorem 2.15). Definition 2.2. Let G be a groupoid. A subset U ⊆ G0 is G-invariant if r−1(U ) = s−1(U ). In this case, U and r−1(U ) = s−1(U ) are the object and arrow spaces of a subgroupoid of G, which we denote by GU . The canonical projection p: G0 → G0/G induces a bijection between G-invariant subsets U ⊆ G0 and subsets p(U ) ⊆ G0/G. We are mainly interested in open invariant subsets. Since p is open and continuous, open G-invariant subsets of G0 correspond to open subsets of G0/G. Lemma 2.3. Let G and H be topological groupoids. A partial equivalence X from H to G is the same as an equivalence from HV to GU for open, invariant subsets U ⊆ G0, V ⊆ H 0. Here U = r(X), V = s(X). Proof. Let U ⊆ G0 be G-invariant. A left GU -action is the same as a left G-action U = U and G1 ×s,G0,r X ∼= for which the anchor map takes values in U because G0 G1 U ×s,G0,r X if r(X) ⊆ U . Thus the commuting actions of GU and HV for an equivalence from HV to GU may also be viewed as commuting actions of G and H, respectively. This gives a partial equivalence (see Definition 2.1). Conversely, let X be a partial equivalence. Let U := r(X) ⊆ G0 and V := s(X) ⊆ H 0. These are open subsets because r and s are open, and they are invariant by (P1). The actions of G and H are equivalent to actions of GU and HV , respectively. After replacing G and H by GU and HV , respectively, all conditions (E1) -- (E5) in Proposition A.5 hold; thus X is an equivalence from HV to GU . (cid:3) 6 ALCIDES BUSS AND RALF MEYER Lemma 2.4. Let X be a partial equivalence from H to G and let U ⊆ G0 and V ⊆ H 0 be invariant open subsets. Then U XV := {x ∈ X r(x) ∈ U, s(x) ∈ V } is again a partial equivalence from H to G. We also write U X and XV for U XH0 and G0XV , respectively. Proof. The subset U XV is open because r and s are continuous and U and V are open, and it is invariant under the actions of G and H because U and V are invariant and the two anchor maps are either invariant or equivariant with respect to the two actions. Hence we may restrict the actions of G and H to U XV . Conditions (P1) -- (P2) and (P4) in Definition 2.1 are inherited by an open invariant subspace. The inverse to the first homeomorphism in (P3) maps U XV ×s,H0,s U XV into G1 ×s,G0,r U XV , and the inverse to the second one maps U XV ×r,G0,r U XV into U XV ×s,H0,r H 1. Thus U XV also inherits (P3) and is a partial equivalence from H to G. (cid:3) Equivalences are partial equivalences, of course. In particular, the identity equiv- alence G1 with G acting by left and right multiplication is also a partial equivalence. Let X and Y be partial equivalences from H to G and from K to H, respectively. Their composite is defined as for global equivalences, and still denoted by ×H: X ×H Y := X ×s,H0,r Y / (x · h, y) ∼ (x, h · y), equipped with the quotient topology and the induced actions of G and K by left and right multiplication. The canonical map X ×s,H0,r Y → X ×H Y is a principal H-bundle for the H-action defined by (x, y) · h := (x · h, h−1 · y); this follows from the general theory in [21]. Example 2.5. We associate an equivalence Hf from G to H to a groupoid isomor- phism f : G → H. The functor f consists of homeomorphisms f i : Gi → H i for i = 0, 1. We take X = H 1 with the usual left H-action and the right G-action by h · g := h · f 1(g) for all h ∈ H 1, g ∈ G1 with s(h) = r(f 1(g)) = f 0(r(g)); so the right anchor map is (f 0)−1 ◦ s = s ◦ (f 1)−1. We claim that an equivalence is of this form if and only if it is isomorphic to H 1 as a left H-space. Since H\H 1 ∼= H 0, the right anchor map gives a homeomorphism H 0 → G0 in this case; let f 0 : G0 → H 0 be its inverse. The right action of g ∈ G1 on h ∈ H 1 with s(h) = f 0(r(g)) must be of the form h · g = h · f 1(g) for a unique f 1(g) ∈ H 1 with r(f 1(g)) = f 0(r(g)) and s(f 1(g)) = s(h · g) = f 0(s(g)). It is routine to check that f 0 and f 1 give a topological groupoid isomorphism. ∼−→ H 1 When do two isomorphisms f, ϕ: G → H give isomorphic equivalences? Let ϕ be an isomorphism. Define a continuous map σ : G0 → H 1 by u : H 1 f σ(x) := u(1f 0(x)) for all x ∈ G0. This satisfies r(σ(x)) = f 0(x) and s(σ(x)) = ϕ0(x) for all x ∈ G0 because u is compatible with anchor maps. Since u is left H-invariant, u(h) = u(h · 1s(h)) = h · (σ ◦ (f 0)−1 ◦ s)(h) for all h ∈ H 1, so σ determines u. The right G-invariance of u translates to σ(r(g)) · ϕ1(g) = f 1(g) · σ(s(g)) for all g ∈ G. Thus (1) ϕ0(x) = s(σ(x)), ϕ1(g) = σ(r(g))−1 · f 1(g) · σ(s(g)). Roughly speaking, f and ϕ differ by an inner automorphism. Let an equivalence f : G → H and a continuous map σ : G0 → H 1 with r(σ(x)) = f 0(x) for all x ∈ H 0 be given. Assume that H 0 → G0, x 7→ s(σ(x)), is a homeomorphism. Then (1) defines an isomorphism ϕ: G → H such that h 7→ h · σ((f 0)−1(s(h))) is an isomorphism between the equivalences Hf and Hϕ. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 7 Example 2.6. If G and H are minimal groupoids in the sense that G0 and H 0 have no proper open invariant subsets, then any partial equivalence is either empty or a full equivalence G ∼−→ H. This holds, in particular, if G and H are groups. Example 2.7. Any non-empty (partial) equivalence between two groups is isomor- phic to one coming from a group isomorphism G ∼= H. Indeed, since X/H ∼= G0 and G\X ∼= H 0 are a single point, both actions on X are free and transitive. Fix x0 ∈ X. Since the actions are free and transitive and part of principal bundles, the maps G → X, g 7→ g · x0, and H → X, h 7→ x0 · h, are homeomorphisms. The composite map G ∼−→ X ∼−→ H is an isomorphism of topological groups. This isomorphism depends on the choice of x0. The isomorphisms G ∼−→ H for different choices of x0 differ by an inner automorphism. Lemma 2.8. The composition ×H is associative and unital with the identity equiv- alence as unit, up to the usual canonical bibundle isomorphisms (X ×H Y ) ×K Z ∼= X ×H (Y ×K Z), G1 ×G X ∼= X ∼= X ×H H 1. Proof. For global equivalences with arbitrary topological spaces, this is contained in [21, Proposition 7.10]. The proofs in [21] can be extended to the partial case as well. Alternatively, we may reduce the partial to the global case by restricting our partial equivalences to global equivalences between open subgroupoids as in Lemma 2.4. This works because U (X ×H Y )V ∼= (U X) ×H (Y V ) for U ⊆ G0, V ⊆ K 0 open and invariant and partial equivalences X from H to G and Y from K to H. Details are left to the reader. (cid:3) Proposition 2.9. Let G and H be topological groupoids. Let X1 and X2 be partial equivalences from H to G. There is no bibundle map X1 → X2 unless r(X1) ⊆ r(X2) and s(X1) ⊆ s(X2). Any G, H-bibundle map ϕ: X1 → X2 is an isomorphism onto the open sub-bibundle r(X1)X2 = X2s(X1). The map ϕ is invertible if r(X2) ⊆ r(X1) or s(X2) ⊆ s(X1). In this case, r(X2) = r(X1) and s(X2) = s(X1). Proof. Since rX2 ◦ ϕ = rX1 and sX2 ◦ ϕ = sX1, we must have r(X1) ⊆ r(X2) and s(X1) ⊆ s(X2) if there is a bibundle map ϕ: X1 → X2. Assume this from now on. The image of a bibundle map is contained in r(X1)X2 and in X2s(X1). Since r(X1) ⊆ r(X2) and s(X1) ⊆ s(X2), we have r(r(X1)X2) = r(X1) and s(X2s(X1)) = s(X1). All remaining assertions now follow once we prove that a bibundle map ϕ: X1 → X2 is invertible if r(X1) = r(X2) or s(X1) = s(X2). We treat the case r(X2) = r(X1); the other one is proved in the same way, exchanging left and right. Since Xi is a partial equivalence, it is a principal H-bundle over Xi/H ∼= r(Xi). The map ϕ induces a homeomorphism on the base spaces because r(X2) = r(X1) both carry the subspace topology from G0. Hence ϕ is a homeomorphism by [21, Proposition 5.9]. (cid:3) In particular, the restricted multiplication maps G1 U ×H X ⊆ G1 ×G X → X and V ⊆ X ×H H 1 → X are bibundle maps. Proposition 2.9 shows that they X ×H H 1 induce bibundle isomorphisms (2) G1 U ×G X ∼= U X, X ×H H 1 V ∼= XV . Partial equivalences carry extra structure similar to an inverse semigroup. The adjoint operation is the following: Definition 2.10. Given a partial equivalence X from H to G, we define the dual partial equivalence X ∗ by exchanging the left and right actions on X. More precisely, X ∗ is X as a space, the anchor maps r∗ : X ∗ → H 0 and s∗ : X ∗ → G0 are r∗ = sX 8 ALCIDES BUSS AND RALF MEYER and s∗ = rX , and the left H- and right G-actions are defined by h ·∗ x = x · h−1 and x ·∗ g := g−1 · x, respectively. If X gives an equivalence from HV to GU for open invariant subsets U ⊆ G0, V ⊆ H 0, then X ∗ gives the "inverse" equivalence from GU to HV . The following properties of duals are trivial: • naturality: a bibundle map X → Y induces a bibundle map X ∗ → Y ∗; • (X ∗)∗ = X; • there is a natural isomorphism σ : (X ×H Y )∗ ∼= Y ∗ ×H X ∗, (x, y) 7→ (y, x), with σ2 = Id. Let Map(Y1, Y2) be the space of bibundle maps between two partial equivalences Y1, Y2 from H to G. Proposition 2.11. Let X be a partial equivalence from H to G. Then there are natural isomorphisms X ∗ ×G X ∼= H 1 that make the following diagrams of isomorphisms commute: X ×H X ∗ ∼= G1 r(X), s(X) (3) X ×H X ∗ ×G X X ×H H 1 s(X) G1 r(X) ×G X X, X ∗ ×G X ×H X ∗ X ∗ ×G G1 r(X) H 1 s(X) ×G X ∗ X ∗. If K is another groupoid and Y and Z are partial equivalences from K to G and from K to H, respectively, with r(Y ) ⊆ r(X) and r(Z) ⊆ s(X), then there are natural isomorphisms Map(X ×H Z, Y ) ∼= Map(Z, X ∗ ×G Y ), Map(Y, X ×H Z) ∼= Map(X ∗ ×G Y, Z). Both map the subsets of bibundle isomorphisms onto each other. r(X) and X ∗ ×G X ∼= H 1 Proof. Lemma 2.3 shows that X is an equivalence from Hs(X) to Gr(X). Hence the usual theory of groupoid equivalence gives canonical isomorphisms X ×H X ∗ ∼= G1 s(X). The first one maps the class of (x1, x2) with s(x1) = s(x2) to the unique g ∈ G1 with x1 = g · x2. In particular, it maps [x, x] 7→ 1r(x). The second one maps the class of (x1, x2) with r(x1) = r(x2) to the unique h ∈ H 1 with x2 = x1 · h. 7→ 1s(x). Then the composite isomorphisms X ×H X ∗ ×G X → X and X ∗ ×G X ×H X ∗ → X ∗ map [x, x, x] 7→ x, respectively. Since any element in X ×H X ∗ ×G X or X ∗ ×G X ×H X ∗ has a representative of the form (x, x, x), we get the two commuting diagrams in (3). In particular, it maps [x, x] The assumption r(Y ) ⊆ r(X) implies s(X ∗ ×G Y ) = s(Y ) because for any y ∈ Y there is x ∈ X ∗ with (x, y) ∈ X ∗ ×G Y . Similarly, r(Z) ⊆ s(X) implies s(X ×H Z) = s(Z). By Proposition 2.9, a bibundle map X ×H Z → Y exists only if s(Z) ⊆ s(Y ), and then it is an isomorphism onto Y s(Z); and a bibundle map Z → X ∗ ×G Y exists only if s(Z) ⊆ s(Y ), and then it is an isomorphism onto X ∗ ×G Y s(Z). Thus we may as well replace Y by Y s(Z) to achieve s(Y ) = s(Z); then all bibundle maps X ×H Z → Y or Z → X ∗ ×G Y are bibundle isomorphisms. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 9 The second isomorphism reduces in a similar way to the case where also s(Y ) = s(Z) and where we are dealing only with bibundle isomorphisms. A bibundle map ϕ: X ×H Z → Y induces IdX ∗ ×G ϕ: X ∗ ×G X ×H Z → X ∗ ×G Z; we compose this with the natural isomorphism X ∗ ×G X ×H Z ∼= H 1 s(X) ×H Z ∼= s(X)Z = Z to get a bibundle map Z → X ∗ ×G Y ; here we used s(X) ⊇ r(Z). We claim that this construction gives the desired bijection between Map(X ×H Z, Y ) and Map(Z, X ∗ ×G Y ). Since composing with an isomorphism is certainly a bijection, it remains to show that Map(X ×H Z, Y ) → Map(X ∗ ×G X ×H Z, X ∗ ×G Y ), ϕ 7→ IdX ∗ ×G ϕ, is bijective. Since X ×H X ∗ ∼= G1 r(X) and r(X) ⊇ r(Y ), we have natural isomor- phisms X ×H X ∗ ×G Y ∼= Y and X ×H X ∗ ×G X ×H Z ∼= X ×H Z. Naturality means that they intertwine ϕ 7→ IdX×H X ∗ ×G ϕ and ϕ. Since IdX×H X ∗ ×G ϕ = IdX ×H IdX ∗ ×G ϕ, we see that ϕ 7→ IdX ∗ ×G ϕ is injective and has ψ 7→ IdX ×H ψ for ψ : Z → X ∗ ×G Y as a one-sided inverse. The same argument also shows that ψ 7→ IdX ×H ψ is injective, so both constructions are bijective. (cid:3) Applying duality, we also get bijections Map(Z ∗×H X ∗, Y ∗) ∼= Map(Z ∗, Y ∗×GX) and Map(Y ∗, Z ∗ ×H X ∗) ∼= Map(Y ∗ ×G X, Z ∗) under the same hypotheses. The canonical isomorphisms (4) X ×H X ∗ ×G X ∼= X, X ∗ ×G X ×H X ∗ ∼= X ∗ from Proposition 2.11 characterise X ∗ uniquely in the following sense: Proposition 2.12. Let X and Y be partial equivalences from H to G and from G to H, respectively. If there are bibundle isomorphisms X ×H Y ×G X ∼= X, Y ×H X ×G Y ∼= Y, then there is a unique bibundle isomorphism X ∗ ∼= Y such that the composite map (5) X ∼= X ×H X ∗ ×G X ∼= X ×H Y ×G X ∼= X is the identity map. Proof. When we multiply the inverse of the isomorphism X ×H Y ×G X ∼= X on both sides with X ∗ and use (2), we get an isomorphism X ∗ ∼= X ∗ ×G X ×H X ∗ ∼= X ∗ ×G X ×H Y ×G X ×H X ∗ ∼= H 1 s(X) ×H Y ×G G1 r(X) ∼= s(X)Y r(X). This implies s(X) = r(X ∗) ⊆ r(Y ) and r(X) = s(X ∗) ⊆ s(Y ) by Proposition 2.9. Exchanging X and Y , the isomorphism Y ×H X ×G Y ∼= Y gives s(Y ) ⊆ r(X) and r(Y ) ⊆ s(X). Hence r(Y ) = s(X) and s(Y ) = r(X), so s(X)Y r(X) = Y . This gives an isomorphism α : X ∗ ∼−→ Y . A diagram chase using the commuting diagrams in (3) shows that the composite of the map X ×H X ∗ ×G X → X ×H Y ×G X induced by the isomorphism α and the given isomorphism X ×H Y ×G X → Y (which we used to construct α) is the canonical map X ×H X ∗ ×G X → X as in (4). Hence the composite in (5) is the identity map for the isomorphism α. The isomorphisms in Proposition 2.11 give a canonical bijection Map(X ∗, Y ) ∼= Map(X ∗ ×G X ×H X ∗, Y ) ∼= Map(X ×H X ∗, X ×H Y ) ∼= Map(X, X ×H Y ×G X) ∼= Map(X, X). 10 ALCIDES BUSS AND RALF MEYER Inspection shows that it maps an isomorphism X ∗ ∼−→ Y to the composite map in (5). Hence there is only one isomorphism X ∗ ∼−→ Y for which the composite map in (5) is the identity map. (cid:3) Proposition 2.13. Let X be a partial equivalence from G to itself and let µ: X ×G X → X be a bibundle isomorphism. Then there is a unique isomorphism ϕ: X ∼−→ U for an open G-invariant subset U ⊆ G0 such that the following diagram com- G1 mutes: (6) X ×G X ϕ ×G ϕ G1 U ×G G1 U µ µ0 X ϕ G1 U , µ0(g1, g2) = g1 · g2. Hence r(X) = s(X) and µ is associative. Proof. The isomorphism µ induces an isomorphism X ×G X ×G X µ×GIdX −−−−−→ X ×G X µ −→ X Hence Y = X satisfies the two conditions in Proposition 2.11 that ensure X = Y ∼= X ∗. This gives an isomorphism ϕ: X ∼= X ×G X ∼= X ×G X ∗ ∼= G1 r(X). Since ϕ is a bibundle map, the diagram (6) commutes if and only if µ is the composite map X ×G X ϕ×GIdX −−−−−→ G1 r(X) ×G X ∼= X, where the map G1 Sending an isomorphism ϕ: X → G1 of the bijections in Proposition 2.11, namely, the first one for X = Y = Z: r(X) ×G X ∼= X is the left multiplication map, [g, x] 7→ g · x. ∼= X ×G X ∗ to this composite map is one r(X) Map(X, G1 s(X)) ∼= Map(X, X ∗ ×G X) ∼= Map(X ×G X, X). Hence there is exactly one isomorphism ϕ that corresponds under this bijection to µ. (cid:3) Proposition 2.11 implies that isomorphism classes of partial equivalences from G to itself form an inverse semigroup fpeq(G). The idempotents in this inverse semi- group are in bijection with G-invariant open subsets of G0 by Proposition 2.13. These are, in turn, in bijection with open subsets of the orbit space G0/G by the definition of the quotient topology on G0/G. These also correspond to the idempo- tents of the inverse semigroup pHomeo(G0/G) of partial homeomorphisms of the topological space G0/G. A partial equivalence X from H to G induces a partial homeomorphism X∗ : H 0/H ⊆ s(X) → r(X) ⊆ G0/G by X∗([h]) = [g] if there is x ∈ X with s(x) ∈ [h], r(x) ∈ [g]. If Y is another partial equivalence from K to H, then (X ×H Y )∗ = X∗ ◦ Y∗ by definition. This gives a canonical homomorphism of inverse semigroups fpeq(G) → pHomeo(G0/G). Remark 2.14. The homomorphism fpeq(G) → pHomeo(G0/G) is neither injective nor surjective in general, although it is always an isomorphism on the semilattice of idempotents. Consider, for instance, the disjoint union G = Z/3 ⊔ {pt}. This groupoid is a group bundle, and G0/G has two points. The partial homeomorphism that maps one point to the other does not lift to a partial equivalence because the stabilisers are not the same and equivalences must preserve the stabiliser groups. The group Z/3 has non-inner automorphisms, so there are non-isomorphic partial INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 11 equivalences of G defined on Z/3 that induce the same partial homeomorphism on G0/G. In our definition of an inverse semigroup action (see Sections 3 and 4 below), certain isomorphisms of partial equivalences are a crucial part of the data. We could not construct transformation groupoids and Fell bundles without them. If we identify isomorphic partial equivalences as above, then we can no longer talk about two isomorphisms of partial equivalences being equal. The correct way to take into account isomorphisms of partial equivalences is through a bicategory (see [1, 6, 19]). The following remarks are intended for readers familiar with bicategories. Our bicategory has topological groupoids as objects and partial equivalences as arrows. Let G and H be topological groupoids and let X1 and X2 be partial equivalences from H to G. As 2-arrows X1 ⇒ X2, we take all G, H-bibundle isomorphisms X1 → X2, so all 2-arrows are invertible. The vertical product of 2-arrows is the composition of bibundle maps. Unit 2-arrows are identity maps on partial equivalences. The composition of arrows is ×H. The unit arrow on a topological groupoid G is G1 with the standard bibundle structure. Lemma 2.8 provides invertible 2-arrows (X ×H Y ) ×K Z ⇒ X ×H (Y ×K Z), G1 ×G X ⇒ X ⇐ X ×H H 1, which we take as associator and left and right unit transformations. Let X1, X2 be partial equivalences from H to G and let Y1, Y2 be partial equivalences from K to H. The horizontal product of two bibundle maps f : X1 → X2 and g : Y1 → Y2 is f ×H g : X1 ×H Y1 → X2 ×H Y2. Theorem 2.15. The data above defines a bicategory peq. Proof. It is routine to check that partial equivalences from H to G with bibundle maps between them form a category C(G, H) for the vertical product of bibundle maps, and that the composition of partial equivalences with the horizontal prod- uct of bibundle maps is a functor C(G, H) × C(H, K) → C(G, K). The associator and both unit transformations are natural isomorphisms of functors; the associa- tor is clearly compatible with unit transformations and makes the usual pentagon commute, see [19, p. 2]. (cid:3) Remark 2.16. We still get a bicategory if we allow all bibundle maps as 2-arrows. We restrict to invertible 2-arrows to get the correct notion of inverse semigroup actions below. An arrow f : x → y in a bicategory is called an equivalence if there are an arrow g : y → x and invertible 2-arrows g ◦ f ⇒ Idx and f ◦ g ⇒ Idy. The equivalences in peq are precisely the global bibundle equivalences. The duality X 7→ X ∗ with the canonical flip maps (X ×H Y )∗ ∼−→ Y ∗ ×H X ∗ gives a functor I : peq → peqop with I 2 = Idpeq. It seems useful to formalise the properties of this functor and look for examples in more general bicategories. But we shall not go into this question here. 3. Inverse semigroup actions on groupoids We give two equivalent definitions for actions of inverse semigroups on topological groupoids by partial equivalences. The first is exactly what it promises to be. The second, more elementary, definition does not mention groupoids or partial equivalences. Let S be an inverse semigroup with unit 1. Let G be a topological groupoid. Definition 3.1. An action of S on G by partial equivalences consists of • partial equivalences Xt from G to G for t ∈ S; 12 ALCIDES BUSS AND RALF MEYER • bibundle isomorphisms µt,u : Xt ×G Xu ∼−→ Xtu for t, u ∈ S; satisfying (A1) X1 is the identity equivalence G1 on G; (A2) µt,1 : Xt ×G G1 ∼−→ Xt and µ1,u : G1 ×G Xu ∼−→ Xu are the canonical iso- morphisms, that is, the left and right G-actions, for all t, u ∈ S; (A3) associativity: for all t, u, v ∈ S, the following diagram commutes: (Xt ×G Xu) ×G Xv ass Xt ×G (Xu ×G Xv) µt,u ×G IdXv IdXt ×G µu,v Xtu ×G Xv Xt ×G Xuv µtu,v Xtuv µt,uv If S has a zero object 0, then we may also ask X0 = ∅. Remark 3.2. Let S be an inverse semigroup possibly without 1. We may add a unit 1 formally and extend the multiplication by 1 · s = s = s · 1 for all s ∈ S ∪ {1}. If partial equivalences (Xt)t∈S and bibundle isomorphisms (µt,u)t,u∈S are given satisfying associativity for all t, u, v ∈ S, then we may extend this uniquely to an action of S ∪ {1}: we put X1 := G1 and let µt,1 and µ1,u be the right and left G-action, respectively. The associativity condition is trivial if one of t, u, v is 1, so associativity holds for all t, u, v ∈ S ∪{1}. As a result, an action of S ∪{1} by partial equivalences is the same as (Xt)t∈S and (µt,u)t,u∈S satisfying only Condition (A3). Similarly, we may add a zero 0 to S and extend the multiplication by 0 · s = 0 = s · 0 for all s ∈ S ∪ {0}. We extend an S-action by X0 := ∅, so that X0 ×G Xt = ∅ = Xt ×G X0, leaving no choice for the maps µt,0, µ0,u : ∅ → ∅. This gives an action of S ∪ {0} with X0 = ∅. If 0, 1 ∈ S and we ask no conditions on X0 and X1, then r(Xt), s(Xt) ⊆ r(X1) = s(X1) for all t ∈ S, and Xt restricted to r(X0) = s(X0) is the trivial action where all Xt act by the identity equivalence. Hence all the action is on the locally closed, invariant subset r(X1) \ r(X0) ⊆ G0. The conditions on X0 and X1 merely rule out such degeneracies. Remark 3.3. An inverse semigroup may be viewed as a special kind of category with only one object, which is also a very special kind of bicategory. An inverse semigroup action by partial equivalences is exactly the same as a functor from this category to the bicategory peq of partial equivalences (see [19]). Lemma 3.4. For an inverse semigroup action (Xt, µt,u), we have r(Xt) = r(Xtt∗ ) = s(Xtt∗) = s(Xt∗) and s(Xt) = s(Xt∗t) = r(Xt∗t) = r(Xt∗ ) for each t ∈ S. Proof. If e ∈ S idempotent, then Proposition 2.13 applied to the isomorphism µe,e : Xe ×G Xe ∼= Xe gives r(Xe) = s(Xe). The existence of an isomorphism µt,t∗ : Xt ×G Xt∗ ∼= Xtt∗ implies r(Xt) ⊇ r(Xtt∗ ) and s(Xt∗) ⊇ s(Xtt∗). Similarly, the isomorphism µtt∗,t gives r(Xtt∗ ) ⊇ r(Xt), and µt,t∗t gives s(Xt∗t) ⊇ s(Xt). Now everything follows. (cid:3) Definition 3.5. Let S be an inverse semigroup with unit. A simplified action of S on a topological groupoid consists of • a topological space G0; • topological spaces Xt for t ∈ S; • continuous maps s, r : Xt → G0; • continuous maps µt,u : Xt ×s,G0,r Xu → Xtu, (x, y) 7→ x · y, for t, u ∈ S; INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 13 satisfying (S1) s(x · y) = s(y), r(x · y) = r(x) for all t, u ∈ S, x ∈ Xt, y ∈ Xu with s(x) = r(y); (S2) r : Xt → G0 and s: Xt → G0 are open for all t ∈ S; (S3) the maps r, s: X1 → G0 are surjective; (S4) µt,u is surjective for each t, u ∈ S; (S5) the map Xt ×s,G0,r Xu → Xu ×s,G0,s Xtu, (x, y) 7→ (y, x · y), is a homeomorphism if t = 1 and u ∈ S; (S6) the map Xt ×s,G0,r Xu → Xt ×r,G0,r Xtu, (x, y) 7→ (x, x · y), is a homeomorphism if t ∈ S and u = 1; (S7) for all t, u, v ∈ S, the following diagram commutes: (Xt ×s,G0,r Xu) ×s,G0,r Xv µt,u ×s,G0,r IdXv Xtu ×s,G0,r Xv (7) ass µtu,v µt,uv Xtuv Xt ×s,G0,r (Xu ×s,G0,r Xv) IdXt ×s,G0,r µu,v Xt ×s,G0,r Xuv If S has a zero element, we may also ask X0 = ∅. This definition is more elementary because it does not mention groupoids or partial equivalences. It seems less elegant than Definition 3.1, but is simpler because much of the complexity of Definition 3.1 is hidden in the conditions (P1) -- (P4) defining partial equivalences of topological groupoids. It is clear that an inverse semigroup action by partial equivalences gives a sim- plified action: forget the multiplication on G1 and the left and right actions of G on the spaces Xt. The isomorphisms in (S5) for t = 1 and in (S6) for u = 1 are those in (P3), and all other conditions in Definition 3.5 are evident. The converse is more remarkable: Proposition 3.6. Any simplified inverse semigroup action on groupoids comes from a unique action by partial equivalences. Thus actions and simplified actions of inverse semigroups by partial equivalences are equivalent. Furthermore, the maps in (S5) and (S6) are isomorphisms and the maps µt,u are open for all t, u ∈ S. Proof. The spaces G0 and G1 := X1 with range and source maps r and s and mul- tiplication µ1,1 satisfy the conditions (G1) -- (G4) in Proposition A.1 because these are special cases of our conditions (S1) -- (S7). Hence this data defines a topological groupoid. Similarly, the anchor maps r : Xt → G0 and s: Xt → G0 and the multi- plication maps µ1,t and µt,1 satisfy conditions (P1) -- (P4) in Definition 2.1 and thus turn Xt into a partial equivalence from G to itself. Let t, u ∈ S. The associativity of the maps µ for t, 1, u, 1, t, u and t, u, 1 implies that µt,u descends to a G, G-bibundle map ¯µt,u : Xt ×G Xu → Xtu. Since µt,u is surjective by (S4), so is ¯µt,u. Hence it is a bibundle isomorphism by Proposition 2.9. The groupoid structure on X1 and the left and right actions on Xt are defined so that X1 is the identity equivalence on G and the maps ¯µ1,u and ¯µt,1 are the canon- ical isomorphisms. The associativity condition for the bibundle isomorphisms ¯µt,u follows from the corresponding property of the maps µt,u. Thus we have got an action by partial equivalences. This is the only action that simplifies to the given data because of the assumptions about X1, µ1,u, and µt,1 in Definition 3.1. 14 ALCIDES BUSS AND RALF MEYER By definition, Xt ×G Xu is the orbit space of the G-action on Xt ×s,G0,r Xu by (x1, x2) · g := (x1 · g, g−1 · x2). The canonical projection Xt ×s,G0,r Xu → Xt ×G Xu is open by Proposition A.3. The map µt,u is the composite of this projection with the homeomorphism ¯µt,u : Xt ×G Xu → Xtu, hence it is also open. Finally, we check that the maps in (S5) are isomorphisms for all t, u ∈ S; ex- changing left and right gives the same for the maps in (S6). The map in (S5) is G-equivariant if we let G act on Xt ×s,G0,r Xu by g · (x, y) := (xg−1, gy) and on Xu ×s,G0,s Xtu by g · (y, x) := (gy, x). Both actions are part of principal bundles: the bundle projection on Xt ×s,G0,r Xu is the canonical map to Xt ×G Xu, and the bundle projection on Xu×s,G0,sXtu is s×G0,sIdXtu to Xtur(Xu). Our G-equivariant map induces the map µt,u on the base spaces, which is a homeomorphism. Hence so is the map on the total spaces by [21, Proposition 5.9]. (cid:3) 3.1. Compatibility with order and involution. Let S be an inverse semigroup with unit. Define a partial order on S by t ≤ u if t = tt∗u or, equivalently, t = ut∗t. The multiplication and involution preserve this order: t1t2 ≤ u1u2 and t∗ 1 if t1 ≤ u1 and t2 ≤ u2 (see [18]). 1 ≤ u∗ Let (Xt)t∈S, (µt,u)t,u∈S be an action of S on G. We are going to prove that the action is compatible with this partial order and the involution on S. To prepare for the proofs of analogous statements for inverse semigroup actions on C∗-algebras, we give rather abstract proofs, which carry over literally to the C∗-algebraic case. Proposition 3.7. There are unique bibundle maps ju,t : Xt → Xu for t, u ∈ S with t ≤ u such that the following diagrams commute for all t1, t2, u1, u2 ∈ S with t1 ≤ u1, t2 ≤ u2: (8) Xt1 ×G Xt2 ju1,t1 ×G ju2,t2 Xu1 ×G Xu2 µt1,t2 µu1,u2 Xt1t2 ju1u2,t1t2 Xu1u2 The map ju,t is a bibundle isomorphism onto Xus(Xt) = r(Xt)Xu. We have jt,t = IdXt for all t ∈ S and jv,u ◦ ju,t = jv,t for t ≤ u ≤ v in S. Proof. Let E(S) ⊆ S be the subset of idempotents and let e ∈ E(S). Proposi- tion 2.13 gives a unique isomorphism Xe ∼= G1 Ue intertwining µe,e : Xe ×G Xe → Xe Ue; here Ue := r(Xe) = s(Xe) is an open G-invariant and the multiplication in G1 subset of G0. The diagram (8) for (e, e) ≤ (1, 1) shows that j1,e has to be this particular isomorphism Xe ∼= G1 Ue ⊆ G1. To simplify notation, we now identify Xe with G1 Ue for all e ∈ E(S) using these unique isomorphisms, and we transfer the multiplication maps µs,t for idempotent s, t or st accordingly. This gives an iso- morphic action of S by partial equivalences. So we may assume that Xe = G1 Ue and that µe,e : Xe ×G Xe → Xe is the usual multiplication map on G1 Ue for all e ∈ E(S). Let e ∈ E(S) and let t, u ∈ S satisfy t∗t ≤ e and uu∗ ≤ e. Thus te = t, eu = u ∼−→ and teu = tu. We show that µt,e : Xt ×G G1 Xu are the obvious maps µ0 e,u from the left and right G-actions in this case. Associativity of the multiplication maps gives us a commuting diagram of isomorphisms Ue → Xt and µe,u : G1 t,e or µ0 Ue ×G Xu Xt ×G Xe ×G Xu µt,e ×G IdXu Xt ×G Xu IdXt ×G µe,u µt,u Xt ×G Xu µt,u Xtu INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 15 We may cancel the isomorphism µt,u to get IdXt ×G µe,u = µt,e ×G IdXu. Now we consider two cases: t = e or e = u. If t = e, then µt,e = µ0 t,e is the multiplication map on G1 Ue. Hence so is µt,e ×G IdXu. Thus µe,u and µ0 e,u induce the same map G1 Ue ×G G1 Ue ×G Xu → G1 Ue because s(Xe) = Ue ⊇ r(Xu) ⊇ r(G1 e,u if e is idempotent and e ≥ uu∗. A similar argument in the other case e = u gives µt,e = µ0 Ue ×G Xu. We may use (2) to cancel the factor G1 Ue ×G Xu). Thus µe,u = µ0 t,e if t∗t ≤ e. Now let t ≤ u, that is, t = tt∗u = ut∗t. Then we get two candidates for the bibundle map ju,t : Xt → Xu: (9) Xt Xt µtt∗ ,u←−−−− ∼= µu,t∗ t←−−−− ∼= Xtt∗ ×G Xu = G1 Utt∗ ×G Xu Xu ×G Xt∗t = Xu ×G G1 Ut∗ t µ0 tt∗ ,u−−−−→ ∼= µ0 u,t∗ t−−−−→ ∼= Utt∗ Xu ⊆ Xu, XuUt∗ t ⊆ Xu. We claim that both maps Xt → Xu are equal, so we get only one map ju,t : Xt → Xu. Let e = tt∗ and f = t∗t. Then there is a commuting diagram of isomorphisms Xe ×G Xu ×G Xf IdXe ×G µu,f µ0 ×G IdXf IdXe ×G µ0 Ue Xu ×G Xf (10) µe,u ×G IdXf Xe ×G Xt µe,t = µ0 e,t Xt ×G Xf Xe ×G XuUf µu,f µe,u Xt µt,f = µ0 t,f The large rectangle commutes by associativity. The argument above gives µe,t = e,t and µt,f = µ0 µ0 t,f . The lower left and upper right triangles commute because µe,u and µu,f are bibundle maps, so they are compatible with µ0. Hence the interior quadrilateral commutes. Thus the two definitions of ju,t in (9) are equal. The first construction of ju,t in (9) gives the unique map for which the diagram (8) commutes for (e, t) ≤ (1, u) and the inclusion map j1,e. Since we already saw that j1,e is unique, the diagrams (8) characterise the bibundle maps ju,t uniquely for all t ≤ u in S. The map jt,t is the identity on Xt because µtt∗,t = µ0 Now let t ≤ u ≤ v, define e = tt∗ and f = uu∗ and identify Xe and Xf with subsets of G1. In the following diagram, we abbreviate ×G to ∗, and µ0 denotes the left and right actions for subsets of G1: tt∗,t. ju,t µe,u Xt Xe ∗ Xu µ0 Ue Xu µe,v Id ∗ µf,v Ue µf,v jv,t Xe ∗ Xv µe,f ∗ Id Xe ∗ Xf ∗ Xv µ0 ∗ Id Ue Xf ∗ Xv Ue jv,u µ0 Ue Xv µ0 Id ∗ µ0 Xe ∗ (Uf Xv) µ0 µ0 Ue Xv 16 ALCIDES BUSS AND RALF MEYER The top left square commutes because the multiplication maps are associative, the top right square because they are bibundle maps. The bottom left square commutes because µe,f = µ0, and the bottom right square commutes for trivial reasons. The bent composite arrows are the maps j by construction. Thus the whole diagram commutes, and this means that jv,u ◦ ju,t = jv,t. If t1 ≤ u1 and t2 ≤ u2 in S, then there is a commuting diagram of isomorphisms Xt1 ∗ Xt2 µt1 ,t2 Xt1t2 µt1 t∗ 1 ,u1 ∗µu2 ,t∗ 2 t2 µt1 t∗ 1 ,u1u2 ,t∗ 2 t2 (11) Xt1t∗ 1 ∗ Xu1 ∗ Xu2 ∗ Xt∗ 2 t2 Id∗µu1 ,u2 ∗Id Xt1t∗ 1 ∗ Xu1u2 ∗ Xt∗ 2 t2 µ0 µ0 Ut1 t∗ 1 Xu1 ∗ Xu2Ut∗ 2 t2 µu1 ,u2 Ut1 t∗ 1 Xu1u2 Ut∗ 2 t2 t ,u1,u2,t∗ Here we abbreviate ×G to ∗, µ0 denotes the left and right actions for subsets of G1, 2t2 denotes the appropriate combination of two multiplication maps, and µttt∗ which is well-defined by associativity. The upper square commutes by associativity. The lower square commutes because µu1,u2 is a bibundle map. The left vertical is ju1,t1 ∗ ju2,t2 because the two isomorphism from Xt1 ∗ Xt2 to Ut1 t∗ constructions in (9) coincide. It remains to see that the right vertical isomorphism from Xt1t2 to Ut1 t∗ Xu1 ∗ Xu2Ut∗ is ju1u2,t1t2. Xu1u2 Ut∗ t2 t2 1 2 The proof of this is similar to the proof that the two maps in (9) coincide. 1. Since r(Xt1t2) = Ue and (11) is a diagram of 2t2. Furthermore, ∗Xu1u2 ∗Xt∗ Let e = (t1t2)(t1t2)∗, so e ≤ t1t∗ isomorphisms, we have Xe∗Xt1t∗ the isomorphism ∗Xu1u2 ∗Xt∗ ∼= Xt1t∗ 2t2 1 1 1 2 µe,t1t∗ 1 ∗ Id : Xe ∗ Xt1t∗ 1 ∗ Xu1u2 ∗ Xt∗ 2 t2 → Xe ∗ Xu1u2 ∗ Xt∗ 2 t2 is equal to the standard multiplication map µ0 1. This fact and associativity show that the right vertical isomorphism in (11) is equal to the composite map ∗ Id because e ≤ ttt∗ e,t1t∗ 1 Xt1t2 µe,u1 u2 ,t2 t∗ 2 ←−−−−−−−− ∼= Xe ∗ Xu1u2 ∗ Xt∗ 2 t2 µ0 −→ ∼= Ue Xu1u2 Ut∗ 2 t2 = Ue Xu1u2. Similarly, we get the same composite map if we replace t∗ 2t2 on the right by the smaller idempotent f = (t1t2)∗(t1t2). Now the diagram (10) shows that the map we get is ju1u2,t1t2 as desired. Hence (8) commutes. (cid:3) Remark 3.8. Let E be a semilattice with unit 1, viewed as an inverse semigroup. An E-action on a topological groupoid G is the same as a unital semilattice map from E to the lattice of open G-invariant subsets of G0, that is, a map e 7→ Ue satisfying U1 = G0 and Ue ∩ Uf = Uef for all e, f ∈ E. The corresponding action by partial equivalences is defined by Xe := G1 → G1 . Proposition 3.7 implies that every action of E is isomorphic to one of this form. Ue and µe,f = µ0 : G1 Ue ×G G1 Uf Uef Proposition 3.9. There are unique bibundle isomorphisms Jt : X ∗ which the following composite map is the identity: t → Xt∗ for (12) Xt ∼= Xt ×G X ∗ t ×G Xt IdXt ×GJt×GIdXt −−−−−−−−−−−−→ Xt ×G Xt∗ ×G Xt µt,t∗ ,t−−−−→ Xt. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 17 These involutions also make the following diagrams commute: Xt ×G X ∗ t G1 Utt∗ (13) IdXt ×G Jt t ×G Xt X ∗ Jt ×G IdXt G1 Ut∗ t Xt ×G Xt∗ Xtt∗ µt,t∗ Xt∗ ×G Xt Xt∗t µt∗,t Here the unlabelled arrows are the canonical isomorphisms from Propositions 2.11 and 2.13. Furthermore, (Jt∗)∗ ◦ Jt : X ∗ is the identity map for all t ∈ S and the following diagrams commute for all t, u, v ∈ S with t ≤ u: t → Xt∗ → X ∗ t (14) u ×G X ∗ v X ∗ Ju ×G Jv Xu∗ ×G Xv∗ µ∗ v,u µu∗,v∗ X ∗ vu Jvu Xu∗v∗ X ∗ t Jt Xt∗ j∗ u,t X ∗ uUtt∗ JuUtt∗ Xu∗Utt∗ ju∗,t∗ Write x∗ := Jt(x) for x ∈ Xt and µt,u(x, y) = x · y for x ∈ Xt, y ∈ Xu with s(x) = r(y). The above diagrams and equations of maps mean that the involution is characterised by x · x∗ · x = x for all x ∈ Xt and has the properties x · x∗ = 1r(x), x∗ · x = 1s(x), (x∗)∗ = x, (x · y)∗ = y∗ · x∗, and ju∗,t∗(x∗) = ju,t(x)∗. Proof. The two isomorphisms µt,t∗,t : Xt ×G Xt∗ ×G Xt → Xt and µt∗,t,t∗ : Xt∗ ×G Xt ×G Xt∗ → Xt∗ that we may build from µ are equal by associativity. Propo- ∼= Xt∗ for sition 2.12 for these isomorphisms gives a unique isomorphism Jt : X ∗ t which (12) becomes the identity map. We claim that (12) is the identity if and only if either of the diagrams in (13) commutes. The proofs for both cases differ only by exchanging left and right, so we only write down one of them. Assume that the first diagram in (13) commutes. Applying the functor ␣×GIdXt to it, we get that the isomorphism (12) is the identity map because the multiplication map µtt∗,t : Xtt∗ ×G Xt → Xt is just the left action if we identify Xtt∗ ∼= G1 Utt∗ as usual. Conversely, assume that the isomorphism in (12) is the identity map. Take a further product with Xt∗ and then identify Xt ×G Xt∗ ∼= Xtt∗ via µt,t∗. Using again that the multiplication with Xtt∗ is just the G-action, this gives the first diagram in (13). Next we show that Jt∗ = (J −1 t commuting diagram )∗, which implies J ∗ t∗ ◦ Jt = IdXt . We use the X ∗ t∗ ×G Xt∗ (J −1 t )∗ ×G J −1 t Xt ×G X ∗ t IdXt ×G Jt Xt ×G Xt∗ G1 Utt∗ G1 Utt∗ µt,t∗ Xtt∗ t The top rectangle commutes because the pairing X ×G X ∗ → G1 r(X) is natural. The bottom diagram is the first one in (13). The large rectangle is the second diagram in (13) for t∗ with (J −1 )∗ instead of Jt∗. Since this diagram characterises Jt∗, we get Jt∗ = (J −1 )∗ as asserted. t Since the involution Jvu is uniquely characterised by a diagram like the first one in (13), we may prove the first diagram in (14) by showing that the composite map µu∗,v∗ ◦ (Ju ×G Jv) ◦ (µ∗ vu → Xu∗v∗ also makes the diagram in (13) for t = vu commute. This is a routine computation using the same diagrams for Ju and Jv and that the multiplication maps involving Xe for idempotent e ∈ S are always given by the left or right action because of the compatibility with j1,e. This v,u)−1 : X ∗ 18 ALCIDES BUSS AND RALF MEYER proof is a variant of the usual proof that (xy)−1 = y−1x−1 in a group because y−1x−1 · (xy) = 1. Similarly, we get the second diagram in (14) by showing that the composite t → Xt∗ satisfies the defining condition for Jt because ju,t (cid:3) map j−1 and ju∗,t∗ are compatible with the multiplication maps. u∗,t∗ ◦ Ju ◦ j∗ u,t : X ∗ 3.2. Transformation groupoids. Let (Xt, µt,u)t,u∈S be an action of a unital in- verse semigroup S on a topological groupoid G by partial equivalences. Define the embeddings ju,t : Xt → Xu for t ≤ u in S and the involutions X ∗ t → Xt∗ as in Propositions 3.7 and 3.9. Let X := Ft∈S Xt and define a relation ∼ on X by (t, x) ∼ (u, y) for x ∈ Xt, y ∈ Xu if there are v ∈ S with v ≤ t, u and z ∈ Xv with jt,v(z) = x and ju,v(z) = y. Lemma 3.10. The relation ∼ is an equivalence relation. Equip X∼ := X/∼ with the quotient topology. The quotient map π : X → X∼ is a local homeomorphism. It restricts to a homeomorphism from Xt onto an open subset of X∼ for each t ∈ S. Thus X∼ is locally quasi-compact or locally Hausdorff if and only if all Xt are. Proof. It is clear that ∼ is reflexive and symmetric. For transitivity, take (t1, x1) ∼ (t2, x2) ∼ (t3, x3). Then there are t12 ≤ t1, t2, t23 ≤ t2, t3, x12 ∈ Xt12, and x23 ∈ Xt23 with jti,t12(x12) = xi for i = 1, 2 and jti,t23(x23) = xi for i = 2, 3. Thus s(x12) = s(x2) = s(x23) ∈ s(t23) = s(t∗ 23t23, so that t ≤ ∼= Xt. t12 and t ≤ t2t∗ Let x be the image of x12 under this isomorphism. Then jt12,t(x) = x12. Hence jti,t(x) = jti,t12(jt12,t(x)) = xi for i = 1, 2. Since jt2,t23(x23) = x2 = jt2,t23(jt23,t(x)) and jt2,t23 is injective by Proposition 2.9, we get jt23,t(x) = x23 and hence also jt3,t(x) = x3. Thus x1 ∼ x3 as desired. 2t23 = t23. We have x12 ∈ Xt12 Ut∗ 23t23). Let t := t12t∗ ∼= Xt12 ×G Xt∗ 23t23 t23 23 We prove that π is open. Any open subset of X is a disjoint union of open subsets of the spaces Xt; so π is open if and only if all the maps Xt → X∼ are open. Let U ⊆ Xt be open, then we must check that π−1(π(U )) is open. This set is a union over the set of triples t, v, w ∈ S with w ≤ t, v, where the set for t, v, w is contained in Xv and consists of all jv,w(x) with x ∈ j−1 t,w(U ). The map jv,w is open by Proposition 2.9, and jt,w is continuous, so jv,w(j−1 t,w(U )) is open. Hence π−1(π(U )) is open as a union of open subsets of X, showing that π is open. If (t, x) ∼ (t, y), then there are u ≤ t and z ∈ Xu with x = jt,u(z) = y; so the map from Xt to X∼ is injective. Since π is open and continuous, it restricts to a homeomorphism from Xt onto an open subset of X∼. Thus π is a local homeomorphism. Since being locally Hausdorff or locally quasi-compact are local properties and π is a local homeomorphism, X∼ has one of these two properties if and only if X has, if and only if each Xt has. (cid:3) The space X∼ need not be Hausdorff, just as for étale groupoids constructed from inverse semigroup actions on spaces, where X∼ will be the groupoid of germs of the action (by Theorem 3.18). From now on, we identify Xt with its image in X∼, using that πXt : Xt → X∼ is a homeomorphism onto an open subset by Lemma 3.10. We are going to turn X∼ into a topological groupoid with the same object space G0 as G. Since ju,t is a bibundle map, it is compatible with range and source maps. So the maps r, s: Xt ⇒ G0 induce well-defined maps r, s: X∼ ⇒ G0. The multiplication maps µt,u give a continuous map X ×s,G0,r X → X, by map- ping the t, u-component of X ×s,G0,r X to the tu-component of X by µt,u. Equa- tion (8) shows that this descends to a well-defined continuous map µ: X∼ ×s,G0,r X∼ → X∼. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 19 Lemma 3.11. The maps r, s: X∼ ⇒ G0 and µ: X∼ ×s,G0,r X∼ → X∼ define a topological groupoid X∼. It contains G as an open subgroupoid. Hence X∼ is étale if and only if G is. Proof. The multiplication is associative already on X by (A3) and the associativity of S. The maps r and s are open on X∼ because they are so on each Xt. The maps r, s, µ restricted to G1 = X1 reproduce the groupoid structure on G by (A1). Even more, (A2) implies that multiplication in X∼ with elements of X1 is the same as the G-action. In particular, unit elements in G1 act identically, so they remain If x ∈ Xt, then x∗ ∈ Xt∗ satisfies µt,t∗(x, x∗) = 1r(x) by unit elements in X∼. Proposition 3.9. Hence π(x, t) · π(x∗, t∗) := π(µt,t∗ (x, x∗), tt∗) = π(1r(x), tt∗). This is equivalent to the unit element (1r(x), 1) in X∼ because j1,tt∗ is the usual inclusion map (more precisely, the computation above assumes that we identify Xtt∗ ∼= G1 Utt∗ ⊆ G1 using j1,tt∗ ). Similarly, π(x, t) · π(x∗, t∗) ∼ (1s(x), 1) is a unit element. Thus π(x∗, t∗) is inverse to π(x, t). The map π(x, t) 7→ π(x∗, t∗) is continuous. Thus we have a topological groupoid. We have seen above that it contains G as an open subgroupoid. Therefore, X∼ is étale if and only if G is. (cid:3) Definition 3.12. The groupoid X∼ is called the transformation groupoid of the S-action (Xt, µt,u) on G and denoted by G ⋊ S, or by G ⋊Xt,µt,u S if the action must be specified. Our proof shows that G ⋊ S with the family of open subsets (Xt)t∈S encodes all the algebraic structure of our action by partial equivalences. The next defi- nition characterises when a groupoid H with a family of subsets (Ht)t∈S is the transformation groupoid of an inverse semigroup action. Definition 3.13. Let S be an inverse semigroup. A (saturated) S-grading on a topological groupoid H is a family of open subsets (Ht)t∈S of H 1 such that (Gr1) Ht · Hu = Htu for all t, u ∈ S; (Gr2) H −1 t = Ht∗ for all t ∈ S; (Gr3) Ht ∩ Hu =Sv≤t,u Hv for all t, u ∈ S; (Gr4) H 1 =St∈S Ht. If S has a zero element 0, we may also require H0 = ∅. The conditions (Gr1) and (Gr2) imply that H1 is a subgroupoid of H, called the unit fibre of the grading. (Gr4) and (Gr1) imply that s(H1) = r(H1) = H 0. (Gr3) implies Hv ⊆ Hu for v ≤ u. A non-saturated S-grading would be defined by weakening (Gr1) to Ht·Hu ⊆ Htu for all t, u ∈ S. We only use saturated gradings and drop the adjective. Theorem 3.14. Let S be an inverse semigroup with unit. The transformation groupoid G⋊S of an S-action on a groupoid G by partial equivalences is an S-graded groupoid. Any S-graded groupoid (H, (Ht)t∈S) is isomorphic to one of this form, where G0 = H 0 and G1 = H1 ⊆ H. Two actions by partial equivalences are isomorphic if and only if their transformation groupoids are isomorphic in a grading- preserving way. Here an isomorphism between actions (Xt)t∈S and (Yt)t∈S by partial equiva- lences on two groupoids G and H means the obvious thing: a family of homeo- morphisms Xt ∼= Yt compatible with the range, source, and multiplication maps in Definition 3.5. 20 ALCIDES BUSS AND RALF MEYER Proof. It follows directly from our construction that the subspaces Xt ⊆ G ⋊ S for an S-action by partial equivalences satisfy (Gr1) -- (Gr4). It is also clear that the transformation groupoid construction is natural for isomorphisms of S-actions. Let H with the subspaces Ht for t ∈ S be an S-graded topological groupoid. Then G1 := H1 with G0 = H 0 is an open subgroupoid of H. Let Xt = Ht with the restriction of the range and source map of H, and with the G-action and maps µt,u : Xt ×s,G0,r Xu → Xtu from the multiplication map in H. This satisfies (S3) by definition, (S4) because Ht · Hu = Htu, (S1) and (S7) because H is a groupoid, and (S2) because Xt is open in H and the range and source maps of H are open. If (y, z) ∈ Xu ×s,G0,s Xtu, then zy−1 ∈ XtuXu∗ = Xtuu∗ ⊆ Xt because tuu∗ ≤ t. Hence (y, z) 7→ (zy−1, y) gives a continuous inverse for the map in (S5), so that the latter is a homeomorphism. A similar argument shows that the map in (S6) is a homeomorphism. Thus we get an S-action by partial equivalences. This construction is natural in the sense that isomorphic S-graded groupoids give isomorphic actions by partial equivalences actions. If we start with an action by partial equivalences, turn it into a graded group- oid, and then back into an action by partial equivalences, then we get an isomor- phic action by construction. When we start with a graded groupoid, go to an action by partial equivalences and back to a graded groupoid, then we also get back our original S-graded groupoid. The only non-trivial point is that the map π : Ft∈S Ht →(cid:0)Ft∈S Ht(cid:1)∼ identifies x ∈ Ht and y ∈ Hu for t, u ∈ S if and only if x = y in H; this is exactly the meaning of (Gr3). (cid:3) 3.3. Examples: group actions and actions on spaces. The equivalence be- tween actions by partial equivalences and graded groupoids makes it easy to de- scribe all actions of groups on groupoids and all actions of inverse semigroups on spaces. Theorem 3.15. Let G be a topological groupoid and let S be a group, viewed as an inverse semigroup. Then an S-action on G by (partial) equivalences is equivalent to a groupoid H containing G as an open subgroupoid with H 0 = G0, and with a continuous groupoid homomorphism π : H ։ S such that π−1(1) = G and, for each x ∈ H 0 and t ∈ S there is h ∈ H 1 with s(h) = x and π(h) = t. In this situation, H is the transformation groupoid G ⋊ S. If G is also a group, this is the same as a group extension G ֌ H ։ S. Proof. Since tt∗ = 1 for any t ∈ S, any action of S by partial equivalences will be an action by global equivalences. By Theorem 3.14, we may replace an S-action by partial equivalences by an S-graded groupoid (H, (Ht)t∈S). We have H 0 = G0 by construction. Since S is a group, (Gr3) says that Ht ∩ Hu = ∅ for t 6= u. Thus we get a well-defined map π : H 1 → S with π−1(t) = Ht; in particular, G = π−1(1). The map π is continuous because the subsets Ht are open. The condition on the existence of h for given x, t says that the map s: Ht → H 0 is onto, that is, Ht is a global equivalence. Thus an S-action on G gives π : H → S with the asserted properties. For the converse, let π : H → S be a groupoid homomorphism as in the statement. Define Ht := π−1(t) ⊆ H 1. These are open subsets because π is continuous. If t, u ∈ S, then HtHu ⊆ Htu is trivial. If h ∈ Htu, then our technical assumption gives h2 ∈ Hu with s(h2) = s(h). Then h1 := hh−1 2 ∈ Ht, so h ∈ HtHu. Thus (Gr1) holds. The remaining conditions for an S-grading are trivial in this case, and H is the transformation groupoid G ⋊ S by construction. If G is also a group, then so is H because G0 = H 0, and then the condition on π simply says that it is a surjection π : H ։ S with kernel G. This is the same as a group extension. (cid:3) INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 21 The obvious definition of a group action by automorphisms on another group only covers split group extensions. We need some kind of twisted action by auto- morphisms to allow for non-trivial group extensions as well. Our notion of action by equivalences achieves this very naturally. For groupoid extensions, one usually requires the kernel to be a group bundle; this need not be the case here. There are many examples of groupoid homomor- phisms (or 1-cocycles) with the properties required in Theorem 3.15. We mention one typical case: Example 3.16. Let H be the groupoid associated to a self-covering σ : X → X of a compact space X as in [9]. The canonical Z-valued cocycle π : H → Z on it clearly has the properties needed to define a Z-grading on H. The subgroupoid G := π−1(0) is the groupoid that describes the equivalence relation generated by x ∼ y if σk(x) = σk(y) for some k ∈ N. The action of σ on X preserves this equivalence relation and hence gives an endomorphism of G; this endomorphism is an equivalence, and our Z-action on G by equivalences is generated by this self- equivalence of G. But unless σ is a homeomorphism, σ is not invertible on G, so it gives no action of Z by automorphisms. Now we turn to actions of inverse semigroups on topological spaces. Let S be an inverse semigroup with unit and let Z be a topological space. First we recall Exel's construction of the groupoid of germs for an inverse semigroup action by partial homeomorphisms [12]. Let pHomeo(Z) be the inverse semigroup of partial homeomorphisms of Z. An action of S on Z by partial homeomorphisms is a monoid homomorphism θ : S → pHomeo(Z). This gives partial homeomorphisms θt : Dt∗t → Dtt∗ for t ∈ S with open subsets De ⊆ Z for e ∈ E(S). The groupoid of germs has object space Z, and its arrows are the "germs" [t, z] for t ∈ S, z ∈ Dt∗t; by definition, [t, z] = [u, z′] if and only if there is e ∈ E(S) with z = z′ ∈ De and te = ue. The groupoid structure is defined by s[t, z] = z, r[t, z] = θt(z), [t, z] · [u, z′] = [tu, z′] if z = θu(z′), and [t, z]−1 = [t∗, θt(z)]. The subsets {[t, z] z ∈ U } for t ∈ S and an open subset U ⊆ Dt∗t form a basis for the topology on the arrow space. Remark 3.17. Many authors use another germ relation that only requires an open subset V of Z with z ∈ V and θtV = θuV . This may give a different groupoid, of course. Exel's germ groupoids need not be essentially principal (see [32]). Theorem 3.18. Let Z be a topological space viewed as a topological groupoid, and let S be an inverse semigroup with unit. Isomorphism classes of actions of S on Z by partial equivalences are in natural bijection with actions of S on Z by partial homeomorphisms. The transformation groupoid Z ⋊ S for an action by partial equivalences is the groupoid of germs defined by Exel [12]. Proof. Let θ : S → pHomeo(Z) be an action of S by partial homeomorphisms. Exel's groupoid of germs carries an obvious S-grading by the open subsets Xt := {[t, z] z ∈ Dt∗t} with X1 = Z. The conditions in Definition 3.13 are trivial to check. Hence Exel's groupoid is the transformation groupoid Z ⋊ S for an action of S on Z by partial equivalences by Theorem 3.14. Conversely, an S-action on Z is equivalent to the S-graded groupoid Z ⋊S. This groupoid is étale. The assumptions of an S-grading imply that the subsets Xt ⊆ Z ⋊ S form an inverse semigroup of bisections that satisfies the assumptions in [12, Proposition 5.4], which ensures that the groupoid of germs is Z ⋊ S. (cid:3) Corollary 3.19. Let G be an étale groupoid, let S be an inverse semigroup, and let f : S → Bis(G) be a semigroup homomorphism. This induces an isomorphism 22 ALCIDES BUSS AND RALF MEYER G0 ⋊ S ∼= G if and only if St∈S f (t) = G and f (t) ∩ f (u) =Sv∈S,v≤t,u f (v) for all t, u ∈ S. Here G0 ⋊ S uses the action of S on G0 induced by f and the usual action of Bis(G). Proof. Add a unit to S and map it to the unit bisection G0 ⊆ G, so that we may apply Theorem 3.14. For t ∈ S, let Gt := f (t) ⊆ G1; these are open subsets because each f (t) is a bisection. Since f is a semigroup homomorphism, (Gr1) and (Gr2) hold. The other two conditions are exactly the technical assumptions of the corollary. Thus these two assumptions are equivalent to (Gt)t∈S being an S-grading on G. If they hold, then Theorem 3.14 says that G ∼= G1 ⋊ S = Z ⋊ S. Conversely, the transformation groupoid Z ⋊ S is S-graded by Theorem 3.14, so if G ∼= Z ⋊ S, then it satisfies the two technical assumptions. (cid:3) A subsemigroup S ⊆ Bis(G) with the properties required in Corollary 3.19 is called wide. Corollary 3.19 explains why they appear so frequently (see, for instance, [3, 12, 28]). [12, Proposition 5.4] already shows that Z ⋊ S = G if S is wide, but we have not seen the converse statement yet. Since the proof of Theorem 3.18 is not explicit, we give another pedestrian proof. Let θt : Dt∗t → Dtt∗ for t ∈ S give an action of S on Z by partial homeomor- phisms. This is a groupoid isomorphism from Dt∗t to Dtt∗, which we turn into a partial equivalence from Z to itself as in Example 2.5. Here this means that we take X ′ (z). Since all arrows in Z are units, the range and source maps determine the partial equivalence. The homeo- t to Xt = Dt∗t with r(z) := θt(z) morphism θt gives a bibundle isomorphism from X ′ and s(z) := z. The comparison with Exel's groupoid is more obvious for the second choice, which we take from now on. t := Dtt∗ with anchor maps r′(z) := z, s′(z) := θ−1 t There is an obvious homeomorphism Xt ×Z Xu ∼−→ {z ∈ Dtt∗ θt(z) ∈ Duu∗} = D(tu)∗(tu), such that the range and source maps are θtu and the inclusion map, respectively. We choose this isomorphism for µt,u to define our action by partial equivalences. Actually, this is no choice at all because the range and source maps are injective here, so there is at most one bibundle map between any two partial equivalences. (We will see more groupoids with this property in Section 3.5.) Hence the associativity condition in the definition of an inverse semigroup action holds automatically. Thus we have turned an action by partial homeomorphisms on Z into an action by partial equivalences on Z, viewed as a topological groupoid. Our construction of the transformation groupoid above is exactly the construc- tion of the groupoid of germs in this special case, so the isomorphism between Z ⋊S and the groupoid of germs from [12] is trivial. Next we check that every partial equivalence X of Z is isomorphic to one coming from a partial homeomorphism. Since all arrows in Z are units, we have X/Z = X = Z\X. Hence the anchor maps G0 ← X → H 0 are continuous, open and injective by condition (P3) in Proposition A.5. The map θ := r ◦ s−1 : s(X) → r(X) is a partial homeomorphism from G0 to H 0, and r : X → r(X) is an isomorphism of partial equivalences from X to Xθ. Since there is always only one isomorphism between partial equivalences coming from the same partial homeomorphism, an inverse semigroup action on Z is deter- mined uniquely by the isomorphism classes of the Xt, which are in bijection with partial homeomorphisms of Z. This proves the first statement in Theorem 3.18. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 23 3.4. Morita invariance of actions by partial equivalences. Proposition 3.20. Let Y be an equivalence from H to G, and let (Xt, µt,u) be an t := Y ×G Xt ×G Y ∗ and let action of an inverse semigroup with unit on G. Let X ′ µ′ t,u : X ′ tu be the composite isomorphism t ×H X ′ u → X ′ Y ×G Xt ×G Y ∗ ×H Y ×G Xu ×G Y ∗ ∼−→ Y ×G Xt ×G G1 ×G Xu ×G Y ∗ ∼−→ Y ×G Xt ×G Xu ×G Y ∗ µt,u−−→ ∼= Y ×G Xtu ×G Y ∗, where the first two isomorphisms are the canonical ones from Proposition 2.11 and Lemma 2.8. Then µ′ t,u is an action of S on H by partial equivalences. Its transfor- mation groupoid H ⋊ S is equivalent to G ⋊ S. When we translate the action on Y back to X using the inverse equivalence Y ∗, we get an action on G that is isomorphic to the original one. 1 := H 1 for t = 1, and let µ′ 1 as defined above is only isomorphic to H 1 in a very Proof. More precisely, X ′ obvious way. We should only use the above definition of X ′ t for t 6= 1 and let X ′ t,1 be the canonical isomorphisms. We should also put in associators for the composition of partial equivalences, which only cause notational complications, however. Up to these technicalities, it is clear that the maps µ′ inherit associativity from the maps µ. The action that we get by translating µ′ back to G with Y ∗ is canonically isomorphic to the original action because Y ∗ ×H Y ∼= G1. 1,t and µ′ It remains to prove the equivalence of the transformation groupoids G ⋊ S and H ⋊ S. Here we use the linking groupoid L of the equivalence; its object space is L0 = G0 ⊔ H 0, its arrow space is G1 ⊔ Y ⊔ Y ∗ ⊔ H 1, its range and source maps are r and s on each component, and its multiplication consists of the multiplica- tions in G and H, the G, H-bibundle structure on Y , the H, G-bibundle structure on Y ∗, and the canonical isomorphisms Y ×G Y ∗ ∼−→ H 1 and Y ∗ ×H Y ∼−→ G1 from Proposition 2.11. This gives a topological groupoid L. There is a canonical right action of L on G1 ⊔ Y = r−1(G1) ⊆ L1 that provides an equivalence from L to G when combined with the left actions of G on G1 and Y ; there is a similar canonical equivalence H 1 ⊔ Y ∗ from L to H. We may transport the S-action on G to L because it is equivalent to G. When we transport this action on L further to H, we get the action described above because the composite equivalence (G1 ⊔ Y ) ×L (H 1 ⊔ Y ∗)∗ from H to G is isomorphic to Y . The action on L is given by bibundles (G1 ⊔ Y )∗ ×G Xt ×G (G1 ⊔ Y ) ∼= Xt ⊔ (Xt ×G Y ) ⊔ (Y ∗ ×G Xt) ⊔ (Y ∗ ×G Xt ×G Y ), where we cancelled factors of G1 using Lemma 2.8. When we restrict the trans- formation groupoid L ⋊ S to G0 ⊆ L0 or to H 0 ⊆ L0, then we only pick the components Xt and Y ∗ ×G Xt ×G Y in the above decomposition, so we get the transformation groupoids G ⋊ S and H ⋊ S, respectively. Routine computations show that the other two parts r−1(G0) ∩ s−1(H 0) and r−1(H 0) ∩ s−1(G0) of L ⋊ S give an equivalence from H ⋊ S to G ⋊ S, such that L ⋊ S is the resulting linking groupoid. It can be shown with less routine computations that the embedding G ⋊ S ֒→ L ⋊ S is fully faithful and essentially surjective. We checked "fully faithful" above. Being "essentially surjective" means that the map G0 ×⊂,L0,r L1 → L0, (x, l) 7→ s(l), is open and surjective. It is open because r : L1 → L0 is open and G0 ⊂ L0 is open, and surjective because already G0 ×L0 Y ⊂ G0 ×L0 L1 surjects onto H 0. Since both G ⋊ S ֒→ L ⋊ S and H ⋊ S ֒→ L ⋊ S are fully faithful and essentially surjective, 24 ALCIDES BUSS AND RALF MEYER they induce equivalence bibundles by [21, Proposition 6.8], which we may compose to an equivalence from H ⋊ S to G ⋊ S. Of course, this gives the same equivalence as the argument above. (cid:3) Corollary 3.21. Let S be an inverse semigroup with unit. Let f : X → Z be an open continuous surjection and let G(f ) be its covering groupoid, see Defini- tion A.8. Then S-actions by partial equivalences on G(f ) are canonically equivalent to S-actions on Z by partial homeomorphisms, such that G(f ) ⋊ S is equivalent to Z ⋊ S. Here "equivalent" means an equivalence of categories, where the arrows are iso- morphisms of S-actions that fix the underlying groupoid. Proof. G(f ) is canonically equivalent to Z viewed as a groupoid, so the assertion follows from Theorem 3.18 and Proposition 3.20. (cid:3) In particular, Corollary 3.21 applies to the Čech groupoid GU of an open cov- ering U of a locally Hausdorff space Z by Hausdorff open subsets. Thus we may replace an S-action by partial homeomorphisms on a locally Hausdorff space Z by an "equivalent" action by partial equivalences on a Hausdorff groupoid GU, and the resulting transformation groupoids Z ⋊ S and GU ⋊ S are equivalent. The quickest way to describe the resulting S-action on GU explicitly is by de- scribing GU ⋊ S and an S-grading on it. Let X := FU∈U U and let p: X → Z be the canonical map, which is an open surjection. The pull-back p∗(Z ⋊ S) of Z ⋊ S along p is a groupoid with object space X, arrow space X ×p,Z,r (Z ⋊ S)1 ×s,Z,p X, r(x1, g, x2) = x1, s(x1, g, x2) = x2, and (x1, g, x2) · (x2, h, x3) = (x1, g · h, x3) (see [21, Example 3.13]). Let Xt := {(x1, g, x2) ∈ X ×p,Z,r (Z ⋊ S)1 ×s,Z,p X g ∈ t}. Proposition 3.22. The subspaces Xt ⊆ p∗(Z ⋊S)1 form an S-grading on p∗(Z ⋊S). The resulting S-graded groupoid is the transformation groupoid for the S-action on GU that we get by translating the S-action on Z along the equivalence to GU. Proof. The subspaces Xt form an S-grading because the bisections t ∈ S give an S-grading on Z ⋊ S and p is surjective. Hence they describe an S-action on GU. The equivalence from GU to Z is given by the canonical action of GU on G0 U = X and the projection p: X → Z. Hence Xt is exactly what we get when we translate t ⊆ Z ⋊ S along the equivalence. (cid:3) For instance, let H be an étale groupoid with locally Hausdorff arrow space and let S be some inverse semigroup of bisections with H ∼= H 0 ⋊ S; we could take S = Bis(H). Let Z = H 1 with the action of H by left multiplication. This induces an action of S on Z. Its transformation groupoid H 1 ⋊S is H 1 ⋊H with the obvious S-grading by H 1 ×H0 t for t ∈ S. The left multiplication action of H on H 1 with the bundle projection s: H 1 → H 0 is a trivial principal bundle. In particular, the transformation groupoid H 1⋊S ∼= H 1 ⋊ H is isomorphic to the covering groupoid of the cover s: H 1 ։ H 0. Hence it is equivalent to the space H 0, viewed as a groupoid with only unit arrows. The S-grading on H 1 ⋊ S does, however, not carry over to H 0. If we replace the S-action on H 1 by an equivalent S-action on GU for a Hausdorff open cover of H 1, then the transformation groupoid GU ⋊ S is equivalent to H 1 ⋊ S and hence also equivalent to the space H 0. In particular, the groupoid GU ⋊ S is basic (see Section A.2). If a groupoid is equivalent to a space, then this space has to be its orbit space. So if H 0 is Hausdorff, then the groupoid GU ⋊ S is a free and proper, Hausdorff groupoid by Proposition A.7. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 25 3.5. Local centralisers. We are going to show that for many groupoids G the bibundles Xt already determine the multiplication maps µt,u and thus the inverse semigroup action. This happens, among others, for essentially principal topological groupoids (meaning that the isotropy group bundle has no interior; see [32]) and for groups with trivial centre. Definition 3.23. A local centraliser of G is a map γ : U → G1 defined on an open G-invariant subset U of G0 with s(γ(x)) = r(γ(x)) = x for all x ∈ U and γ(r(g)) · g = g · γ(s(g)) for all g ∈ G. We say that G has no local centralisers if all local centralisers are given by γ(x) = 1x for x ∈ U and some U as above. Local centralisers defined on the same subset U form an Abelian group under pointwise multiplication. All local centralisers form an Abelian inverse semigroup. It is the centre of Bis(G) if G is étale. Lemma 3.24. Let X be a partial equivalence from H to G. Then Map(X, X) is isomorphic to the group of local centralisers of G defined on r(X), and to the group of local centralisers of H defined on s(X). If G has no local centralisers and X and Y are partial equivalences from G or to G, then there is at most one bibundle map X → Y , so bibundle isomorphisms are unique if they exist. Proof. The two descriptions of Map(X, X) are equivalent by taking X ∗, so we only prove one. Every bibundle map X → X is invertible by Proposition 2.9. The canonical group homomorphisms Map(X, X) ␣×H X ∗ −−−−−→ Map(X ×H X ∗, X ×H X ∗) ∼= Map(G1 r(X)) r(X), G1 ␣×GX−−−−→ Map(X, X) U , G1 are inverse to each other by the proof of Proposition 2.11. Thus it remains to iden- U ) for an open G-invariant subset U of G0 with the group tify the set Map(G1 of local centralisers defined on U . We may view G1 U to itself associated to the identity functor on G1 U . We described all bibundle isomor- phisms between such equivalences in Example 2.5. Specialising Example 2.5 to the automorphisms of the identity functor gives exactly the local centralisers defined on U . A quick computation shows that the composition of bibundle isomorphisms corresponds to the pointwise multiplication of local centralisers. U as the equivalence from G1 Let f1, f2 : X → Y be bibundle maps. Then both are bibundle isomorphisms X → Y s(X), and we may form a composite bibundle isomorphism f −1 2 ◦f1 : X → X. Since there are no local centralisers, the first part of the lemma shows that this is the identity map, so f1 = f2. In particular, if two partial equivalences G → H or H → G are isomorphic, then the isomorphism is unique. (cid:3) tial equivalences on G. More precisely, isomorphism classes of S-actions on G by partial equivalences are Theorem 3.25. Let G be a topological groupoid without local centralisers. An ac- Recall that fpeq(G) denotes the inverse semigroup of isomorphism classes of par- tion of an inverse semigroup S on G is equivalent to a homomorphism S → fpeq(G). in canonical bijection with homomorphisms S → fpeq(G). Proof. A homomorphism f : S → fpeq(G) gives us bibundles Xt with Xt ×G Xu ∼= Xtu and X1 ∼= G1; we may as well assume X1 = G1. By Lemma 3.24, the isomor- ∼−→ Xtu above are unique, so there is no need to specify them. phisms µt,u : Xt×G Xu The conditions (A2) and (A3) hold because any two parallel bibundle isomorphisms are equal. Thus f determines an S-action by partial equivalences. Conversely, an 26 ALCIDES BUSS AND RALF MEYER action by partial equivalences determines such a homomorphism by taking the iso- morphism classes of the Xt and forgetting the µt,u. Since isomorphisms of partial equivalences are unique if they exist, this forgetful functor is actually not forget- ting anything here, so we get a bijection between isomorphism classes of actions by (cid:3) partial equivalences and homomorphisms S → fpeq(G). The results in this section are inspired by the notion of a "quasi-graphoid" used by Debord in [10]. Debord already treats partial equivalences of groupoids as arrows between groupoids and uses them to glue together groupoids constructed locally. She restricts, however, to a situation where bibundle isomorphisms are uniquely determined. Even more, she wants the range and source maps to already determine a partial equivalence uniquely. For this, she assumes that a smooth map γ : U → G1 defined on an open subset U of G0 has to be the unit section already if it only satisfies s(γ(x)) = r(γ(x)) = x for all x ∈ U . This condition holds for holonomy groupoids of foliations -- even for the mildly singular foliations that she is considering. 3.6. Decomposing proper Lie groupoids. A manifold may be constructed by taking a disjoint union of local charts and gluing them together along the coordi- nate change maps, which are partial homeomorphisms, or diffeomorphisms in the smooth case. When constructing groupoids locally, it is more likely that the coordi- nate change maps are no longer partial isomorphisms but only partial equivalences. Actually, it may well be that the local pieces are, to begin with, only local group- oids and not groupoids (see [10]); this is not covered by our theory. Therefore, we know no good examples where groupoids have been constructed by gluing together smaller groupoids along partial equivalences. Instead, we take a groupoid as given and analyse it using local information. The local information should say that the groupoid locally is equivalent to one of a par- ticularly simple form. Then the groupoid is globally equivalent to a transformation groupoid for an inverse semigroup action by partial equivalences on a disjoint union of groupoids having the desired simple form. We now get more concrete and consider a proper Lie groupoid H. To formulate stronger results, we shall work with (partial) equivalences of Lie groupoids in this section; that is, spaces are replaced by smooth manifolds, continuous maps by smooth maps, and open maps by submersions. This does not change the theory significantly, see [21]. First we formulate the local linearisability of proper Lie groupoids. This was conjectured by Weinstein [37] and proved by Zung [38]. Both authors try to de- scribe the local structure of proper Lie groupoids up to isomorphism. Following Trentinaglia [35], we only aim for a description up to Morita equivalence: Theorem 3.26. Let H be a proper Lie groupoid. For every x ∈ H 0 there are an open H-invariant neighbourhood Ux of x in H 0, a linear representation of the stabiliser group Hx on a finite-dimensional vector space Wx, and a Lie groupoid equivalence from the transformation groupoid Wx ⋊ Hx to HUx. The vector space Wx is the normal bundle to the H-orbit Hx of x, with its canonical representation of Hx. Weinstein and Zung impose extra assumptions on H to describe HUx up to isomorphism. The argument in [35, Section 4] shows how to deduce Theorem 3.26 quickly from [38, Theorem 2.3] without extra assumptions. Actually, we do not need H to be proper. Since we only need local structure, it is enough for H to be locally proper, that is, each x ∈ H 0 has an H-invariant open neighbourhood U such that HU is proper; this allows the orbit space H 0/H to be a locally Hausdorff but non-Hausdorff manifold. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 27 Assume now that H is a locally proper groupoid. By Theorem 3.26, there is a covering U of H 0 by open, H-invariant subsets and, for each U ∈ U, a Lie groupoid equivalence XU from a transformation groupoid WU ⋊ GU for a compact Lie group GU and a linear representation WU of GU to the restriction HU . Now let G := GU∈U WU ⋊ GU . Let K be the covering groupoid of H 0 for the covering U. Since HU U∩V = This disjoint union is a groupoid with object space F WU . HU∩V = HV U∩V , the inverse semigroup S := Bis(K) acts on FU∈U HU : each The disjoint union X :=FU∈U XU gives an equivalence from G toFU∈U HU , so we element of Bis(K) acts by the identity equivalence between the appropriate restric- tions of HU and HV , and all the multiplication maps are the canonical isomorphisms. may transfer this S-action to G. We make the action on G more concrete. Any bisection of K is a disjoint union of bisections of the form (U1, D, U2) := {(U1, x, U2) x ∈ D} for U1, U2 ∈ U and an open subset D ⊆ U1 ∩ U2. The product (U1, D1, U2) · (U ′ 2, D2, U3) is empty if U2 6= U ′ The partial equivalence XU1,D,U2 on G associated to (U1, D, U2) is the composite 2, and is equal to (U1, D1 ∩ D2, U3) if U2 = U ′ 2. partial equivalence G ⊇ WU1 ⋊ GU1 D X ∗ U1−−−−→ HD XU2 D−−−−→ WU2 ⋊ GU2 ⊆ G. The composite of XU1,D1,U2 and XU ′ be. If U2 = U ′ 2, then there is a canonical isomorphism of partial equivalences 2,D2,U3 is clearly empty for U2 6= U ′ 2, as it should µ(U1,D1,U2),(U2,D2,U3) : XU1,D1,U2 ×G XU2,D2,U3 → XU1,D1∩D2,U3, using the restriction of the canonical pairing XU2 ×G X ∗ U2 → HU2 to remove the extra two factors in the middle. This is exactly what happens if we translate the "trivial" action of S on F HU described above to G along the equivalence F XU . Theorem 3.27. The locally proper Lie groupoid H is equivalent to the transfor- mation groupoid G ⋊ S for the action of S on G described above. Proof. Since we constructed the action of S on G by translating the action on FU∈U HU , Proposition 3.20 shows that G⋊S is equivalent toFU∈U HU ⋊S. Since S acts "trivially" on F HU , this transformation groupoid is easy to understand: it is the pull-back p∗(H) of H for the canonical map p: FU∈U U → H 0. Since p is a surjective submersion, p∗(H) is equivalent to H. (cid:3) As a result, any locally proper Lie groupoid is equivalent to a transformation groupoid for an inverse semigroup action on a disjoint union of linear actions of compact groups. Such transformation groupoids need not be locally proper, how- ever, so we do not have a characterisation of locally proper Lie groupoids. The groupoid G ⋊ S is étale if and only if G is, if and only if the stabilisers Hx are finite. This means that H is an orbifold (see [22]). 4. Inverse semigroup actions on C∗-algebras We now define inverse semigroup actions on C∗-algebras by Hilbert bimodules, in parallel to actions on groupoids by partial equivalences. Definition 4.1. A Hilbert A, B-bimodule H is a left Hilbert A-module and a right Hilbert B-module such that the left and right multiplications commute, and 28 ALCIDES BUSS AND RALF MEYER hhxyiiA ·z = x·hyziB for all x, y, z ∈ H. A Hilbert A, B-bimodule map is a bimodule map that also intertwines both inner products. Let H be a Hilbert A, B-bimodule. Let I ⊳ A and J ⊳ B be the closed linear spans of the elements hhxyiiA and hxyiB with x, y ∈ H, respectively. These are closed ideals in A and B, and H is an I, J-imprimitivity bimodule by restricting the left multiplications to I and J. Ideals in a C∗-algebra are in bijection with open subsets of its primitive ideal space, so ideals are the right analogues of open invariant subsets of groupoids. Hence we denote the ideals I and J above as I := r(H) and J := s(H), and we think of Hilbert A, B-bimodules as partial Morita equivalences from B to A. Given an ideal K ⊳ A, we define the restriction of a Hilbert bimodule H to K as KH := K · H ⊆ H, which is canonically isomorphic to K ⊗A H. We restrict to ideals in B in a similar way. The left action of A on a Hilbert bimodule is by a nondegenerate ∗-homomor- phism A → B(H) into the adjointable operators on H. Thus a Hilbert A, B- bimodule becomes a correspondence by forgetting the left inner product. Lemma 4.2. A correspondence H carries a Hilbert bimodule structure if and only if there is an ideal I ⊳ A such that the left action ϕ: A → B(H) restricts to an isomorphism from I onto K(H). This ideal and the left inner product are uniquely determined by the correspondence. Proof. First let H be a Hilbert bimodule. Then H is an imprimitivity bimodule from s(H) to r(H), so ϕr(H) is an isomorphism from r(H) onto K(H). If I ⊳ A is another ideal with ϕ(I) = K(H), then ϕ(r(H) · I) = K(H) as well. Thus r(H) is the minimal ideal that ϕ maps onto K(H), and the only one on which this happens isomorphically. Thus r(H) is already determined by the underlying correspondence. Let H′ be another Hilbert A, B-bimodule with the same underlying correspon- dence as H and with left A-valued inner product hhxyii′ A. Then ϕ(hhxyii′ A)z = x · hyziB = ϕ(hhxyiiA)z for all x, y, z ∈ H. Since r(H) = r(H′) depends only on the correspondence and the restriction of ϕ to r(H) is faithful, we get H = H′ as Hilbert bimodules. Now let H be a correspondence and let I ⊳ A be an ideal that is mapped isomorphically onto K(H). Transfer the usual K(H)-valued left inner product on H through this isomorphism to one with values in A ⊇ I. This turns H into a Hilbert A, B-bimodule. (cid:3) Proposition 4.3. Let H and H′ be Hilbert A, B-bimodules. If there is a Hilbert bimodule map f : H → H′, then s(H) ⊆ s(H′) and r(H) ⊆ r(H′). Such a Hilbert bimodule map is an isomorphism from H onto the submodule H′ · s(H) = r(H) · H′ in H′. So it is an isomorphism onto H′ if and only if s(H′) ⊆ s(H), if and only if r(H′) ⊆ r(H), if and only if the map K(H) → K(H′) induced by f is an isomorphism. Proof. Since the norm on a Hilbert bimodule is generated by the inner products, Hilbert bimodule maps are norm isometries and thus injective. Moreover, f (H) = f (r(H) · H) = r(H) · f (H) ⊆ r(H) · H′. Thus r(H′) ⊆ r(H) is necessary for f to be an isomorphism. Conversely, if r(H′) ⊆ r(H), then even r(H′) = r(H) because a bimodule map preserves the left inner product. Then the map from r(H) ∼= K(H) to r(H′) ∼= K(H′) that sends ξihη to f (ξ)ihf (η) for ξ, η ∈ H is an isomorphism K(H) ∼= r(H′) ∼= K(H′). Since K(H′) · H′ = H′, the linear span of elements of the form f (ξ)ihf (η)ζ′ = f (ξ) · hf (η)ζ′i for ξ, η ∈ H, ζ′ ∈ H′ is dense in H′. Since f (H) is a right B-module, this implies INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 29 that f is surjective. Hence it is an isomorphism of Hilbert bimodules. A similar argument for the right inner product instead of the left one shows that all the listed conditions for f are indeed equivalent to f being an isomorphism. If r(H) 6= r(H′), then we may restrict f to a Hilbert bimodule map H → r(H)·H′. Since r(H) · r(H) · H′ = r(H) · H′, this is an isomorphism by the first statement. A similar argument on the other side shows that f is an isomorphism onto H′ · s(H), so H′ · s(H) = r(H) · H′. (cid:3) A Hilbert A, B-bimodule H has a dual Hilbert B, A-bimodule H∗, where we exchange left and right structures using adjoints: b · x∗ · a := (a∗ · x · b∗)∗ for a ∈ A, b ∈ B, x ∈ H, and hx∗y∗iA = hhyxiiA, hhx∗y∗iiB = hyxiB. We will see that this construction has the same formal properties as the dual for partial equivalences of groupoids. To begin with, a Hilbert bimodule map X → Y remains a Hilbert bimodule map X ∗ → Y ∗, and (X ∗)∗ = X. Furthermore, (ξ ⊗ η)∗ 7→ η∗ ⊗ ξ∗ defines a Hilbert bimodule map σ : (X ⊗B Y )∗ → Y ∗ ⊗B X ∗ with dense range, hence an isomorphism. Applying σ twice gives the identity map. (More precisely, σY ∗,X ∗ ◦ σX,Y = Id(X⊗B Y )∗.) Proposition 4.4. Let H be a Hilbert A, B-bimodule. The inner products on H give Hilbert bimodule isomorphisms H ⊗B H∗ ∼= r(H) and H∗ ⊗A H ∼= s(H), and the restrictions of the left and right actions give Hilbert bimodule isomorphisms r(H) ⊗A H ∼= H ∼= H ⊗B s(H), s(H) ⊗B H∗ ∼= H∗ ∼= H∗ ⊗A r(H). that make the following diagrams of isomorphisms commute: H ⊗B H∗ ⊗A H H ⊗B s(H) H∗ ⊗A H ⊗B H∗ H∗ ⊗A r(H) (15) r(H) ⊗A H H, s(H) ⊗B H∗ H∗. Let D be another C∗-algebra, let K be a Hilbert A, D-bimodule and let L be a Hilbert B, D-bimodule with r(K) ⊆ r(H) and r(L) ⊆ s(H). Then Hilbert A, D- bimodule maps H ⊗B L → K are naturally in bijection with Hilbert B, D-bimodule maps L → H∗ ⊗A K, and this bijection maps isomorphisms again to isomorphisms. Similarly, Hilbert A, D-bimodule maps H ⊗B L ← K are naturally in bijection with Hilbert B, D-bimodule maps L ← H∗ ⊗A K. Proof. The Hilbert bimodule isomorphisms H ⊗B H∗ ∼= r(H), H∗ ⊗A H ∼= s(H), r(H) ⊗A H ∼= H ∼= H ⊗B s(H) and s(H) ⊗B H∗ ∼= H∗ ∼= H∗ ⊗A r(H) are routine to check using that H is full as a Hilbert r(H), s(H)-bimodule. The diagrams in (15) are equivalent to the requirement hhxyiiA · z = x · hyziB in the definition of a Hilbert bimodule. The claim about Hilbert bimodule maps is proved like the analogous one about partial equivalences of groupoids in Proposition 2.12; now we use the canonical isomorphisms just established and Proposition 4.3 instead of Proposition 2.9. (cid:3) Proposition 4.5. Up to isomorphism, H∗ is the unique Hilbert B, A-bimodule K for which there are isomorphisms H ⊗B K ⊗A H ∼= H, K ⊗A H ⊗B K ∼= K. More precisely, if there are such isomorphisms then there is a unique Hilbert bimod- ule isomorphism H∗ ∼−→ K such that the following map is the identity map: H ∼−→ H ⊗B K ⊗A H ∼−→ H ⊗B H∗ ⊗A H ∼−→ r(H) ⊗A H ∼−→ H. Proof. Repeat the proof of Proposition 2.12, replacing ×G by ⊗A. (cid:3) 30 ALCIDES BUSS AND RALF MEYER Proposition 4.6. Let H be a Hilbert A, A-bimodule and let µ: H ⊗A H → H be a bimodule isomorphism. Then there is a unique isomorphism from H onto an ideal I ⊳ A that intertwines µ and the multiplication map I ⊗A I ∼−→ I. We have I = r(H) = s(H), and the multiplication µ is associative. Proof. This is proved exactly like Proposition 2.13. Definition 4.7. Let S be an inverse semigroup with unit and let A be a C∗-algebra. An S-action on A by Hilbert bimodules consists of (cid:3) • Hilbert A, A-bimodules Ht for t ∈ S; • bimodule isomorphisms µt,u : Ht ⊗A Hu ∼−→ Htu for t, u ∈ S; satisfying (AH1) H1 is the identity Hilbert A, A-bimodule A; (AH2) µt,1 : Ht ⊗A A ∼−→ Ht and µ1,u : A ⊗A Hu ∼−→ Hu are the canonical isomor- phisms for all t, u ∈ S; (AH3) associativity: for all t, u, v ∈ S, the following diagram commutes: (Ht ⊗A Hu) ⊗A Hv ass Ht ⊗A (Hu ⊗A Hv) µt,u ⊗A IdHv IdHt ⊗A µu,v Htu ⊗A Hv Ht ⊗A Huv µtu,v Htuv µt,uv If S has a zero element 0, we may also require H0 = {0}. Theorem 4.8. Let S be an inverse semigroup with unit, let A be a C∗-algebra. Then actions of S on A by Hilbert bimodules are equivalent to saturated Fell bundles over S (as defined in [13]) with unit fibre A. More precisely, let (Ht)t∈S and (µt,u)t,u∈S be an S-action by Hilbert bimodules on A. Then there are unique Hilbert bimodule maps ju,t : Ht → Hu for t ≤ u that make the following diagrams commute for all t1, t2, u1, u2 ∈ S with t1 ≤ u1, t2 ≤ u2: (16) Ht1 ⊗A Ht2 µt1,t2 Ht1t2 ju1,t1 ⊗A ju2,t2 Hu1 ⊗A Hu2 µu1,u2 ju1u2,t1t2 Hu1u2 The map ju,t is a Hilbert bimodule isomorphism onto Hu · s(Ht) = r(Ht) · Hu. We have jt,t = IdHt for all t ∈ S and jv,u ◦ ju,t = jv,t for t ≤ u ≤ v in S. And ∼−→ Ht∗, x 7→ x∗, such there are unique Hilbert bimodule isomorphisms Jt : H∗ t that µt,t∗,t(x, x∗, x) = x · hxxiA = hhxxiiA · x for all x ∈ Ht. These also satisfy µt,t∗(x ⊗ x∗) = hhxxiiA, µt∗,t(x∗, x) = hxxiA and (x∗)∗ = x for all x ∈ Ht; µt,u(x, y)∗ = µu∗,t∗(y∗, x∗) for all x ∈ Ht, y ∈ Hu, t, u ∈ S; and ju,t(x)∗ = ju∗,t∗(x∗) for all t ≤ u in S, x ∈ Ht. Conversely, a saturated Fell bundle (At)t∈S over S with A = A1 becomes an S-action by Hilbert bimodules by taking Ht = At with the multiplication maps µt,u and the A-bimodule structure induced by the Fell bundle multiplication, and the left and right inner products hhxyiiA := x · y∗, hxyiA := x∗ · y for x, y ∈ Ht. Proof. We construct the inclusion maps jt,u and the involutions Jt and show their properties exactly as in the proofs of Propositions 3.7 and 3.9. (cid:3) With Theorem 4.8, it becomes easier to construct saturated Fell bundles over in- verse semigroups because Definition 4.7 needs far less data and has correspondingly fewer conditions to check. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 31 Remark 4.9. The correspondence bicategory introduced in [6] is not suitable for our purposes by the following observation: Let I ֒→ A ։ A/I be a split extension of C∗-algebras. Then p: A → A/I → A is an idempotent endomorphism. It re- mains an idempotent arrow in the correspondence bicategory. More generally, if A is Morita equivalent to an ideal in a C∗-algebra B, then we can translate p to a correspondence H from B to itself that is idempotent in the sense that H⊗B H ∼= H with an associative isomorphism. Thus there are more idempotent endomorphisms in the correspondence bicategory than usual for inverse semigroup actions. Fur- thermore, the idempotent arrows no longer commute up to isomorphism; thus a very basic assumption for inverse semigroups fails in this case. This is why we only allowed Hilbert bimodules above. Proposition 4.10. There is a bicategory with C∗-algebras as objects, Hilbert bimod- ules as arrows, Hilbert bimodule isomorphisms as 2-arrows, and ⊗B as composition of arrows. Proof. The correspondence bicategory is constructed already in [6]. Lemma 4.2 allows to identify Hilbert bimodules with a subset of correspondences. It is well- known that composites of Hilbert bimodules are again Hilbert bimodules. Hence the Hilbert bimodules form a sub-bicategory in the opposite of the correspondence bicategory. (cid:3) 5. Fell bundles from actions of inverse semigroups All groupoids in this section are assumed to be locally quasi-compact, locally Hausdorff and with (locally compact) Hausdorff object space and a Haar system, so that they have groupoid C∗-algebras. Let G be such a groupoid and let S be a unital inverse semigroup acting on G by partial equivalences. We want to turn this into an action of S on C∗(G) by Hilbert bimodules; equivalently, we want a Fell bundle over S with unit fibre C∗(G). There are two closely related ways to construct this. We are going to explain one approach in detail and only sketch the other one briefly in Section 5.3. We give details for the construction of the Fell bundle using the transforma- tion groupoid L = G ⋊ S because this also suggests how to describe the section C∗-algebra of the resulting Fell bundle. The transformation groupoid L comes with an S-grading (Lt)t∈S. Roughly speaking, our Fell bundle over S will involve the subspaces of C∗(L) of elements supported on the open subsets Lt. Since G1 = L1, the unit fibre of the Fell bundle will be C∗(G). This also suggests that the section C∗-algebra of the Fell bundle over S is C∗(L). This is indeed the case, but the technical details need some care. First, we need a Haar system on L. We show in Proposition 5.1 that the Haar system on G extends uniquely to a Haar system on L. Secondly, it is non-trivial that C∗(G) is contained in C∗(L): this means that the maximal C∗-norm that defines C∗(G) extends to a C∗-norm on C∗(L). A related issue is to show that an element of C∗(L) supported in G actually belongs to C∗(G). These problems become clearer if we construct a pre-Fell bundle using the dense ∗-algebra that defines C∗(L) and then complete it. In the non-Hausdorff case, continuous functions with compact support are re- placed by finite linear combinations of certain functions that are not continuous. The identification of C∗(L) with the section C∗-algebra of the Fell bundle requires a technical result about these functions. We prove it in Appendix B in the more general setting of sections of upper semicontinuous Banach bundles because this is not more difficult and allows us to generalise our main results to Fell bundles over groupoids. 32 ALCIDES BUSS AND RALF MEYER We write S(X) for the space of linear combinations of compactly supported functions on Hausdorff open subsets of a locally Hausdorff, locally quasi-compact space X. This is the space of compactly supported continuous functions on X if and only if X is Hausdorff, and it is often denoted by Cc(X). We find this notation misleading, however, because its elements are not continuous functions. 5.1. A Haar system on the transformation groupoid. Before we enter the construction of Haar systems, we mention an important trivial case: if G is étale, then so is L. Therefore, L certainly has a canonical Haar system if G is étale. This already covers many examples, and the reader only interested in étale groupoids may skip the construction of the Haar system on L. We define Haar systems as in [31, Section 1]. Thus our Haar system (λx on G is left invariant, so supp λx for all g ∈ G. The continuity requirement for (λx G)x∈G0 G = λr(g) G)x∈G0 is that the function λG(f ) G(g) is continuous on G0 for all f ∈ S(G). By the definition of S(G) (see Definition B.1), it suffices to check continuity if f is a continuous function with compact support on a Hausdorff open subset U of G. on G0 defined by λG(f )(x) :=RG f (g) dλx G = Gx = {g ∈ G1 r(g) = x} and g∗λs(g) G Proposition 5.1. The Haar system on G extends uniquely to a Haar system on the transformation groupoid L. Proof. Fix x ∈ G0 = L0. We are going to describe the measure λx Haar system. Since L = St∈S Lt is an open cover, the measure λx L on Lx in the L is determined by its restrictions to Lt for all t ∈ S. If x /∈ r(Lt), then there is nothing to do, so t := Lt ∩ Lx consider t ∈ S with x ∈ r(Lt), and fix g ∈ Lt with r(g) = x. If A ⊆ Lx is measurable, then A = g · (g−1 · A) with g−1 · A ⊆ L−1 · Lt = L1 = G. Since we want (λx L(A) = λs(g) G (g−1 · A) if g ∈ Lt satisfies r(g) = x and A ⊆ Lx t is measurable. Hence there is at most one Haar measure on L extending the given Haar measure on G. L) to be left invariant and to extend (λx G), we must have λx t · g2 ∈ L−1 1 G) with respect to G implies that λs(g) If g1, g2 ∈ Lt satisfy r(g1) = r(g2) = x, then g−1 t Lt = L1 = G; the left invariance of (λx G (g−1 · A) does not depend on the choice of g. If ∅ 6= A ⊆ Lx u, then we may pick the same element g ∈ A to define the measure of A as a subset of Lx u. Thus the definitions t for t ∈ S are compatible. Thus there is a unique measure λx of λx L on Lx with λx t is measurable and g ∈ Lt satisfies r(g) = x. If l ∈ L has s(l) = x, then l∗(λx) is a measure on Lr(l) with the same properties that characterise λr(l) uniquely; so we get the left invariance of our family of measures: l∗(λs(l)) = λr(l) for all l ∈ L. G (g−1 · A) whenever A ⊆ Lx L on the sets Lx L(A) = λs(g) t and of Lx t ∩ Lx Checking continuity by hand is unpleasant, so we use a different description of the same Haar system for this purpose. Recall that Lt is an equivalence between restrictions of G to open invariant subsets of G0. The proof that equivalent group- oids have Morita -- Rieffel equivalent groupoid C∗-algebras uses a family of measures on the equivalence bibundle in order to define the right inner product; this measure on Lt is exactly the one described above (see the proof of [31, Corollaire 5.4]), and its continuity is known, even in the non-Hausdorff case. Thus our family of mea- sures restricts to a continuous family on each Lt. Since the map L S(Lt) → S(L) in Proposition B.2 is surjective, the family of measures (λx L) is continuous. (cid:3) 5.2. Construction of the Fell bundle. We know now that L has a Haar system. So we get a ∗-algebra structure on S(L) as in [25,31]. Since the Haar measure on L extends the one on G, the map S(G) → S(L) induced by the open embedding G → L is a ∗-algebra isomorphism onto its image. The groupoid C∗-algebras of L and G are the completions of S(L) and S(G) for suitable C∗-norms. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 33 Lemma 5.2. The involution on S(L) maps S(Lt) onto S(L−1 convolution product maps S(Lt) × S(Lu) to S(Ltu). t ) = S(Lt∗). The Proof. The claim for the involution is trivial. The claim for the convolution product follows, of course, from Lt · Lu ⊆ Ltu, but requires some care in the non-Hausdorff case because the convolution product is not defined directly, see the proof of [25, Proposition 4.4]. If f1 ∈ S(U ), f2 ∈ S(V ) for Hausdorff open subsets U ⊆ Lt and V ⊆ Lu, and if U · V is also Hausdorff, then we directly get f1 ∗ f2 ∈ S(U · V ) with U · V ⊆ Ltu. If U · V is non-Hausdorff, a partition of unity is used to write f1 and f2 as finite sums of functions on smaller Hausdorff open subsets U ′ ⊆ U , V ′ ⊆ V for which U ′ · V ′ is Hausdorff. Since U ′ · V ′ ⊆ U · V ⊆ Ltu, we get S(Lt) ∗ S(Lu) ⊆ S(Ltu) as desired. (cid:3) 1 ∗ f2 ∈ S(G) and f1 ∗ f ∗ Lemma 5.2 gives S(G) ∗ S(Lt) ⊆ S(Lt) and S(Lt) ∗ S(G) ⊆ S(Lt), so S(Lt) is a S(G)-bimodule; it also implies f ∗ 2 ∈ S(G) for all f1, f2 ∈ S(Lt), which gives S(G)-valued left and right inner products on S(Lt). We also have f1∗f2 ∈ S(Ltu) for f1 ∈ S(Lt) and f2 ∈ S(Lu), and these multiplication maps are associative and "isometric" with respect to the S(G)-valued inner products. We put "isometric" in quotation marks because we have not yet talked about norms. Lemma 5.3. f ∗ ∗ f ∈ S(G) is positive in C∗(G) for each t ∈ S, f ∈ S(Lt), and the closed linear span of f ∗ 1 ∗ f2 for f1, f2 ∈ S(Lt) is dense in C∗(Gr(Lt)). Proof. We have already used in the proof of Proposition 5.1 that the Haar measure on L restricts to the usual family of measures on the partial equivalence space Lt. In that context, the positivity of such inner products is already proved in [25, 31] in order to show that S(Lt) may be completed to a Hilbert C∗(G)-bimodule. The proof that an equivalence induces a Morita -- Rieffel equivalence also shows that the inner product defined above is full, that is, the closed linear span of f ∗ 1 ∗ f2 for f1, f2 ∈ S(Lt) is dense in C∗(Gr(Lt)). (cid:3) Hence we may complete S(Lt) to a Hilbert bimodule C∗(Lt) over C∗(G). The densely defined convolution map S(Lt) × S(Lu) → S(Ltu) extends to a Hilbert bimodule map µt,u : C∗(Lt) ⊗C∗(G) C∗(Lu) → C∗(Ltu) because it is isometric for the S(G)-valued inner products. Since C∗(Lt) is full as a Hilbert bimodule over C∗(Gr(Lt)) and C∗(Gs(Lt)), it follows that the maps µt,u above are surjective. The associativity of the multiplication on the dense subspaces S(Lt) extends to C∗(Lt). Thus we have constructed an action of S by Hilbert bimodules on C∗(G). By Theorem 4.8, this is equivalent to a saturated Fell bundle C∗(Lt)t∈S over S. Theorem 5.4. The section C∗-algebra C∗(S, C∗(Lt)t∈S) is naturally isomorphic to the groupoid C∗-algebra C∗(L). This theorem looks almost trivial from our construction; but the proof requires a technical result about S(L) to be proved in Appendix B. Before we turn to that, we first add coefficients in a Fell bundle over L. The above construction still works in almost literally the same way if we replace S(Lt) by S(Lt, B) everywhere, where B is a Fell bundle over the groupoid L. Unfortunately, we could not find a reference for the generalisation of Lemma 5.3 to this context. The references on groupoid crossed products we could find consider either Fell bundles over Hausdorff groupoids (such as [24]) or a more restrictive class of actions for non-Hausdorff groupoids (such as [25,31]), but not both. In particular, the positivity of the inner product on S(Lt, B) for a partial equivalence Lt is only proved in some cases: for arbitrary upper semicontinuous Fell bundles over 34 ALCIDES BUSS AND RALF MEYER Hausdorff groupoids in [24]; for Green twisted actions of non-Hausdorff groupoids on continuous fields of C∗-algebras over G0 in [31]; and for untwisted actions by automorphisms of non-Hausdorff groupoids on C0(G0)-algebras in [25]. This is probably only a technical issue that will be resolved eventually, but not in this paper. So we add an assumption about it in our next theorem. Theorem 5.5. Let B be a Fell bundle over L. Assume that f ∗ ∗ f ∈ S(G, B) is positive in C∗(G, B) for all f ∈ S(Lt, B), t ∈ S, and that the linear span of these inner products is dense in C∗(Gs(Lt), B). Then there is a Fell bundle C∗(Lt, B)t∈S over S that has the section C∗-algebra of the restriction C∗(G, BG) as unit fibre. The section C∗-algebra C∗(S, C∗(Lt, B)t∈S) is naturally isomorphic to the section C∗-algebra of the groupoid Fell bundle C∗(L, B). Theorem 5.4 is a special case of Theorem 5.5 for the constant Fell bundle C. It remains to prove Theorem 5.5. This will be done in Appendix B.1, after some preliminary results about Banach bundles in Appendix B. Corollary 5.6. Let L be an étale topological groupoid with Hausdorff locally com- pact object space and with a Haar system. Let S be a wide inverse subsemigroup = t1 ∩ t2 for all t1, t2 ∈ S. Then of Bis(L), that is, St∈S t = L and St∈S,t⊆t1∩t2 the groupoid C∗-algebra of L is isomorphic to the crossed product C0(L0) ⋊ S. More generally, if B is a Fell bundle over L, then the section C∗-algebra C∗(L, B) is isomorphic to the section C∗-algebra of the associated Fell bundle over S. Proof. The assumptions on S ensure that L is an S-graded groupoid by Lt := t with unit fibre G = L0. So Theorem 5.4 gives the first assertion, and Theorem 5.5 gives the second one. In this case, positivity is not an issue because we are dealing with a space G, so positivity in C∗(G, B) is equivalent to pointwise positivity in all x ∈ L0 = G0. The value (f ∗ ∗ f )(x) for f ∈ S(Lt, B) is either zero or f (l)∗f (l) for the unique l ∈ Lt with s(l) = x. This is assumed to be positive in the definition of a Fell bundle over a groupoid. (cid:3) The isomorphism C∗(L) ∼= C0(L0) ⋊ S is already proved in [12, Theorem 9.8] (if L0 is second countable and S is countable). The more general result for (sepa- rable) Fell bundles over (second countable) étale groupoids is proved in [4, Theo- rem 2.13]. Another special case worth mentioning are group extensions. Let G ֌ H ։ S be an extension of locally compact groups with discrete S. This gives an action of S, viewed as an inverse semigroup, on G by Theorem 3.15. We get a Fell bundle over S with unit fibre C∗(G) and section C∗-algebra C∗(H). More generally, we get a similar result for a Fell bundle over H (compare with [5, Example 3.9]). Our Fell bundle also comes from a Green twisted action of (H, G), and in this formulation, our theorem is well-known in this case (see [7, 14]). Corollary 5.7. In the situation of Theorem 5.5, the canonical map from C∗(G, B) to C∗(L, B) is injective. Proof. The unit fibre of the Fell bundle in Theorem 5.5 is C∗(G, B) and the section C∗-algebra is C∗(L, B). The unit fibre always embeds into the section C∗-algebras of a Fell bundle over an inverse semigroup, see [13, Corollary 8.10]. (cid:3) Next, we note a useful variant of Theorem 5.4 for group-valued cocycles. Let L be a locally quasi-compact, locally Hausdorff groupoid, let S be a group, and let c: L → S be a 1-cocycle. Let Lt := c−1(t) ⊆ L for t ∈ S, and let G = L1 = c−1(1). Since we do not assume anything about c, this need not be an S-grading (compare Theorem 3.15). Nevertheless, we may complete S(Lt) to a INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 35 Hilbert bimodule over C∗(G) and thus get a Fell bundle over S. The difference to the situation above is that this Fell bundle need not be saturated any more. Theorem 5.8. The section C∗-algebra of the Fell bundle over S with unit fi- bre C∗(G) just described is isomorphic to C∗(L). Hence the canonical map C∗(G) → C∗(L) is faithful. Proof. The proofs of Theorem 5.4 and Corollary 5.7 still work for non-saturated Fell bundles (even over inverse semigroups). Alternatively, we may replace our non- saturated Fell bundle over G by a saturated Fell bundle over an inverse semigroup associated to G, just as for partial actions (see [11]). This does not change the section C∗-algebra, and afterwards Theorem 5.4 applies literally. (cid:3) 5.3. Another construction of the Fell bundle. The construction of the Fell bundle over S in Section 5.2 used the transformation groupoid. Now we construct this Fell bundle using the abstract functorial properties of actions on groupoids and their corresponding actions on C∗-algebras. Actually, some aspects of this have been used to prove Lemma 5.3 above. It is well-known that two equivalent groupoids have Morita -- Rieffel equivalent C∗-algebras (see [23]), even in the non-Hausdorff case (see [31]). The proof is constructive: given an equivalence X from H to G, the space S(X) is completed to a C∗(G)-C∗(H)-imprimitivity bimodule, using certain natural formulas for a S(G)-S(H)-bimodule structure and S(G)- and S(H)-valued inner products. An important ingredient here is that the Haar measures on G and H give canonical families of measures on the fibres of the range and source maps of X, which may be used to integrate functions on X. Even if X is only a partial equivalence, the same formulas still work and give a Hilbert bimodule C∗(X) from C∗(H) to C∗(G) by completing S(X). If f : X → X ′ is an isomorphism between two partial equivalences, then f∗ : S(X) → S(X ′) defined by f∗(h) = h ◦ f −1 is an isomorphism that preserves all structure, so it extends to an isomorphism C∗(X) ∼−→ C∗(X ′). Theorem 5.9. The maps G 7→ C∗(G) from groupoids to C∗-algebras, X 7→ C∗(X) from partial equivalences to Hilbert bimodules, and f 7→ f∗ from bibundle isomor- phisms to Hilbert bimodule isomorphisms are part of a functor from the bicategory of partial groupoid equivalences to the bicategory of C∗-algebras and Hilbert bimodules. Proof. The above map is strictly compatible with unit arrows: the unit arrow G1 on G is sent to C∗(G1) = C∗(G), and the unit transformations in both bicategories are also preserved. To complete the above data to a functor of bicategories, it remains to give natural isomorphisms C∗(X) ⊗C∗(H) C∗(Y ) ∼= C∗(X ×H Y ) and check that they satisfy the expected associativity condition for three composable partial equivalences. They are constructed by writing down the "convolution map" S(X) ⊙ S(Y ) → S(X ×H Y ) given by the formula (17) ξ(x · h)η(h−1 · y) dλu(h), (ξ · η)(x, y) :=ZH1 for all ξ ∈ S(X), η ∈ S(Y ) and (x, y) ∈ X ×H Y , where u = s(x) = r(y). It is routine to check that the map (17) has dense range and is a bimodule map and an isometry for both inner products; thus it extends to an isomorphism between the completions: C∗(X) ⊗C∗(H) C∗(Y ) ∼= C∗(X ×H Y ). One way to construct the convolution maps and check their properties is like our construction above using the transformation groupoid: build an appropriate linking groupoid containing all the data. For two composable equivalences Y and X from K to H and from H to H, this linking groupoid has object space G0⊔H 0⊔K 0; its arrow 36 ALCIDES BUSS AND RALF MEYER space is a disjoint union of G1, H 1, K 1, X, Y , X ∗, Y ∗, X ×H Y , and Y ∗ ×H X ∗, the source and range maps are the obvious ones, and the multiplication is defined using the left and right actions of G, H and K and canonical maps. This is indeed a topological groupoid, and it inherits a canonical Haar system if G, H and K have Haar systems. The convolution map is the restriction of the convolution in this larger groupoid to X ×H Y . Given three composable partial equivalences, there is a similar linking groupoid combining all the relevant data, and the associativity of its convolution product on X ×H Y ×K Z gives the associativity coherence of the isomorphisms C∗(X) ⊗C∗(H) C∗(Y ) ∼= C∗(X ×H Y ). (cid:3) Remark 5.10. The above theorem is extended in the thesis of Rohit Holkar [15], where a similar functor from a bicategory of groupoid correspondences to the bi- category of C∗-correspondences is constructed. This construction is more difficult because the family of measures needed to write down the right inner product is no longer canonical and becomes part of the data. Hence the behaviour of the measures under composition has to be studied as well. An inverse semigroup action by partial equivalences may be defined as a functor (of bicategories) from the inverse semigroup to the bicategory of groupoids and partial equivalences. Composing it with the functor in the theorem gives a functor from the inverse semigroup to the bicategory of Hilbert bimodules, which is the same thing as an action by Hilbert bimodules. This is the same as a saturated Fell bundle over the inverse semigroup by Theorem 4.8. This is the second construction of the Fell bundle over S. It gives an isomorphic Fell bundle because the Haar measure on L used above is the same as the combination of the measure families on the partial equivalences Lt that are used to define the convolution maps in Theorem 5.9. More concretely, an action (Xt, µt,u) of S on G yields the action on C∗(G) given by the Hilbert bimodules C∗(Xt) with the multiplication maps C∗(Xt) ⊗A C∗(Xu) ∼−→ C∗(Xt ×G Xu) C∗(µt,u) −−−−−→ ∼= C∗(Xtu), which involve the convolution isomorphisms C∗(Xt) ⊗A C∗(Xu) ∼−→ C∗(Xt ×G Xu). This is associative by the associativity coherence of these convolution isomorphisms. 6. Actions of inverse semigroups and groupoids Let H be an étale groupoid with locally compact Hausdorff object space. So far, we have constructed actions of the inverse semigroup Bis(H) on certain C∗-algebras. Instead, we would like to construct actions of H itself. In this section, we are going to see that both kinds of actions are very closely related. Here an action of Bis(H) is as above: an action by Hilbert bimodules or, equivalently, a saturated Fell bundle over Bis(H). The corresponding "actions" of H are saturated Fell bundles over H. First we explain how to turn a Fell bundle over H into one over Bis(H). So let B = (Bh)h∈H be a Fell bundle over H (see [4, 17]). Let A := C0(H 0, B) be the C∗-algebra of C0-sections of B over H 0; this is a C0(H 0)-C∗-algebra by construction. If t ∈ Bis(H), then the Fell bundle operations turn Ht := C0(t, B) into a Hilbert C0(r(t), B)-C0(s(t), B)-bimodule. The multiplication in the Fell bundle induces multiplication maps µt,u : Ht ⊗A Hu → Htu. This gives an action of Bis(H) on A by Hilbert bimodules. Not every action of Bis(H) by Hilbert bimodules is of this form. The obstruction lies in how idempotents in Bis(H) act. Idempotents in Bis(H) are the same as open subsets of H 0. We identify the idempotent semilattice E(Bis(H)) with the complete lattice O(H 0) of open subsets of H 0. So the action of idempotents in Bis(H) becomes a map from O(H 0) to the complete lattice I(A) of ideals in A. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 37 Theorem 6.1. An action (Ht, µt,u)t∈Bis(H) of the inverse semigroup Bis(H) on a C∗-algebra A by Hilbert A-bimodules comes from a Fell bundle over H if and only if the map from E(Bis(H)) ∼= O(H 0) to I(A) commutes with suprema. This Fell bundle over H is unique up to isomorphism, and the Fell bundles over Bis(H) and H have the same section C∗-algebras. Proof. A map O(H 0) → I(A) comes from a continuous map Prim(A) → H 0 if and only if it commutes with finite infima and arbitrary suprema by [20, Lemma 2.25]; here we need H 0 to be a sober space, a very mild condition that certainly allows all locally Hausdorff spaces. Compatibility with finite infima says that it is a mor- phism of semilattices, which we assume anyway; compatibility with suprema is an extra condition. A continuous map Prim(A) → H 0 is equivalent to an isomor- phism between A and the C∗-algebra of C0-sections of an upper semicontinuous field (Ax)x∈H0 of C∗-algebras over H 0 (see [26]). Thus the criterion in the theorem is necessary and sufficient for A to come from such an upper semicontinuous field. This gives a Fell bundle over H 0 ⊆ H. It remains to extend this to all of H. Let t ∈ Bis(H). Then Ht is a Hilbert A-bimodule. For h ∈ t ⊆ H 1, we define Hh,t := Ht ⊗A As(h); this is a Hilbert As(h)-module. If ξ ∈ Ht, then kξk2 = khξ, ξik, and for hξ, ξi ∈ A, the norm is the supremum of the norms of its images in Ax for all x ∈ H 0. Therefore, the canonical map from Ht to Qh∈t Hh,t is isometric. Thus we view Ht as a space of sections of the bundle of Banach spaces Hh,t over t. This is an upper semicontinuous bundle on t because (Ax)x∈H0 is and the norm on Ht is given by kξk2 = khξ, ξik with hξ, ξi ∈ A. If t, u ∈ Bis(H) and h ∈ t ∩ u, then both Hh,t and Hh,u are candidates for the fibre Hh of our Fell bundle at h. These are isomorphic through the canonical isomorphisms jt,t∩u : Ht∩u → Hts(t∩u) and ju,t∩u : Ht∩u → Hus(t∩u) from Theo- rem 4.8. For each h ∈ H 1, choose some th ∈ Bis(H) with h ∈ th and define Hh := Hh,th. If t ∈ Bis(H), then there are canonical isomorphisms Hh ∼= Hh,t for all h ∈ t. We use them to transport the topology on the bundle (Hh,t)h∈t to the bundle (Hh)h∈t. These topologies are compatible on t ∩ u for all t, u ∈ Bis(H). Since the subsets t ∈ Bis(H) form an open cover of H 1, there is a topology on the whole bundle (Hh)h∈H1 that coincides with the topology on (Hh)h∈t described above for each t ∈ Bis(H). In particular, the space of C0-sections of (Hh)h∈H1 on t coincides naturally with Ht. Let A(U ) for U ∈ O(G) be the ideal of C0-sections of (Ax) vanishing outside U . Then A(U ) = HU if we view U ∈ E(Bis(G)). We have (18) Ht ⊗A A(U ) = Ht·U = Ht(U)·t = A(t(U )) ⊗A Ht for all t ∈ Bis(H), U ∈ O(H 0) with U ⊆ s(t). Here we view each t ∈ Bis(H) as a partial homeomorphism s(t) → r(t) and write t(U ) for the image of U under this map. This is exactly how Bis(H) acts on H 0. Equation (18) implies that ∼= Ar(h) ⊗A Ht. Thus Hh is a Hilbert Ar(h)-As(h)-bimodule. The isomorphism Hh,t Ht ⊗A Hu → Htu is A-linear and hence C0(H 0)-linear. Thus it restricts to an isomorphism on the fibres, Hg,t ⊗A Hh,u → Hgh,tu for all g ∈ t, h ∈ u with s(g) = r(h). The compatibility of the multiplication with the inclusion maps from Theorem 4.8 shows that these maps on the fibres do not depend on the choice of t and u with h ∈ t and h ∈ u. Thus we get well-defined isomorphisms Hg ⊗As(g) Hh → Hgh for all g, h ∈ H 1 with s(g) = r(h). Since they can be put together to maps Ht ⊗A Hu → Htu for all t, u ∈ Bis(H) and since Bis(H) covers H 1, they are locally ∼= Ht∗ must come continuous, hence continuous. Similarly, the isomorphisms H∗ t from well-defined, continuous maps H∗ h → Hh−1 for h ∈ H 1 by restricting them to fibres. The remaining algebraic conditions needed for a Fell bundle over the groupoid H 1 all follow easily because (Ht, µt,u) gives a Fell bundle over Bis(H). 38 ALCIDES BUSS AND RALF MEYER If we turn the Fell bundle over H constructed above into a Fell bundle over Bis(H) again, we clearly get back the original Fell bundle over Bis(H) because Ht is the space of C0-sections of (Hh)h∈t. Conversely, if we start with a Fell bundle over H, turn it into a Fell bundle over Bis(H), and then use the above construction to go back, we get an isomorphic Fell bundle over H. Hence we get a bijection between isomorphism classes of the two types of Fell bundles. Theorem 5.5 shows that the passage from Fell bundles over H to Fell bundles over Bis(H) does not change the section C∗-algebras. (cid:3) We assumed G0 to be Hausdorff and locally compact so far because Fell bundles over groupoids have not yet been defined in greater generality. We suggest to use the necessary and sufficient criterion in Theorem 6.1 as a definition: Definition 6.2. Let G be an étale topological groupoid for which G0 (and hence G1) is sober. An action of G on a C∗-algebra A is an action of Bis(G) by Hilbert bimod- ules for which the resulting map O(G0) → I(A) commutes with arbitrary suprema. Sobriety of G0 is needed to turn a map O(G0) → I(A) that commutes with suprema into a continuous map Prim(A) → G0 (see [20, Lemma 2.25]). Let G be a sober space G0 viewed as a groupoid. Then an action of G is the same as a continuous map Prim(A) → G0. In the notation of [20], this turns A into a C∗-algebra over G0. It is unclear what the "fibres" of such a C∗-algebra over G0 should be if G0 is badly non-Hausdorff. Therefore, it is not clear how to describe actions of étale sober groupoids in the sense of Definition 6.2 as Fell bundles over G. If G0 is locally Hausdorff and locally quasi-compact, then Definition 6.2 seems to work quite well; we plan to discuss this in greater detail elsewhere. The criterion in Theorem 6.1 also suggests how to define actions of étale group- oids on other groupoids: Definition 6.3. Let G be an étale topological groupoid for which G0 (and hence G1) is sober, and let H be an arbitrary topological groupoid. An action of G on H is an action of Bis(G) on H by partial equivalences for which the map O(G0) → O(H 0/H) that describes the action of E(Bis(G)) commutes with arbitrary suprema. The extra assumption in Definition 6.3 and [20, Lemma 2.25] ensure that the map O(G0) → O(H 0/H) for an action of G on H comes from a continuous map H 0/H → G0 or, equivalently, an H-invariant continuous map H 0 → G0. Proposition 6.4. Let H be a locally quasi-compact, locally Hausdorff groupoid with Hausdorff object space and with a Haar system. An action of G on H induces an action of G on C∗(H) as well. Proof. In Section 5.2, we turn an action of Bis(G) on H into an action of Bis(G) on C∗(H). For any open H-invariant subset U of H 0, the closure of S(HU ) in C∗(H) is an ideal C∗(HU ) in C∗(H). The map O(H 0/H) → I(C∗(H)), U 7→ C∗(HU ), com- mutes with suprema. Hence Theorem 6.1 applies to the action of Bis(G) on C∗(H) if the action of Bis(G) satisfies the condition in Definition 6.3. (cid:3) 6.1. The motivating example. Now we consider our motivating example: an action of a locally Hausdorff, locally quasi-compact, étale groupoid H on a locally Hausdorff, locally quasi-compact space Z. Let U be a Hausdorff open covering of Z and let GU be the associated covering groupoid, which is étale, locally compact and Hausdorff. Its C∗-algebra C∗(GU) is our noncommutative model for the non- Hausdorff space Z. We want to construct an "action" of H on it that models the given action of H on Z. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 39 To construct it, we use the inverse semigroup S := Bis(H) of bisections of H. First we turn the action of H on Z into an action of S on Z by partial home- omorphisms in the usual way: a bisection t ∈ S acts by the homeomorphism r−1(s(t)) → r−1(r(t)), z 7→ gr(z)·z, where gx is the unique arrow in t with s(gx) = x. We have seen in Corollary 3.21 that the S-action on Z induces an S-action on GU by partial equivalences. The transformation groupoid GU ⋊ S for this action is Hausdorff, étale and locally compact. It is equivalent to Z ⋊ S by Corollary 3.21. Let p: X := FU∈U U → Z be the canonical map. Then GU = p∗(Z). An idempotent U ∈ O(H 0) in Bis(H) acts on GU by the identity map on the open invariant subgroupoid GU(r◦p)−1(U). That is, O(H 0) acts on GU through the map U/GU), U 7→ (r ◦ p)−1(U ); this commutes with suprema and infima. O(H 0) → O(G0 Thus our action of Bis(H) on GU is also an action of H in the sense of Definition 6.3. We may identify Z ⋊ S ∼= Z ⋊ H using the obvious S-grading on Z ⋊ H and Theorem 3.14, so GU ⋊ S is equivalent to Z ⋊ H. The S-action on GU induces a Fell bundle over S with unit fibre C∗(GU), which we view as an action of S on C∗(GU). Theorem 5.4 gives an isomorphism between its section C∗-algebra C∗(GU) ⋊ S and the groupoid C∗-algebra C∗(GU ⋊ S). We may turn our Fell bundle over Bis(H) into a Fell bundle over the groupoid H by Proposition 6.4. Theorem 6.1 also says that the section C∗-algebra of the Fell bundle over H is isomorphic to C∗(GU) ⋊S ∼= C∗(GU ⋊S). The restriction to the unit fibre is C∗(GU) by construction. We are going to describe this Fell bundle over H. We have GU ⋊ S ∼= p∗(Z ⋊ H), that is, the object space of GU ⋊ S is X and the arrow space is homeomorphic to the space of triples (x1, h, x2), x1, x2 ∈ X, h ∈ H 1 with r(p(x1)) = r(h) and r(p(x2)) = s(h) in H 0. Here (x1, h, x2) is an arrow from x2 to x1, and the multiplication is (x1, h1, x2)·(x2, h2, x3) = (x1, h1h2, x3). For h ∈ H 1, let Kh be the subspace of triples (x1, h, x2) for x1, x2 ∈ X, r(p(x1)) = r(h) and r(p(x2)) = s(h). Since p and H are étale, this is a discrete set. The fibre at h of our Fell bundle over H is the completion of the space Cc(Kh) of finitely supported functions on Kh to a Hilbert bimodule over C∗(K1r(h) ) and C∗(K1s(h)). Proposition 6.5. Let Z be a basic action of H with Hausdorff quotient space H\Z, for instance, Z = H 1 with the action by left or right multiplication and quotient space H 0. Then the groupoid GU ⋊ S is equivalent to H\Z and C∗(GU) ⋊ S is Morita equivalent to C0(H\Z). Proof. The groupoid GU ⋊ S is equivalent to Z ⋊ S. This is the same as Z ⋊ H by Theorem 3.14, using the evident S-grading on Z ⋊ H. Since the H-action on Z is basic, Z ⋊ H is equivalent to H\Z. This space is assumed to be Hausdorff, and GU ⋊ S is also a groupoid with Hausdorff object space. So the equivalence between them is of the usual type, involving free and proper actions, by Proposition A.7. Hence it induces a Morita -- Rieffel equivalence from C0(H\Z) to C∗(GU) ⋊ S. (cid:3) In the situation of Proposition 6.5, GU ⋊ S has Hausdorff arrow space because it must be isomorphic to the covering groupoid of the open surjection G0 U → (GU ⋊ ∼= H\Z between two Hausdorff spaces. In this case, it is also easy to see that S)\G0 U any Fell bundle over the groupoid GU ⋊ S is a pull-back of a Fell bundle over H\Z, which is the same as a C0(H\Z)-C∗-algebra B. The section C∗-algebra of the Fell bundle over GU ⋊ S is Morita -- Rieffel equivalent to this C0(H\Z)-C∗-algebra B. By Theorem 5.5, this is also the section C∗-algebra of the Fell bundle over S associated to B. Many properties like properness, amenability, essential principality are shared by an action of a groupoid on a space and its transformation groupoid. This suggests how to extend these notions to inverse semigroup actions on groupoids. We take 40 ALCIDES BUSS AND RALF MEYER this as a definition for proper actions of inverse semigroups on locally compact groupoids: Definition 6.6. An action of an inverse semigroup S on a topological groupoid G is proper if the groupoid G ⋊ S is proper, that is, the following map is proper (that is, stably closed): (s, r): (G ⋊ S)1 → G0 × G0, g 7→ (s(g), r(g)). The action is called free if this map is injective. Let L be a proper groupoid such that L0 is a locally compact Hausdorff space. Then the image of L1 in L0 × L0 is locally compact and Hausdorff because it is a closed subspace of a locally compact Hausdorff space. Since this subspace is closed and the orbit space projection L0 → L\L0 is open, it also follows that L\L0 is locally compact Hausdorff (see Proposition A.3). The groupoid L itself need not be Hausdorff: the non-Hausdorff group bundle in Section 8 is proper in this sense because it is quasi-compact and the image of (s, r) is closed. If L acts freely and properly on a Hausdorff space L0, however, then L1 must be Hausdorff. In this case, we also get information about any open subgroupoid, which leads to the following proposition: Proposition 6.7. Let S act properly and freely on a locally Hausdorff, locally quasi-compact groupoid G. Then G is a basic groupoid, so that G is equivalent to the locally Hausdorff, locally quasi-compact space G\G0. Proof. The map in Definition 6.6 is a homeomorphism onto its image because it is continuous, injective, and closed. Hence its restriction to the open subspace G1 ⊆ (G ⋊ S)1 is still a homeomorphism onto its image. This means that G is a basic groupoid, so G is equivalent to G0/G. This is locally Hausdorff and locally quasi-compact by Proposition A.14. (cid:3) Thus the free and proper actions of S all come from actions on locally Hausdorff spaces that are desingularised by replacing the space by a Hausdorff groupoid G. 6.2. Inverse semigroup models for étale groupoids. Let G be an étale group- oid. So far, we have described actions of G through actions of the inverse semi- group Bis(G). Since Bis(G) is usually quite big, even uncountable, we now replace it by smaller inverse semigroups. The following definition describes which inverse semigroups we allow as "models" for G: Definition 6.8. An inverse semigroup model for an étale groupoid G consists of an inverse semigroup S, an S-action on the space G0 by partial homeomorphisms, and an isomorphism G0 ⋊ S ∼= G of étale groupoids that is the identity on objects. In particular, if S ⊆ Bis(G) is a wide inverse subsemigroup, then S with its usual action on G0 and the canonical isomorphism G0 ⋊ S ∼= G from Corollary 3.19 is a model for G. Lemma 6.9. An inverse semigroup model for G is equivalent to an inverse semi- group S with a homomorphism ϕ: S → Bis(G) that induces an isomorphism G0 ⋊ S → G0 ⋊ Bis(G) ∼= G, where we use the canonical action of Bis(G) on G0 and ϕ to let S act on G0. Proof. Let S act on G0. There is a canonical homomorphism S → Bis(G0 ⋊ S), see [12]. Combined with an isomorphism G0 ⋊ S ∼= G, we get a homomorphism ϕ: S → Bis(G). Conversely, such a homomorphism induces an action of S on G0 and then a continuous groupoid homomorphism G0 ⋊ S → G0 ⋊ Bis(G) ∼= G. Routine computations show that these two constructions are inverse to each other. (cid:3) INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 41 The following lemma characterises inverse semigroup models more concretely when we take S = Bis(G). Lemma 6.10. Let S and S be inverse semigroups, let ϕ: S → S be a homo- morphism, and let S act on Z by partial homeomorphisms. The induced groupoid homomorphism ϕ : Z ⋊ S → Z ⋊ S is an isomorphism if and only if (1) for all t1, t2 ∈ S and every z ∈ Z with z ∈ Dt∗ 2 t2 and every f ∈ E( S) with z ∈ Df and ϕ(t1)f = ϕ(t2)f , there is e ∈ E(S) with z ∈ De and t1e = t2e; (2) for every u ∈ S and every z ∈ Z with z ∈ Du∗u, there is t ∈ S with z ∈ Dt∗t 1 t1 ∩ Dt∗ and there is f ∈ E( S) with z ∈ Df and uf = ϕ(t)f . In this case, we call ϕ a Z-isomorphism. Proof. The groupoid homomorphism ϕ is the identity on objects and always contin- uous and open on arrows, so the only issue is whether ϕ is bijective on arrows. It is routine to check that (1) is equivalent to injectivity and (2) to surjectivity of ϕ. (cid:3) Let S and ϕ: S → Bis(G) be an inverse semigroup model for an étale topolog- ical groupoid G. Which actions of S on groupoids by partial equivalences or on C∗-algebras by Hilbert bimodules come from actions of G? First we consider a trivial special case to see why we need more data. Let G be just a topological space, viewed as a groupoid. In this case, the trivial inverse semi- group {1} is an inverse semigroup model. An action of S contains no information. An action of G on a topological groupoid H or a C∗-algebra is simply a continuous map ψ : H 0/H → G0 or ψ : Prim(A) → G0, respectively. Theorem 6.11. Let G be a sober étale topological groupoid and let S and ϕ: S → Bis(G) be an inverse semigroup model for G. Let H be a topological groupoid. An action of G on H by partial equivalences is equivalent to a pair consisting of an action of S on H by partial equivalences and an S-equivariant map ψ : H 0/H → G0. The transformation groupoid for an action of G (that is, Bis(G)) and its restriction to S are the same. The S-equivariance of ψ refers to the actions of S on H 0/H and G0 by partial homeomorphisms induced by the action on H and by ϕ. Proof. First let G act on H; more precisely, Bis(G) acts on H and the resulting map O(G0) = E(Bis(G)) → O(H 0/H) commutes with suprema (Definition 6.3). [20, Lemma 2.25] shows that it comes from a continuous map ψ : H 0/H → G0. This map is Bis(G)-equivariant and hence S-equivariant. Now let S act on H and let ψ : H 0/H → G0 be an S-equivariant map. Let L := H ⋊ S with its canonical S-grading (Lt)t∈S. We claim that there is a unique Bis(G)-grading ( ¯Lt)t∈Bis(G) on L with ¯Lϕ(t) = Lt for all t ∈ S, and ¯LU = H 1 ψ−1(U) for U ∈ O(G0). These two conditions on the Bis(G)-grading say exactly that it corresponds to the given S-action and map ψ. So the proof of the claim will finish the proof of the theorem. For t ∈ Bis(G) and u ∈ S, we may form t ∩ ϕ(u) ∈ Bis(G). We have t ∩ ϕ(u) = t · Vt,u = ϕ(u) · Vt,u for Vt,u = s(t ∩ ϕ(u)) ∈ O(G0); here we also view Vt,u as an idempotent element of Bis(G). Since S models G, we have t = Su∈S t ∩ ϕ(u) and hence s(t) = Su∈S Vt,u. Any Bis(G)-grading with ψ−1(V ) for all V ∈ O(G0) satisfies ¯LV = H 1 ¯Ltψ−1(Vt,u) = ¯Lt · ¯LVt,u = ¯Lt∩ϕ(u) = ¯Lϕ(u)ψ−1(Vt,u) 42 ALCIDES BUSS AND RALF MEYER for all t ∈ Bis(G), u ∈ S. Since s(t) =Su∈S Vt,u and ψ is S-equivariant, this shows that there is at most one Bis(G)-grading with the required properties, namely, ¯Lt = [u∈S Luψ−1(Vt,u). More explicitly, l ∈ ¯Lt if and only if l ∈ Lu for some u ∈ S for which t and ϕ(u) have the same germ at ψ(s(l)). We must prove that ( ¯Lt)t∈Bis(G) is a grading with all desired properties. First we check ¯Lϕ(u) = Lu for u ∈ S. The inclusion ⊇ is trivial. If l ∈ ¯Lϕ(u), then l ∈ Lu′ for some u′ ∈ S for which ϕ(u) and ϕ(u′) have the same germ at ψ(s(l)) ∈ G0. Hence there is an idempotent element e ∈ S with ψ(s(l)) ∈ ϕ(e) and ue = u′e. Since Le = H 1 ψ−1(e), we get l ∈ Lu′Le = Lu′e = Lue = LuLe ⊆ Lu. This finishes the proof that ¯Lϕ(u) = Lu for all u ∈ S. Next we check ¯LW = H 1 ψ−1(W ) for W ∈ O(G0). The inclusion ⊇ holds because VW,1 = W . Conversely, let l ∈ ¯LW . Then l ∈ Lu for some u ∈ S for which ϕ(u) and IdW have the same germ at ψ(s(l)). Since G0 ⋊ S ∼= G, there is an idempotent e ∈ S with ψ(s(l)) ∈ ϕ(e) and ue = e. An argument as in the previous paragraph shows that l ∈ LuLe = Le ⊆ H 1. Thus ¯LW = H 1 ψ−1(W ) for all W ∈ O(G0). If t ∈ Bis(G), u ∈ S, then (ϕ(u) ∩ t)∗ = ϕ(u∗) ∩ t∗. Hence Vt∗,u∗ = t(Vt,u) = ϕ(u)(Vt,u). This implies Lt∗ = L−1 t for all t ∈ Bis(G). Let t1, t2 ∈ Bis(G). We claim that ¯Lt1 · ¯Lt1 = ¯Lt1t2. The inclusion ⊆ follows because (ϕ(u1) ∩ t1) · (ϕ(u2) ∩ t2) ⊆ ϕ(u1u2) ∩ t1t2. For the converse inclusion, take l ∈ ¯Lt1t2. Then t ∈ Lu for some u ∈ S for which t1t2 and ϕ(u) have the same germ at ψ(s(l)). Factor this germ as g1g2 with gj ∈ tj for j = 1, 2. There are uj ∈ S with gj ∈ ϕ(uj) for j = 1, 2 because G ∼= G0 ⋊ S. Thus ϕ(u1)ϕ(u2) = ϕ(u1u2) and t1t2 have the same germ g1g2 at ψ(s(l)). Then u1u2 and u also have the same germ there, and an argument as above shows that l ∈ Lu1u2 as well. Using (Gr1) for the S-grading, we get lj ∈ Luj for j = 1, 2 with l = l1l2. Then s(l2) = s(l) and , so that l1 ∈ ¯Lt1. Hence r(l1) = r(l). This allows to prove l2 ∈ ¯Lt2 and l−1 the Bis(G)-grading satisfies (Gr1). It is clear that ¯Lt1 ⊆ ¯Lt2 if t1 ≤ t2 in Bis(G), so ¯Lt1 ∩ ¯Lt2 ⊇Sv≤t1,t2 ¯Lv = ¯Lt1∩t2 for all t1, t2 ∈ Bis(G). For the converse inclusion, take l ∈ ¯Lt1 ∩ ¯Lt2. Then there are u1, u2 ∈ S with l ∈ Lu1 ∩ Lu2, such that tj and ϕ(uj) have the same germ at ψ(s(l)) for j = 1, 2. (Gr3) for the S-grading gives v ∈ S with v ≤ u1, u2 and l ∈ Lv. Since ψ is S-equivariant, ψ(s(l)) belongs to the domain of ϕ(v), so the germs of ϕ(v) and ϕ(ui) at ψ(s(l)) are equal. Then the germs of t1 and t2 at ψ(s(l)) are equal as well, that is, t1 ∩ t2 is defined at ψ(s(l)) and has the same germ there as ϕ(v). This means that l ∈ ¯Lt1∩t2. This verifies (Gr3) for the Bis(G)-grading. Since ¯Lϕ(u) = Lu for all u ∈ S andSu∈S Lu = L1, we also getSt∈Bis(G) ¯Lt = L1, (cid:3) which is (Gr4). 1 ∈ ¯Lt∗ 1 The following lemma is needed to formulate a similar result for actions on C∗-algebras: Lemma 6.12. An action of S on a C∗-algebra A by Hilbert bimodules induces an action of S on Prim(A) by partial homeomorphisms. Proof. The Rieffel Correspondence (see [29, Corollary 3.33]) says that an imprim- itivity bimodule H from B to A induces a homeomorphism Prim(B) ∼−→ Prim(A). The corresponding lattice isomorphism I(B) = O(Prim(B)) ∼−→ O(Prim(A)) = I(A) INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 43 sends an ideal J ⊆ B to the unique ideal I ⊆ A with I · H = H · J. A Hilbert A, B-bimodule induces a partial homeomorphism Prim(B) → Prim(A) because it is an imprimitivity bimodule between certain ideals in A and B, which correspond to open subsets of the primitive ideal spaces. Isomorphic Hilbert bimodules induce the same partial homeomorphism, of course. The partial homeomorphism associated to a tensor product bimodule H1 ⊗B H2 is the composite of the partial homeo- morphisms associated to H1 and H2. Thus the map from S to pHomeo(Prim(A)) induced by an action on A by Hilbert bimodules is a homomorphism. (cid:3) Theorem 6.13. Let G be a sober étale topological groupoid and let S and ϕ: S → Bis(G) be an inverse semigroup model for G. Let A be a C∗-algebra. An action of G on A by Hilbert bimodules is equivalent to a pair consisting of an action of S on A by Hilbert bimodules and an S-equivariant map ψ : Prim(A) → G0. The section C∗-algebras of the corresponding Fell bundles over Bis(G) and S are the same. The S-equivariance of ψ refers to the action of S on Prim(A) from Lemma 6.12. Proof. Assume first that G0 is locally compact Hausdorff. In that case, an action of G is the same as a Fell bundle over G by Theorem 6.1. This determines an action of Bis(G), which we may compose with ϕ to get an action of S; we also get an S-equivariant map ψ. Conversely, let an action of S and a continuous S-equivariant map ψ : Prim(A) → G0 be given. Since G ∼= G0 ⋊ S, we may carry over the proof of Theorem 6.1. The S-equivariance of ψ gives the compatibility condition (18). Hence literally the same argument still works. If G0 is only a sober topological space, we need a different proof because we cannot describe G-actions fibrewise. We first construct the section C∗-algebra B of the Fell bundle over S corresponding to the action by Theorem 4.8. This C∗-algebra is S-graded by construction: it is the Hausdorff completion of the ∗-algebraLt∈S Ht in the maximal C∗-seminorm that vanishes on ju,t(ξ)δu − ξδt for all t, u ∈ S with t ≤ u and all ξ ∈ Ht, and we let Bt ⊆ B be the image of Ht in B. In particular, we may identify A = B1. Now we must construct a Bis(G)-grading ( ¯Bt)t∈Bis(G) on B with ¯Bϕ(t) = Bt for all t ∈ S and ¯BU = A(U ) for all U ∈ O(G0), where A(U ) denotes the ideal in A corresponding to ψ−1(U ) ∈ O(Prim(A)). This is done similarly to the proof of Theorem 6.11. Since this is rather technical and we already have another proof in the locally compact Hausdorff case, we leave it to the determined reader to spell out the details of this argument. (cid:3) 7. Actions by automorphisms are not enough The following theorem shows that the multiplication action of a non-Hausdorff groupoid on its own arrow space cannot be described by a continuous groupoid action by automorphisms. Theorem 7.1. Let G be a locally quasi-compact, locally Hausdorff, étale groupoid with Hausdorff G0, such that G1 is not Hausdorff. Let A be a C∗-algebra with Prim(A) ∼= G1. There is no continuous (twisted) action of G on A by automor- phisms that induces the left multiplication action on Prim(A) ∼= G1. Proof. Since Prim(A) ∼= G1, the lattice of ideals in A is order-isomorphic to the lattice of open subsets in G1. Let A(U ) ⊳ A for an open subset U ⊆ G1 be the corresponding ideal in A. Then Prim(A(U )) ∼= U . Part of a continuous action of G on A is a continuous map Prim(A) → G0. (This is equivalent to a C0(G0)-algebra structure.) Since we want to have the left multiplication action of G1 on Prim(A), we assume that this map becomes the range map G1 → G0 when we identify Prim(A) ∼= G1. The fibre at x ∈ G0 is the restriction of A to the closed subset Gx = {g ∈ G1 r(g) = x}, which 44 ALCIDES BUSS AND RALF MEYER we denote by AGx; we have Prim(AGx) = Gx. A G-action on A must provide isomorphisms αg : AGs(g) → AGr(g) for g ∈ G1. We assume that αg induces the map Gs(g) → Gr(g), h 7→ gh, on the primitive ideal space. What does continuity of g 7→ αg mean? Let U, V ⊆ G1 be bisections, then U · V is also a bisection. If g ∈ U , h ∈ V satisfy s(g) = r(h), then αg restricts to an isomorphism αg,h : Ah → Agh. Any element of U · V is of the form g · h for unique g ∈ U , h ∈ V . Continuity of (αg) means that for all bisections U, V and all a = (ah)h∈V in A(V ), the section (g · h) 7→ αg,h(ah) for g ∈ U , h ∈ V is continuous on U · V , that is, it belongs to A(U · V ) (see also [27, Definition 2.3]). Thus we get isomorphisms αU : A(V ) → A(U · V ). In brief, Bis(G) acts on A by partial isomorphisms. Since G1 is non-Hausdorff, there are g1, g2 ∈ G1 that cannot be separated by open subsets. Then r(g1) = r(g2) and s(g1) = s(g2). Let U1 and U2 be bisections of G containing g1 and g2, respectively. Shrinking them, we may achieve that s(U1) = s(U2). Let V := U ∗ 1 U1 = {1x x ∈ s(U1)} = U ∗ 2 U2; then U1V = U1 and U2V = U2. Since g1 and g2 cannot be separated, there is a net (hn) in U1 ∩ U2 that converges both to g1 and to g2. Let f ∈ A(V ) with f (1s(g1)) 6= 0. Then αU1(f ) ∈ A(U1V ) and αU2(f ) ∈ A(U2V ) by our continuity assumption. Thus ψ := αU1(f ) · αU2(f )∗ ∈ A(U1V ) ∩ A(U2V ) = A(U1 ∩ U2), so ψ vanishes at g1 and g2. At hn ∈ U1 ∩ U2, we have αU1(f )(hn) = αU1∩U2(f )(hn) = αU2(f )(hn) = αhn(f (1s(hn))). Since each αhn is an isomorphism, we get kψ(hn)k = kαhn(f (1s(hh))f (1s(hn))∗)k = kf (1s(hh))k2. If U ⊆ G1 is Hausdorff and a ∈ A(U ), then U ∋ x 7→ kakx is continuous by [26, Corollary 2.2] because the map Prim A(U ) → U is open and U is Hausdorff and locally compact. Therefore, kψ(hn)k converges towards kψ(g1)k = 0. At the same time, kψ(hn)k converges towards kf (1s(g1))k2 6= 0 because s(hn) → s(g1) inside the Hausdorff open subset V . This contradiction shows that there is no continuous action of G on A that lifts the multiplication action on Prim(A) ∼= G1. (cid:3) Remark 7.2. More generally, if we only assume an open continuous surjection p: Prim(A) → G1, then there is no continuous action of G on A such that p is G-equivariant for the induced action of G on Prim(A) and the left multiplication action on G1; the proof is exactly the same. The proof of Theorem 7.1 does not care about the multiplicativity of the action, so allowing "twisted" actions of G does not help. There are only two ways around this. First, we may allow Fell bundles over G. Secondly, we may allow actions of the inverse semigroup Bis(G). After stabilisation every Fell bundle becomes a twisted action by partial automorphisms (see [3]). We cannot remove the twist, however, because an untwisted action of Bis(G) by automorphisms would give an action of G by automorphisms as well, which cannot exist by Theorem 7.1. 8. A simple explicit example Let G be the group bundle over G0 = [0, 1] with trivial isotropy groups G(x) for x 6= 0 and with G(0) ∼= Z/2 = {1, −1}. So, as a set, G is (0, 1] ∪ {0+, 0−} with 0+ corresponding to +1 ∈ Z/2 and 0− to −1 ∈ Z/2. The topology on G is the quotient topology from [0, 1] × Z/2, where we divide by the equivalence INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 45 relation generated by (x, 1) ∼ (x, −1) for x 6= 0. With this topology, G is an étale, quasi-compact, second countable, locally Hausdorff, non-Hausdorff groupoid (even a group bundle). The points 0+ and 0− cannot be topologically separated: any net in (0, 1] converging to 0+ also converges to 0−, and vice versa. Let H be the groupoid of the equivalence relation ∼ on [0, 1] × Z/2 just defined. Its C∗-algebra C∗(H) ∼= C∗ r (H) is A := {f ∈ C([0, 1], M2) : f (0) is diagonal}. (This can be proved using the same idea as in [8, Example 7.1].) This is a C∗-algebra ∼= C2, and it has A ∼= Prim(A) ∼= G1 over [0, 1] with fibres Ax (this is a special case of [8, Corollary 5.4]). Theorem 7.1 shows that there is no action of G on A by automorphisms that would model the left multiplication action of G on G1. ∼= M2 at x 6= 0 and A0 Since A is the groupoid C∗-algebra of the Čech groupoid for the covering [0+, 1]∪ [0−, 1] = H 1, our main results give an action of G on A by Hilbert bimodules. We first describe it as an inverse semigroup action for a very small inverse semigroup S that models G. We consider three special bisections of G: 1 = [0+, 1] = G1 \ {0−}, g = [0−, 1] = G1 \ {0+}, e = (0, 1] = g ∩ 1. The bisection 1 is the unit bisection of G, so 1x = x = x1 for all x ∈ {1, g, e}. Moreover, g2 = 1, e2 = e, and eg = ge = e. Thus S := {1, e, g} is an inverse semigroup with x∗ = x for all x ∈ {1, e, g}. A bisection t of G cannot contain both 0+ and 0−. Hence either 0+ ∈ t ⊆ 1, 0− ∈ t ⊆ g, or t ⊆ e = 1 ∩ g. The groupoid G is the étale groupoid associated to the trivial action of S on G0; here the trivial action has 1 and g acting by the identity on G0 and e acting by the identity on (0, 1] ⊆ G0. An action of G on a groupoid or a C∗-algebra is equivalent to an action of S together with a compatible action of G0 = [0, 1] (Theorem 6.13). The transformation groupoid L of the S-action on H may be identified with the groupoid of the equivalence relation on [0, 1] ⊔ [0, 1] that identifies the two copies of (0, 1], so that L1 = [0, 1] × {(+, +), (+, −), (−, +), (−, −)} ⊆ ([0, 1] × {(+, +), (+, −), (−, +), (−, −)})2. The S-grading on L has L1 = (0, 1] × {(+, +), (+, −), (−, +), (−, −)} ⊔ {0} × {(+, +), (−, −)}, Lg = (0, 1] × {(+, +), (+, −), (−, +), (−, −)} ⊔ {0} × {(+, −), (−, +)}, Le = (0, 1] × {(+, +), (+, −), (−, +), (−, −)} = L1 ∩ Lg. So L1 ∼= H is open but not closed. The C∗-algebra of L is B := C([0, 1], M2). To let S act on the C∗-algebra A of H, we use the transformation groupoid C∗-algebra B and the involution u := (cid:0) 0 1 1 0(cid:1) ∈ B. We have u = u∗ and u2 = 1, u · A(0, 1] = A(0, 1] = A(0, 1] · u and uA = Au as subsets of B. Let A1 := A, Ae := A(0, 1] ⊆ A1, and Ag := uA = Au. These subspaces Ax for x ∈ S satisfy A∗ x = Ax = Ax∗ for all x ∈ S and Ax · Ay = Axy for all x, y ∈ S; in particular, Ag is a full Hilbert bimodule over A1 with inner products given by the usual formulas a∗ 1 · a2 and a1 · a∗ 2. Furthermore, A1 ∩ Ag = Ae and A1 + Ag = B because elements of Ag are precisely those f ∈ B with off-diagonal f (0). Hence the map g 7→ Ag defines an action of S on A by Hilbert bimodules. Since A1 +Ag is already complete in the C∗-norm of B, there is only one C∗-norm on A1 + Ag that extends the given C∗-norm on A1. Thus the sectional C∗-algebra for the resulting Fell bundle over S is B, which is Morita -- Rieffel equivalent to C[0, 1]. 46 ALCIDES BUSS AND RALF MEYER The S-action on A extends to all bisections of G because they are all contained in 1 or g: if t ⊆ G1 is a bisection, then let At = A1s(t) if t ⊆ 1 and At = Ags(t) if t ⊆ g; this is consistent for t ⊆ 1 ∩ g = e because Ae = A1 ∩ Ag. Next we describe a twisted S-action by partial automorphisms of A that induces the S-action by Hilbert bimodules described above. (This is possible by [3, Corol- lary 4.16] because our saturated Fell bundle is regular in the notation of [3].) A twisted S-action by partial automorphisms is given by ideals A1 = A and Ae with isomorphisms αx : Axx∗ → Ax∗x and unitary multipliers (the twists) ω(x, y) in M(Axyy∗x∗) for x, y ∈ S. For the idempotent elements x = e, 1, the isomorphism αx is the identity; for x = g, it is the order-2-automorphism αg : A → A, a 7→ uau, because a1 · ua2 = u · (ua1u · a2) for all a1 ∈ A1, ua2 ∈ Ag. The automorphism αg is not inner on A1 because u ∈ B does not belong to M(A). The restriction of αg to the ideal Ae becomes inner, however, because u ∈ M(A(0, 1]). This unitary u enters in the twisting unitaries ω(x, y) for x, y ∈ S; they are 1 if x = 1 or y = 1, or if (x, y) is (e, e) or (g, g) (α2 g = IdA = α1). The remaining cases are ω(e, g) = ω(g, e) = uAe, that is, u viewed as a multiplier of the ideal Ae = A(0, 1]. It is routine to check that this data gives a twisted action of S on A in the sense of [3, Definition 4.1] and that the resulting saturated Fell bundle over S is isomorphic to the one described above. Incidentally, this is not a twisted action in the sense of Sieben [34] because ω(e, g) and ω(g, e) are non-trivial although e is idempotent. This twisted S-action cannot be turned into a groupoid action of G by partial automorphisms because for x ∈ 1∩g, the restrictions of αg and α1 to As(x) differ by a non-trivial inner automorphism. This impossibility is in accord with Theorem 7.1. Remark 8.1. The Packer -- Raeburn Stabilisation Trick replaces a twisted group ac- tion by an untwisted action on a suitable C∗-stabilisation. We claim that this cannot be done for the above inverse semigroup twisted action. Let D be a C∗-algebra with an untwisted action of S by automorphisms. Then 1 and e act by the identity on D and by some ideal De ⊳ D, respectively, and g acts by some automorphism αg on D. If there is no twist, then αgDe = α1De is the identity on De because eg = e = ge. Suppose that D is also a C∗-algebra over [0, 1], with D((0, 1]) = De. Then this allows to define an action of the groupoid G on D by letting elements of g or 1 act by the fibre restrictions of αg and IdD, respectively. This gives a well-defined, untwisted action of G on D. Theorem 7.1 implies that Prim(D) 6∼= G1, so that A and D cannot be Morita -- Rieffel equivalent. This example therefore shows that the Packer -- Raeburn Stabilisation Trick cannot be extended from groups to inverse semigroups or non-Hausdorff groupoids. Appendix A. Preliminaries on topological groupoids This appendix defines topological groupoids and equivalences between them, fol- lowing [21]. The point is that all this works smoothly without assuming topological spaces to be Hausdorff or locally (quasi)compact, if we choose appropriate defini- tions. The theory of possibly non-Hausdorff topological groupoids becomes very natural if one treats topological groupoids, Lie groupoids, infinite-dimensional Lie groupoids (modelled on Banach or Fréchet manifolds), and other types of groupoids simultaneously as in [21]. Here we recall the results and definitions from [21] that are relevant for us. The theory of topological groupoids and their principal bundles and equivalences depends on a choice of "covers" in the category of topological spaces (see [21]). We choose the open surjections as covers. This means that we require the range and source maps in a topological groupoid, the bundle projection in a principal bundle, and the anchor maps in a (bibundle) equivalence to be open surjections. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 47 Following Bourbaki, we require compact and locally compact spaces to be Haus- dorff. Since many authors allow non-Hausdorff locally compact spaces, we usually speak of "Hausdorff locally compact" spaces to avoid confusion. A topological space is locally quasi-compact if every point has a neighbourhood basis consisting of quasi- compact neighbourhoods. This is strictly more than having a single quasi-compact neighbourhood, but both notions coincide in the locally Hausdorff case, which is the case we are interested in. Recall that a topological space is locally Hausdorff if every point has a Hausdorff neighbourhood (and thus a neighbourhood basis consisting of Hausdorff neighbourhoods). A space is locally Hausdorff, locally quasi-compact if and only if every point has a compact (hence Hausdorff) neighbourhood. It would make sense to call such spaces "locally compact," if it were not for the conflict with other established notation. A.1. Topological groupoids, principal bundles, and equivalences. We now specialise the general definitions of groupoids, groupoid actions, principal bundles, basic groupoid actions and bibundle equivalences in [21] to the category of (all) topological spaces with open surjections as covers. Proposition A.1. A topological groupoid consists of topological spaces G0 and G1 and continuous maps r, s: G1 ⇒ G0 and m: G1 ×s,G0,r G1 → G1, (g1, g2) 7→ g1 · g2, such that (G1) s(g1 · g2) = s(g2) and r(g1 · g2) = r(g1) for all g1, g2 ∈ G1; (G2) m is associative: (g1 · g2) · g3 = g1 · (g2 · g3) for all g1, g2, g3 ∈ G1 with s(g1) = r(g2), s(g2) = r(g3); (G3) the following two maps are homeomorphisms: G1 ×s,G0,r G1 → G1 ×s,G0,s G1, G1 ×s,G0,r G1 → G1 ×r,G0,r G1, (g1, g2) 7→ (g1 · g2, g2), (g1, g2) 7→ (g1, g1 · g2), (G4) r and s are open surjections. Then m is open and surjective and there are continuous maps G0 → G1 and G1 → G1 with the usual properties of unit and inversion. Conversely, the maps in (G3) are homeomorphisms if G has continuous unit and inversion maps. Proof. Our definition of a groupoid is exactly [21, Definition 3.4]. It implies that m is open and surjective and is equivalent to the usual one with unit and inverse by [21, Proposition 3.6]. (cid:3) Let G be a topological groupoid as above. Proposition A.2. A (right) G-action is a space X with continuous maps s: X → G0 and m: X ×s,G0,r G1 → X, (x, g) 7→ x · g, such that (A1) s(x · g) = s(g) for all x ∈ X, g ∈ G1 with s(x) = r(g); (A2) m is associative: (x · g1) · g2 = x · (g1 · g2) for all x ∈ X, g1, g2 ∈ G1 with s(x) = r(g1) and s(g1) = r(g2); (A3) m is surjective. Condition (A3) holds if and only if x · 1s(x) = x for all x ∈ X, if and only if m is an open surjection, if and only if the following map is a homeomorphism: X ×s,G0,r G1 → X ×s,G0,s G1, (x, g) 7→ (x · g, g). Proof. This is contained in [21, Definition and Lemma 4.1]. (cid:3) Left actions are defined similarly and are equivalent to right actions by g · x = x · g−1. The transformation groupoid X ⋊ G of a groupoid action is a topological groupoid by [21, Definition and Lemma 4.11]. Any groupoid acts on G0 by r(g) · 48 ALCIDES BUSS AND RALF MEYER g := s(g) for all g ∈ G1, and on G1 both on the left and right by left and right multiplication. Proposition A.3. For any G-action on a topological space X, the orbit space projection X → X/G is an open surjection, and X/G is Hausdorff if and only if X ×X/G X = {(x1, x2) ∈ X there is g ∈ G1 with s(x1) = r(g) and x1 · g = x2} is a closed subset of X × X. Proof. The orbit space projection is open by [21, Proposition 9.31] because the range and source maps of G are open. By [21, Proposition 9.18], X/G is Hausdorff if and only if X ×X/G X is closed in X × X (open surjections are clearly biquotient maps, see the discussion in [21, Section 9.6]). (cid:3) We now specialise the general concepts of basic actions and principal bundles from [21] to our context. Proposition A.4. A right G-action is basic if the map (19) X ×s,G0,r G1 → X × X, (x, g) 7→ (x, x · g), is a homeomorphism onto its image with the subspace topology. A principal right G-bundle is a space X with continuous maps s: X → G0, p: X → Z, and m: X ×s,G0,r G1 → X, (x, g) 7→ x · g, such that (Pr1) s(x · g) = s(g) and p(x · g) = p(x) for all x ∈ X, g ∈ G1 with s(x) = r(g); (Pr2) m is associative: (x · g1) · g2 = x · (g1 · g2) for all x ∈ X, g1, g2 ∈ G1 with s(x) = r(g1) and s(g1) = r(g2); (Pr3) the map X ×s,G0,r G1 → X ×p,Z,p X, (x, g) 7→ (x, x · g), is a homeomorphism; (Pr4) the map p is open and surjective. Then x · 1s(x) = x for all x ∈ X, and there is a unique homeomorphism Z ∼= X/G intertwining p and the canonical projection X → X/G. Thus a principal G-bundle is equivalent to a basic G-action with a homeomorphism X/G ∼= Z. Proof. A principal bundle in the sense above also satisfies x · 1s(x) = x for all x ∈ X because of (Pr3) (see [21, Lemma 5.3]). Hence s and m give a right G-action, and all conditions for a principal bundle in [21] are met. [21, Lemma 5.3] also gives the unique homeomorphism X/G ∼= Z intertwining p and the canonical map X → X/G. A groupoid action is called basic in [21] if it becomes a principal bundle with X → X/G as bundle projection. The canonical map X → X/G is automatically G-invariant, and it is an open surjection by [21, Proposition 9.31]. Thus the second half of (Pr1) and (Pr4) hold for any G-action with this choice of p. The first half of (Pr1) and (Pr2) are part of the definition of a groupoid action. The image of the map in (19) is X ×X/G X by the definition of X/G, so that (Pr3) is equivalent to (19) being a homeomorphism onto its image. (cid:3) Next, we consider the notion of equivalence between groupoids as defined in [21]. We will relate it to notions of equivalence by other authors in Appendix A.2. Proposition A.5. Let G and H be topological groupoids. A bibundle equivalence from H to G consists of a topological space X, continuous maps r : X → G0, s: X → H 0 (anchor maps), G1 ×s,G0,r X → X and X ×s,H0,r H 1 → X (multi- plications), satisfying the following conditions: INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 49 (E1) s(g · x) = s(x), r(g · x) = r(g) for all g ∈ G1, x ∈ X with s(g) = r(x), and s(x · h) = s(h), r(x · h) = r(x) for all x ∈ X, h ∈ H 1 with s(x) = r(h); (E2) associativity: g1 · (g2 · x) = (g1 · g2) · x, g2 · (x · h1) = (g2 · x) · h1, x · (h1 · h2) = (x · h1) · h2 for all g1, g2 ∈ G1, x ∈ X, h1, h2 ∈ H 1 with s(g1) = r(g2), s(g2) = r(x), s(x) = r(h1), s(h1) = r(h2); (E3) the following two maps are homeomorphisms: G1 ×s,G0,r X → X ×s,H0,s X, X ×s,H0,r H 1 → X ×r,G0,r X, (g, x) 7→ (x, g · x), (x, h) 7→ (x, x · h); (E4) s and r are open; (E5) s and r are surjective. Then 1r(x) · x = x = x · 1s(x) for all x ∈ X, and the anchor maps descend to homeomorphisms G\X ∼= H 0 and X/H ∼= G0. Proof. Condition (E1) and (E3) are equivalent to (Pr1) and (Pr3) for both the left G-action with p = s and the right H-action with p = r, respectively. Condi- tion (E2) means that the left G-action and the right H-action satisfy (Pr2) and commute. Conditions (E4) and (E5) together are equivalent to (Pr4) for both actions. Thus the conditions (E1) -- (E5) characterise bibundle equivalences in the notation of [21]. The last sentence follows from the general properties of principal bundles, see Proposition A.4. (cid:3) In the following, we abbreviate "bibundle equivalence" to "equivalence" because we do not use any other equivalences between groupoids. We have switched the direction of a bibundle equivalence compared to [21] be- cause this is convenient here. Going from right to left is also consistent with our notation s and r for the right and left anchor maps. A.2. Basic actions versus free and proper actions. We now compare our basic actions with free and proper actions. A continuous map f : X → Y is closed if it maps closed subsets of X to closed subsets of Y , and proper if IdZ × f : Z × X → Z ×Y is closed for all topological spaces Z or, equivalently, f is closed and f −1(y) is quasi-compact for all y ∈ Y (see [2, Theorem 1 in I.10.2]). A map from a Hausdorff space X to a Hausdorff locally compact space Y is proper if and only if preimages of compact subsets are compact. In this case, X is necessarily locally compact ([2, Proposition 7 in I.10.3]). Definition A.6. A right action of a topological groupoid G on a topological space X is proper if the map in (19) is proper. The action is free if the map (19) is injective. Groupoids for which the action on its unit space is free (that is, for which the map s × r : G1 → G0 × G0 is injective) are often called principal (see [30]). This terminology conflicts, however, with the usual notion of a principal bundle, which requires extra topological conditions besides freeness of the action. We call a groupoid basic if its canonical action on the object space is basic, that is, the map s × r : G1 → G0 × G0 is a homeomorphism onto its image. Proposition A.7. A groupoid action is free and proper if and only if it is basic and has Hausdorff orbit space. If G and H are topological groupoids with Hausdorff object spaces, then an equiva- lence from H to G in our sense is the same as a topological space X with commuting free and proper actions of G and H, such that the anchor maps induce homeomor- phisms G\X ∼= H 0 and X/H ∼= G0. 50 ALCIDES BUSS AND RALF MEYER Proof. The characterisation of free and proper actions is [21, Corollary 9.32]; the main point of the proof is that the orbit space is Hausdorff if and only if the orbit equivalence relation is closed in X ×X (Proposition A.3). The left and right actions on an equivalence are basic with X/H ∼= G0 and G\X ∼= H 0; hence they are free and proper if and only if G0 and H 0 are Hausdorff, respectively. Conversely, if the actions of G and H on X are free and proper, then both actions are basic, and both anchor maps are open because they are equivalent to orbit space projections; thus we have an equivalence in our sense. (cid:3) For a general action of a groupoid G on a space X, the image of the map (19) is the orbit equivalence relation X ×X/G X ⊆ X × X. Thus the map (19) is a homeomorphism (the action is basic) if and only if the action is free and the map that sends (x1, x2) ∈ X ×X/G X to the unique g ∈ G1 with s(x1) = r(g) and x1 · g = x2 is continuous. If G, H and X are locally compact Hausdorff, then an equivalence in our sense is the same as a (G, H)-equivalence in the notation of [23]; the main result of [23] is that such an equivalence induces a Morita equivalence between the groupoid C∗-algebras of G and H (for any Haar systems). For non-Hausdorff groupoids, Jean-Louis Tu defined a notion of equivalence in [36], using a technical variant of proper actions: he calls a groupoid G ρ-proper with respect to a G-invariant continuous map ρ: G0 → T if the map (r, s): G1 → G0 ×ρ,T,ρ G0, g 7→ (r(g), s(g)), is proper. If T is non-Hausdorff, then G0 ×ρ,T,ρ G0 need not be closed in G0 ×G0, so that this is weaker than properness. In the definition of equivalence, he takes ρ to be the anchor map on the other side, so he requires the maps in (E3) to be proper. These maps are continuous bijections because the actions are free. A continuous, proper bijection, being closed, must be a homeomorphism. Thus Tu's notion of equivalence is equivalent to ours. A.3. Covering groupoids and equivalence. Definition A.8. Let f : X → Z be a continuous, open surjection. The covering groupoid G(f ) has object space X, arrow space X ×f,Z,f X, range and source maps r(x1, x2) := x1, s(x1, x2) := x2, and multiplication (x1, x2) · (x2, x3) := (x1, x3) for all x1, x2, x3 ∈ X with f (x1) = f (x2) = f (x3). The assumption on f implies that it is a quotient map, that is, we may identify Z with the quotient space X/∼ by the following equivalence relation: x ∼ y if and only if f (x) = f (y); and f becomes the quotient map X → X/∼. The covering groupoid G(f ) is the groupoid associated to this equivalence relation. In particular, Z can be identified with the orbit space X/G(f ) for the canonical action of G(f ) on its unit space X. Every covering groupoid is basic, that is, its action on the unit space is basic. Conversely, if G is a basic groupoid, then it is isomorphic to a covering groupoid. The map r×s: G1 → G0×G0 gives a homeomorphism from G1 onto G0×f,G0/G,f G0, where f : G0 → G0/G denotes the quotient map. This yields an isomorphism of topological groupoids G ∼= G(f ). Example A.9 (Čech groupoids). Let Z be a topological space and let U be an open covering of Z. Let X := FU∈U U and let f : X → Z be the canonical map: f is the inclusion map on each U ∈ U. This map is an open surjection. It is even étale, that is, a local homeomorphism. We denote the covering groupoid of f by GU and call it the Čech groupoid of the covering. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 51 Assume that Z is locally Hausdorff and choose the open covering U to consist of Hausdorff open subsets U ⊂ Z. Then the Čech groupoid GU is a Hausdorff, étale topological groupoid (see also [8, Lemma 4.2]). If, in addition, Z is locally quasi-compact, then GU is a (Hausdorff) locally compact, étale groupoid. This is the situation we are mainly interested in. Proposition A.10. Let fi : Xi → Z for i = 1, 2 be two continuous, open surjec- tions. Then X1×f1,Z,f2 X2 with the obvious left and right actions of G(f1) and G(f2) gives an equivalence from G(f2) to G(f1). Proof. This is [21, Example 6.4]. (cid:3) If G(f1) and G(f2) are Hausdorff locally compact, then so is the equivalence X1 ×f1,Z,f2 X2 between them. If the maps f1 and f2 are both étale -- for instance, if they come from open coverings of Z -- then the groupoids G(f1) and G(f2) are étale, and the anchor maps X1 ← X1 ×f1,Z,f2 X2 → X2, x1 ← (x1, x2) → x2, are étale as well. Proposition A.11. The covering groupoid G(f ) of a continuous open surjection f : X → Z is always equivalent (as a topological groupoid) to the space Z viewed as a groupoid with only identity arrows. In particular, the Čech groupoid of a covering of Z is equivalent to Z. Conversely, if X is an equivalence from a space Z to a topological groupoid G, then G is isomorphic to the covering groupoid of the anchor map s: X → Z. Hence covering groupoids are exactly the groupoids that are equivalent to spaces. Proof. The first part is a consequence of Proposition A.10 applied to f1 = f and f2 = IdZ (see also [21, Example 6.3]). For the second part, observe that the action of Z on X is simply the anchor map s: X → Z, which must be an open surjection. The anchor map r : X → G0 must be a homeomorphism (because it must be the projection map X → Z\X = X), so we may as well assume X = G0. Then G1 ×s,G0,r X ∼= G1, and the first isomorphism in (E3) identifies G1 with X ×s,Z,s X. This yields an isomorphism from G to the covering groupoid G(s) of s: X → Z. (cid:3) Let Z be a space, view Z as a groupoid with only identity arrows. When is Z equivalent to a locally compact, Hausdorff groupoid? If Z is equivalent to a topo- logical groupoid G, then G is necessarily the covering groupoid G(f ) of a cover f : X → Z by Proposition A.11. Given a space Z, we thus seek a locally compact, Hausdorff space X and an open, continuous surjection f : X → Z such that X ×f,Z,f X is locally compact. The question when X ×f,Z,f X is locally compact is also asked in [8] at the end of Section 4. We answer this question in Proposition A.14 below: X ×f,Z,f X is locally compact if and only if Z is locally Hausdorff. Proposition A.16 says that the only topological spaces Z that are equivalent to locally compact Hausdorff groupoids are the locally Hausdorff, locally quasi-compact ones; for them, Example A.9 gives such an equivalence, where the groupoid is even étale. We need some preparation in order to prove Proposition A.14. Definition A.12 ([2, I.3.3, Définition 2, Proposition 5]). A subset S of a topolog- ical space X is locally closed if it satisfies the following equivalent conditions: (1) any x ∈ S has a neighbourhood U such that S ∩ U is relatively closed in U ; (2) S is open in its closure; (3) S is an intersection of an open and a closed subset of X. The following proposition generalises [2, I.9.7, Propositions 12 and 13] to the locally Hausdorff case. 52 ALCIDES BUSS AND RALF MEYER Proposition A.13. A subset S of a locally Hausdorff, locally quasi-compact space X is locally quasi-compact in the subspace topology if and only if it is locally closed. Proof. First let S be locally closed. Write S = A∩U with A closed and U open in X. Let x ∈ S. Since X is locally quasi-compact, the quasi-compact neighbourhoods of x in X form a neighbourhood basis of X. Since x ∈ U , those quasi-compact neighbourhoods of x that are contained in U form a neighbourhood basis in U . Their intersections with A remain quasi-compact because A is closed in X. They form a neighbourhood basis of x in S, proving that S is locally quasi-compact. Conversely, assume that S is locally quasi-compact in the subspace topology. Let x ∈ S. Let U be a Hausdorff open neighbourhood of x in X. Then S ∩ U is a neighbourhood of x in S and hence contains a quasi-compact neighbourhood K of x in S because S is locally quasi-compact. We have K = S ∩ V for some neighbourhood V of x in X, and we may assume V ⊆ U because K ⊆ U . The subset S ∩ V is relatively closed in V because U ⊇ V is Hausdorff and S ∩ V is quasi-compact. Thus S is locally closed. (cid:3) Proposition A.14. Let f : X → Z be a continuous, open surjection. The equiva- lence relation Xf,Z,f X ⊆ X × X defined by f is locally closed if and only if Z is locally Hausdorff. In particular, if X is locally quasi-compact and locally Hausdorff, then Xf,Z,f X is locally quasi-compact if and only if Z is locally Hausdorff. Proof. Assume Z to be locally Hausdorff first. Let (x1, x2) ∈ X ×f,Z,f X and let U ⊆ Z be a Hausdorff open neighbourhood of f (x1) = f (x2). Then f −1(U ) ⊆ X is an open subset such that f : f −1(U ) → U is an open map onto a Hausdorff space. Hence f −1(U ) ×f,U,f f −1(U ) =(cid:0)X ×f,Z,f X(cid:1) ∩(cid:0)f −1(U ) × f −1(U )(cid:1) is relatively closed in f −1(U ) × f −1(U ) by [21, Proposition 9.15]. Thus X ×f,Z,f X is locally closed in X × X. Conversely, assume X ×f,Z,f X to be locally closed in X × X. Let x ∈ X. Then (x, x) has a neighbourhood in X × X so that X ×f,Z,f X restricted to it is relatively closed. Shrinking this neighbourhood, we may assume that it is of the form U × U for an open neighbourhood of x, by the definition of the product topology on X ×X. The map f U : U → f (U ) is open, and (X ×f,Z,f X) ∩ (U × U ) = U ×f U ,f (U),f U U . Since this is relatively closed by assumption, [21, Proposition 9.15] shows that f (U ) is Hausdorff. Since x was arbitrary, this means that Z is locally Hausdorff. The last sentence follows from the first one and Proposition A.13. (cid:3) Corollary A.15. A topological space X is locally Hausdorff if and only if the diagonal {(x, x) x ∈ X} is a locally closed subset in X × X. Proof. Apply Proposition A.14 to the identity map. (cid:3) Proposition A.16. Let G be a locally quasi-compact, locally Hausdorff groupoid and let X be a basic right G-action. Then X/G is locally quasi-compact and locally Hausdorff. If X is an equivalence from a space Z to G, then Z ∼= X/G is locally quasi- compact and locally Hausdorff. Proof. Since G and X are locally quasi-compact and locally Hausdorff, so is their product X × G1. Since G0 is locally Hausdorff, the diagonal in G0 is locally closed by Corollary A.15. The fibre product X ×s,G0,r G1 is the preimage of the diagonal in G0 × G0 under the continuous map r × s: X × G1 → G0 × G0; hence X ×s,G0,r G1 is locally closed in X × G1. Thus X ×s,G0,r G1 is locally quasi-compact and locally Hausdorff by Proposition A.13. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 53 Since the G-action on X is basic, X ×s,G0,r G1 is homeomorphic to the subset X ×X/G X ⊆ X × X. Now Proposition A.13 shows that X ×X/G X is locally closed in X × X. Then X/G is locally Hausdorff by Proposition A.14. Since continuous images of quasi-compact subsets are again quasi-compact, X/G is also locally quasi- compact. An equivalence from a space Z to G is the same as a basic G-action with a homeomorphism X/G ∼= Z. If this exists, then Z must be locally Hausdorff and locally quasi-compact by the above argument. (cid:3) Appendix B. Fields of Banach spaces over locally Hausdorff spaces Let X be a locally quasi-compact, locally Hausdorff space. Thus any Hausdorff open subset of X is locally compact. Definition B.1 (see [25] and the references there). An upper semicontinuous field of Banach spaces on X is a family of Banach spaces (Bx)x∈X with a topology on B = Fx∈X Bx such that, for each Hausdorff open subset U of X, BU is an upper semicontinuous field of Banach spaces on U . In particular, the norm of any continuous section of BU is an upper semicontinuous scalar-valued function on U . Let S(U, B) denote the vector space of continuous, compactly supported sections of BU . This is the union (hence inductive limit) of the subspaces S0(K, B) of continuous sections on K vanishing on ∂K, where K runs through the directed set of compact subsets of U and ∂K = K ∩ U \ K is the boundary of K in U . Each S0(K, B) is a Banach space for the supremum norm kf k∞ := sup{kf (x)k x ∈ K}. We call a subset of S(U, B) bounded if it is the image of a norm-bounded subset of S0(K, B) for some K. If f ∈ S(U, B) for a Hausdorff open subset U of X, then we always extend f to a section of B on all of X by taking f (x) := 0 for x /∈ U . Let S(X, B) be the vector i=1 fi for fi ∈ S(Ui, B) and Hausdorff open subset Ui of X. We call such sections of B quasi-continuous. space of all sections of B that may be written as finite linear combinationsPm A subset A of S(X, B) is bounded if there are Hausdorff open subsets U1, . . . , Um of X and bounded subsets Ai ⊆ S(Ui, B) for i = 1, . . . , m such that every element To simplify our proofs, we use bornological language, that is, we speak of bounded instead of open subsets. For a Hausdorff locally compact space X, S(X, B) with its usual topology is an inductive limit of Banach spaces. The inductive limit topology is determined by its continuous seminorms. A seminorm is continuous if and only if it is bounded in the sense that its supremum over each bounded subset is finite; this is so because a seminorm on a Banach space is continuous if and only if it is bounded. For locally Hausdorff X, the bounded seminorms are those that restrict to bounded seminorms on all the subspaces S(U, B) for U ⊆ X open and Hausdorff; this is the same as the quotient topology from the map LU S(U, B) → S(X, B), where U runs through the Hausdorff open subsets of X. Thus the usual topology on S(X, B) -- which is the quotient topology induced by the inductive limit topologies on the direct sums of the spaces S(U, B) -- is the topology generated by all bounded seminorms. Let U be a family of open subsets of X with the following two properties: (1) X =SU∈U U , that is, for each x ∈ X there is U ∈ U with x ∈ U ; (2) U1 ∩U2 =S{U ∈ U U ⊆ U1 ∩U2} for all U1, U2 ∈ U; that is, if x ∈ U1 ∩U2, then there is U ∈ U with U ⊆ U1 ∩ U2 and x ∈ U . of A may be written as a sum Pm i=1 fi with fi ∈ Ai for i = 1, . . . , m. 54 ALCIDES BUSS AND RALF MEYER In our main application, the open subsets in U will not be Hausdorff. Thus S(U, B) for U ∈ U is defined in the same way as S(X, B), by taking finite linear combina- tions of continuous compactly supported sections on Hausdorff open subsets of U . We view S(U, B) as a subspace in S(X, B) by extending functions on U by 0 out- side U . This gives an injective, bounded linear map S(U, B) → S(X, B). Being bounded means that it maps bounded subsets to bounded subsets. Let ιU : S(U, B) → LU∈U S(U, B) for U ∈ U denote the inclusion map of the U -summand. We call a subset A of LU∈U S(U, B) bounded if there are finitely of A may be written as Pm many U1, . . . , Um ∈ U and bounded subsets Ai of S(Ui, B) such that any element i=1 ιUi(fi) with fi ∈ Ai. Proposition B.2. The map E : MU∈U S(U, B) → S(X, B) is bounded linear and a bornological quotient map in the sense that any bounded subset of S(X, B) is the image of a bounded subset ofLU∈U S(U, B); in particular, it is surjective. The kernel of E is the closed linear span of the set of elements of the form ιU (f ) − ιV (f ) for f ∈ S(U, B), U, V ∈ U with U ⊆ V . The "closure" in the description of the kernel is the bornological one, defined using Mackey's notion of convergence in a bornological vector space. For any el- ement g ∈ ker E, we will find a bounded subset A ⊆ LU∈U S(U, B), and linear combinations gn of ιU (f ) − ιV (f ) for f ∈ S(U, B), U, V ∈ U with U ⊆ V such that g − gn ∈ 2−n · A. This implies convergence in any bounded seminorm. Remark B.3. Proposition B.2 implies that E is a quotient map with respect to the canonical topologies on the spaces involved. That is, a seminorm p on S(X, B) is continuous if and only if p ◦ E is a continuous seminorm on LU∈U S(U, B). The proof uses that continuity and boundedness are equivalent for seminorms on both spaces and that E is a bornological quotient map. It seems inconvenient, however, to prove this directly without bornological language. Proof. In the proof, we abbreviate S(U ) := S(U, B) because the Banach space bundle is fixed throughout. We first show that E is a bornological quotient map. Let A ⊆ S(X) be bounded. By definition, there are finitely many Hausdorff open subsets V1, . . . , Vm ⊆ X, compact subsets Ki ⊆ Vi and scalars Ci > 0 such i=1 fi with fi ∈ S0(Ki) having kfik∞ ≤ Ci. Since the subsets U ∈ U cover X, they cover the compact subset Ki. Since compact spaces are paracompact, there is a finite subordinate partition of unity (ψi,U )U∈U, that is, ψi,U : Ki → [0, 1] is continuous and has compact support Li,U that any f ∈ A may be written as Pm If fi ∈ S0(Ki), then fi ·ψi,U ∈ S0(Ki ∩Li,U ) ⊆ S(Vi ∩U ) and kfi ·ψi,U k∞ ≤ kfik∞. i=1 fi with fi ∈ S0(Ki) having kfik∞ ≤ Ci, and then contained in U ∩ Ki, only finitely many ψi,U are non-zero, and PU∈U ψi,U (x) = 1. Now write f ∈ A first asPm i=1PU∈U fi · ψi,U . This sum is still finite because only finitely many ψi,U are asPn non-zero for each i, and each summand fi · ψi,U runs through a bounded subset of S(Vi ∩ U ) and hence of S(U ) because we have uniform control on the supports supp fiψi,U ⊆ Ki ∩ Li,U and norms kfi · ψi,U k∞ ≤ Ci of the summands. Hence A is contained in the E-image of a bounded subset in L S(U ). Now we describe the kernel of E. Let N be the linear span of elements of the form ιU (f ) − ιV (f ) for all f ∈ S(U ), U, V ∈ U with U ⊆ V . Since E(ιU (f ) − ιV (f )) = 0, we have N ⊆ ker E. If U1, U2, V ∈ U satisfy V ⊆ U1 ∩ U2 and f ∈ S(V ), then ιU1 (f ) − ιU2(f ) = −(ιV (f ) − ιU1(f )) + (ιV (f ) − ιU2(f )) ∈ N. INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 55 i i let K ◦ 1, . . . , m be such that f = Pm Pm i We are going to modify a given element of ker E by adding elements of N so that the norms of its constituents become arbitrarily small, without enlarging their supports. A generic element f ∈LU∈U S(U ) is of the form f =P ιU (fU ) with fU ∈ S(U ) and fU = 0 for all but finitely many U . Each non-zero fU is a sum fU =PkU j=1 fU,j with fU,j ∈ S(VU,j) for finitely many Hausdorff open subsets VU,1, . . . , VU,kU ⊆ U . We renumber the finitely many Hausdorff open subsets VU,j consecutively as V1, . . . , Vm and relabel our sections fi ∈ S(Vi) accordingly. Let Ui ∈ U for i = i=1 ιUi(fi); so Vi ⊆ Ui. Let Ki := supp fi ⊆ Vi and i be the interior of Ki inside Vi; thus x ∈ K ◦ i for all x ∈ X with fi(x) 6= 0. ) with f (0) = f , f (j+1) − f (j) ∈ N , and kf (m) Now assume f ∈ ker(E) and let ǫ > 0. We will construct a finite sequence f (j) = i=1 ιUi(f (j) k < ǫ for all i = 1, . . . , m. Furthermore, our construction ensures that the support of f (j) is contained in Ki for all i, j. Letting ǫ run through a sequence going to 0, the differences f −f (m) in N will converge to f in the sense explained above because each constituent fi − f (m) converges to fi in the normed space S0(Ki). Our construction will be such that i = f (i) f (j) for j ≥ i, that is, in the jth step we keep f1, . . . , fj−1 fixed. To make the following steps possible, we aim for stronger norm estimates kf (j) i k < 2j−mǫ. ) with f − f (j) ∈ N i k < 2j−mǫ for i = 1, . . . , j; for j = 0, this is satisfied for f (0) = f . We ) with f (j) − f (j+1) ∈ N and hence Assume that we have already constructed f (j) =Pm are going to construct f (j+1) =Pm i=1 ιUi(f (j+1) for i = 1, 2, . . . , j, and kf (j+1) f − f (j+1) ∈ N , with f (j+1) i=1 ιUi(f (j) and kf (j) = f (j) j+1(x)k ≥ 2j+1−mǫ}. This is a closed subset of K ◦ j+1 k < 2j+1−mǫ. j+1 because the norm function is upper semicontinuous. Since Kj+1 is compact, Aj+1 (x) = 0 for all Let Aj+1 = {x ∈ Vj+1 kf (j) x ∈ X. If x ∈ Aj+1, then this gives is compact. Since E(f ) = 0 and E(f − f (j)) = 0, we have Pm jXi=1 j+1(x)k − j+1Xi=1 f (j) i f (j) i i=1 f (j) kf (j) i k∞ > 0. i i i i i (cid:13)(cid:13)(cid:13)(cid:13) mXi=j+2 (x)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13) (x)(cid:13)(cid:13)(cid:13)(cid:13) ≥ kf (j) i i i Hence there must be i > j + 1 with f (j) i . Thus the open subsets K ◦ i , then x ∈ Ui ∩ Uj+1. By our assumption on U, there is U ∈ U with x ∈ U and U ⊆ Ui ∩ Uj+1. Thus the open subsets K ◦ i ∩ U for i > j + 1 and U ∈ U with U ⊆ Ui ∩ Uj+1 cover Aj+1. i for i > j + 1 cover Aj+1. If x ∈ Aj+1 ∩ K ◦ (x) 6= 0, so that x ∈ K ◦ Since Aj+1 is compact and contained in the Hausdorff locally compact space Vj+1, there is a subordinate finite partition of unity (ψi,U ). That is, all but finitely many ψi,U are non-zero, ψi,U : Aj+1 → [0, 1] is a continuous function with compact support contained in K ◦ non-zero ψi,U from Aj+1 to a continuous function ¯ψi,U : Kj+1 → [0, 1] vanishing in a neighbourhood of ∂Kj+1 and on Kj+1 \(K ◦ i ∩U ) because these two compact subsets of Aj+1 are disjoint from the compact support of ψi,U in K ◦ i ∩ U . If necessary, we i ∩U , andP ψi,U (x) = 1 for x ∈ Aj+1. We may extend each multiply all ¯ψi,U with a suitable cut-off function so that P ¯ψi,U (x) ≤ 1 for all x ∈ Kj+1. Now we let f (j+1) = f (j) +Xi,U ιUi (f (j) j+1 ¯ψi,U ) − ιUj+1 (f (j) j+1 ¯ψi,U ). By construction, f (j) j+1 K ◦ f (j+1) − f (j) ∈ N as desired. Since only i > j + 1 appear in the sum, f (j+1) ¯ψi,U is continuous and supported in a compact subset of ¯ψi,U ) ∈ N , so = f (j) i ∩ U with U ⊆ Ui ∩ Uj+1, U ∈ U. Hence ιUi(f (j) ¯ψi,U ) − ιUj+1 (f (j) j+1 j+1 i i 56 ALCIDES BUSS AND RALF MEYER for i < j + 1. We get ¯ψi,U (x)(cid:19). This has supremum norm less than 2j+1−mǫ because 1−Pi,U j+1(x) ·(cid:18)1 −Xi,U j+1 (x) = f (j) f (j+1) ¯ψi,U (x) vanishes where is j+1(x)k ≥ 2j+1−mǫ and is at most 1 everywhere else. The support of f (j+1) kf (j) still contained in Kj+1 by construction. j+1 For i > j + 1, we get f (j+1) i = f (j) i +XU j+1 · ¯ψi,U . f (j) This still has support Ki because ¯ψi,U is supported there. This completes the induction step and thus the proof. (cid:3) Remark B.4. If X is Hausdorff, then a partition-of-unity argument as in the proof of [4, Theorem 2.13] shows that ker(E) is the linear span without closure of ιU (f ) − ιV (f ) with U, V ∈ U. Hence this linear span is already closed for the natural topology onLU∈U S(U, B). Convergent infinite series are needed to generate ker E from ιU (f ) − ιV (f ) with U, V ∈ U. This happens in simple examples, such as the space X = [0, 1] ⊔(0,1] [0, 1] discussed in Section 8 with the trivial bundle C and the standard open cover by two Hausdorff open subsets with their intersection (0, 1]. B.1. Proof of Theorem 5.5. We apply Proposition B.2 to X = L, the cover (Lt)t∈S, and the given Fell bundle B as in the statement of Theorem 5.5. The subsets B∗ and B1∗B2 for bounded subsets B, B1, B2 ⊆ S(L, B) are again bounded; this is routine to check. Thus S(L, B) is a bornological ∗-algebra. (The continuity of the operations for the "inductive limit topology" is also known but somewhat more difficult.) We are going to cite some results of [31] below, which follow from the Disintegra- tion Theorem and the Morita Equivalence Theorem. We assume that they hold for the Fell bundle B in question and its restriction to G; this is not yet proved in the literature, see the discussion before Theorem 5.5. Remark B.6 sketches a slightly more complicated proof that uses only the Morita Equivalence Theorem, that is, the assumptions in Theorem 5.5. Lemma B.5. The C∗-algebra C∗(L, B) is the completion of S(L, B) in the maxi- mal bounded C∗-seminorm. Proof. Usually, C∗(L, B) is defined as the completion of S(L, B) in the maxi- mal C∗-seminorm that is bounded with respect to the I-norm, a certain norm on S(L, B). [31, Corollaire 4.8] shows that a representation of S(L, B) that is con- tinuous with respect to the "inductive limit topology" is bounded for the I-norm. Hence a C∗-seminorm on S(L, B) is continuous with respect to the "inductive limit topology" if and only if it is bounded with respect to the I-norm. The topology on S(L, B) called "inductive limit topology" in [31] is really the quotient topology induced by the inductive limit topology on LU∈U S(U, B), where U is the set of all Hausdorff open subsets of L andLU∈U S(U, B) is viewed as the inductive limit of the Banach subspaces LU∈F S0(KU , B) where F is a finite subset of U and KU ⊆ U for U ∈ F are compact subsets. As we remarked above, a seminorm is continuous in this sense if and only if it is bounded in the canonical bornology on S(L, B) introduced in Appendix B. (cid:3) Let D := Lt∈S S(Lt, B). This carries a canonical direct sum bornology as in Appendix B. The Fell bundle operations turn it into a ∗-algebra. The multiplication INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 57 and involution are bounded, so we even have a bornological ∗-algebra. The map E : D → S(L, B) from Proposition B.2 is a bounded ∗-homomorphism. Since E is a bornological quotient map by Proposition B.2, a C∗-seminorm p on S(L, B) is bounded if and only if p ◦ E is a bounded C∗-seminorm on D. A bounded C∗-seminorm on D is of the form p ◦ E for a C∗-seminorm p on S(L, B) if and only if it vanishes on the kernel of E. By Proposition B.2, a bounded seminorm on D vanishes on ker E if and only if it vanishes on ιt(f )−ιu(f ) for all f ∈ S(Lt, B), t, u ∈ S, t ≤ u. Thus C∗(L, B) is isomorphic to the completion of D in the maximal C∗-seminorm q on D that is bounded and vanishes on ιt(f ) − ιu(f ) for all f, t, u as above. The restriction of this C∗-seminorm q to S(G, B) ⊆ D is a bounded C∗-seminorm. Since C∗(G, B) is defined as the completion of S(G, B) with respect to the maxi- mal bounded C∗-seminorm on S(G, B), q extends to a C∗-seminorm on C∗(G, B). Since q(f )2 = q(f ∗ ∗ f ) for f ∈ S(Lt, B), the restriction of q to S(Lt, B) is domi- nated by the Hilbert module norm from C∗(Lt, B). Thus q automatically extends f ∈ C∗(Lt, B), t, u ∈ S, t ≤ u because S(Lt, B) is dense in C∗(Lt, B). Conversely, a t, u ∈ S, t ≤ u restricts to a C∗-seminorm q on D that annihilates ιt(f ) − ιu(f ) to the sumLt∈S C∗(Lt, B)t∈S. Furthermore, q still annihilates ιt(f ) − ιu(f ) for all C∗-seminorm onLt∈S C∗(Lt, B) that annihilates ιt(f )−ιu(f ) for all f ∈ C∗(Lt, B), for all f ∈ S(Lt, B), t, u ∈ S, t ≤ u. Since D is dense in Lt∈S C∗(Lt, B), this says that C∗(L, B) is isomorphic to the completion of Lt∈S C∗(Lt, B) in the max- imal C∗-seminorm that annihilates ιt(f ) − ιu(f ) for all f ∈ C∗(Lt, B), t, u ∈ S, t ≤ u. This is exactly the definition of the section C∗-algebra of the Fell bundle C∗(Lt, B)t∈S over S. This finishes the proof of Theorem 5.5. Remark B.6. We may also prove Theorem 5.5 without Lemma B.5, using the usual definition of C∗(L, B) involving the I-norm on D. This variant of the proof has the advantage that it does not require the Disintegration Theorem. We still need the Morita equivalence theorem for our Fell bundles, however, so that our inner products are positive and generate the expected ideals. We only explain the new points in this alternative proof. The I-norm on S(L, B) restricts to the I-norm on S(G, B). Consider a C∗-seminorm q on D that annihi- lates ιt(f ) − ιu(f ) for all f ∈ S(Lt, B), t, u ∈ S, t ≤ u and satisfies q(f ) ≤ kf kI for all f ∈ S(G, B). Then q(f ) = q(f ∗ ∗ f )1/2 ≤ kf ∗ ∗ f k1/2 I ≤ kf kI for all f ∈ S(Lt, B), t ∈ S. Thus q is bounded with respect to our bornology as well, so it factors as q ◦ E for a bounded seminorm q on S(L, B) by Proposition B.2. This seminorm satisfies q(f ) ≤ kf kI for all f ∈ S(Lt, B), t ∈ S. But then q(f ) ≤ kf kI follows for all f ∈ S(L, B), t ∈ S. References [1] Jean Bénabou, Introduction to bicategories, Reports of the Midwest Category Seminar, Springer, Berlin, 1967, pp. 1 -- 77, doi: 10.1007/BFb0074299. MR 0220789 [2] Nicolas Bourbaki, Topologie générale. Chapitres 1 à 4, Éléments de mathématique, Hermann, Paris, 1971. MR 0358652 [3] Alcides Buss and Ruy Exel, Twisted actions and regular Fell bundles over inverse semigroups, Proc. Lond. Math. Soc. (3) 103 (2011), no. 2, 235 -- 270, doi: 10.1112/plms/pdr006.MR 2821242 group- at oids, http://www.theta.ro/jot/archive/2012-067-001/2012-067-001-007.pdf. MR 2881538 semigroups no. twisted 153 -- 205, étale available over Theory inverse 67 Fell bundles and 1, (2012), [4] , J. Operator [5] Alcides Buss and Ralf Meyer, Crossed products for actions of crossed modules on C∗-algebras, J. Noncommut. Geom. (2016), accepted. arXiv: 1304.6540. [6] Alcides Buss, Ralf Meyer, and Chenchang Zhu, A higher category approach to twisted (2) 56 (2013), no. 2, 387 -- 426, actions on C∗-algebras, Proc. Edinb. Math. Soc. doi: 10.1017/S0013091512000259. MR 3056650 58 ALCIDES BUSS AND RALF MEYER [7] Jérôme Chabert and Siegfried Echterhoff, Twisted equivariant KK-theory and the Baum -- Connes conjecture for group extensions, K-Theory 23 (2001), no. 2, 157 -- 200, doi: 10.1023/A:1017916521415. MR 1857079 [8] Lisa Orloff Clark, Astrid an Huef, and Iain Raeburn, The equivalence relations of local homeomorphisms and Fell algebras, New York J. Math. 19 (2013), 367 -- 394, available at http://nyjm.albany.edu/j/2013/19_367.html. MR 3084709 [9] Valentin Deaconu, Groupoids associated with endomorphisms, Trans. Amer. Math. Soc. 347 (1995), no. 5, 1779 -- 1786, doi: 10.2307/2154972. MR 1233967 [10] Claire Debord, Holonomy groupoids of singular foliations, J. Differential Geom. 58 (2001), no. 3, 467 -- 500, available at http://projecteuclid.org/euclid.jdg/1090348356.MR 1906783 [11] Ruy Exel, Partial actions of groups and actions of inverse semigroups, Proc. Amer. Math. Soc. 126 (1998), no. 12, 3481 -- 3494, doi: 10.1090/S0002-9939-98-04575-4. MR 1469405 [12] [13] , Inverse semigroups and combinatorial C ∗-algebras, Bull. Braz. Math. Soc. (N.S.) 39 (2008), no. 2, 191 -- 313, doi: 10.1007/s00574-008-0080-7. MR 2419901 , Noncommutative Cartan subalgebras of C ∗-algebras, New York J. Math. 17 (2011), 331 -- 382, available at http://nyjm.albany.edu/j/2011/17-17.html. MR 2811068 [14] Philip Green, The local structure of twisted covariance algebras, Acta Math. 140 (1978), no. 3-4, 191 -- 250, doi: 10.1007/BF02392308. MR 0493349 [15] Rohit Dilip Holkar, Topological construction of C∗-correspondences for groupoid C∗-algebras, Ph.D. Thesis, Georg-August-Universität Göttingen, 2014. [16] Gennadi G. Kasparov and Georges Skandalis, Groups acting on buildings, operator K-theory, and Novikov's conjecture, K-Theory 4 (1991), no. 4, 303 -- 337, doi: 10.1007/BF00533989. MR 1115824 [17] Alex Kumjian, Fell bundles over groupoids, Proc. Amer. Math. Soc. 126 (1998), no. 4, 1115 -- 1125, doi: 10.1090/S0002-9939-98-04240-3. MR 1443836 [18] Mark V. Lawson, Inverse semigroups: the theory of partial symmetries, World Scientific Publishing Co., River Edge, NJ, 1998. [19] Tom Leinster, Basic Bicategories (1998), eprint. arXiv: math/9810017. [20] Ralf Meyer bootstrap 215 -- 252, the http://nbn-resolving.de/urn:nbn:de:hbz:6-10569452982. MR 2545613 and Ryszard class, Münster Nest, J. Math. C ∗-Algebras (2009), over 2 topological available spaces: at [21] Ralf Meyer and topology, http://www.tac.mta.ca/tac/volumes/30/55/30-55abs.html. Theory Categ. 30 Chenchang Appl. Zhu, Groupoids (2015), categories in 1906 -- 1998, with available pre- at [22] Ieke Moerdijk, Orbifolds as groupoids: an introduction, Orbifolds in mathematics and physics (Madison, WI, 2001), Contemp. Math., vol. 310, Amer. Math. Soc., Providence, RI, 2002, pp. 205 -- 222, doi: 10.1090/conm/310/5405. MR 1950948 [23] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomor- phism for groupoid C ∗-algebras, J. Operator Theory 17 (1987), no. 1, 3 -- 22, available at http://www.theta.ro/jot/archive/1987-017-001/1987-017-001-001.pdf. MR 873460 [24] Paul S. Muhly and Dana P. Williams, Equivalence and disintegration theorems for Fell bundles and their C ∗-algebras, Dissertationes Math. (Rozprawy Mat.) 456 (2008), 1 -- 57, doi: 10.4064/dm456-0-1. MR 2446021 [25] , Renault's equivalence theorem for groupoid crossed products, NYJM Monographs, vol. 3, State University of New York University at Albany, Albany, NY, 2008. available at http://nyjm.albany.edu/m/2008/3.htm MR 2547343 [26] May Nilsen, C ∗-bundles and C0(X)-algebras, Indiana Univ. Math. J. 45 (1996), no. 2, 463 -- 477, doi: 10.1512/iumj.1996.45.1086. MR 1414338 [27] Radu Popescu, Equivariant E-theory for groupoids acting on C ∗-algebras, J. Funct. Anal. 209 (2004), no. 2, 247 -- 292, doi: 10.1016/j.jfa.2003.04.001. MR 2044224 [28] John Quigg and Nándor Sieben, C ∗-actions of r-discrete groupoids and inverse semigroups, J. Austral. Math. Soc. Ser. A 66 (1999), no. 2, 143 -- 167, doi: 10.1017/S1446788700039288. MR 1671944 [29] Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Mathematical Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. MR 1634408 [30] Jean Renault, A groupoid approach to C∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer, Berlin, 1980. doi: 10.1007/BFb0091072 MR 584266 [31] [32] , Operator Représentation Theory des 18 produits (1987), croisés no. J. http://www.theta.ro/jot/archive/1987-018-001/1987-018-001-005.pdf. MR 912813 available 1, d'algèbres 67 -- 97, de groupoïdes, at , Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bull. 61 (2008), 29 -- 63, available at http://www.maths.tcd.ie/pub/ims/bull61/S6101.pdf. MR 2460017 INVERSE SEMIGROUP ACTIONS ON GROUPOIDS 59 [34] , [33] Nándor Sieben, C ∗-crossed products by partial actions and actions of inverse semigroups, J. Austral. Math. Soc. Ser. A 63 (1997), no. 1, 32 -- 46, doi: 10.1017/S1446788700000306.MR 1456588 ac- at tions, http://www.theta.ro/jot/archive/1998-039-002/1998-039-002-009.pdf. MR 1620499 C ∗-crossed J. Operator by (1998), twisted no. available semigroup 2, 361 -- 393, inverse products Theory 39 [35] Giorgio Trentinaglia, Tannaka duality for proper Lie groupoids, Ph.D. Thesis, Utrecht Uni- versity, 2008. arXiv: 0809.3394. [36] Jean-Louis K-theory, http://www.mathematik.uni-bielefeld.de/documenta/vol-09/26.html. MR 2117427 Non-Hausdorff Math. 9 groupoids, (2004), proper 565 -- 597, actions available Tu, Doc. and at [37] Alan Weinstein, Linearization of regular proper groupoids, J. Inst. Math. Jussieu 1 (2002), no. 3, 493 -- 511, doi: 10.1017/S1474748002000130. MR 1956059 linearization, affin- (4) 39 (2006), no. 5, 841 -- 869, [38] Nguyen Tien Zung, Proper groupoids and momentum maps: ity, and convexity, Ann. Sci. École Norm. Sup. doi: 10.1016/j.ansens.2006.09.002. MR 2292634 E-mail address: [email protected] Departamento de Matemática, Universidade Federal de Santa Catarina, 88.040-900 Florianópolis-SC, Brazil E-mail address: [email protected] Mathematisches Institut, Georg-August-Universität Göttingen, Bunsenstrasse 3 -- 5, 37073 Göttingen, Germany
1706.00554	1	1706	2017-06-02T04:53:01	Quantum groups, property (T), and weak mixing	[ "math.OA", "math.FA", "math.QA" ]	For second countable discrete quantum groups, and more generally second countable locally compact quantum groups with trivial scaling group, we show that property (T) is equivalent to every weakly mixing unitary representation not having almost invariant vectors. This is a generalization of a theorem of Bekka and Valette from the group setting and was previously established in the case of low dual by Daws, Skalsi, and Viselter. Our approach uses spectral techniques and is completely different from those of Bekka--Valette and Daws--Skalski--Viselter. By a separate argument we furthermore extend the result to second countable nonunimodular locally compact quantum groups, which are shown in particular not to have property (T), generalizing a theorem of Fima from the discrete setting. We also obtain quantum group versions of characterizations of property (T) of Kerr and Pichot in terms of the Baire category theory of weak mixing representations and of Connes and Weiss in term of the prevalence of strongly ergodic actions.	math.OA	math	QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING MICHAEL BRANNAN AND DAVID KERR Abstract. For second countable discrete quantum groups, and more generally sec- ond countable locally compact quantum groups with trivial scaling group, we show that property (T) is equivalent to every weakly mixing unitary representation not having almost invariant vectors. This is a generalization of a theorem of Bekka and Valette from the group setting and was previously established in the case of low dual by Daws, Skalsi, and Viselter. Our approach uses spectral techniques and is completely different from those of Bekka -- Valette and Daws -- Skalski -- Viselter. By a separate argument we furthermore extend the result to second countable nonuni- modular locally compact quantum groups, which are shown in particular not to have property (T), generalizing a theorem of Fima from the discrete setting. We also ob- tain quantum group versions of characterizations of property (T) of Kerr and Pichot in terms of the Baire category theory of weak mixing representations and of Connes and Weiss in term of the prevalence of strongly ergodic actions. 1. Introduction Introduced by Kazhdan in the 1960s for the purpose of showing that many lattices are finitely generated, Property (T) has come to play a foundational role in the study of rigidity in Lie groups, ergodic theory, and von Neumann algebras through work of Margulis, Zimmer, Connes, Popa, and others [34, 5, 28]. Over the last twenty-five years it has been extended in stages to the realm of quantum groups, first via Kac algebras [15], then in the algebraic [3] and discrete [13, 24] settings, and finally in the general framework of locally compact quantum groups as defined by Kusterman and Vaes [10]. In one notable recent application, Arano showed in [1, 2] that the Drinfeld double of a q-deformation of compact simple lie group has property (T) and that this implies that the duals of these q-deformations have a central version of property (T), a fact which has inspired progress in the theory of C∗-tensor categories and underpins Popa and Vaes's construction of subfactors with property (T) standard invariant that do not come from groups [29, 26]. By definition, a locally compact group G does not have property (T) if it admits a unitary representation which does not have a nonzero invariant vector (ergodic- ity) but does have a net of unit vectors which is asymptotically invariant on each group element (having almost invariant vectors). Because ergodicity has poor per- manence properties, it can be hard to leverage this definition so as to obtain global information about the representation theory of a group without property (T), and in particular to determine to what extent the kind of flexible behaviour exhibited by 1 2 MICHAEL BRANNAN AND DAVID KERR amenable groups persists in this more general setting. Bekka and Valette provided a remedy for this in the separable case by showing that one can equivalently replace ergodicity above with weak mixing, which is characterized by the absence of nonzero finite-dimensional subrepresentions, or alternatively by the ergodicity of the tensor product of the representation with its conjugate [4]. This leads for example to a short proof of a theorem of Wang that characterizes property (T) in terms of the isolation of finite-dimensional representations in the spectrum ([4], Section 4) and a stream- lined proof of the Connes -- Weiss characterization of property (T) in terms of strongly ergodic probability-measure-preserving actions ([5], Section 6.3). Using the fact that weak mixing is preserved under tensor products with arbitrary representations, Kerr and Pichot applied the Bekka -- Valette theorem to show that if a second countable locally compact group does not have property (T) then within the set of all unitary representations of the group on a fixed separable infinite-dimensional Hilbert space the weakly mixing ones form a dense Gδ in the weak topology [20], strengthening a result of Glasner and Weiss that gave the same conclusion for ergodic representations [14]. The idea is that any representation will approximately absorb a representation with almost invariant vectors under tensoring (since locally it is as if we were tensoring with the trival representation) and so such a tensor product will be "close" to the original representation while also inheriting any properties of the second one that are preserved under tensoring, such as weak mixing. By a similar principle requiring a more subtle implementation, Kerr and Pichot also established an analogous conclusion for the measure-preserving actions of the group on a fixed standard atomless probability space. Using the theory of positive-definite functions as in [16, 27], Daws, Skalsi, and Visel- ter demonstrated in [11] that the conclusion of the Bekka -- Valette theorem also holds for second countable discrete unimodular quantum groups with low dual, and as an ap- plication they derive analogues of the Connes -- Weiss theorem and the representation- theoretic Kerr -- Pichot theorem. Low dual is the rather restrictive assumption that there is a bound on the dimensions of the irreducible representations of the quantum group, and the authors of [11] wonder, somewhat pessimistically, whether it can be removed. In the present paper we show that the Bekka -- Valette and Kerr -- Pichot the- orems actually hold for all second countable discrete quantum groups, and even more generally for all second countable locally compact quantum groups with trivial scal- ing group (Theorems 4.8 and 4.9) as well as for all second countable nonunimodular locally compact quantum groups (Theorem 6.3). The methods of Daws, Skalsi, and Viselter can then also be applied to extend their version of the Connes -- Weiss theorem to all second countable locally compact quantum groups with trivial scaling group (Theorem 5.1). Our approach is completely different from those of Daws -- Skalsi -- Viselter and Bekka -- Valette and consists in applying the quantum group version of Wang's characterization of property (T) mentioned above in order to reduce the problem to a purely spectral question concerning C∗-algebras. In Theorems 3.7 and 3.8 we prove that the following QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 3 hold for a separable unital C∗-algebra A and a fixed separable infinite-dimensional Hilbert space H : (i) if the spectrum of A contains no isolated finite-dimensional representations then the set of weakly mixing unital representations of A on H is a dense Gδ, and (ii) if the set of finite-dimensional representations in the spectrum of A is nonempty and contains only isolated points then the set of weakly mixing unital repre- sentations of A on H is closed and nowhere dense. A version of the argument establishing (i) for unitary representations of countable discrete groups has also been included in the book [19] by Li and the second author. Theorems 4.8 and 4.9 then follow from (i) and (ii) whenever Wang's characterization of property (T) holds in the quantum group context, and this is known to be the case when the scaling group is trivial (see Section 4). By a completely different argument we also prove in Theorem 6.3 that the conclusions Theorems 4.8 and 4.9 are valid for second countable nonunimodular locally compact quantum groups, which we show in particular not to have property (T), generalizing a result of Fima from the discrete case [13]. We begin in Section 2 by reviewing some of the basic theory of locally compact quantum groups and their unitary representations as developed by Kustermans and Vaes [22, 23, 30, 21]. In Section 3 we study weak mixing for C∗-algebra representa- tions and establish the two key spectral results (i) and (ii) concerning separable unital C∗-algebras. In Section 4 we discuss weak mixing and property (T) for quantum groups, record the quantum group incarnation of Wang's theorem, and then establish our versions of the Bekka -- Valette and Kerr -- Pichot theorems. Section 5 contains the Connes -- Weiss-type dynamical characterization of property (T). Finally, the nonuni- modular case is treated in Section 6. Acknowledgements. M.B. was partially supported by NSF grant DMS-1700267. D.K. was partially supported by NSF grant DMS-1500593. 2. Preliminaries For a C∗-algebra A we write M(A) for its multiplier algebra. A representation of A is understood to mean a ∗-homomorphism from A into the C∗-algebra of bounded linear operators on some Hilbert space. When working with tensor products of Hilbert spaces H and K , we denote by Σ the tensor flip map from H ⊗ K to K ⊗ H . For linear operators on multiple tensor products, we use leg notation. For example, if U is a unitary operator on a Hilbert space tensor product H ⊗ K we write U13 for the unitary operator on a Hilbert space tensor product of the form H ⊗ J ⊗ K which is given by V (U ⊗ id)V −1 where V is the shuffle map H ⊗ K ⊗ J → H ⊗ J ⊗ K defined on elementary tensors by ξ ⊗ ζ ⊗ κ 7→ ξ ⊗ κ ⊗ ζ, i.e., V = idH ⊗ Σ. 4 MICHAEL BRANNAN AND DAVID KERR 2.1. Locally compact quantum groups. Our main references for generalities on locally compact quantum groups are [22, 23, 30]. Formally speaking, a (von Neumann algebraic) locally compact quantum group is a von Neumann algebra with coassociative coproduct and left and right Haar weights, but as usual we use the simple notation G so that we can conveniently and suggestively refer to the various objects that are canonically attached to it just as one does for locally compact groups, although there is no longer anything like an underlying group. The von Neumann algebra itself is thus written L∞(G), and the coproduct is a unital normal ∗-homomorphism ∆ : L∞(G) → L∞(G)⊗L∞(G) satisfying the coassociativity condition (∆ ⊗ id)∆ = (id ⊗ ∆)∆. The left and right Haar weights are normal semifinite weights ϕ and ψ on L∞(G) such that for every ω ∈ L∞(G)+ ∗ one has for all a ∈ L∞(G)+ with ϕ(a) < ∞ and ϕ((ω ⊗ id)∆(a)) = ϕ(a)ω(1) ψ((id ⊗ ω)∆(a)) = ψ(a)ω(1) for all a ∈ L∞(G)+ with ψ(a) < ∞. The predual of L∞(G) is written as L1(G), and becomes a completely contractive Banach algebra with respect to the convolution product ω1 ⋆ ω2 = (ω1 ⊗ ω2) ◦ ∆, ω1, ω2 ∈ L1(G). Associated to G is a canonical weakly dense sub-C∗-algebra of L∞(G), written C0(G), which plays the role of the C∗-algebra of continuous functions vanishing at infinity in the case of ordinary groups. We say that G is second countable if C0(G) is separable. The coproduct restricts to a unital ∗-homomorphism ∆ : C0(G) → M(C0(G) ⊗ C0(G)). The algebras C0(G) and L∞(G) are standardly represented on the GNS Hilbert space L2(G) associated to the left Haar weight. In the case of a locally compact group, the notations L∞(G), L1(G), C0(G), and L2(G) have their ordinary meaning. k·k There is a (left) fundamental unitary operator W on L2(G) ⊗ L2(G) which satisfies the pentagonal relation W12W13W23 = W23W12 and unitarily implements the coproduct ∆ on L∞(G) via the formula ∆(x) = W ∗(1 ⊗ x)W . Using W one has C0(G) = {(id ⊗ ω)W : ω ∈ B(L2(G))∗} , and one can define the antipode of G as the (generally only densely defined) linear operator S on C0(G) (or L∞(G)) satisfying the identity (S⊗id)W = W ∗. The antipode admits a polar decomposition S = R◦τ−i/2 where R is an antiautomorphism of L∞(G) (the unitary antipode) and {τt}t∈R is a one-parameter group of automorphisms (the scaling group). In the case of a locally compact group, the scaling group is trivial and the antipode is the antiautomorphism sending a function f ∈ C0(G) to the function s 7→ f (s−1). Using the antipode S one can endow the convolution algebra L1(G) with a densely defined involution by considering the norm- dense subalgebra L1 ♯ (G) of L1(G) consisting of all ω ∈ L1(G) for which there exists an QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 5 k·k equivalent to C0(G) being a direct sum of matrix algebras. cally identified with the original quantum group G. One says that a locally compact ω♯ ∈ L1(G) with hω♯, xi = hω, S(x)∗i for each x ∈ D(S). It is known from [21] and Section 2 of [23] that L1 ♯ (G) is an involutive Banach algebra with involution ω 7→ ω♯ and norm kωk♯ = max{kωk,kω♯k}. Associated to any locally compact quantum group G is its dual locally compact quantum group bG, whose associated algebras, coproduct, and fundamental unitary are ⊆ B(L2(G)), L∞(bG) = C0(bG)′′, given by C0(bG) = {(ω ⊗ id)W : ω ∈ B(L2(G))∗} ∆(x) = W ∗(1 ⊗ x) W , and W = ΣW ∗Σ. Then in fact W ∈ M(C0(G) ⊗ C0(bG)), and the Pontryagin duality theorem asserts that the bidual quantum group bbG is canoni- quantum group G is compact if C0(G) is unital, and discrete if bG is compact, which is For a locally compact quantum group G, we can always assume that the left and right Haar weights are related by ψ = ϕ◦ R, where R is the unitary antipode. If the left and right Haar weights ϕ and ψ of G coincide then we say that G is unimodular. In general, the failure of ψ to be left-invariant is measured by the modular element, which is a strictly positive element δ affiliated with L∞(G) satisfying the identities ∆(δ) = δ ⊗ δ and ψ(·) = ϕ(δ1/2 · δ1/2). Compact quantum groups are always unimodular, and the corresponding Haar weight can always be chosen to be a state. Although discrete groups are always unimodular, discrete quantum groups need not be. We recall that a discrete quantum group G is said to be of Kac type (or a Kac algebra) if it is unimodular, which is equivalent to the Haar state on bG being a trace. 2.2. Unitary representations. Definition 2.1. A unitary representation of a locally compact quantum group G on a Hilbert space H is a unitary U ∈ M(C0(G)⊗ K (H )) ⊆ B(L2(G)⊗ H ) such that (∆ ⊗ id)(U) = U13U23. In the above definition one can replace M(C0(G)⊗ K (H )) with the larger algebra L∞(G)⊗B(H ), for if U is a unitary in the latter which satisfies (∆⊗ id)(U) = U13U23 then U automatically belongs to the former (see for example Theorem 4.12 of [6]). Associated to a unitary representation U ∈ M(C0(G) ⊗ K (H )) is an adjointable operator on the Hilbert module C0(G)⊗ H which we write using the boldface version U of the symbol in question. The relation between U and U is given by hU(a ⊗ ξ), b ⊗ ζi = b∗(id ⊗ ωξ,ζ)(U)a for all a, b ∈ C0(G) and ξ, ζ ∈ H , where ωξ,ζ is the vector functional x 7→ hxξ, ζi. Associated to G are two distinguished unitary representations, the one-dimensional trivial representation 1G ∈ M(C0(G)) given by the unit of L∞(G), and the left regular representation given by the fundamental unitary W ∈ M(C0(G) ⊗ C0(bG)). Two unitary representations U ∈ M(C0(G)⊗K (H1)) and V ∈ M(C0(G)⊗K (H2)) of G are (unitarily) equivalent if there is a unitary isomorphism u : H1 → H2 such that 6 MICHAEL BRANNAN AND DAVID KERR V = (id ⊗ Ad u)(U). A subrepresentation of a unitary representation U ∈ M(C0(G) ⊗ K (H )) is a unitary representation of the form Q = (1⊗P )U(1⊗P ) ∈ L∞(G)⊗B(H0) where H0 is a closed subspace of H , P is the orthogonal projection of H onto H0, and 1 ⊗ P commutes with U. In this case, we write Q ≤ U. The subrepresentation is said to be finite-dimensional if H0 is finite-dimensional. Let G be a unimodular locally compact quantum group. Let U ∈ M(C0(G) ⊗ K (H )) be a unitary representation of G. Write H for the conjugate of H , i.e., the Hilbert space which is the same as H as an additive group but with the scalar multiplication (c, ξ) 7→ ¯cξ for c ∈ C and inner product hξ, ζiH = hζ, ξiH . Letting T : B(H ) → B(H ) be the transpose map T (a)(ξ) = a∗(ξ), we define the conjugate of U, written U, to be the unitary representation (R ⊗ T )(U) ∈ M(C0(G) ⊗ K (H ))). The tensor product of two unitary representations U ∈ M(C0(G) ⊗ K (H )) and V ∈ M(C0(G) ⊗ K (K )) is the unitary representation U ⊙ V := U12V13 ∈ M(C0(G) ⊗ K (H ⊗ K )) ⊆ L∞(G)⊗B(H )⊗B(K ). There is a bijective correspondence between unitary representations U ∈ M(C0(G)⊗ ♯ (G) → B(H) ([21], Corol- K (H )) and nondegenerate ∗-representations πU : L1 lary 2.13). This correspondence is given by πU (ω) = (ω ⊗ id)U ∈ B(H), ω ∈ L1 ♯ (G). k·k Let C u ♯ (G). This is a univer- k·k ♯ (G)) and C0(G) = λ(L1 C0(bG) = λ(L1 At the level of ∗-representations of L1 ♯ (G), the trivial representation 1G corresponds to the ∗-character ω 7→ ω(1), and the left regular representation is written as ω 7→ λ(ω) = (ω ⊗ id)W ∈ C0(bG) ⊆ B(L2(G)). As expected, we have the dual relations where λ is the left regular representation of bG. sal version of C0(bG) which encodes the unitary representation theory of G (since ♯ (bG)) 0 (bG) denote the universal enveloping C∗-algebra of L1 its nondegenerate representations are in bijective correspondence with the nonde- generate ∗-representations of L1 In par- ticular, the left regular representation W gives rise to a surjective representation λ : C u (dual) counit εu : C u ∗-representations of L1 0 (bG) → C0(bG) ⊆ B(L2(G)), and the trivial representation gives rise to the 0 (bG) → C. We will generally use the same symbols to denote 0 (bG)) 0 (bG), ∆u) into a universal C∗-algebraic locally compact which can be used to turn (C u quantum group. For our purposes, we only need the fact that ∆u allows one to express the tensor product U ⊙ V of two unitary representations U ∈ M(C0(G)⊗ K (H )) and 0 (bG) admits a coproduct ∆u : C u 0 (bG). 0 (bG) → M(C u ♯ (G) as bounded Hilbert space operators). ♯ (G) and their unique extensions to C u 0 (bG)⊗ C u As was shown in [21], C u QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 7 V ∈ M(C0(G) ⊗ K (K )) in terms of the representation (πU ⊗ πV ) ◦ σ ∆u : C u 0 (bG) → 0 (bG) ⊗ C u B(H ⊗ K ), where σ denotes the tensor flip map on C u 0 (bG) (or of ∗-representations Finally, note that at the level of representations of C u of L1 ♯ (G)) the notions of subrepresentation and unitary equivalence of unitary repre- sentations reduce to their standard meanings. Indeed, given a unitary representation U ∈ M(C0(G)⊗K (H )) and a projection P in B(H ), the projection 1⊗P commutes 0 (bG)), in which case the representation with U if and only if P commutes with πU (C u 0 (bG) → B(P H ) associated to Q = (1⊗ P )U(1⊗ P ) is given by a 7→ P πU (a)P . πQ : C u Similarly, if V ∈ M(C0(G)⊗K (K )) is another unitary representation, then a unitary isomorphism u : H → K implements an equivalence between U and V if and only if it implements a unitary equivalence between πU and πV in the sense that πV = Ad u◦πU . 0 (bG). 3. Weak mixing and representations of C∗-algebras This section is purely C∗-algebraic and aims to establish two results concerning the prevalence of weak mixing among unital representations of a separable unital C∗- algebra on a fixed Hilbert space (Theorems 3.7 and 3.8). Throughout this section A will denote a separable unital C∗-algebra. For a fixed Hilbert space H , the set of all unital representations of A on H will be written Rep(A, H ). We equip Rep(A, H ) with the point-strong operator topology, which is equivalent to the point-weak operator topology, and also to the point-∗-strong operator topology since the strong and ∗-strong operator topologies agree on the unitary group of B(H ) and A is linearly spanned by its unitaries. representations, will be thought of as actual representations via their representatives, Points in the spectrum bA, while formally defined as equivalence classes of irreducible following convention. The set of finite-dimensional representations in bA will be written bAfin. We begin with a discussion of weak mixing for unital representations of unital C∗- algebras. Definition 3.1. We say that a unital representation of A on a Hilbert space is weakly mixing if it has no nonzero finite-dimensional subrepresentations. Recall that weak mixing for a unitary representation π : G → U (H ) of a group can be expressed in either of the following equivalent ways (see Theorem 2.23 in [19], and note that the countability assumption there is not needed): (i) for every finite set Ω ⊆ H and ε > 0, there exists an s ∈ G such that (ii) π has no nonzero finite-dimensional subrepresentations. hπ(s)ξ, ζi < ε for all ξ, ζ ∈ Ω, Since a unital C∗-algebra is linearly spanned by its unitaries, from (ii) we immediately obtain the following, justifying the terminology of Definition 3.1. 8 MICHAEL BRANNAN AND DAVID KERR Proposition 3.2. A unital representation of A is weakly mixing if and only if its restriction to the unitary group of A is weakly mixing. Next we consider a C∗-algebra version of Zimmer's notion [34] of weak containment for unitary representations of groups. Definition 3.3. Let π : A → B(H ) and ρ : A → B(K ) be unital representations. We write π ≺ ρ if for every finite set Ω ⊆ A, orthonormal set {ξ1, . . . , ξn} ⊆ H , and ε > 0 there is an orthonormal set {ζ1, . . . , ζn} ⊆ K such that hπ(a)ξi, ξii − hρ(a)ζi, ζii < ε for all a ∈ Ω and i = 1, . . . , n. This is the same as the usual notion of weak containment when the representation π is irreducible, but is different in general. In fact π is weakly contained in ρ if and only if π ≺ ρ⊕N. The following is a straightforward consequence of Definition 3.3. The version for unitary group representations was noted in the remark after Proposition H.2 in [18]. Lemma 3.4. Let π : A → B(H ) and ρ : A → B(K ) be unital representations on separable infinite-dimensional Hilbert spaces. Then π ≺ ρ if and only if π ∈ {κ ∈ Rep(A, H ) : κ ∼= ρ}. Denote by WM(A, H ) ⊆ Rep(A, H ) the subcollection of all weakly mixing repre- sentations. Lemma 3.5. Let H be a separable Hilbert space. Then WM(A, H ) is a Gδ in Rep(A, H ). Proof. Write G for the unitary group of A. Take an increasing sequence Ω1 ⊆ Ω2 ⊆ . . . of finite subsets of H with dense union in H . For every n ∈ N define Γn to be the set of all ϕ ∈ Rep(A, H ) such that there exists a u ∈ G satisfying hϕ(u)ξ, ζi < 1/n for all ξ, ζ ∈ Ωn. Then Γn is open, and so the set Γ = T∞ n=1 Γn is a Gδ. By the characterization of weak mixing for unitary group representations described before Proposition 3.2, Γ is precisely the set of all representations whose restriction to G is weakly mixing, which is equal to WM(A, H ) by Proposition 3.2. (cid:3) Lemma 3.6. Suppose that each point in bAfin is isolated in bA. Let ρ : A → B(H ) be a representation in bAfin. Then there exists a weakly mixing representation θ of A on a separable Hilbert space such that ρ ≺ θ. Proof. We may assume that ρ is not the limit of a sequence {πn} of infinite-dimensional representations in bA, for in that case the representation π =L∞ n=1 πn is weakly mixing and ρ ≺ π. Since bA is second countable ([12], Proposition 3.3.4), we can then find a countable neighbourhood basis {Un}n∈N for ρ in bA such that no Un contains an Let n ∈ N. We construct a weakly mixing representation θn in bA as follows. First we argue that Un is uncountable. Suppose that this is not the case. As bA is a Baire space infinite-dimensional representation. QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 9 ([12], Theorem 3.4.13) and Un is open, Un is itself a Baire space. For every ω ∈ Un the singleton {ω} is closed by finite-dimensionality (see Section 3.6 of [12]), and so by the Baire property there exists an ω0 ∈ Un such that {ω0} is open, which means that ω0 is isolated, contradicting our hypothesis. Thus Un is uncountable. We can consequently find a dn ∈ N such that Un contains uncountably many dn-dimensional representations. Fix a Hilbert space Hn of dimension dn. Denote by Irr(A, Hn) the set of irreducible representations in Rep(A, Hn). We observe the following: are equivalent to some element of U is open. (i) Irr(A, Hn) is open in Rep(A, Hn), (ii) every equivalence class in Irr(A, Hn) is closed in Irr(A, Hn), (iii) for every open set U ⊆ Irr(A, Hn) the set of all elements in Irr(A, Hn) which Assertion (iii) is clear. To verify (i), let {πk} be a convergent sequence in Rep(A, Hn) whose terms are not irreducible and let us show that its limit π is not irreducible. For every k choose a nonzero projection Pk ∈ πk(A)′ of rank less than dn. In view of the finite-dimensionality of Hn, we may assume by passing to a subsequence that the sequence {Pk} converges in B(Hn), in which case its limit P is a nonzero pro- jection which commutes with π(A). This means that π is not irreducible, yielding (i). Finally, to verify (ii) we let {πk} be a convergent sequence in Irr(A, Hn) such that for every k there exists a unitary operator Zk which conjugates πk to π1. By the finite-dimensionality of Hn, there is a subsequence {Zkj}j that converges to a unitary operator Z, which must then conjugate the limit of {πk} to π1, yielding (ii). Since open subsets of Polish spaces are themselves Polish spaces ([17], Theorem 3.11), we infer from (i) that Irr(A, Hn) is a Polish space. Assertions (ii) and (iii) then per- mit us to apply a standard selection theorem ([17], Theorem 12.16) which provides a Borel set Bn ⊆ Irr(G, Hn) of representatives for the relation of unitary equivalence. Write Wn for the set of all π ∈ Irr(A, Hn) which, as elements in bA, belong to Un. This is clearly an open set in Irr(A, Hn), and it is uncountable by our choice of dn. Thus Bn ∩ Wn is an uncountable Borel set, which means that it is isomorphic as a Borel space to the unit interval with Lebesgue measure and hence admits an atomless Borel probability measure of full support. Let µn be the push forward of this measure under the inclusion Bn∩ Wn ֒→ Irr(A, Hn), and note that µn(C) = 0 for every unitary equivalence class C in Irr(A, Hn), Setting Yn = Irr(A, Hn), we next consider the Hilbert space L2(Yn, Hn) of (classes of) Hn-valued functions on Yn with inner product hf, gi =ZYnhf (π), g(π)i dµn(π). Finally, we define the representation θn : A → B(L2(Yn, Hn)) by setting (θn(a)ζ)(π) = π(a)ζ(π) for a ∈ A, ζ ∈ L2(Yn, Hn), and π ∈ Yn. 10 MICHAEL BRANNAN AND DAVID KERR We next verify that θn is weakly mixing. Suppose that this is not the case. Then there is a finite-dimensional irreducible representation π : A → B(H ) and an isomet- ric operator Z : H → L2(Yn, Hn) such that Zπ(a) = θn(a)Z for all a ∈ A. Then for a.e. ρ ∈ Yn and every a ∈ A and η ∈ H we have (Zπ(a)η)(ρ) = (θi(a)Zη)(ρ) = ρ(a)((Zη)(ρ)). so that the operator Zρ : H → Hn given by Zρη = (Zη)(ρ) satisfies Zρπ(a) = ρ(a)Z for all a ∈ A. As Z is isometric, the operator Zρ must be nonzero for all ρ in a nonull subset of Yn. But each such ρ is equivalent to π by irreducibility, contradicting the fact that the measure of every unitary equivalence class is zero. Therefore θn is weakly mixing. Now set θ = L∞ n=1 θn. Then θ is weakly mixing since each summand is weakly It remains to show that ρ ≺ θ. Let Ω be a finite subset of A and ε > mixing. 0. Then we can find an n ∈ N such that for every π ∈ Un there is an isometry V : H → Hn such that kV ρ(a) − π(a)V k < ε/2 for all a ∈ Ω. Since bounded sets in B(Hn) are precompact and representations are contractive, we can find an open set U ⊆ Yn with µ(U) > 0 such that for all a ∈ Ω and π, π′ ∈ U one has kπ(a) − π′(a)k < ε/2. Choose a π0 ∈ U and an isometry V : H → Hn such that for all a ∈ Ω we have kV ρ(a) − π0(a)V k < ε/2 and hence kV ρ(a) − π(a)V k < ε for every π ∈ U. Writing 1U for the indicator function of U, we set f = µn(U)−1/21U , which is a unit vector in L2(Yn, Hn). Define an isometry V : H → L2(Yn, Hn) by V ξ = f ⊗ V ξ ∈ L2(Yn, µn) ⊗ Hn ∼= L2(Yn, Hn). Then for all a ∈ Ω and norm-one vectors ξ ∈ H we have k( V ρ(a) − θn(a) V )ξk2 = kf ⊗ V ρ(a)ξ − θn(a)(f ⊗ V ξ)k2 1 1 µn(U)ZU k(V ρ(a) − π(a)V )ξk2 dµn(π) µn(U)ZU ε2 dµn(π) = ≤ = ε2, so that k V ρ(a) − θn(a) V k < ε. We conclude that ρ ≺ θ, as desired. We now come to the main theorems of this section. (cid:3) Theorem 3.7. Suppose that no point in bAfin is isolated in bA. Let H be a separable Hilbert space. Then WM(A, H ) is a dense Gδ in Rep(A, H ). Proof. By Lemma 3.5 it suffices to show the density of WM(A, H ). Let π ∈ Rep(A, H ). By a maximality argument involving the collection of direct sums of finite-dimensional subrepresentations of π, we can write π = π0 ⊕Li∈I πi where π0 is weakly mixing and πi is finite-dimensional for every i ∈ I. By decomposing further we may assume that πi is irreducible for each i ∈ I. By Lemma 3.6, for each i ∈ I we can find a weak mixing QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 11 (cid:3) the desired density. H be a separable infinite-dimensional Hilbert space. Then WM(A, H ) is closed and nowhere dense in Gδ in Rep(A, H ). representation θi on a separable Hilbert space such that πi ≺ θi. Set ρ = π0 ⊕Li∈I θi. Then π ≺ ρ, and ρ acts on a separable Hilbert space since I is countable by the sep- arability of H . It follows by Lemma 3.4 that π belongs to the closure of the set of all κ ∈ Rep(G, H ) such that κ ∼= ρ, and hence to the closure of WM(A, H ), yielding Theorem 3.8. Suppose that bAfin 6= ∅ and each point in bAfin is isolated in bA. Let Proof. By assumption there exists a ρ ∈ bAfin. Let π ∈ Rep(A, H ). Then clearly π ≺ π ⊕ ρ and so by Lemma 3.4 the representation π belongs to the closure of the set of all κ ∈ Rep(G, H ) such that κ ∼= π ⊕ ρ, showing that the complement of WM(A, H ) is dense in Rep(A, H ). Now let π be a representation in Rep(A, H ) which is not weakly mixing. Then we can write π = π0 ⊕ π1 where π1 is finite-dimensional, and we may assume that π1 is irreducible. Now suppose that {ρn} is a sequence in Rep(A, H ) converging to ρ and set ρ = L∞ n=1 ρn. Then π ≺ ρ and hence π1 ≺ ρ, which implies that π1 is a subrepresentation of ρ since π1 is isolated in bAfin ([33], Theorem 1.7). Since π1 is irreducible, there must exist an n ∈ N such that π1 is a subrepresentation of ρn. We deduce from this that π has a neighbourhood in Rep(A, H ) which does not intersect WM(A, H ). This shows that the complement of WM(A, H ) is open and hence completes the proof. (cid:3) 4. Weak mixing and Property (T) for quantum groups We now return to the context of quantum groups and discuss the notions of weak mixing and property (T). Let G be a locally compact quantum group. Definition 4.1. A unitary representation U ∈ M(C0(G) ⊗ K (H )) of G is weakly mixing if it contains no nonzero finite-dimensional subrepresentation. Since finite-dimensionality is preserved under the canonical correspondence between 0 (bG), by Propo- unitary representations of G and nondegenerate representations of C u sition 3.2 we obtain the following. Proposition 4.2. Let U ∈ M(C0(G) ⊗ K (H )) be a unitary representation, let πU be the corresponding representation of C u U be the canonical extension 0 (bG)+. Then U is weakly mixing of πU to a unital representation of the unitization C u 0 (bG)+ is weakly mixing. if and only if the restriction of π+ Let H be a fixed Hilbert space. Write Rep(G, H ) for the collection of all unitary representations of G on H , and equip it with the point-strict topology it inherits as a subset of M(C0(G) ⊗ K (H )). Let WM(G, H ) ⊆ Rep(G, H ) be the subcollection 0 (bG), and let π+ U to the unitary group of C u 12 MICHAEL BRANNAN AND DAVID KERR 0 (bG), H ) for the collection of non- of all weakly mixing representations. Write Rep(C u 0 (bG) on H (which is consistent with our notation in degenerate representations of C u Section 3 for unital C∗-algebras). In Proposition 5.1 of [10] it is shown that the topol- 0 (bG), H ) under the canonical bijection between Rep(G, H ) ogy T induced on Rep(C u 0 (bG), H ) is the point-strict topology. Since the strict topology and the and Rep(C u ∗-strong operator topology coincide on bounded subsets of B(H ), this is the same as 0 (bG), H ). Since the ∗-strong operator the point-∗-strong operator topology on Rep(C u topology and the strong operator topology agree on the unitary group of B(H ), and a unital C∗-algebra is spanned by its unitaries, we therefore have πn → π in the topology n → π+ in the point-strong operator topology T on Rep(C u n and π+ are the canonical unital extensions of πn and π on Rep(C u 0 (bG)+. Combining these observations with Lemma 3.5, we obtain: to the unitization C u 0 (bG), H ) if and only if π+ 0 (bG)+, H ), where π+ Proposition 4.3. Let H be a separable infinite-dimensional Hilbert space. Then WM(G, H ) is a Gδ in Rep(G, H ). Next we recall the definition of property (T). Definition 4.4. Let G be a locally compact quantum group and let U ∈ M(C0(G) ⊗ K (H )) be a unitary representation of G. A vector ξ ∈ H is said to be invariant for U if U(η ⊗ ξ) = η ⊗ ξ for all η ∈ L2(G). We say that U has almost invariant vectors if there is a net {ξi}i of unit vectors in H such that for all η ∈ L2(G), which by Proposition 3.7 of [10] is equivalent to kU(η ⊗ ξi) − η ⊗ ξik → 0 kπU (a)ξi − εu(a)ξik → 0 for all a ∈ C u Definition 4.5. A locally compact quantum group G has property (T) if every unitary representation G having almost invariant vectors has a nonzero invariant vector. 0 (bG). In order to establish the two main results of this section, Theorems 4.8 and 4.9, we require a characterization of property (T) in terms of the isolation of finite-dimensional representations in the spectrum. Specifically, we will need the equivalence of (i) and (iv) in Theorem 4.7 below. For locally compact groups the equivalence between (i), (iii), and (iv) in Theorem 4.7 is due to Wang [33], and is known more generally for locally compact quantum groups with trivial scaling group, although it does not seem to be explicitly stated in this generality in the literature (see Remark 5.3 of [24] and Section 3 of [7]). For quantum groups the idea is to adapt the argument of Bekka, de la Harpe, and Valette for groups in Section 1.2 of [5]. This requires Lemma 4.6, which generalizes a well known fact for locally compact groups. The discrete case of Lemma 4.6 appears in Section 2.5 of [24], although the proof there works more QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 13 generally, as observed in Proposition 7.2 of [9]. See also Section 3 of [7] and the first paragraph of the proof of Proposition 3.5 in [32]. Lemma 4.6. Let U and V be finite-dimensional unitary representations of a second countable locally compact quantum group G with trivial scaling group. Then 1G ≤ U⊙ ¯V if and only if U and V contain a common nonzero subrepresentation. Armed with Lemma 4.6, one can now establish the following result by repeating mutatis mutandis the argument in Section 1.2 of [5], as was done in Section 3 of [7] in the discrete unimodular case. Theorem 4.7. For a second countable locally compact quantum group G with trivial scaling group the following are equivalent: (i) G has property (T), (ii) εu is isolated in the spectrum of C u (iii) every finite-dimensional representation in the spectrum of C u (iv) there exists a finite-dimensional representation in the spectrum of C u 0 (bG), 0 (bG) is isolated, 0 (bG) which is isolated. Theorem 4.8. A second countable locally compact quantum group G with trivial scal- ing group has property (T) if and only if every weakly mixing unitary representation of G fails to have almost invariant vectors. Proof. For the nontrivial direction, if G does not have property (T) then by Theo- rem 4.7 and Lemma 3.6 there is a weakly mixing unitary representation of G with almost invariant vectors. (cid:3) Theorem 4.9. Let G be a second countable locally compact quantum group with trivial scaling group. Let H be a separable infinite-dimensional Hilbert space. If G does not have property (T) then WM(G, H ) is a dense Gδ in Rep(G, H ), while if G has property (T) then WM(G, H ) is closed and nowhere dense in Rep(G, H ). Proof. Apply Theorem 4.7 in conjunction with Theorems 3.7 and 3.8. (cid:3) 5. Property (T) and strongly ergodic actions We establish in Theorem 5.1 a quantum group version of a result of Connes and Weiss [8] for countable discrete groups. It was verified by Daws, Skalski, and Viselter under the additional hypothesis that the quantum group is discrete and has low dual ([11], Theorem 9.3). In fact to obtain the conclusion we can simply apply the argument of Daws, Skalsi, and Viselter by replacing their Theorem 7.3 with our Theorem 4.8, or rather a slight strengthening of the latter in line with Remark 7.4 of [11], as we explain below. An n.s.p. (normal-state-preserving) action G yα (N, σ) is a normal injective unital ∗-homomorphism α : N → L∞(G)⊗N, where N is a von Neumann algebra with a 14 MICHAEL BRANNAN AND DAVID KERR faithful normal state σ, such that (id⊗ α)α = (∆⊗ id)α and (id⊗ σ)α(x) = σ(x)1 for all x ∈ N. We drop the symbol α if we don't need to refer to it explicitly. Let G yα (N, σ) be an n.s.p. action. A bounded net {xi} ⊂ N in said to be asymptotically invariant if for every normal state ω ∈ L1(G) we have (ω⊗ id)(α(xi))− xi → 0 strongly, and trivial if xi − σ(xi)1 → 0 strongly. The action is said to be strongly ergodic if every asymptotically invariant net is trivial. Recall that R denotes the unitary antipode of G, which acts on B(L2(G)) as R(x) = JRx∗JR where JR is the modular conjugation associated to the left Haar weight on L∞(G). We say that a unitary representation U ∈ M(C0(G) ⊗ K (H )) of G is self- conjugate (referred to as condition R in [11]) if there exists an anti-unitary operator J : H → H such that the anti-isomorphism j : B(H ) → B(H ) given by x 7→ Jx∗J ∗ satisfies (R ⊗ j)(U) = U. Let U be a unitary representation of G on a Hilbert space H . Recall that the conjugate representation U on the conjugate Hilbert space H is defined as (R⊗T )(U) where T : B(H ) → B(H ) is the map given by T (b)ξ = b∗ξ. Let V = U ⊙ U be the tensor product of U and U . Write J for the anti-unitary operator on H ⊗ H given on elementary tensors by J(ξ ⊗ ζ) = ζ ⊗ ξ. and let j : B(H ) → B(H ) be the anti- isomorphism given by x 7→ Jx∗J ∗. Let y =Pi∈I ai ⊗ bi be a finite sum of elementary tensors in L∞(G)⊗ B(H ). Since R(aiR(aj)) = ajR(ai) and j(bi ⊗ T (bj)) = bj ⊗ T (bi) for all i, j ∈ I, it follows that the element x = y12[(R ⊗ T )y]13 = Xi,j∈I aiR(aj) ⊗ bi ⊗ T (bj) ∈ L∞(G) ⊗ B(H ) ⊗ B(H ) satisfies (R ⊗ j)(x) = x. Since the maps R⊗ j and R⊗ T are ∗-strongly continuous, multiplication is ∗-strongly continuous on the unit ball of B(L2(G) ⊗ H ⊗ H ), and U is a ∗-strong limit of operators of norm at most one in L∞(G)⊗ B(H ) by Kaplansky density, we conclude that (R ⊗ j)(V ) = V , so that V is self-conjugate. Now if U has almost invariant vectors then so does V , as is easily seen, and if U is weakly mixing and G has trivial scaling group then V is weakly mixing by Theo- rem 3.11 of [32]. It thus follows from Theorem 4.8 that if G has trivial scaling group and does not have property (T) then it admits a weakly mixing self-conjugate unitary representation with almost invariant vectors. This extra self-conjugacy condition is needed in the argument of Daws -- Skalski -- Viselter, who apply it in Lemma 9.2 of [11] so as to permit the use of Vaes's construction of actions on the free Araki -- Wood factors from [31]. Now that we have also have it in our more general setting, we can apply the argument in Section 9 of [11] to deduce the following theorem. Note that the assumption of trivial scaling group is not merely required for the application of Theorem 3.11 of [32] in the previous paragraph, but is also a hypothesis in Lemma 9.2 of [11]. Here (L F∞, τ ) is the von Neumann algebra of QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 15 the free group on a countably infinite set of generators along with its canonical tracial state. Theorem 5.1. For a second countable locally compact quantum group G with trivial scaling group the following are equivalent: (i) G has property (T), (ii) every weakly mixing n.s.p. action G y (N, σ) is strongly ergodic, (iii) every ergodic n.s.p. action G y (N, σ) is strongly ergodic. One can also replace (N, σ) with the fixed pair (L F∞, τ ) in (ii) and (iii). 6. The general nonunimodular case In this final section, our aim is to prove Theorem 6.3, which extends Theorems 4.8 and 4.9 so as to cover the general nonunimodular case. It shows in particular that a nonunimodular second countable locally compact quantum group cannot have property (T), which in the discrete situation was established in [13]. Let G be a second countable locally compact quantum group. Recall that the mod- ular element of G is strictly positive unbounded operator δ on L2(G) affiliated with the von Neumann algebra L∞(G). The map t 7→ δit from R to L∞(G) ⊆ B(L2(G)) is continuous in the strong operator topology by Stone's theorem and so it defines a uni- tary operator U on L2(R, L2(G)) ∼= L2(R)⊗ L2(G) which, when viewing L2(R, L2(G)) as the Hilbert space direct integral of copies of L2(G) over R with respect to Lebesgue measure, is a decomposable operator and as such is expressed by the direct integral R ⊕ R δit dt. Thus for all η1, η2 ∈ L2(G) and ξ1, ξ2 ∈ L2(R) we have hU(η1 ⊗ ξ1), η2 ⊗ ξ2i =ZRhξ1(t)δitη1, ξ2(t)η2i dt (1) ξ1(t)ξ2(t)hδitη1, η2i dt. =ZR Since ∆(δit) = δit ⊗ δit (see for example the proof of Proposition 1.9.11 in [30]), we see that (∆⊗id)U and U13U12 are both decomposable operators on L2(R, L2(G)⊗L2(G)) ∼= L2(G) ⊗ L2(G) ⊗ L2(R) which can be expressed as the direct integral R ⊕ R δit ⊗ δit dt. Thus U is a unitary representation of G on L2(R). (In fact, U ∈ L∞(G)⊗L∞(R) can be regarded as a unitary representation of both G and R simultaneously.) Lemma 6.1. The unitary representation U of G has almost invariant vectors. Proof. For every n ∈ N, writing 1[0, 1 √n1[0, 1 formula (1), for every n we have n ] we set ξn = n ], which is a unit vector in L2(R). Let η ∈ L2(G) be a unit vector. Using the n ] for the indicator function of [0, 1 hU(η ⊗ ξn), η ⊗ ξni − 1 = nZ[0, 1 n ]h(δit − 1)η, ηi dt 16 MICHAEL BRANNAN AND DAVID KERR and the expression on the right converges to zero as n → ∞ by the strong operator continuity of the map t 7→ δit. Thus U has almost invariant vectors. (cid:3) Let V be any unitary representation of G on a separable Hilbert space H , and con- sider the tensor product representation Z = V ⊙ U = V12U13. Viewing L2(R, L2(G) ⊗ H ) ∼= L2(G)⊗ H ⊗ L2(R) as the Hilbert space direct integral of copies of L2(G)⊗ H over R with respect to Lebesgue measure, the operator V12 is decomposable and can expressed by the direct integral R ⊕ R V dt, so that for η ∈ L2(G) and ζ ∈ L2(R, H ) ∼= H ⊗ L2(R) the vector V ∗ 12(η ⊗ ζ), viewed as an element of L2(R, L2(G)⊗ H ), is equal to t 7→ V ∗(η ⊗ ζ(t)). Thus for η1, η2 ∈ L2(G) and ζ1, ζ2 ∈ L2(R, H ) ∼= H ⊗ L2(R) we have the formula (2) hZ(η1 ⊗ ζ1), η2 ⊗ ζ2i = hU13(η1 ⊗ ζ1i), V ∗ 12(η2 ⊗ ζ2i) =ZRhδitη1 ⊗ ζ1(t), V ∗(η2 ⊗ ζ2(t))i dt. Lemma 6.2. Suppose that G is not unimodular. Let V be any unitary representation of G on a separable Hilbert space H . Then the tensor product representation Z = V ⊙ U is weakly mixing. Moreover, if H is infinite-dimensional then the closure of the set of unitary conjugates of Z in Rep(G, H ) contains V . Proof. Since G is not unimodular, there is a real number c 6= 0 such that ec belongs to the spectrum of the modular element δ. The equation ∆(δ) = δ ⊗ δ then implies, via elementary spectral theory, that enc belongs to the spectrum of δ for every n ∈ N. Let E be a nonzero finite-dimensional subspace of L2(R)⊗ H . Choose an orthonor- mal basis {ζ1, . . . , ζK} for E . Let ε > 0. Then there are a b > 0 and an N ∈ N k ∈ L2([−b, b], H ) ∼= L2([−b, b]) ⊗ H such that for each k = 1, . . . , K we can find a ζ ′ which is an H -valued step function on [−b, b] taking at most N different values and satisfying kζ ′ k(t)k. By the proof of the Riemann -- Lebesgue lemma for step functions, there is an n ∈ N depending only on N and M such that, setting s = nc, every step function f : [−b, b] → C which takes at most N 2 values and is bounded in modulus by M 2 satisfies k − ζkk < ε/3. Set M = maxk=1,...,K maxt∈[−b,b] kζ ′ (3) f (t)eist dt(cid:12)(cid:12)(cid:12)(cid:12) < ε 6 Z[−b,b] (cid:12)(cid:12)(cid:12)(cid:12) Write K for the range of the spectral projection of δ corresponding to [0, 2es]. Since es belongs to the spectrum of δ, we can find a norm-one vector η ∈ K such that kδη − esηk is small enough so that by the continuous functional calculus, applied to δ acting (boundedly) on K , we have kδitη−eistηk < ε/(6M 2) for all t ∈ [−b, b]. For every norm-one vector θ ∈ L2(G) and k = 1, . . . , n the function t 7→ hη⊗ ζ ′ k(t))i on [−b, b] is a step function which takes at most N 2 values and is bounded in modulus 1(t), V ∗(θ⊗ ζ ′ QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 17 by M 2, and so using the formula (2) and applying (3) we obtain hZ(η ⊗ ζ ′ 1), θ ⊗ ζ ′ ki =(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12) 1(t), V ∗(θ ⊗ ζ ′ Z[−b,b]hδitη ⊗ ζ ′ Z[−b,b]heistη ⊗ ζ ′ +(cid:12)(cid:12)(cid:12)(cid:12) Z[−b,b]hη ⊗ ζ ′ t∈[−b,b](cid:0)kδitη − eistηkkζ ′ ε 6M 2 · M 2 1(t), V ∗(θ ⊗ ζ ′ Z[−b,b]hδitη − eistη ⊗ ζ ′ 1(t), V ∗(θ ⊗ ζ ′ + sup k(t))i dt(cid:12)(cid:12)(cid:12)(cid:12) k(t))i dt(cid:12)(cid:12)(cid:12)(cid:12) k(t))ieist dt(cid:12)(cid:12)(cid:12)(cid:12) k(t)k(cid:1) 1(t), V ∗(θ ⊗ ζ ′ 1(t)kkζ ′ + k(t))i dt(cid:12)(cid:12)(cid:12)(cid:12) < = and hence ε 6 ε 3 (4) 1), θ ⊗ ζ ′ hZ(η ⊗ ζ1), θ ⊗ ζki ≤ hZ(η ⊗ ζ ′ ε = ε. + 3 ε 3 ε 3 + < ki + kζ1 − ζ ′ 1k + kζk − ζ ′ kk Now given a norm-one vector κ ∈ L2(G) ⊗ E we can write it asPK PK k=1 kθkk2 = 1, and so if we take ε = 1/K then from (4) we get k=1 θk ⊗ ζk where hZ(η ⊗ ζ1), κi ≤ hZ(η ⊗ ζ1), θk ⊗ ζki < KXk=1 KXk=1 1 nkθkk ≤ Kε = 1. Since the vector Z(η⊗ζ1) has norm one by the unitarity of Z, it follows that Z(η⊗ζ1) /∈ E , which shows that L2(G)⊗E is not U-invariant. We conclude that U has no nonzero finite-dimensional subrepresentations. Suppose now that H is infinite-dimensional and let us show that the closure of the set of unitary conjugates of Z in Rep(G, H ) contains V . Fix an orthonormal basis {ζk}∞ k=1 of H . Let Ω be a finite set of norm-one elements in C0(G). Let K ∈ N and ε > 0. Working in the Hilbert module C0(G) ⊗ H , for each k = 1, . . . , K there exist xk,1, . . . , xk,Lk, yk,1, . . . , yk,Lk ∈ H such that kV(a ⊗ ζk) −PLk l=1 xk,l ⊗ ζlk < ε/3 and kV∗(a ⊗ ζk) −PLk l=1 yk,l ⊗ ζlk < ε/3 for all a ∈ Ω. Set L = max{k, L1, . . . , LK}. Using the characterization of having almost invariant vectors given in Proposition 3.7(v) of [10], Lemma 6.1 yields a unit vector ξ ∈ H such that kU(a ⊗ ξ) − a ⊗ ξk < ε/3 for all a ∈ Ω. Since H is separable and infinite-dimensional we can find a unitary isomorphism u : H ⊗ L2(R) → H which sends ζk ⊗ ξ to ζk for every k = 1, . . . , L. 18 MICHAEL BRANNAN AND DAVID KERR Then for every a ∈ Ω and k = 1, . . . , K we have k((id ⊗ u)Z(id ⊗ u)−1 − V)(a ⊗ ζk)k ≤ k(id ⊗ u)V12(U13(a ⊗ ζk ⊗ ξ) − a ⊗ ζk ⊗ ξ)k +(cid:13)(cid:13)(cid:13)(cid:13)(id ⊗ u)(cid:18)V12(a ⊗ ζk ⊗ ξ) − xk,l ⊗ ζl − V(a ⊗ ζk)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13) LkXl=1 = ε LkXl=1 xk,l ⊗ ζl ⊗ ξ(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) < ε 3 + + ε 3 ε 3 and similarly k((id ⊗ u)Z∗(id ⊗ u)−1 − V∗)(a ⊗ ζk)k < ε. We conclude that V belongs to the closure of the set of unitary conjugates of Z in Rep(G, H ). (cid:3) We can now conclude with the main result of the section. Theorem 6.3. Let G be a second countable nonunimodular locally compact quantum group. Then there exists a a weakly mixing unitary representation of G which has almost invariant vectors. Moreover, if H is a separable infinite-dimensional Hilbert space then WM(G, H ) is a dense Gδ in Rep(G, H ). Proof. The representation U has almost invariant vectors by Lemma 6.1, and it is weakly mixing by Lemma 6.2, as we can take V there to be the trivial representation. Finally, if H is a separable infinite-dimensional Hilbert space then WM(G, H ) is (cid:3) a dense Gδ in Rep(G, H ) by Lemmas 3.5 and 6.2. References [1] Y. Arano. Unitary spherical representations of Drinfeld doubles. To appear in J. Reine Angew. Math. [2] Y. Arano. Comparison of unitary duals of Drinfeld doubles and complex semisimple Lie groups. Comm. Math. Phys. 351 (2017), 1137 -- 1147. [3] E. B´edos, R. Conti, and L. Tuset. On amenability and co-amenability of algebraic quantum groups and their corepresentations. Canad. J. Math. 57 (2005), 17 -- 60. [4] M. E. B. Bekka and A. Valette. Kazhdan's property (T) and amenable representations. Math. Z. 212 (1993), 293 -- 299. [5] B. Bekka, P. de la Harpe, and A. Valette. Kazhdan's Property (T). New Mathematical Mono- graphs, 11. Cambridge University Press, Cambridge, 2008. [6] M. Brannan, M. Daws, and E. Samei. Completely bounded representations of convolution algebras of locally compact quantum groups. Munster J. Math. 6 (2013), 445 -- 482. [7] X. Chen and C.-K. Ng. Property T for locally compact groups. Intl. J. Math. 26 (2015), 1550024, 13 pp. [8] A. Connes and B. Weiss. Property T and asymptotically invariant sequences. Israel J. Math. 37 (1980), 209 -- 210. [9] B. Das and M. Daws. Quantum Eberlein compactifications and invariant means. Indiana Univ. Math. J. 65 (2016), 307 -- 352. QUANTUM GROUPS, PROPERTY (T), AND WEAK MIXING 19 [10] M. Daws, P. Fima, A. Skalski, S. White. The Haagerup property for locally compact quantum groups. J. Reine Angew. Math. 711 (2016), 189 -- 229. [11] M. Daws, A. Skalsi, and A. Viselter. Around property (T) for quantum groups. arXiv:1605.02800v1. [12] J. Dixmier. C∗-Algebras. Translated from the French by Francis Jellett. North-Holland Mathe- matical Library, Vol. 15. North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977. [13] P. Fima. Kazhdan's property T for discrete quantum groups. Internat. J. Math. 21 (2010), 47 -- 65. [14] E. Glasner and B. Weiss. Kazhdan's property T and the geometry of the collection of invariant measures. Geom. Funct. Anal. 7 (1997), 917 -- 935. [15] M. Joit¸a and S. Petrescu. Property (T) for Kac algebras. Rev. Roumaine Math. Pures Appl. 37 (1992), 163 -- 178. [16] P. Jolissaint. On property (T) for pairs of topological groups. Enseign. Math. (2) 51 (2005), 31 -- 45. [17] A. S. Kechris. Classical Descriptive Set Theory. Graduate Texts in Mathematics, 156. Springer- Verlag, New York, 1995. [18] A. S. Kechris. Global Aspects of Ergodic Group Actions. Mathematical Surveys and Monographs, 160. American Mathematical Society, Providence, RI, 2010. [19] D. Kerr and H. Li. Ergodic Theory: Independence and Dichotomies. Springer, 2016. [20] D. Kerr and M. Pichot. Asymptotic Abelianness, weak mixing, and property T. J. Reine Angew. Math. 623 (2008), 213 -- 235. [21] J. Kustermans. Locally compact quantum groups in the universal setting. Internat. J. Math. 12 (2001), 289 -- 338. [22] J. Kustermans and S. Vaes. Locally compact quantum groups. Ann. Sci. ´Ecole Norm. Sup. (4) 33 (2000), 837 -- 934. [23] J. Kustermans and S. Vaes. Locally compact quantum groups in the von Neumann algebraic setting. Math. Scand. 92 (2003), 68 -- 92. [24] D. Kyed and P. Soltan. Property (T) and exotic quantum group norms. J. Noncommut. Geom. 6 (2012), 773 -- 800. [25] E. C. Lance. Hilbert C∗-Modules. A Toolkit for Operator Algebraists. London Mathematical Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995. [26] S. Neshveyev and M. Yamashita. Drinfeld center and representation theory for monoidal cate- gories. Comm. Math. Phys. 345 (2016), 385 -- 434. [27] J. Peterson and S. Popa. On the notion of relative property (T) for inclusions of von Neumann algebras. J. Funct. Anal. 219 (2005), 469 -- 483. [28] S. Popa. Deformation and rigidity for group actions and von Neumann algebras. In: Interna- tional Congress of Mathematicians. Vol. I., 445 -- 477. Eur. Math. Soc., Zurich, 2007. [29] S. Popa, and S. Vaes. Representation theory for subfactors, λ-lattices and C∗-tensor categories. Comm. Math. Phys. 340 (2015), 1239 -- 1280. [30] S. Vaes. Locally compact quantum groups. Thesis, Catholic University of Leuven, 2001. [31] S. Vaes. Strictly outer actions of groups and quantum groups. J. Reine Angew. Math. 578 (2005), 147 -- 184. [32] A. Viselter. Weak mixing for locally compact quantum groups. To appear in Ergodic Theory Dynam. Systems. [33] P. S. Wang. On isolated points in the dual spaces of locally compact groups. Math. Ann. 218 (1975), 19 -- 34. [34] R. J. Zimmer. Ergodic Theory and Semisimple Groups. Monographs in Mathematics, 81. Birkhauser Verlag, Basel, 1984. 20 MICHAEL BRANNAN AND DAVID KERR Michael Brannan, Department of Mathematics, Mailstop 3368, Texas A&M Univer- sity, College Station, TX 77843-3368, USA E-mail address: [email protected] David Kerr, Department of Mathematics, Mailstop 3368, Texas A&M University, College Station, TX 77843-3368, USA E-mail address: [email protected]
1611.01556	2	1611	2016-11-08T14:59:31	Dirac type operators on the quantum solid torus with global boundary conditions	[ "math.OA" ]	We define a noncommutative space we call the quantum solid torus. It is an example of a noncommutative manifold with a noncommutative boundary. We study quantum Dirac type operators subject to Atiyah-Patodi-Singer like boundary conditions on the quantum solid torus. We show that such operators have compact inverse, which means that the corresponding boundary value problem is elliptic.	math.OA	math	DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS WITH GLOBAL BOUNDARY CONDITIONS SLAWOMIR KLIMEK AND MATT MCBRIDE Abstract. We define a noncommutative space we call the quantum solid torus. It is an example of a noncommutative manifold with a noncommutative boundary. We study quantum Dirac type operators subject to Atiyah-Patodi-Singer like boundary conditions on the quantum solid torus. We show that such operators have compact inverse, which means that the corresponding boundary value problem is elliptic. 1. Introduction In this paper we are continuing our study of quantum analogs of Dirac type operators on manifolds with boundary and global boundary conditions of Atiyah, Patodi, and Singer type [2]. Despite similarities with our previous papers on the subject, the problem studied here is much more complex. We consider a higher dimensional space, namely three dimensional solid torus, and its quantization. The resulting space is noncommutative and its boundary is also noncommutative. Dirac type operators are matrix valued operators and the Atiyah- Patodi-Singer condition cannot be generalized straightforwardly. One of the main objectives of our previous work was to find appropriate non-commutative analogs of Dirac type operators on planar domains with boundary and to investigate their functional analytic properties. Some examples of Dirac type operators in simple domains, such as the disk, annulus, and punctured disk were described in [5] and [6]. These papers exhibited strong similarities in the setup and the results between the commutative and quantum cases. Also in those papers the global boundary condition imposed on Dirac-like operators was essentially the classical APS boundary condition. In reference [8] we discussed Dirac type operators on the solid torus, in the commutative sense, with a different nonlocal boundary condition, that was inspired by [9]. This is because the natural metric on the disk does not have the product structure near boundary, required for APS theory. This was not a problem for two dimensional domains but becomes an issue in dimension three. In this paper we follow up the analysis [8] of the operators on the commutative solid torus, with same type of investigation for their quantum analogs. Instead of constructing a quantum analog of the non-local boundary condition that was used in [8], we consider a large class of boundary conditions that yield desired analytical properties of the parametrices of the quantum Dirac operators. The boundary condition of [8] required extending functions in the domain beyond the boundary of the solid torus, which makes sense geometrically in the classical case. However in the quantum analog, there are obvious obstructions, for example what does we mean by the "outside" of the boundary of the quantum solid torus. This will be addressed in our future work. Date: January 16, 2018. 1 2 SLAWOMIR KLIMEK AND MATT MCBRIDE The solid torus is the product of a disk and a circle. In quantum case we take a noncom- mutative disk and then a twisted product of it with the unit circle to produce the quantum solid torus. The operators we consider respect this decomposition. Unlike our previous work on d-bar operators on the quantum disk [5], we consider here a bigger variety of related Dirac like operators on the disk. They yield a wide class of operators on the solid torus. The crux of the analysis presented here is the proof that the parametrices of our Dirac type operators are compact operators, like it was shown in [8] that the parametrix of the classical Dirac type operator was also compact. The paper is organized as follows. In section 2 the quantum solid torus is introduced and discussed. This is followed by section 3 on partial Fourier series on the quantum solid torus. The relevant Hilbert spaces are also defined there. Section 4 introduces our Dirac type operators, while section 5 contains a discussion of some properties of their coefficients. The boundary conditions are defined in section 6, based on properties of special solutions introduced there. The computation of the parametrix of the Dirac type operator is described in section 7. Finally, section 8 contains analysis of a parametrix culminating in the proof of the main theorem on compactness. 2. Non-commutative solid Torus In this section we define our version of the quantum solid torus. From the topological point of view such a quantum torus is a noncommutative C ∗−algebra which has a structure similar in some ways to the ordinary solid torus. The idea here is simple: the solid torus can be thought of as the product of the disk and the circle. To obtain a quantum solid torus we take the quantum disk of [4] and take its twisted product with the circle. The result can be described either as a suitable crossed product algebra or as a universal C ∗−algebra with generators and relations. We proceed to describe them now. Definition: Let Aθ be the universal C ∗−algebra with generators U and V such that U is an isometry i.e. U ∗U = I, V is a unitary i.e. V ∗V = V V ∗ = I and such that they satisfy the commutation relation V U = e2πiθUV. Let {ek} be the canonical basis for ℓ2(Z≥0) and let W be the unilateral shift, i.e. W ek = ek+1. Let T be the C ∗−algebra generated by W . The algebra T is called the Toeplitz algebra and by Coburn's theorem it is the universal C ∗−algebra with generator W satisfying the relation W ∗W = I, i.e. an isometry. Reference [4] shows that this algebra can be thought of as a quantum unit disk. It's structure is described by a short exact sequence, namely: (2.1) where K is the ideal of compact operators in ℓ2(Z≥0). In fact K is the commutator ideal of the algebra T . 0 −→ K −→ T −→ C(S1) −→ 0. ∼= T ⋊θ Z Proposition 2.1. For any θ we have the following isomorphism of algebras: Aθ where the crossed product on the right hand side is defined, for n ∈ Z, by the automorphisms Θn(U) = e2πinθU for some θ ∈ [0, 2π). DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 3 Note that because of the universality of the Toeplitz algebra, we only need to define an automorphism on its generator and check that it satisfies the defining relation. Proof. By definition, a representation π of the universal C ∗−algebra Aθ, consists of a Hilbert space, H and bounded operators π(U) and π(V ) on H, that satisfy the defining relations of Aθ. Then Aθ is the completion of the algebra of polynomials a in U and V and their adjoints with respect to the maximal norm, i.e. kakmax = sup π kπ(a)k. On the other hand, the covariant representation of the dynamical system (Z, Θn, T ) consists of a Hilbert space X, a bounded representation ρ of T on X, and a unitary operator ρ(V ) on X, such that the following commutation relation holds: ρ(Θ1(b)) = ρ(V )ρ(b)ρ(V )∗, for any b ∈ T . By the universality of the Toeplitz algebra, its representation ρ is completely determined by an isometry ρ(U) i.e. ρ(U)ρ(U)∗ = I. Moreover, the above commutation relation with ρ(V ) becomes the following: e2πiθρ(U) = ρ(V )ρ(U)ρ(V )∗. Consequently the crossed product T ⋊θ Z is obtained by completion of the algebra of poly- nomials a in U and V and their adjoints with respect to the maximal norm, i.e. kakmax = sup ρ kρ(a)k. So we see that the two concepts are identical, establishing an isomorphism between the algebras, and completing the proof. (cid:3) By the amenability of Z, the crossed product of the Toeplitz algebra with Z is equal to the reduced crossed product of the Toeplitz algebra with Z. Consequently, the norm of an element of the crossed product can computed from a single representation by combining a faithful representation of the Toeplitz algebra with the left regular representation of Z. Explicitly, choosing the defining representation of the Toeplitz algebra, we can view the reduced crossed product as the C ∗−algebra generated by the operators described below acting in the Hilbert space ℓ2(Z, ℓ2(Z≥0)) = ℓ2(Z) ⊗ ℓ2(Z≥0) = ℓ2(Z≥0 × Z), where the equality of Hilbert spaces follows from [10]. Corollary 2.2. Let {ek,l} be the canonical basis in ℓ2(Z≥0 × Z). Defining two operators: Uek,l = e−2πilθek+1,l and V ek,l = ek,l+1, we have Aθ = C ∗(U, V ). For θ ∈ [0, 2π) let T 2 θ is the universal C ∗−algebra with two unitary generators u and v such that vu = e2πiθuv, see for example [3]. The following proposition describes the structure of Aθ. θ be the two dimensional quantum torus. In other words, T 2 Proposition 2.3. We have the following short exact sequence: 0 −→ K ⊗ C(S1) −→ Aθ −→ T 2 θ −→ 0. 4 SLAWOMIR KLIMEK AND MATT MCBRIDE Proof. Consider the crossed product of the short exact sequence (2.1) for the Toeplitz algebra with Z. We get: 0 −→ K ⋊θ Z −→ T ⋊θ Z −→ C(S1) ⋊θ Z −→ 0. ∼= C(S1) ⋊θ Z, see [3]. Moreover, it is easy to see that K ⋊θ Z ∼= It is well known that T 2 θ K ⊗ C(S1), see [12] for the details. Explicitly, the cross product K ⋊θ Z can be "untwisted" by noticing that, in the tensor product of Hilbert spaces ℓ2(Z) ⊗ ℓ2(Z≥0), if a is compact in ℓ2(Z≥0) then: (I ⊗ a)(V ′)n = (I ⊗ e−2πiKθa) · (V ⊗ I)n. The operators above are Kek,l = kek,l and V ′ek,l = e−2πikθek,l+1. In addition define the operator U ′ek,l = ek+1,l, then we have Aθ = C ∗(U ′, V ′) by universality, while e−2πiKθa is still compact. Combination of all those facts above yields the above short exact sequence. (cid:3) In view of the results about Aθ contained in this section it natural to call this algebra the quantum solid torus. 3. Fourier series The purpose of this section is to introduce a partial Fourier transform on the quantum torus. As stated in Corollary 2.2, the algebra Aθ is generated by the following two operators: Uek,l = e−2πilθek+1,l and V ek,l = ek,l+1, satisfying the relations: V U = e2πiθUV , U ∗U = I and V ∗V = V V ∗ = I. We also reuse the following diagonal label operator Kek,l = kek,l, so, for a bounded function f : Z≥0 → C, we have f (K)ek,l = f (k)ek,l. We have the following two useful commutation relations for a diagonal operator f (K): f (K) U = Uf (K + 1) and f (K) V = V f (K). (3.1) Let P ol(U, V ) be the set of all polynomials in U, U ∗, V and V ∗. We call a function f : Z≥0 → C eventually constant, if there exists a natural number k0 such that f (k) is constant for k ≥ k0. The set of all such functions will be denoted by EC. Using the above notation we consider the following operators in ℓ2(Z≥0 × Z) expressed as finite sums: N N a = Xn=0,m≤M Xn=1,m≤M V mU nf + m,n(K) + f − m,n(K)V m(U ∗)n for some M, N ≥ 0 and f ± following proposition: m,n(k) ∈ EC. Let A be the set of all such operators. We have the Proposition 3.1. A = P ol(U, V ). Proof. Using the commutation relations (3.1) it is easy to see that a product of two elements of A and the adjoint of an element of A are still in A. It follows that A is a ∗−subalgebra of Aθ. Since U and V are in A, it follows that P ol(U, V ) ⊂ A. To prove the other way inclusion it suffices to show that for any f ∈ EC the operator f (K) is in P ol(U, V ), as the remaining parts of the sum are already polynomials in U, V , U ∗ and V ∗. To show that f (K) ∈ P ol(U, V ), we decompose any f (K) ∈ EC in the following way: DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 5 k0−1 f (K) = f (k)Pk + f∞P≥k0, Xk=0 where f∞ = limk→∞ f (k), Pk is the orthogonal projection onto span{ek,l}l∈Z and P≥k0 is the orthogonal projection onto span{ek,l}k≥k0,l∈Z. A straightforward calculation shows that U k(U ∗)k = P≥k and Pk = P≥k − P≥k+1. This completes the proof. (cid:3) The above considerations tell us that a ∈ A are completely determined by the coefficients f ± m,n(K) ∈ EC and so it motivates the following definition of a partial Fourier series. For f ∈ Aθ, define the formal series: fseries = Xn≥0,m∈Z V mU nf + m,n(K) + Xn≥1,m∈Z m,n(K)V m(U ∗)n, f − where f + m,n(k) = hek,0, (U ∗)nV −mf ek,0i and f − m,n(k) = hek,0, f U nV −mek,0i. Similarly to the usual theory of Fourier series, fseries determines f even though in general the series is not norm convergent. Other types of convergence results can be obtained along the lines of the usual Fourier analysis. We now proceed to describing the Hilbert spaces hosting our Dirac type operators. They are analogs of the L2 spaces on the classical solid torus. Let {an(k)} be a sequence of positive numbers, which we call weights, such that the sum exists and such that s(n) goes to zero as n → ∞. For any formal series define a norm by s(n) := 1 an(k) ∞ Xk=0 kfseriesk2 = ∞ Xk=0 Xn≥0,m∈Z 1 an(k) f + m,n(k)2 + ∞ Xk=0 Xn≥1,m∈Z 1 an(k) f − m,n(k)2. Let H0 be the Hilbert space whose elements are the above formal series fseries such that kfseriesk is finite. Let Q0 be the orthogonal projection in ℓ2(Z≥0 ×Z) onto span{ek,0}k∈Z≥0. Define the linear functional τ on B1(ℓ2(Z≥0 × Z)), the space of trace class operators, by τ (a) = tr(aQ0). Lemma 3.2. For a bounded operator a, and b ∈ B1(ℓ2(Z≥0 × Z)) we have the following inequality: Proof. We estimate as follows: τ (ab) ≤ kak(τ (b∗b))1/2 := kakkbkτ . τ (ab) = tr(abQ0) ≤ kak(tr((bQ0)∗bQ0))1/2 = kak(tr(Q∗ 0b∗bQ0))1/2 = kak(tr(b∗bQ0))1/2 = kak(τ (b∗b))1/2, where the inequality follows from the usual estimate for the trace. (cid:3) 6 SLAWOMIR KLIMEK AND MATT MCBRIDE We use this lemma to show that Aθ is dense in H0. Proposition 3.3. If f ∈ Aθ, then fseries converges in H0. Moreover the map Aθ ∋ f 7→ fseries ∈ H0 is continuous, one-to-one, and the image is dense in H0. Proof. We follow here the similar argument from [7]. For f ∈ Aθ we need to estimate the norm of fseries. Notice first that if f ∈ A, then fseries = f . Since such f 's are dense in Aθ, it suffices to estimate the norm of the finite sums. Without loss of generality, we assume that f has only the U ∗ and V terms as the same proof works for the remaining terms. For such f we have: N kfseriesk2 H0 = Xk=0 = τ Xn=1,m≤M   = τ Xn=1,m≤M   N Next we estimate: Consequently, using the inequality from Lemma 3.2, we get ∞ N Xn=1,m≤M 1 an(k) f − N m,n(k)2 = τ Xn=1,m≤M  m,n(K)  1 an(K) m,n(K)V −m (U ∗)n V mU nf − f − 1 an(K) f − m,n(K)2  kfseriesk2 kf k. 1 an(K) f −  f ∗ m,n(K)V m(U ∗)n  . m,n(K)V m(U ∗)n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)τ an(K) f − N 2 1 H0 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn=1,m≤M m,n(K)V m(U ∗)n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) f − 1 = τ f − m,n(K)V m(U ∗)n an(K) 1 an(k)an(k) N Xl=1,j≤M V −jU lf − j,l(K) 1 an(K)  n,k (cid:18) 1 an(k)(cid:19) . f − m,n(k)2 ≤ kfseriesk2 H0 sup an(K) N N 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn=1,m≤M = τ Xn=1,m≤M  Xn=1,m≤M Xk=0 = ∞ N Using the the summability conditions on the weights we obtain: sup n,k (cid:18) 1 an(k)(cid:19) ≤ sup n ∞ Xk=0 1 an(k)! = sup n (s(n)) ≤ const, and hence we get kfseriesk2 Aθ ∋ f 7→ fseries ∈ H0, and consequently fseries converges in H0 for every f ∈ Aθ. H0 ≤ constkfserieskH0kf k. This shows the continuity of the map Next we show that the map Aθ ∋ f 7→ fseries ∈ H0 is one-to-one. Let f and g belong to Aθ and suppose that fseries = gseries. This means that the Fourier coefficients are equal, that is: f + m,n(k) = g+ m,n(k) for all n ≥ 1 and all m,n(k) for all n ≥ 0 and all m and k, and f − m,n(k) = g− DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 7 m and k. From f + m,n(k) = hek,0, f U nV −mek,0i it follows that all matrix coefficients of f and g are the same so we must have f = g. Thus the map f 7→ fseries ∈ H0 is one-to-one. m,n(k) = hek,0, (U ∗)nV −mf ek,0i and f − To prove density we define the following indicator function δl(k) =(cid:26) 1 l = k 0 l 6= k . It is clear that V mU nδl(K) and δl(K)V m(U ∗)n are in Aθ, and moreover they form an orthog- onal basis for H0. Finally finite linear combinations of V mU nδl(K) and δl(K)V m(U ∗)n form a dense subspace of Aθ as they are polynomials in U and V and hence are in A, making it a dense subspace of H0. Thus the proof is complete. (cid:3) 4. Dirac type operators The purpose of this section is to introduce the main object of study of this paper: Dirac type operators on the quantum solid torus. We start with reviewing the definition of such operators, contained in [8], on the classical solid torus. In that paper we considered the following formally self-adjoint Dirac operator Dcl defined in L2(ST 2, C2) ∼= L2(ST 2) ⊗ C2 by where ST 2 = D × S1 is the solid torus, and D = {z ∈ C : Dcl =(cid:18) 1 ∂ i ∂θ −2 ∂ 2 ∂ ∂z ∂z − 1 ∂ i ∂θ (cid:19) , z ≤ 1} is the unit disk, while S1 =(cid:8)eiθ ∈ C : 0 ≤ θ ≤ 2π(cid:9) is the unit circle. Using Fourier decomposition: gm,n(r) (cid:19) einϕ+imθ, F = Xm,n∈Z(cid:18) fm,n(r) eiϕ(cid:0) ∂ m ∂r + n e−iϕ(cid:0)− ∂ r(cid:1) the operator D becomes: DclF = Xm,n∈Z(cid:18) ∂r − n −m r(cid:1) (cid:19)(cid:18) fm,n(r) gm,n(r) (cid:19) einϕ+imθ. The boundary condition we studied in [8], geometrically very much the same as the APS condition, considered the solid torus as a subset of the bigger noncompact space of the plane cross the unit circle. We defined the domain of the Dirac operator Dcl as consisting of those sufficiently regular (first Sobolev class) vectors F which extend to square integrable solutions of DclF = 0 on the complement of the solid torus. Denote by In, Kn the modified Bessel functions of the first and second kind respectively. Without boundary conditions, the operator Dcl on the solid torus has an infinite dimen- sional kernel consisting of linear combinations of special solutions In for which, in Fourier decomposition, the only nonzero components for m 6= 0 are: fm,n+1(r) = − m m In+1(mr), gm,n(r) = In(mr). Outside of the solid torus in the plane cross the unit circle the square integrable functions in kernel of Dcl are linear combinations of solutions Kn for which the only non vanishing 8 SLAWOMIR KLIMEK AND MATT MCBRIDE components when m 6= 0 are: fm,n+1(r) = m m Kn+1(mr), gm,n(r) = Kn(mr). Consequently, see [8] for details, using the Fourier decomposition of F , the boundary condi- tion can be rephrased as mKn+1(m)gm,n(1) − mKn(m)fm,n+1(1) = 0, if m 6= 0 and f0,n(1) = 0 for n ≤ 0. Additionally, for m = 0, we have g0,n(1) = 0 for n ≥ 0. We now proceed to define our Dirac type operators for the noncommutative solid torus, in a way that is analogous to the commutative case. Let c1,n(k) ≤ 1 and c2,n(k) ≤ 1 be sequences of positive numbers such that Qk c1,n(k), Qk c2,n(k) exist and are not zero, and such that there exists a constant, κ, such that for all k or n we have 1/κ ≤ 1/ci,n(k) ≤ 1, i = 1, 2. Let ℓ2 an(Z≥0) be the following Hilbert space of sequences: ℓ2 an(Z≥0) =({h(k)}k≥0 : Xk 1 an(k) h(k)2 < ∞) . We introduce the Jacobi type difference operators by formulas: Bnh(k) = an(k)(h(k) − c2,n(k − 1)h(k − 1)) : ℓ2 Bnh(k) = an+1(k)(h(k) − c1,n(k)h(k + 1)) : ℓ2 an+1(Z≥0) → ℓ2 an(Z≥0) an(Z≥0) → ℓ2 an+1(Z≥0) : kBhkℓ2 an+1(Z≥0) an (Z≥0) < ∞} and the (4.1) defined on maximal domains, i.e. dom(B) = {h ∈ ℓ2 domain of B is defined in the same way. Using the above general one-step difference operators we define operators δ0 and δ2 which are noncommutative analogs of ∂ ∂z and ∂ V mU n+1Bnf + ∂z as: δ0(f ) = − Xn≥0,m∈Z δ2(f ) = − Xn≥1,m∈Z m,n(K) + Xn≥1,m∈Z m,n(K) + Xn≥0,m∈Z Bn−1f − m,n(K)V m (U ∗)n−1 V mU n−1Bn−1f + Bnf − m,n(K)V m (U ∗)n+1 . Those operators are more general versions of the d-bar like operators considered in [5]. Next define another diagonal label operator L on ℓ2(Z≥0 × Z) by Lek,l = lek,l. We use it for the definition of δ1 = [L, · ], the analog of 1 i ∂ ∂θ . It is easy to see that δ1(f ) = Xn≥0,m∈Z mV mU nf + m,n(K) + Xn≥1,m∈Z again considered on its maximal domain. mf − m,n(K)V m (U ∗)n , Let H = H0 ⊗ C2. Any F ∈ H has the following Fourier decomposition: F =(cid:18) Pn≥1,m∈Z V mU nf + Pn≥0,m∈Z V mU ng+ m,n(K) +Pn≥0,m∈Z f − m,n(K) +Pn≥1,m∈Z g− We define the quantum Dirac type operator D to be m,n(K)V m(U ∗)n m,n(K)V m(U ∗)n (cid:19) . (4.2) DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 9 Initially we set the domain of D to be the maximal domain: D =(cid:18) δ1 δ2 −δ1 (cid:19) . δ0 dom(D) = {F ∈ H : kDF k < ∞}. (4.3) (4.4) Using Fourier series (4.2) we relate D to finite difference operators with matrix coefficients. For brevity throughout the rest of the paper we only study the "positive" part of the Fourier transform of the elements of the Hilbert space, as the "negative" part can be analyzed in an analogical fashion, but writing down all the repetative details would have significantly increase the length of the paper (and decrease its readability). Because of this, the "+" superscript is dropped for simplicity. Define the following Jacobi type difference operator with matrix valued coefficients Am,n(cid:18) f g (cid:19) (k) = Am,n(k + 1)(cid:20)(cid:18) f (k + 1) g(k + 1) (cid:19) − Cm,n(k)(cid:18) f (k) g(k) (cid:19)(cid:21) , Am,n(k + 1) =(cid:18) an+1(k)c1,n(k) m 0 an(k + 1) (cid:19) , where and Cm,n(k) = 1 c1,n(k) −m an(Z≥0) ⊗ ℓ2 an(k+1)c1,n(k) an+1(Z≥0) → ℓ2 −m an+1(k)c1,n(k) m2 an(k+1)an+1(k)c1,n(k) ! . c2,n(k) + (4.5) (4.6) (4.7) Notice that Am,n : ℓ2 domain for Am,n to be an+1(Z≥0) ⊗ ℓ2 an(Z≥0). Initially we define the The domain will be modified later to include a boundary condition. dom Am,n =(cid:8)h ∈ ℓ2 an(Z≥0) ⊗ ℓ2 an+1(Z≥0) : kAm,nhkan+1⊗an < ∞(cid:9) . (4.8) Proposition 4.1. For F, G ∈ H where F = (f, g)t, G = (p, q)t, and f, g, p, q ∈ H0, the "positive" part of the equation DF = G is equivalent to the following equations: Am,n(cid:18) gm,n fm,n+1 (cid:19) (k) =(cid:18) pm,n+1(k) −qm,n(k + 1) (cid:19) , for n ≥ 0, m ∈ Z, k ≥ 0, and an(0)fm,n+1(0) + mgm,n(0) = qm,n(0). (4.9) (4.10) The last equation (4.10) will be referred to below as the initial regularity condition as it is both a starting point of recurrence proofs and it is analogous to the regularity condition at origin for solutions of Dirac operator for the classical solid torus. Proof. In the Fourier series for F and G there is an ambiguity whether n = 0 term should go with positive or with negative terms. Because we change summation indices in the calculation below, the ambiguity is resolved differently for different components of F and G. 10 SLAWOMIR KLIMEK AND MATT MCBRIDE Using the definition of D and shifting n in the first sum we get, ignoring the "negative" part, DF = Xn≥0,m∈Z V m(cid:18) U n+1(mfm,n+1(k) − Bngm,n(k)) U n(−Bnfm,n+1(k) − mgm,n(k)) (cid:19) = G. Shifting k in the second equation for k ≥ 1 we get the following equivalent system of equations mfm,n+1(k) − an+1(k)(gm,n(k) − c1,n(k)gm,n(k + 1)) = pm,n+1(k) an(k + 1)(fm,n+1(k + 1) − c2,n(k)fm,n+1(k)) + mgm,n(k + 1) = −qm,n(k + 1), (cid:26) while the case k = 0 in the second equation leads to equation (4.10). The system above can be rewritten in matrix form: m (cid:18) an+1(k)c1,n(k) −(cid:18) an+1(k) 0 0 an(k + 1) (cid:19)(cid:18) gm,n(k + 1) fm,n+1(k + 1) (cid:19) an(k + 1)c2,n(k) (cid:19)(cid:18) gm,n(k) −m fm,n+1(k) (cid:19) =(cid:18) pm,n+1(k) −qm,n(k + 1) (cid:19) . The first matrix above is Am,n(k +1), therefore factoring it out of the left side of the equation gives the desired result and completes the proof. (cid:3) 5. Properties of the coefficients It follows from Proposition 4.1 that to analyze D we need to study the properties of the operators Am,n subject to the initial regularity condition. This section discusses important properties of the matrix coefficients of those operators. Unless specified differently, in all formulas in the section, n ≥ 0, m ∈ Z, k ≥ 0. First notice that det Am,n(k + 1) = an+1(k)an(k + 1)c1,n(k) 6= 0 for any k and n which means that the inverse of the matrices Am,n(k + 1) exists for any k and n. Notice also that det Cm,n(k) = c2,n(k)/c1,n(k). Additionally we have: Proposition 5.1. The infinite product exists and is invertible. Cm,n := Cm,n(k) ∞ Yk=0 Here and everywhere below a product of matrices is understood from right to left, that is: Cm,n(i) = Cm,n(k)Cm,n(k − 1) · · · Cm,n(0). k Yi=0 Proof. We consider Cm,n(k) − I, since by [11], if the series Pk kCm,n(k) − Ik converges and det Cm,n(k) 6= 0, then the infinite product of Cm,n(k) converges. We already have det Cm,n(k) 6= 0 for all k, m, n, so consider DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 11 Cm,n(k) − I = − 1 c1,n(k)−1 m an(k+1)c1,n(k) − an+1(k)c1,n(k) m an(k+1)an+1(k)c1,n(k) ! . m2 c2,n(k) − 1 + Estimating the matrix norm by the sum of absolute values of the coefficients, we have kCm,n(k) − Ik ≤(cid:12)(cid:12)(cid:12)(cid:12) 1 c1,n(k) − 1(cid:12)(cid:12)(cid:12)(cid:12) + m + m an+1(k)c1,n(k) an(k + 1)c1,n(k) + c2,n(k) − 1 + m2 an(k + 1)an+1(k)c1,n(k) . By assumption, the infinite products of c1,n(k) and c2,n(k) exist. Using this and the summa- bility criteria on 1/an(k), we see thatPk kCm,n(k) − Ik converges. Therefore we can deduce that the infinite product Qk Cm,n(k) exists, thus completing proof. Below we use simplified notation for the products: (cid:3) J1(n) = ∞ Yk=0 c1,n(k) and J2(n) = c2,n(k). ∞ Yk=0 We have k k det Cm,n = lim k→∞ det Cm,n(i) = lim k→∞ Yi=0 c2,n(i) c1,n(i) Yi=0 = Q∞ Q∞ i=0 c2,n(i) i=0 c1,n(i) = J2(n) J1(n) . The next proposition describes the structure of the infinite product of the Cm,n(k) matrices. Proposition 5.2. The infinite product Cm,n has the following structure: Cm,n =(cid:18) Q∞ i=0 1 c1,n(i) + F0(m2) −mF2(m2) −mF1(m2) i=0 c2,n(i) + F3(m2) (cid:19) , Q∞ where the Fj(m2) are power series in m2 with positive coefficients for j = 0, 1, 2, 3, and additionally ∞ 1 form c1,n(i) F3(m2) = ∞ Yi=0 + F0(m2)!−1 m2F1(m2)F2(m2) − F0(m2) Yi=0 Proof. It is easy to verify by induction that for each k the product Qk Yi=0 Cm,n(i) = Qk i=0 u2(n, k)(m2)i i=0 u0(n, k)(m2)i −mPk i=0 c2,n(i) +Pk Qk c1,n(i) +Pk −mPk i=0 k 1 where the above sums are polynomials in m2 with positive coefficients. Moreover as k → ∞, the polynomials converge to a power series, by the Weierstrass Analytic Convergence Theorem, see [1]. Thus the result follows. (cid:3) c2,n(i)! . i=0 Cm,n(i) is of the i=0 u1(n, k)(m2)i i=0 u3(n, k)(m2)i ! , 12 SLAWOMIR KLIMEK AND MATT MCBRIDE 6. The Class of Boundary Conditions As defined above, the operator D is not a Fredholm operator on its maximal domain: it has an infinite dimensional kernel, as is typical for differential operators on manifolds with boundary. In this section we single out two types of solutions in the kernel of Am,n that play the role of similar solutions in [8] made of modified Bessel functions of first and second kind. Like in the commutative case, they are then used to define a class boundary conditions for D that turns D into an invertible operator. First we describe the kernel of D through the kernel of Am,n. Proposition 6.1. Let Am,n be the operator given by equation (4.5), then for some vector v. Ker Am,n =( k−1 Yi=0 Cm,n(i)! v) Proof. By formula (4.5) we need to solve the equation: Am,n(k + 1)(cid:20)(cid:18) f (k + 1) g(k + 1) (cid:19) − Cm,n(k)(cid:18) f (k) g(k) (cid:19)(cid:21) =(cid:18) 0 0 (cid:19) . Solving recursively we see that for some arbitrary vector v. g(k + 1) (cid:19) = k (cid:18) f (k + 1) Yi=0 Cm,n(i)! v (cid:3) If Fm,n(k) ∈ Ker Am,n, then by Proposition 5.1 limk→∞ Fm,n(k) = Fm,n(∞) exists and is finite. We define Im,n(k) to be a special element of Ker Am,n, so: Im,n(k) = k−1 Yi=0 Cm,n(i)! Im,n(0) = k−1 Yi=0 Cm,n(i)! I (1) m,n(0) ! , m,n(0) I (2) (6.1) which additionally satisfies the initial regularity equation (4.10) with zero right-hand side. We normalize Im,n(k) in the following way: we set I (1) m,n(0) = −1 and I (2) m,n(0) = m an(0) . We also want to define a class of complementary solutions, denoted Km,n(k), for which Km,n(k) = k−1 Yi=0 Cm,n(i)! Km,n(0) = k−1 Yi=0 Cm,n(i)! K (1) m,n(0) ! , m,n(0) K (2) (6.2) and such that for every k, n ≥ 0 and m ∈ Z, the set {Im,n(k), Km,n(k)} is linearly inde- pendent. This can be achieved by requesting the following first three sign conditions at infinity: DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 13 1.) m > 0 : K (1) 2.) m < 0 : K (1) 3.) m = 0 : K (1) K (1) K (2) 4.) m 6= 0 : m,n(∞) > 0 and K (2) m,n(∞) < 0 and K (2) 0,n(∞) = 0 and K (2) m,n(∞) m,n(∞) m,n(∞) > 0 m,n(∞) > 0 0,n(∞) 6= 0 → 0 as m → ∞ uniformly in n. (6.3) Additionally, the sign conditions, and the normalization of Im,n, lead to monotonicity prop- erties of Im,n and Km,n that are crucially used in estimates in the last section. The key asymptotic fourth condition above is also required to prove compactness of the resolvent. Any collection Km,n(k) satisfying (6.2) and (6.3) will be referred to, abusing the terminology, as a quantum K function. We have: Proposition 6.2. Let Km,n(k) be any quantum K function. For every k, n ≥ 0 and m ∈ Z the vectors Im,n(k) and Km,n(k) are linearly independent in C2. Moreover for all n ≥ 0 and m ∈ Z, Im,n(∞) and Km,n(∞) are also linearly independent. Proof. Consider the case m > 0. Recall that we have Im,n(k) = k−1 Yi=0 Cm,n(i)!(cid:18) −1 an(0) (cid:19) and Km,n(k) = k−1 Yi=0 m Cm,n(i)! K (1) m,n(0) ! . m,n(0) K (2) Using the formulas for the product of Cm,n(k) in the proof of Proposition 5.2 we write out the components to get u0(n, k)(m2)i + m an(0) k Yi=0 i=0 u1(n, k)(m2)i an(0) ! , u3(n, k)(m2)i! . m2Pk Xi=0 k c2,n(i) + I (1) m,n(k) = − k Yi=0 Xi=0 I (2) m,n(k) = m k 1 c1,n(i) + k Xi=0 u2(n, k)(m2)i + Since all coefficients are positive and m > 0, we see that I (1) positive. m,n(k) is negative and I (2) m,n(k) is To analyze Km,n(k) we use an alternative, equivalent, way to write it as: Km,n(k) = ∞ Yi=k Cm,n(i)!−1 Km,n(∞). Since the components of Km,n(∞) are both positive, and the matrix Cm,n(i)−1 has all positive entries for m > 0, it follows that K (1) m,n(k) are positive. Therefore since one of the components of Im,n(k) is negative and both components of Km,n(k) are positive, it is impossible for them to be linearly dependent. m,n(k) and K (2) In the case m < 0 both components of Im,n(k) are negative and one of the components of Km,n(k) is positive again showing they can not be linearly dependent. When m = 0 14 SLAWOMIR KLIMEK AND MATT MCBRIDE then all matrices Cm,n(k) are diagonal and so K (1) m,n(k) = 0, which implies independence. Finally the same arguments are valid when k = ∞, thus the proof is complete. (cid:3) m,n(k) = 0 and I (2) Now we can define the class of boundary conditions we consider for D. Given an F ∈ H such that DF ∈ H, the Fourier coefficients fm,n+1(k) and gm,n(k) given by equation (4.2), have limits as k → ∞ for all n ≥ 0 and m ∈ Z. This follows from the calculation of the resolvent in the next section, see equation (7.3) for details. These limits will be denoted by fm,n+1(∞) and gm,n(∞). Any quantum K function defines a boundary condition in the following way. Definition: Given Km,n(k) satisfying (6.2) and (6.3) we set the domain dom(D) of D to be the space of all F ∈ H such that DF ∈ H and such that for all m ∈ Z, n ≥ 0 there exist βm,n ∈ C such that gm,n(∞) = βm,nK (1) fm,n+1(∞) = βm,nK (2) m,n(∞) m,n(∞). (6.4) The above conditions at infinity can be restated to mimic the boundary condition in the classical case: gm,n(∞)K (2) m,n(∞) − fm,n+1(∞)K (1) m,n(∞) = 0. The following theorem is the main result of this paper and is proved at the end of this paper. Theorem 6.3. The quantum Dirac operator D, defined by (4.3) subject to the boundary condition defined in equation (6.4), is an invertible operator whose inverse Q, is a compact operator. We end this section with listing of useful recurrence relations for Im,n(k) and Km,n(k) that are easy consequences of their definitions. Let Hm,n(k) be Im,n(k) or Km,n(k) with components Hm,n(k) = (H (1) m,n(k))t, then we have m,n(k), H (2) Hm,n(k + 1) = Cm,n(k)Hm,n(k). (6.5) This gives the following recurrence relation: H (1) m,n(k + 1) − 1 c1,n(k) H (1) m,n(k) = − H (2) m,n(k + 1) − c2,n(k)H (2) m,n(k) = − H (1) m,n(k) (6.6) H (2) m,n(k) m an+1(k)c2,n(k) m an(k + 1)c1,n(k) + m2 an+1(k)an(k + 1)c1,n(k) H (2) m,n(k). Similarly, using the relation we can produce the following two (equivalent) equations: Hm,n(k) = (Cm,n(k))−1Hm,n(k + 1), DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 15 H (2) m,n(k) − 1 c2,n(k) H (2) m,n(k + 1) = m an(k + 1)c2,n(k) H (1) m,n(k + 1) H (1) m,n(k) − c1,n(k)H (1) m,n(k + 1) = m an+1(k)c2,n(k) H (2) m,n(k + 1)+ (6.7) + m2 an+1(k)an(k + 1)c2,n(k) H (1) m,n(k + 1). The recurrence relations above are extensively used in the next two section in the analysis of the parametrix of D. 7. Parametrix of the quantum Dirac type operator In the previous section we introduced a new domain of D. To account for this, we redefine an(Z≥0) ⊗ the domain of Am,n in the following way: dom(Am,n) is the space of all h ∈ ℓ2 ℓ2 an+1(Z≥0) such that kAm,nhkan+1⊗an < ∞, and such that there exists β ∈ C such that x(∞) = βK (1) y(∞) = βK (2) m,n(∞) m,n(∞), (7.1) where h = (x, y)t. The existence of such limits x(∞) and y(∞) follows from the calculation of the resolvent in this section, see equation (7.3). Proposition 7.1. Given h ∈ dom Am,n, if h ∈ Ker Am,n and h = (x, y)t satisfies an(0)y(0) + mx(0) = 0, then h(k) = 0 for every k ≥ 0. In other words, the kernel of Am,n subject to initial regularity condition (4.10) is trivial. Proof. First notice that we can write any element h(k) ∈ Ker Am,n as h(k) = c1Im,n(k) + c2Km,n(k), since by Proposition 6.2, the nonzero two-vectors Im,n(k) and Km,n(k) are linearly indepen- dent. However if k = 0 then both h(0) and Im,n(0) satisfy the equation an(0)y(0)+mx(0) = 0 so there exists some constant c3 such that: c3Im,n(0) = h(0) = c1Im,n(0) + c2Km,n(0). This would imply that Km,n(0) is a scalar multiple of Im,n(0), which is impossible by the linear independence of the solutions, hence c2 = 0. If k = ∞ then, since h ∈ dom Am,n, there is some constant c4, such that: c4Km,n(∞) = h(∞) = c1Im,n(∞), which implies that Km,n(∞) is a scalar multiple of Im,n(∞), but again this is impossible by Proposition 6.2, hence c1 = 0. Therefore h(k) = 0 for every k. This completes the proof. (cid:3) 16 SLAWOMIR KLIMEK AND MATT MCBRIDE Next we discuss the non-homogeneuous equation DF = G, the solution of which leads to the parametrix (in our case the inverse) of the quantum Dirac type operator D, subject to boundary conditions (6.4). We use the following notation: for a vector v = (v1, v2)t we write v⊥ := (v2, −v1)t. Proposition 7.2. Let Am,n be the one-step matrix difference operator defined by equation (4.5), then Am,n, subject to the boundary conditions given by the equation (7.1), and subject to the initial regularity condition (4.10) is an invertible operator with inverse Q(m,n) given by (7.7) below. Proof. The goal is to solve the equation which becomes the following difference equation with matrix coefficients Am,n(cid:18) x y (cid:19) (k) =(cid:18) p(k) −q(k + 1) (cid:19) , Am,n(k + 1)(cid:20)(cid:18) x(k + 1) y(k + 1) (cid:19) − Cm,n(k)(cid:18) x(k) y(k) (cid:19)(cid:21) =(cid:18) p(k) −q(k + 1) (cid:19) with Am,n(k + 1) and Cm,n(k) defined in formulas (4.6) and (4.7) respectively, while the initial regularity condition is an(0)y(0) + mx(0) = q(0). Relabeling h(k) = (x(k), y(k))t and r(k + 1) = (p(k), −q(k + 1))t, the system becomes Am,n(k + 1)(h(k + 1) − Cm,n(k)h(k)) = r(k + 1). Before applying the boundary condition, the solution h(k) is not unique, with non-uniqueness due to one-dimensional kernel Im,n(k). With this in mind we choose h(0) to be the following particular solution of the initial regularity condition: Solving the difference equation by using variation of constants method we get h(0) =(cid:18)0, q(0) an(0)(cid:19)t . Q(m,n)r(k) := h(k) = Cm,n(i) Cm,n(j)!−1 (Am,n(i))−1 r(i) + αIm,n(k) (7.2) k−1 Yi=0 k Xi=0 i−1 Yj=0 for some parameter α. Here, for convenience, we set the Qk−1 an(0)(cid:17)t and also we define (Am,n(0))−1 r(0) := h(0) =(cid:16)0, q(0) j=k Cm,n(j) = 1 for any m and n, from above. To apply the boundary conditions in equation (6.4), we need to know that Q(m,n)r(∞) is well defined; this follows from Proposition 5.1 and the summability of Am,n(k)−1. Therefore the boundary condition, equation (6.4), is well defined, and gives: Q(m,n)r(∞) = Cm,n ∞ Xi=0 i−1 Yj=0 = βKm,n(∞) Cm,n(j)!−1 (Am,n(i))−1 r(i) + αIm,n(∞) = (7.3) DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 17 for some constant β. The goal is to solve for α. Below we use the following notation: (7.4) Multiplying by (Cm,n)−1 and taking the inner product of both sides of equation (7.3) with τm,n := hKm,n(0), Im,n(0)⊥i. Km,n(0)⊥, we get * ∞ Xi=0 i−1 Yj=0 Cm,n(j)!−1 (Am,n(i))−1 r(i), Km,n(0)⊥+ + ατm,n = 0, which can now be solved for α: α = −1 τm,n * ∞ Xi=0 i−1 Yj=0 Cm,n(j)!−1 (Am,n(i))−1 r(i), Km,n(0)⊥+ . There is a partial cancellation between the two terms in equation (7.2). To see this, we notice that, since Im,n(0) and Km,n(0) are linearly independent, we can decompose any two-vector v as where v = v1Im,n(0) + v2Km,n(0), v1 = −(cid:10)v, Km,n(0)⊥(cid:11) τm,n Applying this decomposition to vector and v2 = (cid:10)v, Im,n(0)⊥(cid:11) τm,n , v = k Xi=0 i−1 Yj=0 Cm,n(j)!−1 (Am,n(i))−1 r(i), and using the formula for α, and the formulas for Im,n(k) and Km,n(k), equations (6.1) and (6.2) respectively, we get k−1 1 Cm,n(i) Q(m,n)r(k) = Yi=0 Xi=k+1 i−1 τm,n * ∞ Yj=0 Cm,n(j)!−1 Xi=0 i−1 Yj=0 × − τm,n * k + 1 Cm,n(j)!−1 (Am,n(i))−1 r(i), Km,n(0)⊥+ Im,n(0) (Am,n(i))−1 r(i), Im,n(0)⊥+ Km,n(0)  . (Am,n(i))−1 r(i), Km,n(0)⊥+ Cm,n(j)!−1 (7.5) The coefficients in the formula above will be denoted by: e(1) m,n(k) := and −1 τm,n * ∞ Xi=k+1 i−1 Yj=0 18 SLAWOMIR KLIMEK AND MATT MCBRIDE e(2) m,n(k) := 1 τm,n * k Xi=0 i−1 Yj=0 Cm,n(j)!−1 (Am,n(i))−1 r(i), Im,n(0)⊥+ . With this notation we get the formula for Q(m,n): Q(m,n)r(k) = e(1) m,n(k)Im,n(k) + e(2) k Q(m,n)r(k) = The proof is complete. Xi=0 (cid:18) 0 0 1 an(i)Qk−1 m,n(k)Km,n(k) m 6= 0 0 j=i c2,n(j) (cid:19) r(i) m = 0. (7.6) (7.7) (cid:3) As computed above, the operator Q(m,n) is not easy to analyze, mostly because it contains products of non-commuting matrices Cm,n(k). Our strategy is to re-write the formulas for Q(m,n) in terms of quantum I and K functions, similarly to the commutative case, and then estimate the Hilbert-Schmidt norm of the parametrics using the recurrence relations (6.6) and (6.7). We need the following observations. For vectors v = (v1, v2)t and u = (u1, u2)t we have hv⊥, ui = −hv, u⊥i, where h·, ·i is the standard Euclidean inner product on R2. If R is a 2 × 2 matrix and v = (v1, v2)t, then we have (Rv)⊥ = (det R) (Rt)−1 v⊥. As a consequence of the above and using the formula for the determinant of the Cm,n(k) matrix, we get: Im,n(k)⊥ = Km,n(k)⊥ = k−1 Yi=0 Yi=0 k−1 c2,n(i) c1,n(i)  c1,n(i)  c2,n(i) k−1 Yi=0 k−1 Yi=0 t t Cm,n(i)!−1  Cm,n(i)!−1  Im,n(0)⊥ Km,n(0)⊥. (7.8) Consequently, inserting equations (7.8) into formulas for the coefficients e(1) (7.6), gives: m,n, e(2) m,n, (7.5) and e(1) m,n(k) = 1 τm,n Similarly we also have: ∞ Xi=k+1 i−1 Yj=0 c1,n(j) c2,n(j) (K (2) m,n(i), −K (1) m,n(i)) (Am,n(i))−1 r(i). e(2) m,n(k) = 1 τm,n c1,n(j) c2,n(j) (I (2) m,n(i), −I (1) m,n(i)) (Am,n(i))−1 r(i). k i−1 Xi=0 Yj=0 These are to be understood as a row of a matrix times a column vector. It is convenient to rewrite Q(m,n) in the matrix notation: DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 19 Q(m,n)rn(k) = e(1) m,n(k)Im,n(k) + e(2) m,n(k)Km,n(k) = (Im,n(k), Km,n(k)) e(1) m,n(k) ! = I (1) m,n(k) e(2) m,n(k) K (1) I (2) m,n(k) K (2) m,n(k) m,n(k) ! e(1) m,n(k) ! . m,n(k) e(2) Inserting the above formulas for the coefficients e(1) m,n, e(2) m,n we get: Q(m,n)rn(k) = j=0 i=k+1Qi−1 P∞ i=0Qi−1 Pk j=0 1 c1,n(j) m,n(k) τm,n I (1) m,n(k) K (1) I (2) m,n(k) K (2) c2,n(j)K (2) c2,n(j)I (2) m,n(i) m,n(k) ! × i=k+1Qi−1 m,n(i) −P∞ i=0Qi−1 −Pk c1,n(j) j=0 j=0 c1,n(j) m,n(i) c2,n(j)K (1) c2,n(j)I (1) m,n(i) ! (Am,n(i))−1 r(i). c1,n(j) Next we focus on multiplying the second matrix with (Am,n(i))−1. We use the recurrence relations, (6.6) and (6.7), to get: Q(m,n)rn(k) = 1 τm,n I (1) m,n(k) K (1) I (2) m,n(k) K (2) an+1(i−1) K (2) an+1(i−1) I (2) m,n(k) m,n(k) ! × i=k+1Qi−1 m,n(i − 1) P∞ i=0Qi−1 Pk m,n(i − 1) j=0 1 1 j=0 j=0 i=k+1Qi−2 P∞ i=0Qi−2 Pk j=0 c1,n(j) c2,n(j) c1,n(j) c2,n(j) c1,n(j) c2,n(j) c1,n(j) c2,n(j) 1 m,n(i) an(i) K (1) an(i) I (1) m,n(i) ! r(i). 1 (7.9) This formula finally expresses Q(m,n) through the quantum I and K functions. Its primary benefit is that it shows that the entries the matrix of Q(m,n) are integral operators in weighted ℓ2 spaces. Below we explicitly write down those operators. For α, β = 1, 2, m 6= 0 and n ≥ 0, we define the following integral operators X αβ ℓ2 an−1+β (Z≥0) → ℓ2 an−1+α(Z≥0) by: m,n, Y αβ m,n : X αβ m,nr(k) = I (α) m,n(k) Y αβ m,nr(k) = K (α) m,n(k) m,n(i − β + 1) an−1+β(i − β + 1) r(i) (7.10) k ∞ c1,n(j) c1,n(j) Xi=k+1 i−β c2,n(j)! K (β) Yj=0 Xi=0 i−β c2,n(j)! I (β) Yj=0 Xi=0 k−1 Yj=i c2,n(j)! r(i) an(Z≥0) → ℓ2 k an(i) m,n(i − β + 1) an−1+β(i − β + 1) r(i). Z0,nr(k) = . (7.11) Also, for n ≥ 0, define integral operator Z0,n : ℓ2 an+1(Z≥0) by Lemma 7.3. The parametrix, Q(m,n) for the operator Am,n from above for m 6= 0 and n ≥ 0 is given by the following equivalent formula: Q(m,n)(cid:18) r(1)(k) r(2)(k) (cid:19) = 1 τm,n p(1) m,n(k) p(2) m,n(k) ! , 20 where SLAWOMIR KLIMEK AND MATT MCBRIDE p(1) m,n(k) = X 12 p(2) m,n(k) = X 22 m,nr(1)(k) + Y 12 m,nr(1)(k) + Y 22 m,nr(1)(k) + X 11 m,nr(1)(k) + X 21 m,nr(2)(k) + Y 11 m,nr(2)(k) + Y 21 m,nr(2)(k) m,nr(2)(k). Moreover when m = 0 and n ≥ 0 the parametrix is given by the following formula: Q(0,n)(cid:18) r(1)(k) r(2)(k) (cid:19) =(cid:18) 0 Z0,nr(2)(i) (cid:19) . Proof. Multiplying out the two matrices in (7.9) and applying it to the vector r(i) = (cid:0)r(1)(i), r(2)(i)(cid:1), and using the definitions of the integral operators, equation (7.10), gives the desired result. The case m = 0 immediately follows from equations (7.7) and (7.11). Thus the proof is complete. (cid:3) 8. Analysis of the Parametrix This section contains analysis of the parametrix Q(m,n), culminating in the proof of the main theorem 6.3. The case m = 0 is the easiest: all matrices are diagonal and the analysis is not any different than the analysis for the quantum disk [5] and [7]. Below we only consider the case m > 0 since for m < 0 the computations are virtually identical. We start by gathering the basic information about the quantum I and K functions. Lemma 8.1. For m > 0 the components −I (1) positive for all n, k ≥ 0. m,n(k), and K (2) m,n(k), K (1) m,n(k), I (2) m,n(k) are all Proof. This immediately follows from the definitions and the proof of Proposition 6.2. (cid:3) Lemma 8.2. We have the following inequalities for all n, k ≥ 0 and m > 0: − I (1) m,n(k) < −I (1) m,n(k + 1), I (2) m,n(k) < 1 c2,n(k) I (2) m,n(k + 1), K (1) m,n(k + 1) < 1 c1,n(k) K (1) m,n(k), K (2) m,n(k + 1) < K (2) m,n(k). Proof. Using recurrence relation (6.6), (6.7), the assumption that m > 0 and the previous lemma we have 0 < m an(k + 1) K (1) m,n(k + 1) = c2,n(k)K (2) m,n(k) − K (2) m,n(k + 1). Consequently, using c2,n(k) ≤ 1 , we have m,n(k + 1) < c2,n(k)K (2) proving one of the inequalities. Similarly, we have K (2) m,n(k) ≤ K (2) m,n(k), 0 < m an+1(k)c(k) K (2) m,n(k) = 1 c1,n(k) K (1) m,n(k) − K (1) m,n(k + 1), implying the other inequality involving Km,n. The proofs for the Im,n(k) are identical. (cid:3) DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 21 The next lemma is the crux of the compactness argument. It establishes estimates on the components of quantum I and K functions in a similar fashion to the estimates of the modified Bessel functions in [8]. Consider the quantity: The series above is convergent by summability assumptions on an(k). ε(m, n) = an+1(k) m2 + an(k)an+1(k) . ∞ Xk=0 Lemma 8.3. For m 6= 0 and n, k ≥ 0 the following inequalities hold: I (2) m,n(k) ≤ −mε(m, n)I (1) m,n(k+1) and K (1) m,n(k+1) ≤ m ε(m, n) + K (1) K (2) m,n(∞) m,n(∞)! K (2) m,n(k). Moreover ε(m, n) → 0 as m, n → ∞. Proof. Assuming m > 0 and using the recurrence relations (6.6), (6.7), we have I (2) m,n(k) −I (1) m,n(k + 1) = m an+1(k) + c1,n(k) 1 m an(k) +c2,n(k−1) (2) m,n (k−1) I −I (1) m,n(k) ≤ c1,n(k) 1 + m an(k) m2 an(k)an+1(k) + c2,n(k − 1) I (2) m,n(k − 1) −I (1) m,n(k) ! ≤ 1 + m an(k) m2 an(k)an+1(k) + I (2) m,n(k − 1) −I (1) m,n(k) , since c1,n(k) ≤ 1 and c2,n(k) ≤ 1 for all k ≥ 0, assuming I (2) rearranging the terms, that m,n(−1) = 0. It follows, by m,n(k) 1 m I (2) −I (1) m,n(k + 1) − I (2) m,n(k − 1) −I (1) m,n(k) ! ≤ 1 + 1 an(k) m2 . an(k)an+1(k) Summing both sides and telescoping the left side we get 1 m,n(k) m I (2) m,n(k + 1)! ≤ −I (1) 1 an(k) m2 an(k)an+1(k) 1 + = ε(m, n). ∞ Xk=0 This establishes the first inequality. To obtain the second inequality we proceed similarly, using the recurrence relations (6.6), (6.7), but this time telescoping sum goes to infinity. The next step is to show ε(m, n) goes to zero as m and n go to infinity. It follows immediately from the definition that: and the sum on the right goes to zero as n → ∞ from the condition on an(k), and hence ε(m, n) → 0 as n → ∞. Now for any η > 0 pick N > 0 such that we have ε(m, n) ≤ 1 an(k) , ∞ Xk=0 22 SLAWOMIR KLIMEK AND MATT MCBRIDE and pick M > 0 such that 1 an(k) ≤ η 2 , Xk>N 1 an(k) m2 an(k)an+1(k) ≤ η 2 1 + Xk≤N for m > M. It now follows that ε(m, n) → 0 as n, m → ∞. The case m < 0 is analogous, and thus the desired result follows. (cid:3) The next technical lemma deals with sums appearing in the definitions of the integral operators which comprise the parametrix. Lemma 8.4. The following summation estimates are true for m 6= 0 and n, k ≥ 0: ∞ Xi=k+1 i−2 Yj=0 c1,n(j)! K (2) m,n(i − 1) an+1(i − 1) ≤ 1 m k−1 Yj=0 c1,n(j)K (1) m,n(k) and K (1) m,n(i) an(i) ≤ 1 m K (2) m,n(k). ∞ Xi=k+1 Proof. As usual we consider only m > 0. Using equation (6.6), multiplying both sides of the equation by a product of c1,n(j)'s, and summing we get: ∞ Xi=k+1 i−2 Yj=0 c1,n(j)! K (2) m,n(i − 1) an+1(i − 1) = ≤ 1 m 1 m c1,n(j)K (1) m,n(i − 1) − c1,n(j)K (1) m,n(i)! i−1 Yj=0 ∞ Xi=k+1 i−2 Yj=0 Yj=0 k−1 c1,n(j)K (1) m,n(k), where the last inequality is true because the difference above is a telescoping sum. The second inequality follows from the recurrence relation (6.7) and the same telescoping sum trick. This completes the proof. (cid:3) The next and final lemma in preparation for the proof of the main result deals with pointwise estimates of the products of different components of quantum I and K functions through the quantity τm,n = hKm,n(0), Im,n(0)⊥i. Lemma 8.5. If m 6= 0 then, for all n ≥ 0, K (1) m,n(k)I (2) m,n(k) ≤ τm,n −I (1) m,n(k + 1)K (2) m,n(k) ≤ τm,n c1,n(k) c2,n(i) c1,n(i) c2,n(i) c1,n(i) k−1 k−1 Yi=0 Yi=0 and − K (2) m,n(k)I (1) m,n(k) ≤ τm,n k−1 Yi=0 and I (2) m,n(k)K (1) m,n(k + 1) ≤ τm,n c1,n(k) c2,n(i) c1,n(i) c2,n(i) c1,n(i) . k−1 Yi=0 DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 23 Proof. Assume as usual that m > 0. Using the matrix valued recurrence relation (6.5) and properties of the inner product we get hKm,n(k + 1), Im,n(k + 1)⊥i = = c2,n(i) c1,n(i)* k Yi=0 Cm,n(i)Km,n(0), c2,n(i) c1,n(i) hKm,n(0), Im,n(0)⊥i = k k Yi=0 Yi=0 Writing out the above inner product gives k Yi=0(cid:0)Cm,n(i)−1(cid:1)t Im,n(0)⊥+ Yi=0 c2,n(i) c1,n(i) τm,n. k (8.1) hKm,n(k + 1), Im,n(k + 1)⊥i = K (1) m,n(k + 1)I (2) m,n(k + 1) + K (2) m,n(k + 1)[−I (1) m,n(k + 1)]. Consequently, the above equality (8.1) implies that K (1) m,n(k)I (2) m,n(k) ≤ τm,n c2,n(i) c1,n(i) and − K (2) m,n(k)I (1) m,n(k) ≤ τm,n c2,n(i) c1,n(i) . k−1 Yi=0 k−1 Yi=0 To get the second set of inequalities we use the above equation (8.1) and recurrence relations (6.6), (6.7) to get τm,n c2,n(i) c1,n(i) k−1 Yi=0 = −I (1) m,n(k)K (2) m,n(k) + I (2) m,n(k)K (1) m,n(k) m,n(k + 1) − m an+1(k) m,n(k + 1) + + I (2) =(cid:18)−c1,n(k)I (1) m,n(k)(cid:18)c1,n(k)K (1) = c1,n(k)(cid:0)−I (1) I (2) m K (2) m,n(k) m,n(k)(cid:19) K (2) m,n(k)(cid:19) m,n(k + 1)(cid:1) . m,n(k)K (1) an+1(k) m,n(k + 1)K (2) m,n(k) + I (2) From the last equality it follows that: I (1) m,n(k + 1)K (2) m,n(k) ≤ τm,n c1,n(k) and the lemma is proved. c2,n(i) c1,n(i) k−1 Yi=0 and I (2) m,n(k)K (1) m,n(k + 1) ≤ τm,n c1,n(k) k−1 Yi=0 c2,n(i) c1,n(i) , (cid:3) We have gathered now enough information to analyze compactness of the inverse of D. The Fourier transform decomposes that inverse into a direct sum of parametrices Qm,n, which in turn consist of the integral operators X αβ m,n, Z0,n. Those integral operators are compact operators, in fact even Hilbert-Schmidt operators as stated in the following theorem. m,n, Y αβ Theorem 8.6. Let n ≥ 0. If m 6= 0, then the integral operators X αβ m,n defined in equation (7.10) are Hilbert-Schmidt operators for α, β = 1, 2, and, if m = 0, the integral m,n and Y αβ 24 SLAWOMIR KLIMEK AND MATT MCBRIDE operator Z0,n defined in equation (7.11) is a Hilbert-Schmidt operator. Moreover the Hilbert- Schmidt norms of X αβ m,n and Y αβ m,n go to zero as m, n → ∞, and the Hilbert-Schmidt norm of Z (0,n) goes to zero as n → ∞. Proof. We start with the case m 6= 0. Using the definition of X αβ that for α, β = 1, 2 we have m,n and Y αβ m,n, it is easy to see kX αβ m,nk2 HS = kY αβ m,nk2 HS = an−1+α(k) m,n(k)(cid:17)2 (cid:16)I (α) m,n(k)(cid:17)2 (cid:16)K (α) an−1+α(k) ∞ Xk=0 Xk=0 ∞ ∞ c1,n(j) c2,n(j)!2 Xi=k+1 i−β Yj=0 c2,n(j)!2 Xi=0 i−β Yj=0 c1,n(j) k an−1+β(i − β + 1) m,n(i − β + 1)(cid:17)2 · (cid:16)K (β) m,n(i − β + 1)(cid:17)2 · (cid:16)I (β) an−1+β(i − β + 1) . (8.2) There are eight norms to estimate; however that number can be reduced to four, since by Fubini's Theorem we have: kX 11 kX 12 m,nk2 m,nk2 HS = kY 11 HS = kY 21 m,nk2 m,nk2 HS, HS, kX 22 kX 21 m,nk2 m,nk2 HS = kY 22 HS = kY 12 m,nk2 m,nk2 HS HS. As usual consider the case m > 0. Using Lemmas 8.2, 8.4, 8.5 in the formula for the Hilbert-Schmidt norm of X 11 m,n, we obtain: kX 11 m,nk2 HS = ≤ ≤ ∞ ∞ Xk=0 J1(n)2 J2(n)2 κ an(k) m,n(k)(cid:17)2 (cid:16)I (1) Xi=k+1 i−1 Yj=0 m,n(k)(cid:17)2 (cid:16)I (1) h−I (1) Xk=0 J1(n) J2(n) τm,nκ m ∞ ∞ · an(k) Xk=0 c1,n(j) c2,n(j)!2 · (cid:16)K (1) m,n(i)(cid:17)2 an(i) K (1) m,n(k + 1) K (1) m,n(i) an(i) ∞ Xi=k+1 m,n(k)i K (1) an(k) m,n(k + 1) . We estimate the last term in the following way: τm,nκJ1(n) mJ2(n) h−I (1) ∞ Xk=0 ≤ J2(n) J1(n) τ 2 m,nκJ1(n) J2(n) = τ 2 m,nκ ε(m, n) + m,n(k + 1) ∞ an(k) m,n(k)i K (1) Xk=0 ε(m, n) + m,n(∞)! s(n), K (1) K (2) m,n(∞) · K (2) K (2) m,n(k) m,n(k) K (1) K (2) m,n(∞) m,n(∞)! 1 an(k) by using Lemmas 8.3, 8.5, and the inequality 1/c1,n(k) ≤ κ. Therefore we have DIRAC TYPE OPERATORS ON THE QUANTUM SOLID TORUS 25 kX 11 m,nk2 HS ≤ τ 2 m,nκ ε(m, n) + K (1) K (2) m,n(∞) m,n(∞)! s(n). Similarly, using Lemmas 8.2, 8.4, and 8.5, we can estimate the Hilbert-Schmidt norm of m,n by: X 22 kX 22 m,nk2 HS = ∞ Xk=0 ∞ an+1(k) m,n(k)(cid:17)2 (cid:16)I (2) Xi=k+1 i−2 Yj=0 m,n(k)(cid:17)2 (cid:16)I (2) an+1(k) ∞ J1(n) J2(n)2 Xk=0 an+1(i − 1) c1,n(j) c2,n(j)!2 (cid:16)K (2) Xi=k+1 i−2 Yj=0 m,n(i − 1)(cid:17)2 c1,n(j)! K (2) ∞ K (2) m,n(k) m,n(i − 1) an+1(i − 1) ≤ ≤ J1(n)2 J2(n)2 τm,n m · J2(n) J1(n) I (2) m,n(k)K (2) an+1(k) m,n(k) . ∞ Xk=0 The above is then equal to τm,n m · J1(n) J2(n) ∞ Xk=0 m,n(k)K (2) I (2) an+1(k) m,n(k) · I (1) m,n(k + 1) I (1) m,n(k + 1) ≤ τ 2 m,nκ ε(m, n) an+1(k) = τ 2 m,nκε(m, n)s(n + 1), ∞ Xk=0 again by using Lemmas 8.3 and 8.5 and the fact that 1/c1,n(k) is less than κ. Thus we have Very similar arguments, using the same steps as above, show that kX 22 m,nk2 HS ≤ τ 2 m,nκε(m, n)s(n + 1). kX 12 m,nk2 HS ≤ τ 2 m,nκ ε(m, n) + K (1) K (2) m,n(∞) m,n(∞)! s(n) and kX 21 m,nk2 HS ≤ τ 2 m,nκε(m, n)s(n + 1). Therefore for m 6= 0 the Hilbert-Schmidt norms are finite for all of the operators. When m = 0 we have: kZ0,nk2 HS = since again c2,n(k) ≤ 1. 1 an+1(k) ∞ Xk=0 k Xi=0 k−1 Yj=i c2,n(j)!2 1 an(i) ≤ s(n)s(n + 1), By assumption s(n) goes to zero as n → ∞, Lemma 8.3 implies that ε(m, n) → 0 as m, n → ∞, and the boundary condition given in equation (6.4) requires that K (1) K (2) m,n(∞) m,n(∞) → 0 as m → ∞. uniformly in n. Consequently, the Hilbert-Schmidt norms of X αβ m,n go to zero as m, n → ∞, and the Hilbert-Schmidt norm of Z (0,n) goes to zero as n → ∞, and the proof is finished. (cid:3) m,n and Y αβ 26 SLAWOMIR KLIMEK AND MATT MCBRIDE The following statement is an immediate consequence of the previous theorem, since the parametrix Q(m,n) is comprised of the integral operators estimated in it. Corollary 8.7. The parametrix Q(m,n) for m 6= 0 and n ≥ 0 is a Hilbert-Schmidt operator and the Hilbert-Schmidt norm of Q(m,n) goes to zero as m, n → ∞. Moreover for m = 0 the parametrix Q(0,n) is a Hilbert-Schmidt operator and its Hilbert-Schmidt norm goes to zero as n → ∞. We can now close out this section by proving the main theorem of this paper, that the Dirac operator defined by equation (4.3) subject to the boundary condition (6.4) has a compact inverse. Proof. (Proof of Theorem 6.3) It follows from Proposition 4.1 that inverse Q of the Dirac operator D is (essentially) a direct sum of Q(m,n) and its analogs for "negative" terms in the Fourier decomposition. Corollary 8.7 shows that Q(m,n) is a Hilbert-Schmidt operator for all m ∈ Z and n ≥ 0. Moreover the same corollary stipulates that the Hilbert-Schmidt norms of Q(m,n) go to zero as m, n → ∞. This means that since Q is a direct sum of compact operators with norms going to zero, which implies that Q is a compact operator. Thus the proof is complete. (cid:3) References [1] Ahlfors, L., Complex Analysis, McGraw-Hill, 1979. [2] Atiyah, M. F., Patodi, V. K. and Singer I. M., Spectral asymmetry and Riemannian geometry I, II, III, Math. Proc. Camb. Phil. Soc. 77, 43-69, 78, 43-432, 79, 71-99, 1975-1976. [3] Davidson, K., C ∗−algebras by Example, American Mathematical Society, 1991. [4] Klimek, S. and Lesniewski, A., Quantum Riemann surfaces, I. The unit disk, Comm. Math. Phys., 146, 103-122, 1992. [5] Klimek, S. and McBride M., D-bar Operators on Quantum Domains. Math. Phys. Anal. Geom., 13, 357 - 390, 2010. [6] Klimek, S. and McBride M., A note on Dirac Operators on the Quantum Punctured Disk. SIGMA, 6, 56 - 68, 2010. [7] Klimek, S. and McBride M., A Note on Gluing Dirac Type Operators on a Mirror Quantum Two-Sphere. SIGMA, 10, 36 - 51, 2014. [8] Klimek, S. and McBride M., Global boundary conditions for a Dirac operator on the solid torus. Jour. Math. Phys., 52, 1 - 14, 2011. [9] Mishchenko, A. V. and Sitenko, Yu, Spectral Boundary Conditions and Index Theorem for Two- Dimensional Compact Manifold with Boundary. Annals of Physics, 218, 199 - 232, 1992. [10] Reed, M. and Simon, B., Functional Analysis, Academic Press, 1980. [11] Trgo, A., Monodromy Matrix for Linear Difference Operators with Almost Constant Coefficients. Jour. Math. Anal., 194, 697 - 719, 1995. [12] Williams, D., Crossed Products of C ∗−algebras, American Mathematical Society, 2007. Department of Mathematical Sciences, Indiana University-Purdue University Indianapo- lis, 402 N. Blackford St., Indianapolis, IN 46202, U.S.A. E-mail address: [email protected] Department of Mathematics and Statistics, Mississippi State University, 175 President's Cir., Mississippi State, MS 39762, U.S.A. E-mail address: [email protected]
1602.04807	1	1602	2016-02-15T09:53:32	Woronowicz's Tannaka-Krein duality and free orthogonal quantum groups	[ "math.OA", "math.CT", "math.QA" ]	Given a finite dimensional Hilbert space H and a collection of operators between its tensor powers satisfying certain properties, we give a category-free proof of the existence of a compact quantum group G with a fundamental representation U on H such that the intertwiners between the tensor powers of U coincide with the given collection of operators. We then explain how the general version of Woronowicz's Tannaka-Krein duality can be deduced from this.	math.OA	math	WORONOWICZ'S TANNAKA-KREIN DUALITY AND FREE ORTHOGONAL QUANTUM GROUPS SARA MALACARNE Abstract. Given a finite dimensional Hilbert space H and a collection of operators between its tensor powers satisfying certain properties, we give a category-free proof of the existence of a compact quantum group G with a fundamental representation U on H such that the intertwiners between the tensor powers of U coincide with the given collection of operators. We then explain how the general version of Woronowicz's Tannaka-Krein duality can be deduced from this. Introduction The aim of the paper is to give a category-free proof of Woronowicz's Tannaka-Krein duality Theorem [5]. We consider a finite dimensional Hilbert space H and a collection of operators between its tensor powers satisfying certain properties. Categorically speaking, we deal with a C∗-tensor category with conjugates that is a subcategory of the category of finite dimensional Hilbert spaces, Hilbf , and assume that such category is generated by one self-conjugate Hilbert space. We prove the existence of a compact quantum group G, such that its representation category Rep G is our given category. The proof consists of an explicit reconstruction of the Hopf ∗-algebra C[G], sometimes denoted by Pol G, generated by the coefficients of all finite dimensional representations of G. The relations defining such Hopf ∗-algebra are directly obtained through morphisms in the category, or equivalently, through the collection of operators between tensor powers of H. The version of Woronowicz's Tannaka-Krein Theorem that we prove is essentially formulated in the paper by T. Banica and R. Speicher [1], where the duality is used for the construction of new examples of free quantum groups, via subcategories of the so-called categories of noncrossing partitions. Even though the proof here presented is, in many respects, similar to the proofs of Woronowicz's Tannaka-Krein duality appearing in [5] and [2], we wish to point out that this version is more algebraic, mostly category-free and the key part is based on simple duality statements for finite dimensional vector spaces. Related to this reconstruction process, it is also important to mention P. Schauenburg's paper [3], in which a proof of Tannaka-Krein duality is given in a more general setting: monoidal categories that are not semisimple are considered and, correspondingly, arbitrary Hopf algebras are recovered. The paper is structured in the following way: in Section 1 we define a bialgebra, which we will later prove to be the Hopf ∗-algebra C[G]. The relations defining this bialgebra are obtained from the collection of operators between tensor powers of H, denoted by C. In Section 2 we prove that the bialgebra defined in Section 1 can be equipped with a Hopf ∗-algebra structure. For this we first consider a smaller collection of operators, CF , and show that they define the free orthogonal quantum group O+ F , [4]. We want to stress that nothing, apart from the fact that C[O+ F ] is a well-defined Hopf ∗-algebra, is used. In Section 3 we prove the equivalence between the C∗-tensor category generated by one selfdual Hilbert space H and Rep G. In Section 4 we show how the particular case analysed in Section 3 can be extended to the general case of a not necessarily finitely generated C∗-tensor category. Acknowledgments. I would like to thank my supervisor Sergey Neshveyev for his help and precious advice throughout this work. I am grateful to Teodor Banica for useful suggestions. Thanks also to Marco Matassa for fruitful discussions. Date: February 15, 2016. Supported by the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013)/ ERC Grant Agreement no. 307663 (PI: S. Neshveyev). 1 1. Singly generated categories of Hilbert spaces Our goal is to prove the following version of Woronowicz's Tannaka-Krein duality. Theorem 1.1. Let H be a finite dimensional Hilbert space. Suppose we are given a collection C of spaces C(k, l) of operators H ⊗k → H ⊗l for all k, l ≥ 0 satisfying the following properties: (1) if T, S ∈ C, then T ⊗ S ∈ C; (2) if T, S ∈ C are composable, then T S ∈ C; (3) T ∈ C implies T ∗ ∈ C; (4) C(k, k) contains the identity operator for all k ≥ 0; (5) C(0, 2) contains an operator R such that (R∗ ⊗ ι)(ι ⊗ R) = ±ι on H. Then there exists a unique up to isomomorphism compact quantum group G with a self-conjugate funda- mental representation U on H such that HomG(H ⊗k, H ⊗l) = C(k, l) for all k, l ≥ 0. In the last section we will discuss how the general form of Woronowicz's Tannaka-Krein duality can be easily deduced from this. Denote by A the tensor algebra of the space of linear functionals on B (H), i.e., A := T (cid:0)B (H)∗(cid:1) = ∞ Mk=0 B(cid:0)H ⊗k(cid:1)∗ . Let U ∈ B (H) ⊗ B(H)∗ ⊂ B (H) ⊗ A be the "fundamental matrix" of A, so U is characterized by the property In other words, (ι ⊗ T )(U ) = T ∀T ∈ B(H). (1.1) where the eij 's are matrix units in B (H) and {uij}i,j is the dual basis of B (H)∗ such that uij (ekl) = δikδjl. The tensor algebra A is a bialgebra with comultiplication ∆ defined by duality from the multiplication on B(H), so that eij ⊗ uij, U :=Xi,j or equivalently, using the leg-numbering notation, (ι ⊗ ∆)(U ) = U12U13. ∆ (uij) =Xk uik ⊗ ukj, Next, denote by An ⊂ A the subspace given by An := and denote by Bn the commutant Bn :=  n Mk,l=0 Finally, let B(cid:0)H ⊗k(cid:1)∗ n Mk=0 C(k, l)  ′ ⊆ n Mk=0 = n Mk=0 B(cid:0)H ⊗k(cid:1)!∗ B(H ⊗k) ⊂ B n Mk=0 , H ⊗k! . and denote by I the union I :=S∞ Lemma 1.2. I is a bi-ideal in the bialgebra A. In :=(cid:8)a ∈ An : aBn = 0(cid:9) , n=0 In. Note that In+1 ∩ An = In, so I is a subspace of A. Proof. We will first prove that I is an ideal. Assume a ∈ In and b ∈ B(H ⊗m)∗; we have to check that a ⊗ b vanishes on (cid:16)⊕n+m k,l=mC(k, l)(cid:17)′ . Since (cid:16)⊕n+m k,l=mC(k, l)(cid:17)′ simply follows from the assumption that a vanishes on (⊕n ⊆ (cid:16)⊕n k,l=0C(k, l)(cid:17)′ k=0C(k, l))′. ⊗(cid:16)⊕m k,l=0C(k, l)(cid:17)′ , the statement To prove that I is a coideal we have to show that ∆(I) ⊆ I ⊗ A + A ⊗ I. For this purpose we use an equivalent definition of I, that is, we consider the space spanned by the slices (ω ⊗ ι)(cid:0)(T ⊗ 1)U ⊗k − U ⊗l(T ⊗ 1)(cid:1) 2 (1.2) for all ω ∈ B(H ⊗k, H ⊗l)∗, T ∈ C(k, l) and k, l ≥ 0. This space indeed coincides with I, since using (1.1) k=0B(H ⊗k) vanishes on the elements (1.2) for all k, l ≤ n if and only it lies we see that an operator S ∈ ⊕n in Bn. We choose an orthonormal basis of H ⊗k, {ξj}j, and of H ⊗l, {ηi}i, and assume ω is of the form ωij = h · ξj , ηii. We set V := U ⊗k and W := U ⊗l. Then, using the leg-numbering notation, ∆(ωij ⊗ ι) ((T ⊗ 1)V − W (T ⊗ 1)) is equal to The expression in the parentheses can be written as (ωij ⊗ ι ⊗ ι) ((T ⊗ 1 ⊗ 1)V12V13 − W12W13(T ⊗ 1 ⊗ 1)) . ((T ⊗ 1 ⊗ 1)V12 − W12(T ⊗ 1 ⊗ 1)) V13 + W12 ((T ⊗ 1 ⊗ 1)V13 − W13(T ⊗ 1 ⊗ 1)) . Now, if we just consider the first part of the sum (ωij ⊗ ι ⊗ ι) (((T ⊗ 1 ⊗ 1)V12 − W12(T ⊗ 1 ⊗ 1)) V13) , it can be expressed as (ωik ⊗ ι ⊗ ι) ((T ⊗ 1 ⊗ 1)V12 − W12(T ⊗ 1 ⊗ 1)) (ωkj ⊗ ι ⊗ ι)(V13), Xk which belongs to I ⊗ A. Similarly the other part lies in A ⊗ I, as wanted. By the previous Lemma, A/I is a bialgebra. What we wish to prove is that A/I ∼= C[G], for a compact quantum group G, and to do so we need for A/I to be a Hopf ∗-algebra and for U to be unitary (see Theorem 1.6.7 of [2]), and this is not obvious written in this manner. In fact it is not even clear whether A/I has a ∗-structure. We shall proceed with an intermediate step. The idea is the following: we will introduce another bi-ideal in A, IF , and show that A/IF ∼= C[O+ F is the free orthogonal quantum group. Thus, A/IF will automatically inherit a Hopf ∗-algebra structure. Finally we will show that I/IF is a Hopf ∗-ideal in A/IF and then conclude that there exists a compact quantum group G such that F ], where O+ (cid:3) again by Theorem 1.6.7 of [2]. A/I ∼= A/IF.I/IF ∼= C[G], 2. Representation category of a free orthogonal quantum group Following the strategy described above, we now consider the case when C is the smallest collection of spaces as in Theorem 1.1 containing a fixed operator R : C → H ⊗2 such that (R∗ ⊗ ι)(ι ⊗ R) = ±ι. It is known, and not difficult to see, that if we fix an orthonormal basis e1, . . . , en in H, then R has the form (ι ⊗ F )r, where r : C → H ⊗ H is given by r(1) =Pi ei ⊗ ei and F ∈ GLn(C) is such that F ¯F = ±1, where ¯F is the matrix obtained from F by taking the complex conjugate of every entry. We will use the subindex F for the constructions of the previous section related to this smallest collection, so we write CF , BF,n, IF,n, etc. Consider the universal unital algebra C[O+ F ] generated by entries of a matrix U = (uij )i,j satisfying the relations U F tU t(F −1)t = 1, F tU t(F −1)tU = 1. for an invertible n by n matrix F . It is again known and easy to see that this is a Hopf ∗-algebra with comultiplication ∆(uij) =Pk uik ⊗ ukj and involution given by U ∗ = F tU t(F −1)t. The compact quantum group O+ F thus defined is known in literature as the free orthogonal quantum group [4], but we do not need to know any properties of this quantum group apart from the fact that it is well-defined. The following Lemma is a simple consequence of our definitions. Lemma 2.1. We have A/IF ∼= C[O+ F ]. 3 Proof. By definition, the bialgebra C[O+ elements F ] can be written as A/L where L is the ideal generated by the with U ∈ B(H) ⊗ A being the fundamental matrix of A. In order to prove the Lemma we need to show that IF = L. To show that L ⊆ IF , consider the linear functionals ω1,ij := h · 1, ei ⊗ eji ∈ B(C, H ⊗2)∗ and ωij,1 := h · ei ⊗ ej, 1i ∈ B(H ⊗2, C)∗. Then (cid:0)F t − (U F tU t)(cid:1)ij , (cid:0)(F −1)t − U t(F −1)tU(cid:1)ij ∀i, j and (cid:0)F t − (U F tU t)(cid:1)ij = (ω1,ij ⊗ ι)(cid:0)(R ⊗ ι) − U ⊗2(R ⊗ ι)(cid:1) where we recall that R = (ι ⊗ F )r and in the second equality we use that F ∗ = ±(F −1)t. Hence L ⊆ IF . (cid:0)(F −1)t − U t(F −1)tU(cid:1)ij = ±(ωij,1 ⊗ ι)(cid:0)(R∗ ⊗ ι) − (R∗ ⊗ ι)U ⊗2(cid:1) , Conversely, let us show that IF ⊆ L. As follows from the above identities, R and R∗ are morphisms in F . But this implies IF ⊆ L F . It follows that any operator in CF is a morphism in Rep O+ the category Rep O+ because any relation defined by elements of CF has to be satisfied in C[O+ F ]. Therefore in order to prove Theorem 1.1 for C = CF it remains to establish the following Lemma. We remark that the Lemma itself is not needed for the general case, but its proof will be reused. Lemma 2.2. We have HomO+ F (H ⊗k, H ⊗l) = CF (k, l) for all k, l ≥ 0. (cid:3) Proof. We have to show that for every n n As both sides are (finite dimensional) von Neumann algebras, the equality is equivalent to the equality of their commutants: CF (k, l) = EndO+ Mk,l=0 F n Mk=0 H ⊗k! . BF,n = EndO+ F n Mk=0 H ⊗k!′ . Recall now that for any finite dimensional representation V ∈ B(HV ) ⊗ C[G] of a compact quantum group G we have a representation πV of the algebra C[G]∗ on HV defined by πV (φ) = (ι ⊗ φ)(V ), and then πV (C[G]∗) = EndG(HV )′. Therefore we have to show that πLn immediately follows from the previous Lemma, as C[O+ F ]∗(cid:1) = BF,n. But this k=0 U ⊗k(cid:0)C[O+ F ] = A/IF ⊃ An/IF,n = B∗ F,n. (cid:3) 3. Proof of the Theorem We now turn to a general C as in Theorem 1.1. Let R ∈ C(0, 2) be an operator such that (R∗ ⊗ ι)(ι ⊗ R) = ±ι. As in the previous section, we fix an orthonormal basis e1, . . . , en in H and write R as (ι ⊗ F )r. Denote by J the bi-ideal I/IF in C[O+ F ] = A/IF . Note that J can still be described as the space spanned by the elements (ω ⊗ ι)(cid:0)(T ⊗ 1)U ⊗k − U ⊗l(T ⊗ 1)(cid:1), for T ∈ C(k, l) and ω ∈ B(H ⊗k, H ⊗l)∗, where we slightly abuse the notation and denote by the same symbol U the fundamental matrix of A and its image in C[O+ F ]. Lemma 3.1. J is a Hopf ∗-ideal in C[O+ F ]. Proof. We denote by S the antipode of C[O+ under taking the adjoints and is invariant under S. Let a∗ = (ω ⊗ ι)(cid:0)(T ⊗ 1)U ⊗k − U ⊗l(T ⊗ 1)(cid:1)∗ We have to show that it lies in J for any k, l ≥ 0. We first note that C(1, 1) is closed under the operation . F ]. Since J is a bi-ideal, we only need to check that J is closed ∨ defined by (ι ⊗ T )R = (T ∨ ⊗ ι)R, since T ∨ = ±(ι ⊗ R∗)(ι ⊗ T ⊗ ι)(R ⊗ ι). As (ι ⊗ T )r = (T t ⊗ ι)r, we have T ∨ = (F −1T F )t, and the inverse operation, still preserving C(1, 1), is T 7→ F T tF −1. We recall from the previous section that we also have U ∗ = F tU t(F −1)t. Analogous formulas hold for T ∈ C(k, l). In fact, if we denote by Fk = F ⊗k and Uk = U ⊗k, 4 then T ∨ = (F −1 l T Fk)t ∈ C(l, k) and U ∗ k = F t kU t k(F −1 k )t. Therefore, choosing an orthonormal basis of H ⊗k, {ξj}j, and of H ⊗l, {ηi}i, and assuming ω is of the form ωij = h · ξj , ηii, we have a∗ =(ωij ⊗ ι) ((T ⊗ 1)Uk − Ul(T ⊗ 1))∗ = (ωji ⊗ ι) (U ∗ k (T ∗ ⊗ 1) − (T ∗ ⊗ 1)U ∗ l ) = k )t − (U ∗ k )t − (U ∗ k UkFk) − (F −1 l )t((T ∗)t ⊗ 1)(cid:1)t(cid:17) = l )t((T ∗)t ⊗ 1)(cid:1) = l UlFl)((T ∗)t ⊗ 1)(cid:1) = k ⊗ 1)Uk − Ul(Fl(T ∗)tF −1 (ωji ⊗ ι)(cid:16)(cid:0)((T ∗)t ⊗ 1)(U ∗ (ωij ⊗ ι)(cid:0)((T ∗)t ⊗ 1)(U ∗ (ωij ⊗ ι)(cid:0)((T ∗)t ⊗ 1)(F −1 (ωij ⊗ ι)(cid:0)(F −1 Xm,n (F −1 l l ⊗ 1)(cid:0)(Fl(T ∗)tF −1 )im(Fk)nj (ωmn ⊗ ι)(cid:16)( T ⊗ ι)Uk − Ul( T ⊗ ι)(cid:17) , k ⊗ 1)(cid:1) (Fk ⊗ 1)(cid:1) = where T = Fl(T ∗)tF −1 operation of ∨. Hence, a∗ ∈ J . k ∈ C(k, l), since T 7→ FlT tF −1 k is a map from C(l, k) to C(k, l), being the inverse The invariance of J under the antipode immediately follows from its invariance under involution. If a = (ωij ⊗ ι) ((T ⊗ 1)Uk − Ul(T ⊗ 1)) then S(a) =(ωij ⊗ ι) ((T ⊗ 1)(ι ⊗ S)(Uk) − (ι ⊗ S)(Ul)(T ⊗ 1)) =(ωij ⊗ ι) ((T ⊗ 1)U ∗ =(ωji ⊗ ι) (Uk(T ∗ ⊗ 1) − (T ∗ ⊗ 1)Ul)∗ ∈ J . l (T ⊗ 1)) k − U ∗ Given the above Lemma, we conclude that there exists a compact quantum group G such that A/I ∼= C[G]. It remains to show that HomG(H ⊗k, H ⊗l) = C(k, l). But this is done in exactly the same way as in Lemma 2.2. (cid:3) To finish the proof of Theorem 1.1 we have to show that the compact quantum group G is unique up to isomorphism. Let G′ be another compact quantum group satisfying the assumptions of Theorem 1.1, that is, having a fundamental representation V = (vij )ij on H such that HomG′(H ⊗k, H ⊗l) = C(k, l). We can identify C[G′] with A/I ′, for a bi-ideal I ′ ⊂ A. The only thing to check is that the bi-ideal I ′ is completely determined by the operator spaces C(k, l). Since C(k, l) and HomG′(H ⊗k, H ⊗l) coincide, from the proof of Lemma 2.2 we see that this implies that An/I ′ n. Hence, the spaces I ′ n are completely determined by the spaces C(k, l). Thus Theorem 1.1 is proved. n = I ′ ∩ An and I ′ = Sn≥0 I ′ n, where I ′ n = B∗ 4. General version of the Tannaka-Krein duality In this section we want to explain, without too many details, how using Theorem 1.1 one can recover the following result. Theorem 4.1 (Woronowicz's Tannaka-Krein duality). Let C be an essentially small C∗-tensor category with conjugates, τ : C → Hilbf be a unitary fiber functor. Then there exists a compact quantum group G and a unitary monoidal equivalence θ : C → Rep G such that τ is naturally unitarily monoidally isomorphic to the composition of the canonical fiber functor π : Rep G → Hilbf with θ. Furthermore, the Hopf ∗-algebra (C[G], ∆) for such a G is uniquely determined up to isomorphism. We remark that for C∗-tensor categories we follow the conventions of [2], in particular, we assume that they are closed under finite direct sums and subobjects. We concentrate only on the existence of G. We may assume that C is a subcategory of Hilbf and τ is the embedding functor. If there exist an object H in C such that any other object is isomorphic to a subobject of H ⊗n for some n ≥ 0, and a morphism R : C → H ⊗ H such that (R∗ ⊗ ι)(ι ⊗ R) = ±ι , then the result follows from Theorem 1.1. For general C let us distinguish between three cases: (i) C is generated, as a C∗-tensor category with conjugates, by one object; 5 (ii) C is generated by a finite number of objects; (iii) C is infinitely generated. (i) Assume C is generated by one object K, so every object of C is isomorphic to a subobject of a tensor product of copies of K and an object ¯K conjugate to K. Let (R′, ¯R′) be a solution of the conjugate equations for K and ¯K. Then letting H = K ⊕ ¯K and R = R′ ⊕ ¯R′, considered as a morphism C → H ⊗ H, we have (R∗ ⊗ ι)(ι ⊗ R) = ι, so we are back to the case covered by Theorem 1.1. (ii) The case when C is generated by a finite number of objects H1, . . . , Hn is not much different from (i), as then C is generated by H1 ⊕ · · · ⊕ Hn. (iii) For general C, choose a generating set F in C and let E be the family of finite subsets of F ordered by inclusion. For each E ∈ E let CE be the full rigid C∗-tensor subcategory of C generated by the finite set of objects in E. By the previous case, for each subcategory CE we get a compact quantum group GE with representation category CE. Moreover, if E ⊂ E′, then, since CE ⊂ CE′, by the uniqueness part of Theorem 1.1, the quantum group GE is a quotient of GE′ , that is, we have an embedding C[GE] ֒→ C[GE′ ] of Hopf ∗-algebras. Then C[G] is defined as the inductive limit of the Hopf ∗-algebras C[GE]. In the following example we can see how to recover the free unitary quantum group following the procedure explained in point (i) of the above. unitary quantum group. Example 4.2 (Free Unitary quantum group). We denote by C[U + Q ] the universal unital ∗-algebra generated by the entries of matrices V = (vij )i,j and ¯V = (¯vij )i,j such that V and ¯V are unitary with involution defined by V ∗ = Qt ¯V t(Q−1)t and ¯V ∗ = (Q−1)∗V tQ∗, for an invertible n by n matrix Q. The algebra C[U + Q ] Q is known in literature as the free is a Hopf ∗-algebra with comultiplication ∆(vij ) = Pk vik ⊗ vkj and U + We wish to prove the equivalence between the representation category of the free unitary quantum group and a concrete C∗-tensor category having certain properties. More specifically, consider the Hilbert space K = Cn and its complex conjugate ¯K. Let CQ be the smallest collection of operators between tensor powers of H := K ⊕ ¯K, as in Theorem 1.1, containing the operator R : C → H ⊗2 such that (R∗ ⊗ ι)(ι ⊗ R) = ι, and the projection p : K ⊕ ¯K → K. The operator R is equal to (ι ⊗ F )r for F ∈ GL2(Mn(C)) with entries F11 = F22 = 0, F12 = ¯Q−1 and F21 = Q, where ¯Q is the matrix whose coefficients are the complex conjugates of the entries of Q. We claim that HomU + Q (H ⊗k, H ⊗l) = CQ(k, l) for all k, l ≥ 0. We show that C[U + Q ] ∼= A/I, where I is the ideal generated by slice maps (ω ⊗ ι)(cid:0)(T ⊗ 1)U ⊗k − U ⊗l(T ⊗ 1)(cid:1), for T ∈ CQ(k, l) and ω ∈ B(H ⊗k, H ⊗l)∗. The claim will then follow from Theorem 1.1. By definition, C[U + ideal generated by the relations Q ] can be written as A/L where L is the U F tU t(F −1)t = 1, F tU t(F −1)tU = 1, U12 = 0, U21 = 0. We already know that the ideal I contains slices of the first two relations, since we showed in Lemma 2.1 that they correspond to the slice maps (ω1,ij ⊗ ι)(cid:0)(R ⊗ ι) − U ⊗2(R ⊗ ι)(cid:1)i,j , (ωij,1 ⊗ ι)(cid:0)(R∗ ⊗ ι)U ⊗2 − (R∗ ⊗ ι)(cid:1)i,j . The other two relations correspond to (ωij ⊗ι) ((p ⊗ ι)U − U (p ⊗ ι))i,j. Hence L ⊆ I. The opposite inclusion follows analogously to the second part of the proof of Lemma 2.1. References [1] T. Banica and R. Speicher. Liberation of Orthogonal Lie groups. Advances in Mathematics, 222: 1461-1501, 2009. [2] S. Neshveyev and L. Tuset. Compact quantum groups and their representation categories, volume 20 of Cours Sp´ecialis´es. Societe Math´ematique de France, Paris, 2013. [3] P. Schauenburg. Tannaka duality for Arbitrary Hopf Algebras. Algebra Berichte 66, Fischer, Munich, 1991. [4] A. Van Daele and S. Wang. Universal Quantum Groups. Int. J. Math., 7(2): 255-263, 1996. [5] S. L. Woronowicz. Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU(N) groups. Invent. Math., 93(1): 35-76, 1988. 6 E-mail address: [email protected] Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, NO-0316 Oslo, Norway 7
1203.5063	3	1203	2012-09-06T20:26:21	Closed quantum subgroups of locally compact quantum groups	[ "math.OA", "math.FA" ]	We investigate the fundamental concept of a closed quantum subgroup of a locally compact quantum group. Two definitions - one due to S.Vaes and one due to S.L.Woronowicz - are analyzed and relations between them discussed. Among many reformulations we prove that the former definition can be phrased in terms of quasi-equivalence of representations of quantum groups while the latter can be related to an old definition of Podle\'s from the theory of compact quantum groups. The cases of classical groups, duals of classical groups, compact and discrete quantum groups are singled out and equivalence of the two definitions is proved in the relevant context. A deep relationship with the quantum group generalization of Herz restriction theorem from classical harmonic analysis is also established, in particular, in the course of our analysis we give a new proof of Herz restriction theorem.	math.OA	math	CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN Dedicated to Leonid Vainerman on the occasion of his 65th birthday Abstract. We investigate the fundamental concept of a closed quantum subgroup of a locally compact quantum group. Two definitions -- one due to S. Vaes and one due to S.L. Woronowicz -- are analyzed and relations between them discussed. Among many reformulations we prove that the former definition can be phrased in terms of quasi-equivalence of representations of quantum groups while the latter can be related to an old definition of Podle´s from the theory of compact quantum groups. The cases of classical groups, duals of classical groups, compact and discrete quantum groups are singled out and equivalence of the two definitions is proved in the relevant context. A deep relationship with the quantum group generalization of Herz restriction theorem from classical harmonic analysis is also established, in particular, in the course of our analysis we give a new proof of Herz restriction theorem. 1. Introduction In this paper we study the notion of a closed quantum subgroup of a locally compact quan- tum group. The theory of quantum groups phrased in operator algebra language is already well established as a rapidly developing field on the border between noncommutative geometry and abstract harmonic analysis. Nevertheless, the fundamental notion of a closed (quantum) subgroup has not received enough attention so far. There have been several "working definitions" of such an object, but most efforts were directed toward developing other aspects of the theory. The first to look at quantum subgroups of (compact) quantum groups was P. Podle´s ([31, 32], see also a later discussion in [30]). His view was motivated by the straightforward noncommutative generalization of the inclusion homomorphism from the subgroup to the group and required the existence of a surjective ∗-homomorphism between the algebras of continuous functions on respective quantum groups. This point of view, however, has many disadvantages and drastically limits the number of subgroups (e.g. many quantum groups do not have the trivial subgroup in this sense). Soon it was realized that in the context of compact quantum groups one should rather require the ex- istence of a surjective ∗-homomorphism between the universal versions of algebras of continuous functions on respective quantum groups. This approach, adopted for example in [3] and [4], avoids the problems mentioned above and also enables a purely algebraic reformulation in terms of the underlying Hopf ∗-algebras. It was not clear, however, whether it would lead to a satisfactory notion for arbitrary locally compact quantum groups. In 2005 in [43] S. Vaes proposed another definition of a closed quantum subgroup of a locally compact quantum group, phrased in the language of von Neumann algebras. This definition was used in the same paper to develop the full force of the theory of induced representations and homogeneous spaces for quantum groups. Earlier another definition of a closed subgroup of a locally compact quantum group was proposed in [44, Definition 2.9] by Vaes and Vainerman. We show that the definition of Vaes and that given by Vaes and Vainerman are equivalent. It should be stressed that the argument needed to show that the definitions of a closed quantum subgroup proposed in [43, 44] give the standard notion of a closed subgroup in classical case is quite subtle. It can be formulated as saying that an inclusion of a closed subgroup H into a locally compact Date: November 6, 2018. 2010 Mathematics Subject Classification. Primary: 20G42, 22D25, Secondary: 43A30, 46L89. Key words and phrases. Quantum group, quantum subgroup, representation, quasi-equivalence, Herz restriction theorem. 1 2 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN group G induces a normal inclusion of respective group von Neumann algebras vN(H) ֒→ vN(G) and is equivalent to the fact that the restriction to H of regular representation of G is quasi- equivalent to the regular representation of H. This is, in turn, equivalent to the conclusion of the Herz restriction theorem which says that the map of Fourier algebras associated to H ֒→ G is a surjective contraction ([15], cf. also Section 4). All this has been known to the experts for a long time (cf. [16, 42, 10, 11, 44]); a detailed proof can be found in the 2008 thesis of C. Zwarich ([52]). The definition given in [43] is very well adapted to the problems studied in that paper, but it was not clear whether it is optimal in other contexts and how it relates to the notion studied earlier for compact quantum groups. As mentioned above it is also relatively difficult to see that it actually generalizes the classical notion of a closed subgroup. Yet another possible definition, related to the recently introduced notion of morphisms between quantum groups ([27]), was suggested to us by S.L. Woronowicz. Woronowicz's definition is phrased entirely in the language of C∗-algebras and it is notably easier to see that it generalizes the ordinary notion of a closed subgroup of a locally compact group (see Section 4). The main focus of this paper is on understanding the relations between the definitions of a closed quantum subgroup of a locally compact quantum group given by Vaes and Woronowicz and providing their equivalent reformulations. The definition of Woronowicz is deeply connected with the notion of a C∗-algebra generated by a quantum family of multipliers (which we analyze in Subsection 1.1) and turns out to be equivalent to the reformulation of the original idea of Podle´s, i.e. corresponds to the existence of a surjective ∗- homomorphism between the universal versions of the algebras of continuous, vanishing at infinity, functions on respective locally compact quantum groups. On the other hand the definition of Vaes can be rephrased in a simplified way (still in the von Neumann algebraic language) and turns out to be intimately connected with the notion of quasi-equivalence of representations of quantum groups ([40], cf. Theorem 3.4). Moreover, we show that this definition of a closed subgroup of a quantum group is strongly tied to the generalization to quantum groups of the Herz restriction theorem (cf. Remark 3.8). We show that the definition of Vaes is stronger than the definition of Woronowicz, in the sense that if H and G are locally compact quantum groups and H is a closed quantum subgroup of G in the sense of Vaes, then it is also a closed quantum subgroup of G in the sense of Woronowicz. Further we prove that they are equivalent in all special cases one usually considers: classical groups (both definitions describe the standard notion of a closed subgroup), duals of classical groups (both definitions describe a group epimorphism in the opposite direction), compact quantum groups and discrete quantum groups (Sections 4, 5 and 6). In particular this opens the way to finding all compact quantum subgroups of a given locally compact quantum group G via the theory of idempotent states (as studied for example in [36]) since each compact quantum subgroup of G gives rise to a state on the algebra of functions on G, which is idempotent with respect to the convolution product, and such states can be sometimes computed directly using Fourier transform methods. In the context of compact quantum groups this strategy was employed in [14] to re-establish the list of all quantum subgroups of SUq(2), originally found by Podle´s in [32]. In the course of our investigation we make crucial use of the quantum group versions of the Fourier and Fourier-Stieltjes algebras (cf. [9, Section 8]). It is worth noting that our work produces a new proof of the classical Herz restriction theorem (see Section 4). In the group-dual case we use the results of M. Ilie and R. Stokke on weak∗-continuous maps of Fourier-Stieltjes algebras ([18]) which we are also able to generalize (to some extent) to the quantum group setting (Proposition 5.3). This exemplifies the connections of our article with recent extensions of noncommutative harmonic analysis to the context of locally compact quantum groups (see for example [17] and references therein). Finally let us note that the differences between the definitions of a closed quantum sub- group according to Vaes and Woronowicz bear a striking similarity to the interplay between the Kustermans-Vaes definition of a locally compact quantum group (formulated in [23]) and the def- inition of a quantum group used in [40, 27] and based on the theory of manageable and modular multiplicative unitaries ([50, 39]). Again, the former definition is stronger and in all examples one finds that the two approaches are equivalent. Moreover, in special cases of classical groups, duals CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 3 of classical groups, compact and discrete quantum groups we have results on existence of Haar measures, so the Kustermans-Vaes approach is equivalent with the one used by So ltan-Woronowicz. At the present stage of research in the theory of quantum groups it is very difficult to predict whether the definitions of a closed quantum subgroup given by Vaes and Woronowicz are equiv- alent. We conjecture that in the full generality they are different. However, it seems very likely that in large classes of well-behaved locally compact quantum groups, e.g. the regular or even semi-regular ones, the two definitions will turn out to be equivalent. Let us give now a brief description of the paper. In the remainder of this section we collect necessary preliminaries from the theory of C∗-algebras (Subsection 1.1), locally compact quantum groups (Subsection 1.2) and homomorphisms of quantum groups as defined in [27] (Subsection 1.3). Section 2 focuses on the theory of representations of quantum groups and the notion of quasi- equivalence of such representations. We also relate this notion to the problem of generation of C∗- algebras by quantum families of multipliers, which later turns out to be crucial for understanding the interplay between the definitions of closed quantum subgroups given by Vaes and Woronowicz. These are introduced in Section 3 with the relations between them unraveled. We provide several equivalent reformulations of either definition and show that the former implies the latter (in the sense described above). We also give sufficient conditions for the two definitions to be equivalent. Section 4 is devoted to the study of both definitions of a closed quantum subgroup in the special case of classical groups. We prove there in detail that both are equivalent to the standard definition of a closed subgroup and discuss the direct connection between the definition of Vaes and the Herz restriction theorem. Then in Section 5 we conduct a similar investigation for the case of duals of classical groups. In this case also the definitions of Vaes and Woronowicz agree. Finally in Section 6 we show that the two definitions are equivalent for compact and discrete quantum groups (more precisely a compact quantum group H is a closed subgroup of a locally compact quantum group G in the sense of Vaes if and only if it is a closed subgroup of G in the sense of Woronowicz, and a similar result holds for subgroups of discrete quantum groups). 1.1. C∗-algebras and morphisms. Throughout the paper we will use the language of the theory of C∗-algebras as introduced in [47, 48, 49, 24]. In particular for C∗-algebras A and B a morphism from A to B is a ∗-homomorphism Φ from A into the multiplier algebra M(B) of B which is non-degenerate, i.e. the set Φ(A)B of linear combinations of products of the form Φ(a)b (a ∈ A, b ∈ B) is dense in B (by the Cohen factorization theorem this is equivalent to the condition that Φ(A)B = B). The set of all morphisms from A to B will be denoted by Mor(A, B). The non-degeneracy of morphisms ensures that each Φ ∈ Mor(A, B) extends uniquely to a unital ∗-homomorphism M(A) → M(B) which we will sometimes denote by ¯Φ. This also defines the operation of composition of morphisms (see [47, 49, 24]). For a Hilbert space H the C∗-algebra of compact operators on H will be denoted by K(H). Any C∗-algebra A acting on H (written A ⊂ B(H)) will act non-degenerately, so that the identity map idA : A → A is a morphism from A to K(H). More generally a representation of A on H is by definition an element of Mor(cid:0)A, K(H)(cid:1). The notion of a morphism of C∗-algebras generalizes that of a continuous map between locally compact Hausdorff spaces. We have the following well known result: Theorem 1.1. Let X and Y be locally compact Hausdorff spaces and let B = C0(X) and A = C0(Y ). Then (1) any continuous φ : X → Y defines a morphism Φ ∈ Mor(A, B) via Φ(f ) = f ◦φ, (f ∈ A); (1.1) (2) for any Φ ∈ Mor(A, B) there exists a continuous φ : X → Y such that (1.1) holds. Fixing Φ and φ so that (1.1) holds we moreover have (3) the range of Φ is contained in B = C0(X) if and only if φ is a proper map, (4) φ has dense image if and only if Φ is injective, (5) φ is injective if and only if Φ has strictly dense range. 4 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN The strict topology on a multiplier algebra mentioned in Theorem 1.1 is described e.g. in [49, Section 2] or [24, Chapter 1]. The proof of the above theorem is a simple exercise in elementary topology and we leave it to the reader (see e.g. [46, Exercises to Chapter 2]). Let A be a C∗-algebra. The dual space A∗ is naturally a module over A and we will denote the natural left action of a ∈ A on ϕ ∈ A∗ by a · ϕ, so that (a · ϕ)(b) = ϕ(ba) for all b ∈ A. Note that if C ⊂ B(H) then C acts in a natural way on the functionals in B(H)∗ = K(H)∗ and we have C · K(H)∗ = C K(H) · K(H)∗ = K(H) · K(H)∗ = K(H)∗. (1.2) (all sets above are automatically closed by the Cohen factorization theorem). For C∗-algebras A and B their minimal tensor product will be denoted by A ⊗ B. For von Neumann algebras M and N the von Neumann algebra tensor product of M and N will be denoted by M ¯⊗N. The tensor flip a ⊗ b 7→ b ⊗ a will be denoted by σ regardless of which C∗-algebras are being considered. We will also use the same symbol "⊗" to denote tensor product of Hilbert spaces. In [49] S.L. Woronowicz introduced a very important notion of a C∗-algebra generated by elements which do not necessarily belong to it. We will use a crucial part of his theory dealing with C∗-algebras "generated by a quantum family of multipliers". Let A and C be C∗-algebras and let T ∈ M(C ⊗ A). By analogy with the classical situation (when C is commutative) the element T is referred to as a quantum family of elements of M(A) labeled by the spectrum of C (cf. [49, Formula (2.5)]). Definition 1.2 ([49, Definition 4.1]). Let A and C be C∗-algebras. We say that A is generated by T ∈ M(C ⊗ A) if for any Hilbert space H, any representation ρ of A on H and any C∗-algebra B ⊂ B(H) the condition that (id ⊗ ρ)(T ) ∈ M(C ⊗ B) implies that ρ ∈ Mor(A, B). Examples of the situation described in Definition 1.2 are plentiful. For the simplest case consider a C∗-algebra A generated by a finite set of elements a1, . . . , an ∈ A (in the usual sense, i.e. the closure of the set of algebraic combinations of the elements a1, . . . , an and their adjoints coincides with A). Then A is generated by T ∈ M(Cn ⊗ A) with T = nXi=1 ei ⊗ ai, where {e1, . . . , en} is the standard basis of Cn. More complicated examples of C∗-algebras gener- ated by quantum families of multipliers are given in [49, Section 4]. In this paper we will be mostly interested in examples of this situation arising from representations of locally compact quantum groups to be studied in Subsection 1.2, Section 2 and Section 3. Remark 1.3. Let H and K be Hilbert spaces and consider C∗-algebras A1, A2 ⊂ B(H) and C ⊂ B(K). Suppose that T ∈ B(K ⊗ H) is such that T ∈ M(C ⊗ A1) ∩ M(C ⊗ A2) and T generates both A1 and A2. Then A1 = A2, as the identity representation of A1 is a morphism in Mor(A1, A2), and similarly the identity representation of A2 is a morphism in Mor(A2, A1). This argument appeared already in [49]. Usually it is difficult to check that a given T ∈ M(C⊗A) generates A. For the needs of this paper it will be very useful to apply the following criterion. Note that if T ∈ M(C ⊗ A) and C ⊂ B(H), then each functional ω ∈ B(H)∗ defines an element of C∗, so that, in particular, (ω ⊗id)(T ) ∈ M(A) ([24, Proposition 8.3], [25, Lemma A.3]). Lemma 1.4. Let A and C be C∗-algebras with C ⊂ B(H) for a Hilbert space H. Let T ∈ M(C ⊗ A) be unitary and define If S ⊂ A and S generates A (as a subset of the C∗-algebra A) then T ∈ M(C ⊗ A) generates A. S =(cid:8)(ω ⊗ id)(T ) ω ∈ B(H)∗(cid:9) ⊂ M(A). Proof. Let K be a Hilbert space and let ρ be a representation of A on K such that (id ⊗ ρ)(T ) ∈ M(C ⊗ B) for a certain C∗-algebra B ⊂ B(K). It is easily seen that ρ(S) ⊂ M(B), which implies CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 5 that ρ(A) ⊂ M(B) because S generates A. Furthermore (cid:0)ρ(S)B(cid:1) -- k·k =(cid:8)ρ(cid:0)(c · ω ⊗ id)(T )(cid:1)b c ∈ C, b ∈ B, ω ∈ B(H)∗(cid:9) -- k·k =(cid:8)(ω ⊗ id)(cid:0)(id ⊗ ρ)(T ) (c ⊗ b)(cid:1) c ∈ C, b ∈ B, ω ∈ B(H)∗(cid:9) -- k·k =(cid:8)(ω ⊗ id)(c ⊗ b) c ∈ C, b ∈ B, : ω ∈ B(H)∗(cid:9) -- k·k = B, where in the first equality we used the formula (1.2) and in the last one we used the fact that (id ⊗ ρ)(T ) is unitary in M(C ⊗ B). This shows that ρ ∈ Mor(A, B) and ends the proof. (cid:3) Remark 1.5. Sometimes it is important to use the notion of a C∗-algebra generated by a quantum family of multipliers in a different version. More precisely let A and C be C∗-algebras and let T ∈ M(A ⊗ C) (note the different order of tensor factors from the one in Definition 1.2). We will say that T ∈ M(A ⊗ C) generates A if σ(T ) ∈ M(C ⊗ A) generates A in the sense described in Definition 1.2. It can happen that a given T ∈ M(C ⊗ A) generates A and at the same time T ∈ M(C ⊗ A) generates C. Coming back to the analogy with classical situation we would say that in the first statement T is a quantum family of multipliers of A labeled by the spectrum of C and in the second statement T is a quantum family of multipliers of C labeled by the spectrum of A. Throughout the paper we will use the so-called leg-numbering notation. This is explained in a number of texts on quantum groups, e.g. [33, 2]. 1.2. Locally compact quantum groups and their universal versions. For the theory of locally compact quantum groups we refer the reader to [23] and to [25] for an equivalent approach with different initial axioms. Most results of this paper are true in a potentially more general setting of quantum groups defined by modular multiplicative unitaries ([39, 40, 27]), but we will stay within the theory of Kustermans and Vaes. For a locally compact quantum group G the corresponding C∗-algebra of "continuous functions on G vanishing at infinity" will be denoted by C0(G). This C∗-algebra is equipped with a comultiplication ∆G ∈ Mor(cid:0)C0(G), C0(G) ⊗ C0(G)(cid:1). There is also the reduced bicharacter WG ∈ M(cid:0)C0(bG) ⊗ C0(G)(cid:1) (see [17, Page 53]), where bG denotes the dual of G. The Haar weights provide a realization of both C0(G) and C0(bG) on the Hilbert space L2(G). Then WG ∈ B(cid:0)L2(G) ⊗ L2(G)(cid:1) is a multiplicative unitary ([2]) called the Kac-Takesaki operator of G ([25]). The comultiplication is then implemented by WG: ∆G(f ) = WG(f ⊗ 1)(WG)∗ for all f ∈ C0(G) (note that we are using the conventions of [2, 50, 25, 40, 27] favoring right Haar weights over left ones). The embedding of C0(G) into B(cid:0)L2(G)(cid:1) defines also the von Neumann algebra L∞(G) as C0(G)′′. Moreover we have C0(G) =(cid:8)(ω ⊗ id)(WG) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k. In fact C0(G) is generated by the quantum family WG ∈ M(cid:0)C0(bG)⊗ C0(G)(cid:1) in the sense described in Definition 1.2 ([50]). Moreover the C∗-algebra C0(bG) is generated by quantum family WG ∈ M(cid:0)C0(bG) ⊗ C0(G)(cid:1) (note the difference, cf. Remark 1.5). The dense subspace (1.3) (1.4) AG =(cid:8)(ω ⊗ id)(WG) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) ⊂ C0(G) (no closure) is called the Fourier algebra of G ([9, Section 8]). Note that the vector space AG is by the functionals which vanish on C0(G) with L∞(G)∗. It is clear that one can use this space of indeed a subalgebra of C0(G) ([2, Proposition 1.4]). We will identify the quotient of B(cid:0)L2(G)(cid:1)∗ functionals instead of B(cid:0)L2(G)(cid:1)∗ in all formulas of the form (1.4) or (1.3). M(cid:0)C0(bG)(cid:1) is non-zero. Lemma 1.6. Let G be a quantum group and let η ∈ C0(G)∗ be non-zero. Then (id ⊗ η)(WG) ∈ 6 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN Proof. If η 6= 0 then it must be non-zero on the norm dense set AG. Therefore there is a normal functional ω on B(cid:0)L2(G)(cid:1) such that η(cid:0)(ω ⊗ id)(WG)(cid:1) 6= 0. Consequently ω(cid:0)(id ⊗ η)(WG)(cid:1) 6= 0 which clearly implies that (id ⊗ η)(WG) 6= 0. (cid:3) of view of classical harmonic analysis ([12]). We will view the Fourier algebra both as a Banach space and a subspace of C0(G). The first consequence of Lemma 1.6 is that AG is isomorphic as a vector space to L∞(bG)∗; in particular it is a Banach space with the norm transported from L∞(bG)∗. Indeed this is the point 0(G)(cid:1). This object was introduced and analyzed a comultiplication ∆u in [22]. In the more general setting of quantum groups defined by modular multiplicative unitaries the universal C∗-algebra corresponding to G is studied in [40, Section 5]. The reduced bicharacter lifts to the universal level, i.e. we have the universal bicharacter The universal object related to G is a C∗-algebra which we will denote by Cu G ∈ Mor(cid:0)Cu 0(G), endowed with 0(G) ⊗ Cu 0(G), Cu ([22, Proposition 3.8] and [27, Proposition 4.8]). Following the conventions of [40] the reducing mor- 0(bG) ⊗ Cu phisms for G and bG will be denoted by ΛG ∈ Mor(cid:0)Cu respectively (see [40, Definition 35]). We have VVG ∈ M(cid:0)Cu 0(G)(cid:1) 0(G), C0(G)(cid:1) and Λ bG ∈ Mor(cid:0)Cu 0(bG), C0(bG)(cid:1) (Λ bG ⊗ ΛG)(VVG) = WG. The elements (id ⊗ ΛG)(VV G) and (Λ bG ⊗ id)(VV G) will be denoted by WG ∈ M(cid:0)Cu and W G ∈ M(cid:0)C0(bG) ⊗ Cu 0(bG) ⊗ C0(G)(cid:1) 0(G) =(cid:8)(ω ⊗ id)(W G) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k 0(G)(cid:1) (by Lemma 1.4, cf. [40] and Proposition 2.1). Cu 0(G)(cid:1) respectively. We have W G ∈ M(cid:0)C0(bG) ⊗ Cu The universal dual is determined by the quantum group G only up to isomorphism, so when ΛG 0(G) = C0(G) and ΛG = id. is an isomorphism (i.e. G is coamenable) then we can declare that Cu Then ([40, Formula (5.14)]) and consequently the C∗-algebra Cu 0(G) is generated by the quantum family (1.5) (1.6) (1.7) VVG = W G and W bG = W bG. Similarly, when bG is coamenable then VV bG = W bG and W G = WG. (1.8) Note that quantum groups which are classical (i.e. quantum groups G for which C0(G) is com- mutative) are always coamenable. Proposition 1.7. Let G be a locally compact quantum group. Then the reducing map ΛG is injective on the subspace The proof is obvious: (cid:16)(id ⊗ η)(W (cid:8)(id ⊗ η)(W bG) η ∈ B(cid:0)L2(bG)(cid:1)∗(cid:9). bG) 6= 0(cid:17) =⇒ (cid:0)η 6= 0(cid:1) =⇒ (cid:16)(id ⊗ η)(W (1.9) bG) 6= 0(cid:17) very useful [51, Proposition 3.2]. The image of ΛG on the subspace (1.9) is exactly the Fourier by Lemma 1.6 applied to bG. Note that Proposition 1.7 can be viewed as a generalization of the algebra AG. It follows that ΛG(cid:0)Cu 0(G)(cid:1) = C0(G). The Fourier-Stieltjes algebra of G is the space BG =(cid:8)(η ⊗ id)(VVG) η ∈ Cu 0(bG)∗(cid:9) ⊂ M(cid:0)Cu 0(G)(cid:1) (see [9, Section 8], note that in that paper BG was embedded into M(cid:0)C0(G)(cid:1) and not into 0(G)(cid:1)). A reasoning analogous to that in the proof of Lemma 1.6 shows that BG is isomorphic M(cid:0)Cu CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 7 as a vector space to Cu In what follows we shall utilize both pictures of AG and BG -- as Banach spaces of functionals 0(bG)∗. Indeed, as (cid:8)(cid:0)id ⊗ [ω ◦ ΛG](cid:1)(WG) ω ∈ L∞(G)∗(cid:9) is dense in Cu 0(bG) ([40, Section 5]), a non-zero η must be non-zero on some element of the form(cid:0)id ⊗ [ω◦ΛG](cid:1)(WG), so (ω ◦ΛG)(cid:0)(η ⊗ id)(VVG)(cid:1) 6= 0. In particular (η ⊗ id)(VVG) 6= 0. and at the same time as (non-closed) subspaces of C0(G) and M(cid:0)Cu instead of C0(G). Dually, G is discrete if bG is compact. In this case C0(G) is a c0-direct sum of matrix algebras and we write c0(G) instead of C0(G). We also write in this case ℓ∞(G) for L∞(G). Discrete quantum groups are always coamenable ([33]). We refer to [51] for the complete account of the theory of compact quantum groups and to [33, Section 3] for a thorough treatment of discrete quantum groups. A quantum group G is compact if the C∗-algebra C0(G) is unital. In this case we write C(G) 0(G)(cid:1) respectively. Finally let us mention that on the level of bicharacters the duality between G and bG is imple- mented by the tensor flip and the adjoint operation: VV bG = σ(VVG)∗. (1.10) It follows that W bG = σ(WG)∗ and W bG = σ(WG)∗. 1.3. Homomorphisms of locally compact quantum groups. Let G and H be locally compact quantum groups. In [27] it is shown that the following three classes of objects are in a one-to-one correspondence: (1) strong quantum homomorphisms: morphisms (2) bicharacters (from H to G): unitaries π ∈ Mor(cid:0)Cu 0(G), Cu 0(H)(cid:1) (π ⊗ π)◦∆u G = ∆u H ◦π; V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) such that such that such that (∆bG ⊗ idC0(H))(V ) = V23V13, (idC0( bG) ⊗ ∆H)(V ) = V12V13. (1.11) (3) right quantum homomorphisms: morphisms ρ ∈ Mor(cid:0)C0(G), C0(G) ⊗ C0(H)(cid:1) (∆G ⊗ id)◦ρ = (id ⊗ ρ)◦∆G, (id ⊗ ∆H)◦ρ = (ρ ⊗ id)◦ρ. All these should be thought of as alternative descriptions of a fixed homomorphism from H to G. Note that the reduced bicharacter WG of G introduced in Subsection 1.2 is a bicharacter from G to G in the above sense and describes the identity homomorphism. Sometimes, to simplify the language, we will refer to a strong quantum homomorphism π as above as a homomorphism from H to G. A strong quantum homomorphism π is related to the bicharacter V via the formula V =(cid:0)Λ bG ⊗ [ΛH◦π](cid:1)(VV G), (1.12) while the right quantum homomorphism ρ is given by for any x ∈ C0(G) ⊂ B(cid:0)L2(G)(cid:1). ρ(x) = V (x ⊗ 1C0(H))V ∗ 8 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN One can also check (see [27, Lemma 3.4]) that for a unitary V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) the conditions (1.11) are equivalent to the following "twisted" pentagonal equations: 12 = WG V23WG WH 12V13V23, 23V12 = V12V13WH 23, in M(cid:0)C0(bG) ⊗ K(cid:0)L2(G)(cid:1) ⊗ C0(H)(cid:1), in M(cid:0)C0(bG) ⊗ K(cid:0)L2(H)(cid:1) ⊗ C0(H)(cid:1). (1.13a) (1.13b) cπ2 ◦cπ1 = \π1 ◦π2, The next result, namely [27, Proposition 3.14], describes in the simplest way the construction of the dual homomorphisms (cf. (1.11)). Proposition 1.8. If V is a bicharacter from H to G, the unitary bV = σ(V ∗) ∈ M(cid:0)C0(H)⊗C0(bG)(cid:1) is a bicharacter from bG to bH. Proposition 1.8 makes possible the following definition: Definition 1.9. Let π be a morphism from H to G with corresponding bicharacter V . Then the strong quantum homomorphism defined by bV is called the dual of π and will be denoted by bπ, so that bπ ∈ Mor(cid:0)Cu Let us note the most fundamental equality relating π to bπ (and determining bπ uniquely) con- tained in [27, Theorem 4.15]: 0(bH), Cu 0(bG)(cid:1). By applying (Λ bG ⊗ ΛH) to both sides and using (1.12) we obtain (id ⊗ π)(VVG) = (bπ ⊗ id)(VVH). Moreover if π1 and π2 are strong quantum homomorphisms associated with homomorphisms from G1 to G2 and from G2 to G3 respectively then (cid:0)Λ bG ⊗ [ΛH ◦π](cid:1)(VVG) = V =(cid:0)[Λ bG ◦bπ] ⊗ ΛH(cid:1)(VVH). (1.14) (1.15) (1.16) since (1.14) characterizes the dual strong quantum homomorphism. Thus if π is an isomorphism of C∗-algebras then so is bπ. from H to G such that the corresponding π ∈ Mor(cid:0)Cu 0(G) onto Cu a one-to-one map from Cu C0(H) such that πr ◦ΛG = ΛH ◦π. Theorem 1.10. Let G and H be locally compact quantum groups. Consider a homomorphism 0(H). Then there exists an isomorphism πr of C0(G) onto 0(G), Cu 0(H)(cid:1) is an isomorphism, i.e. π is Theorem 1.10 says that isomorphisms in the category of locally compact quantum groups con- sidered in [27] drop down to C∗-algebraic isomorphisms of the reduced level. In what follows we will refer to this situation by simply saying that G and H are isomorphic. A proof of this result may be given along the lines of [22, Proposition 8.7] (cf. also [22, Proposition 7.1]). In Section 3 we will give a short proof of Theorem 1.10 using representation theory of locally compact quantum groups and techniques developed in this paper. Let us note that these techniques make no use of the existence of Haar weights and are equally applicable to quantum groups arising from modular multiplicative unitaries. 2. Representations of locally compact quantum groups In this section we recall some basic notions of the representation theory of locally compact quan- tum groups ([2], [40, Section 3]) and establish alternative characterizations of quasi-equivalence of two representations of a given quantum group (Theorem 2.2). Let G be a locally compact quantum group and let H be a Hilbert space. A strongly continuous unitary representation of G on H is a unitary element U ∈ M(cid:0)K(H) ⊗ C0(G)(cid:1) such that usually decorated by the subscript U , so that U ∈ M(cid:0)K(HU ) ⊗ C0(G)(cid:1). For such a representation U of G the subspace We will usually write simply of "representations of G". Moreover the Hilbert space H will be (id ⊗ ∆G)(U ) = U12U13. AU =(cid:8)(id ⊗ ω)(U ) ω ∈ L∞(G)∗(cid:9) -- k·k CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 9 is a non-degenerate C∗-subalgebra of B(H) (it was denoted "BU " in [40]). In fact U is a multiplier of AU ⊗ C0(G) and the quantum family U ∈ M(cid:0)AU ⊗ C0(G)(cid:1) generates AU ([50, 40]). We will also use at some point the notation AU =(cid:8)(id ⊗ ω)(U ) ω ∈ L∞(G)∗(cid:9). It is easy to check that AU is an algebra -- this is a quantum group analogue of the "Fourier space of a representation" defined in [1, Definition (2.1)]; note for example that AWG is the Fourier algebra AbG of bG. Observe further that a bicharacter from H to G is a representation of H on L2(G) and it follows from Proposition 1.8 that bV is a representation of bG on L2(H). The generating property for representations can be reformulated in terms of their slices. In the following proposition note the use of the notion of a C∗-algebra generated by a quantum family of multipliers in the version described in Remark 1.5. Proposition 2.1. Let U be a representation of G on a Hilbert space H and let A be a non- degenerate C∗-subalgebra of B(H). Assume that U ∈ M(cid:0)A ⊗ C0(G)(cid:1). Then the following are equivalent: (1) U ∈ M(cid:0)A ⊗ C0(G)(cid:1) generates A; (2) A = AU . Proof. A direct consequence of the fact that U generates AU and Remark 1.3 (cf. [40, Subsection 3.5]). (cid:3) The standard notions of representation theory were all collected in [40, Section 3]. • Two representations U and V of G are equivalent if there exists a unitary operator T ∈ B(HU , HV ) such that (T ⊗ 1)U = V (T ⊗ 1). • If H is a Hilbert space then the trivial representation of G on H is • The tensor product of two representations U and V is the representation IH = 1B(H) ⊗ 1C0(G) ∈ M(cid:0)K(H) ⊗ C0(G)(cid:1). U ⊤ V ∈ M(cid:0)K(HU ⊗ HV ) ⊗ C0(G)(cid:1) defined by • Representation U and V are quasi-equivalent if there exists a Hilbert space H such that U ⊤ V = U13V23. IH ⊤ U and IH ⊤ V are equivalent ([40, Proposition 13]). The following theorem will be crucial in the next section, when we analyze a definition of a closed quantum subgroup proposed in [43]. The implication (1)⇒(2) is [40, Corollary 15]. Theorem 2.2. Let U and V be representations of G on HU and HV respectively. The following three conditions are equivalent: (1) U is quasi-equivalent to V ; (2) there exists a (necessarily unique) normal ∗-isomorphism γ : A′′ U → A′′ V such that (γ ⊗ id)(U ) = V ; (3) we have (cid:8)(η ⊗ idB(L2(G)))(U ) η ∈ B(HU )∗(cid:9) =(cid:8)(µ ⊗ idB(L2(G)))(V ) µ ∈ B(HV )∗(cid:9). Proof. (1)⇒(2). Suppose that U and V are quasi-equivalent. Let K be a Hilbert space and let T : K ⊗ HU → K ⊗ HV be a unitary such that Take ω ∈ B(cid:0)L2(G)(cid:1)∗ and put x = (id ⊗ ω)(U ) and y = (id ⊗ ω)(V ). Equation (2.2) shows that T (1K ⊗ x)T ∗ = 1K ⊗ y. This implies that T (1K ⊗ A′′ in the converse direction we observe that actually T (1K ⊗ A′′ V . Applying a similar argument V so there exists a U )T ∗ = 1K ¯⊗A′′ U )T ∗ ⊂ 1K ¯⊗A′′ T12U23T ∗ 12 = V23. (2.1) (2.2) 10 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN normal ∗-isomorphism γ : A′′ equation (2.2) we see that (γ ⊗ id)(U ) = V . U → A′′ V such that 1K ⊗ γ(x) = T (1K ⊗ x)T ∗ for all x ∈ A′′ U . Using U → A′′ (2)⇒(1). Let γ : A′′ V be an normal ∗-isomorphism such that (γ ⊗ id)(U ) = V . It is a well known fact (see e.g. [7, Theorem III.2.2.8]) that γ is of the form 1K ⊗ γ(x) = T (1K ⊗ x)T ∗ for some Hilbert space K and a unitary operator T : K ⊗ HU → K ⊗ HV . It is then easy to check that T12U23T ∗ 12 = V23, which proves the quasi-equivalence of U and V . (2)⇒(3). Since for any µ ∈ B(HV )∗ the composition µ◦γ is a normal functional on A′′ V , there exists η ∈ B(HV )∗ such that µ◦γ = η(cid:12)(cid:12)A′′ V (cid:8)(η ⊗ idB(L2(G)))(U ) ω ∈ B(HU )∗(cid:9) ⊃(cid:8)(µ ⊗ idB(L2(G)))(V ) µ ∈ B(HV )∗(cid:9). Exchanging the roles of U and V we get the opposite inclusion; hence (3) follows. . This shows that (3)⇒(2). Let κ be the extension of the antipode κ of G to an unbounded operator acting on M(cid:0)C0(G)(cid:1) (see [50, Theorem 1.6]). Recall that κ is a densely defined operator acting on its domain D(κ) ⊂ M(cid:0)C0(G)(cid:1) such that for any representation U ∈ M(cid:0)K(HU ) ⊗ C0(G)(cid:1) of G and any η ∈ B(HU )∗ we have (η ⊗ id)(U ) ∈ D(κ) and (2.3) Consider the set X ⊂ L∞(G)∗ defined so that ω ∈ X if and only if ω∗ ◦ κ extends to a bounded normal functional on L∞(G). Define further AX the fact that U is a representation ensures that AX U = (cid:8)(id ⊗ ω)(U ) ω ∈ X(cid:9). Equation (2.3) and U is a weakly dense ∗-subalgebra of A′′ U . κ(cid:0)(η ⊗ id)(U )(cid:1) = (η ⊗ id)(U ∗). Let us define a map γ0 : AX U → A′′ V by the following formula: γ0(cid:0)(id ⊗ ω)(U )(cid:1) = (id ⊗ ω)(V ), ω ∈ X. Fix ω ∈ X. Since V ∈ M(cid:0)AV ⊗ C0(G)(cid:1), the expression (id ⊗ ω)(V ) makes sense. Moreover if (id ⊗ ω)(U ) = 0 then for any η ∈ B(HU )∗ we have ω(cid:0)(η ⊗ id)(U )(cid:1) = 0 and by our assumption ω(cid:0)(µ ⊗ id)(V )(cid:1) = 0 for any µ ∈ B(HV )∗. The last property means that (id ⊗ ω)(V ) = 0 and shows that γ0 is well-defined (cf. Lemma 1.6). It can be checked that γ0 is a ∗-homomorphism, for example γ0(cid:0)(id ⊗ ω)(U )(cid:1)∗ =(cid:0)(id ⊗ ω)(V )(cid:1)∗ =(cid:0)id ⊗ [ω∗◦ κ](cid:1)(V ) = γ0(cid:0)(cid:0)(id ⊗ ω)(U )(cid:1)∗(cid:1). to a certain element y ∈ A′′ In the next step we shall show that γ0 may be extended to a normal ∗-isomorphism γ : A′′ U → A′′ V . Take x ∈ AU . Using Kaplansky's density theorem, we may find a bounded net (xi)i∈I of elements in AX xi = x. Let M ∈ R+ be the U , say xi = (id ⊗ ωi)(U ) with ωi ∈ X such that w-lim i∈I corresponding bound: kxik ≤ M . In what follows we shall prove that(cid:0)γ0(xi)(cid:1)i∈I weakly converges V )∗. For each i ∈ I we have µ(cid:0)γ0(xi)(cid:1) = ωi(cid:0)(µ ⊗ id)(V )(cid:1). For η ∈ (A′′ U )∗ such that (µ ⊗ id)(V ) = (η ⊗ id)(U ) we obtain µ(cid:0)γ0(xi)(cid:1) = η(xi). In particular(cid:12)(cid:12)µ(cid:0)γ0(xi)(cid:1)(cid:12)(cid:12) ≤ M kηk and lim we conclude that the family (cid:0)γ0(xi)(cid:1)i∈I of functionals on (A′′ Banach-Steinhaus theorem it is norm bounded. Let N ∈ R+ be a bound: (cid:13)(cid:13)γ0(xi)(cid:13)(cid:13) ≤ N for all µ(cid:0)γ0(xi)(cid:1) ∈ C is a bounded functional with the i ∈ I. Noting that the map (A′′ norm not greater than N we conclude the existence of y ∈ A′′ V )∗ V )∗ is pointwise bounded. By the µ(cid:0)γ0(xi)(cid:1) = η(x). Interpreting A′′ V . Take now µ ∈ (A′′ V as the dual of (A′′ V )∗ ∋ µ 7→ lim i∈I i∈I V , such that y = w-lim i∈I 0 then for each γ0(xi). This The equality γ(x∗) = γ(x)∗ for any x ∈ AX continuous imply that γ(x∗) = γ(x)∗ for any x ∈ A′′ that for any x, x′ ∈ A′′ U and the fact that the star operation is weakly U . We will now show using once again (2.1) U we have γ(xx′) = γ(x)γ(x′). Note that, in the notation of the previous U )∗. Passing to the limit paragraph, for any i ∈ I we have µ(cid:0)γ0(xi)(cid:1) = η(xi) for a certain η ∈ (A′′ we get µ(cid:0)γ(x)(cid:1) = η(x), for any x ∈ A′′ i)(cid:1) =(cid:0)µ · γ(x)(cid:1)(cid:0)γ(x′ i)(cid:1) i) = µ(cid:0)γ(xx′ U . Note also that i) = η(xx′ (η · x)(x′ enables us to define the aforementioned extension by putting γ(x) = y. If xi −−→ i∈I η(xi) = 0, so that y = 0. This implies that γ is well defined. µ as above lim i∈I µ(cid:0)γ0(xi)(cid:1) = lim i∈I CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 11 for any x, x′ i ∈ AX U . Again, passing to the limit, we get (η · x)(x′) =(cid:0)µ · γ(x)(cid:1)(cid:0)γ(x′)(cid:1) for any x ∈ AX elements of AX Finally we compute: U and x′ ∈ A′′ U . Replacing x ∈ AX U and passing to the limit yields (η · x)(x′) = (cid:0)µ · γ(x)(cid:1)(cid:0)γ(x′)(cid:1) for any x, x′ ∈ A′′ µ(cid:0)γ(xx′)(cid:1) = η(xx′) = (η · x)(x′) =(cid:0)µ · γ(x)(cid:1)(cid:0)γ(x′)(cid:1) = µ(cid:0)γ(x)γ(x′)(cid:1) U with a bounded, weakly convergent net (xi) of U . which shows that γ(xx′) = γ(x)γ(x′) for any x, x′ ∈ A′′ U . Exchanging the roles of U and V leads to the inverse ∗-homomorphism γ−1 : A′′ U . This shows that γ is normal, since isomorphisms of von Neumann algebras are automatically normal ([41, Corollary 3.10, page 135]). (cid:3) V → A′′ It was shown in [40] that a unitary representation U of G is quasi-equivalent to WG if it is right absorbing, i.e. for any other representation V of G the tensor product V ⊤ U is equivalent to ⊤ U (this can be viewed as a version of the Fell absorption principle). We finish the section IHV with a proposition which describes relation between quasi-equivalence of a given representation U of G with WG and the fact that U ∈ M(cid:0)K(HU ) ⊗ C0(G)(cid:1) generates C0(G). element U ∈ M(cid:0)AU ⊗ C0(G)(cid:1) generates C0(G). On the other hand a representation U which Proposition 2.3. Let U be a representation of G quasi-equivalent to WG. Then the unitary generates C0(G) need not be quasi-equivalent to WG (even when G is a locally compact group). Proof. From Theorem 2.2(3) it follows that (cid:8)(ω ⊗ idB(L2(G)))(U ) ω ∈ B(HU )∗(cid:9) =(cid:8)(µ ⊗ idB(L2(G)))(WG) µ ∈ B(cid:0)L2(G)(cid:1)∗(cid:9). For the second part it suffices to observe the following fact: let U1 and U2 be representations of a locally compact quantum group G and let U be their direct sum ([40, Subsection 3.3.1]). If Since for Y = (cid:8)(µ ⊗ idB(L2(G)))(WG) µ ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) we have Y = C0(G), we see that Y generates C0(G) as a C∗-algebra. Lemma 1.4 implies that U ∈ M(cid:0)AU ⊗ C0(G)(cid:1) generates C0(G). U1 ∈ M(cid:0)K(HU1 ) ⊗ C0(G)(cid:1) generates C0(G) then so does U ∈ M(cid:0)K(HU1 ⊕ HU2 ) ⊗ C0(G)(cid:1). U1 ⊕ U2 ∈ M(cid:0)K(ℓ2(Z) ⊕ C) ⊗ c0(Z)(cid:1) generates c0(Z). It cannot be quasi-equivalent to WZ, as Let G = Z and let U1 and U2 be the regular and trivial representation of G. Then U = then, according to Theorem 2.2, we would have a (normal) ∗-isomorphism between von Neumann algebras A′′ WZ = L∞(T). However, the latter algebra is non-atomic, so we would have a contradiction. U = L∞(T) ⊕ C and A′′ (cid:3) 3. Closed quantum subgroups of locally compact quantum groups This section is central to our paper. We begin by introducing two possible definitions of a closed quantum subgroup of a given quantum group, the first of which appears in [43] and the second was suggested to us by S.L. Woronowicz. Then we provide alternative, simplified descriptions for both of them (Theorems 3.4 and 3.6) and analyze their mutual relations (Theorems 3.5 and 3.7). We also present here a proof of Theorem 1.10. The aforementioned definitions are as follows: Definition 3.1 ([43, Definition 2.5]). Let G, H be locally compact quantum groups. Then H is said to be a closed quantum subgroup of G in the sense of Vaes if there exists a morphism π from H to G and a normal injective ∗-homomorphism γ : L∞(bH) → L∞(bG) such that (3.1) Definition 3.2 (Woronowicz). Let G, H be locally compact quantum groups. Then H is said to be a closed quantum subgroup of G in the sense of Woronowicz if there exists a morphism π from H to G such that the associated bicharacter V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) generates C0(H). γ(cid:12)(cid:12)C0( bH)◦Λ bH = Λ bG ◦bπ. 12 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN The conditions above take as a starting point a morphism π from H to G. We will sometimes say that H is a closed quantum subgroup of G in the sense of Vaes (respectively, in the sense of Woronowicz) via the morphism π. In Section 4 we will explain why when both H and G are locally compact groups both definitions are equivalent to the classical notion of H being (homeomorphic to) a closed subgroup of G. We will see later that the various examples of quantum subgroups considered in the literature are all closed quantum subgroups in the sense of both Vaes and Woronowicz. The case of compact and discrete subgroups is treated in Section 6. The non-compact examples of quantum subgroups in [38, Sections 3 and 4] and those coming from Rieffel deformation presented in [20] are all closed subgroups in the sense of Vaes and Woronowicz ([20, Section 6], cf. also Theorem 3.5). Another class of examples is provided by the bicrossed product construction (see e.g. [44, 45]). If (G1, G2) is a matched pair of locally compact quantum groups in the sense of [45, Definition 2.1] then bG1 is a closed quantum subgroup of the bicrossed product of G1 and G2 both in the sense of Vaes and Woronowicz. In the next theorem we note that the definition of Vaes can be reformulated in various simplified ways (note especially condition (2), which does not assume a priori the existence of a homomor- phism between H and G). In particular the definition of Vaes-Vainerman ([44, Definition 2.9]) is equivalent to Definition 3.1. Theorem 3.3. Let G, H be locally compact quantum groups. Then the following conditions are equivalent: (1) H is a closed quantum subgroup of G in the sense of Vaes; (γ ⊗ γ)◦∆bH = ∆bG ◦γ; (2) there exists a normal injective ∗-homomorphism γ : L∞(bH) → L∞(bG) such that (3) there exists a normal injective ∗-homomorphism γ : L∞(bH) → L∞(bG) such that the unitary (γ ⊗ id)(WH) ∈ L∞(bG) ¯⊗L∞(H) is a bicharacter from H to G -- in particular it belongs to M(cid:0)C0(bG) ⊗ C0(H)(cid:1). It will become clear from the proof of Theorem 3.3 that the map γ mentioned in point (2) is the same as the one in (3) and still the same as the map γ from Definition 3.1. Moreover we show in the proof that γ restricted to C0(bH) is an element of Mor(cid:0)C0(bH), C0(bG)(cid:1). (2)⇒(3). The map γ′ = γ(cid:12)(cid:12)C0( bH) is naturally a representation of the C∗-algebra C0(bH) on L2(G). Consider the unitary V = (γ′ ⊗ id)(WH) ∈ M(cid:0)K(cid:0)L2(G)(cid:1) ⊗ C0(H)(cid:1). Applying γ ⊗ γ ⊗ id to both Proof of Theorem 3.3. (1)⇒(2) -- trivial. sides of the equality (3.2) (∆bH ⊗ id)(WH) = WH 23WH 13 see that (3.3) 12V ∗ 23. V13 = (WG 12)∗V23WG precisely (1.13a). The application of γ ⊗ id ⊗ id to the pentagonal equation for WH implies that (1.13b) holds and V is a bicharacter. (viewed as an equality of operators in L∞(bH) ¯⊗ L∞(bH) ¯⊗ L∞(H)) and using the equation (3.2) we The right side of the above expression belongs to M(cid:0)C0(bG) ⊗ K(cid:0)L2(G)(cid:1) ⊗ C0(H)(cid:1) and V13 has legs only in the first and third tensor factor. Thus V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1). Note that (3.3) is (3)⇒(1). Note first that as V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1), it follows that because WH ∈ M(cid:0)C0(bH) ⊗ C0(H)(cid:1) generates C0(bH). Let π ∈ Mor(cid:0)Cu acter V and let bπ ∈ Mor(cid:0)Cu 0(H)(cid:1) be the strong quantum homomorphism associated with the bichar- 0(bH), C0(bG)(cid:1) be the dual quantum homomorphism. Then on one hand γ′ = γ(cid:12)(cid:12)C0( bH) ∈ Mor(cid:0)C0(bH), C0(bG)(cid:1) 0(G), Cu CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 13 we have (recall the dependencies between V , bV and bπ listed in Subsection 1.3) (id ⊗ γ′)(W and on the other hand bH) = bV bV =(cid:0)ΛH ⊗ [Λ bG◦bπ](cid:1)(VV W bH = (ΛH ⊗ Λ bH)(VV bH) bH). (this is (1.15) combined with (1.10)). Comparing the above and using the fact that This can be rewritten as (i.e. (1.5) for the quantum group bH) we obtain (cid:0)ΛH ⊗ [γ′◦Λ bH](cid:1)(VV (cid:0)id ⊗ [γ′ ◦Λ bH](cid:1)(W (γ′◦Λ bH)(cid:0)(ω ⊗ id)(W and upon application of (ω ⊗ id) with ω ∈ B(cid:0)L2(H)(cid:1)∗ yields bH) =(cid:0)ΛH ⊗ [Λ bG ◦bπ](cid:1)(VV bH) =(cid:0)id ⊗ [Λ bG ◦bπ](cid:1)(W bH)(cid:1) = (Λ bG ◦bπ)(cid:0)(ω ⊗ id)(W for any such ω. By (1.7) this implies that (3.1) holds. This ends the proof. bH)(cid:1) bH). bH). (cid:3) Theorem 3.3 and a straightforward application of Theorem 2.2 yields the following result. Theorem 3.4. Let G, H be locally compact quantum groups and suppose that V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) is a bicharacter describing a morphism π from H to G. Then the following conditions are equivalent: (1) H is a closed quantum subgroup of G in the sense of Vaes via the morphism π; (2) the bicharacter V is quasi-equivalent to WH (as a representation of H); (3) A bV = AW bH . Proof. The equivalence of (2) and (3) follows immediately from Theorem 2.2, as the sets appearing in (3) here coincide with the analogous sets in condition (3) of that theorem (recall that bV = σ(V ∗), bH = σ(WH)∗). The equivalence of (1) and (2) follows again from Theorem 2.2 and (the proof (cid:3) W of) Theorem 3.3. It now follows from Theorem 3.4 and Proposition 2.3 that there is a natural relation between Definitions 3.1 and 3.2. Theorem 3.5. If H is a closed quantum subgroup of G in the sense of Vaes, it is also a closed quantum subgroup of G in the sense of Woronowicz. Proof. Immediate consequence of Theorem 3.4 and Proposition 2.3. (cid:3) It is not clear if Definitions 3.1 and 3.2 are equivalent; in other words, whether Theorem 3.5 admits the converse. This would follow if we could show that a bi-character U ∈ M(cid:0)C0(bG)⊗C0(H)(cid:1) describing a homomorphism from H to G which generates C0(H) must be quasi-equivalent to WH (the example in Proposition 2.3 showed it need not be the case if we only assume that U is a representation of H). In the following sections we will show that in fact the equivalence of Definitions 3.1 and 3.2 holds in many natural cases. Now we show that Definition 3.2 also admits several natural equivalent reformulations. We collect them in the next theorem. 0(G), Cu Theorem 3.6. Let G, H be locally compact quantum groups and consider a homomorphism from H to G described by a bicharacter V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1), a strong quantum homomorphism π ∈ Mor(cid:0)Cu 0(H)(cid:1) and a right quantum homomorphism ρ ∈ Mor(cid:0)C0(G), C0(G) ⊗ C0(H)(cid:1). Then the following conditions are equivalent (recall that bV := σ(V )∗ is a representation of bG): (1) V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) generates C0(H) (in other words H is a closed quantum subgroup of G in the sense of Woronowicz); 14 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN (3.4) (2) A bV = C0(H); (3) the right quantum homomorphism ρ is strongly non-degenerate: ρ(cid:0)C0(G)(cid:1)(cid:0)C0(G) ⊗ 1C0(H)(cid:1) = C0(G) ⊗ C0(H) 0(H); (2)⇔(3). We compute: 0(G)(cid:1) = Cu (in particular the left hand side of (3.4) is contained in the right hand side); 0(G)(cid:1) = C0(H). (4) π(cid:0)Cu (5) (ΛH ◦π)(cid:0)Cu Proof. (1)⇔(2). This follows from Proposition 2.1 and an obvious fact that V ∈ M(cid:0)C0(bG)⊗C0(H)(cid:1) generates C0(H) if and only if bV ∈ M(cid:0)C0(H) ⊗ C0(bG)(cid:1) generates C0(H) (cf. Remark 1.5). ρ(cid:0)C0(G)(cid:1)(cid:0)C0(G) ⊗ 1(cid:1) =(cid:8)V(cid:0)(x · ω ⊗ id)(WG)∗ ⊗ 1(cid:1)V ∗(y ⊗ 1) x ∈ C0(bG), y ∈ C0(G), ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k 23(x ⊗ y ⊗ 1)(cid:1) x ∈ C0(bG), y ∈ C0(G), ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k =(cid:8)(ω ⊗ id ⊗ id)(cid:0)V23(WG =(cid:8)(ω ⊗ id ⊗ id)(cid:0)V ∗ 12)∗(x ⊗ y ⊗ 1)(cid:1) x ∈ C0(bG), y ∈ C0(G), ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k =(cid:8)(ω ⊗ id ⊗ id)(cid:0)V ∗ = C0(G) ⊗(cid:8)(ω ⊗ id)(V ) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k = C0(G) ⊗ A bV 13(x ⊗ y ⊗ 1)(cid:1) x ∈ C0(bG), y ∈ C0(G), ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k In the third equality we used the bicharacter property of V (Eq. (1.13a)) and in the fourth equality we used the fact that WG ∈ M(cid:0)C0(bG) ⊗ C0(G)(cid:1) is unitary. The above computation shows that A bV = C0(H) if and only if ρ(cid:0)C0(G)(cid:1)(cid:0)C0(G) ⊗ 1(cid:1) = C0(G) ⊗ C0(H). (2)⇔(5). Taking into account (1.7), (1.6) and (1.12) we find that V =(cid:0)id ⊗ [ΛH◦π](cid:1)(WG) and (ΛH ◦π)(cid:0)(ω ⊗ id)(WG)(cid:1) = (ω ⊗ id)(V ). 12)∗V ∗ 13(WG (3.5) Thus (ΛH ◦π)(cid:0)Cu 0(G)(cid:1) = (ΛH ◦π)(cid:0)(cid:8)(ω ⊗ id)(WG) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k(cid:1) =(cid:8)(ΛH ◦π)(cid:0)(ω ⊗ id)(WG)(cid:1) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k = A bV . (4)⇔(5). Since (4)⇒(5) is clear, it remains to show the converse implication. Consider the universal lift of V u ∈ M(cid:0)C0(bG) ⊗ Cu 0(H)(cid:1) defined as V u = (id ⊗ π)(W G), cf. [27, Section 4]. To show the desired implication it suffices to establish the following equality: (cf. the proof of (2)⇔(5)). Noting that1 (cid:8)(ω ⊗ id)(V u) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k = Cu 0(H) 13 = V ∗ V u 12 W H 23V12(W H 23)∗ (3.6) we compute: (cid:8)(ω ⊗ id)(V u) ω ∈ B(cid:0)L2(G)(cid:1)∗(cid:9) -- k·k =(cid:8)(ω ⊗ µ ⊗ id)(cid:0)V ∗ =(cid:8)(ω ⊗ µ ⊗ id)(cid:0)WH =(cid:8)(η ⊗ µ ⊗ id)(cid:0)W H =(cid:8)(η ⊗ µ ⊗ id)(WH =(cid:8)(η · x ⊗ id)(W H 12 W H 23V12(W H 12 23V12(WH 23WH 12(W H W H 23)∗(cid:1) ω ∈ B(cid:0)L2(G)(cid:1)∗, µ ∈ B(cid:0)L2(H)(cid:1)∗(cid:9) -- k·k 23)∗(cid:1) ω ∈ B(cid:0)L2(G)(cid:1)∗, µ ∈ B(cid:0)L2(H)(cid:1)∗(cid:9) -- k·k 23)∗(cid:1) η ∈ B(cid:0)L2(H)(cid:1)∗, µ ∈ B(cid:0)L2(H)(cid:1)∗(cid:9) -- k·k 13) η ∈ B(cid:0)L2(H)(cid:1)∗, µ ∈ B(cid:0)L2(H)(cid:1)∗(cid:9) -- k·k 13) η ∈ B(cid:0)L2(H)(cid:1)∗, x ∈ C0(bH)(cid:9) -- k·k = Cu 0(H). 1To prove (3.6) we first note that we have (∆bH 0(H)(cid:1) described in [27, Proposition 4.14]. This can be rewritten as WH in M(cid:0)C0(bH) ⊗ Cu Slicing with ω ∈ B(cid:0)L2(H)(cid:1)∗ on the left leg we obtain the formula (ΛH ⊗ id)∆u x ∈ Cu H)(V u) = V u reads W H 0(H). Now we apply id ⊗ ΛH ⊗ id to both sides of (id ⊗ ∆u 13 which is (3.6). 23)∗ = V12 V u V12(W H V u 23 as WH is the unique lift of WH to a bicharacter 23)∗. H(x) = WH(cid:0)ΛH(x) ⊗ 1(cid:1)(WH)∗ for all 13. By the previous formula this 13 = W H 12(W H 23WH W H ⊗id)WH = WH 12 WH 13 12 23 CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 15 W H W H 12(W H 23)∗ = WH 12 13 (see derivation of formula (3.6)). The second equality follows from the unitarity of V , in the third one we used the fact that V ∈ M(cid:0)C0(bG) ⊗ C0(H)(cid:1) generates C0(H) and in the fourth equality we used the equation 23WH Condition (3) in Theorem 3.6 classically corresponds to properness and freeness of the natural action of H on G induced by the homomorphism from H to G, see Section 4. It was introduced in the context of quantum groups by Podle´s in his thesis [31, Definicja 2.2], see also [32] and [38, Proposition 2.3] for a complete discussion. Condition (4) is a natural reflection of a general principle that injectivity on the level of point transformations is equivalent to surjectivity on the level of induced transformations on algebras of functions (but cf. Theorem 1.1(5)). One can ask the following question: are the conditions (1)-(3) above equivalent in general to the surjectivity of the extension of π to multiplier algebras ¯π : M(cid:0)Cu 0(G)(cid:1) → M(cid:0)Cu 0(H)(cid:1)? The Pedersen-Tietze theorem ([46, Theorem 2.3.9]) implies that ¯π is surjective if π is, provided that Cu Cu 0(H)) is σ-unital. With the results obtained in this section in hand we are ready to give a proof of Theorem 1.10. 0(G) (and hence (cid:3) 0(G), Cu Proof of Theorem 1.10. We are assuming that Mor(cid:0)Cu bπ ∈ Mor(cid:0)Cu with a Vaes-closed subgroup of bH. Let bV be the bicharacter associated to bπ: 0(H)(cid:1) is an isomorphism. Consider the dual strong quantum homomorphism 0(bG)(cid:1). By (1.16) bπ is an isomorphism. We will show that bπ identifies bG 0(bH), Cu the strong quantum homomorphism π ∈ bH). bV =(cid:0)ΛH ⊗ [Λ bG◦bπ](cid:1)(VV As for all x ∈ Cu W bH(cid:0)Λ bH(x) ⊗ 1(cid:1)(W 0(bH) (see Footnote 1), it follows that bV(cid:0)Λ bH(x) ⊗ 1(cid:1)(bV )∗ =(cid:0)Λ bH ⊗ [Λ bG ◦bπ](cid:1)∆u bH)∗ = (Λ bH ⊗ id)∆u bH(x) bH(x). so (id ⊗ ∆u bH). Then bH) = Y . Set X = (ϕ ⊗bπ−1)(VV bH)(X) = X12X13 because bπ−1 intertwines the coproducts. Further Let Y ∈ M(cid:0)K(HY ) ⊗ C0(H)(cid:1) be a representation of H. By the results of [22, 27], there exists a unique ϕ ∈ Mor(cid:0)Cu 0(H), K(H)(cid:1) such that (ϕ ⊗ Λ bH)(VV (cid:0)id ⊗ Λ bH ⊗ [Λ bG ◦π](cid:1)(X12X13) = bV23(cid:0)(id ⊗ Λ bH)(X)(cid:1)12(bV23)∗, 12bV23(cid:0)(id ⊗ Λ bH)(X)(cid:1)12 (cid:0)id ⊗ [Λ bG◦π](cid:1)(X) ⊤bV =(cid:0)(id ⊗ Λ bH)(X)(cid:1)∗ ⊤bV . However(cid:0)id⊗[Λ bG◦π](cid:1)(X) = Y , so which means that(cid:0)id⊗[Λ bG◦π](cid:1)(X) ⊤bV is equivalent to IHY that, as Y was arbitrary, bV is right-absorbing. It follows from [40] (see remark before Proposition 2.3) that bV is quasi-equivalent to W bH. By Theorem 3.4 bG is a closed quantum subgroup of bH in By Theorem 3.3 and the comment after it, there exists a morphism bγ1 ∈ Mor(cid:0)C0(G), C0(H)(cid:1) bγ1◦ΛG = ΛH ◦π. Applying identical reasoning to bπ−1 we obtain the existence ofbγ2 ∈ Mor(cid:0)C0(H), C0(G)(cid:1) such that bγ2◦ΛH = ΛG ◦π−1 (note that we use once again the fact that bπ−1 = dπ−1). Since ΛG and ΛH are surjections we see that bγ1 and bγ2 are mutually inverse and we can set πr =bγ1. We recall from Subsection 1.2 that the Fourier algebra and the Fourier-Stieltjes algebra of a the sense of Vaes. such that (cid:3) locally compact quantum group are the Banach spaces AG = L∞(bG)∗ and BG = Cu 0(bG)∗ 16 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN which we embedded into C0(G) and M(cid:0)Cu 0(G)(cid:1) respectively with the maps AG ∋ ω 7−→ (ω ⊗ id)(WG), G). BG ∋ η 7−→ (η ⊗ id)(VV We also note that AG embeds in BG via Λ∗ bG ◦ı∗, where ı is the embedding C0(G) ֒→ L∞(G) (we will also use the symbol "ı" to denote the analogous embedding for other quantum groups). It is easy to check that this embedding is isometric. Moreover the induced embedding of AG into M(cid:0)Cu 0(G)(cid:1) actually embeds the Fourier algebra into Cu 0(G). Theorem 3.7. Let H be a closed subgroup of G in the sense of Woronowicz via the morphism π : Cu 0(G) → Cu (1) π restricts to a map T : AG → AH which has dense range (for the AH norm); 0(H). Then the following are equivalent: (2) bπ∗ : BG → BH restricts to a map S : AG → AH which has dense range; (3) H is a closed subgroup of G in the sense of Vaes. Moreover we can replace "dense range" by "surjection" in (1) and (2). If these conditions hold, then S and T are the same map, which is nothing but the pre-adjoint of the implicit map bG(cid:0)ı∗(ω)(cid:1) ∈ Cu γ : L∞(bH) → L∞(bG) appearing in (3). Proof. Let ω ∈ L∞(bG)∗ = AG, set µ = Λ∗ 0(G)(cid:1), so a = (µ ⊗ id)(VVG). Then BG ⊂ M(cid:0)Cu so that π(a) is (the image of) bπ∗(µ) in BH. It is now clear that (1) and (2) are equivalent. isometries, S must be bounded. Set γ = S∗ : L∞(bH) → L∞(bG), so as S has dense range, γ is π(a) = (µ ⊗ π)(VVG) =(cid:0)bπ∗(µ) ⊗ id(cid:1)(VVH), 0(bG)∗, and let a be the image of ω in If (2) holds then the map S satisfies Λ∗ injective. Then we have that bG ◦ ı∗, and as Λ∗ bG ◦ ı∗ and Λ∗ bH ◦ ı∗ are bH ◦ ı∗ ◦ S = bπ∗ ◦ Λ∗ bH (cid:12)(cid:12)Cu 0 ( bH) = ı∗∗ ◦Λ∗∗ bG ◦bπ∗∗(cid:12)(cid:12)Cu 0 ( bH) = ı◦Λ bG◦bπ. γ ◦Λ bH = S∗ ◦Λ bH = S∗◦ı∗∗◦Λ∗∗ As γ is weak∗-continuous, it now follows that γ is a ∗-homomorphism, and so (3) holds. Finally, if (3) holds, then we have a normal injective ∗-homomorphism γ : L∞(bH) → L∞(bG) with γ ◦ı◦Λ bH = ı◦Λ bG◦bπ. Thus, for ω ∈ L∞(bG)∗, we have that and so bπ∗ restricts to a map S : AG → AH. As γ is injective and hence an isometry, γ∗ is a surjection and so S, which agrees with γ∗ once appropriate identifications are made, is also a surjection. This shows (2), and also demonstrates the claim about replacing "dense range" by "surjection". (cid:3) bG ◦ı∗)(ω)(cid:1) = (Λ∗ bπ∗(cid:0)(Λ∗ bH ◦ı∗◦γ∗)(ω), Remark 3.8. From Theorem 3.7 we immediately see that H is a closed subgroup of G in the 0(H), if and only if π restricts to a surjection sense of Vaes, via the morphism π : Cu AG → AH. In the classical case, the Herz restriction theorem ([15, 1]) says exactly that if H is a closed subgroup of G, then the restriction map (which is nothing but π : C0(G) → C0(H)) gives a surjection AG → AH . In other words the definition of a Vaes-closed subgroup is tailored exactly so that the quantum version of the Herz restriction theorem holds. 0(G) → Cu 4. Commutative case Let now G and H be locally compact groups, so in particular Cu 0(H) = C0(H). Any homomorphism from H to G (in the sense of quantum groups -- as defined in Subsection 1.3) is then described by a π ∈ Mor(cid:0)C0(G), C0(H)(cid:1). Moreover π is necessarily of the form π(f ) = f ◦θ, where θ : H → G is a continuous homomorphism (cf. Theorem 1.14). 0(G) = C0(G) and Cu Given a situation as above, consider the natural right action of H on the topological space G given by G × H ∋ (g, h) 7−→ g · h = g θ(h) ∈ G. (4.1) CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 17 Let us also introduce the so called canonical map γ : G × H → G × G for this action γ(g, h) = (g, g · h) =(cid:0)g, g θ(h)(cid:1). (4.2) ([37]). Let ρ ∈ Mor(cid:0)C0(G), C0(G) ⊗ C0(H)(cid:1) and Γ ∈ Mor(cid:0)C0(G) ⊗ C0(G), C0(G) ⊗ C0(H)(cid:1) be the morphisms of C∗-algebras corresponding to (4.1) and (4.2): f ∈ C0(G), ρ(f )(g, h) = f (g · h), Γ(F )(g, h) = F(cid:0)γ(g, h)(cid:1), F ∈ C0(G) ⊗ C0(G), g ∈ G, h ∈ H, g ∈ G, h ∈ H. Lemma 4.1. Let θ : H → G be a continuous homomorphism with corresponding action of H on G as in (4.1) and canonical map γ : G × H → G × G. Then the following are equivalent: (1) θ is a homeomorphism onto its closed image; (2) the action of H on G is free and proper i.e. γ is injective and proper; (3) (cid:0)ρ(cid:0)C0(G)(cid:1)(cid:0)C0(G) ⊗ 1(cid:1) = C0(G) ⊗ C0(H). Proof. (1)⇒(2). If θ is a homeomorphism, then γ is a homeomorphism onto its range, as for (g, g′) in the range of γ (i.e. g′ = g · h for some h ∈ H) we have γ−1(g, g′) =(cid:0)g, θ−1(g−1g′)(cid:1). Hence γ is in particular injective and proper. (2)⇒(1). Assume that γ is injective and proper. Clearly θ is then injective. Similarly, if θ were not proper, then there would be a compact set K ⊂ G with θ−1(K) non-compact. But then γ−1(cid:0){e} × K(cid:1) = {e} × θ−1(K) would not be compact either. Hence θ is injective and proper. Proper continuous maps between locally compact spaces are automatically closed ([8, Chapter 1, §10]). Hence θ has a closed image, and as a bijective continuous closed map is in fact a homeomorphism. (2)⇔(3). Note first that (cid:0)ρ(cid:0)C0(G)(cid:1)(cid:0)C0(G) ⊗ 1(cid:1) = Γ(cid:0)C0(G) ⊗ C0(G)(cid:1). Hence (3) is equivalent to the fact that γ is injective and proper by Theorem 1.1. (cid:3) Theorem 4.2. Suppose that G and H are locally compact groups. Then the following conditions are equivalent: (1) H is a closed quantum subgroup of G in the sense of Vaes; (2) H is a closed quantum subgroup of G in the sense of Woronowicz; (3) H is homeomorphic to a closed subgroup of G. Proof. Condition (1) implies (2) by Theorem 3.5. Conditions (2) and (3) are equivalent by Lemma 4.1 and Theorem 3.6. It remains to note that (3) implies (1). By theorem 3.3 this is precisely [52, Corollary 4.2.6] (which is a consequence of [1, Theorem (3.23)]). (cid:3) Remark 4.3. (1) Let G be a locally compact group and let H be a locally compact quantum group. If H is a closed subgroup of G in the sense of Woronowicz then by Theorem 3.6 there is a surjection from C0(G) onto C0(H), so that H is in fact a classical group. By Theorem 4.2, H is then also a closed subgroup of G in the usual sense. (2) Let H be a locally compact group and let G be a locally compact quantum group. If H is a closed subgroup of G in the sense of Woronowicz then the associated morphism 0(G) → C0(H) factors through the algebra C0( G), where G is the intrinsic group of π : Cu G as defined by Kalantar and Neufang (a locally compact group associated to G, see [19]). It follows from Theorem 3.6 that H is a closed subgroup of G (again in the usual sense). Theorem 3.7 shows that the existence of an injective normal ∗-homomorphism from vN(H) to vN(G) is naturally very closely related to the Herz restriction theorem (cf. Remark 3.8). To analyze the situation closer assume that H is a closed subgroup of a locally compact group G and consider the following statements: (1) the restriction map from C0(G) to C0(H) yields a surjective map from AG to AH ; 18 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN (2) the prescription λh 7→ λ(G) L2(H) and λ(G) injective ∗-homomorphism from vN(H) to vN(G); h h , h ∈ H (where λh denotes the (unitary) left shift by h on the corresponding (unitary) left shift by h on L2(G)) extends to a normal (3) the restriction of the left regular representation of G to H is quasi-equivalent to the left regular representation of H. It is very easy to see that they are all logically equivalent ((2) is essentially the definition of quasi-equivalence in (3), and the equivalence of (1) and (2) follows from a basic functional analytic argument, appearing already in [16, Section 0]). The first condition is the Herz restriction theorem. The third one can be viewed as a statement related to the theory of induced representations, and the induction-restriction procedure, as [26, Theorem 4.2] states that the left regular representation of G is the induction of the left regular representation of H. Interestingly, we could not locate an explicit statement of the condition (3) in literature. In the remainder of this section we will give an alternative proof of the implication (3)⇒(1) in Theorem 4.2. In particular this gives a new proof of Herz restriction theorem (cf. Remark 3.8). Our reasoning is based on existence of locally Baire cross-sections for the canonical projection G → G/H ([21]). Following the notation of [21, Section 4] we let q : G/H → G be a locally bounded Baire cross-section to the canonical quotient map G → G/H. We denote by Φ the bijection Let µ and β be Haar measures on G and H respectively and let λ be a quasi-invariant measure on G/H with associated ρ-function ρ ([13, Section 2.6]). In [21, Section 4] E.T. Kehlet shows that (G/H) × H ∋(cid:0)[g], h(cid:1) 7−→ q([g])h ∈ G. is a unitary map L2(G, µ) → L2(cid:0)(G/H) × H, λ × β(cid:1). We note that and the function ρ satisfies ψ 7−→ ρ− 1 2 · ψ◦Φ Φ(cid:0)[g], h(cid:1)h′ = Φ(cid:0)[g], hh′(cid:1), ρ(gh′) = ∆H (h′) ∆G(h′) ρ(g) (4.3) (4.4) (4.5) ([13]). We identify L2(G/H, λ) ⊗ L2(H) with L2(cid:0)(G/H) × H, λ × β(cid:1) in the usual way (the respective measures are regular) and define a unitary T : L2(G/H, λ) ⊗ L2(H) → L2(G) as the inverse of (4.3), i.e. (T ψ)(g) = ρ(g) 1 2 ψ(cid:0)Φ−1(g)(cid:1). 1 1 1 1 2 ∆H (h′) 2(cid:16)(cid:0)(1 ⊗ Rh′ )T ∗(cid:1)ψ(cid:17)(cid:0)Φ−1(g)(cid:1) 2 (T ∗ψ)(cid:0)[g0], h0h′(cid:1) 2 ρ(cid:0)Φ(cid:0)[g0], h0h′(cid:1)(cid:1)− 1 2 (cid:16) ∆H (h′) ∆G(h′)(cid:17)− 1 2 ∆H (h′) 2 ∆H (h′) 2 ρ(gh′)− 1 2 ∆H (h′) 2 ψ(gh′) 2 1 1 1 1 1 = ρ(g) = ρ(g) = ρ(g) = ρ(g) = ∆G(h′) 1 2 ψ(gh′) = (RG h′ ψ)(g). 2 ψ(cid:0)Φ(cid:0)[g0], h0h′(cid:1)(cid:1) ρ(g)− 1 2 ψ(gh′) For h′ ∈ H let Rh′ be the unitarized operator of right translation by h′ on L2(H) and let RG h′ denote the operator of right translation by h′ on L2(G). Fix g ∈ G and let Φ−1(g) = (cid:0)[g0], h0(cid:1). Taking into account (4.4) and (4.5) we compute (cid:16)(cid:0)T (1 ⊗ Rh′ )T ∗(cid:1)ψ(cid:17)(g) = ρ(g) Thus T (1 ⊗ Rh′ )T ∗ = RG h′ . This means that the (right) group von Neumann algebra vN(H) is isomorphic to the von Neumann subalgebra of vN(G) generated by the right shifts on G by elements from the subgroup H. This embedding is the map γ from Definition 3.1. In particular H is a closed subgroup of G in the sense of Vaes. CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 19 5. Cocommutative case Let again G and H be locally compact groups. Recall that in this case the dual locally compact r(G), quantum groups bG and bH of G and H are respectively defined by putting C0(bG) = C∗ C0(bH) = C∗ Theorem 5.1. Let π be a morphism from bH to bG and let, as usual, bπ denote the dual morphism r(H). We have the following result. from G to H, so that bπ : C0(H) ∋ f 7−→ f ◦θ ∈ M(cid:0)C0(G)(cid:1), (f ∈ C0(H)) for some continuous homomorphism θ : G → H. Then the following conditions are equivalent: (1) bH is a closed quantum subgroup of bG in the sense of Vaes (via the morphism π); (2) bH is a closed quantum subgroup of bG in the sense of Woronowicz (via the morphism π); (3) θ maps G onto H and the induced map θ : G/ker θ → H is a homeomorphism. Proof. That (1)⇒(2) is Theorem 3.5. Suppose that (2) holds, so that the morphism π : C∗(G) → C∗(H) maps to C∗(H) and is surjective. Since G and H are classical groups, the algebras AH , BH , AG and BG (as defined in Subsection 1.2) are the classical Fourier and Fourier-Stieltjes algebras of H and G respectively. In particular we have that BH = C∗(H)∗, and similarly BG = C∗(G)∗. Thus π∗ : BH → BG is weak∗-weak∗-continuous. Let G0 be the closure of the image of θ in H, and let θ0 : G → G0 be the corestriction of θ. By [18, Lemma 4.2] it follows that θ0 is an open surjection. We claim that G0 = H, from which (3) will follow. Indeed, if G0 6= H then as AH (and hence also BH ) is a regular algebra of functions on H (see [12, Lemme 3.2] or [52, Proposition 4.1.8]) we can find a non-zero b ∈ BH with b(s) = 0 for all s ∈ G0. As a map between function algebras, π∗ is simply π∗(b) = b◦θ, and so π∗(b) = 0. However, as π is surjection, π∗ is an isometry, and so π∗(b) 6= 0, a contradiction. Thus G0 = H as required. If (3) holds then as both K = ker θ and G/K are locally compact groups in their own right, they carry Haar measures, which we may normalize so that the Weyl formula holds: for f ∈ C00(G), Z f (s) ds = Z Z G G/K K It is not hard to see that the map IK : L1(G) ∋ f 7−→Z K f (st) dt d(sK). f (st) dt ∈ L1(G/K) is an algebra homomorphism and a metric surjection; see [28, Section 1.9.12] for example. We notice that then I ∗ K : L∞(G/K) → L∞(G) is an injective normal ∗-homomorphism which inter- twines the coproducts. As G/K is homeomorphic to H, the Haar measures on H and on G/K are proportional, and so the map γ0 : L∞(H) ∋ F 7−→ F ◦ θ ∈ L∞(G/K) is well-defined, and is hence a normal ∗-isomorphism which intertwines the coproduct. Then set γ = I ∗ K ◦γ0 : L∞(H) → L∞(G). So γ is an injective normal ∗-homomorphism which intertwines (cid:3) the coproducts, and a simple check shows that γ(cid:12)(cid:12)C0(H) = bπ, so (1) holds. H is a closed subgroup of bG in the sense of Woronowicz then π : C∗(G) → Cu Remark 5.2. Let G be a locally compact group and let H be a locally compact quantum group. If 0(H) is a surjection, 0(H) is cocommutative, hence of the form C∗(H) for some H, and it then follows that H and so Cu is a quotient of G. Let us give some indications of how the proof of [18, Lemma 4.2] proceeds. Firstly, arguing as in the proof of (3)⇒(1) above, it is not hard to reduce the problem to the case when θ is an injection. The key result is then [6, Theorem 1.3] which tells us that π∗(BH ) contains AG (as it is a weak∗-closed, conjugate closed, C∗(G)-module which, as a space of functions on G, separates the points of G; this final claim uses the assumption that θ is injective). 20 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN These ideas can be readily generalized to the setting of locally compact quantum groups. Proposition 5.3. Let G and H be locally compact quantum groups and let π : Cu 0(H) be a strong quantum homomorphism identifying H as a Woronowicz-closed subgroup of G. Furthermore, suppose that π∗(BH) contains the image of AG under the map Λ∗ G. Then G and H are isomorphic. 0(G) → Cu Proof. As π is onto, π∗ is an isometry onto its range, and so there is an isometric map φ : L∞(G)∗ → Cu G; clearly φ is a Banach algebra homomorphism. Let 0(H)∗ with π∗ ◦ φ = Λ∗ ψ = φ∗(cid:12)(cid:12)Cu G (cid:12)(cid:12)Cu Cu 0(G) π / / Cu 0(H) Then ψ◦π = φ∗ ◦π∗∗(cid:12)(cid:12)Cu 0 (H) : Cu 0(H) −→ L∞(G). 0 (G) = Λ∗∗ 0 (G) = ΛG, so we have the diagram ✉ ✉ ✉ ✉ ψ ✉ ✉ ✉ ✉ ΛG z✉ C0(G) 0(H) → Cu As π is onto, it follows that ψ is a ∗-homomorphism. Therefore it follows easily that ψ intertwines the coproducts. From the results of [27, Section 4] there is a strong quantum homomorphism 0(G) with ΛG ◦ ψ0 = ψ. Thus ΛG ◦ ψ0 ◦ π = ΛG. By passing to bicharacters and ψ0 : Cu 0(G). In particular, π must be applying [27, Lemma 4.13] it follows that ψ0◦π is the identity on Cu injective, and so an isomorphism. Thus the quantum groups G and H are isomorphic by Theorem 1.10. (cid:3) 6. Compact and discrete cases In this section we establish the equivalence of Definitions 3.1 and 3.2 when a potential quantum subgroup is compact (Theorem 6.1) and when the "larger" quantum group is discrete (Theorem 6.2). The first of these results shows in particular that if both quantum groups in question are compact, the definitions studied in this paper coincide with the one currently adopted in literature (see [3], [4], etc.); the second can be thought of as the generalization of the Herz restriction theorem to the context of discrete quantum groups. 6.1. Compact subgroups. Let G and H be locally compact quantum groups and assume further that H is compact. We will show in Theorem 6.1 that H is a closed subgroup of G in the sense of Vaes if and only if it is a closed subgroup of G in the sense of Woronowicz. However, before proceeding with this theorem let us make the following observation: consider a homomorphism Theorem 1.1(5) one could define injectivity of the homomorphism from H to G as the property from H to G described by π ∈ Mor(cid:0)Cu 0(G), Cu(H)(cid:1) (remember that H is compact). Based on that the range of π is strictly dense in M(cid:0)Cu(H)(cid:1). But Cu(H) is unital, so strict density of the range of π is equivalent to its norm-density. Moreover, since the image of a C∗-algebra under a ∗-homomorphisms is closed, π must be a surjection. By Theorem 3.6 this means that H is a closed subgroup of G in the sense of Woronowicz. In other words, the above argument shows that a compact quantum group H with an injective homomorphism into G is automatically a closed subgroup of G in the sense of Woronowicz (thus by Theorem 6.1 it is also closed in the sense of Vaes). In particular the notion of a quantum subgroup used e.g. in [38, Sections 4 and 5], [36], [4] is identical to those given in Definitions 3.1 and 3.2. Before proceeding let us also quickly note that a Woronowicz-closed subgroup of a compact quantum group is automatically compact (so, by Theorem 6.1 it is also Vaes-closed). The reason for this is that a quotient of a unital C∗-algebra is obviously unital (cf. Theorem 3.6(4)). Theorem 6.1. Let H be a closed subgroup of G in the sense of Woronowicz and assume that H is compact. Then H is a closed subgroup of G in the sense of Vaes. Proof. The subgroup H is compact, so we can write Cu 0(bH) = c0(bH), as the quantum group bH is discrete and hence coamenable. Moreover the C∗-algebra c0(bH) is a c0-direct sum of matrix algebras. It is not difficult to see the following / / z CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 21 • the multiplier algebra M(cid:0)c0(bH)(cid:1) is canonically isomorphic to the double dual c0(bH)∗∗, • for any C∗-algebra C of operators and any Φ ∈ Mor(cid:0)c0(bH), C(cid:1) the extension of Φ to a mapping M(cid:0)c0(bH)(cid:1) → M(C) ⊂ C′′ is σ-weakly continuous; in fact the extension of Φ to multipliers coincides with its normal extension ([29, Theorem 3.7.7]). Now, just as in the proof of Theorem 3.6 (Eq. (3.5)) we have Since V ∈ M(cid:0)C0(G) ⊗ C0(H)(cid:1) generates C0(H), we have by Proposition 2.1 that (ΛH ◦π)(cid:0)(ω ⊗ id)(WG)(cid:1) = (ω ⊗ id)(V ). (6.1) is dense in C(H). follows that Therefore if ω ∈ L∞(H)∗ is non-zero then it must be non-zero on some element (η ⊗ id)(V ). It (cid:8)(η ⊗ id)(V ) η ∈ L∞(bG)∗(cid:9) η(cid:0)(id ⊗ ω)(V )(cid:1) 6= 0, so (id ⊗ ω)(V ) 6= 0. In view of (6.1) this means that Λ bG◦bπ is injective on the subspace which coincides with the Fourier algebra (cid:8)(id ⊗ ω)(WH) ω ∈ L∞(H)∗(cid:9) ⊂ c0(bH) AbH =(cid:8)(id ⊗ ω)(WH) ω ∈ L∞(H)∗(cid:9), as bH is coamenable (cf. (1.8)). This last subspace contains the Pedersen ideal of c0(bH) (cf. (6.3)). By [7, Proposition II.8.2.4] this implies injectivity of Λ bG◦bπ on all of c0(bH). Finally Λ bG◦bπ remains injective after extension to ℓ∞(bH) = M(cid:0)c0(bH)(cid:1) because this extension coincides with the extension to the multiplier algebra and such extensions always preserve injectivity ([24, Proposition 2.1]). (cid:3) The arguments similar to these above appeared earlier in [35], an article which studies the relations between compact quantum subgroups of a coamenable locally compact quantum group G and left invariant C∗-subalgebras of C0(G). 6.2. Subgroups of discrete quantum groups. The main result of this subsection is the fol- lowing: Theorem 6.2. Let H be a closed subgroup of G in the sense of Woronowicz and assume that G is discrete. Then H is discrete and H is a closed subgroup of G in the sense of Vaes. We will prove Theorem 6.2 by generalizing to the setting of discrete quantum groups the theorem of Herz [15], [1, Proposition 3.23] and using Theorem 3.7 (cf. Remark 3.8). Since G is a discrete quantum group, the C∗-algebra C0(G) = Cu 0(G) = c0(G) is a c0-direct sum: Mnα c0(G) = Mα∈R 0(G)(cid:1) is in this case and the embedding of BG into M(cid:0)Cu (and AG is then mapped to the space of slices of WG = WG with normal functionals on L∞(bG)). BG = Cu(bG)∗ ∋ η 7−→ (η ⊗ id)(VVG) ∈ M(cid:0)c0(G)(cid:1) = ℓ∞(G) In particular one can use the functionals dual to the canonical basis of the Hopf ∗-algebra sitting inside C(G) ([51, Theorem 2.2], [5, Theorem 5.1]). These are normal and we easily see that their image in the mapping (6.2) spans the Pedersen ideal c00(G) of c0(G). The ideal c00(G) is the algebraic direct sum of the same family of matrix algebras. On the other hand these functionals (6.2) are linearly dense in L∞(bG)∗ (they correspond to density matrices on L2(bG) which are of finite rank). Therefore c00(G) ⊂ AG (6.3) with AG viewed as a subspace of c0(G). Theorem 6.3. The space AG is the closure in BG of c00(G). 22 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN Proof. As we mentioned before stating Theorem 6.3 the space c00(G) viewed inside BG is the space of functionals which are normal on L∞(bG) and whose density matrix is a finite rank operator. The closure of this space of functionals inside the space of all functionals on Cu(bG) is the space of all functionals which are normal on L∞(bG), i.e. the space AG. Proof of Theorem 6.2. The C∗-algebra C0(G) = c0(G) is a c0-direct sum of matrix algebras: (cid:3) c0(G) = Mα∈R Mnα. By [7, Proposition II.8.2.4] any ideal in c0(G) is of the form Mα∈R0 Mnα. for some R0 ⊂ R (the direct sum is still in c0-sense). Now if π : c0(G) → Cu 0(H) is the epimorphism corresponding to the embedding of H into G and R0 corresponds to the kernel of π, we see that Cu 0(H) is the c0-direct sum Cu 0(G) = Mα∈R\R0 Mnα. For the same reason the algebra C0(H) which is a (potentially proper) quotient of Cu 0(H) is also a c0-direct sum of matrix algebras. In particular C0(H) is an ideal in C0(G)∗∗. By [34, Theorem 4.4] H is a discrete quantum group. In particular Cu 0(H) = C0(H) = c0(H). Consider the adjoint of the map bπ : Cu(bH) → Cu(bG), i.e. bπ∗ : BG −→ BH. We now note that bπ∗ maps c00(G) into c00(H). Indeed, bπ∗ is the operation of pre-composing a functional with bπ. In particular, on the level of ℓ∞(G), where BG is embedded, we have bπ∗(cid:0)(η ⊗ id)(VVH)(cid:1) =(cid:0)[η◦bπ] ⊗ id(cid:1)(VVH) = (η ⊗ id)(cid:0)(bπ ⊗ id)(VVH)(cid:1) = (η ⊗ id)(cid:0)(id ⊗ π)(VVG)(cid:1) G)(cid:1), = ¯π(cid:0)(η ⊗ id)(VV where ¯π is the canonical extension of π to M(cid:0)c0(G)(cid:1) = ℓ∞(G). Also bπ∗ is a contraction for the norms on BG and BH (as an adjoint map of a contraction bπ : Cu(bH) → Cu(bG)). It follows from Theorem 6.3 that bπ∗ restricts to a contraction T : AG −→ AH with dense range; this completes the proof by applying Theorem 3.7. (cid:3) Acknowledgments The authors wish to thank Nico Spronk for helpful suggestions and pointing out several im- portant references. The second and third authors were supported by the National Science Centre (NCN) grant no. UMO-2011/01/B/ST1/06474. The fourth author was supported by National Sci- ence Centre (NCN) grant no. 2011/01/B/ST1/05011. The third author's visit to the University of Leeds was supported by EPSRC grant EP/I002316/1. References [1] G. Arsac: Sur l'espace de Banach engendr´e par les coefficients d'une repr´esentation unitare. Pub. D`ep. Math. Lyon 13 (1976), 1 -- 101. [2] S. Baaj & G. Skandalis: Unitaires multiplicatifs et dualit´e pour les produits crois´es de C∗-alg`ebres. Ann. Scient. ´Ec. Norm. Sup., 4e s´erie, t. 26 (1993), 425 -- 488. [3] T. Banica & J. Bichon: Quantum groups acting on 4 points. J. Reine Angew. Math. 626 (2009), 75 -- 114. [4] T. Banica, A. Skalski & P.M. So ltan: Noncommutative homogeneous spaces: the matrix case. J. Geom. Phys. 62 (2012), 1451 -- 1466. CLOSED QUANTUM SUBGROUPS OF LOCALLY COMPACT QUANTUM GROUPS 23 [5] E. Bedos, G.J. Murphy & L. Tuset: Co-amenability for compact quantum groups. J. Geom. Phys. 40 (2001), 130 -- 153. [6] M.E.B. Bekka, A.T. Lau & G. Schlichting: On invariant subalgebras of the Fourier-Stieltjes algebra of a locally compact group. Math. Ann. 294 (1992), 513 -- 522. [7] B. Blackadar: Operator algebras. Theory of C∗-algebras and von Neumann algebras. Encyclopedia of Mathematical Sciences, Vol. 122, Springer-Verlag 2006. [8] N. Bourbaki: Elements of Mathematics. General Topology. Chapters 1-4. Springer-Verlag 1989. [9] M. Daws: Multipliers, Self-Induced and Dual Banach Algebras. Dissertationes Mathematicae (Rozprawy Matematyczne) 470 (2010). [10] J. Delaporte & A. Derighetti: On Herz' extension theorem. Boll. Un. Mat. Ital. A 6 (1992), 245-247. [11] A. Derighetti: Relations entre les convoluteurs d'un groupe localement compact et ceux d'un sous-groupe ferm´e. Bull. Sci. Math. 106 (1982), 69-84. [12] P. Eymard: L'alg`ebre de Fourier d'un groupe localement compact. Bull. Soc. Math. France 92 (1964), 181 -- 286. [13] G.B. Folland: A course in abstract harmonic analysis. CRC Press 1995. [14] U. Franz, A. Skalski & R. Tomatsu: Idempotent states on the compact quantum groups and their classi- q(3). To appear in Journal of Noncommutative Geometry, available at q(2) and SO fication on U arXiv:0903.2363v2 [math.OA]. q(2), SU [15] C. Herz: Le rapport entre l'alg`ebre A p d'un groupe et d'un sous-groupe. C. R. Acad. Sci. Paris S´er. A-B 271 (1970), 244 -- 246. [16] C. Herz: Harmonic synthesis for subgroups. Ann. Inst. Fourier (Grenoble) 23 (1973), 91 -- 123. [17] Z. Hu, M. Neufang & Z.-J. Ruan: Completely bounded multipliers over locally compact quantum groups. Proc. Lond. Math. Soc. (3) 103 (2011), 1 -- 39. [18] M. Ilie & R. Stokke: Weak∗-continuous homomorphisms of Fourier-Stieltjes algebras, Math. Proc. Cam- bridge Philos. Soc. 145 (2008), 107 -- 120. [19] M. Kalantar & M. Neufang: From Quantum Groups to Groups. Available at arXiv:1110.5129v1 [math.OA]. [20] P. Kasprzak: Rieffel deformation of homogeneous spaces. J. Funct. Anal. 260 (2011), 146 -- 163. [21] E.T. Kehlet: Cross sections for quotient maps of locally compact groups.Math. Scand. 55 (1984), 152 -- 160. [22] J. Kustermans: Locally compact quantum groups in the universal setting. Int. J. Math. 12 (2001) 289 -- 338. [23] J. Kustermans & S. Vaes: Locally compact quantum groups. Ann. Scient. ´Ec. Norm. Sup. 4e s´erie, t. 33 (2000), 837 -- 934. [24] C.E. Lance: Hilbert C∗-modules: a toolkit for operator algebraists. London Mathematical Society Lecture Note Series 210, Cambridge University Press 1995. [25] T. Masuda, Y. Nakagami & S.L. Woronowicz: A C∗-algebraic framework for the quantum groups. Int. J. Math. 14 (2003), 903 -- 1001. [26] G. Mackey: Induced representations of locally compact groups I. Ann. Math. 55 (1952), 101 -- 139. [27] R. Meyer, S. Roy & S.L. Woronowicz: Homomorphisms of quantum groups, Munster J. Math. 4 (2011), 101 -- 124. [28] T.W. Palmer: Banach algebras and the general theory of ∗-algebras, Volume 1, Cambridge University Press 1994. [29] G.K. Pedersen: C∗-algebras and their automorphism groups. Academic Press 1979. [30] C. Pinzari: Embedding ergodic actions of compact quantum groups on C∗-algebras into quotient spaces. Int. J. Math. 18 (2007), 137 -- 164. [31] P. Podle´s: Przestrzenie kwantowe i ich grupy symetrii (Quantum spaces and their symmetry groups). Ph.D. Thesis, Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw (1989) (in Polish). [32] P. Podle´s: Symmetries of quantum spaces. Subgroups and quotient spaces of quantum SU(2) and SO(3) groups. Commun. Math. Phys. 170 (1995), 1 -- 20. [33] P. Podle´s & S.L. Woronowicz: Quantum deformation of Lorentz group. Comm. Math. Phys. 130 (1990), 381 -- 431. [34] V. Runde: Characterizations of compact and discrete quantum groups through second duals. J. Op. Th. 60 (2008), 415 -- 428. [35] P. Salmi: Compact quantum subgroups and left invariant C∗-subalgebras of locally compact quantum groups. J. Funct. Anal. 261 (2011), 1 -- 24. [36] P. Salmi & A. Skalski: Idempotent states on locally compact quantum groups. To appear in Quarterly Journal of Mathematics, available at arXiv:1102.2051v2 [math.OA]. [37] H.-J. Schneider: Principal homogeneous spaces for arbitrary Hopf algebras. Isr. J. Math. 72 (1990), 167 -- 195. [38] P.M. So ltan: Examples of non-compact quantum group actions. J. Math. Anal. Appl. 372 (2010), 224 -- 236. [39] P.M. So ltan & S.L. Woronowicz: A remark on manageable multiplicative unitaries. Lett. Math. Phys. 57 (2001), 239 -- 252. [40] P.M. So ltan & S.L. Woronowicz: From multiplicative unitaries to quantum groups II. J. Funct. Anal. 252 (2007), 42 -- 67. 24 MATTHEW DAWS, PAWE L KASPRZAK, ADAM SKALSKI, AND PIOTR M. SO LTAN [41] M. Takesaki: Theory of Operator Algebras I. Springer-Verlag 1979. [42] M. Takesaki & N. Tatsuuma: Duality and subgroups. Ann. Math. 93 (1971), 344 -- 364. [43] S. Vaes: A new approach to quantum and imprimitivity results. J. Funct. Anal. 229 (2005), 317 -- 374. [44] S. Vaes & L. Vainerman: On low-dimensional locally compact quantum groups. In Locally compact quan- tum groups and groupoids, IRMA Lect. Math. Theor. Phys. 2, de Gruyter, Berlin, (2003), pp. 127 -- 187. [45] S. Vaes & L. Vainerman: Extensions of locally compact quantum groups and the bicrossed product construction. Adv. Math. 175 (2003), 1 -- 101. [46] N.E. Wegge-Olsen: K-theory and C∗-algebras: a friendly approach. Oxford University Press 1993. [47] S.L. Woronowicz: Pseudogroups, pseudospaces and Pontryagin duality. Proceedings of the International Conference on Mathematical Physics, Lausanne 1979 Lecture Notes in Physics, 116, pp. 407 -- 412. [48] S.L. Woronowicz: Unbounded elements affiliated with C∗-algebras and non-compact quantum groups. Commun. Math. Phys. 136 (1991), 399 -- 432. [49] S.L. Woronowicz: C∗-algebras generated by unbounded elements. Rev. Math. Phys. 7, (1995), 481 -- 521. [50] S.L. Woronowicz: From multiplicative unitaries to quantum groups. Int. J. Math. 7, (1996), 127 -- 149. [51] S.L. Woronowicz: Compact quantum groups. In: Sym´etries quantiques, les Houches, Session LXIV 1995, Elsevier 1998, pp. 845 -- 884. [52] C. Zwarich: Von Neumann algebras for abstract harmonic analysis. Thesis at University of Waterloo, available at http://hdl.handle.net/10012/3920. School of Mathematics, University of Leeds, Leeds LS2 9JT, United Kingdom E-mail address: [email protected] Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw, Poland and Institute of Mathematics of the Polish Academy of Sciences, ul. ´Sniadeckich 8, 00 -- 956 Warszawa, Poland E-mail address: [email protected] Institute of Mathematics of the Polish Academy of Sciences, ul. ´Sniadeckich 8, 00 -- 956 Warszawa, Poland E-mail address: [email protected] Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw, Poland E-mail address: [email protected]
1806.03189	1	1806	2018-06-07T14:11:51	Local Lie derivations on von Neumann algebras and algebras of locally measurable operators	[ "math.OA" ]	Let $\mathcal{A}$ be a unital associative algebra and $\mathcal{M}$ be an $\mathcal{A}$-bimodule. A linear mapping $\varphi$ from $\mathcal{A}$ into an $\mathcal{A}$-bimodule $\mathcal{M}$ is called a Lie derivation if $\varphi[A,B]=[\varphi(A),B]+[A,\varphi(B)]$ for each $A,B$ in $\mathcal{A}$, and $\varphi$ is called a \emph{local Lie derivation} if for every $A$ in $\mathcal{A}$, there exists a Lie derivation $\varphi_{A}$ (depending on $A$) from $\mathcal{A}$ into $\mathcal{M}$ such that $\varphi(A)=\varphi_{A}(A)$. In this paper, we prove that every local Lie derivation on von Neumann algebras is a Lie derivation; and we show that if $\mathcal M$ is a type I von Neumann algebra with atomic lattice of projections, then every local Lie derivation on $LS(\mathcal M)$ is a Lie derivation.	math.OA	math	Local Lie derivations on von Neumann algebras and algebras of locally measurable operators Jun He1 and Guangyu An2∗ 1Department of Mathematics, Anhui Polytechnic University Wuhu 241000, China 2Department of Mathematics, Shaanxi University of Science and Technology 8 1 0 2 n u J 7 ] . A O h t a m [ 1 v 9 8 1 3 0 . 6 0 8 1 : v i X r a Xi'an 710021, China Abstract Let A be a unital associative algebra and M be an A-bimodule. A linear mapping ϕ from A into an A-bimodule M is called a Lie derivation if ϕ[A, B] = [ϕ(A), B] + [A, ϕ(B)] for each A, B in A, and ϕ is called a local Lie derivation if for every A in A, there exists a Lie derivation ϕA (depending on A) from A into M such that ϕ(A) = ϕA(A). In this paper, we prove that every local Lie derivation on von Neumann algebras is a Lie derivation; and we show that if M is a type I von Neumann algebra with atomic lattice of projections, then every local Lie derivation on LS(M) is a Lie derivation. Keywords: Lie derivation, local Lie derivation, von Neumann algebra, locally measurable operator. Mathematics Subject Classification(2010): 46L57; 47L35; 46L50 1 Introduction Let A be a unital associative algebra over the complex field C and M be an A- bimodule. An linear mapping δ from A into M is called a derivation if δ(AB) = δ(A)B + Aδ(B) for each A and B in A. In particular, a derivation δM defined by ∗Corresponding author. E-mail address: [email protected] 1 δM (A) = M A − AM for every A in A is called an inner derivation, where M is a fixed element in M. In [31], S. Sakai proves that every derivation on von Neumann algebras is an inner derivation. In [13], E. Christensen shows that every derivation on nest algebras on a Hilbert space H is an inner derivation. For more information on derivations and inner derivations, we refer to [14, 15, 19]. In [23, 25], R. Kadison and D. Larson introduce the concept of local derivations. A linear mapping δ from A into M is called a local derivation if for every A in A, there exists a derivation δA (depending on A) from A into M such that δ(A) = δA(A). In [23], R. Kadison proves that every continuous local derivation from a von Neu- mann algebra into its dual Banach module is a derivation. In [25], D. Larson and A. Sourour prove that if X is a Banach space, then every local derivation on B(X) is a derivation. In [21], B. Jonson shows that every local derivation from a C ∗-algebra into its Banach bimodule is a derivation. In [17, 18], D. Hadwin and J. Li characterize lo- cal derivations on non self-adjoint operator algebras such as nest algebras and CDCSL algebras. A linear mapping ϕ from A into an A-bimodule M is called a Lie derivation if ϕ[A, B] = [ϕ(A), B] + [A, ϕ(B)] for each A and B in A, where [A, B] = AB − BA is the usual Lie product. A Lie derivation ϕ is said to be standard if it can be decomposed as ϕ = δ + τ , where δ is a derivation from A into M and τ is a linear mapping from A into Z(A, M) such that τ [A, B] = 0 for each A and B in A, where Z(A, M) = {M ∈ M : AM = M A for every A in A}. In [22], B. Johnson proves that every continuous Lie derivation from a C ∗-algebra into its Banach bimodule is standard. In [28], M. Mathieu and A. Villena prove that every Lie derivation on a C ∗-algebra is standard. In [12], W. Cheung characterizes Lie derivations on triangular algebras. In [27], F. Lu proves that every Lie derivation on a completely distributed commutative subspace lattice algebra is standard. In [4], D. Benkovic proves that every Lie derivation on matrix algebra Mn(A) is standard, where n ≥ 2 and A is a 2-torsion free unital algebra. Similar to local derivations, In [10], L. Chen, F. Lu and T. Wang introduce the concept of local Lie derivations. A linear mapping ϕ from A into M is called a local Lie derivation if for every A in A, there exists a Lie derivation ϕA (depending on A) from A into M such that ϕ(A) = ϕA(A). In [10], L. Chen, F. Lu and T. Wang prove that every local Lie derivation on B(X) is a Lie derivation, where X is a Banach space of dimension exceeding 2. In [11], L. Chen and F. Lu prove that every local Lie derivation on nest algebras is a Lie derivation. In [26], D. Liu and J. Zhang prove that under certain conditions, every local Lie derivation on triangular algebras is a Lie derivation. In [20], J. He, J. Li, G. An and W. Huang prove that every local Lie derivation on some algebras such as finite von Neumann 2 algebras, nest algebras, Jiang-Su algebra and UHF algebras is a Lie derivation. Compare with the characterizations of derivations on Banach algebras, investigation of derivations on unbounded operator algebras begin much later. In [32], I. Segal studies the theory of noncommutative integration, and introduces various classes of non-trivial ∗-algebras of unbounded operators. In this paper, we mainly consider the ∗-algebra S(M) of all measurable operators and the ∗-algebra LS(M) of all locally measurable operators affiliated with a von Neumann algebra M. In [32], I. Segal shows that the algebraic and topological properties of the measurable operators algebra S(M) are similar to the von Neumann algebra M. If M is a com- mutative von Neumann algebra, then M is ∗-isomorphic to the algebra L∞(Ω, Σ, µ) of all essentially bounded measurable complex functions on a measure space (Ω, Σ, µ); and S(M) is ∗-isomorphic to the algebra L0(Ω, Σ, µ) of all measurable almost every- where finite complex-valued functions on (Ω, Σ, µ). In [5], A. Ber, V. Chilin and F. Sukochev show that there exists a derivation on L0(0, 1) is not an inner derivation, and the derivation is discontinuous in the measure topology. This result means that the properties of derivations on S(M) are different from the derivations on M. In [1, 2], Albeverio, Ayupov and Kudaybergenov study the properties of derivations on various classes of measurable algebras. If M is a type I von Neumann algebra, in [1], the authors prove that every derivation on LS(M) is an inner derivation if and only if it is Z(M) linear; in [2], the authors give the decomposition form of derivations on S(M) and LS(M); they also prove that if M is a type I∞ von Neumann algebra, then every derivation on S(M) or LS(M) is an inner derivation. If M is a properly infinite von Neumann algebra, in [6], A. Ber, V. Chilin and F. Sukochev prove that every derivation on LS(M) is continuous with respect to the local measure topology t(M); and in [7], the authors show that every derivation on LS(M) is an inner derivation. In [3], S. Albeverio and S. Ayupov give a characterization of local derivations on S(M), where M is an abelian von Neumann algebra. In [16], D. Hadwin and J. Li prove that if M is a von Neumann algebra without abelian direct summands, then every local derivation on LS(M) or S(M) is a derivation. In [9], V. Chilin and I. Juraev show that every Lie derivation on LS(M) or S(M) is standard. This paper is organized as follows. In Section 2, we recall the definitions of algebras of measurable operators and local measurable operators. In Section 3, we generalize the Corollary 3.2 in [20] and prove that every local Lie derivation on von Neumann algebras is a Lie derivation. In Section 4, we prove that if M is a type I von Neumann algebra with an atomic lattice of projections, then every local Lie derivation on LS(M) is a Lie derivation. 3 2 Preliminaries Let H be a complex Hilbert space and B(H) be the algebra of all bounded linear operators on H. Suppose that M is a von Neumann algebra on H and Z(M) = M∩M′ is the center of M, where M′ = {a ∈ B(H) : ab = ba for every b in M}. Denote by P(M) = {p ∈ M : p = p∗ = p2} the lattice of all projections in M and by Pf in(M) the set of all finite projections in M. For each p and q in P(M), if we define the inclusion relation p ⊂ q by p ≤ q, then P(M) is a complete lattice. Suppose that {pl}l∈λ is a family of projections in M, we denote pl = [ sup l∈λ l∈λ plH and inf l∈λ pl = \ plH. l∈λ If {pl}l∈λ is an orthogonal family of projections in M, then we have that pl = X pl. sup l∈λ l∈λ Let x be a closed densely defined linear operator on H with the domain D(x), where D(x) is a linear subspace of H. x is said to be affiliated with M, denote by xηM, if u∗xu = x for every unitary element u in M′. A linear operator affiliated with M is said to be measurable with respect to M, if there exists a sequence {pn}∞ n=1 ⊂ P(M) such that pn ↑ 1, pn(H) ⊂ D(x) and p⊥ n = 1 − pn ∈ Pf in(M) for every n ∈ N, where N is the set of all natural numbers. Denote by S(M) the set of all measurable operators affiliated with the von Neumann algebra M. A linear operator affiliated with M is said to be locally measurable with respect to M, if there exists a sequence {zn}∞ n=1 ⊂ P(Z(M)) such that zn ↑ 1 and znx ∈ S(M) for every n ∈ N. Denote by LS(M) the set of all locally measurable operators affiliated with the von Neumann algebra M. In [29], Muratov and Chilin prove that S(M) and LS(M) are both unital ∗-algebras and M ⊂ S(M) ⊂ LS(M); the authors also show that if M is a finite von Neumann algebra or dim(Z(M)) < ∞, then S(M) = LS(M); if M is a type III von Neumann algebra and dim(Z(M)) = ∞, then S(M) = M and LS(M) 6= M. 3 Local Lie derivations on von Neumann alge- bras 4 In this section, we consider local Lie derivations on von Neumann algebras. To prove our main theorem, we need the following lemma. Lemma 3.1. Let A1 and A2 be two unital algebras and A = A1 L A2. If the following five conditions hold: (1) each Lie derivation on A is standard; (2) each derivation on A is inner; (3) each local derivation on A is a derivation; (4) Z(A1) ∩ [A1, A1] = {0}; (5) A2 = [A2, A2], then every local Lie derivation on A is a Lie derivation. Proof. Denote the units of A, A1 and A2 by I, P and Q, respectively. For each A in A, we have that A = P A + QA = A1 + A2, where Ai ∈ Ai, i = 1, 2. In the following we suppose that ϕ is a local Lie derivation on A. By the definition of local Lie derivation, we know that for every A1 in A1, there exists a Lie derivation ϕA1 on A such that ϕ(A1) = ϕA1(A1). Since ϕA1 is standard and each derivation on A is inner, we can obtain that ϕ(A1) = ϕA1(A1) = δA1(A1) + τA1(A1) = [A1, TA1 ] + P τA1(A1) + QτA1(A1), where δA1 is a derivation on A, TA1 is an element in A, and τA1 is a linear mapping from A into Z(A) such that τA1([A, A]) = 0. It means that ϕ has a decomposition at A1. Next we show that the decomposition at A1 is unique. Assume there is another decomposition at A1, that is ϕ(A1) = ϕ A1(A1) = δ A1(A1) + τ A1(A1) = [A1, T A1 ] + P τ ′ ′ ′ ′ ′ A1(A1) + Qτ ′ A1(A1), ′ A1 is a derivation on A, T where δ from A into Z(A) such that τ ′ A1([A, A]) = 0. ′ A1 is an element in A and τ ′ A1 is a linear mapping Then we have that [A1, TA1] + P τA1(A1) + QτA1(A1) = [A1, T ′ A1] + P τ ′ A1(A1) + Qτ ′ A1(A1). Thus [A1, TA1] − [A1, T A1] = P τ ′ ′ A1(A1) − P τA1(A1) + Qτ ′ A1(A1) − QτA1(A1). Since [A1, TA1] − [A1, T A1(A1)−QτA1(A1) belongs to A2, we have that Qτ A1(A1) − P τA1(A1) ∈ A1 belong to A1, and A1(A1)−QτA1(A1) = 0. Moreover, A1] and P τ ′ ′ ′ ′ Qτ we can obtain that [A1, TA1] − [A1, T A1] = [A1, P TA1] − [A1, P T A1] ∈ [A1, A1], ′ ′ 5 and P τ ′ A1(A1) − P τA1(A1) ∈ Z(A1). By condition (4), it follows that [A1, TA1] − [A1, T It implies that δA1(A1) = δ unique. ′ A1(A1) and τA1(A1) = τ ′ ′ A1] = P τ A1(A1) − P τA1(A1) = 0. A1(A1). Hence the decomposition is ′ Now we have ϕA1 = δ1 + τ1, where δ1 is a mapping from A1 into A1 such that δ1(A1) = [A1, SA1] for some element SA1 in A1, and τ1 is a mapping from A1 into Z(A) such that τ1([A1, A1]) = 0. Next we prove that δ1 and τ1 are linear mappings. For each A1 and B1 in A1, we have that and ϕ(A1) = δ1(A1) + τ1(A1) = [A1, SA1] + τ1(A1), ϕ(B1) = δ1(B1) + τ1(B1) = [B1, SB1 ] + τ1(B1), ϕ(A1 + B1) = δ1(A1 + B1) + τ1(A1 + B1) = [A1 + B1, SA1+B1] + τ1(A1 + B1). Since ϕ is additive, through a discussion similar to that before, it implies that [A1 + B1, SA1+B1] = [A1, SA1] + [B1, SB1] and τ1(A1 + B1) = τ1(A1) + τ1(B1). It means that δ1 and τ1 are additive mappings. Using the same technique, we can prove that δ1 and τ1 are homogeneous. Hence δ1 and τ1 are linear mappings. For every A2 in A2, we have that ϕ(A2) = ϕA2(A2) = δA2(A2) + τA2(A2) = [A2, TA2] + τA2(A2), where δA2 is a derivation on A, TA2 is an element in A and τA2 is a linear mapping from A into Z(A) such that τA2([A, A]) = 0. By condition (5), we have that τA2(A2) = 0. Thus ϕ(A2) = [A2, TA2] = [A2, QTA2]. Let ϕA2 = δ2. Then we have δ2(A2) = [A2, SA2] for some element SA2 in A2. And obviously, δ2 is linear. Define two linear mappings as follows: δ(A) = δ1(A1) + δ2(A2), τ (A) = τ1(A1), for all A = A1 + A2 ∈ A. By the previous discussion, τ is a linear mapping from A into Z(A) such that τ ([A, A]) = 0. In addition, δ(A) = δ1(A1) + δ2(A2) = [A1, SA1] + [A2, SA2] = [A1 + A2, SA1 + SA2] = [A, SA1 + SA2]. 6 It means that δ is a local derivation. By condition (3), δ is a derivation. Notice that ϕ(A) = ϕ(A1) + ϕ(A2) = δ1(A1) + τ1(A1) + δ2(A2) = δ(A) + τ (A). Hence ϕ is a standard Lie derivation. By Lemma 3.1, we have the following result. Theorem 3.2. Every local Lie derivation on a von Neumann algebra is a Lie deriva- tion. Proof. Let A be a von Neumann algebra. It is well known that A = A1 L A2, where A1 is a finite von Neumann algebra, and A2 is a proper infinite von Neumann algebra. By [28, Theorem 1.1], we know that every Lie derivation on A is standard, by [31, Theorem 1], we know that every derivation on A is inner, and by [21, Theorem 5.3], we know that every local derivation on A is a derivation. Since A2 is a proper infinite von Neumann algebra, we known that A2 = [A2, A2](see in [33]). Hence it is sufficient to prove that Z(A1) ∩ [A1, A1] = {0}. Since A1 is finite and by [24, Theorem 8.2.8], it follows that there is a center-valued trace τ on A1 such that τ (Z) = Z for every Z in Z(A1) and τ ([A, B]) = 0 for each A and B in A1. Suppose that A ∈ Z(A1) ∩ [A1, A1], then we have that τ (A) = A and τ (A) = 0. it implies that A = 0. By Lemma 3.1, we know that every local Lie derivation on a von Neumann algebra is a Lie derivation. 4 Local Lie derivations on algebras of locally measurable operators In this section, we mainly consider local Lie derivations on algebras of all locally measurable operators affiliated with a type I von Neumann algebra. To prove the main result, we need the following lemmas. Lemma 4.1. Suppose that A is a commutative unital algebra and J = Mn(A). Then Z(J ) ∩ [J , J ] = {0} Proof. Let {ei,j}n A in J , we have that A = Pn i,j=1 aij eij, where aij ∈ A. i,j=1 be the system of matrix units in Mn(A). Then for every element Define a linear mapping τ from J into A by τ (A) = Pn i=1 aii for every A = i,j=1 aijeij ∈ J . Since A is commutative, it is not difficult to verify that τ ([A, B]) = 0 Pn for each A and B in J . 7 It should be noticed that Z(J ) = {A : A = Pn i=1 aeii, a ∈ A}. Suppose that i=1 aeii is an element in Z(J ) ∩ [J , J ], then by the definition of τ , we have that A = Pn τ (A) = na and τ (A) = 0. It implies that A = 0. Lemma 4.2. Suppose that A = Qi∈Λ Ai. If Z(Ai) ∩ [Ai, Ai] = {0} for every i ∈ Λ, then we have that Z(A) ∩ [A, A] = {0}. Proof. Let A = {ai}i∈Λ be an element in Z(A) ∩ [A, A]. Then for every i ∈ Λ, we have that ai ∈ Z(Ai) ∩ [Ai, Ai]. By assumption, it follows that ai = 0. Hence A = 0. Lemma 4.3. Suppose that M is a type I∞ von Neumann algebra. Then LS(M) = [LS(M), LS(M)]. Proof. By [30], we know that for every x in LS(M), there exists a sequence {zn} of mutually orthogonal central projections in M with P∞ n=1 zn = I, such that x = P∞ n=1 znx, and znx ∈ M for every n ∈ N. Since M is a proper infinite von Neumann algebra, it is well known that M = [M, M]. Thus we have that znx = Pk i , bn i ], where an i ∈ M for each n and i. i=1[an i , bn Set si = P∞ n=1 znan i and ti = P∞ n=1 znbn i . By the definition of locally measurable operators, it is easy to show that si and ti are two elements in LS(M). Since that {zn} are mutually orthogonal central projections, we can obtain that ∞ [si, ti] = [ X znan i , n=1 ∞ X n=1 znbn i ] = ∞ X n=1 zn[an i , bn i ], moreover, we have that k X [si, ti] = i=1 k ∞ X X i=1 n=1 zn[an i , bn i ] = ∞ X n=1 k zn( X i=1 [an i , bn i ]) = ∞ X n=1 znx = x. It follows that x ∈ [LS(M), LS(M)]. In the following we show the main result of this section. Theorem 4.4. Suppose that M is a type I von Neumann algebra with an atomic lattice of projections. Then every local Lie derivation from LS(M) into itself is a Lie derivation. Proof. By [24, Theorem 6.5.2], we know that M = M1 L M2, where M1 is a type If inite von Neumann algebra and M2 is a type I∞ von Neumann algebra. Hence by [2, Proposition 1.1], we have that LS(M) ∼= LS(M1) L LS(M2). In the following we will verify the conditions (1) to (5) in Lemma 3.1 one by one. By [9, Theorem 1], we know that every Lie derivation on LS(M) is standard; by [2, Corollary 5,12], we know that every derivation on LS(M) is inner for a von Neumann algebra with atomic lattice of projections. 8 It is proved in [17] that every local derivation on LS(M) is a derivation for a von Neumann algebra without abelian direct summands. While for an abelian von Neumann algebra with atomic lattice of projections, by [3, Theorem 3.8] we know that every local derivation on LS(M) is a derivation. Associated the two results, we can obtain each local derivation on LS(M) is a derivation for a von Neumann algebra with atomic lattice of projections. Since M1 is a type If inite von Neumann algebra, we know that M1 = L∞ n=1 An, where each An is a homogenous type In von Neumann algebra. Hence LS(M1) ∼= Q∞ n=1 LS(An). Since An is a homogenous type In von Neumann algebra, by [2] we know that LS(An) ∼= Mn(Z(LS(An))). By Lemmas 4.1 and 4.2, we know that the condition (4) in Lemma 3.1 holds. And by Lemma 4.3, the condition (5) in Lemma 3.1 holds. References [1] S. Albeverio, S. Ayupov, K. Kudaybergenov. Derivations on the algebra of measur- able operators affiliated with a type I von Neumann algebra. Siberian Adv. Math., 2008, 18: 86-94. [2] S. Albeverio, S. Ayupov, K. Kudaybergenov. Structure of derivations on various algebras of measurable operators for type I von Neumann algebras. J. Func. Anal., 2009, 256: 2917-2943. [3] S. Albeverio, S. Ayupov, K. Kudaybergenov, B. Nurjanov. Local derivations on algebras of measurable operators. Comm. In Contem. Math., 2011, 13: 643 -- 657. [4] D. Benkovic. Lie triple derivations of unital algebras with idempotents. Linear Multilinear Algebra, 2015, 63: 141 -- 165. [5] A. Ber, V. Chilin, F. Sukochev. Non-trivial derivation on commutative regular algebras. Extracta Math., 2006, 21: 107 -- 147. [6] A. Ber, V. Chilin, F. Sukochev. Continuity of derivations of algebras of locally measurable operators. Integr. Equ. Oper. Theory, 2013, 75: 527 -- 557. [7] A. Ber, V. Chilin, F. Sukochev. Continuous derivations on algebras of locally measurable operators are inner. Proc. London Math. Soc., 2014, 109: 65 -- 89. [8] M. Bresar, E. Kissin, S. Shulman, Lie ideals: from pure algebra to C-algebras, J. reine angew. Math., 2008, 623: 73 -- 121. [9] V. Chilin, I. Juraev. Lie derivations on the algebras of locally measurable operators. 2016, arXiv: 1608. 03996v1. [10] L. Chen, F. Lu, T. Wang. Local and 2-local Lie derivations of operator algebras on Banach spaces. Integr. Equ. Oper. Theory, 2013, 77: 109 -- 121. 9 [11] L. Chen, F. Lu. Local Lie derivations of nest algebras. Linear Algebra Appl., 2015, 475: 62 -- 72. [12] W. Cheung. Lie derivations of triangular algebra. Linear Multilinear Algebra, 2003, 512: 299 -- 310. [13] E. Christensen. Derivations of nest algebras. Math. Ann., 1977, 229: 155 -- 161. [14] H. Du, J. Zhang. Derivations on nest-subalgebras of von Neumann algebras. Chin. Ann. Math., 1996, A 17: 467 -- 474. [15] H. Du, J. Zhang. Derivations on nest-subalgebras of von Neumann algebras II. Acta. Math., 1997, A 40: 357 -- 362. [16] D. Hadwin, J. Li, Q. Li, X. Ma. Local derivations on rings containing a von Neu- mann algebra and a question of Kadison. 2013, arXiv:1311.0030v1. [17] D. Hadwin, J. Li. Local derivations and local automorphisms on some algebras. J. Oper. Theory, 2008, 60: 29 -- 44. [18] D. Hadwin, J. Li. Local derivations and local automorphisms. J. Math. Anal. Appl., 2004, 290: 702 -- 714. [19] J. He, J. Li, D. Zhao. Derivations, Local and 2-Local derivations on some algebras of operators on Hilbert C-bodules. Mediterr. J. Math., 2017, 14: article 230. [20] J. He, J. Li, G. An, W. Huang. Characterizations of 2-local derivations and local Lie derivations on some algebras. Sib. Math. J., 2017, accepted. [21] B. Johnson. Local derivations on C∗-algebras are derivations. Trans. Amer. Math. Soc., 2001, 353: 313 -- 325. [22] B. Johnson. Symmetric amenability and the nonexistence of Lie and Jordan deriva- tions. Math. Proc. Cambd. Philos. Soc., 1996, 120: 455 -- 473. [23] R. Kadison. Local derivations. J. Algebra, 1990, 130: 494 -- 509. [24] R. Kadison, J. Ringrose, Fundamentals of the Theory of Operator Algebras, I, Pure Appl. Math. 100, Academic Press, New York, 1983. [25] D. Larson, A. Sourour. Local derivations and local automorphisms. Proc. Sympos. Pure Math., 1990, 51: 187 -- 194. [26] D. Liu, J. Zhang. Local Lie derivations on certain operator algebras. Ann. Funct. Anal., 2017: 270 -- 280. [27] F. Lu. Lie derivation of certain CSL algebras. Israel J. Math., 2006, 155: 149 -- 156. [28] M. Mathieu, A. Villena. The structure of Lie derivations on C ∗-algebras. J. Funct. Anal., 2003, 202: 504 -- 525. [29] M. Muratov, V. Chilin. Algebras of measurable and locally measurable operators. Kyiv, Pratse In-ty matematiki NAN ukraini., 2007, 69: 390 pp, (Russian). 10 [30] M. Muratov, V. Chilin. Central extensions of *-algebras of measurable operators. Reports of the National Academy of Science of Ukraine, 2009, 7: 24 -- 28. (Russian). [31] S. Sakai. Derivations of W ∗-algebras. Ann. Math., 1966, 83: 273 -- 279. [32] I. Segal. A non-commutative extension of abstract integration. Ann, Math., 1953, 57: 401 -- 457. [33] H. Sunouchi. Infinite Lie rings. Tohoku Math. J., 1956, 8: 291 -- 307. 11
1711.09466	1	1711	2017-11-26T21:42:46	Measures of noncompactness on the standard Hilbert $C^*$-module	[ "math.OA" ]	We define a measure of noncompactness $\lambda$ on the standard Hilbert $C^$-module $l^2(\mathcal A)$ over a unital $C^$-algebra, such that $\lambda(E)=0$ if and only if $E$ is $\mathcal A$-precompact (i.e.\ it is $\varepsilon$-close to a finitely generated projective submodule for any $\varepsilon>0$) and derive its properties. Further, we consider the known, Kuratowski, Hausdorff and Istr\u{a}\c{t}escu measure of noncomapctnes on $l^2(\mathcal A)$ regarded as a locally convex space with respect to a suitable topology, and obtain their properties as well as some relationship between them and introduced measure of noncompactness $\lambda$.	math.OA	math	MEASURES OF NONCOMPACTNESS ON THE STANDARD HILBERT C∗-MODULE DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Abstract. We define a measure of noncompactness λ on the standard Hilbert C ∗-module l2(A) over a unital C ∗-algebra, such that λ(E) = 0 if and only if E is A-precompact (i.e. it is ε-close to a finitely generated projective submodule for any ε > 0) and derive its properties. Further, we consider the known, Kura- towski, Hausdorff and Istrăţescu measure of noncomapctnes on l2(A) regarded as a locally convex space with respect to a suitable topology, and obtain their properties as well as some relationship between them and introduced measure of noncompactness λ. 1. Introduction Measures of noncompactness (MNCs in further) have been studied almost for a century. Roughly speaking, a MNC is a function which assigns a real number to any bounded set in a given metric space, and this real number can be regarded as a characteristic of the extent to which A is not totally bounded, (that is relatively compact when completeness is supposed). There are many different MNCs on metric spaces, among them the most cited are: Kuratowski, Hausdorf and Istrăţescu MNC. Their definitions is given by: Definition 1.1. Let (M, d) be a metric space, and let A ⊆ M be a bounded set. a) The Kuratowski measure of noncompactness of A, denoted by α(A), is the infimum of all d > 0 such that A admits a partitioning into finitely many subsets whose diameters are less than d ([8], see also [9]). b) The Hausdorff measure of noncompactness of A, denoted by χ(A), is the infimum of all ε > 0 such that A has a finite ε-mash in M . (If we require that this ε-mash belongs to A, we refer to inner Hausdorff measure of noncompactness, denoted by χi(A).) c) The Istrăţescu measure of noncompactness, denoted by I(A) , is the infimum of all d > 0 such that there is an ε-discrete sequence in A, that is, a sequence xn ∈ A, such that d(xm, xn) ≥ ε for all m 6= n ([6]). Although different in general, these three functions share some common proper- ties, e.g. the following: Proposition 1.1. Let (M, d) be a metric space, and let µ be any of functions α, χ, I defined above. Then µ has the following properties: (a) regularity: µ(A) = 0 iff A is relatively compact; 2010 Mathematics Subject Classification. Primary: 46L08, Secondary: 47H08, 54E15. Key words and phrases. Hilbert module, measures of noncompactness, uniform spaces. The authors was supported in part by the Ministry of education and science, Republic of Serbia, Grant #174034. 1 2 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ (b) non-singularity: µ is equal to zero on any single-element set; (c) monotonicity: µ(A) ≤ µ(B), whenever A ⊆ B; (d) subadditivity: µ(A ∪ B) = max{µ(A), µ(B)}, for all A and B; If, in addition, M is a normed space, i.e., M is a linear space and the metric d is defined via d(x, y) = x − y then µ has additional properties: (e) algebraic semi-additivity: µ(A + B) ≤ µ(A) + µ(B), for all A, B; (f) semi-homogeneity: µ(tA) = tµ(A), for t ∈ C and all A ⊆ M ; (g) invariance under translations: µ(x+A) = µ(A), for all A ⊆ M and x ∈ M ; (h) Lipschitz continuity: µ(A) − µ(B) ≤ LdH (A, B), where L is an absolute constant (Lα = 2, Lχ = 1, LI = 2) and d(A, B) is the so called Hausdorff distance, that is dH (A, B) = max{supx∈A d(x, B), supy∈B d(y, A)}; (i) invariance to the transition to the closure and to the convex hull: µ(A) = µ(A) = µ(co A); (j) The functions α, χ and I are equivalent to each other, that is, (1.1) χ(A) ≤ I(A) ≤ α(A) ≤ 2χ(A) for all bounded A. These properties are well known and their proofs can be found throughout liter- ature, for instance in [16]. (Though, it should be mentioned that inner Hausdorff MNC does not satisfy the equality χi(co A) = χi(A), see [1, page 9]) Some of the mentioned properties were singled out in order to establish the axiomatic definition of the abstract notion of the MNC, and this was done in various different manners. For more detailed exposition on MNCs on metric or normed spaces, the reader is referred to [1], [4] or [17]. There is some extension of the MNC theory to the framework of uniform spaces which will be described in Section 3. Among all uniform spaces, we are specially interested in the standard Hilbert module l2(A) over a unital C∗-algebra A. It is defined by l2(A) = nx = (ξ1, ξ2, . . . ) ξj ∈ A, +∞ X j=1 j ξj converges in the norm topologyo, ξ∗ and it is equipped with the A-valued inner product l2(A) × l2(A) ∋ (x, y) 7→ +∞ X j=1 ξ∗ j ηj ∈ A, x = (ξ1, ξ2, . . . ), y = (η1, η2, . . . ). Since an arbitrary A-linear bounded operator on l2(A) does not need to have an adjoint, the natural algebra of operators is Ba(l2(A)) - the algebra of all A-linear bounded operators on l2(A) having an adjoint. It is known that Ba(l2(A)) is a C∗-algebra, as well. Among all operators in Ba(l2(A)), those that belong to the linear span of the operators of the form x 7→ Θy,z(x) = z hy, xi (y, z ∈ l2(A)) are called finite rank operators. The norm closure of finite rank operators is known as the algebra of all "compact" operators. The quotation marks are usually written in order to emphasize the fact that "compact" operators does not maps bounded sets into relatively compact sets, as it is the case in the framework of Hilbert (and also Banach) spaces, though they share many properties of proper compact operators on a Hilbert space, [12], [13]. MEASURES OF NONCOMPACTNESS 3 (For general literature concerning Hilbert modules over C∗ algebras, including the standard Hilbert module, the reader is referred to [10] or [14].) In our earlier work [7], we construct a locally convex topology τ on l2(A) such that T ∈ Ba(l2(A)) is "compact" implies that its image of the unit ball is totally bounded with respect to τ . The converse is obtained in the special case A = B(H). The aim of this note is to introduce a MNC on the standard Hilbert module l2(A) and to derive its properties. In Section 2, we introduce the MNC λ such that for any bounded set E ⊆ l2(A) the equality λ(E) = 0 holds if and only if E is precompact in the sense of [14, Proposition 2.6], which implies that the corresponding MNC of a given operator T is equal to 0 if and only if T is "compact". In Section 3 we list some known results on MNCs on uniform spaces and also derive some simple generalizations. In Section 4, we discuss generalizations of Kuratowski, Hausdorff and Istrăţescu measures of noncomapctness on l2(A) and obtain their relationship with MNC λ. We use the following basic and also simple facts on Hilbert modules, that can be found throughout the literature. To make proofs more easy we list and prove them. (F1) Let z1 ⊥ z2. Then z1 + z2 ≥ z1. Indeed, we have hz1 + z2, z1 + z2i = hz1, z1i + hz2, z2i ≥ hz1, z1i , in the order defined by the positive cone in A. Therefore z1 + z22 = hz1 + z2, z1 + z2i ≥ hz1, z1i = z12. (F2) Let M be a projective finitely generated submodule of l2(A), and let x ∈ l2(A) be arbitrary. Then d(x, M ) = x − PM x, where PM is orthogonal projection onto M with null-space M ⊥. Indeed, by [14, Theorem 1.4.5] M is orthogonally complemented, l2(A) = M ⊕ M = PM , the range of PM M ⊥, and there is PM : l2(A) → l2(A) such that P 2 is M and the kernel of PM is M ⊥. Let y ∈ M be arbitrary. Then M = P ∗ x − y = (x − PM x) + (PM x − y) ≥ x − PM x, by (F1), because x − PM x ∈ M ⊥ and PM x − y ∈ M . 2. Measure of noncompactness λ Throughout this section, A will always denote a unital C∗-algebra, and its unit will be denoted by 1. Also, l2(A) will denote the standard Hilbert C∗-module over A defined in the introduction. In [14, Proposition 2.6] A-precompact sets were defined as those bounded sets E such that for all ε > 0 there is a free finitely generated module M ∼= An such that d(E, M ) := sup x∈E d(x, M ) = sup x∈E inf y∈M d(x, y) < ε. We generalize this notion in the following way. Definition 2.1. Let E ⊆ l2(A) be a bounded set. The measure of noncompactness of E, denoted by λ(E) is the greatest lower bound of all η > 0 for which there is a 4 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ free finitely generated module M ≤ l2(A) such that d(E, M ) := sup x∈E inf y∈M d(x, y) < η. Proposition 2.1. The measure of noncompactness λ(E) can be computed as: (1) λ(E) = inf M∈F supx∈E d(x, M ), where F is the set of all free finitely gen- erated modules; (2) λ(E) = limn→+∞ supx∈E x − Pnx = inf n≥1 supx∈E x − Pnx, where Pn : l2(A) → l2(A) is given by Pn(x1, x2, . . . ) = (x1, x2, . . . , xn, 0, 0, . . . ). Proof. The equation (1) is obvious. The sequence I − Pn is decreasing. Hence limn→+∞ supx∈E x − Pnx = inf n≥1 supx∈E x − Pnx. Since Pnl2(A) is a free finitely generated module, we have immediately λ(E) ≤ inf n≥1 sup x∈E x − Pnx. To get the opposite inequality, let ε > 0. Then, there is a free finitely generated module M such that d(x, M ) < λ(E) + ε. Denote the projection on M by Q. By [14, Proposition 2.2.1.], Q − PnQ → 0 as n → +∞. Since PnQx ∈ Pnl2(A), we have by (F2) x − Pnx =d(x, Pnl2(A)) ≤ x − PnQx ≤ ≤x − Qx + Qx − PnQx ≤ λ(E) + ε + KQ − PnQ, where x ≤ K for all x ∈ E (E is bounded). Thus x − Pnx ≤ λ(E) + ε + KQ − PnQ → λ(E) + ε, as n → +∞, sup x∈E which finishes the proof. (cid:3) Proposition 2.2. The measure of noncompactness λ has the following properties (1) if E ⊆ F then λ(E) ≤ λ(F ); (2) λ(E ∩ F ) ≤ min{λ(E), λ(F )}; (3) λ(E ∪ F ) ≤ max{λ(E), λ(F )}; (4) λ(E + F ) ≤ λ(E) + λ(F ). (5) λ(Ea) ≤ λ(E)a, where a ∈ A. If, in addition, a is invertible, then also In particular, when a is unitary then λ(Ea) = a−1−1λ(E) ≤ λ(Ea). λ(E). (6) λ(co E) = λ(E), where co E = {Pn is the convex hull of E. i=1 tixi 0 ≤ ti ∈ R,Pn i=1 ti = 1, xi ∈ E} Proof. (1) It is obvious. (2) This follows from (1). (3) Let d = max{λ(E), λ(F )}. Then for all x ∈ E, as well as for all x ∈ F , we have x − Pnx ≤ d + ε, for all ε > 0 and n large enough. Hence the result. (4) Let z ∈ E + F . Then z = x + y for some x ∈ E, y ∈ F . We have z − Pnz ≤ x − Pnx + y − Pny ≤ λ(E) + ε + λ(F ) + ε, for n large enough. (5) Any z ∈ Ea is of the form z = ya for some y ∈ E. Therefore z − Pnz ≤ y − Pny a and the inequality follows by taking a limit. If a has the inverse a−1, then E = (Ea)a−1 and by previous λ(E) ≤ λ(Ea)a−1. MEASURES OF NONCOMPACTNESS 5 i=1 tixi for some xi ∈ E, and positive ti such that Pn (6) Let x ∈ co E. Then x = Pn i=1 ti = 1. We have n x − Pnx = (cid:13)(cid:13)(cid:13) ti(xi − Pnxi)(cid:13)(cid:13)(cid:13) X i=1 ≤ n X i=1 tixi − Pnxi ≤ sup x∈E x − Pnx. Thus, supx∈co E x−Pnx ≤ supx∈E x−Pnx. The opposite inequality is obvious. The required follows from Proposition 2.1-(2). (cid:3) Proposition 2.3. Let B denote the unit ball in l2(A). Then λ(B) = 1. Proof. Any submodule contains the origin. Hence λ(B) ≤ 1. Let 0 < δ < 1, and let M be some free finitely generated submodule of l2(A). Then, there is nontrivial y ∈ M ⊥ and δy−1y ∈ B ∩ M ⊥. We have d(B, M ) ≥ d(δy−1y, M ). However, δy−1y ⊥ M which implies that for all x ∈ M we have δy−1y − x2 = δ2 + x2 ≥ δ2. Hence d(B, M ) ≥ δ. Thus λ(B) ≥ δ. Corollary 2.4. If E ⊆ F + δB then λ(E) ≤ λ(F ) + δ. Proof. Immediately follows from Proposition 2.2-(4) and Proposition 2.3. (cid:3) (cid:3) Proposition 2.5. The measure of noncompactness λ has the following continuity properties: (1) λ(E) − λ(F ) ≤ dH (E, F ) = max{d(E, F ), d(F, E)} (dH stands for the so called Hausdorf distance.) (2) λ(E) = λ(E) (E stands for the norm closure of E); (3) λ(E) = 0 iff E is A-precompact; (4) λ(E) ≤ supx∈E x. Proof. (1) Let d = dH (E, F ). Then E ⊆ F + dB and by Corollary 2.4 λ(E) ≤ λ(F ) + d, i.e. λ(E) − λ(F ) ≤ d. Similarly, F ⊆ E + dB implying λ(F ) − λ(E) ≤ d. (2) As it is easy to see dH (E, E) = 0 we can apply the previous item. (3) Follows directly from the definition. (4) Follows from E ⊆ (supx∈E x) · B. (cid:3) Example 2.1. In Proposition 2.2-(5) the strict inequalities might hold. Indeed, let the algebra A contain a nontrivial projection, say p, and let A(1 − p) is isomorphic to A. (For instance A = L∞(0, 1) and p = χ[0,1/2].) Let E = n(a1p + b1(1 − p), a2p + b2(1 − p), . . . ) ∞ X n=1 a∗ nan ≤ 1, ∞ X n=1 nbn ≤ 4o b∗ and let a = p. Then p = 1 and λ(E) ≥ 2, since E contains a copy of a ball of radius 2 in l2(A) (when aj = 0). On the other hand λ(Ep) ≤ 1. Indeed, Ep is contained in the unit ball of l2(A). Finally, we want to define MNC of an operator T ∈ Ba(l2(A)). As it is expected, it will be the MNC of its image of the unit ball. 6 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Definition 2.2. Let T ∈ Ba(l2(A)) be an adjointable operator. We set where B1 is the unit ball in l2(A). λ0(T ) = λ(T (B1)), Proposition 2.6. The function λ0 has the following properties: (a) λ0 is subadditive, i.e. λ0(T1 + T2) ≤ λ0(T1) + λ0(T2); (b) λ0 is positively homogeneous, i.e. λ0(cT ) = cλ0(T ), for all c > 0 and all T ∈ Ba(l2(A)); (c) λ0(T ) ≤ T , for all T ∈ Ba(l2(A)); Proof. Direct verification. (cid:3) We will be able to say more on the MNC λ0 in section 4. 3. Measure of noncompactness on uniform spaces - known results Let us recall some basic definitions and facts concerning uniform spaces. For more details see [18] or [5]. Uniform spaces are those topological spaces in which one can deal with notions such as Cauchy sequence, Cauchy net or uniform continuity. Although it is usual to define them as spaces endowed with a family of sets in X × X given as some kind of neighborhoods of the diagonal, so called entourages, for our purpose it is more convenient to give an equivalent definition, via a family of semi-metrics. Definition 3.1. A nonempty set endowed with a family of semi-metrics, functions dα : X × X → [0, +∞) satisfying (i) dα(x, y) ≥ 0; (ii) dα(x, y) = dα(y, x); (iii) dα(x, z) ≤ dα(x, y) + dα(y, z) is called a uniform space. Remark 3.1. There is some ambiguity in literature; sometimes, functions dα from the previous definition are called "pseudo-metrics", whereas the term "semi-metric" is reserved for a different notion. All dα are metrics except they do not distinguish points, i.e. there might be dα(x, y) = 0 for some x = y. However it is provided that for all x = y there is an α such that dα(x, y) > 0. The family of sets Bdα(x; ε) = {y ∈ Xdα(x, y) < ε} makes a basis for some topology. It is well known that a topological space X is a uniform space if and only if it is completely regular. Any locally convex topological vector space is a uniform space. Indeed, there is a family of semi-norms generating its topology. This family can be obtained by Minkowski functionals of basic neighborhoods of zero. And an arbitrary semi-norm define a semi-metric in a natural way. Conversely, any family of semi-norms that distinguishes points leads to a locally convex Hausdorff topological vector space. We point out two generalizations of the notion of MNC to the framework of uniform spaces, i.e, those topological spaces that arise from the family of semi- metrics. Sadovskii [18] considered a uniform space X and a family P of semi-metrics that are uniformly continuous on X × X. Starting with Kuratowski and Hausdorff MNC on a semi-metric space (which is defined exactly as on a metric space), Sadovskii MEASURES OF NONCOMPACTNESS 7 defined corresponding MNCs α, χ : Υ → G, where Υ denotes the family of all subsets that are bounded with respect to any semi-metric p ∈ P on the given uniform space, and G denotes the set of functions g : P → [0, +∞) with uniformity generated by pointwise convergence and the natural partial ordering: g1 ≤ g2 ⇔ (∀p ∈ P ) g1(p) ≤ g2(p). Their definitions are [α(E)](p) = inf nd > 0 : E = m [ j=1 Ej, for some Ej, diam(Ej) < do, [χ(E)](p) = inf nε > 0 : E ⊆ m [ j=1 Bp(xj ; ε) for some xj ∈ Xo. For such defined α and χ, in [18, §1.2.3. and §1.2.5.], the following is proved, provided that the family P generates the topology on X. Theorem 3.1. The Kuratowski and Hausdorff measures of noncompactness (µ = α or µ = χ) have the following properties: (a) µ is non-singular, that is, they are zero on any single-element set; (b) µ is continuous, that is, for all E ∈ Υ, p ∈ P and ε > 0 there is an entourage V in X such that for all E1 that is V -close to E there holds µ(E1)(p) − µ(E)(p) < ε; (c) µ is semi-additive, that is, for all E1, E2 we have µ(E1 ∪ E2) = max{µ(E1), µ(E2)}; (d) The function is algebraically semi-additive, that is, µ(E1 + E2)(p) ≤ µ(E1)(p) + µ(E2)(p) for all E1, E2 ∈ Υ; (e) µ is invariant under shifts, that is, µ(x + E) = µ(E) for all E ∈ Υ, x ∈ X; (f) µ is invariant under the transition to its closure and to the convex hull of the set, that is, µ(co E) = µ(E) = µ(E) for all E ∈ Υ; (g) µ is uniformly continuous, that is, for all p ∈ P and ε > 0 there is an entourage V in X such that for all V -close E1 and E2 there holds µ(E1)(p) − µ(E2)(p) < ε; (h) The functions α, χ and I are equivalent to each other, that is, χ(Ω) ≤ I(Ω) ≤ α(Ω) ≤ 2χ(Ω) for all Ω ∈ Υ. Although it was not done in [18], one can also define the Istrăţescu MNC in a similar way, i.e. [I(E)](p) = inf{ε > 0 : E ∋ xn, p(xn, xm) ≥ ε for all m 6= n}. By (1.1) one can easily derive (3.1) [χ(E)](p) ≤ [I(E)](p) ≤ [α(E)](p) ≤ 2[χ(E)](p) for all bounded E. Also, Theorem 3.1 hold for µ = I. Part (a) is obvious, parts (d) and (e) follow from [3, Proposition 1], part (c) follows from [6], parts (b) and (g) from (3.1) and the corresponding properties of χ, and finally (f) can be derived from [1, Theorem 1.3.4]. Arandjelović [2] dealt with an arbitrary uniform space, and gave the following: 8 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Definition 3.2. Let X be a uniform space, metric space or semi-metric space. Any function Φ defined on the partitive set of X, which satisfies the following: (1) Φ(E) = +∞ if and only if E is unbounded; (2) Φ(E) = Φ(E); (3) from Φ(E) = 0 follows that E is totally bounded set; (4) from E ⊆ F it follows Φ(E) ≤ Φ(F ); (5) if X is complete, and if {En}n∈N is a sequence of closed subsets of X such that En+1 ⊆ En for each n ∈ N and limn→∞ Φ(En) = 0, then K = Tn∈N En is a nonempty compact set. is called a measure of noncompactness on X. Remark 3.2. Note that the only nontrivial requirement in (5) is that K is nonempty. Moreover, condition (5) can be replaced by a weaker one - Φ(A ∪ {x}) = Φ(A). It was shown in [15], see also []. Theorem 3.2. [2, Theorem 3] Let X be a uniform space and let {dii ∈ I} be a family of semi-metrics which defines topology on X. Denote by µi arbitrary MNC on the semi-metric space (X, di) for each i ∈ I. Then the function µ∗ : X → [0, +∞] defined by for each E ∈ X, is a measure of noncompactness on X. µ∗(E) = sup i∈I µi(E) Uniform spaces make a proper subclass of all topological spaces, but still wide enough. For instance all topological vector spaces are uniform spaces. Hence, we can apply results of Arandjelović to topological vector spaces. 4. Measures of "noncompactness" over standard Hilbert W ∗-modul l2(B(H)) In this section we shall discuss standard Hilbert modules over a W ∗-algebra A, a narrower class then that considered in section 2. As it is well known, A always has a unit. In our earlier work [7], we construct a locally convex topology τ on l2(A) such that T ∈ Ba(l2(A)) is "compact" implies its image of the unit ball is totally bounded with respect to τ . This topology is defined via the family of semi-norms pϕ,y (4.1) ∞ pϕ,y(x) = vuut X j=1 ϕ(η∗ j ξj)2, x = (ξ1, ξ2, . . . ) ∈ l2(A), where ϕ ∈ A∗ is a normal state and y = (η1, η2, ...) is a sequence of elements in A such that (4.2) ϕ(η∗ j ηj) = 1. sup j≥1 Also, in special case, where A = B(H) is the full algebra of all bounded operators on a Hilbert space H, the converse is also proved, i.e. that any T ∈ Ba(l2(A)) whose image of the unit ball is totally bounded with respect to τ must be "compact". Construction described in the previous section endows the space l2(A) by the corresponding Kuratowski, Hausdorff and Istrăţescu measure of noncompactness, MEASURES OF NONCOMPACTNESS 9 α, χ, I : Υ → G, [α(E)](pϕ,y) = inf nε > 0 : E = n [ i=1 Si, pϕ,y(x′ − x′′) < ε, ∀x′, x′′ ∈ Sio, [χ(E)](pϕ,y) = inf nε > 0 : E ⊂ [ i=1 where Bpϕ,y (xi, ε) = {ypϕ,y(y − xi) < ε}, and [I(E)](pϕ,y) = sup{ε > 0 : there is S ⊂ E such that pϕ,y(x′−x′′) ≥ ε, ∀x′, x′′ ∈ S}. Bpϕ,y (xi, ε), xi ∈ l2(A)o, n The function α, χ and I can be regarded as functions depending on two variables, on the bounded set Ω and on the semi-norm pϕ,y. If we want to obtain a MNC that not depends on a particular semi-norm, we can use the functions χ∗, α∗, I ∗ : Υ → [0, +∞) defined by (4.3) χ∗(E) = sup pϕ,y ∈P α∗(E) = sup pϕ,y ∈P I ∗(E) = sup pϕ,y∈P [χ(E)](pϕ,y), [α(E)](pϕ,y), [I(E)](pϕ,y) for each E ∈ Υ, where P is the set of all semi-norms of the form (4.1). Since α, χ and I annihilates all singletons (Theorem 3.1-(a)), they satisfy condition in Remark 3.2, and hence they are measures of noncompactness in the sense of Definition 3.2. By Theorem 3.2 α∗, χ∗ and I ∗ are measures of noncompactness on (l2(A), τ ) in the sense of Definition 3.2, as well. Also, properties (c), (d), (e) and (f) in Theorem 3.1 are easily transferred to α∗, χ∗ and I ∗, by taking a supremum. Finally, by (3.1), we have (4.4) χ∗(E) ≤ I ∗(E) ≤ α∗(E) ≤ 2χ∗(E). Remark 4.1. Note that l2(A) is rarely complete, due to [7, Proposition 3.3]. There- fore, the condition (5) in Definition 3.2 is vague, unless l2(A)′ ∼= l2(A) which is equivalent to the condition that A is finite dimensional. We want to place the MNC λ, discussed in section 2, somewhere in the preceding chain of inequalities. Proposition 4.1. For any bounded set E ⊆ l2(A), we have χ∗(E) ≤ λ(E). Proof. Let E be a bounded set, and let Pn denote the projection to the first n coordinates in l2(A), i.e. Pn(ξ1, ξ2, . . . ) = (ξ1, ξ2, . . . , ξn, 0, . . . ). Since E ⊆ PnE + (I − Pn)E, and since χ∗ is subadditive, we have χ∗(E) ≤ χ∗(PnE) + χ∗((I − Pn)E). However, by ([7, Proposition 3.4.]), the set PnE is totally bounded, and we have χ∗(PnE) = 0. Hence χ∗(E) ≤ χ∗((I − Pn)E) ≤ sup x∈E (I − Pn)x, for all n ∈ N. Therefore, by Proposition 2.1-(2), χ∗(E) ≤ λ(E). (cid:3) 10 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ The preceding Proposition establishes a lower bound of λ. Before we obtained an upper bound for λ, in a special case, we introduce balanced sets. Definition 4.1. Let E ⊂ l2(A) be a bounded set. (a) We say that E is A-balanced if x · u ∈ E whenever x ∈ E and u ∈ A is unitary. (This definition is motivated by the notion of balanced sets on topological vector spaces over the field C, where u unitary is reduced to u = 1.) (b) By A-balanced hull of E we assume the minimal balanced set containing E, that is S Eu, where the union is taken over all unitaries u ∈ A. In Proposition 2.2-(5) we proved λ(Eu) = λ(E). We give two extensions of this statement, the first of them concerning the balanced hull. Proposition 4.2. Let E ⊆ l2(A) be a bounded set and let F be its A-balanced hull. Then λ(F ) = λ(E). Proof. For all x ∈ E and all unitaries u, we have u = 1, and hence xu−Pnxu ≤ x − Pnxu. Therefore, by Proposition 2.1-(2), we have λ(F ) = lim n→+∞ sup x∈E,u−unitary xu − Pnxu ≤ lim n→+∞ sup x∈E x − Pnx = λ(E). The opposite inequality λ(E) ≤ λ(F ) follows from E ⊆ F . (cid:3) Proposition 4.3. Let E ⊆ l2(A) be a bounded set, let u ∈ A be a unitary and let µ stands for any of Kuratowski, Hausdorff or Istrăţescu MNC. Then µ∗(Eu) = µ∗(E). Proof. First, observe that given a normal state ϕ on A and a unitary u ∈ A, the mapping ϕu, ϕu(x) = ϕ(u∗xu) is also a normal state. Obviously, ϕu(1) = 1 and ϕu(x) ≥ 0 whenever x ≥ 0. Hence, it suffices to prove that ϕu is normal. Let xα be an increasing net with the least upper bound x. Then uxαu∗ is also an increasing net, bounded by uxu∗. Thus, its least upper bound is less then uxu∗. Moreover, it is equal to uxu∗ by interchanging roles. Therefore sup ϕu(xα) = sup ϕ(uxαu∗) = ϕ(uxu∗) = ϕu(x). Let ε > 0 be arbitrary. By (4.3) there is a semi-norm pϕ,y ∈ P such that [µ(Eu)](pϕ,y) > µ∗(Eu) − ε. Next, we have pϕ,y(xu) = pϕu,yu, where yu∗ = (η1u∗, η2u∗, . . . ). Indeed +∞ +∞ pϕ,y(xu)2 = ϕ(η∗ j ξju)2 = X j=1 X j=1 ϕ(u∗(ηj u∗)∗ξju)2 = pϕu,yu∗(x)2. (Note that the pair (ϕu, yu∗) trivially satisfies (4.2).) Therefore, [µ(E)](pϕu,yu∗) = [µ(Eu)](pϕ,y) > µ∗(Eu) − ε and hence µ∗(E) ≥ µ∗(Eu). The opposite inequality follows by E = (Eu)u−1. (cid:3) Remark 4.2. We don't know whether Kuratowski, Hausdorff and Istrăţescu MNCs are stable with respect to balanced hull. Now, we are going to derive an upper bound for the MNC λ. Namely, in a special case, where A = B(H), we can obtain the exact position of λ in the inequality chain (4.4) for balanced sets. MEASURES OF NONCOMPACTNESS 11 Theorem 4.4. Let A = B(H). Let E ⊆ l2(A) be an A-balanced set. Then where E = supx∈E x. λ(E) ≤ pEI ∗(E), Proof. Let Pk denote the projection to the first k coordinates, i.e. Pk(ξ1, ξ2, . . . ) = (ξ1, . . . , ξk, 0, 0, . . . ). It is well known that all Pk are "compact". If inf k≥1 (I − Pk)E = 0, then from Proposition 4.1 and (4.4), it follows λ∗(E) = χ∗(E) = I ∗(E) = 0. So, let δ = inf k≥1 (I − Pk)E > 0. Then immediately, δ ≤ E. Choose ε > 0 such that ε < δ2. Define the sequence of projections Qn ∈ {P1, P2, . . . } and the sequences of vectors xn and zn ∈ l2(A) in the following way. Let Q0 = 0. If Qn−1 is already defined, there is xn ∈ E 2 (cid:16)δ + pδ2 − ε/2(cid:17). Since such that xn ≥ (I − Qn−1)xn > C1, where C1 = 1 limk→+∞ (I − Pk)(I − Qn−1)xn = 0, there is a positive integer kn such that 2 (cid:16)δ − pδ2 − ε/2(cid:17)o. (I − Pkn )(I − Qn−1)xn < C2, where C2 = minn ε Define Qn = Pkn and 2E , 1 (4.5) zn = Qn(I − Qn−1)xn. The sequences xn and zn have the following properties: Firstly, by definition, there hold the inequalities (4.6) (4.7) (4.8) (I − Qn)(I − Qn−1)xn < C2, zn ≤ xn ≤ E, zn ≥ (I − Qn−1)xn − (I − Qn)(I − Qn−1)xn > C1 − C2. Secondly, (4.9) hzn, xni = hzn, zni . Indeed, since zn = Qn(I − Qn−1)xn, we have hzn, xni = hQn(I − Qn−1)xn, xni = = hQn(I − Qn−1)xn, (I − Qn−1)Qnxni = hzn, zni . Thirdly, for m > n we have (4.10) hzm, xni < C2E. Indeed, for such m and n we have Qn−1 ≤ Qn ≤ Qm−1, i.e. I − Qm−1 ≤ I − Qn ≤ I − Qn−1, implying I − Qm−1 = (I − Qm−1)(I − Qn)(I − Qn−1), and thus hzm, xni = h(I − Qm−1)zm, xni = = hzm, (I − Qm−1)(I − Qn)(I − Qn−1)xni = = hzm, (I − Qn)(I − Qn−1)xni . Therefore, by (4.6) and (4.8) hzm, xni ≤ zm · (I − Qn)(I − Qn−1)xn ≤ C2E. 12 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Let us construct a semi-norm p, continuous in τ , and a totally discrete sequence from E. Since by (4.8) zn2 = hzn, zni > (C1 − C2)2, we can choose a normal state ϕ and υj, νj ∈ A according to [7, Lemma 4.6.]), such that (4.11) ϕ(υ∗ n hzn, zni νn) > (C1 − C2)2. Consider the semi-norm p given by +∞ p(x) = vuut X j=1 ϕ(hzjυj, xi)2. By (4.5) there is a sequence ζj ∈ A such that Define ωj = ζjυn/ϕ(υ∗ Also, for x = (ξ1, ξ2, . . . ) we have zn = (0, . . . , 0, ζkn−1+1, . . . , ζkn , 0, . . . ). nζ∗ j ζjυn)1/2, for kn−1 + 1 ≤ j ≤ kn. Obviously ϕ(ω∗ j ωj) = 1. ϕ(hznυn, xi)2 = (cid:12)(cid:12)(cid:12) kn X j=kn−1+1 ϕ(υ∗ nζ∗ j ζjυn)1/2ϕ(ω∗ kn ≤ X j=kn−1+1 ϕ(υ∗ nζ∗ j ζjυn) kn X j=kn−1+1 ≤ 2 j ξj)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)ϕ(ω∗ 2 = j ξj)(cid:12)(cid:12) = ϕ(υ∗ n hzn, zni υn) kn X j=kn−1+1 2 . (cid:12)(cid:12)ϕ(ω∗ j ξj)(cid:12)(cid:12) Including (4.7) we obtain ϕ(υ∗ hence n hzn, zni υn) ≤ υ∗ n hzn, zni υn = zn2 ≤ E2 and p(x)2 = +∞ X n=1 ϕ(hznυn, xi)2 ≤ E2 +∞ X j=1 ϕ(ω∗ j ξj)2 = E2pϕ,ω1,...,ωn,...(x)2. Thus, we conclude that p is well defined and also that it is continuous with respect to τ . Also, E is A-balanced, so xnνn ∈ E. Finally we shall prove that xnνn is a totally discrete sequence. Indeed, for m > n we have p(xmνm − xnνn) ≥ ϕ(hzmυm, xmνm − xnνni) ≥ m hzm, xmi νm) − ϕ(υ∗ ≥ ϕ(υ∗ m hzm, xni νn) . However, by (4.9) and (4.11), ϕ(υ∗ m hzm, zmi νm) > (C1 − C2)2 and, by (4.10) Therefore and ϕ(υ∗ m hzm, xni νn) ≤ hzm, xni < C2E. p(xmνm − xnνn) > (C1 − C2)2 − C2E ≥ δ2 − ε pϕ,ω1,...,ωn,...(xmνm − xnνn) > δ2 − ε E . For ε ∈ (0, δ2), we have I ∗(E) ≥ δ2−ε λ(E) ≤ pEI ∗(E). E and hence I ∗(E) ≥ δ2 E = λ(E)2 E . Thus, (cid:3) MEASURES OF NONCOMPACTNESS 13 Remark 4.3. The preceding proof is adapted proof of our earlier result [7, Theorem 4.10]. Corollary 4.5. On the standard Hilbert module l2(B(H)) over the algebra B(H) of all bounded operators on a Hilbert space H, there holds χ∗(E) ≤ λ(E) ≤ pEI ∗(E) ≤ pEα∗(E) ≤ p2Eχ∗(E), for any balanced set, i.e. a set for which x ∈ E, u-unitary implies xu ∈ E. Finally, we discuss some relationship between the MNC of an arbitrary ad- jointable operator λ0 introduced in Definition 2.2, and corresponding MNCs derived from α∗, χ∗ and I ∗. Definition 4.2. Let A is an arbitrary W ∗-algebra and let T ∈ Ba(l2(A)) be an adjointable operator. The functions α∗ 0 : Ba(l2(A)) → [0, +∞) defined by 0, χ∗ 0, I ∗ α∗ 0(T ) = α∗(T (B1)), χ∗ 0(T ) = χ∗(T (B1)), I ∗ 0 (T ) = I ∗(T (B1)) are called, respectively, Kuratowski, Hausdorff and Istrăţescu measure of noncom- pactness of the operator T . Proposition 4.6. Let A be an arbitrary W ∗-algebra, let T , S ∈ Ba(l2(A)) and let µ stands for any of MNCs α, χ, I. Then 0, χ∗ (a) All α∗ 0 and I ∗ µ∗ 0(T + S) ≤ µ∗ (b) The functions α∗ 0(S), µ∗ 0(cT ) = cµ∗(T ), 0 are subadditive and positivel homogeneous, i.e. there holds 0(T ) + µ∗ 0, χ∗ 0 and I ∗ χ∗ 0(T ) ≤ I ∗ 0(T ) ≤ λ0(T ). 0 are equivalent to each other, that is, 0 (T ) ≤ α∗ 0(T ) ≤ 2χ∗ for all c > 0. 0(T ). Also, there holds χ∗ 0(T ), λ0(T ) ≤ T and α∗ 0(T ), I ∗ 0 (T ) ≤ 2T . (c) χ∗ (d) If T is "compact", i.e. T belongs to the closed linear space generated by x 7→ z hy, xi, then λ0(T ) = χ0(T ) = α0(T ) = I(T ) = 0. In general, the converse might not hold. 0(T + K) = µ∗ operators K. 0(T ), as well as λ0(T + K) = λ0(T ) for all "compact" (e) µ∗ Proof. Part (a) follows easily from Theorem 3.1-(d), whereas part (b) follows from (4.4) and Proposition 4.1. Since T (B1) ⊆ B(0; T ) = T B1, it follows λ0(T ) ≤ T according to Propo- sition 2.3. Other inequalities in part (c) follows from part (a). If T is "compact", then T is norm limit of finite rank operators. Hence T (B1) is A-precompact. Therefore λ0(T ) = 0. This, together with (b) proves (d). The converse does not always hold due to [7, Example 5.1.]. Finally, (e) follows from (d). (cid:3) In the case A = B(H), we can obtain more. Proposition 4.7. Let A = B(H) and let T ∈ Ba(l2(B(H))). Then (a) There holds 0(T ) ≤ λ0(T ) ≤ qT I ∗ χ∗ 0(T ) = 0, µ ∈ {α, χ, I} iff λ0(T ) = 0 iff T is a "compact" operator; 0(T ) ≤ q2T χ∗ 0 (T ) ≤ qT α∗ 0(T ). (b) µ∗ 14 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Proof. Part (a) follows from Proposition 4.6-(b) and Theorem 4.4, since T (B1) is a balanced set (y = T x ∈ T (B1) implies yu = T (xu) ∈ T (B1)). Part (b) follows from part (a) and Proposition 4.6-(d). Indeed, if any of four 0(T ) = 0, and hence, T (B1) is totally (cid:3) MNCs annihilate T , then, by part (a), α∗ bounded in the topology τ . By [7, Theorem 4.10.], T is "compact". 5. Three questions Question 5.1. It is easy to obtain the following: Let E ⊆ l2(A) be bounded. For any ε > 0 there is an A-precompact set Cε such that E ⊆ Cε + (λ(E) + ε)B. (For instance Cε = E ∩ M for a suitable free finitely generated M .) Is it possible to get something stronger: There is an A-precompact set C such that E ⊆ C + λ(E) · B? Question 5.2. Among all properties of MNCs on a Banach space, it turns out that the most important is µ(co E) = µ(E). This was proved in this note for λ if we co E 1 cjxj P cj = 1, cj ∈ R, xj ∈ E}. regard as a real field convex hull, i.e. co E = {Pn However, there is a notion of A-convex hull (see for instance [11]) coA E = n n X j=1 j xjaj xj ∈ E, aj ∈ A,X a∗ a∗ j aj = 1o. Is it possible to obtain λ(coA E) = λ(E)? Question 5.3. As it was mentioned in Remark 4.2, we ask for the following: Is it true µ∗(E) = µ∗(F ), where F denotes the balanced hull of E, i.e. F = Su Eu (the union runs through all unitaries u), and µ denotes any of Kuratowski, Hausdorff or Istrăţescu MNC? Acknowledgement. The authors was supported in part by the Ministry of edu- cation and science, Republic of Serbia, Grant #174034. References [1] R. R. Akhmerov, M. I. Kamenskii, A. S. Potapov, A. E. Rodkina, and B. N. Sadovskii. Measures of Noncompactness and Condensing Operators. Birkhäuser Verlag, Basel, Boston, Berlin, 1992. Translated from Russian. [2] Ivan D. Arandjelović. Measure of noncompactness on uniform spaces. Math. Morav., 2:1–8, 1998. [3] Ivan D. Arandjelović. Some properties of Istratescu's measure of noncompactness. Filomat, 13:99–104, 1999. [4] J. M. Ayerbe Toledano, T. Dominguez Benavides, and G. Lopez Acedo. Measures of Non- compactness in Metric Fixed Point Theory. Springer Basel AG, 1997. [5] N. Bourbaki. General Topology, chapters 1–4. Springer, 1987. [6] Vasile Ion Istrăţescu. On a measure of noncompactness. Bull. Math. Soc. Sci. Math. R.S. Roumanie (N.S), 16:195–197, 1972. [7] Dragoljub J. Kečkić and Zlatko Lazović. Compact and "compact" operators on the standard Hilbert module over a W ∗ algebra. Ann. Funct. Anal., to appear:13 pages. [8] C. Kuratowsky. Sur les espaces complets. Fund. Math., 15:301–309, 1930. [9] C. Kuratowsky. Topologie I. Monografje Matematyczne, Warszaw, 4th edition, 1958. [10] K. L. Lance. Hilbert C ∗-Modules: A toolkit for operator algebraists. Cambridge University Press, 1995. [11] Bojan Magajna. C ∗-convex sets and completely positive maps. Integral Eq. Op. Th., 85:37– 62, 2016. MEASURES OF NONCOMPACTNESS 15 [12] V. M. Manuilov. Diagonalization of compact operators in Hilbert modules over finite W ∗- algebras. Ann. Global Anal. Geom., 13(3):207–226, 1995. [13] V. M. Manuilov. Diagonalization of compact operators on Hilbert modules over C ∗-algebras of real rank zero. Math. Notes, 62(6):726–730, 1997. [14] V. M. Manuilov and E. V. Troitsky. Hilbert C ∗-modules. Translations of mathematical mono- graphs Vol. 226. AMS, Providence, Rhode Island, 2005. [15] Marina M. Milovanović-Arandjelović. Measures of noncompactness on uniform spaces–the axiomatic approach. Filomat, 15:221–225, 2001. [16] Mohammad Mursaleen, Syed M. H. Rizvi, and Bessem Samet. Measures of noncompactness and their applications. In J. Banas, M. Jleli, M. Mursaleen, B. Samet, and Vetro C., editors, Advances in Nonlinear Analysis via the Concept of Measure of Noncompactness, chapter 2, pages XIII, 487. Springer, 2017. [17] Vladimir Rakočević. Measures of noncompactness and some applications. Filomat, 12. [18] A. I. Sadovskii. Limit-compact and condensing operators. Russian Math. Surveys, 27(1):85– 155, 1972. University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected] University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected]
1806.10110	1	1806	2018-06-26T17:08:20	Infinite characters on $GL_n(\mathbf{Q})$, on $SL_n(\mathbf{Z}),$ and on groups acting on trees	[ "math.OA", "math.GR" ]	Answering a question of J. Rosenberg, we construct the first examples of infinite characters on $GL_n(\mathbf{K})$ for a global field $\mathbf{K}$ and $n\geq 2.$ The case $n=2$ is deduced from the following more general result. Let $G$ a non amenable countable subgroup acting on locally finite tree $X$. Assume either that the stabilizer in $G$ of every vertex of $X$ is finite or that the closure of the image of $G$ in ${\rm Aut}(X)$ is not amenable. We show that $G$ has uncountably many infinite dimensional irreducible unitary representations $(\pi, \mathcal{H})$ of $G$ which are traceable, that is, such that the $C^*$-subalgebra of $\mathcal{B}(\mathcal{H})$ generated by $\pi(G)$ contains the algebra of the compact operators on $\mathcal{H}.$ In the case $n\geq 3,$ we prove the existence of infinitely many characters for $G=SL_n(R)$, where $n\geq 3$ and $R$ is an integral domain such that $G$ is not amenable. In particular, the group $SL_n(\mathbf{Z})$ has infinitely many such characters for $n\geq 2.$	math.OA	math	INFINITE CHARACTERS ON GLn(Q), ON SLn(Z), AND ON GROUPS ACTING ON TREES BACHIR BEKKA Abstract. Answering a question of J. Rosenberg from [Ros -- 89], we construct the first examples of infinite characters on GLn(K) for a global field K and n ≥ 2. The case n = 2 is deduced from the following more general result. Let G a non amenable countable subgroup acting on locally finite tree X. Assume either that the stabilizer in G of every vertex of X is finite or that the closure of the image of G in Aut(X) is not amenable. We show that G has uncountably many infinite dimensional irreducible unitary representations (π, H) of G which are traceable, that is, such that the C ∗-subalgebra of B(H) generated by π(G) contains the algebra of the compact operators on H. In the case n ≥ 3, we prove the existence of infinitely many characters for G = GLn(R), where n ≥ 3 and R is an integral domain such that G is not amenable. In particular, the group SLn(Z) has infinitely many such characters for n ≥ 2. 1. Introduction is, the set of equivalence classes of irreducible unitary representations standard Borel space exactly when G is virtually abelian, by results of Glimm and Thoma (see [Gli -- 61] and [Tho -- 68]). So, unless G is virtually abelian (in which case the representation theory of G is well Let G be a countable discrete group and bG the unitary dual of G, that of G. The space bG, equipped with a natural Borel structure, is a understood), a description of bG is hopeless or useless. There are at than bG: least two other dual objects of G, which seem to be more accessible • Thoma's dual space E(G), that is, the set of indecomposable positive definite central functions on G; • the space Char(G) of characters of G, that is, the space of lower semi-continuous semi-finite (not necessarily finite) traces t on the maximal C ∗-algebra C ∗(G) of G (see Subsection 2.1) The author acknowledges the support by the ANR (French Agence Nationale de la Recherche) through the projects Labex Lebesgue (ANR-11-LABX-0020-01) and GAMME (ANR-14-CE25-0004). 1 2 BACHIR BEKKA which satisfies the following extremality condition: every lower semi-continuous semi-finite trace on C ∗(G) dominated by t on the ideal of definition of t is proportional to t. The space Char(G) parametrizes the quasi-equivalence classes of fac- torial representations of C ∗(G) which are traceable ; recall that a unitary representation π is factorial if the von Neumann algebra M generated by π(G) is a factor and that a factorial representation π is traceable if there exists a faithful normal (not necessarily finite) trace τ on M and a positive element x ∈ C ∗(G) such that 0 < τ (π(x)) < +∞. If this is the case, then t = τ ◦ π belongs to Char(G). Conversely, every element of Char(G) is obtained in this way. Traceable representations are also called normal representations. Two traceable factorial representations π1 and π2 are quasi-equivalent if there exists an isomorphism Φ : M1 → M2 such that Φ(π1(g)) = π2(g) for all g ∈ G, where Mi is the factor generated by πi(G). Observe that an irreducible unitary representation (π, H) of G is traceable if and only if π(C ∗(G)) contains the algebra of compact op- erators on H. The character associated to such a representation is given by the usual trace on B(H) and so does not belong to E(G) whenever H is infinite dimensional; in this case, the character is said to be of type I∞, in accordance with the type classification of von Neumann algebras. Observe also that two irreducible traceable representations of a group G are quasi-equivalent if and only if they are unitarily equivalent. Thoma's dual space E(G) is a subspace of Char(G) and classifies the quasi-equivalence classes of the factorial representations π of C ∗(G) for which the factor M generated by π(G) is finite, that is, such that the trace τ on M takes only finite values (for more detail on all of this, see Chapters 6 and 17 in [Dix -- 77]). Thoma's dual space E(G) was determined for several examples of countable groups G, among them G = GLn(K) or G = SLn(K) for an infinite field K and n ≥ 2 ([Kiri -- 65]; see also [PeT -- 16]), and G = SLn(Z) for n ≥ 3 ([Bek -- 07]); a procedure is given in [How -- 77, Proposition 3] to compute E(G) when G is a nilpotent finitely gener- ated group. The space Char(G) has been described for some amenable groups G: • when G is nilpotent, we have E(G) = Char(G) (see [CaM -- 84, Theorem 2.1]); • the space Char(G) is determined in [Guic -- 63] for the Baumslag- Solitar group BS(1, 2) and in [VeK -- 91] for the infinite symmet- ric group; INFINITE CHARACTERS 3 • for G = GLn(K) and n ≥ 2, it is shown in [Ros -- 89] that E(G) = Char(G) in the case where K an algebraic extension of a finite field. (Observe that GLn(K) is amenable if and only if K an algebraic extension of a finite field; see Proposition 9 in [HoR -- 89] or Proposition 11 below.) J. Rosenberg asked in [Ros -- 89, Remark after Th´eor`eme 1] whether there exists an infinite character on G = GLn(K), that is, whether Char(G) 6= E(G), for a field K which is not an algebraic extension of a finite field. We will show below that the answer to this question is positive, by exhibiting as far we know the first examples of such characters. The case where n = 2 and K is a global field (see below) will be deduced from a general result concerning groups acting on trees, which we now state. Recall that a graph X is locally finite if every vertex on X has only finitely many neighbours. In this case, the group Aut(X) of automor- phisms of X, equipped with the topology of pointwise convergence, is a locally compact group for which the vertex stabilizers are compact. Concerning the notion of weakly equivalent representations, see Chap- ters 3 and 18 in [Dix -- 77] (see also Section 2.1). Theorem 1. Let X be a tree and G a countable subgroup acting on X. Assume that (a) either G is not amenable and the stabilizer in G of every vertex of X is finite, or (b) X is locally finite and the closure of the image of G in Aut(X) is not amenable. There exists an uncountable family (πt)t of irreducible unitary represen- tations of G with the following properties: πt is infinite dimensional, is traceable and is not weakly equivalent to πt′ for t′ 6= t. Recall that a global field is a finite extension of either the field Q of rational numbers or of the field Fp(T ) of rational functions in T over the finite field Fp (see Chapter III in [Wei -- 67]). Corollary 2. Let G be either (i) GL2(K) or SL2(K) for a global field K, or (ii) SL2(Z), or (iii) Fn, the free non abelian group over n ∈ {2, . . . , +∞} genera- tors. There exists an uncountable family (πt)t of unitary representations of G with the properties from Theorem 1; moreover, in case G = Fn, the representations πt are all faithful. 4 BACHIR BEKKA Turning to the case n ≥ 3, we prove a result for G = GLn(R) or G = SLn(R), valid for every integral domain R such that G is not amenable. Theorem 3. Let R be a countable unital commutative ring which is an integral domain; in case the characteristic of R is positive, assume that the field of fractions of R is not an algebraic extension of its prime field. For n ≥ 3, let G = GLn(R) or G = SLn(R). There exists an infinite dimensional irreducible unitary representation of G which is traceable. In the case where R is a field or the ring of integers, we can even pro- duce infinitely many non equivalent representations as in Theorem 3. Corollary 4. (i) For n ≥ 3, let G = GLn(K) for a countable field K which is not an algebraic extension of a finite field. There exists an uncountable family (πt)t of pairwise non equiv- alent infinite dimensional irreducible unitary representations of G which are traceable. Moreover, the representations πt all have a trivial central character, that is, the πt's are representations of PGLn(K). (ii) Let G = SLn(Z) for n ≥ 3. There exists an infinite family of pairwise non equivalent infinite dimensional irreducible unitary representations of G which are traceable. The methods of proofs of Theorem 1 and Theorem 3 are quite dif- ferent in nature: • the proof of Theorem 1 is based on properties of a remark- able family of unitary representations of groups acting on trees constructed in [JuV -- 84] and used to show their K-theoretic amenability, a notion which originated from [Cun -- 83] in the case of free groups; • the traceable representations we construct in Theorem 3 are induced representations from suitable subgroups. The case n ≥ 4 uses the existence of appropriate subgroups of GLn(R) with Kazhdan's Property (T). Remark 5. For a group G as in Theorem 1 or Theorem 1, our results show that the set Char(G) contains characters of type I∞. For, say, G = GLn(Q), we do not know whether Char(G) contains characters of type II∞, that is, characters for which the corresponding factorial representation generates a factor of type II∞. INFINITE CHARACTERS 5 This paper is organized as follows. In Section 2, we establish some preliminary facts which are necessary to the proofs of our results. Sec- tion 3 is devoted to the proofs of Theorem 1 and Corollary 2; Theorem 3 and Corollary 4 are proved in Section 4. 2. Some preliminary results . 2.1. C ∗-algebras. Let G be a countable group. Recall that a unitary representation of G is a homomorphism π : G → U(H) from G to the unitary group of a complex separable Hilbert space H. From now on, we will simply write representation of G instead of "unitary repre- sentation of G". Every representation (π, H) of G extends naturally to a ∗-representation, denoted again by π, of the group algebra C[G] by bounded operators on H. Recall that the maximal C ∗-algebra C ∗(G) of G is the completion of C[G] of G with respect to the norm f 7→ sup kπ(f )k, π∈Rep(G) where Rep(G) denotes the set of representations (π, H) of G in a sep- arable Hilbert space H. We can view G as subset of C[G] and hence as a subset of C ∗(G). The C ∗-algebra C ∗(G) has the following universal property: every rep- resentation (π, H) of G extends to a unique representation (that is, ∗- homomorphism) π : C ∗(G) → B(H). The correspondence G → C ∗(G) is functorial: every homomorphism ϕ : G1 → G2 between two count- able groups G1 and G2 extends to a unique morphism ϕ∗ : C ∗(G1) → C ∗(G2) of C ∗-algebras. In particular, given a subgroup H of a group G, the injection map i : H → G extends to a morphism i∗ : C ∗(H) → C ∗(G); the map i∗ is injective and so C ∗(H) can be viewed naturally as a subalgebra of C ∗(G): indeed, this follows from the fact that every representation σ of H occurs as subrepresentation of the restriction to H of some representation π of G (one may take as π the induced representation IndG H σ, as shown below in Proposition 9). The following simple lemma will be one of our tools in order to show that π(C ∗(G)) contains a non-zero compact operator for a representa- tion π of G. 6 BACHIR BEKKA Let A be a C ∗-algebra. Recall that a representation π : A → B(H) weakly contains another representation ρ : A → B(K) if kρ(a)k ≤ kπ(a)k for all a ∈ A, or, equivalently, ker π ⊂ ker ρ (see Chapter 3 in [Dix -- 77]). Two repre- sentations π and ρ are weakly equivalent if π weakly contains ρ and ρ weakly contains π, that is, if ker π = ker ρ. Lemma 6. Let A be a C ∗-algebra and π : A → B(H) a representation of A. Assume that H contains a non-zero finite dimensional π(A)- invariant subspace K and that the restriction π1 of π to K is not weakly contained in the restriction π0 of π to the orthogonal complement K⊥. Then π(A) contains a non-zero compact operator. Proof. The ideal ker π0 is not contained in ker π1, since π1 is not weakly contained in π0. Hence, there exists a ∈ A with π0(a) = 0 and π1(a) 6= 0. Then π(a) = π1(a) has a finite dimensional range and is non-zero. (cid:3) Knowing that a representation of A contains in its image a non- zero compact operator, the following lemma enables us to construct an irreducible representation of A with the same property. Lemma 7. Let A be a C ∗-algebra and π : A → B(H) a representation of A in a separable Hilbert space H Let a ∈ A be such that π(a) is a non-zero compact operator. Then there exists an irreducible subrepre- sentation σ of π such that σ(a) is a compact operator and such that kσ(a)k = kπ(a)k. Ω πωdµ(ω) of irre- ducible representations πω; thus, we can find a probabilility measure µ on a standard Borel space Ω, a measurable field ω → πω of irreducible representations of A in a measurable field ω → Hω of separable Hilbert Ω Hωdµ(ω) Proof. We can decompose π as a direct integral R ⊕ spaces on Ω, and a Hilbert space isomorphism U : H →R ⊕ such that Uπ(x)U −1 =Z ⊕ Ω πω(x)dµ(ω). for all x ∈ A R ⊕ (see [Dix -- 77, §8.5]). Without loss of generality, we will identify π with Ω πωdµ(ω). Let a ∈ A be such that π(a) is a non-zero compact operator. Since kσ(a∗a)k = kσ(a)k2 for every representation σ of A, upon replacing a by a∗a, we can assume that a is a positive element of A. So π(a) is a positive selfadjoint compact operator on H with π(a) 6= 0. INFINITE CHARACTERS 7 There exists an orthonormal basis (Fn)n≥1 of H =R ⊕ Ω Hωdµ(ω) con- sisting of eigenvectors of π(a), with corresponding eigenvalues (λn)n≥1, counted with multiplicities. For every ω ∈ Ω and every n ≥ 1, we have (∗) πω(a)(Fn(ω)) = λnFn(ω). Let n0 ≥ 1 be such that λn0 = max{λn n ≥ 1}. Then kπ(a)k = λn0. Set Ω0 = {ω ∈ Ω Fn0(ω) 6= 0.}. Since Fn0 6= 0, we have µ(Ω0) > 0. We claim that Ω0 is a finite subset of Ω consisting of atoms of µ. Indeed, assume by contradiction that is not the case. Then there exists an infinite sequence (Ak)k of pairwise disjoint Borel subsets of Ω0 with µ(Ak) > 0. Observe that 1AkFn0 is a non-zero vector in H and that h1AkFn0, 1AlFn0i = 0 for every k 6= l. Moreover, we have π(a)(1AkFn0) =Z ⊕ = λn0Z ⊕ Ak Ak = λn01AkFn0. πω(a)(Fn0(ω))dµ(ω) Fn0(ω)dµ(ω) Since π(a) is a compact operator and λn0 6= 0, this is a contradiction. Let ω0 ∈ Ω0 be such that µ({ω0}) > 0. We claim that the linear span of {Fn(ω0) n ≥ 1} is dense in Hω0. Indeed, let v ∈ Hω0 be such that hv Fn(ω0)i = 0 for all n ≥ 1. Let F = 1ω0 ⊗ v ∈ H be defined by F (ω0) = v and F (ω) = 0 for ω 6= ω0. Then hF Fni = 0 for all n ≥ 1. Hence, F = 0, that is, v = 0, since (Fn)n≥1 is a basis of H. By (∗), Fn(ω0) is an eigenvector of πω0(a) with eigenvalue λn for every n ≥ 1 such that Fn(ω0) 6= 0. Since {Fn(ω0) n ≥ 1} is a total subset of Hω0, it follows that there exists a basis of Hω0 consisting of eigenvectors of πω0(a). As lim n→∞ λn = 0 (in case the sequence (λn)n≥1 is infinite), it follows that πω0(a) is a compact operator on Hω0. Moreover, we have kπω0(a)k = max{λn n ≥ 1} = λn0 = kπ(a)k. Finally, an equivalence between πω0 and a subrepresentation of π is provided by the unitary linear map Hω0 → H, v 7→ 1ω0 ⊗ v. 8 BACHIR BEKKA (cid:3) 2.2. Induced representations of groups. In the sequel, we will of- ten consider group representations which are induced representations. Let G be a countable group, H a subgroup of G and (σ, K) a represen- tation of H. Recall that the induced representation IndG H σ of G may be realized as follows. Let H be the Hilbert space of maps f : G → K with the following properties (i) f (hx) = σ(h)f (x) for all x ∈ G, h ∈ H; (ii) Px∈H\G kf (x)k2 < ∞. (Observe that kf (x)k only depends on the coset of x in H\G.) The induced representation π = IndG H σ is given on H by right trans- lation: (π(g)f )(x) = f (xg) for all g ∈ G, f ∈ H and x ∈ G. Recall that the commensurator of H in G is the subgroup, denoted by CommG(H), of the elements g ∈ G such that gHg−1 ∩ H is of finite index in both H and g−1Hg. The following result appeared in [Mac -- 51] in the case where σ is of dimension 1 and was extended to its present form in [Kle -- 61] and [Cor -- 75]. Theorem 8. Let G be a countable group and H a subgroup of G such that CommG(H) = H. (i) For every finite dimensional irreducible representation σ of H, the induced representation IndG H σ is irreducible. (ii) Let σ1 and σ2 be non equivalent finite dimensional irreducible H σ2 representations of H. The representations IndG are non equivalent. H σ1 and IndG We will need to decompose the restriction to a subgroup of an in- duced representation IndG H σ as in Theorem 8. For g ∈ G, we denote by σg the representation of g−1Hg defined by σg(x) = σ(gxg−1) for x ∈ g−1Hg. For the convenience of the reader, we give a short and elemen- tary proof of the following special case of the far more general result [Mac -- 52, Theorem 12.1]. Proposition 9. Let G be a countable group, H, L subgroups of G and (σ, K) a representation of H. Let S be a system of representatives for the double coset space H\G/L. The restriction πL to L of the induced INFINITE CHARACTERS 9 representation π = IndG H σ is equivalent to the direct sum IndL s−1Hs∩L(σss−1Hs∩L) Ms∈S Proof. Let H be the Hilbert space of π, as described above. For every s ∈ S, let Hs be the space of maps f ∈ H such that f = 0 outside the double coset HsL. We have an orthogonal L-invariant decomposition H =Ms∈S Hs. Fix s ∈ S. The Hilbert space H′ s of IndL s−1Hs∩L(σss−1Hs∩L) consists • f (tx) = σ(sts−1)f (x) for all t ∈ s−1Hs ∩ L, x ∈ L; of the maps f : L → K such that • Px∈s−1Hs∩L\L kf (x)k2 < ∞. Define a linear map U : Hs → H′ Uf (x) = f (sx) f ∈ Hs, x ∈ L. s by for all Observe that, for t ∈ s−1Hs ∩ L, x ∈ L and f ∈ Hs, we have Uf (tx) = f (stx) = f ((sts−1)sx) = σ(sts−1)f (sx) = σs(t)Uf (x) and that Xx∈s−1Hs∩L\L kUf (x)k2 = Xx∈s−1Hs∩L\L kf (sx)k2 = Xy∈H\G kf (y)k2 < ∞, so that Uf ∈ H′ map U is invertible, with inverse given by s and U is an isometry. It is easy to check that the U −1f (y) =(σ(h)f (x) 0 if y = hsx ∈ HsL otherwise , for f ∈ H′ IndL s. Moreover, U intertwines the restriction of πL to Hs and s−1Hs∩L(σss−1Hs∩L) : for g, x ∈ L and f ∈ H′ s, we have (cid:0)Uπ(g)U −1f(cid:1) (x) =(cid:0)π(g)U −1f(cid:1) (sx) = (U −1f (sxg) = f (xg) =(cid:0)IndL s−1Hs∩L(σss−1Hs∩L)(g)f(cid:1) (x) (cid:3) We will need the following elementary lemma about induced repre- sentations containing a finite dimensional representation. Recall that a representation π of a group G contains another representation σ of G if σ is equivalent to a subrepresentation of π. Recall also that, if π is 10 BACHIR BEKKA finite dimensional representation of a group G, then π ⊗ ¯π contains the trivial representation 1G, where ¯π is the conjugate representation of π and π ⊗ ρ denotes the (inner) tensor product of the representations π and ρ (see [BHV -- 08, Proposition A. 1.12]). Proposition 10. Let G be a countable group, H a subgroup of G, and σ a representation of H. Assume that the induced representation IndG H σ contains a finite dimensional representation of G. Then H has finite index in G. Proof. By assumption, π := IndG resentation σ. Hence, π ⊗ ¯π contains 1G. On the other hand, H σ contains a finite dimensional rep- π ⊗ ¯π = (IndG H σ) ⊗ ¯π is equivalent to IndG H(ρ), where ρ = σ⊗(¯πH); see [BHV -- 08, Proposition E. 2.5]. So, there exists a non-zero map f : G → K in the Hilbert space of IndG H(ρ) which is G-invariant, that is, such that f (xg) = f (x) for all g, x ∈ G. This implies that the L2-function x 7→ kf (x)k2 is constant on H\G. This is only possible if H\G is finite. (cid:3) 2.3. Amenability. Let G be a topological group and UCB(G) the Banach space of the left uniformly continuous bounded functions on G, equipped with the uniform norm. Recall that G is amenable if there exists a G-invariant mean on UCB(G) (see Appendix G in [BHV -- 08]). The following proposition characterizes the integral domains R for which GLn(R) or SLn(R) is amenable; the proof is an easy extension of the proof given in Proposition 9 in [HoR -- 89] for the case where R is a field. Proposition 11. Let R be a countable unital commutative ring which is an integral domain. Let K be the field of fractions of R and G = GLn(R) or G = SLn(R) for an integer n ≥ 2. The following properties are equivalent: (i) G is not amenable. (ii) K is not an algebraic extension of a finite field. (iii) R contains Z if the characteristic of K is 0 or the polynomial ring Fp[T ] if the characteristic of K is p > 0. Proof. Assume that K is an algebraic extension of a finite field Fq. Then K = Sm Km for an increasing family of finite extensions Km of Fq; hence, GLn(K) = Sm GLn(Km) is the inductive limit of the finite and hence amenable groups GLn(Km); it follows that GLn(K) is amenable and therefore GLn(R) and SLn(R) are amenable. This shows that (i) implies (ii). INFINITE CHARACTERS 11 Assume that (ii) holds. If the characteristic of K is 0, then K con- tains Q and hence R contains Z. So, we can assume that the charac- teristic of K is p > 0. We claim that R contains an element which is not algebraic over the prime field Fp. Indeed, otherwise, every element in R is algebraic over Fp. As the set of elements in K which are alge- braic over Fp is a field, it would follow that the field fraction field K is algebraic over Fp. This contradiction shows that (ii) implies (iii). Assume that (iii) holds. Then SLn(R) contains a copy of SL2(Z) or a copy of SL2(Fp[T ]). It is well-known that both SL2(Z) and SL2(Fp[T ]) contain a subgroup which is isomorphic to the free group on two gen- erators. Therefore, G is not amenable and so (iii) implies (i). (cid:3) Let G be a locally compact group, with Haar measure m. Recall that the amenability of G is characterized by the Hulanicki-Reiter the- orem (see [BHV -- 08, Theorem G.3.2]): G is amenable if and only if the regular representation (λG, L2(G, m)) weakly contains the trivial representation 1G, where m is Haar measure on G; when this is the case, λG weakly contains every representation of G The following result shows the amenability of G can be detected by the restriction of λG to a dense subgroup; for a more general result, see [Guiv -- 80, Proposition 1] or [Bek -- 16, Theorem 5.5]. Proposition 12. Let G be a locally compact group and G a countable dense subgroup of G. Assume that the restriction to G of the regular representation λG of G weakly contains the trivial representation 1G. Then G is amenable. Proof. By assumption, there exists a sequence (fn)n in L2(G, m) with kfnk = 1 such that lim n kλG(g)fn − fnk = 0 for all g ∈ G. Then, since fn(g−1x) − fn(x) ≤ fn(g−1x) − fn(x) for g, x ∈ G, we have (∗) kλG(g)fn − fnk = 0 for all g ∈ G. lim n Set ϕn :=pfn. Then ϕn ≥ 0 andRG ϕndm = 1. Every ϕn defines a mean Mn : f 7→RG f ϕndm on UCB(G). Let M be a limit of (Mn)n for the weak--topology on the dual space of UBC(G). It follows from (∗) that M is invariant under G. Since, for every f ∈ UCB(G), the map G → UCB(G), g 7→g f is continuous (where gf denotes left translation by g ∈ G), it follows that M is invariant under G. Hence, G is amenable. 12 BACHIR BEKKA 2.4. Special linear groups over a subring of a field. We will use the following elementary lemma about subgroups of SLn(K) which stabilize a line in Kn. Lemma 13. For an infinite field K and n ≥ 2, let L be a subgroup of SLn(K) which stabilizes a line ℓ in Kn. Then L ∩ SLn(R) has infinite index in SLn(R) for every infinite unital subring R of K. Proof. Let {v1, . . . , vn} be a basis of Kn with ℓ = Kv1. Fix i, j ∈ {1, . . . , n} with i 6= j and, for λ ∈ K, let Eij(λ) be the corresponding elementary matrix in SLn(K), that is, Eij(λ) = In + λ∆ij, where ∆ij denotes the matrix with 1 at the position (i, j) and 0 other- wise. For every l = 1, . . . , n, let ϕl : K → K be defined by Eij(λ)(v1) = nXi=1 ϕl(λ)vi for λ ∈ K. Every ϕl is a polynomial function (in fact, an affine function) on K and, for l = 2, . . . , n, we have ϕl(λ) = 0 for every λ ∈ K such that Eij(λ) ∈ L. Assume, by contradiction, that L∩SLn(R) has finite index in SLn(R) for an infinite subring R of K. Then the subgroup Li,j(R) := L ∩ {Eij(λ) λ ∈ R} has finite index in the subgroup {Eij(λ) λ ∈ R} of SLn(R). In partic- ular, Li,j(R) is infinite. It follows that ϕl has infinitely many roots in K and hence that ϕl = 0, for every l = 2, . . . , n. Therefore, every ele- mentary matrix Eij(λ) fixes the line ℓ, for i, j ∈ {1, . . . , n} and λ ∈ K. Since SLn(K) is generated by elementary matrices, it follows that every matrix in SLn(K) fixes the line ℓ; this of course is impossible. (cid:3) 3. Proofs of Theorem 1 and Corollary 2 3.1. Proof of Theorem 1. Let X be a tree, with X 0 the set of vertices and X 1 the set of edges of X. Let G be a locally compact group acting on X. Julg and Valette constructed in [JuV -- 84] (see also [Szw -- 91] and [Jul -- 15]) a remarkable family of representations (πt)t∈[0,1] of G, all de- fined on ℓ2(X 0), with the following properties: INFINITE CHARACTERS 13 (i) π0 is the natural representation of G on ℓ2(X 0) and π1 is equiv- alent to 1G ⊕ ρ1, where ρ1 is the natural representation of G on ℓ2(X 1); (ii) for every t ∈ [0, 1], there exists a bounded operator Tt on ℓ2(X 0) with inverse T −1 defined on the subspace of functions of X 0 with finite support such that πt(g) := T −1 t π0(g)Tt extends to a unitary operator on ℓ2(X 0) for every g ∈ G; so, a unitary representation g 7→ πt(g) of G is defined on ℓ2(X 0); t (iii) πt(g) − π0(g) is a finite-rank operator on ℓ2(X 0), for every t ∈ [0, 1] and g ∈ G; (iv) we have hπt(g)T −1 t δx T −1 t δyi = td(gx,y), for every t ∈ (0, 1), g ∈ G and x, y ∈ X 0, where d denotes the natural distance on X 0; (v) the map [0, 1] → R+, t 7→ kπt(g) − π0(g)k is continuous for every g ∈ G. (Our representation πt is g 7→ Uλρλ(g)U −1 representation ρλ and the operator Uλ appearing in §2 of [JuV -- 84].) t with λ = − log t, for the Let G be a countable group acting on X. Assume that (a) either G is not amenable and the stabilizer in G of every vertex of X is finite or (b) X is locally finite and the closure of the image of G in Aut(X) is not amenable. Set G = G in case (a) and let G be the closure of G in Aut(X) in case (b). Let (πt)t∈[0,1] be the family of representations of G as above. • First step. For every a ∈ C ∗(G) and every t ∈ [0, 1], the operator πt(a) − π0(a) is compact and the map [0, 1] → R+, t 7→ kπt(a) − π0(a)k is continuous. Indeed, this follows from Properties (iii) and (v) of the family (πt)t and from the fact that C[G] is dense in C ∗(G). • Second step. The restriction π0G of π0 to G does not weakly contain the trivial representation 1G. Indeed, the representation π0 of G is equivalent to the direct sum ⊕s∈T λG/Gs, where S is a system of representatives for the G-orbits in X 0 and Gs is the stabilizer in G of s ∈ S. Since Gs is compact (and even finite in case (a)) and hence amenable, λG/Gs = IndG Gs 1Gs 14 BACHIR BEKKA is weakly contained in the regular representation λG of G and so π0 is weakly contained in λG. Hence, π0 does not weakly contain the triv- ial representation 1G in case (a). In case (b), the claim follows from Proposition 12, since G is not amenable and G is dense in G. • Third step. There exists an element a ∈ C ∗(G) and 0 ≤ t0 < 1 with the following properties: πt0(a) = 0, πt(a) is a non zero compact operator for every t ∈ (t0, 1], and the map [t0, 1] → R+, t 7→ kπt(a)k is continuous. Indeed, by the second step, there exists a ∈ C ∗(G) such that π0(a) = 0 and 1G(a) 6= 0. Therefore, π1(a) 6= 0 and πt(a) = πt(a) − π0(a) for every t ∈ [0, 1] and so the claim follows from the first step. • Fourth step. Let a ∈ C ∗(G) and 0 ≤ t0 < 1 be as in the third step. There exists an irreducible infinite dimensional subrepresentation σt of πt such that σt(a) is a compact operator and such that kσt(a)k = kπt(a)k for every t ∈ (t0, 1). Indeed, it follows from the third step and Lemma 7 that πtG con- tains an irreducible subrepresentation σt such that σt(a) is a compact operator with kσt(a)k = kπt(a)k. It remains to show that σt is infinite dimensional for every t ∈ (t0, 1). Assume, by contradiction, σt is finite dimensional for some t ∈ (t0, 1). Since G is dense in G, the closed subspace Kt of ℓ2(X 0) defining σt is invariant under G and so σt is the restriction to G of a dimensional subrepresentation of πt, again denoted by σt. On the one hand, G acts properly on X 0, since the stabilizers of vertices are compact (and even finite in case (a)). So, we have lim g→+∞:g∈G d(gx, x) = 0 for all x ∈ X 0. It follows from Property (iv) of the family (πt)t that πt (and hence σt) is a C0-representation, that is, lim hπt(g)v wi = 0 g→+∞:g∈G for every v, w ∈ ℓ2(X 0). On the other hand, since σt is finite dimen- sional, σt ⊗σt contains 1G. As G is not compact, this is a contradiction to the fact that σt is a C0-representation. • Fifth step. There exists uncountably many real numbers t ∈ (t0, 1) such that the subrepresentations σt of πtG as in the fourth step are pairwise non weakly equivalent. INFINITE CHARACTERS 15 Indeed, by the third step, the function f : t 7→ kπt(a)k is continuous on [t0, 1], with f (t0) = 0 and f (1) > 0. So, the range of f contains a whole interval. Let t, s ∈ (t0, 1) be such that f (t) 6= f (s). Then kσt(a)k = kπt(a)k = f (t) 6= f (s) = kπs(a)k = kσs(a)k, and so σt and σs are not weakly equivalent. 3.2. Proof of Corollary 2. The following remarks show how Corol- lary 2 follows from Theorem 1. (i) Let K be global field K. Choose a non trivial discrete valuation v : K∗ → Z. The completion of K at v is a non archimedean local field Kv. The tree Xv associated to v (see Chapter II in [Ser -- 80]) is a locally finite regular graph. The group G = GL2(K) acts as a group of automorphisms of Xv, with vertex stabilizers conjugate to GL2(Ov∩K), where Ov is the compact subring of the integers in Kv. The closure of the image of G in Aut(Xv) coincides with PGL2(Kv) and is therefore non amenable. A similar remark applies to G = SL2(K). (ii) As is well-known, the group G = SL2(Z) is an amalgamated product Z/4Z∗Z/2ZZ/6Z. It follows that G acts on a tree with vertices of valence 2 or 3 with vertex stabilizers of order 4 or 6 (see Chapter I, Examples 4.2. in [Ser -- 80]) (iii) The free non abelian group F2 acts freely on its Cayley graph X, which is a 4-regular tree. It follows that Fn acts freely on X for every n ∈ {2, . . . , +∞}. Observe that, in this case, the representations πt and σt as in the proof of Theorem 1 are faithful for t 6= 1 (since there are even C0-representations). (cid:3) 4. Proofs of Theorem 3 and Corollary 4 4.1. Proof of Theorem 3. Let R be a countable unital commutative ring which is an integral domain and K its field of fractions. In case the characteristic of K is positive, assume that K is not an algebraic extension of its prime field. Let n ≥ 3 and G = GLn(R). We consider the natural action of G on the projective space P(Kn). Let ℓ0 = Ke1 ∈ P(Kn) be the line defined by the first unit vector e1 in Kn. The stabilizer of ℓ0 in G is H =(cid:18)R× 0 GLn−1(R)(cid:19) . Rn−1 16 BACHIR BEKKA Let σ be a finite dimensional representation of H and π := IndG H σ. We claim that π is irreducible and that π(C ∗(G)) contains a non zero com- pact operator. For the proof of this claim, we have to treat separately the cases n = 3 and n ≥ 4. 4.1.1. Case n = 3. • First step. We claim that gHg−1∩H is amenable, for every g ∈ G \ H. Indeed, let g ∈ G \ H. Then ℓ0 and gℓ0 are distinct lines in Kn and are both stabilized by gHg−1 ∩ H. Hence, gHg−1 ∩ H is isomorphic to a subgroup of the solvable group  K∗ 0 K 0 K∗ K 0 0 K∗ and is therefore amenable. • Second step. We claim that the representation π is irreducible. Indeed, in view of Theorem 8, we have to show that CommG(H) = H. Let g ∈ G \ H. On the one hand, gHg−1 ∩ H is amenable, by the first step. On the other hand, H is non amenable, by Proposition 11. This implies that gHg−1 ∩ H is not of finite index in H and so g is not in the commensurator of H in G. • Third step. We claim that the C ∗-algebra π(C ∗(G)) contains a non-zero compact operator. Indeed, let S be a system of representatives for the double cosets space H\G/H with e ∈ S. By Proposition 9, the restriction πH of π to H is equivalent to the direct sum Ms∈S IndH s−1Hs∩H(σss−1Hs∩H) = σ ⊕ Ms∈S\{e} IndH s−1Hs∩H(σss−1Hs∩H) Let s ∈ S \ {e}. By the first step, s−1Hs ∩ H is amenable and hence σss−1Hs∩H is weakly contained in the regular representation λs−1Hs∩H of s−1Hs ∩ H, by the Hulanicki-Reiter theorem. By continuity of induc- tion (see [BHV -- 08, Theorem F.3.5]), it follows that IndH is weakly contained in the regular representation λH of H. Therefore, s−1Hs∩H(σss−1Hs∩H) π0 := Ms∈S\{e} IndH s−1Hs∩H(σss−1Hs∩H) is weakly contained in λH. It follows that π0 does not weakly contain σ; indeed, assume by contradiction that σ is weakly contained in π0. Then λH ⊗ λH , which is a multiple of λH , weakly contains σ ⊗ σ. However, since σ is finite dimensional, σ ⊗ σ contains 1H. Hence, 1H is weakly INFINITE CHARACTERS 17 contained in λH and this is a contradiction to the non amenability of H. It follows from Lemma 6 that π(C ∗(H)) contains a non-zero compact operator. Since C ∗(H) can be viewed a subalgebra of C ∗(G), the claim is proved for G = GL3(R). 4.1.2. Case n ≥ 4. For every unital subring R′ of R, set L(R′) :=(cid:18)1 0 SLn−1(R′)(cid:19) , 0 which is a subgroup of H isomorphic to SLn−1(R′). • First step. Let g0 ∈ G \ H and R′ an infinite unital subring of R. We claim that g0Hg−1 0 ∩ L(R′) has infinite index in L(R′). Indeed, the group L := g0Hg−1 0 ∩ L(R′) stabilizes the two lines ℓ0 and g0ℓ0. Let V be the linear span of the n − 1 unit vectors e2, . . . , en. Denote by ℓ the projection on V of the line g0ℓ0, parallel to ℓ0. As gℓ0 6= ℓ0, we have ℓ 6= {0}. Moreover, L stabilizes ℓ, since L stabilizes ℓ0 and V . So, identifying L(R′) with the group SLn−1(R′), we see can view L as a subgroup of SLn−1(K) which stabilizes a line in Kn−1. Lemma 13 shows that L has infinite index in SLn−1(R′), as claimed. • Second step. We claim that the representation π is irreducible. In view of Theorem 8, it suffices to show that CommG(H) = H. Let g0 ∈ G \ H. By the first step, g0Hg−1 0 ∩ L(R) has infinite index 0 ∩ H has infinite index in H, since L(R) is a in L(R); hence, g0Hg−1 subgroup of H. • Third step. We claim that π(C ∗(G)) contains a non-zero compact operator. Indeed, since K is not an algebraic extension over its prime field, R contains a subring R′ which is a copy Z or a copy of the polynomial ring Fp[T ], by Proposition 11. The corresponding subgroup L := L(R′) of G is isomorphic to SLn−1(Z) or SLn−1(Fp[T ]). Observe that L is a lattice in the locally group G = SLn−1(R) or G = SLn−1(Fp((T −1))), where Fp((T −1)) is the local field of Laurent series over Fp. Since n − 1 ≥ 3, the group G and hence L has Kazhdan's Property (T); see [BHV -- 08, §. 1.4, 1.7]. Let S be a system of representatives for the double cosets space H\G/H with e ∈ S. By Proposition 9, the restriction πL to L of π is 18 BACHIR BEKKA equivalent to the direct sum σL ⊕ π0, where π0 := Ms∈S\{e} IndL s−1Hs∩L(σss−1Hs∩L). We claim that π0 does not weakly contain σL. Indeed, assume by con- tradiction that π0 weakly contains σL. Since σ is finite dimensional and L has Property (T), it follows that π0 contains σL (see [BHV -- 08, Theorem 1.2.5]). Therefore, IndL s−1Hs∩L(σss−1Hs∩L) contains a subrep- resentation of σL for some s ∈ S \ {e}. Hence, s−1Hs ∩ L has finite index in L, by Proposition 10. Since L = L(R′) for an infinite unital subring R′ of R, this is a contradiction to the first step. As in the proof for the case n = 3, we conclude that π(C ∗(G)) contains a non-zero compact operator. This proves Theorem 3 for G = GLn(R) when n ≥ 3. The case G = SLn(R) is proved in exactly the same way. 4.2. Proof of Corollary 4. For n ≥ 3, let G = GLn(R) for a ring R as above. The irreducible traceable representations of G constructed in the proof of Theorem 3 are of the form π = IndG H σ for a finite dimensional representation of the subgroup H = (cid:18)R× 0 GLn−1(R)(cid:19). Rn−1 Observe that, π = IndG contains Z. H σ is trivial on the center Z of G, since H By Theorem 8, there are infinitely (respectively, uncountably) many non equivalent such representations π, provided there exists infinitely (respectively, uncountably) non equivalent finite dimensional irreducible representations of H. This will be the case if GLn−1(R) ⋉ Rn−1, which is a quotient of H, has infinitely (respectively, uncountably) many non equivalent finite dimensional irreducible representations. (i) Assume that R = K is an infinite field. It is easy to show that the finite dimensional irreducible representations of GLn−1(K) ⋉ Kn−1 are all of the form (cid:18)∗ ∗ 0 A(cid:19) 7→ χ(det A), A ∈ GLn−1(K) for some χ in the unitary dual cK∗ of K∗; as K∗ is infinite, cK∗ is a compact infinite group and is therefore uncountable. (ii) Assume that R = Z. • Case n = 3. The free group F2 is a subgroup of finite index in GL2(Z). There exists uncountably many unitary characters (that is one-dimensional unitary representations) of F2. For every such unitary INFINITE CHARACTERS 19 i i F2 character χ, the representation IndGL2(Z) χ is finite dimensional and so has a decomposition ⊕iσ(χ) as a direct sum of finite dimensional irre- ducible representations σ(χ) of GL2(Z). One can choose uncountably many pairwise non equivalent representations among the σ(χ) 's and we obtain in this way uncountably many non equivalent finite dimensional irreducible representations of GL2(Z) and hence of GL2(Z) ⋉ Z2. • Case n ≥ 4. The group GLn−1(Z) ⋉Zn−1 has Kazhdan's property (T) and so has at most countably many non equivalent finite dimensional representations (see [Wan -- 75, Theorem 2.1]). There are indeed infin- itely many such representations: for every integer N ≥ 1, the finite group i GN = GLn−1(Z/N Z)) ⋉ (Z/N Z)n−1 is a quotient of GLn−1(Z)⋉Zn−1; infinitely many representations among the irreducible representations of the GN 's are pairwise non equivalent when viewed as representations of GLn−1(Z) ⋉ Zn−1. References [Bek -- 07] B. Bekka. Operator-algebraic superrigidity for SLn(Z), n ≥ 3. Invent. Math. 169 (2007), 401 -- 425. [Bek -- 16] B. Bekka. Spectral rigidity of group actions on homogeneous spaces. Prepint 2017, to appear in "Handbook of group actions, Volume III" (Editors: L. Ji, A. Papadopoulos, S-T Yau); ArXiv 1602.02892. [BHV -- 08] B. Bekka, P. de la Harpe, A. Valette. Kazhdan's Property (T). Cambridge University Press 2008. [CaM -- 84] A. Carey and W. Moran. Characters of nilpotent groups. Math. Proc. Camb. Phil. Soc. 96 (1984), no. 1, 123 -- 137. [Cor -- 75] L. Corwin. Induced representations of discrete groups. Proc. Amer. Math. Soc. 47 (1975), 279 -- 287. [Cun -- 83] J. Cuntz. K-theoretic amenability for discrete groups. J. Reine Angew. Math. 344 (1983), 180 -- 195 [Dix -- 77] J. Dixmier. C ∗-algebras. North-Holland 1977. [Gli -- 61] [Guic -- 63] A. Guichardet. Caract`eres des alg`ebres de Banach involutives. Ann. Inst. J. Glimm. Type I C-algebras. Ann. of Math. (2) 73 (1961), 572 -- 612. Fourier 13 (1963), 1 -- 81. [Guiv -- 80] Y. Guivarc'h. Quelques propri´et´es asymptotiques des produits de matri- ces al´eatoires. Lecture Notes Math. 774, 17-250, Springer, 1980. [How -- 77] R. Howe. On representations of discrete, finitely generated, torsion-free, nilpotent groups. Pacific J. Math. 73 (1977), no. 2, 281 -- 305. [HoR -- 89] R. Howe, J. Rosenberg. The unitary representation theory of GL(n) of an infinite discrete field. Israel J. Math. 67 (1989), 67 -- 81. [Jul -- 15] P. Julg. A new look at the proof of K-theoretic amenability for groups acting on trees. Bull. Belg. Math. Soc. Simon Stevin 22 (2015), 263 -- 269. [JuV -- 84] P. Julg, A. Valette. K-theoretic amenability for SL2(Qp), and the action on the associated tree. J. Funct. Anal. 58 (1984),194 -- 215. 20 BACHIR BEKKA [Kiri -- 65] A.A. Kirillov. Positive definite functions on a group of matrices with elements from a discrete field. Soviet. Math. Dokl. 6 (1965), 707 -- 709. [Kle -- 61] A. Kleppner. On the intertwining number theorem. Proc. Amer. Math. Soc. 12 (1961), 731 -- 733. [Mac -- 51] G.W. Mackey. On induced representations of groups. Amer. J. Math. 73 (1951), 576 -- 592. [Mac -- 52] G.W. Mackey. Induced representations of locally compact groups. I. Ann. of Math. (2) 55 (1952), 101 -- 139. [PeT -- 16] J. Peterson, A. Thom. Character rigidity for special linear groups. J. reine angew. Math. 716 (2016), 207 -- 228. [Ros -- 89] J. Rosenberg. Un compl´ement `a un th´eor`eme de Kirillov sur les caract`eres de GL(n) d'un corps infini discret. C.R. Acad. Sci. Paris 309 (1989), S´erie I, 581 -- 586. J-P. Serre. Trees. Springer-Verlag, Berlin-New York, 1980. [Ser -- 80] [Szw -- 91] R. Szwarc. Groups acting on trees and approximation properties of the Fourier algebra. J. Funct. Anal. 95 (1991), 320 -- 343. [VeK -- 91] A.M. Vershik, S.V. Kerov. Asymptotic theory of the characters of a sym- metric group. Functional Anal. Appl. 15 (1982), 246 -- 255. [Tho -- 68] E. Thoma. Eine Charakterisierung diskreter Gruppen vom Typ I. Invent. Math. 6 (1968), 190 -- 196. [Wan -- 75] S.P. Wang. On isolated points in the dual spaces of locally compact groups. Math. Ann. 218 (1975), 19 -- 34. [Wei -- 67] A. Weil. Basic number theory. Die Grundlehren der mathematischen Wissenschaften 144, Springer 1967. Bachir Bekka, Univ Rennes, CNRS, IRMAR -- UMR 6625, Campus Beaulieu, F-35042 Rennes Cedex, France E-mail address: [email protected]
1602.04760	1	1602	2016-02-15T18:44:21	Piecewise conjugacy for multivariable dynamics over the Jacobson spectrum of a C*-algebra	[ "math.OA" ]	We show that if (A,a) and (B,b) are automorphic multivariable C*-dynamical systems with isometrically isomorphic tensor algebras (or semi crossed products), then the systems are piecewise conjugate over their Jacobson spectrum. This answers a question of Kakariadis and the author.	math.OA	math	PIECEWISE CONJUGACY FOR MULTIVARIABLE DYNAMICS OVER THE JACOBSON SPECTRUM OF A C∗-ALGEBRA ELIAS G. KATSOULIS Abstract. We show that if (A, α) and (B, β) are automorphic multivariable C∗-dynamical systems with isometrically isomorphic tensor algebras (or semi crossed products), then the systems are piecewise conjugate over their Jacob- son spectrum. This answers a question of Kakariadis and the author. 1. Introduction The concept of a non-selfadjoint operator algebra associated with a multivari- able dynamical system is new and yet fruitfull. Such algebras appeared for the first time in the work of Power [8] and Donsig, Katavolos and Manoussos [5] but their systematic study only started recently with the Memoirs of Davidson and the author [2]. In that paper, given a multivariable dynamical system over a locally compact Hausdorff space, we isolated two associated operator algebras, the tensor algebra and the semi crossed product, and we made the case that various properties of the dynamical system are encoded in these algebras. Rather surprisingly, the classification of such algebras in [2] has had an impact beyond operator algebras, as witnessed in the work of Cornelissen and Marcolli [1] on class field theory. Inspired by [1, 2], Kakariadis and the author [7] extended the study of mul- tivariable dynamics beyond commutative C∗-algebras. It turns out that the non- commutative context allows for questions that do not materialize in the commuta- tive setting. One such question revolves around the various notions of a spectrum for a C∗-algebra. In [7] we showed that if (A, α) and (B, β) are multivariable dynamical systems with isometrically isomorphic tensor algebras (or semi crossed products), then the systems are piecewise conjugate over the spectrum, as described by Ernest in [6]. However the conjugacy over the Jacobson spectrum, i.e., the prim- itive ideal space with the hull-kernel topology, was left open and it was asked as a question at the end of the paper [7, Question 3]. In this note we observe that the presence of a continuous open map between the spectra, combined with the results of [7] settles this question in the affirmative. 2. The main result We adhere to the notation of [7] and use as references for the properties of the various spectra of C∗-algebras the book of Dixmier [4], and Ernest's paper [6]. If A is a C∗-algebra, then Prim(A) will denote its Jacobson spectrum. Let A be the collection of all (equivalence classes of non-trivial) irreducible representations 2000 Mathematics Subject Classification. 47L55, 47L40, 46L05, 37B20. Key words and phrases: C∗-algebra, multivariable dynamical system, piecewise conjugacy, spectrum. 1 2 of A and (1) E.G.KATSOULIS θ : A → Prim(A); A ∋ π 7→ ker π. The space A is equipped with the smallest topology that makes θ continuous. This forces θ to be an open mapping as well. In [7] we worked exclusively with J. Ernest's picture for the spectrum for a C∗- algebra. Let R(A) be the collection of all railway representations of A; these are representations equivalent to appropriately large ampliations of irreducible repre- sentations of A, all acting on the same Hilbert space. The space R(A) is equipped with the topology of pointwise-SOT convergence. If ϕ : R(A) → A is the map that associates a railway representation to the equivalence class of its associated irreducible representation, then [6] shows that ϕ is continuous and open. Lemma 2.1. Let X, Y, Z be topological spaces and f, g, h maps so that the following diagram f X ❅ ❅ ❅ ❅ g ❅ ❅ ❅ ❅ Y h Z commutes. Assume that f is continuous and open, Z is equipped with the quotient topology for g and h is a bijective surjection. Then h is a homeomorphism. Proof. Assume that U ⊆ Z is open. Since g is continuous, g−1(U ) = f −1(h−1(U )) is also open. Since f is open, we obtain f (g−1(U )) = f (cid:0)f −1(h−1(U ))(cid:1) = h−1(U ) is open and so h is continuous. Now let U ⊆ Y be open. Then, g−1(h(U )) = (cid:0)g−1(h−1)−1(cid:1) (U ) = (h−1g)−1(U ) = f −1(U ) is open. Since Z is equipped with the quotient topology for g, h(U ) is open, i.e., h is open. If we take X = R(A), Y = A, Z = R(A)/ ∼, i.e., unitary equivalence classes of railway representations, f = ϕ, g the quotient map and h the map that assigns an equivalence class of irreducible representations to the equivalence class of the corresponding railway representation, then Lemma 2.1 shows that the map h is a canonical homeomorphism between A and R(A)/ ∼. In the sequel, we will not be distinguishing between these two spectra. Let X and Y be topological spaces and let σ = (σ1, σ2, . . . , σn) and τ = (τ1, τ2, . . . , τn) be multivariable dynamical systems consisting of selfmaps of X and Y respectively. Davidson and Katsoulis [2, Definition 3.16] define (X, σ) and (Y, τ ) to be piecewise conjugate if there exists a homeomorphism h : X → Y and an open cover {Ug g ∈ Sn} of Y so that τi(y) = h σg(i) h−1(y), for each y ∈ Ug, g ∈ Sn and 1 ≤ i ≤ n. / / JACOBSON SPECTRUM 3 An (automorphic) multivariable C∗-dynamical system (A, α) consists of a C∗- algebra A and ∗-automorphisms α = (α1, α2, . . . , α) acting on A. Any automor- phism (resp. multivariable system) α of A induces a homeomorphism (resp. multi- variable dynamical system) α on A, that maps the equivalence class [ρ] of a railway representation to [ρα]. Similarly, there is a map α on Prim(A) that maps ker π to ker πα. Below we answer affirmatively [7, Question 3]. We will not be explaining the operator algebras appearing in the Theorem below as we do not require any of their defining properties. Instead we direct the reader to [7, Section 2]. Theorem 2.2. Let (A, α) and (B, β) be multivariable dynamical systems consisting of ∗-automorphisms. If the associated operator algebras alg(A, α) and alg(B, β) are isometrically isomorphic as operator algebras, then the multivariable dynamical systems (Prim(A), α) and (Prim(B), β) are piecewise conjugate. Proof. We begin by noticing that for any choice of C∗-algebras C, D and a ∗-automorphism δ : C → D we have a commutative diagram (2) C δ  y D θ−−−−→ Prim(C) δ  y θ−−−−→ Prim(D) where θ is defined in (1). In [7, Theorem 4.9] we proved that if alg(A, α) and alg(B, β) are isometrically isomorphic as operator algebras, then the dynamical systems ( A, α) and ( B, β) are piecewise conjugate. Furthermore, it follows from the proof of [7, Theorem 4.9] that the homeomorphism h implementing the piecewise conjugacy comes from a ∗-automorphism γ : A → B, i.e., h = γ. Let {Ug g ∈ Sn} be an open cover of B so that βi(y) = γ αg(i) γ−1(y), for each y ∈ Ug, g ∈ Sn and 1 ≤ i ≤ n. Since θ is open, {θ(Ug) g ∈ Sn} be an open cover of Prim(B). Furthermore, repeated use of (2) for the appropriate choices of C, D and δ implies that βiθ(y) = θ βi(y) = θγ αg(i) γ−1(y) = γθ αg(i)γ−1(y) = · · · = γ αg(i) γ−1θ(y). Hence, βi θ(Ug)= γ αg(i) γ−1 θ(Ug ) and we are done. References [1] G. Cornelissen and M. Marcolli, Quantum Statistical Mechanics, L-series and Anabelian Geometry, manuscript. [2] K. Davidson, E. Katsoulis, Operator algebras for multivariable dynamics, Mem. Amer. Math. Soc. 209 (2011), no. 982, viii+53 pp. [3] K. Davidson, E. Katsoulis, Isomorphisms between topological conjugacy algebras, J. reine angew. Math. 621 (2008), 29-51. [4] J. Dixmier, C∗-algebras, North-Holland Mathematical Library; v. 15, 1977. [5] A. Donsig, A. Katavolos and A. Manoussos, The Jacobson radical for analytic crossed prod- ucts J. Funct. Anal. 187 (2001), 129 -- 145. [6] J. Ernest, On the topology of the spectrum of a C∗-algebra, Math. Ann. 216 (1975), 149 -- 153. 4 E.G.KATSOULIS [7] E.T.A. Kakariadis and E. Katsoulis, Isomorphism invariants for multivariable C∗-dynamics. J. Noncommut. Geom. 8 (2014), 771 -- 787. [8] S.C. Power, Classification of analytic crossed product algebras, Bull. London Math. Soc. 24 (1992), 368 -- 372. Dept. Math., East Carolina University, Greenville, NC 27858, USA E-mail address: [email protected]
1601.01259	2	1601	2016-01-13T17:14:16	A "quantum" Ramsey theorem for operator systems	[ "math.OA", "math.CO", "math.FA", "math.RA", "quant-ph" ]	Let V be a linear subspace of M_n(C) which contains the identity matrix and is stable under the formation of Hermitian adjoints. We prove that if n is sufficiently large then there exists a rank k orthogonal projection P such that dim(PVP) = 1 or k^2.	math.OA	math	A "QUANTUM" RAMSEY THEOREM FOR OPERATOR SYSTEMS NIK WEAVER Abstract. Let V be a linear subspace of Mn(C) which contains the identity matrix and is stable under the formation of Hermitian adjoints. We prove that if n is sufficiently large then there exists a rank k orthogonal projection P such that dim(P V P ) = 1 or k2. An operator system in finite dimensions is a linear subspace V of Mn(C) with 1. Background the properties • In ∈ V • A ∈ V ⇒ A∗ ∈ V where In is the n × n identity matrix and A∗ is the Hermitian adjoint of A. In this paper the scalar field will be complex and we will write Mn = Mn(C). Operator systems play a role in the theory of quantum error correction. In clas- sical information theory, the "confusability graph" is a bookkeeping device which keeps track of possible ambiguity that can result when a message is transmitted through a noisy channel. It is defined by taking as vertices all possible source mes- sages, and placing an edge between two messages if they are sufficiently similar that data corruption could lead to them being indistinguishable on reception. Once the confusability graph is known, one is able to overcome the problem of information loss by using an independent subset of the confusability graph, which is known as a "code". If it is agreed that only code messages will be sent, then we can be sure that the intended message is recoverable. When information is stored in quantum mechanical systems, the problem of error correction changes radically. The basic theory of quantum error correction was laid down in [3]. In [2] it was suggested that in this setting the role of the confusability graph is played by an operator system, and it was shown that for every operator system a "quantum Lov´asz number" could be defined, in analogy to the classical Lov´asz number of a graph. This is an important parameter in classical information theory. See also [5] for much more along these lines. The interpretation of operator systems as "quantum graphs" was also proposed in [8], based on the more general idea of regarding linear subspaces of Mn as "quantum relations", and taking the conditions In ∈ V and A ∈ V ⇒ A∗ ∈ V to respectively express reflexivity and symmetry conditions. The idea is that the edge structure of a classical graph can be encoded in an obvious way as a reflexive, symmetric relation on a set. This point of view was explicitly connected to the quantum error correction literature in [9]. Date: Jan. 13, 2016. 1 2 NIK WEAVER Ramsey's theorem states that for any k there exists n such that every graph with at least n vertices contains either a k-clique or a k-anticlique, i.e., a set of k vertices among which either all edges are present or no edges are present. Simone Severini asked the author whether there is a "quantum" version of this theorem for operator systems. The natural notions of k-clique and k-anticlique are the following. Definition 1.1. Let V ⊆ Mn be an operator system. A quantum k-clique of V is an orthogonal projection P ∈ Mn (i.e., a matrix satisfying P = P 2 = P ∗) whose rank is k, such that P VP = {P AP : A ∈ V} is maximal; that is, such that P VP = P MnP ∼= Mk, or equivalently, dim(P VP ) = k2. A quantum k- anticlique of V is a rank k projection P such that P VP is minimal; that is, such that P VP = C · P ∼= M1, or equivalently, dim(P VP ) = 1. The definition of quantum k-anticlique is supported by the fact that in quantum error correction a code is taken to be the range of a projection satisfying just this condition, P VP = C · P [3]. As mentioned earlier, classical codes are taken to be independent sets, which is to say, anticliques. See also Section 4 of [9], where intuition for why P VP is correctly thought of as a "restriction" of V is given. The main result of this paper is a quantum Ramsey theorem which states that for every k there exists n such that every operator system in Mn has either a quantum k-clique or a quantum k-anticlique. This answers Severini's question positively. The quantum Ramsey theorem is not merely analogous to the classical Ramsey theorem; using the bimodule formalism of [8], we can formulate a common generalization of the two results. This will be done in the final section of the paper. I especially thank Michael Jury for stimulating discussions, and in particular for conjecturing Proposition 2.3 and improving Lemma 4.2. Part of this work was done at a workshop on Zero-error information, Operators, and Graphs at the Universitat Aut`onoma de Barcelona. 2. Examples If G = (V, E) is any finite simple graph, without loss of generality suppose V = {1, . . . , n} and define VG to be the operator system VG = span{Eij : i = j or {i, j} ∈ E} ⊆ Mn. Here we use the notation Eij for the n × n matrix with a 1 in the (i, j) entry and 0's elsewhere. Also, let (ei) be the standard basis of Cn, so that Eij = eie∗ j . The inclusion of the diagonal Eii matrices in VG corresponds to including a loop at each vertex in G. In the error correction setting this is natural: we place an edge between any two messages that might be indistinguishable on reception, and this is certainly true of any message and itself. Once we adopt the convention that every graph has a loop at each vertex, an anticlique should no longer be a subset S ⊆ V which contains no edges, it should be a subset which contains no edges except loops. Such a set corresponds to the projection PS onto span{ei : i ∈ S}, which has the property that PSVGPS = span{Eii : i ∈ S}. Or course this is very different from a quantum anticlique where P VP is one-dimensional. To illustrate the dissimilarity between classical and quantum cliques and anti- cliques, consider the diagonal operator system Dn ⊆ Mn consisting of the diagonal n × n complex matrices. In the notation used above, this is just the operator sys- tem VG corresponding to the empty graph on n vertices. It might at first appear to falsify the desired quantum Ramsey theorem, because of the following fact. A QUANTUM RAMSEY THEOREM 3 Proposition 2.1. Dn has no quantum k-anticlique for k ≥ 2. Proof. Let P ∈ Mn be a projection of rank k ≥ 2. Since rank(Eii) = 1 for all i, it follows that rank(P EiiP ) = 0 or 1 for each i. If P EiiP = 0 for all i then i=1 P EiiP = 0, contradiction. Thus we must have rank(P EiiP ) = 1 for some i, but then P EiiP cannot belong to C · P = {aP : a ∈ C}, since every matrix in this set has rank 0 or k. So P DnP 6= C · P . (cid:3) P = Pn Since every operator system of the form VG contains the diagonal matrices, none of these operator systems has nontrivial quantum anticliques. The surprising thing is that for n sufficiently large, they all have quantum k-cliques. This follows from the next result. Proposition 2.2. If n ≥ k2 + k − 1 then Dn has a quantum k-clique. Proof. Without loss of generality let n = k2 + k − 1. Start by considering Mk acting on Ck. Find k2 vectors v1, . . . , vk2 in Ck such that the rank 1 matrices viv∗ i are linearly independent. (For example, we could take the k standard basis vectors ei plus the k2−k vectors ei + iej for i 6= j. The corresponding rank 1 matrices span Mk and thus they must be independent since dim(Mk) = k2.) Making the identification Cn ∼= Ck ⊕ Ck2−1, we can extend the vi to orthogonal vectors wi ∈ Cn as follows: take w1 = v1 ⊕ (1, 0, . . . , 0), w2 = v2 ⊕ (a1, 1, 0, . . . , 0), w3 = v3 ⊕ (b1, b2, 1, 0, . . . , 0), etc., with a1, b1, b2, . . . successively chosen so that hwi, wj i = 0 for i 6= j. We need k2 − 1 extra dimensions to accomplish this. Now let P be the rank k projection of Cn onto Ck and let Dn be the diagonal operator system relative to any orthonormal basis of Cn that contains the vectors wi i for all i, so dim(P DnP ) = k2. (cid:3) kwik for 1 ≤ i ≤ k2. Then P DnP contains P wiw∗ vectors ei + ej for i 6= j plus the k2−k i P = viv∗ 2 2 A stronger version of this result will be proven in Lemma 4.3. The value n = k2 + k − 1 may not be optimal, but note that in order for Dn to have a quantum k-clique n must be at least k2, since dim(Dn) = n and we need dim(P DnP ) = k2. Next, we show that operator systems of arbitrarily large dimension may lack quantum 3-cliques. Proposition 2.3. Let Vn = span{In, E11, E12, . . . , E1n, E21, . . . , En1} ⊆ Mn. Then Vn has no quantum 3-cliques. Proof. Let P ∈ Mn be any projection. If P e1 = 0 then P E1iP = P Ei1P = 0 for all i, so P is a quantum anticlique. Otherwise let k = rank(P ) and let f1, . . . , fk be an orthonormal basis of ran(P ) with f1 = P e1 i P = f1v∗ i where vi = kP e1kP ei. The span of these matrices f1v∗ i }, i since the projections of the ei span ran(P ). Similarly, the span of the matrices 1 }. So P VnP is just Vk ⊆ Mk ∼= P MnP , relative to P Ei1P is precisely span{fif ∗ If k ≥ 3 then dim(Vk) = 2k < k2, so P cannot be a quantum the (fi) basis. clique. (cid:3) kP e1k . Then P E1iP = P e1e∗ is precisely span{f1f ∗ 3. Quantum 2-cliques In contrast to Proposition 2.3, we will show in this section that any operator system whose dimension is at least four must have a quantum 2-clique. This result is clearly sharp. It is somewhat analogous to the trivial classical fact that any graph that contains at least one edge must have a 2-clique. 4 NIK WEAVER Define the Hilbert-Schmidt inner product of A, B ∈ Mn to be Tr(AB∗). Denote the set of Hermitian n × n matrices by M h n . Observe that any operator system is spanned by its Hermitian part since any matrix A satisfies A = Re(A) + iIm(A) where Re(A) = 1 2 (A + A∗) and Im(A) = 1 2i (A − A∗). Lemma 3.1. Let V ⊆ Mn be an operator system and suppose dim(V) ≤ 3. Then its Hilbert-Schmidt orthocomplement is spanned by rank 2 Hermitian matrices. Proof. Work in M h matrices in V ⊥ complex spans then yields the desired result. n . Let V0 = V ∩M h 0 whose rank is 2. We will show that W0 = V ⊥ n and let W0 be the real span of the Hermitian n ); taking 0 (in M h Suppose to the contrary that there exists a nonzero Hermitian matrix B ∈ V ⊥ 0 which is orthogonal to W0. Say V0 = span{In, A1, A2}, where A1 and A2 are not necessarily distinct from In. Since B ∈ V ⊥ 0 , we have Tr(InB) = Tr(A1B) = Tr(A2B) = 0, but Tr(B2) 6= 0. We will show that there is a rank 2 Hermitian matrix C whose inner products against In, A1, A2, and B are the same as their inner products against B. This will be a matrix in W0 which is not orthogonal to B, a contradiction. Since B is Hermitian, we can choose an orthonormal basis (fi) of Cn with respect to which it is diagonal, say B = diag(b1, . . . , bn). We may assume b1, . . . , bj ≥ 0 and bj+1, . . . , bn < 0. Let B+ = diag(b1, . . . , bj, 0, . . . , 0) and B− = diag(0, . . . , 0, −bj+1, . . . , −bn) be the positive and negative parts of B, so that B = B+ − B−. Let α = Tr(B+) = Tr(B−) (they are equal since Tr(B) = Tr(InB) = 0). Then 1 α B+ is a convex combination of the rank 1 ma- α Tr(AB+) is a convex trices f1f ∗ combination of the linear functionals A 7→ hAfi, fii for 1 ≤ i ≤ j. By the convexity of the joint numerical range of three Hermitian matrices [1], there exists a unit vector v ∈ Cn such that 1 α Tr(AB+) = hAv, vi for A = A1, A2, and B. Similarly, there exists a unit vector w such that 1 α Tr(AB−) = hAw, wi for A = A1, A2, and B. Then C = α(vv∗ − ww∗) is a rank 2 Hermitian matrix whose inner products against In, A1, A2, and B are the same as their inner products against B. So C has the desired properties. (cid:3) j ; that is, the linear functional A 7→ 1 1 , . . ., fjf ∗ Lemma 3.2. Let V ⊆ M3 be an operator system and suppose dim(V) = 4. Then V has a quantum 2-clique. Proof. The proof is computational. Say V = span{A0, A1, A2, A3} where A0 = I3 and the other Ai are Hermitian. It will suffice to find two vectors v, w ∈ C3 such that the four vectors [hAiv, vi, hAiv, wi, hAiw, vi, hAiw, wi] ∈ C4 for 0 ≤ i ≤ 3 are independent. That is, we need the 4×4 matrix whose rows are these vectors to have nonzero determinant. Then letting P be the orthogonal projection onto span{v, w} will verify the lemma. We can simplify by putting the Ai in a special form. First, by choosing a basis of eigenvectors, we can assume A1 is diagonal. By subtracting a suitable multiple of A0 from A1, multiplying by a nonzero scalar, and possibly reordering the basis vectors, we can arrange that A1 has the form diag(0, 1, a). (Note that dim(V) = 4 implies that A1 cannot be a scalar multiple of A0.) These operations do not affect span{A0, A1, A2, A3}. Then, by subtracting suitable linear combinations of A0 and A QUANTUM RAMSEY THEOREM 5 b12 0 ¯b23 b13 b23 b33   . 1 0 β   , 0 ¯a12 ¯a13   a12 0 ¯a23 A1, we can arrange that A2 and A3 have the forms  0 ¯b12  ¯b13 w =   v =       a13 a23 a33 1 α 0 and and Let then evaluate the determinant of the 4 × 4 matrix described above and expand it as a polynomial in α, β, ¯α, and ¯β. We just need this determinant to be nonzero for some values of α and β; if this fails, then the polynomial coefficients must all be zero, and direct computation shows that this forces one of A2 and A3 to be a scalar multiple of the other. We omit the tedious but straightforward details. (cid:3) Theorem 3.3. Let V ⊆ Mn be an operator system and suppose dim(V) ≥ 4. Then V has a quantum 2-clique. Proof. Without loss of generality we can suppose that dim(V) = 4. Say V = span{In, A1, A2, A3} where the Ai are Hermitian and linearly independent. We first claim that there is a projection P of rank at most 3 such that P InP , P A1P , and P A2P are linearly independent. If A1 and A2 are jointly diagonalizable then we can find three common eigenvectors v1, v2, and v3 such that the vectors (1, 1, 1), (λ1, λ2, λ3), (µ1, µ2, µ3) ∈ C3 are linearly independent, where λi and µi are the eigenvalues belonging to vi for A1 and A2, respectively. Then the projection onto span{v1, v2, v3} verifies the claim. If A1 and A2 are not jointly diagonalizable, then we can find two eigenvectors v1 and v2 of A1 such that hA2v1, v2i 6= 0. Let- ting v3 be a third eigenvector of A1 with the property that the eigenvalues of A1 belonging to v1, v2, and v3 are not all equal, we can again use the projection onto span{v1, v2, v3}. This establishes the claim. Now let P be as in the claim and find B ∈ Mn such that P InP , P A1P , P A2P , and P BP are linearly independent. By Lemma 3.2 we can then find a rank 2 pro- jection Q ≤ P such that QInQ, QA1Q, QA2Q, and QBQ are linearly independent. If QInQ, QA1Q, QA2Q, and QA3Q are linearly independent then we are done. Otherwise, let α, β, and γ be the unique scalars such that QA3Q = αQInQ + βQA1Q + γQA2Q. By Lemma 3.1 we can find a rank 2 Hermitian matrix C such that Tr(InC) = Tr(A1C) = Tr(A2C) = 0 but Tr(A3C) 6= 0. Then C = vv∗ − ww∗ for some orthogonal vectors v and w. Thus, hAv, vi = hAw, wi for A = In, A1, and A2, but not for A = A3. It follows that the two conditions hA3v, vi = αhInv, vi + βhA1v, vi + γhA2v, vi and hA3w, wi = αhInw, wi + βhA1w, wi + γhA2w, wi cannot both hold. Without loss of generality suppose the first fails. Then letting Q′ be the projection onto span(ran(Q) ∪ {v}), we cannot have Q′A3Q′ = αQ′InQ′ + βQ′A1Q′ + γQ′A2Q′. Thus rank(Q′) = 3 and dim(Q′VQ′) = 4. The theorem now follows by applying Lemma 3.2 to Q′VQ′. (cid:3) Theorem 3.3 does not generalize to arbitrary four-dimensional subspaces of Mn. For instance, let V = span{E11, E12, E13, E14} ⊂ M4; by reasoning similar to that in the proof of Proposition 2.3, if P is any rank 2 projection in M4 then dim(P VP ) ≤ 2. 6 NIK WEAVER 4. The main theorem The proof of our main theorem proceeds through a series of lemmas. If dim(V) ≥ Lemma 4.1. Suppose the operator system V is contained in Dn. k2 + k − 1 then V has a quantum k-clique. If dim(V) ≤ n−k k−1 then V has a quantum k-anticlique. If n ≥ k3 − k + 1 then V has either a quantum k-clique or a quantum k-anticlique. Proof. If dim(V) ≥ k2 + k − 1 = m then we can find a set of indices S ⊆ {1, . . . , n} of cardinality m such that dim(P VP ) = m where P is the orthogonal projection onto span{ei : i ∈ S}. Then P VP ∼= Dm ⊆ Mm ∼= P MnP and Proposition 2.2 yields that P VP , and hence also V, has a quantum k-clique. If dim(V) ≤ n−k k−1 then a result of Tverberg [6, 7] can be used to extract a quantum k-anticlique; this is essentially Theorem 4 of [4]. Thus if k2 + k − 1 ≤ n−k k−1 then one of the two cases must obtain, i.e., V must have either a quantum k-clique or a quantum k-anticlique. A little algebra shows that this inequality is equivalent to n ≥ k3 − k + 1. (cid:3) Lemma 4.2. Let v1, . . . , vr be vectors in Cs. Then there are vectors w1, . . . , wr ∈ Cr−1 such that the vectors vi ⊕ wi ∈ Cs+r−1 are pairwise orthogonal and all have the same norm. Proof. Let G be the Gramian matrix of the vectors vi and let kGk be its operator norm. Then rank(kGkIr − G) ≤ r − 1, so we can find vectors wi ∈ Cr−1 whose Gramian matrix is kGkIr − G. The Gramian matrix of the vectors vi ⊕ wi is then kGkIr, as desired. (cid:3) Then next lemma improves Proposition 2.2. Lemma 4.3. Let n = k2 + k − 1 and suppose A1, . . . , Ak2 are Hermitian matrices in Mn such that for each i we have hAiei, eii = 1, and also hAier, esi = 0 whenever max{r, s} > i. Then V = span{I, A1, . . . , Ak2 } has a quantum k-clique. Proof. Let Ai have matrix entries (ai such that the matrices rs). The goal is to find vectors v1, . . . , vk2 ∈ Ck A′ i = X1≤r,s≤k2 ai rsvrv∗ s ∈ Mk are linearly independent. Once we have done this, find vectors wi ∈ Ck2−1 as in N (vi ⊕ wi) ∈ Cn ∼= Ck ⊕ Ck2−1 where N is the common Lemma 4.2 and let fi = 1 norm of the vi ⊕ wi. Then the fi form an orthonormal set in Cn, so they can be extended to an orthonormal basis, and the operators whose matrices for this basis i on the initial Ck, which are linearly are the Ai compress to the matrices independent. So P VP contains k2 linearly independent matrices, where P is the orthogonal projection onto Ck, showing that V has a quantum k-clique. N 2 A′ 1 The vectors vi are constructed inductively. Once v1, . . . , vi are chosen so that A′ 1, . . . , A′ i are independent, future choices of the v's cannot change this since A1, . . . , Ai all live on the initial i × i block. We can let v1 be any nonzero vec- tor in Ck, since A1 = e1e∗ 1 and this only has to be nonzero. Now suppose v1, . . . , vi−1 have been chosen and we need to select vi so that A′ i is rsvrv∗ independent of A′ s . i−1. After choosing vi we will have A′ 1, so that A′ i = P1≤r,s≤i ai 1 = v1v∗ 1, . . . , A′ A QUANTUM RAMSEY THEOREM 7 Let B be this sum restricted to 1 ≤ r, s ≤ i − 1. That part is already determined since vi does not appear. Also let u = ai 1iv1 + · · · + ai (i−1)ivi−1; then we will have A′ (using the assumption that ai i = B + uv∗ ii = 1). That is, i + viu∗ + viv∗ i A′ i = (B − uu∗) + (u + vi)(u + vi)∗ = B′ + uu∗ where u = u + vi is arbitrary, and the question is whether u can be chosen to make this matrix independent of A′ i span Mk - there is no matrix which is Hilbert-Schmidt orthogonal to B′ + uu∗ for all u - so there must be a choice of u which makes A′ i−1, as desired. (cid:3) i−1. But the possible choices of A′ i independent of A′ 1, . . . , A′ 1, . . . , A′ Next we prove a technical variation on Lemma 4.3. Lemma 4.4. Let n = k4 + k3 + k − 1 and let V be an operator system contained in Mn. Suppose V contains matrices A1, . . . , Ak4+k3 such that for each i we have hAiei, ei+1i 6= 0, and also hAier, esi = 0 whenever max{r, s} > i + 1 and r 6= s. Then V has a quantum k-clique. rr, . . . , ai Proof. Let Ai have matrix entries (ai rs). Observe that for each i the compression of Ai to span{ei+2, . . . , en} is diagonal. For each r > i + 1 let the r-tail of Ai be the vector (ai nn). Suppose there exist indices i1, . . . , ik2+k−1 such that the r-tails of the Aij , 1 ≤ j ≤ k2 + k − 1, are linearly independent, where r = maxj{ij + 2}. Then the compression of V to span{er, . . . , en} contains k2 + k − 1 linearly independent diagonal matrices, so it has a quantum k-clique by the first assertion of Lemma 4.1. Thus, we may assume that for any k2 + k − 1 distinct indices ij the matrices Aij have linearly dependent r-tails. We construct an orthonormal sequence of vectors vi and a sequence of Her- mitian matrices Bi ∈ V, 1 ≤ i ≤ k2, such that the compressions of the Bi to span{v1, . . . , vk2 , ek4+k3+1, . . . , ek4+k3+k−1} satisfy the hypotheses of Lemma 4.3. This will ensure the existence of a quantum k-clique. 1 = Pk2+k−1 The first k2 + k − 1 matrices A1, . . . , Ak2+k−1 have linearly dependent r-tails for r = k2 + k + 1. Thus there is a nontrivial linear combination B′ αiAi whose r-tail is the zero vector. Letting j be the largest index such that αj is nonzero, we have hB′ 1ej, ej+1i 6= 0 because hAj ej, ej+1i 6= 0 but hAiej, ej+1i = 0 for i < j. Thus the compression of B′ 1 to span{e1, . . . , ek2+k} is nonzero, so there exists a unit vector v1 in this span such that hB′ 1v1, v1i 6= 0. Then let B1 be a scalar multiple of either the real or imaginary part of B′ 1 which satisfies hB1v1, v1i = 1. Note that hB1er, esi = 0 for any r, s with max{r, s} > k2 + k. Apply the same reasoning to the next block of k2 + k − 1 matrices Ak2+k+1, . . . , A2k2+2k−1 to find v2 and B2, and proceed inductively. After k2 steps, k2(k2 + k) = k4 + k3 indices will have been used up and k − 1 (namely, ek4+k3+1, . . . , ek4+k3+k−1) will remain, as needed. (cid:3) i=1 Theorem 4.5. Every operator system in M8k11 has either a quantum k-clique or a quantum k-anticlique. 8 NIK WEAVER Proof. Set n = 8k11 and let V be an operator system in Mn. Find a unit vector v1 ∈ Cn, if one exists, such that the dimension of Vv1 = {Av1 : A ∈ V} is less than 8k8. Then find a unit vector v2 ∈ (Vv1)⊥, if one exists, such that the dimension of (Vv1)⊥ ∩ (Vv2) is less than 8k8. Proceed in this fashion, at the rth step trying to find a unit vector vr in (Vv1)⊥ ∩ · · · ∩ (Vvr−1)⊥ such that the dimension of (Vv1)⊥ ∩ · · · ∩ (Vvr−1)⊥ ∩ (Vvr) is less than 8k8. If this construction lasts for k3 steps then the compression of V to span{v1, . . . , vk3 } ∼= Mk3 is contained in Dk3 , so this compression, and hence also V, has either a quantum k-clique or a quantum k-anticlique by Lemma 4.1. Otherwise, the construction fails at some stage d. This means that the compres- sion V ′ of V to F = (Vv1)⊥ ∩ · · · ∩ (Vvd)⊥ has the property that the dimension of V ′v is at least 8k8, for every unit vector v ∈ F . Work in F . Choose any nonzero vector w1 ∈ F and find A1 ∈ V ′ such that w2 = A1w1 is nonzero and orthogonal to w1. Then find A2 ∈ V ′ such that w3 = A2w2 is nonzero and orthogonal to span{w1, w2, A1w1, A∗ 1w2}. Continue in this way, at the rth step finding Ar ∈ V ′ such that wr+1 = Arwr is nonzero and orthogonal to span{wj, Aiwj, A∗ i wj : i < r and j ≤ r}. The dimension of this span is at most 2r2 − r, so as long as r ≤ 2k4 its dimension is less than 8k8 and a vector wr+1 can be found. Compressing to the span of the wi then puts us in the situation of Lemma 4.4 with n = 2k4, which is more than enough. So there exists a quantum k-clique by that lemma. (cid:3) 1w1, A1w2, A∗ The constants in the proof could easily be improved, but only marginally. Very likely the problem of determining optimal bounds on quantum Ramsey numbers is open-ended, just as in the classical case. 5. A generalization In this section we will present a result which simultaneously generalizes the classical and quantum Ramsey theorems. This is less interesting than it sounds because the proof involves little more than a reduction to these two special cases. Perhaps the statement of the theorem is more significant than its proof. At the beginning of Section 2 we showed how any simple graph G on the vertex set {1, . . . , n} gives rise to an operator system VG ⊆ Mn. This operator system has the special property that it is a bimodule over Dn, i.e., it is stable under left and right multiplication by diagonal matrices. Conversely, it is not hard to see that any operator system in Mn which is also a Dn-Dn-bimodule must have the form VG for some G ([8], Propositions 2.2 and 2.5). The general definition therefore goes as follows: Definition 5.1. ([8], Definition 2.6 (d)) Let M be a unital ∗-subalgebra of Mn. A quantum graph on M is an operator system V ⊆ Mn which satisfies M′VM′ = V. Here M′ = {A ∈ Mn : AB = BA for all B ∈ M} is the commutant of M. This if M and N are ∗-isomorphic definition is actually representation-independent: unital ∗-subalgebras of two matrix algebras (possibly of different sizes), then the quantum graphs on M naturally correspond to the quantum graphs on N ([8], A QUANTUM RAMSEY THEOREM 9 Theorem 2.7). More properly, one could say that the pair (M, V) is the quantum graph, just as a classical graph is a pair (V, E). If M = Mn then its commutant is C·In and the bimodule condition in Definition 5.1 is vacuous: any operator system in Mn is a quantum graph on Mn. On the other hand, the commutant of M = Dn is itself, so that by the comment made above, the quantum graphs on Dn - the operator systems which are Dn-Dn-bimodules - correspond to simple graphs on the vertex set {1, . . . , n}. In this correspondence, subsets of {1, . . . , n} give rise to orthogonal projections P ∈ Dn, and the k-cliques and k-anticliques of the graph are realized in the matrix picture as rank k orthogonal projections P ∈ Dn which satisfy P VGP = P MnP or P DnP , respectively. This suggests the following definition. Definition 5.2. Let M be a unital ∗-subalgebra of Mn and let V ⊆ Mn be a quantum graph on M. A rank k projection P ∈ M is a quantum k-clique if it satisfies P VP = P MnP and a quantum k-anticlique if it satisfies P VP = P M′P . Since every operator system contains the identity matrix, if V is a quantum graph on M then M′ ⊆ V. So P VP = P M′P is the minimal possibility, as P VP = P MnP is the maximal possibility. Note the crucial requirement that P must belong to M. If M = Mn then M′ = C · In and the preceding definition duplicates the notions of quantum k-clique and quantum k-anticlique used earlier in the paper, whereas if M = Dn it effectively reproduces the classical notions of k-clique and k-anticlique in a finite simple graph. In the classical setting fewer operator systems count as graphs, but one also has less freedom in the choice of P when seeking cliques or anticliques. We require only the following simple lemma. Lemma 5.3. Let V ⊆ Mnd ∼= Mn ⊗ Md be a quantum graph on Mn ⊗ Id. If nd ≥ 8k11 then there is a projection in Mn ⊗ Id whose rank is at least k, and which is either a quantum clique or a quantum anticlique of V. Proof. Since V is a bimodule over (Mn⊗Id)′ = In⊗Md, it has the form V = W ⊗Md for some operator system W ⊆ Mn. If d = 1 then the desired result was proven in Theorem 4.5, and if d ≥ k then any projection of the form P ⊗ Md, where P is a rank 1 projection in Mn, will have rank at least k and be both a quantum clique and a quantum anticlique. So assume 2 ≤ d < k. Now if d ≥ 3 then d10/11 > 2, so d < k(d10/11 − 1). Thus 1 < k d (d10/11 − 1), i.e., k d + 1 < k d · d10/11 = k d1/11 , which implies ( k d . So finally n ≥ 8k11 d > 8(cid:18) k d d + 1)11 < k11 d(cid:25)11 + 1(cid:19)11 > 8(cid:24) k . If d = 2 then d < 3(d10/11 − 1), so the same reasoning leads to the same inequality n ≥ 8⌈ k d ⌉11 provided k ≥ 3, and the inequality is immediate when k = d = 2. So we conclude that in all cases n ≥ 8⌈ k d ⌉- clique or a quantum ⌈ k d ⌉-anticlique Q ∈ Mn. Then Q ⊗ Id is correspondingly either a quantum ⌈ k d ⌉ · d-anticlique of V, which is enough. (cid:3) d ⌉11. By Theorem 4.5, W has a quantum ⌈ k d ⌉ · d-clique or a quantum ⌈ k Note that we cannot promise a quantum k-clique or -anticlique, only a ≥ k-clique or -anticlique, since the rank of any projection in Mn ⊗ Id is a multiple of d. 10 NIK WEAVER Theorem 5.4. For every k there exists n such that if M is a unital ∗-subalgebra of Mn and V ⊆ Mn is an operator system satisfying M′VM′ = V, then there is a projection P ∈ M whose rank is at least k and such that P VP = P MnP or P M′P . Proof. Let R(k, k) be the classical Ramsey number and set n = 8k11 · R(k, k). Now M has the form (Mn1 ⊗ Id1) ⊕ · · · ⊕ (Mnr ⊗ Idr ) for some pair of sequences (n1, . . . , nr) and (d1, . . . , dr) such that n1d1 + · · · + nrdr = n. Thus if r ≤ R(k, k) then for some i we must have nidi ≥ 8k11, and compressing to that block then yields the desired conclusion by appealing to the lemma. Otherwise, if r > R(k, k), then choose a sequence of rank 1 projections Qi ∈ Mni and work in QMnQ where Q = (Q1 ⊗ Id1 ) ⊕ · · · ⊕ (Qr ⊗ Idr ). Then QMnQ ∼= Md1+···+dr , QMQ ∼= C · Id1 ⊕ ∼= Dr, and QVQ is a bimodule over the commutant of QMQ in QMnQ, · · · ⊕ C · Idr i.e., the ∗-algebra Md1 ⊕ · · · ⊕ Mdr . It follows that there is a graph G = (V, E) on the vertex set V = {1, . . . , r} such that QVQ has the form QVQ = X Eij ⊗ Mdidj , taking the sum over the set of pairs {(i, j) : i = j or {i, j} ∈ E} ([8], Theorem 2.7). Since r > R(k, k), there exists either a k-clique or a k-anticlique in G, and this gives rise to a diagonal projection in QMQ whose rank is at least k and which is either a quantum k-clique or a quantum k-anticlique of V. (cid:3) Again, when M = Mn Theorem 5.4 recovers the quantum Ramsey theorem and when M = Dn it recovers the classical Ramsey theorem (though in both cases with worse constants). Theorem 5.4 could also be proven by mimicking the proof of Theorem 4.5. How- ever, in order to accomodate the requirement that P belong to M we need to modify the last part of the proof so as to be sure that each wr belongs to Ww1∩· · ·∩Wwr−1. This means that instead of needing Wv to have sufficiently large dimension for each v, we need it to have sufficiently small codimension. Ensuring that this must be the case if the construction in the first part of the proof fails then requires that construction to take place in a space whose dimension is exponential in k. This explains the dramatic difference between classical and quantum Ramsey numbers (the first grows exponentially, the second polynomially). References [1] Y. H. Au-Yeung and Y. T. Poon, A remark on the convexity and positive definiteness concerning Hermitian matrices, Southeast Asian Bull. Math. 3 (1979), 85-92. [2] R. Duan, S. Severini, and A. Winter, Zero-error communication via quantum channels, noncommutative graphs, and a quantum Lov´asz number, IEEE Trans. Inform. Theory 59 (2013), 1164-1174. [3] E. Knill and R. Laflamme, Theory of quantum error-correcting codes, Phys. Rev. A 55 (1997), 900-911. [4] E. Knill, R. Laflamme, and L. Viola, Theory of quantum error correction for general noise, Phys. Rev. Lett. 84 (2000), 2525-2528. [5] D. Stahlke, Quantum source-channel coding and non-commutative graph theory, arXiv:1405.5254. [6] H. Tverberg, A generalization of Radon's theorem, J. London Math. Soc. 41 (1966), 123-128. [7] ---, A generalization of Radon's theorem, II, Bull. Austral. Math. Soc. 24 (1981), 321-325. [8] N. Weaver, Quantum relations, Mem. Amer. Math. Soc. 215 (2012), v-vi, 81-140. [9] ---, Quantum graphs as quantum relations, arXiv:1506.03892. A QUANTUM RAMSEY THEOREM 11 Department of Mathematics, Washington University, Saint Louis, MO 63130 E-mail address: [email protected]
1211.0239	2	1211	2013-03-25T19:13:51	KMS states for the generalized gauge action on graph algebras	[ "math.OA" ]	Given a positive function on the set of edges of an arbitrary directed graph $E=(E^0,E^1)$, we define a one-parameter group of automorphisms on the C-algebra of the graph $C^(E)$, and study the problem of finding KMS states for this action. We prove that there are bijective correspondences between KMS states on $C^*(E)$, a certain class of states on its core, and a certain class of tracial states on $C_0(E^0)$. We also find the ground states for this action and give some examples.	math.OA	math	KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS Gilles G. de Castro and Fernando de L. Mortari Abstract. Given a positive function on the set of edges of an arbitrary di- rected graph E = (E 0, E 1), we define a one-parameter group of automorphisms on the C-algebra of the graph C ∗(E), and study the problem of finding KMS states for this action. We prove that there are bijective correspondences be- tween KMS states on C ∗(E), a certain class of states on its core, and a certain class of tracial states on C0(E 0). We also find the ground states for this action and give some examples. 1. Introduction Given a directed graph E = (E0, E1), we can associate to it a C-algebra C∗(E), and an interesting problem that arises is to find relations between the algebraic properties of the algebra and the combinatorial properties of the graph [?]. One such problem is to determine the set of KMS states for a certain action on the algebra. Graph algebras are a generalization of Cuntz algebras and Cuntz-Krieger alge- bras. For the Cuntz algebra, there is a very natural action of the circle, the gauge action, which can be extended to an action of the real line. The KMS states for this action are studied in [?] and later generalized to a more general action of the line, that can be thought of as a generalized gauge action [?]. The same is done for the Cuntz-Krieger algebras [?], [?]. Recently there were similar results proven for the C-algebra associated to a finite graph. This is done in [?] for an arbitrary finite graph, in [?] for a certain class of finite graphs via groupoid C-algebras and in [?] for the Toeplitz C-algebra of the graph. Our goal is to generalize these results to the case of an arbitrary graph. First we analyze which conditions the restrictions of a KMS state to the core of C∗(E) and to C0(E0) must satisfy. By using a description of the core as an inductive limit, we can build a KMS state on C∗(E) from a tracial state on C0(E0) satisfying the conditions found. In section 2 we review some of the basic definitions and results about graph algebras as well as the description of the core as an inductive limit. In section 3 Partially supported by Funpesquisa/UFSC The first named author was also supported by project mathamsud U11-MATH05 and Capes/Math-AmSud 013/10. 1 2 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI we establish the results concerning KMS states, followed by a discussion on ground states in section 4. In section 5, some examples are given. 2. Graph algebras Definition 2.1. A (directed) graph E = (E0, E1, r, s) consists of nonempty sets E0, E1 and functions r, s : E1 → E0; an element of E0 is called a vertex of the graph, and an element of E1 is called an edge. For an edge e, we say that r(e) is the range of e and s(e) is the source of e. Definition 2.2. A vertex v in a graph E is called a source if r−1(v) = ∅, and is said to be singular if it is either a source, or r−1(v) is infinite. Definition 2.3. A path of length n in a graph E is a sequence µ = µ1µ2 . . . µn such that r(µi + 1) = s(µi) for all i = 1, . . . , n − 1. We write µ = n for the length of µ and regard vertices as paths of length 0. We denote by En the set of all paths of length n and E∗ = ∪n≥0En. We extend the range and source maps to E∗ by defining s(µ) = s(µn) and r(µ) = r(µ1) if n ≥ 2 and s(v) = v = r(v) for n = 0. Definition 2.4. Given a graph E, we define the C-algebra of E as the uni- versal C-algebra C∗(E) generated by mutually orthogonal projections {pv}v∈E0 and partial isometries {se}e∈E1 with mutually orthogonal ranges such that ese = ps(e); (1) s∗ (2) ses∗ e ≤ pr(e) for every e ∈ E1; (3) pv =Pe∈r−1(v) ses∗ e for every v ∈ E0 such that 0 < r−1(v) < ∞. For a path µ = µ1 . . . µn, we denote the composition sµ1 . . . sµn by sµ, and for v ∈ E0 we define sv to be the projection pv. Propositions 2.5, 2.6, 2.8 and 2.9 below are found in [?] (as Corollary 1.15, Proposition 2.1, Proposition 3.2 and Corollary 3.3, respectively) in the context of row-finite graphs, but their proofs hold just the same for general graphs as above. Proposition 2.5. For α, β, µ, ν ∈ E∗ we have (sµs∗ ν)(sαs∗ β) =  sµα′ s∗ β sµs∗ βν′ 0 if α = να′ if ν = αν′ otherwise and C∗(E) = span{sµs∗ ν : µ, ν ∈ E∗, s(µ) = s(ν)}. Proposition 2.6. Let E be a graph. Then there is an action γ of T on C∗(E), called a gauge action, such that γz(se) = zse for every e ∈ E1 and γz(pv) = pv for every v ∈ E0. Definition 2.7. The core of the algebra C∗(E) is the fixed-point subalgebra for the gauge action, denoted by C∗(E)γ . Proposition 2.8. C∗(E)γ = span{sµs∗ ν : µ, ν ∈ E∗, s(µ) = s(ν), µ = ν}. Proposition 2.9. There is a conditional expectation Φ : C∗(E) → C∗(E)γ such that Φ(sµs∗ ν) = [µ = ν]sµs∗ ν. It is useful to describe the core as an inductive limit of subalgebras, as was done in an appendix in [?]. The idea is as follows. For k ≥ 0 define the sets Fk = span{sµs∗ ν : µ, ν ∈ Ek, s(µ) = s(ν)}, KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS 3 Ek = span{sµs∗ ν : µ, ν ∈ Ek and s(µ) = s(ν) is singular}, Also, for a given vertex v we define Ck = F0 + · · · + Fk. Fk(v) = span{sµs∗ ν : µ, ν ∈ Ek, s(µ) = s(ν) = v} so that (2.1) Fk = Mv∈E0 Fk(v) as a direct sum of C-algebras. Lemma 2.10. Let Λ be the set of all finite subsets of Ek and for λ ∈ Λ define sµs∗ µ. uλ = Xµ∈λ Then {uλ}λ∈Λ is an approximate unit of Fk consisting of projections. Proof. This is a direct consequence of Proposition 2.5. (cid:3) The following result is a combination of Proposition A.1 and Lemma A.2 in [?]. Proposition 2.11. With the notation as above for a graph E, the following hold for k ≥ 0: (a) Ck is a C-subalgebra of C∗(E)γ , Fk+1 is an ideal in Ck, Ck ⊆ Ck+1 and C∗(E)γ = lim−→ Ck. (b) Fk ∩ Fk+1 =L{Fk(v) : 0 < r−1(v) < ∞}. (C-algebraic direct sum) (c) Ck = E0 ⊕ E1 ⊕ · · · ⊕ Ek−1 ⊕ Fk (vector space direct sum) With the above, we can now prove the following. Proposition 2.12. For each k ≥ 0, Ck ∩ Fk+1 = Fk ∩ Fk+1. Proof. Obviously one has Fk ∩ Fk+1 ⊆ Ck ∩ Fk+1. On the other hand, given x ∈ Ck ∩ Fk+1, one can decompose x as sums in Ck and Ck+1 with Proposition 2.11(c), use the fact that Fk = Ek ⊕ Fk ∩ Fk+1 and the uniqueness of the direct sum decompositions of x to conclude that x ∈ Fk. (cid:3) 3. KMS states for the generalized gauge action In this section we define an action on C∗(E) from a function c : E1 → R∗ + similar to what is done in [?] for the Cuntz algebras and in [?] for the Cuntz- Krieger algebras. We will always suppose that there is a constant k > 0 such that c(e) > k for all e ∈ E1 and that β > 0. Observe that in this case c−β is bounded. + by defining c(v) = 1 We extend a function as above to a function c : E∗ → R∗ if v ∈ E0 and c(µ) = c(µ1) . . . c(µn) if µ = µ1 · · · µn ∈ En. Proposition 3.1. Given a function c : E1 → R∗ tinuous action σc : R → Aut(C∗(E)) given by σc σc t (se) = c(e)itse for all e ∈ E1. +, there is a strongly con- t (pv) = pv for all v ∈ E0 and 4 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI e = ses∗ Proof. Let Te = c(e)itse and note that Te is a partial isometry with T ∗ e Te = ese and TeT ∗ s∗ e. It follows that the sets {pv}v∈E0 and {Te}e∈E1 satisfy the same relations as {pv}v∈E0 and {se}e∈E1. By the universal property, there is a homomorphism σc t (pv) = pv for all v ∈ E0 and σc t (se) = Te = c(e)itse for all e ∈ E1. t : C∗(E) → C∗(E) such that σc It is easy to see that σc t1 ◦ σc t2 = σc t1+t2 and σc 0 = Id. Hence σc t is an automor- phism with inverse σc −t. x = Pµ,ν∈E∗ λµ,νsµs∗ λµ,ν 6= 0, there is δµ,ν such that To prove continuity, let a ∈ C∗(E), t ∈ R and ε > 0. Take x to be a finite sum ν such that ka − xk < ε/3. For each pair of paths µ, ν with c(µ)itc(ν)−it − c(µ)iuc(ν)−iu < ε 3Pµ,ν∈E∗ kλµ,νsµs∗ ν k for all u ∈ R with t − u < δµ,ν. If we take δ to be the minimum of all such δµ,ν, then for all u ∈ R with t − u < δ we have kσc t (x) − σc (c(µ)itc(ν)−it − c(µ)iuc(ν)−iu)λµ,ν sµs∗ < u(x)k =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xµ,ν∈E∗ 3Pµ,ν∈E∗ kλµ,νsµs∗ ε kλµ,νsµs∗ νk = ε 3 νk Xµ,ν∈E∗ ν(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and hence kσc ≤ kσc t (a) − σc u(a)k = kσc t (a) − σc t (a − x)k + kσc t (x) − σc t (x) + σc u(x)k + kσc u(x) + σc t (x) − σc u(x − a)k ≤ ε/3 + ε/3 + ε/3 = ε. u(x) − σc u(a)k ≤ From now on, we will write simply σ instead of σc. The next result shows that KMS states on C∗(E) are determined by their values at the core algebra. (cid:3) Proposition 3.2. Suppose c : E1 → R∗ + is such that c(µ) 6= 1 for all µ ∈ If two (σ, β)-KMS states ϕ1, ϕ2 on C∗(E) coincide at the core algebra E∗\E0. C∗(E)γ, then ϕ1 = ϕ2. sµs∗ Proof. Taking an arbitrary sµs∗ ν ∈ C∗(E)γ and thus ϕ1(sµs∗ Suppose then that µ 6= ν, and denote the functional ϕ2 − ϕ1 by ϕ. Using ν such that s(µ) = s(ν), if µ = ν then ν ) = ϕ2(sµs∗ ν). the KMS condition, one obtains ϕ(sµs∗ ν ) = ϕ(s∗ νc(µ)−β sµ) =  c(µ)−βϕ(s∗ c(µ)−β ϕ(s∗ ν′ ) µ′ ) 0 if ν = µν′ if µ = νµ′ otherwise . It is therefore sufficient to show that ϕ(sµ) = ϕ(s∗ notice that if C∗(E) has a unit, then µ) = 0 if µ ≥ 1. To see this, ϕ(sµ) = ϕ(sµ1) = ϕ(1c(µ)−βsµ) = c(µ)−β ϕ(sµ), whence ϕ(sµ) = 0 since c(µ) 6= 1 by hypothesis; the non-unital case is established analogously with the use of an approximate unit. (cid:3) KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS 5 Theorem 3.3. Suppose c : E1 → R∗ + is such that c(µ) 6= 1 for all µ ∈ E∗\E0. If ϕ is a (σ, β)-KMS state on C∗(E) then its restriction ω = ϕC ∗(E)γ to C∗(E)γ satisfies ω(sµs∗ (3.1) conversely, if ω is a state on C∗(E)γ satisfying (3.1) then ϕ = ω◦Φ is a (σ, β)-KMS state on C∗(E), where Φ is the conditional expectation from proposition 2.9. The correspondence thus obtained is bijective and preserves convex combinations. ν) = [µ = ν]c(µ)−βω(ps(µ)); Proof. Let ϕ be a (σ, β)-KMS state on C∗(E) and ω its restriction to C∗(E)γ. If µ, ν are paths such that µ = ν and s(µ) = s(ν) then ν σiβ(sµ)) = ϕ(s∗ ν) = ϕ(sµs∗ ν) = ϕ(s∗ ω(sµs∗ ν c(µ)−β sµ) = = [µ = ν]c(µ)−βϕ(ps(µ)) = [µ = ν]c(µ)−β ω(ps(µ)). Conversely, let ω be a state on C∗(E)γ satisfying (3.1) and ϕ = ω ◦ Φ; we have to show that ϕ satisfies the KMS condition. By continuity and linearity, it is η where µ, ν, ζ, η ∈ E∗ sufficient to verify this for elements x = sµs∗ are paths such that s(µ) = s(ν) and s(ζ) = s(η). ν and y = sζs∗ We need to check that ϕ(xy) = ϕ(yσiβ(x)). First note that xy = (sµs∗ ν)(sζ s∗ and yσiβ(x) = c(µ)−βc(ν)β (sζs∗ η)(sµs∗ if ζ = νζ′ if ν = ζν′ otherwise (1) (2) (3) sµζ′ s∗ η sµs∗ ην′ 0 η) =  ν ) = c(µ)−βc(ν)β  sζµ′ s∗ ν sζs∗ νη′ 0 if µ = ηµ′ if η = µη′ otherwise (a) (b) (c) . There are nine cases to consider. In each case it must be checked whether the resulting paths have the same size, for they will be otherwise sent to 0 by Φ. Case 1-a. In this case ζ = νζ′ and µ = ηµ′ so that ζ = ν + ζ′ and µ = η + µ′. We claim that µζ′ = µ + ζ′ = η if and only if ζµ′ = ζ + µ′ = ν, and in this case µ = η and ν = ζ. In fact, µ + ζ′ = η ⇔ η + µ′ + ζ′ = η ⇔ µ′ + ζ′ = 0 ⇔ ⇔ ν + ζ′ + µ′ = ν ⇔ ζ + µ′ = ν. Observe that, in this case, we have µ′ + ζ′ = 0 so that µ′ = ζ′ = 0, and hence µ = η, ν = ζ. It follows that, if µζ′ 6= η, then ϕ(xy) = ω ◦ Φ(xy) = ω(0) = ω ◦ Φ(yσiβ(x)) = ϕ(yσiβ (x)) and, if µζ′ = η, we get ϕ(xy) = ϕ(sµs∗ µ) = ω(sµs∗ µ) = c(µ)−βω(ps(µ)) and on the other hand ϕ(yσiβ (x)) = c(µ)−βc(ν)β ϕ(sν s∗ ν) = c(µ)−βc(ν)β ω(sνs∗ ν) = = c(µ)−β c(ν)βc(ν)−β ω(ps(µ)) = c(µ)−βω(ps(µ)). Case 1-b. Now, we have that ζ = νζ′ and η = µη′ so that ζ = νζ′ = ν + ζ′ and η = µη′ = µ + η′; as before, we can check that µ + ζ′ = η if and 6 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI only if ζ = ν + η′. If that is not the case then ϕ(xy) = 0 = ϕ(yσiβ (x)). If the equivalent conditions are true then ϕ(xy) = ϕ(sµζ′ s∗ η) = ω(sµζ′ s∗ η) = [µζ′ = η]c(η)−β ω(ps(η)) and ϕ(yσiβ (x)) = c(µ)−βc(ν)β ϕ(sζs∗ νη′ ) = c(µ)−βc(ν)β [ζ = νη′]c(ζ)−β ω(ps(ζ)). Since ζ = νζ′ and η = µη′, we have that µζ′ = η if and only if ζ = νη′ and if both are true, then ζ′ = η′ and c(µ)βc(ν)−β c(ζ)−β = c(µ)−βc(ν)β c(ν)−βc(η′)−β = c(µ)−βc(η′)−β = = c(µ)−βc(ζ′)−β = c(η)−β. From our original hypothesis, we have that s(η) = s(ζ) so we conclude that ϕ(xy) = ϕ(yσiβ (x)). Case 1-c. In this case ϕ(yσiβ (x)) = 0, so we need to check that ϕ(xy) = 0. As with the previous case, we have that ϕ(xy) = [µζ′ = η]c(η)−β ω(ps(η)); however, in case (c) µζ′ 6= η for all ζ′ and therefore ϕ(xy) = 0. The other cases are analogous to these three, except for case 3-c, where ϕ(xy) = 0 = ϕ(yσiβ (x)) since xy = 0 = yσiβ (x). That the correspondence obtained is bijective follows from Proposition 3.2 and (cid:3) that it preserves convex combinations is immediate. Next, we want to show that there is also a bijective correspondence between (σ, β)-KMS states on C∗(E) and a certain class of tracial states on C0(E0). We build this correspondence by first describing a correspondence between this class of tracial states on C0(E0) and states ω on C∗(E)γ satisfying (3.1). The conditions found for the states on C0(E0) are similar to those in [?], al- though as discussed in [?], their results cannot be used directly for an arbitrary graph; nevertheless, the results of Theorem 1.1 of [?] still apply in the general set- ting, and we use them to build a certain kind of transfer operator on the dual of C0(E0). Let us first recall how to construct C∗(E) as C-algebra associated to a C- correspondence [?]. If we let A = C0(E0), then Cc(E1) has a pre-Hilbert A-module structure given by hξ, ηi (v) = Xe∈s−1(v) ξ(e)η(e) for v ∈ E0, (ξa)(e) = ξ(e)a(s(e)) for e ∈ E1, where ξ, η ∈ Cc(E1) and a ∈ A; it follows that the completion X of Cc(E1) with respect to the norm given by kξk = k hξ, ξi k1/2 is a Hilbert A-module. A represen- tation iX : A → L(X) is then defined by by iX (a)(ξ)(e) = a(r(e))ξ(e) for v ∈ E0, where L(X) is the C-algebra of adjointable operators on X. Let K(X) be the C-subalgebra of L(X) generated by the operators θξ,η given by θξ,η(ζ) = ξ hη, ζi. For each e ∈ E1, let χe ∈ Cc(E1) be the characteristic function of {e} and observe that (tλ =Xe∈λ θχe,χe)λ∈Λ , KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS 7 where Λ is the set of all finite subsets of E1, is an approximate unit of K(X) . It is essentially the same approximate unit given by Lemma 2.10. If τ is a tracial state on C0(E0), as in Theorem 1.1 of [?] we define a trace Trτ on L(X) by where T ∈ L(X). Trτ (T ) = lim λ→∞Xe∈λ τ (hχe, T χei) For a function c : E1 → R∗ + as in the beginning of the section and β > 0, we have that c−β ∈ Cb(E1) and so it defines an operator on L(X) by pointwise multiplication. Definition 3.4. Given c and β as above and τ a tracial state on C0(E0), we define a trace Fc,β(τ ) on C0(E0) by Fc,β(τ )(a) = Trτ (iX (a)c−β). Now, observe that C0(E0) ∼= span{pv}v∈E0; regarding this as an equality, for a given tracial state τ on C0(E0) we will write τ (pv) = τv. For v ∈ E0, it can be verified that Fc,β(τ )(pv) = lim c(e)−βτs(e), D→r−1(v)Xe∈D where the limit is taken on finite subsets D of r−1(v), and Fc,β(τ )(pv) = 0 if r−1(v) = ∅. Remark 3.5. By Theorem 1.1 of [?], if Fc,β(τ )(a) < ∞ for all a ∈ C0(E0), then Fc,β(τ ) is actually a positive linear functional; also, if V ⊆ E0 and Fc,β(τ )(pv) < ∞ for all v ∈ V then Fc,β(τ ) is a positive linear functional on span{pv : v ∈ V }. Definition 3.6. For a vertex v ∈ E0 and a positive integer n, we define r−n(v) = {µ ∈ En : r(µ) = v}. Lemma 3.7. If Fc,β(τ )(pv) ≤ τv for all v ∈ E0 then lim D→r−n(v) Xµ∈D for all v ∈ E0 and for all n ∈ N∗. c(µ)−β τs(µ) ≤ τv Proof. This is proved by induction. The case n = 1 is the hypothesis. Now suppose it is true for n, then lim D→r−(n+1)(v) Xµ∈D c(µ)−βτs(µ) = lim D→r−n(v) Xν∈r≤n c(ν)−β Xe∈r−1(s(ν)) c(e)−βτs(e)[νe ∈ D] ≤ ≤ lim D→r−n(v)Xν∈D c(ν)−β τs(ν) ≤ τv where the first inequality is true due to the fact that since c is a positive function then the net Pe∈D c(e)−βτs(e) for finite subsets D of r−1(s(ν)) is nondecreasing and less than or equal to τs(ν) by hypothesis. The last inequality is the induction hypothesis. (cid:3) The next lemma is found in [?] for unital algebras, but their proof carries out the same in the non-unital case by using an approximate unit instead of a unit. 8 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI Lemma 3.8 (Exel-Laca). Let B be a C-algebra, A be a C-subalgebra such that an approximate unit of A is also an approximate unit of B and I a closed bilateral ideal of B such that B = A + I. Let ϕ be a state on A and ψ a linear positive functional on I such that ϕ(x) = ψ(x) ∀x ∈ A ∩ I and ψ(x) ≤ ϕ(x) ∀x ∈ A+, where ψ(x) = limλ ψ(buλ) for an approximate unit {uλ}λ∈Λ of I. Then there is a unique state Φ on B such that ΦA = ϕ and ΦI = ψ. We want to use this lemma for A = Cn, I = Fn+1 and B = Cn+1, defined in section 2. For that, we first note that Fn+1 is indeed an ideal of Cn+1 by Proposition 2.11 and that the approximate unit for F0 given by Lemma 2.10 is also an approximate unit of Cn for all n. We also need to know what the intersection A ∩ I is, and for that we need a preliminary result. Lemma 3.9. Suppose c, β and τ are such that Fc,β(τ )(a) ≤ τ (a) for all a ∈ C0(E0)+, then for each k ≥ 1 there is a unique positive linear functional ψk on Fk defined by (3.2) ψk(sµs∗ ν) = [µ = ν]c(µ)−β τs(µ). Proof. Since {sµs∗ tion 3.2 defines a unique linear functional on span{sµs∗ To extend to the closure, it is sufficient to prove that ψk is continuous. ν : µ, ν ∈ Ek, s(µ) = s(ν)} is linearly independent, equa- ν : µ, ν ∈ Ek, s(µ) = s(ν)}. If x ∈ span{sµs∗ ν : µ, ν ∈ Ek, s(µ) = s(ν)} then x = Xv∈V X(µ,ν)∈Gv µ,νsµs∗ av ν where V is a finite subset of E0 and Gv is a finite subset of {(µ, ν) ∈ En × En : s(µ) = s(ν) = v}. Using the decomposition given by equation 2.1 and observing that {sµs∗ ν : (µ, ν) ∈ Gv} can be completed to generators of a matrix algebra, we have that k(av µ,ν)µ,ν k where the last norm is the matrix norm. If Tr is the usual matrix trace we have = max v∈V kxk = max av µ,νsµs∗ av µ,νsµs∗ ν(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) v∈V (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) X(µ,ν)∈Gv ψk(x) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ψk ν Xv∈V X(µ,ν)∈Gv  =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) µ,ν [µ = ν]c(µ)−β τs(µ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xv∈V X(µ,ν)∈Gv µ,ν )µ,ν diag(c(µ)−βτs(µ))(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xv∈V ≤ Xv∈V (cid:12)(cid:12)Tr((av µ,ν )µ,ν diag(c(µ)−βτs(µ))(cid:12)(cid:12) ≤ ≤ Xv∈V µ,ν)µ,ν k Xµ:(µ,µ)∈Gv Tr((av k(av av = = ≤ c(µ)−βτs(µ) ≤lemma 3.7 KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS 9 ≤ Xv∈V k(av µ,ν)µ,ν kτv ≤ max v∈V (k(av µ,ν )µ,νk)Xv∈V τv = = kxkXv∈V τv ≤ kxk where the last inequality comes from the fact that τ comes from a probability measure on a discrete space. (cid:3) Theorem 3.10. If ω is a state on C∗(E)γ satisfying (3.1) then its restriction τ to C0(E0) satisfies: (K1) Fc,β(τ )(a) = τ (a) for all a ∈ span{pv : 0 < r−1(v) < ∞}, (K2) Fc,β(τ )(a) ≤ τ (a) for all a ∈ C0(E0)+. Conversely, if τ is a tracial state on C0(E0) satisfying (K1) and (K2) then there is unique state ω on C∗(E)γ satisfying (3.1). This correspondence preserves convex combinations. Proof. Let ω be a state on C∗(E)γ satisfying (3.1) and τ its restriction to C0(E0). By Remark 3.5, to establish (K1) it is sufficient to consider a = pv where v ∈ E0 is such that 0 < r−1(v) < ∞, and in this case τ (pv) = ω(pv) = ω = Xe∈r−1(v)  Xe∈r−1(v) ses∗ e  = Xe∈r−1(v) c(e)−βτs(e) = Fc,β(τ )(pv). c(e)−βω(ps(e)) = For (K2), let a ∈ C0(E0)+ and write a =Pv∈E0 avpv; again, by remark 3.5 it is sufficient to show the result for a = pv where v ∈ E0. If 0 < r−1(v) < ∞, then we have an equality as shown above. If r−1(v) = 0, then Fc,β(τ )(pv) = 0 ≤ τ (pv). If r−1(v) = ∞, then Fc,β(τ )(pv) = lim D→r−1(v)Xe∈D D→r−1(v)Xe∈D = lim c(e)−βτs(e) = lim D→r−1(v)Xe∈D ω(ses∗ e) = ω(pvses∗ e) ≤ ω(pv) = τ (pv). To see the inequality above, we observe that ses∗ e are mutually orthogonal projections that commute with pv so that pvses∗ pv −Xe∈D e = pv 1 −Xe∈D ses∗ e! = 1 −Xe∈D ses∗ e! pv 1 −Xe∈D ses∗ e! ≥ 0. Now, let τ be a tracial state on C0(E0) satisfying (K1) and (K2). We will use Lemma 3.8 and the discussion after it. Observe that F0 = C0(E0) and let ψ0 = τ . For n ≥ 1, by Lemma 3.9 there exists a positive linear functional ψn on Fn defined by ψn(sµs∗ ν) = [µ = ν]c(µ)−β τs(µ). Let us show by induction that there is a unique state ϕn on Cn such that the restriction to Fn is ψn. For n = 1, we use Lemma 3.8 with A = C0(E0), I = F1, 10 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI B = C1, ϕ = τ and ψ = ψ1. By Proposition 2.12, in this case A ∩ I = span{pv : v ∈ E0, 0 < r−1(v) < ∞} and if pv ∈ A ∩ I then ψ(pv) = ψ1(pv) = ψ1(svs∗ v) = τv = τ (pv). Using the approximate unit given by Lemma 2.10, for any v ∈ E0 we have ψ(pv) = ψ1(pv) = lim λ→∞ ψ1(pvuλ) = lim D→r−1(v)Xe∈D ψ1(ses∗ e) = = lim D→r−1(v)Xe∈D c(e)−βτs(e) = Fc,β(τ )(pv) ≤ τ (pv), where the last inequality is exactly (K2). Now suppose that there is a unique state ϕn on Cn such that the restriction to Fn is ψn and let us show that this is also true for n + 1. We set A = Cn, I = Fn+1, B = Cn+1, ϕ = ϕn and ψ = ψn+1 on Lemma 3.8. By Proposition 2.12, we have that A ∩ I = span{sµs∗ ν : µ, ν ∈ En, s(µ) = s(ν), µ = ν, 0 < r−1(s(µ)) < ∞}. Let sµs∗ ν ∈ A ∩ I. Since 0 < r−1(s(µ)) < ∞ we have that ψ(sµs∗ ν ) = ψn+1(sµs∗ ν) = Xe∈r−1(s(µ)) ψn+1(sµes∗ νe) = Xe∈r−1(s(µ)) [µe = νe]c(µe)−βτs(µe) = = Xe∈r−1(s(µ)) [µ = ν]c(µ)−β c(e)−βτs(e) = [µ = ν]c(µ)−β Xe∈r−1(s(µ)) c(e)−βτs(e) = = [µ = ν]c(µ)−βFc,β(τ )(ps(µ)) = [µ = ν]c(µ)−β τ (ps(µ)) = ψn(sµs∗ ν ) = ϕn(sµs∗ ν). Again, using the approximate unit given by Lemma 2.10, if sµs∗ ν ∈ Cn, then ψ(sµs∗ ν ) = ψn+1(sµs∗ ν) = lim λ→∞ ψn+1(sµs∗ ν uλ) = lim D→r≤n+1−ν(s(µ))Xζ∈D ψn+1(sµζ s∗ νζ) = [µζ = νζ]c(νζ)−β τs(νζ) = = lim D→r≤n+1−ν(s(ν))Xζ∈D D→r≤n+1−ν(s(ν))Xζ∈D lim = [µ = ν]c(ν)−β c(ζ)−βτs(ζ) = = [µ = ν]c(ν)−β lim D→r≤n+1−ν(s(ν))Xζ∈D c(ζ)−β τs(ζ) ≤ ≤ [µ = ν]c(ν)−β τs(ν) = ϕn(sµs∗ ν), where the inequality is given by Lemma 3.7, which is a consequence of (K2). By the description of the core C∗(E)γ as an inductive limit of the Cn, we can define a state ω as the inductive limit of ϕn. By construction, ω satisfies (3.1) and, since each ϕn is uniquely defined by (3.1), so is ω. Finally, it is easily seen that the correspondence built preserves convex combi- nations by construction. (cid:3) KMS STATES FOR THE GENERALIZED GAUGE ACTION ON GRAPH ALGEBRAS 11 4. Ground states In this section, we let a function c : E1 → R∗ + be given and define a one- parameter group of automorphisms σ as in the last section. The following definition of a ground state will be used [?]. Definition 4.1. We say that φ is a σ-ground state if for all a, b ∈ C∗(E)a, the entire analytic function ζ 7→ φ(aσζ (b)) is uniformly bounded in the region {ζ ∈ C : Im(ζ) ≥ 0}, where C∗(E)a is the set of analytic elements for σ. Proposition 4.2. If τ is a tracial state on C0(E0) such that supp(τ ) ⊆ {v ∈ E0 : v is singular} then there is a unique state φ on C∗(E) such that (i) φ(pv) = τ (pv) for all v ∈ E0; (ii) φ(sµs∗ ν ) = 0 if µ > 0 or ν > 0. Proof. First, observe that a state φ satisfying (ii) is uniquely determined by its values on C∗(E)γ because (ii) implies that φ = φC ∗(E)γ ◦ Φ, where Φ is the conditional expectation given by Proposition 2.9. Given τ as in the statement of the proposition, a state ω on C∗(E)γ can be built in the same way as in the proof of Theorem 3.10. For each n, use Lemma 3.8 with A = Cn, B = Cn+1, I = Fn+1, ψn ≡ 0 and ϕn is given by the previous step, where for the first step we have ϕ0 = τ . For ω = lim−→ ϕn, we have that φ = ω ◦ Φ satisfies (i) and (ii) and is unique by construction. (cid:3) Proposition 4.3. If c is such that c(e) > 1 for all e ∈ E1, then a state φ on ν ) = 0 whenever µ > 0 or C∗(E) is a σ-ground state for σ if and only if φ(sµs∗ ν > 0. Proof. If φ is a ground state then for each pair µ, ν ∈ E∗ the function ζ 7→ ν )) is bounded on the upper half of the complex plane. If ζ = x + iy then φ(sµσζ(s∗ φ(sµσζ(s∗ ν )) = φ(sµc(ν)−iζ s∗ ν) = c(ν)y−ixφ(sµs∗ ν) = c(ν)yφ(sµs∗ ν). If ν > 0, we have that c(ν) > 1 and so the only possibility for the above function to be bounded is if φ(sµs∗ It is shown analogously that if µ > 0 then φ(sµs∗ ν) = 0. ν) = 0. For the converse, observe that if µ = ν = 0 then φ(sµσζ (s∗ 1. It can be now readily verified that if φ(sµs∗ then φ is a ground state. ν ≤ ν) = 0 whenever µ > 0 or ν > 0 (cid:3) ν)) = φ(sµs∗ Theorem 4.4. If c is such that c(e) > 1 for all e ∈ E1 then there is a bijective correspondence, given by restriction, between σ-ground states φ and tracial states τ on C0(E0) such that supp(τ ) ⊆ {v ∈ E0 : v is singular}. Proof. This is an immediate consequence of Propositions 4.2 and 4.3. Just note that if φ is a σ-ground state and v ∈ E0 is not singular then φ(pv) = φ  Xe∈r−1(v) ses∗ e  = 0. (cid:3) 12 GILLES G. DE CASTRO AND FERNANDO DE L. MORTARI 5. Examples In this section we give two examples with infinite graphs and study the KMS states on the C*-algebras associated to these graphs. Example 5.1 (The Cuntz algebra O∞). Let E0 = {v} be any unitary set and E1 = {en}n∈N any countably infinite set with r(en) = s(en) = v ∀n ∈ N, then C∗(E) ∼= O∞. If c(en) = e (Euler's number) then we have the usual gauge action. In this case, Fc,β(τ )(pv) = ∞ so that condition (K2) from Theorem 3.10 is not satisfied and we have no KMS states for finite β. Since we have only one state on C0(E0) and v is a singular vertex, by Theorem 4.4 there exists a unique ground state. n=0 a−β P∞ that Fc,β(τ )(pv) = P∞ n converges, then there exists β0 > 0 such thatP∞ Now if c(en) = an where an ∈ (1, ∞) is such that there is β > 0 for which n = 1. Observing and using again the fact that there exists only one state on C0(E0), we conclude from Theorems 3.3 and 3.10 that there is no KMS state for β < β0, there exists a unique KMS state for each β ≥ β0 and, as with the gauge action, there is a unique ground state. n=0 a−β n=0 a−β n Example 5.2 (A graph with infinitely many sources). Let E0 = {vn}n∈N and E1 = {en}n∈N\{0} be countably infinite sets and define r(en) = v0 and s(en) = vn for all n ∈ N \ {0}. Again, let an ∈ (1, ∞), n ∈ N \ {0}, be such that P∞ converges for some β > 0. For n 6= 0 we have that Fc,β(τ )(pvn ) = 0 and for n = 0 we have n τvn . Condition (K1) of Theorem 3.10 is trivially satisfied, n=1 a−β n n=1 a−β Fc,β(τ )(pv0 ) =P∞ and for condition (K2) we need P∞ n=1 a−β n τvn ≤ τv0 . If τv0 > 0, since 0 ≤ τvn ≤ 1 for all n there exists β0 > 0 such thatP∞ τv0 so that (K2) is verified for all β ≥ β0 and so there are infinitely many KMS states. And for β < β0 (K2) is not verified so that there are no KMS states. n=1 a−β0 n τvn = For ground states, since all vertices are singular, we have no restriction on τv0 ; every state τ on C0(E0) gives a ground state on C∗(E). Departamento de Matem´atica, Universidade Federal de Santa Catarina, 88040-970 Florian´opolis SC, Brazil. E-mail address: [email protected], [email protected]
1604.00713	1	1604	2016-04-04T01:20:08	Notes on ergodic theorems in non-commutative symmetric spaces	[ "math.OA" ]	In this paper we establish individual ergodic theorem for positive kernels (or so called Danford Shwartz (DS+) operators acting on non commutative symmetric spaces.	math.OA	math	NOTES ON ERGODIC THEOREMS IN NON-COMMUTATIVE SYMMETRIC SPACES GENADY YA. GRABARNIK Abstract. In this paper we establish individual ergodic theorem for positive kernels (or so called Danford Shwartz (DS+) operators acting on non commu- tative symmetric spaces. 1. Introduction The goal of the paper is to see that one of the results by Veksler [20] or Muratov, Pashkova and Rubshtein [16] remains valid for the non commutative case. 1 Let M, τ be a semifinite von Neumann algebra with semifinite normal faithful trace τ . In addition we assume that M, τ satisfy homogeneity property, it may be pre- sented in the form of the resonant property on trace, see for example [1]. The space of all measurable operators affiliated with M, τ in the Sigal [19] sense is denoted by L0, see for details [19, 5, 6]. Notions of L1(M, τ ) and L∞(M, τ ) was naturally introduced in the same paper. Since we fix algebra M and trace τ , we omit them from the notations from now on. Let F ∈ M be a set of finite linear combinations of orthogonal projections with finite trace. Space R0 = (L1 + L∞)0 is the closure of F in the norm kxk = inf {kx1k1 + kx2k∞, x = x1 + x2, x1 ∈ L1, x2 ∈ L∞}. Remark 1. The space R0 is not necessary separable, it is sufficient to consider M = B(H) of all bounded operators in the Hilbert space H with not separable H with natural trace τ . Definition 1. Non-commutative re-arrangement invariant (or symmetric) space L for the fully symmetric case were introduced by Yeadon in [24]. For the definition of the symmetric spaces we refer to the recent book of Lord, Sukochev, Zanin on singular traces, with original proofs due to Kalton and Sukochev. Definition 2. Re-arrangement invariant space L over (M, τ ) is called minimal if F is dense in L by norm of L. Date: March 1, 2016. 1991 Mathematics Subject Classification. Primary 05C38, 15A15; Secondary 05A15, 15A18. Key words and phrases. Ergodic theory, Operator algebras, re-arrangement invariant spaces. 1 It become known to author that Litvinov and Chilin also saw this result at the same time and wrote it at the same time. I suggested to them to combine results and names on paper. Waiting for the answer. 1 2 GENADY YA. GRABARNIK 2. Embedding theorem The following refinement of the embedding theorem [12] take place. Theorem 1. Let L be a re-arrangement invariant space over (M, τ ). Then (1) If L is minimal, then (2) If L is not minimal L, then L1 ∩ L∞ ⊆ L ⊆ R0 L∞ ⊆ L ⊆ L1 + L∞ Proof. Proof is given in the forthcoming paper of author and some co-authors. (cid:3) 3. Individual ergodic theorem for minimal symmetric spaces 3.1. The Largest Minimal Symmetric Space. The largest minimal symmetric space is R0 = (L1 + L∞)0 is a set of all measurable integrable with trace operators plus bounded with not increasing re-arrangement functions decreasing to 0. The space R0 is minimal symmetric space. Any minimal symmetric space is a subset of R0. 3.2. Positive double contraction on R0. Any positive kernel (T ∈ DS+) leaves R0 invariant. 3.3. Mean Ergodic Theorem. Von Neumann Mean ergodic Theorem on L2 fol- lows from the general von Neumann ergodic theorem for the contractions on the general Hilbert spaces. Mean convergence on L1 ∩L∞ follows from the the fact that both L1 and L∞ are invariant under positive kernels. The space L1 ∩ L∞ itself is invariant under action of positive kernel, and positive kernel is a contraction of the space L1 ∩ L∞. The von Neumann Ergodic theorem and the closedness of L1 ∩ L∞ in its norm implies mean ergodic theorem on L1 ∩ L∞. 3.4. Mean convergence on R0. Proposition 1. The Cesaro averages Sn(T )x converge to some x in norm of R0. Proof. Follows from von Neumann ergodic theorem and the fact that L2 is dense R0. Details. We show that sl(T )x, l = 1, 2, ... is fundamental sequence in R0. Indeed, each x ∈ R0 may be presented as x = x1,n + x2,n with x1,n ∈ L1 and x2,n ∈ L∞ with kx2,nk∞ < 2−n. Then we can apply von Neumann or Yeadon's Mean Ergodic theorem 4.2 [24] for L1 to x1,n and find l(n) such that for l, m ≥ l(n) holds ksl(T )x1,n − sm(T )x1,nkL1 < 2−n. Then for l, m ≥ l(n) ksl(T )x − sm(T )xkR0 ≤ ksl(T )x1,n − sm(T )x1,nkL1+ +ksl(T )x2,n − sm(T )x2,nkL∞ ≤ 4 ∗ 2−n, and, hence, the sequence sl(T )x is fundamental. Completeness of R0 implies exis- tence of x liml→∞ sl(T )x. (cid:3) NOTES ON NON-COMMUTATIVE ERGODIC THEOREMS 3 Remark 2. Mean ergodic theorem for fully symmetric spaces is due to Yeadon [24], Theorem 4.2. Note that we do not require the space L to be fully symmetric here. Condition ii) of the theorem 4.2 [24] means that the space L is minimal. The space R0 does not satisfy condition iii) in theorem 4.2 [24]. 3.5. Individual Ergodic Theorem in L1. Individual Ergodic theorem for L1 was established by Yeadon [22], among other authors. 3.6. Individual ergodic theorem for R0. The goal of the section is to show double side almost everywhere convergence for the operators from the R0. Definition 3. The sequence xn from L0 is called converging double side almost everywhere to x0 ∈ L0 if for every ǫ > 0 there exist orthogonal projection E ∈ M such that τ (1 − E) < ǫ, E(xn − x0)E ∈ M and E(xn − x0)E → 0. Theorem 2. Let M, τ, R0 are as above and T is positive kernel on M . For x ∈ R0, Cesaro averages Sn(T )x converge d.s.a.e. in R0. Proof. The proof follows the line of the proof of Proposition 1 and uses Yeadon's individual ergodic theorem for L1 [22], see also Chilin, Litvinov [?], [?] We show that sl(T )x, l = 1, 2, ... is fundamental d.s.a.e. sequence in R0. Indeed, each x ∈ R0 may be presented as x = x1,n + x2,n with x1,n ∈ L1 and x2,n ∈ L∞ with kx2,nk∞ < 2−4∗n. In turn, x1,n = x1,1,n + x1,2,n, with x1,1,n ∈ L∞, x1,2,n ∈ L1 and kx1,2,nkL1 < 2−8∗n, n = 1, 2, .... Then we can apply Yeadon's Individual Ergodic theorem 1, [22] for L1 to x1,1,n and find l(n) and projector E(n) ∈ M such that τ (1 − E(n)) < 2−4∗n and for l, m ≥ l(n) holds kE(n)(sl(T )x1,1,n − sm(T )x1,1,n)E(n)kL∞ } → 0. We can represent x1,2,n = P∞ k=1 x2,n,k, with x2,n,k ∈ L∞ and kx2,n,kkL1 < 2−8∗(n+k). Then, we can find E(1, n) = ∧kE(1, n, k), with τ (1 − E(1, n)) ≤ 2−4∗n and E(1, n)sl(x1,2,n)E(1, n) ∈ L∞ and kE(1, n)sl(x2,n)E(1, n)kL∞ < 2−4∗n, where projections E(1, n, k) are obtain by Theorem 1 from [22] applied to x2,n,k and ǫ = 2−4(n+k). By choosing E(2, n) = ∧∞ k=1E(1, n + k) ∧ ∧∞ k=1E(n + k), we have τ (1 − E(2, n)) < 2−4∗n. Moreover, for l, m ≥ l(n) kE(2, n)(sl(T )x − sm(T )x)E(2, n)kL∞ ≤ kE(2, n)(sl(T )x1,n−sm(T )x1,n)E(2, n)kL∞+kE(2, n)(sl(T )x2,n−sm(T )x2,n)E(2, n)kL∞ ≤ kE(2, n)(sl(T )x1,1,n − sm(T )x1,1,n)E(2, n)kL∞ + kE(2, n)(sl(T )x1,2,nE(2, n)kL∞ +kE(2, n)sm(T )x1,2,nE(2, n)kL∞ + 2 ∗ 2−4∗n ≤ 8 ∗ 2−n, and, hence, the sequence sl(T )x is fundamental d.s.a.e. The sequence sl(T )x also converges in norm in R0 to x, hence it converges in measure. This implies conver- gence of sl(T )x d.s.a.e. to x. (cid:3) Remark 3. In the case when space L is not minimal, it contains M . Then it is possible to show [16], that even in the commutative case, there exists ergodic automorphism of the space with measure such that ergodic averages do not converge almost everywhere. 4 GENADY YA. GRABARNIK Corollary 1. Let L be a minimal non-commutative symmetric space. Let T be a positive kernel such that T leaves L invariant and T acts as contraction on L. Then Cesaro averages sn(T )x converge in norm and d.s.a.e. for any x ∈ L. Proof. Since the space L is minimal, the set of L1 ∩ L∞ is dense in L. Since L1 ∩ L∞ ⊆ L, hence kxkL1∩L∞ ≥ C ∗ kxkL for any x ∈ L1 ∩ L∞, which in turn implies convergence of Cesaro averages of sn(T )x in norm of L for x ∈ L1 ∩ L∞. Fix real ǫ > 0. Find xk ∈ L1 ∩ L∞ with kx − xkkL < ǫ/2. Then Cesaro averages are sn(T )x are within ǫ of the xk, where xk = limn→∞ sn(T )xk. This implies norm convergence of sn(T )x. The d.s.a.e. convergence follows from the embedding theorem 1, since a minimal re-arrangement invariant non-commutative function space L is a subspace of R0. (cid:3) Corollary 2. Let L be a minimal fully symmetric space. Let T be a positive kernel on (M, τ ). Then T leaves L invariant and act on L as contraction. Moreover, the Cesaro averages Sn(T )x converge d.a.e. for any x ∈ L. Corollary 3. (see Chilin Litvinov [3]). Let LΨ be a non-commutative Orlicz space with function Ψ satisfying conditions δ2 and ∆2. Let T be a positive kernel. Then the Cesaro averages Sn(T )x converge d.s.a.e. for any x ∈ L. Proof. The Orlicz space LΨ with function Ψ satisfying conditions δ2 and ∆2 is minimal [1, 12, 15]. Since the space LΨ is fully symmetric, it is interpolation space [1, 12] and hence T leaves LΨ invariant and acts on LΨ as a contraction. Then we are in the assumptions of the Corollary 2, and hence Sn(T )x converges d.a.e. . (cid:3) References [1] C. Bennett, R. Sharpley, Interpolation of Operators, Academic Press Inc. (London) LTD, 1988. [2] V. Chilin, S. Litvinov, On pointwise ergodic theorems for infinite measure, arhiv, 2015 [3] V. Chilin, S. Litvinov, Individual ergodic theorems in noncommutative Orlicz spaces, arhiv, 2016 [4] V. I. Chilin, F. A. Sukochev, Weak convergence in non-commutative symmetric spaces, J. Operator Theory, 31 (1994), 35-65. [5] P. G. Dodds, T. K. Dodds, and B. Pagter, Fully symmetric operator spaces, Integr. Equat. Oper. Theory, 15 (1992), 942-972. [6] P. G. Dodds, T. K. Dodds, and B. Pagter, Noncommutative Kothe duality, Trans. Amer. Math. Soc., 339(2) (1993), 717-750. [7] P. G. Dodds, T. K. Dodds, F. A. Sukochev, and O. Ye. Tikhonov, A Non-commutative Yoshida-Hewitt theorem and convex sets of measurable operators closed locally in measure, Positivity, 9 (2005), 457-484. [8] P. G. Dodds, B. Pagter and F. A. Sukochev, Sets of uniformly absolutely continuous norm in symmetric spaces of measurable operators, Trans. Amer. Math. Soc., (2015). [9] N. Dunford and J. T. Schwartz, Linear Operators, Part I: General Theory, John Willey and Sons, 1988. [10] T. Fack, H. Kosaki, Generalized s-numbers of τ -mesaurable operators, Pacific J. Math., 123(1986), 269-300. [11] R. V. Kadison, A generalized Schwarz inequality and algebraic invariants for operator alge- bras, Ann. of Math. (2), 56(1952), 494-503. [12] S. G. Krein, Ju. I. Petunin, and E. M. Semenov, Interpolation of Linear Operators, Translations of Mathematical Monographs, Amer. Math. Soc., 54, 1982. [13] J. Lindenstraus, L. Tsafriri, Classical Banach spaces I-II, Springer-Verlag, Berlin Heidel- berg New York. 1977. NOTES ON NON-COMMUTATIVE ERGODIC THEOREMS 5 [14] E. Nelson, Notes on non-commutative integration, J. Funct. Anal., 15 (1974), 103-116. [15] B.Z. Rubshtein, G. Ya. Grabarnik, M. A. Muratov, Pashkova, SSMF, in press, 2017 [16] M. A. Muratov, J. Pashkova, B-Z. Rubshtein Order Convergence Ergodic Theorems in Re- arrangement Invariant Spaces, Operator Methods in Mathematical Physics: Conference on Operator Theory, Analysis and Mathematical Physics, 2013 [17] B.Z. Rubshtein, M. A. Muratov, Pashkova, Embedding theorem, oral communication, 2017 [18] S. Sakai, C ∗ −algebras, Springer-Verlag, Berlin Heidelberg New York, −algebras and W ∗ 1971. [19] I. E. Segal, A non-commutative extension of abstract integration, Ann. of Math., 57 (1953), 401-457. [20] A. Veksler, An ergodic theorem in symmetric spaces, Subirsk. Mat. Zh, 24 (1985), 189-191 (in Russian). [21] H. Yanhou, T. N. Bekjan, The dual on noncommutative Lorentz spaces, Acta Math. Sci., 31 B(5)(2011), 2067-2080. [22] F. J. Yeadon, Non-commutative Lp −spaces, Math. Proc. Camb. Phil. Soc., 77(1975), 91-102. [23] F. J. Yeadon, Ergodic theorems for semifinite von Neumann algebras-I, J. London Math. Soc., 16 (2)(1977), 326-332. [24] F. J. Yeadon, Ergodic theorems for semifinite von Neumann algebras: II, Math. Proc. Camb. Phil. Soc., 88 (1980), 135-147. (GYaG) St Johns University, Queens, NY, USA E-mail address, GYaG: [email protected]
1611.01632	1	1611	2016-11-05T11:17:59	Characterizations of 2-local derivations and local Lie derivations on some algebras	[ "math.OA" ]	We prove that every 2-local derivation from the algebra $M_n(\mathcal{A})(n>2)$ into its bimodule $M_n(\mathcal{M})$ is a derivation, where $\mathcal{A}$ is a unital Banach algebra and $\mathcal{M}$ is a unital $\mathcal{A}$-bimodule such that each Jordan derivation from $\mathcal{A}$ into $\mathcal{M}$ is an inner derivation, and that every 2-local derivation on a C*-algebra with a faithful traceable representation is a derivation. We also characterize local and 2-local Lie derivations on some algebras such as von Neumann algebras, nest algebras, Jiang-Su algebra and UHF algebras.	math.OA	math	Characterizations of 2-local derivations and local Lie derivations on some algebras Jun He∗, Jiankui Li, Guangyu An, and Wenbo Huang Department of Mathematics, East China University of Science and Technology Shanghai 200237, China Abstract We prove that every 2-local derivation from the algebra Mn(A)(n > 2) into its bimodule Mn(M) is a derivation, where A is a unital Banach algebra and M is a unital A-bimodule such that each Jordan derivation from A into M is an inner derivation, and that every 2-local derivation on a C-algebra with a faithful traceable representation is a derivation. We also characterize local and 2-local Lie derivations on some algebras such as von Neumann algebras, nest algebras, Jiang-Su algebra and UHF algebras. Keywords: 2-local derivation, local Lie derivation, 2-local Lie derivation, matrix algebra, von Neumann algebra Mathematics Subject Classification(2010): 46L57; 47B47; 47C15 1 Introduction Let A be a Banach algebra and M an A-bimodule. We recall that a linear map D : A → M is called a derivation if D(ab) = D(a)b + aD(b) for all a, b ∈ A, and a linear map D : A → M is called a Jordan derivation if D(a2) = D(a)a + aD(a) for all a ∈ A. A derivation Da defined by Da(x) = ax − xa for all x ∈ A is called an inner derivation, where a is a fixed element in M. In [17], R. Kadison introduces the concept of local derivation in the following sense: a linear mapping T from A into M such that for every a ∈ A, there exists a derivation Da : A → X , depending on a, satisfying T (a) = Da(a). Also in [17], the author proves ∗Corresponding author. E-mail address: [email protected] 1 that each continuous local derivation from a von Neumann algebra into its dual Banach module is a derivation. B. Jonson [16] extends the above result by proving that every local derivation from a C-algebra into its Banach bimodule is a derivation. Based on these results, many authors have studied local derivations on operator algebras, for example, see in [12, 14, 19, 27]. In [28] P. Semrl introduces the concept of 2-local derivations. Recall that a map ∆ : A → M (not necessarily linear) is called a 2-local derivation if for each a, b ∈ A, there exists a derivation Da,b : A → M such that ∆(a) = Da,b(a) and ∆(b) = Da,b(b). Moreover, the author proves that every 2-local derivation on B(H) is a derivation. In [18] S. Kim and J. Kim give a short proof of that every 2-local derivation on the algebra Mn(C) is a derivation. Later J. Zhang and H. Li [30] extend the above result for arbitrary symmetric digraph matrix algebras and construct an example of 2-local derivation which is not a derivation on the algebra of all upper triangular complex 2 × 2 matrices. In [2], S. Ayupov and K. Kudaybergenov suggest a new technique and prove that every 2-local derivation on B(H) is a derivation for arbitrary Hilbert space H. Then they consider the cases for several kinds of von Neumann algebras in succession in [3, 4, 5], and finally prove that any 2-local derivation on arbitrary von Neumann algebra is a derivation. Quite recently, in [6] the authors study the case for matrix algebras over unital semiprime Banach algebras, and prove that every 2-local derivation on the algebra M2n(A), n ≥ 2, is a derivation, where A is a unital semiprime Banach algebra with the inner derivation property. In Section 2, we improve the above result([6, Theorem 2.1]) for arbitrary n > 2. More specifically, we prove that if A is a unital Banach algebra and M is a unital A- bimodule such that each Jordan derivation from A into M is an inner derivation, then each 2-local derivation ∆ from Mn(A)(n > 2) into Mn(M) is a derivation. Moreover, if we only consider the case M = A, then the assumption of innerness can be relaxed to spatial innerness. That is, if A is a unital Banach algebra such that each Jordan derivation on A is a derivation and for each derivation D on A, there exists an element a in B such that D(x) = [a, x] for all x ∈ A, where B is an algebra containing A, then each 2-local derivation on Mn(A)(n > 2) is a derivation. Based on these, we obtain some applications. We also prove that every 2-local derivation on a C-algebra with a faithful traceable representation is a derivation. Recall that a linear map ϕ : A → A is called a Lie derivation if ϕ[a, b] = [ϕ(a), b] + [a, ϕ(b)], for all a, b ∈ A, where [a, b] = ab−ba is the usual Lie product. A Lie derivation ϕ is said to be standard if it can be decomposed as ϕ = D + τ , where D is a derivation on A and τ is a linear map from A into the center of A such that τ [a, b] = 0 for all a, b ∈ A. It is natural to ask under which conditions each Lie derivation is standard. This 2 problem has been studied by many authors. M. Mathieu and A. Villena [26] prove that each Lie derivation on a C-algebra is standard. W. Cheung [10] characterizes Lie derivations on triangular algebras. F. Lu [24] proves that each Lie derivation on a CDCSL(completely distributed commutative subspace lattice) algebra is standard. Obviously, one can define local and 2-local Lie derivations in a similar way as the local and 2-local derivations. In [9], L. Chen, F. Lu and T. Wang prove that every local Lie derivation on B(X) is a Lie derivation, where X is a Banach space of dimension exceeding 2. Also in this paper [9], the authors characterize 2-local Lie derivations on B(X). Later, L. Chen, and F. Lu [8] prove that every local Lie derivation from a nest algebra algN into B(H) is a Lie derivation, where N is a nest on the Hilbert space H. Quite recently, L. Liu [21] characterizes 2-local Lie derivations on a semi-finite factor von Neumann algebra with dimension greater than 4. In Section 3, we study local and 2-local Lie derivations on some algebras by using a new technique. On the algebras including factor von Neumann algebras, finite von Neumann algebras, nest algebras, UHF(uniformly hyperfinite) algebras, and the Jiang- Su algebra, we prove that every local Lie derivation is a Lie derivation. On the algebras including factor von Neumann algebras, UHF algebras, and the Jiang-Su algebra, we prove that every 2-local Lie derivation is a Lie derivation. Besides, for a finite von Neumann algebra A which is not a factor, we construct an example of 2-local Lie derivation but not a Lie derivation on A. 2 2-Local derivations Through this paper, we denote by Mn(A) the set of all matrices (xi,j)n×n, where xi,j ∈ A. Clearly, if A is a Banach algebra, then so is Mn(A), and if M is an A- bimodule, then Mn(M) is also an Mn(A)-bimodule. Theorem 2.1. Let A be a unital Banach algebra and M be a unital A-bimodule. If each Jordan derivation from A into M is an inner derivation, then each 2-local derivation ∆ from Mn(A)(n > 2) into Mn(M) is a derivation. Let {ei,j}n the linear span of the set {ei,j}n matrices in Mn(A) by Dn(A). i,j=1 be the system of matrix units in Mn(A) and denote by span{ei,j}n i,j=1 i,j=1. Besides, we denote the subalgebra of all diagonal To prove Theorem 2.1, we need several lemmas. Firstly, we give the following two lemmas. Since the proofs are completely similar as the proofs of Lemma 2.2 and Lemma 2.9 in [6], we omit it. Lemma 2.2. Let A be a unital Banach algebra and M be a unital A-bimodule. If each derivation from A into M is an inner derivation, then each derivation from Mn(A) 3 into Mn(M) is also an inner derivation. Lemma 2.3. Let A be a unital Banach algebra and M be a unital A-bimodule. If each Jordan derivation from A into M is an inner derivation, then for each 2-local derivation ∆ from Mn(A)(n > 2) into Mn(M), there exists an element a ∈ Mn(M) such that ∆Dn(A) = DaDn(A) and ∆span{ei,j}n = Daspan{ei,j}n . i,j=1 i,j=1 In following Lemmas 2.4-2.7, the conditions of Theorem 2.1 hold. We consider a 2- = local derivation δ from Mn(A) into Mn(M) such that δDn(A) = 0 and δspan{ei,j }n 0. i,j=1 In addition, from the definition, it is easy to see that δ is homogeneous, i.e. δ(λx) = λδ(x), for each x ∈ Mn(A) and λ ∈ C. Thus, if necessary, we can assume that kxi,jk < 1, and so e + xi,j is invertible in A, where 1 ≤ i, j ≤ n, xi,j is the (i, j)− entry of x and e is the unit of A. For each x = (xi,j)n×n ∈ Mn(A) , we denote by dxi,j the matrix in Mn(A) such that the (i, j)− entry is xi,j and the others are zero, i.e. dxi,j = ei,ixej,j =   0 . . . 0 ... ... 0 . . . xi,j ... ... 0 0 . . . . . . . . . . . .   0 ... 0 ... 0 Lemma 2.4. For all x ∈ Mn(A), [δ(x)]i,i = 0, i = 1, 2, · · ·, n. Proof. Let y = ei,i + cxi,i. By lemma 2.2, there exists an element a in Mn(M) such that δ(x) = [a, x], δ(y) = [a, y]. Since y ∈ Dn(A), we have δ(y) = 0. For any k 6= i, 0 = [δ(y)]i,k = [a, y]i,k = nX j=1 (ai,jyj,k − yi,jaj,k) = −yi,iai,k = −(e + xi,i)ai,k. Since e + xi,i is invertible, it follows that ai,k = 0. Similarly, ak,i = 0, for any k 6= i. Now we can obtain [δ(x)]i,i = [a, x]i,i = nX k=1 (ai,kxk,i − xi,kak,i) = [ai,i, xi,i] = [a, y]i,i = 0. The proof is complete. Lemma 2.5. For each x ∈ Mn(A), let   0 ... xi,1 ... 0 y = . . . . . . x1,j ... xi,j ... . . . xn,j 4 . . . 0 ... . . . xi,n ... 0 . . .   i.e. y = Pn k=1(dxi,k + dxk,j) −dxi,j. Then [δ(x)]i,j = [δ(y)]i,j . Proof. Take an element a from Mn(M) such that δ(x) = [a, x], and δ(y) = [a, y]. It is easy to verify the above result by direct calculation. Lemma 2.6. For all x ∈ Mn(A), [δ(x)]i,j = [δ(dxi,j )]i,j, whenever i 6= j. Proof. Let By Lemma 2.5, [δ(x)]i,j = [δ(y)]i,j . Let s 6= i, and z = y − dxs,j + es,i, i.e. y = nX k=1 (dxi,k + dxk,j) −dxi,j.   0 ... xi,1 ... 0 ... 0 z = . . . 0 ... . . . xi,i ... e ... 0 . . . . . . . . . . . . x1,j ... xi,j ... 0 ... . . . xn,j . . . . . . 0 ... . . . xi,n ... 0 ... 0 . . . . . .   Take an element a in Mn(M) such that δ(y) = [a, y], δ(z) = [a, z]. Then [δ(y)]i,i − [δ(z)]i,i = [a, y − z]i,i = [a,dxs,j − es,i]i,i = −ai,s. By Lemma 2.4, [δ(y)]i,i = [δ(z)]i,i = 0, and so ai,s = 0. Thus [δ(y)]i,j − [δ(z)]i,j = [a, y − z]i,j = [a,dxs,j − es,i]i,j = ai,sxs,j = 0, i.e. [δ(y)]i,j = [δ(z)]i,j . Let w = z − es,i =   0 ... xi,1 ... 0 ... 0 . . . 0 ... . . . xi,i ... 0 ... 0 . . . . . . . . . . . . x1,j ... xi,j ... 0 ... . . . xn,j . . . . . . 0 ... . . . xi,n ... 0 ... 0 . . . . . .   again by Lemma 2.5, we have [δ(z)]i,j = [δ(w)]i,j . Now we have [δ(y)]i,j = [δ(y−dxs,j)]i,j. Repeating the above steps, we can obtain that [δ(y)]i,j = [δ(u)]i,j , where u = nX k=1dxi,k =   0 ... xi,1 ... 0 . . . 0 ... . . . xi,n ... 0 . . .   . . . 0 ... . . . xi,j ... 0 . . . 5 Similarly, we can show that [δ(u)]i,j = [δ(u −dxi,s + ej,s)]i,j = [δ(u −dxi,s)]i,j , f ors 6= j. It follows that [δ(u)]i,j = [δ(dxi,j)]i,j by taking other indexes successively. Hence [δ(x)]i,j = [δ(y)]i,j = [δ(u)]i,j = [δ(dxi,j )]i,j . The proof is complete. Lemma 2.7. For all x ∈ Mn(A), [δ(x)]i,j = 0, whenever i 6= j. Proof. Assume that i < j. Let v = n−1X k=1 ek+1,k =   0 e 0 0 e ... 0 0 . . . . . . . . . e   0 It is not difficult to check that every matrix a in Mn(M) commuting with v satisfies the following properties: ak,k = as,s, ak,s = 0, whenever k < s. Let y be a matrix in Mn(A) such that yi,i = e+ xi,j, yi,j = xi,j and the other entries of y are all zero. Take an element a from Mn(M) such that δ(y) = [a, y], δ(v) = [a, v]. Since v ∈ span{ei,j}n i,j=1, δ(v) = [a, v] = 0. Thus ai,i = aj,j, and ai,j = 0. Since yj,i = 0, by Lemma 2.6, [δ(y)]j,i = 0. Then 0 = [δ(y)]j,i = [a, y]j,i = aj,iyi,i = aj,i(e + xi,j). Since e + xi,j is invertible, it implies that aj,i = 0. Thus [δ(y)]i,i = [a, y]i,i = ai,iyi,i − yi,iai,i − yi,jaj,i = ai,ixi,j − xi,jai,i. Therefore [δ(y)]i,j = [a, y]i,j = ai,iyi,j − yi,iai,j − yi,jaj,j = ai,ixi,j − xi,jai,i = [δ(y)]i,i = 0. Again by Lemma 2.6, [δ(x)]i,j = [δ(y)]i,j = 0. Similarly we can show that [δ(x)]i,j = 0 if i > j. The proof is complete. Now we are in position to prove Theorem 2.1. 6 Proof of Theorem 2.1. According to Lemma 2.3, there exists a derivation Da such . Let δ = ∆ − Da. that ∆Dn(A) = DaDn(A) and ∆span{ei,j}n = Daspan{ei,j}n Then by Lemmas 2.4 and 2.7, we have δ = 0, i.e. ∆ = Da. It follows that ∆ is a derivation. i,j=1 i,j=1 The following corollaries are some specific examples for applying Theoerm 2.1. Corollary 2.8. Let A be a unital commutative C-algebra and M be a unital A- bimodule. Then each 2-local derivation from Mn(A)(n > 2) into Mn(M∗) is a deriva- tion. Proof. Since M is a unital A-bimodule, so is M∗. We know A is amenable, that is each derivation from A into M∗ is an inner derivation [29]. Besides, there is a classic result that each Jordan derivation from a C-algebra into its bimodule is a derivation (see [15]). By Theorem 2.1, the proof is complete. Corollary 2.9. Let A be a C-algebra. Then each 2-local derivation from Mn(A)(n > 2) into Mn(A∗) is a derivation. Proof. Every C-algebra A is weakly amenable, that is, each derivation from A into A∗ is an inner derivation [29]. By Theorem 2.1, the proof is complete. Corollary 2.10. Each 2-local derivation on the matrix algebra Mn(A)(n > 2) is a derivation, if the algebra A satisfies one of the following conditions: (1) A = algL, where L is a subspace lattice on a Hilbert space H with the property that 0+ 6= 0 or H− 6= H; (2) A = alg(N1N N2N · · ·N Nn), where each Ni is a nest, i = 1, 2, · · ·, n. Proof. (1) According to [22, Theorem 2.1], each Jordan derivation on A is a derivation. And see in [20] , each derivation on A is an inner derivation. (2) It is known that N1N N2N · · ·N Nn is still a CSL, so Jordan derivations on A are derivations [23]. In [13], F. Gilfeather, A. Hopenwasser and D. Larson prove that the cohomology vanishes for CSL algebras whose lattices are generated by finite independent nests. Applying this result, we immediately obtain that derivations on A are inner derivations. By Theorem 2.1, the proof is complete. Corollary 2.11. Let A be a nest algebra of infinite multiplicity. Then each 2-local derivation on A is a derivation. Proof. Since A is of infinite multiplicity, it is isomorphic to Mn(B) for some nest algebra B and some integer n > 2. It is known that each derivation on nest algebras is an inner derivation [11]. By Theorem 2.1, the proof is complete. 7 Lemma 2.12. Let A = L∞ i=1 Ai be a Banach algebra with the inner derivation prop- erty, i.e. all derivations on A are inner derivations. If each 2-local derivation on Ai is a derivation for any i ∈ N, then each 2-local derivation on A is also a derivation. Proof. We only consider the case A = A1L A2. For a 2-local derivation δ on A, denote by δi the restriction of δ in Ai, i = 1, 2. One can easily verify that δ(ai) = δi(ai) ∈ Ai for each ai ∈ Ai, moreover, δi is a 2-local derivation on Ai, and thus a derivation. For each a ∈ A, we write a = a1 + a2, where ai ∈ Ai, i = 1, 2. By the definition of 2- local derivation, there exists an element x ∈ A such that δ(a) = [a, x] = [a1, x1]+[a2, x2] and δ(a1) = [a1, x] = [a1, x1]. Similarly, there exists an element y ∈ A such that δ(a) = [a, y] = [a1, y1] + [a2, y2] and δ(a2) = [a2, y] = [a2, y2]. So we can obtain δ(a) = δ(a1) + [a2, x2] = [a1, y1] + δ(a2). It follows that δ(a1) = [a1, y1], δ(a2) = [a2, x2] and thus δ(a) = δ(a1) + δ(a2). Hence δ is linear. Moreover, we can obtain δ(ab) = δ(a1b1 + a2b2) = δ(a1b1) + δ(a2b2) = δ(a1)b1 + a1δ(b1) + δ(a2)b2 + a2δ(b2) = (δ(a1) + δ(a2))(b1 + b2) + (a1 + a2)(δ(b1) + δ(b2)) = δ(a)b + aδ(b) Hence δ is a derivation on A. The proof is complete. Corollary 2.13. Let A be a von Neumann algebra without direct summand of type I1 and I2. Then each 2-local derivation on A is a derivation. Proof. It is well known that A = L∞ i=1 Ai, where each Ai is isomorphic to a matrix algebra Mn(B) for some von Neumann algebra B and some integer n > 2. By Theorem 2.1 and Lemma 2.12, the result follows. Remark 2.14. In [4] and [5], the authors show that each 2-local derivation on a von Neumann algebra is a derivation in other ways. By comparison, our proof is more simple. However, we can not handle the case for type I2. In addition, for the case of type I1, the result is trival, since each derivation on an abelian von Neumann algebra is zero. Corollary 2.15. Let A be a unital algebra with the inner derivation property and n ≥ 6 be a positive integer but not a prime number. Then each 2-local derivation on the matrix algebra Mn(A) is a derivation. 8 Proof. Suppose n = rt, where r > 2 and t > 1. Then Mn(A) is isomorphic to Mr(Mt(A)). In [1], the author proves that each Jordan derivation on Mt(A)(t > 1) is a derivation. By Theorem 2.1, the proof is complete. Through the same technique with the proof of Theorem 2.1, we can show the fol- lowing theorem. Theorem 2.16. Let A be a unital Banach algebra such that: (1) each Jordan derivation on A is a derivation; (2) for each derivation D on A, there exists an element a in B such that D(x) = [a, x] for all x ∈ A, where B is an algebra containing A. Then each 2-local derivation on Mn(A)(n > 2) is a derivation. Corollary 2.17. Let A be a C-algebra. Then each 2-local derivation on Mn(A)(n > 2) is a derivation. Let A be a C-algebra, π a representation of A, and M the von Neumann algebra generated by π(A). Then π is said to be a traceable representation if there exists a faithful normal trace τ on M+ such that π(A)T Mτ is weakly dense in M, where Mτ denotes the span of the set {M ∈ M+ : τ (M ) < ∞}. Especially, a finite representation the von Neumann algebra generated by π(A) is finite) can be regarded as a π(i.e. specific traceable representation. Theorem 2.18. Let A be a C-algebra with a faithful traceable representation π. Then each 2-local derivation on A is a derivation. Proof. Let M be the von Neumann algebra generated by π(A), and τ a faithful normal trace on M+ such that Aτ is weakly dense in M, where Aτ denotes π(A)T Mτ . It is known that Mτ is a two-side ideal of M, thus Aτ is also a two-side ideal of π(A). For a 2-local derivation ∆ on A, define δ = π ◦∆◦π−1. Then δ is a 2-local derivation on π(A). For each x ∈ π(A) and y ∈ Aτ , there exists an element m ∈ M such that δ(x) = [m, x] and δ(y) = [m, y]. Hence τ (δ(x)y) = τ ((mx − xm)y) = τ (mxy) − τ (xmy) = τ (xym) − τ (xmy) = τ (x(ym − my)) = −τ (xδ(y)). For each a, b ∈ π(A) and x ∈ Aτ , we have τ (δ(a + b)x) = −τ ((a + b)δ(x)) = −τ (aδ(x)) − τ (bδ(x)) = τ (δ(a)x) + τ (δ(b)x) = τ ((δ(a) + δ(b))x). 9 It means that τ (ux) = 0 for any x ∈ Aτ , where u denotes δ(a + b) − δ(a) − δ(b). For any y ∈ Aτ , by taking x = yy∗u∗, we obtain τ (uyy∗u∗) = 0. Since τ is faithful and Aτ is weakly dense, we have uy = 0 and thus u = 0. Hence δ is additive, and thus a local derivation, moreover, a derivation. It follows that ∆ = π−1 ◦ δ ◦ π is a derivation on A. The proof is complete. As direct applications of the above theorem, we have the following corollaries. Corollary 2.19. Each 2-local derivation on a UHF C-algebra is a derivation. Corollary 2.20. Each 2-local derivation on the Jiang-Su algebra is a derivation. 3 Local and 2-local Lie derivations In this section, we study local and 2-local Lie derivations on some algebras. The main results are as follows. Theorem 3.1. Let B be a Banach algebra over C satisfying the following conditions: (1) each Lie derivation on B is standard; (2) each derivation on B is inner; (3) each local derivation on B is a derivation; (4) every nonzero element in Z(B)(the center of B) can not be written in the form Pn i=1[Ai, Bi], where Ai, Bi ∈ B. Then each local Lie derivation ϕ on B is a Lie derivation. Proof. By the definition of local Lie derivation, for any A ∈ B, there exists a Lie derivation ϕA such that ϕA(A) = ϕ(A). Since ϕA is standard, ϕA = δA + τA, where δA is a derivation and τA is a linear mapping from B into Z(B) such that τA[X, Y ] = 0, for all X, Y ∈ B. Thus ϕ(A) = δA(A) + τA(A) = [A, TA] + τA(A) for some TA ∈ B, which means that ϕ has a decomposition at each point. following we prove that the decomposition is unique. Suppose In the ϕ(A) = δ A(A) + τ A(A) = [A, T A] + τ ′ ′ ′ ′ A(A). Then Thus [A, TA] + τA(A) = [A, T ′ A] + τ ′ A(A). [A, TA] − [A, T ′ A] = τ ′ A(A) − τA(A). 10 Since every nonzero element in Z(B) can not be written in the form Pn i=1[Ai, Bi], we obtain [A, TA] − [A, T ′ A] = τ ′ A(A) − τA(A) = 0, and so Now we can define δA(A) = δ A(A), τA(A) = τ ′ ′ A(A). δ(A) = δA(A) = [A, TA] and τ (A) = τA(A), for every A ∈ B. We shall prove δ and τ are linear. For all A, B ∈ B, since ϕ(A + B) = ϕ(A) + ϕ(B), we obtain that [A, TA] + [B, TB] − [A + B, TA+B] = τ (A + B) − τ (A) − τ (B). Hence [A + B, TA+B] = [A, TA] + [B, TB], τ (A + B) = τ (A) + τ (B). It means that δ and τ are additive. Similarly, we can prove that δ and τ are homoge- neous, and so linear. Hence δ is a local derivation, by assumption, a derivation, and τ is a linear mapping from B into Z(B) such that τ [X, Y ] = 0, for all X, Y ∈ B. It follows that ϕ is a Lie derivation. The proof is complete. Corollary 3.2. Each local Lie derivation on a factor von Neumann algebra M is a Lie derivation. Proof. It is known that every element in M can be written in the form λI+Pn Pn every element in M has the form Pn i=1[Ai, Bi], where λ ∈ C and Ai, Bi ∈ M(see, [7]). If the unit I can not be written in the form i=1[Ai, Bi], where Ai, Bi ∈ M, then by Theorem 3.1, the result follows. Otherwise, i=1[Ai, Bi]. Hence each Lie derivation is a deriva- tions, so every local Lie derivation is a Lie derivation. The following result is proved in [8]. We show it by Theorem 3.1 as an alternative method. Corollary 3.3. Each local Lie derivation on a nest algebra A = algN is a Lie deriva- tion. Proof. If A is of infinite multiplicity, then every A ∈ A has the form Pn i=1[Ai, Bi], where Ai, Bi ∈ A([25, Theorem 4.7]). In this case, all Lie derivations are derivations, and so the result is trival. 11 Otherwise, suppose A is not of infinite multiplicity. Then N has a finite-dimensional atom P . It is known that A satisfies the conditions (1) to (3) of Theorem 3.1(see in [24, 11, 14]), so it is sufficient to check the condition (4). We know Z(A) = CI. Assume i=1[Ai, Bi]. It is known that the linear map σ : A → P AP defined by σ(A) = P AP for all A ∈ A is a homomorphism. So we have that the unit I can be written in the form Pn P = P IP = nX i=1 P [Ai, Bi]P = [σ(Ai), σ(Bi)]. nX i=1 It follows that there is no trace on P AP , which contradicts that P AP has finite di- i=1[Ai, Bi]. The proof is mension. Therefore I can not be written in the form Pn complete. Theorem 3.4. Let B be an algebra over C satisfying the following conditions: (1) each Lie derivation on B is standard; (2) each derivation on B is inner; (3) each local derivation on B is a derivation; (4) there exists a center-valued trace on B. Then each local Lie derivation ϕ on B is a Lie derivation. Proof. Since there exists a center-valued trace on B, every nonzero element in Z(B) i=1[Ai, Bi], where Ai, Bi ∈ B. By Theorem 3.1, the can not be written in the form Pn result follows. Corollary 3.5. Each local Lie derivation on a finite von Neumann algebra M is a Lie derivation. Proof. Since M is finite, there exists a central valued trace on M. By Theorem 3.4, the result follows. Remark 3.6. For a properly infinite von Neumann algebra M, since every element i=1[Ai, Bi], where Ai, Bi ∈ M(see, [7]), all Lie derivations are in M has the form Pn derivations, so every local Lie derivation is a Lie derivation. Also as applications of Theorem 3.4, we can immediately obtain the following two corollaries. Corollary 3.7. Each local Lie derivation on a UHF C-algebra A is a Lie derivation. Corollary 3.8. Each local Lie derivation on the Jiang-Su algebra Z is a Lie derivation. Theorem 3.9. Let B be an algebra over C satisfying the following conditions: (1) each Lie derivation on B is standard; (2) each derivation on B is inner; (3) each 2-local derivation on B is a derivation; 12 (4) Z(B) = CI, where Z(B) is the center of B; i=1[Ai, Bi], where λ ∈ C and (5) every element in B can be written in the form λI +Pn Ai, Bi ∈ B. Then each 2-local Lie derivation ϕ on B is a Lie derivation. Proof. Similarly with the foregoing discussion, if I can be written in the formPn then the result is trival. We assume that I can not be written in the form Pn i=1[Ai, Bi]. In this case, we have proved that ϕ has a unique decomposition at each point A in B, i.e. i=1[Ai, Bi], ϕ(A) = ϕA(A) = δA(A) + τA(A), where ϕA is a Lie derivation, δA is a derivation and τA is a linear mapping from B into Z(B) such that τA[X, Y ] = 0, for all X, Y ∈ B. Similarly we can define δ(A) = δA(A) and τ (A) = τA(A), for all A ∈ B. Now it is sufficient to show that δ is a derivation and τ is linear. Given A and B in B, there exists a Lie derivation ϕA,B such that and ϕ(A) = ϕA,B(A) = δA,B(A) + τA,B(A), ϕ(B) = ϕA,B(B) = δA,B(B) + τA,B(B), where δA,B + τA,B is the standard decomposition of ϕA,B. By the uniqueness of the decomposition, δ(A) = δA,B(A) and δ(B) = δA,B(B). Hence δ is a 2-local derivation and thus a derivation. Now Our task is to show that τ is linear. It is easy to see that i=1[Ai, Bi], then the i=1[Ai, Bi], ϕ is homogeneous, and so is τ . If both A and B have the form Pn equation τ (A + B) = τ (A) + τ (B) is trival. Otherwise, suppose A = λI +Pn where λ 6= 0. Take a Lie derivation ϕA,A+B such that ϕ(A) = ϕA,A+B(A) = δA,A+B(A) + τA,A+B(A), and ϕ(A + B) = ϕA,A+B(A + B) = δA,A+B(A + B) + τA,A+B(A + B), where δA,A+B + τA,A+B is the standard decomposition of ϕA,A+B. From τ (A) = τA,B(A) = τA,A+B(A), we have τA,B(I) = τA,A+B(I), and thus τA,B(B) = τA,A+B(B). Now we can obtain that τ (A + B) = τA,A+B(A + B) = τA,A+B(A) + τA,A+B(B) = τ (A) + τA,B(B) = τ (A) + τ (B). Hence τ is additive, and so linear. The proof is complete. 13 Directly applying the above theorem, we can obtain the following corollaries. Corollary 3.10. Each 2-local Lie derivation on the Jiang-Su algebra is a Lie deriva- tion. Corollary 3.11. Each 2-local Lie derivation on a UHF C-algebra is a Lie derivation. Corollary 3.12. Each 2-local Lie derivation on a factor von Neumann algebra is a Lie derivation. Example 3.13. If M is a finite von Neumann algebra but not a factor, then there exists a 2-local Lie derivation on M but not a Lie derivation. Proof. There exist a center-valued trace τ on M and a non-trival central projection P in M. Suppose the Hilbert space that M acts on is H. Take two unital vectors x and y in P H and (I − P )H, respectively. For each element M ∈ M, denote < τ (M )x, x > and < τ (M )y, y > by xM and yM , respectively. Define a functional f on M by follows: f (M ) equals to x2 M yM if yM 6= 0, else 0, and define a mapping δ from M into Z(M) by δ(M ) = f (M )P . Obviously, δ is not linear and thus not a Lie derivation. Now we only need to check that δ is a 2-local Lie for each A, B in M, there exists a Lie derivation φ on M such that derivation, i.e. φ(A) = δ(A) and φ(B) = δ(B). case1 xAyB 6= xByA. Take φ(M ) = (f (A)yB − f (B)yA)xM + (f (B)xA − f (A)xB)yM P. xAyB − xByA It is not difficult to check that φ is a linear mapping from M into Z(M) and φ([X, Y ]) = 0 for all X, Y in M. Hence φ is a Lie derivation. Moreover, φ(A) = δ(A) and φ(B) = δ(B). case2 xAyB = xByA. If both yA and yB are zero, then it is sufficient to take φ ≡ 0. Otherwise, without P . One can easily = 0 and loss of generality, we may assume yA 6= 0, then take φ(M ) = xAxM yA check that φ is a Lie derivation and φ(A) = δ(A). If yB = 0, then xB = xAyB yA φ(B) = 0 = δ(B). If yB 6= 0, then φ(B) = xAxB yA P = x2 B yB P = δ(B). The proof is complete. Acknowledgements. This paper was partially supported by National Natural Science Foundation of China(Grant No. 11371136). References [1] R. Alizadeh, Jordan derivations of full matrix algebras, Linear Algebra Appl. 430 (2009), 574-578 14 [2] S. Ayupov, K. Kudaybergenov, 2-Loacl derivations and automorphisms on B(H), J. Math. Anal. Appl. 395(1)(2012), 15-18. [3] S. Ayupov, K. Kudaybergenov, B. Nurjanov, A. Alauatdinov, Local and 2-local derivations on noncommutative Arens algebras, Mathematica Slovaca 64(2014), 423-432. [4] S. Ayupov, F. Arzikulov, 2-Local derivations on semi-finite von Neumann al- gebras, Glasgow Math. J. 56(2014), 9-12. [5] S. Ayupov, K. Kudaybergenov, 2-Local derivations on von Neumann algebras, Positivity 19(3)(2014), 445-455. [6] S. Ayupov, K. Kudaybergenov, 2-Local derivations on matrix algebras over semi-prime Banach algebras and on AW-algebras, Journal of Physics Conference Series 697(2016). [7] M. Bresar, E. Kissin, S. Shulman, Lie ideals: from pure algebra to C- algebras, J. reine angew. Math. 623 (2008), 73121. [8] L. Chen, F. Lu, Local Lie derivations of nest algebras, Linear Algebra Appl. 475(2015), 62-72. [9] L. Chen, F. Lu, T. Wang, Local and 2-local Lie derivations of operator algebras on Banach spaces, Integr. Equ. Oper. Theory 77(2013), 109-121. [10] W. Cheung, Lie derivations of triangular algebra, Linear Multilinear Algebra 51(2003), 299-310. [11] E. Christensen, Derivations of nest algebras, Math. Ann 229(1977), 155-161. [12] R. Crist, Local derivations on operator algebras, J. Funct. Anal. 135(1996), 72- 92. [13] F. Gilfeather, A. Hopenwasser, D. Larson, Reflexive algebras with finite width lattices: tensor products, cohomology, compact perturbations, J. Funct. Anal. 55(2)(1984), 176-199. [14] D. Hadwin, J. Li, Local derivations and local automorphisms on some algebras, J. Operator Theory 60(1)(2008), 29-44. [15] N. Jacobson, C. Rickart, Jordan homomorphisms of rings, Trans. Amer. Math. Soc. 69(3)(1950), 479-502. [16] B. Johnson, Local derivations on C-algebras are derivations, Trans. Amer. Math. Soc. 353(2001), 313-325. [17] R. Kadison, Local derivations, J. Algebra 130(1990), 494-509. [18] S. Kim, J. Kim, Local automorphisms and derivations on Mn(C), Proc. Amer. Math. Soc. 132(5)(2004), 1389-1392. 15 [19] D. Larson, A. Sourour, Local derivations and local automorphisms, Proc. Sym- pos. Pure Math. 51(1990), 187-194. [20] J. Li, H. Pendharkar, Derivations of certain algebras, Internat. J. Math. 24(5)(2000), 345-349. [21] L. Liu, 2-Local Lie derivations on semi-finite factor von Neumann algebras, Linear Multilinear Algebra 64(2016), 1679-1686. [22] F. Lu, Jordan derivations of reflexive algebras, Integr. Equ. Oper. Theory 67(2010), 51-56. [23] F. Lu, The Jordan structure of CSL algebras, Studia Math. 190(2009), 283-299. [24] F. Lu, Lie derivation of certain CSL algebras, Israel J. Math. 155(2006), 149-156. [25] L. Marcoux, A. Sourour, Conjugation-invariant subspaces and Lie ideals in non-selfadjoint operator algebras, J. London Math. Soc. 65(2)(2002), 493-512. [26] M. Mathieu, A. Villena, The structure of Lie derivations on C-algebras, J. Funct. Anal. 202(2003), 504-525. [27] Y. Pang, W. Yang, Derivations and local derivations on strongly double triangle subspace lattice algebras, Linear Multilinear Algebra 58(2010), 855-862. [28] P. Semrl, Local automorphisms and derivations on B(H), Proc. Amer. Math. Soc. 125(1997), 2677-2680. [29] V. Runde, Lectures on Amenability, I. Springer Verlag, 2002. [30] J. Zhang, H. Li, 2-Loacl derivations on digraph algebras, Acta Math. Sinica, Chinese series 49(2006), 1401-1406. 16
1811.04043	3	1811	2019-05-03T16:19:05	The universal C*-algebra of a contraction	[ "math.OA", "math.FA" ]	We say that a contractive Hilbert space operator is universal if there is a natural surjection from its generated C-algebra to the C-algebra generated by any other contraction. A universal contraction may be irreducible or a direct sum of (even nilpotent) matrices; we sharpen the latter fact and its proof in several ways, including von Neumann-type inequalities for noncommutative -polynomials. We also record properties of the unique C-algebra generated by a universal contraction, and we show that it can be used similarly to C(F_2) in various Kirchberg-like reformulations of Connes' Embedding Problem (some known, some new). Finally we prove some analogous results for universal C-algebras of noncommuting row contractions and universal Pythagorean C*-algebras.	math.OA	math	THE UNIVERSAL C∗-ALGEBRA OF A CONTRACTION KRISTIN COURTNEY AND DAVID SHERMAN Abstract. We say that a contractive Hilbert space operator is universal if there is a natural surjection from its generated C ∗-algebra to the C ∗-algebra generated by any other contraction. A universal contraction may be irreducible or a direct sum of (even nilpotent) matrices; we sharpen the latter fact and its proof in several ways, including von Neumann-type inequalities for noncom- mutative -polynomials. We also record properties of the unique C ∗-algebra generated by a universal contraction, and we show that it can be used simi- larly to C ∗(F2) in various Kirchberg-like reformulations of Connes' Embedding Problem (some known, some new). Finally we prove some analogous results for universal C ∗-algebras of row contractions and universal Pythagorean C ∗- algebras. Contents Introduction 1. 2. Universal contraction operators 3. Von Neumann-type inequalities for noncommutative -polynomials 4. Universal C∗-algebras and projectivity 5. Finite-dimensional and nilpotent representations of A 6. A and Connes' Embedding Problem 6.1. The WEP, LLP, and some results of Kirchberg 6.2. A versus C∗(F2) 7. C∗-algebras generated by multiple universal contractions 7.1. Universal row contractions 7.2. Universal Pythagorean C∗-algebras References 1 3 7 10 12 17 18 20 22 22 25 26 1. Introduction For T a contractive linear operator on a Hilbert space, let C∗(T, I) denote the unital C∗-algebra generated by T and the identity I. We say that T is a universal contraction if for any other contractive Hilbert space operator S there is a unital -homomorphism from C∗(T, I) to C∗(S, I) taking T to S. This is equivalent to requiring that for any noncommutative -polynomial q, the norm of q(T ) is as large Date: May 6, 2019. 2010 Mathematics Subject Classification. 46L05. Key words and phrases. universal contraction, universal C ∗-algebra, von Neumann's inequal- ity, residual finite dimensionality, primitivity, nilpotent operators, Connes Embedding Problem. The research of the first-named author was partially supported by the Eric Nordgren Research Fellowship Fund at the University of New Hampshire. 1 2 KRISTIN COURTNEY AND DAVID SHERMAN as it can be for a contraction. (It should not be confused with "universal (model) operators" as introduced by Rota [47, 48], although we explain a connection in Remark 2.8.) One goal of this paper is to describe universal contractions themselves. On a given separable infinite-dimensional Hilbert space, universal contractions comprise a single approximate unitary equivalence class that is strong* dense in the set of all contractions (Propositions 2.6 and 2.7). But there are qualitative differences among them: a single universal contraction can be irreducible (Theorem 2.5) or a direct sum of nilpotent matrices (Theorem 5.1). The latter result is known, but the existing proof is quite intricate. In Section 5 we give two alternative proofs, and in Section 7.1 we give a generalization. If T1 and T2 are universal contractions, then C∗(T1, I) is -isomorphic to C∗(T2, I) via the map sending T1 to T2. Thus we may call this C∗-algebra (equipped with its distinguished generator) the universal unital C∗-algebra of a contraction. We will denote it here by A, and we will denote the non-unital C∗-algebra C∗(T1) ≃ C∗(T2) by A0. (It is indeed non-unital; see Remark 4.3.) The familiar reader will notice that A can be identified with the universal unital C∗-algebra associated to the op- erator space C, as in [43], and the maximal C∗-dilation of the disk algebra, as in [6]. In the usual generator-relation notation for universal C∗-algebras, we write A0 = C∗hx : kxk ≤ 1i. Adding a unit to the generating set and relations above gives us the unitization of A0, which is exactly A. The two are essentially interchangeable in this paper, but we often make use of the unit and so choose A for the main discussion. A C∗-algebra is called residually finite dimensional (RFD) if it has a separating family of finite-dimensional representations. The fact that a universal contraction can be a direct sum of matrices says that A is RFD and gives us a -polynomial analogue of von Neumann's inequality: the maximal norm of any noncommutative -polynomial whose entry ranges over contractive Hilbert space operators can be determined by considering only contractive (even nilpotent) matrices. In Section 3 we sharpen this by showing that the maximal norm is actually achieved at a contractive matrix whose dimension depends only on the degree of the -polynomial (Theorem 3.1). But not all elements of A attain their norm in a finite-dimensional representation (Proposition 3.3). In Section 4 we discuss the projectivity of A as a unital C∗-algebra, which is stronger than RFD [35, Lemma 8.1.4] and implies that A has trivial topological invariants such as K-theory and shape theory. Note that any separable RFD algebra has a faithful tracial state, so A is a finite C∗-algebra. On the other hand, the algebra A is the mother of all singly-generated unital C∗-algebras, so it also has various "largeness" properties. Because it surjects onto a non-exact algebra, it is not exact and a fortiori not nuclear ([8, Corollary 1] or [6, Section 6.1], although "K ⊗ B" in the latter should be unitized). Because it surjects onto a non-type I algebra (e.g., a UHF algebra, singly generated by [50]), it is not type I. This leads to comparisons with other large C∗-algebras, like the universal C∗-algebra of a partial isometry, to which A is Morita equivalent [8], and full C∗-algebras of free groups. In fact several of the results for A in Section 2 are patterned on corresponding results for C∗(F2) obtained in the beautiful 1980 paper of Choi [12]. Note that C∗(F2) is not singly generated as a unital C∗-algebra, so it cannot be a quotient of A. However, A is a quotient of C∗(F2), and they embed into each other relatively weakly injectively (Theorem 6.10). Using Theorem 6.10 THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 3 and results from Kirchberg's seminal work [31], we give in Theorem 6.11 several characterizations of Connes' Embedding Problem in terms of A (some of these were deduced differently in [43, 6]). In Section 7 we adapt some of our arguments and results for A to other C∗- algebras generated by multiple universal contractions. Theorem 7.4 generalizes Theorem 5.1 to row contractions. This leads to a Popescu-von Neumann inequality for noncommutative -polynomials on row contractions (Corollary 7.6). Section 7.2 establishes that the universal Pythagorean C∗-algebras from [9] are RFD. Our notation throughout the paper is fairly standard. For an operator T ∈ B(H), we denote its spectrum by σ(T ) and its essential spectrum by σe(T ). We denote the open unit disk and unit circle of C by D and T, respectively. We write Mn for the complex n × n matrices and freely associate it with B(Cn). After this article was completed, the authors became aware of unpublished work from 1989 by Froelich and Salas along similar themes. Upon request Prof. Salas graciously shared their manuscript; most of the overlap concerns results in Section 2. We decided to make no changes to the present article except mention of the reference [17]. The authors are grateful to David Blecher, Scott Atkinson, and an anonymous referee for useful comments on a draft of this article, and to Don Hadwin for many valuable discussions and perspectives. Some of this material is taken from the first-named author's 2018 PhD dissertation at the University of Virginia [15]. 2. Universal contraction operators We start with an easy observation. Proposition 2.1. The following are equivalent for a contractive operator T on a Hilbert space. (1) The operator T is a universal contraction; i.e., for any other contractive Hilbert space operator S, the assignment T 7→ S induces a -homomorphism from C∗(T, I) to C∗(S, I). (2) For any noncommutative -polynomial q, (2.1) kq(T )k = sup kq(S)k, S where the supremum is taken over all contractive Hilbert space operators. Proof. We have that (1) implies (2) because -homomorphisms are contractive, and (2) implies (1) because the assignment q(T ) 7→ q(S) densely defines a continuous -homomorphism between the C∗-algebras. (cid:3) How can we produce universal contractions? For the reader versed in universal C∗-algebras defined by generators and relations, it is clear that A is a separable nonzero algebra (see for instance [35, Section 3.1]). So it has a faithful represen- tation on ℓ2, and the image of its distinguished generator is a separably-acting universal contraction. The next proposition exhibits a universal contraction that is a little more con- crete. It is "known to the experts" and seems to have been first mentioned as a tool in the proof of [25, Theorem 5.1], but we lack a reference that provides the details. Since it plays an important role in this paper, for the benefit of the reader we give a short argument using only basic operator theory. (It also can be proved by appealing to the projectivity of A; see the text after Remark 4.3.) 4 KRISTIN COURTNEY AND DAVID SHERMAN Proposition 2.2. (1) The supremum in (2.1) is unchanged if S ranges only over matrix contrac- tions. (2) Let T be any direct sum of contractive matrices such that the n × n sum- mands are dense in (Mn)≤1 for each n. Then T is a universal contraction. (3) The algebra A is RFD. Proof. All of these statements effectively say the same thing, so it suffices to prove (1). (To see that A is RFD, note that restriction to the matrix summands of the operator in (2) gives a separating family of representations.) Let S be a contractive operator on the Hilbert space H. For F a finite-dimensional subspace of H, let PF be the projection onto F . Then the net {PF SPF }, ordered by inclusion of the subspaces F , converges strong to S. For any noncommuta- tive -polynomial q, we also have q(PF SPF ) → q(S) strong. By strong* lower semicontinuity of the norm, we have kq(S)k = ks∗ − lim q(PF SPF )k ≤ lim inf kq(PF SPF )k. Now PF SPF is the direct sum of a matrix and some multiple of the 1 × 1 zero operator. It follows that kq(S)k is not more than the maximal norm of q evaluated on contractive matrices. (In fact it is a maximum; see Theorem 3.1 below.) (cid:3) In Section 5 we give proofs and variations for the much harder fact, basically due to Herrero, that a universal contraction can be built as a direct sum of nilpotent matrices. So must a universal contraction be a direct sum of some sort? No, we show below in Theorem 2.5 that there are irreducible universal contractions. This is equivalent to proving that A is primitive, meaning that it has a faithful irreducible representation. Our argument mimics Choi's proof of primitivity for C∗(F2) [12, Theorem 6], with additional reliance on the fact that A is RFD. Proposition 2.3. The algebra A contains no nontrivial projections. faithful representation of A, then π(A) contains no nonzero compact operators. If π is a Proof. This can be proved in the same way as [12, Theorem 1 and Corollary 2]. (cid:3) Lemma 2.4. Let T be a universal contraction and K be a compact operator. If T + K is a contraction, then it is a universal contraction. Proof. Let T and K be as stated, let q be a noncommutative -polynomial, and let π : B(H) → B(H)/K(H) be the Calkin map. It follows from Proposition 2.3 that the restriction of π to C∗(T, I) is isometric. Since π(q(T )) = q(π(T )) = q(π(T + K)) = π(q(T + K)), we have kq(T + K)k ≥ kπ(q(T + K))k = kπ(q(T ))k = kq(T )k and are done by Proposition 2.1. (cid:3) Theorem 2.5. The algebra A is primitive. Equivalently, there exist irreducible universal contractions. Proof. Choi used the following lemma, based on techniques of Radjavi-Rosenthal [46, Theorem 7.10 and Theorem 8.30]. Let A and B be operators on the same Hilbert space, expressed as matrices with respect to a basis. If A is diagonal with distinct entries, and B has no zero entries in its first column, then A and B have THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 5 no common nontrivial reducing subspace. We will find a universal contraction whose real and imaginary parts can be taken as A and B above. Since a projection commutes with an operator if and only if it commutes with the real and imaginary parts, the contraction will have no nontrivial reducing subspaces. As in Proposition 2.2, construct a universal contraction T = ⊕∞ j=1Tj on H = j=1Cnj as a direct sum of matrices Tj ∈ Mnj . We may choose the matrices Tj ⊕∞ to be strict contractions, and we may choose the ordering so that kTjk < 1 − 1 j . By taking unitary conjugates we may also assume that the real part of each Tj is diagonal. For each j ≥ 1, perturb the diagonal entries of Re Tj by less than 1 2j so that the entries are distinct from each other and all diagonal entries of Ti for i < j. This perturbs T to T ′ so that Re T ′ is diagonal with distinct entries, and the cor- responding summands have kT ′ 2j . Because this is a compact perturbation, T ′ is still a universal contraction by Lemma 2.4. jk < 1 − 1 Now perturb T ′ to T ′′ as follows. We only change the first column and row so that these are everywhere nonzero in Im T ′′. We go one block at a time, taking advantage of the little bit of norm wiggle room. In the first block T ′ 1, change all entries in the first column and row by some small identical pure imaginary amount λ1, making them all nonzero and keeping kT ′′ 3 . Then change the part of the first column and row of T ′ corresponding to the second block - these are all zero in T ′ - by a small identical pure imaginary amount λ2, making them all nonzero and preserving that the submatrix of T ′′ corresponding to the first two blocks has norm < 1 − 1 j. Visually, we are perturbing T ′ by the operator 6 . Continue in this way for all blocks T ′ 1 k < 1 − 1 · · · λ1 λ2 · · · λn · · · 0   R = λ1 λ1 λ1 ... λ1 λ2 ... λn ...   where each λj ∈ iR is repeated nj times. The submatrix of T ′′ = T ′ + R corresponding to the initial string of n blocks has norm < 1 − 1 3n . The operator T ′′ is a strong limit of these submatrices (considered as infinite-dimensional operators by filling in the rest of the matrix with zeroes), so T ′′ is also a contraction. And because we have only changed the first row and column to go from T ′ to T ′′, it is a compact perturbation, and T ′′ is still a universal contraction by Lemma 2.4. Notice that the matrix for Re T ′′ = Re T ′ is diagonal with distinct entries, and the matrix for Im T ′′ has nonzero entries in the first column. By the result mentioned at the beginning of the proof, we are done. (cid:3) It is apparent from the preceding results that separably-acting universal con- tractions need not be unitarily equivalent, but the truth is not so far from that. Recall that two operators are said to be approximately unitarily equivalent (a.u.e.), 6 KRISTIN COURTNEY AND DAVID SHERMAN denoted here ∼a, if one is the norm limit of unitary conjugates of the other. Two representations of the same C∗-algebra are a.u.e. if one representation is the point- norm limit of unitary conjugates of the other. Parts of the next two propositions are noted or implied in Hadwin's work on Voiculescu's noncommutative Weyl-von Neumann theorem [24, 25, 26] and the sub- sequent discussion in [17, Section 7]. We include some citations but make all the reasoning explicit here. Proposition 2.6. Separably-acting universal contractions form a single a.u.e. class. Proof. Let T1 and T2 be separably-acting universal contractions. By sending the distinguished generator of A to Tj, we obtain two faithful representations of A whose ranges contain no nontrivial compact operators by Proposition 2.3. It then follows from Voiculescu's noncommutative Weyl-von Neumann theorem [51, Corollary 1.4] that the representations are a.u.e., which means that T1 ∼a T2. If S1 ∼a S2, then for any noncommutative -polynomial q we have q(S1) ∼a q(S2) and kq(S1)k = kq(S2)k. So any operator a.u.e. to a universal contraction must also be one by Proposition 2.1. (cid:3) Proposition 2.7. For T ∈ B(ℓ2)≤1, the following are equivalent. (1) T is a universal contraction. (2) T ∼a (T ⊕ S) for every S ∈ B(ℓ2)≤1. (3) The strong* closure of the unitary orbit of T is B(ℓ2)≤1. Proof. (1) ⇐⇒ (2): We first show that the operators on ℓ2 satisfying the condition in (2) form a single a.u.e. class. If T1, T2 satisfy the condition, then And if T ∼a T1 and S ∈ B(ℓ2)≤1, then T1 ∼a T1 ⊕ T2 ∼a T2. T ⊕ S ∼a T1 ⊕ S ∼a T1 ∼a T. The conclusion then follows from Proposition 2.6 and Hadwin's observation [26, Example 7.3(1)] that a universal contraction constructed in Proposition 2.2(2) sat- isfies the condition in (2). (2) ⇒ (3): Hadwin shows in [24, Corollary 3.4 and Theorem 4.3] that for any T ∈ B(ℓ2), the strong* closure of the unitary orbit of T is the set of operators that are summands of operators a.u.e. to T . (3) ⇒ (1): Assume (3), and let S be any contraction on ℓ2. Let {Uj} be unitaries with U ∗ j T Uj → S strong, and let q be a noncommutative -polynomial. Then kq(S)k = kq(s∗ − lim(U ∗ j T Uj))k = ks∗ − lim q(U ∗ j T Uj)k = ks∗ − lim U ∗ j q(T )Ujk ≤ lim inf kU ∗ j q(T )Ujk = kq(T )k. Thus T is universal by Proposition 2.1. (cid:3) Remark 2.8. "Universality" has various meanings for operators, some connected to our situation and some not. In the context of operator ideals it is a factorization property (e.g., [41]) -- this is quite different from the present paper. In the definition perhaps most familiar to an operator theorist, a Hilbert space operator T is said to be universal, or be a universal model, if every separably- acting operator can be scaled to become similar to the restriction of T to some invariant subspace. See [33, Chapter 6] for a discussion of the area. The terminology THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 7 originates with Rota [47], who proved in [48] that B, the backward shift with infinite multiplicity, is a universal model in a slightly more precise sense: any operator with spectral radius < 1 is similar to the restriction of B to an invariant subspace. Proposition 2.7(2) characterizes the universality of T (as defined in this paper) analogously: any operator with norm ≤ 1 is unitarily equivalent to the restriction of an operator a.u.e. to T to a reducing subspace. Operator theorists may be tempted to look for universal contractions among the weighted shifts, but that cannot succeed because weighted shifts are centered [39]; they satisfy relations saying that the family {T mT ∗m, T ∗nT n : m, n ≥ 0} is commutative. In fact the authors know of no explicit "naturally-occurring" universal contractions. Our constructions all rely in some way on the direct sum of a dense set. 3. Von Neumann-type inequalities for noncommutative -polynomials Let Mz denote multiplication by z on L2(T, m), and let p be a polynomial. From compactness, the maximum modulus principle, and spectral theory, we have (3.1) [sup or max]¯Dp(λ) = [sup or max]Tp(λ) = kp(Mz)k. The celebrated von Neumann inequality [40] is the fact that for any contraction S, the norm of p(S) is dominated by the quantity in (3.1). We may then say that Mz is a universal contraction for polynomials. (In fact a contraction has this property if and only if its spectrum contains T.)1 Equiva- lently, Mz generates the universal unital operator algebra2 of a contraction, which is isometrically isomorphic as a Banach algebra to the disk algebra A(D) via the map densely defined by p ↔ p(Mz) [7, Example 2.2]. All of the foregoing is still true if we replace p with any f ∈ A(D), and so von Neumann's inequality opens the door for an analytic functional calculus for all contractions. (For any contraction S, f (S) is well-defined by von Neumann's inequality as the limit of pj(S), where {pj} is any sequence of polynomials converging to f in A(D).) In summary, von Neumann's inequality says that there are sufficiently many one- dimensional representations of the universal unital operator algebra of a contraction to determine the norm of a polynomial in the generator (and thus we can identify the algebra with A(D) and the one-dimensional representations with ¯D). In fact the norm is achieved. The smaller set T suffices to determine the norm, which is still achieved. For any element of A(D) the norm is achieved at a one-dimensional representation. We pursue analogues of every clause in the preceding paragraph, in the context of noncommutative -polynomials and C∗-algebras, and call these analogues (when valid) "von Neumann-type inequalities." Let x be the canonical generator of A, S any contraction, and q a noncommu- tative -polynomial. By definition we have that kq(S)k ≤ kq(x)k. We typically cannot determine kq(x)k with the one-dimensional representations of A (i.e., re- placing x with scalars): consider q(z) = z∗z − zz∗. But the finite-dimensional 1The backward implication holds because the operator norm is not smaller than the spectral radius. For the forward implication, let λ ∈ T and T be a contraction with this property. The spectral radius of (T + λ) is lim k(T + λ)nk1/n = lim k(z + λ)nk1/n C(T) = 2, which forces λ ∈ sp(T ). 2Here an operator algebra is a (not necessarily self-adjoint) norm-closed subalgebra of some B(H), or any matrix-normed Banach algebra that can be represented as such. 8 KRISTIN COURTNEY AND DAVID SHERMAN representations of A do suffice (i.e., replacing x with contractive matrices). We already proved this in Proposition 2.2(1), which may thus be considered a von Neumann-type inequality: (3.2) kq(S)k ≤ kq(x)k = sup kq(M )k. M contractive matrix Here are our questions. Note that there is no compactness available. (1) Is the supremum in (3.2) achieved? (2) Can we replace the contractive matrices with a smaller natural set? And will the supremum still be achieved? (3) Since noncommutative -polynomials in x form a dense set in A, it follows from the above that for any a ∈ A, kak = supπ kπ(a)k, as π ranges over the finite-dimensional representations of A. Is this supremum achieved? Here are the answers, the first and third of which we proceed to show in the remainder of this section. (1) Yes (Theorem 3.1). (2) Contractive nilpotent matrices suffice, but the supremum is not achieved in general (Theorem 5.1 and Remark 5.2). (3) No, it is not achieved for all elements of A, but it is for some elements that are not of the form q(x) (Proposition 3.3). In [16, Theorem 3.2] it was shown that a C∗-algebra is RFD exactly when it has a dense subset of elements that attain their norm under a finite-dimensional representation of the algebra. In some cases, this dense subset contains all non- commutative -polynomials in the standard generators. For instance, Fritz, Netzer, and Thom prove in [22, Lemma 2.7] that every element in CFn attains its norm under some finite-dimensional representation of C∗(Fn), where Fn is a free group on n ≤ ∞ generators. Our proof of Theorem 3.1 below is a simplified version of the proof of [22, Lemma 2.7], which itself is an adaptation of Choi's argument in [12, Theorem 7] that C∗(Fn) is RFD. The degree of a noncommutative -polynomial is the length of its longest mono- mial. Theorem 3.1. Let q be a noncommutative -polynomial of degree d, and let x be the canonical generator of A. Then (3.3) kq(x)k = max{kq(M )k : M ∈ M2d+1, kM k ≤ 1}. Proof. Let ϕ be a state on A with ϕ(q(x)∗q(x)) = kq(x)k2, and let (π, H, ξ) be the associated GNS representation. Then kq(x)k = kπ(q(x))ξk. Define H0 = span{π(g(x))ξ : g is a -monomial of length ℓ(g) ≤ d}. Note that dim(H0) ≤ 2d+1. Let V be the inclusion of H0 in H, so that V V ∗ is the projection in B(H) onto H0. By construction (3.4) V q(V ∗π(x)V )V ∗ξ = q(π(x))ξ, since V V ∗ acts as the identity everywhere in the expansion of the left-hand side. We can think of V ∗π(x)V ∈ B(H0) as a contractive matrix of size ≤ 2d+1, so the assignment x 7→ V ∗π(x)V induces a unital -homomorphism π0 : A → B(H0). THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 9 Compute kq(π0(x))(V ∗ξ)k ≥ kV q(π0(x))V ∗ξk = kV q(V ∗π(x)V )V ∗ξk = kq(π(x))ξk = kπ(q(x))ξk = kq(x)k. Thus the norm of q(π0(x)) must be kq(x)k (it cannot be bigger). Finally note that B(H0) can be (not necessarily unitally) included in M2d+1 . (cid:3) The reader may wonder if A has a separating family of representations of bounded (finite) dimension, so that the maximum in (3.3) can be taken in some fixed MN for all q. (For polynomials N = 1 works!) This is not so: for a C∗-algebra, the existence of a separating family of representations of dimension ≤ N implies that all irreducible representations have dimension ≤ N (see [18, Proposition 3.6.3(i)]). Actually, it follows from [16, Theorem 5.1] that there are elements of A whose maximal norm in n-dimensional representations grows according to any prescribed finite pattern in n. But the methods in [16] are nonconstructive, so that one cannot exhibit such elements, and they are unlikely to be noncommutative -polynomials in the generator. As a complement to these results we display here explicit non- commutative -polynomials that achieve their maximal norm at any prescribed dimension and no lower. Example 3.2. For any n ≥ 1, consider the noncommutative -polynomial qn(z) = (z∗z + z∗2z2 + ... + z∗nzn) + (1 − zz∗). We claim that the maximal norm of qn(M ), where M ranges over all contractions, is attained in Mn+1 but no smaller matrix algebra. We have that qn(M ) is a sum of n + 1 positive contractions, so its norm cannot be more than n + 1. Taking Sn to be the forward shift in Mn+1, qn(Sn) sends the first basis element e1 to (n + 1)e1, so it has the maximal norm n + 1. Now suppose that M is a matrix such that the positive matrix qn(M ) has norm n + 1. Let v be a unit eigenvector for qn(M ) for the value n + 1. It follows that v is fixed by all of M ∗M, . . . , M ∗nM n and annihilated by M M ∗; this means that M ∗v = 0 and M jv is a unit vector for j = 1, . . . n. We have M v ∈ ran(M ) ⊥ ker(M ∗) ∋ v. We also have M 2v ∈ ran(M 2) ⊥ ker(M ∗2) ∋ v, M v. (For the last membership, note M ∗2(M v) = M ∗(M ∗M v) = M ∗v = 0.) Continuing, we get that v, M v, M 2v, . . . , M nv form an orthonormal set of (n + 1) vectors on which M acts as the forward shift. Thus M is a matrix of size at least (n + 1) × (n + 1). Let AF denote the set of elements in A that achieve their norm under some finite-dimensional representation, i.e., a ∈ AF if there exists a finite-dimensional representation π of A such that kak = kπ(a)k. Then Theorem 3.1 says that the set P∗ of noncommutative -polynomials in x is contained in AF . It is natural to ask whether every element in A attains its norm under some finite-dimensional representation, i.e., whether or not AF is the entire space. The first-named author and T. Shulman explored this question for general C∗-algebras in [16]. As the next proposition shows, not every element of A achieves its norm under a finite- dimensional representation, but there are some elements in A\P∗ that do. Proposition 3.3. Let AF be defined as above. Then P∗ ( AF ( A. 10 KRISTIN COURTNEY AND DAVID SHERMAN Proof. By [16, Theorem 4.4], we know that AF ( A if and only if A has a simple, infinite-dimensional AF subquotient. The CAR algebra M2∞ is a simple, unital, infinite-dimensional AF C∗-algebra, which is singly generated by [50]. Hence, it is isomorphic to a quotient of A, and so AF ( A. It also follows from [16, Theorem 5.5] that AF is not closed under addition (or multiplication) and hence cannot equal P∗, i.e., P∗ ( AF . (cid:3) The proof of Proposition 3.3 is again nonconstructive. A specific element of AF \ P∗ is ex∗x, which attains its norm of e in the one-dimensional representation of A sending x to 1. More interestingly, we can use Example 3.2 to exhibit an element of A \ AF . Example 3.4. For any contractive element y of a C∗-algebra, consider the norm convergent series q(y) = 2−n qn(y) n + 1 , ∞ Xn=1 where the qn are as in Example 3.2. Since kqn(x)k = n + 1, we have kq(x)k ≤ 1. Taking S to be the unilateral shift on ℓ2 and e1 the first standard basis vector, we also have kq(S)k ≥ kq(S)e1k = ke1k = 1. Thus kq(x)k = 1. But it follows from Example 3.2 that kq(M )k < 1 for any finite-dimensional contraction M . 4. Universal C∗-algebras and projectivity In this section, we consider properties of A as a universal C∗-algebra. Because other universal C∗-algebras will soon make appearances, now is a good time to introduce them formally. We will forgo constructions and refer the reader to those given in [35, Chapter 3] or [5, Section 1]. Though they can be defined in more generality, we will restrict ourselves to separable universal C∗-algebras defined by imposing norm constraints and noncommutative -polynomial relations on the gen- erators. Let G = {x1, ..., xn} be a finite set and R a finite set of relations of the form • r(x1, ..., xn) ≥ 0, or • ks(x1, ..., xn)k ≤ C for some C ≥ 0 where r and s are noncommutative -polynomials. To fend off existence issues, we require that the relations in R enforce norm bounds on the elements of G and that there exists some tuple (A1, ..., An) of operators on some Hilbert space that satisfy R. The universal C∗-algebra with generators G subject to relations R is denoted by C∗hG : Ri and has the following defining universal property: given Hilbert space operators T1, ..., Tn satisfying R, the assignments xi 7→ Ti, 1 ≤ i ≤ n, induce a surjective -homomorphism C∗hG : Ri → C∗(T1, ..., Tn). If C∗hG : Ri has a unit, then the -homomorphism is assumed to be unital. If it is not unital, then we denote its unitization by C∗ uhG : Ri; we also adopt the convention of identifying a unital C∗-algebra with its unitization. Note that unitizing a universal C∗-algebra is the same as adding a unit to the generating set, i.e., C∗h{y1, ..., yn} ∪ {I} : R ∪ {I = I ∗ = I 2, yiI = yi = Iyi, 1 ≤ i ≤ n}i. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 11 To see this, we faithfully represent C∗h{y1, ..., yn} ∪ {I} : R ∪ {I = I ∗ = I 2, yiI = yi = Iyi, 1 ≤ i ≤ n}i as C∗(IH, Y1, ..., Yn) on some Hilbert space H. By universality, we have a surjective -homomorphism C∗hG : Ri → C∗(Y1, ..., Yn) induced by sending xi 7→ Yi. This extends to a surjective unital -homomophism C∗ uhG : Ri → C∗(IH, Y1, ..., Yn), which is injective by the universality of the latter algebra. In particular, this means adding a unit to the generators of A0 and unitizing A0 give the same algebra, namely A. (We shall see in Remark 4.3 why A0 cannot have a unit.) As a convention, we say the unitization of the zero C∗-algebra is C. Now we are ready to change gears and introduce one of the most important properties of A. We call a C∗-algebra A projective if given any C∗-algebra B with closed two-sided ideal I and quotient map π : B → B/I, any -homomorphism φ : A → B/I lifts to a -homomorphsim ψ : A → B so that φ = π ◦ ψ. In other words, a projective C∗-algebra is a projective object in the category of C∗-algebras with -homomorphisms. By [5, Proposition 2.5], a C∗-algebra is projective if and only if its unitization is projective in the category of unital C∗-algebras with unital -homomorphisms. For simplicity, if we call a unital C∗-algebra projective, we mean in the category of unital C∗-algebras with unital -homomorphisms. Proposition 4.1. [35, Lemma 8.1.4] Both A0 and A are projective. The projectivity of A follows from that of A0, which is a consequence of the fact that any contraction in a C∗-quotient lifts to a contraction. Remark 4.2. We mentioned in the introduction that C∗(F2) is not projective. This means exactly that unitaries do not in general lift to unitaries, e.g., unitaries in the Calkin algebra with nonzero Fredholm index. Remark 4.3. By [5, Proposition 2.3], a projective C∗-algebra is contractible, which implies that it is non-unital and has trivial shape in the sense of [5] and trivial K- theory. (We say a unital C∗-algebra is contractible when it is homotopy equivalent to C.) That A is contractible and has trivial K1 is also proved in [8, Theorem 3.1]. Projectivity is a powerful (and rare) property because many other nice properties can be rephrased in terms of lifting problems. For instance, we can quickly establish Propositions 2.2 and 2.3. Proposition 2.2 is a special case of the following. Proposition 4.4. Any separable projective C∗-algebra is RFD. See [35, Theorem 11.2.1] or [27] for a full proof. The main ideas are that any separable C∗-algebra is the quotient of an RFD C∗-algebra, and that residual fi- nite dimensionality passes to subalgebras. For projective objects, being a quotient implies being a subobject. Meanwhile Proposition 2.3 follows from the next proposition, the proof of which uses only the statement of [12, Proposition 1] as opposed to its proof. Proposition 4.5. Any separable projective C∗-algebra B has no nontrivial projec- tions, and if (π, H) is a faithful representation of B, then π is essential (i.e., π(B) contains no nonzero compact operators). Proof. It will suffice to prove the unital case. To that end, let B be a separable unital C∗-algebra that is projective as a unital C∗-algebra. Then B is isomorphic to a quotient of C∗(F∞). By projectivity, this isomorphism lifts to an embedding 12 KRISTIN COURTNEY AND DAVID SHERMAN of B into C∗(F∞). Since C∗(F∞) has no nontrivial projections [12, Theorem 1], neither does B.3 It then follows, by the same argument as for [12, Corollary 2], that any faithful representation of B on a Hilbert space H trivially intersects K(H). (cid:3) Contractibility makes it difficult to distinguish projective C∗-algebras, such as A i , i = 1, 2i. and the free product C[−1, 1] ∗C C[−1, 1] ≃ C∗ (That this particular C∗-algebra is projective follows from [19, Theorem 3.2].) uhx1, x2 : kxik ≤ 1, xi = x∗ Question 4.6. Are A and C[−1, 1] ∗C C[−1, 1] isomorphic? We point out why the most obvious guess at an isomorphism does not answer Question 4.6. Let x be the distinguished generator of A, and b, c be the canonical generators of the C[−1, 1] factors in the free product. Since Re x and Im x are self-adjoint contractions with spectrum [−1, 1], one might propose an isomorphism via Re x 7→ b, Im x 7→ c. But this fails, because x = Re x + iIm x is a contraction, while b+ic has norm 2. (One can deduce this from the representation of C[−1, 1]∗C i 0(cid:19), and b + ic to (cid:18)0 2 0 0(cid:19).) C[−1, 1] sending b to (cid:18)0 1 1 0(cid:19), c to (cid:18) 0 −i Consider the "abelianized" version of this question. The universal unital C∗- algebra of a normal contraction is C(D), and the tensor product C[−1, 1] ⊗ C[−1, 1] is C([−1, 1]2). These two C∗-algebras are isomorphic, because the closed unit disk is homeomorphic to the closed (solid) 2 × 2 square. But the homeomorphism is more complicated than simply rescaling coordinates. Question 4.6 is asking for a noncommutative version of this homeomorphism. With poetic license: are the noncommutative disk and noncommutative solid square homeomorphic? 5. Finite-dimensional and nilpotent representations of A In this section we discuss the rather surprising fact that a universal contrac- tion can be a countable direct sum of nilpotent matrices. This was originally proved by Herrero via intricate computations; here we give two alternative argu- ments. The first, "`a la Choi" and fairly short, proceeds by dilating slightly rescaled finite-dimensional compressions of a universal contraction to contractive nilpotent matrices. The second, which we later generalize in the proof of Theorem 7.4, asymp- totically factorizes a faithful representation of A through the universal C∗-algebras of nilpotents of increasing order, relying on the classical nilpotent approximation theorem of Apostol-Foias-Voiculescu and a result of Shulman on lifting nilpotent contractions. Theorem 5.1. There exists a separating family of finite-dimensional representa- tions of A that map the generator to contractive nilpotent matrices. In other words, for any noncommutative -polynomial q and contractive Hilbert space operator S, kq(S)k ≤ sup{kq(M )k : M ∈ Mn, kM k ≤ 1, M n = 0, n ≥ 1}. Theorem 5.1 gives us a different refinement (from Theorem 3.1) of the von Neumann-type inequality we got from the residual finite-dimensionality of A in Section 3. Since the representations guaranteed by Theorem 5.1 are separating, by [16, Theorem 3.2], A has a dense subset of elements that attain their norm under one 3This fact also follows from [28, Proposition 3]. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 13 of these representations. However, this dense subset need not contain the canonical dense subset P∗ of noncommutative -polynomials in the generator, as the next example shows. Example 5.2. Let x be the canonical generator of A. Then x+x∗ cannot attain its norm under a finite-dimensional representation mapping x to a nilpotent operator. First note that kx + x∗k = 2, which can be seen by mapping x to 1. Now, suppose M ∈ Mn is nilpotent with kM k = 1. Since M + M ∗ is self-adjoint, if kM + M ∗k = 2, then there exists a unit vector v such that either (M + M ∗)v = 2v or (M + M ∗)v = −2v. Because the unit vector v is an extreme point of the unit ball, it follows that M v = v or M v = −v, either way violating the fact that σ(M ) = 0. Now, let us commence with the proofs of Theorem 5.1. The first proof is truly due to Herrero, though he does not draw this explicit conclusion. Let T be a universal contraction as constructed in Proposition 2.2(2). Herrero shows in [30, Corollary 4.8] that T is the norm limit of contractive nilpotent operators that are block diagonal. Hence by Proposition 2.6, the same holds for any universal contraction on B(ℓ2), and we have the following extension of [30, Corollary 4.8]. Proposition 5.3. Every universal contraction in B(ℓ2) is the norm limit of block- diagonal nilpotent contractive operators with finite-dimensional blocks. Theorem 5.1 follows by taking the family of representations induced by mapping the generator x to each of the blocks in the sequence. Remark 5.4. Note that no universal contraction T ∈ B(H) can be the norm limit of nilpotent contractions that lie in C∗(T, I). Indeed, if N ∈ C∗(T, I) is nilpotent and π : C∗(T, I) → C maps T 7→ 1, then kT − N k ≥ kπ(T ) − π(N )k = k1 − 0k = 1. Thanks to Don Hadwin for originally pointing out this fact. Our second argument proves Theorem 5.1 directly by an "asymptotic dilation" argument. This is close in spirit to Choi's proof [12, Theorem 7] that C∗(F2) is RFD, but the details are somewhat different.4 Proof. For each n > 1, let Jn ∈ Mn denote the n×n Jordan block with eigenvalue 0. By [23, Proposition 1], there is a unit vector ξn ∈ Cn so that w(Jn) = hJnξn, ξni = cos π n+1 , where w(Jn) is the numerical radius of Jn. Let Un ∈ Mn be a unitary so that Une1 = ξn (where e1 is from the standard basis of Cn). For each n > 1, define Mn := U ∗ nJnUn = (a(n) ij )1≤i,j≤n. Then Mn is a nilpotent contraction with a(n) 11 = cos π n+1 . Let H = ℓ2 with the standard basis {e1, e2, ...}; let Pn be the orthogonal projec- tion of H onto span{e1, ..., en}; let T be a universal contraction operator on H; and let Tn := PnT Pn for each n > 1. Then Tn ⊗ Mn ∈ PnB(H)Pn ⊗ Mn is a nilpotent contraction, and, if we view PnB(H)Pn as Mn, then the assignments T 7→ Tn ⊗ Mn 4Arveson has characterized in [3, Theorem 1.3.1] the contractive Hilbert space operators that can be power dilated to contractive nilpotent operators -- for powers less than the order of the nilpotent. But we need all -monomials, at least asymptotically. 14 KRISTIN COURTNEY AND DAVID SHERMAN induce a family of finite-dimensional representations πn : C∗(T, I) → Mn2 where πn(T ) is nilpotent. To see that this family is separating, it will suffice to show that ⊕nπn is isometric, which will follow from showing that kq(T )k = supn kq(Tn ⊗Mn)k for any noncommutative -polynomial q. To that end, let q be any nonzero non- commutative -polynomial, and let ε > 0. For each n > 1, define Cn = (c(n) ij ) ∈ Mn−1 by c(n) ij := a(n) i+1,j+1 for each 1 ≤ i, j ≤ n − 1. Since a(n) 11 = cos π n+1 → 1 as n → ∞, we have k[a(n)∗ 1j ]2≤j≤nkℓ2 n as n → ∞, and so → 0 and k[a(n) i1 ]2≤i≤nkℓ2 n → 0 a(n) 12 0 ... 0 0 a(n) 21 ... a(n) 1n ≤ k[a(n)∗ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   1j ... a(n) 1n 0 ... ... 0 ... ... + k[a(n)   (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kMn −(cid:16)cos( π n+1 ) ⊕ Cn(cid:17) k = ]2≤j≤nkℓ2 n i1 ]2≤i≤nkℓ2 n → 0 as n → ∞, which implies that kTn ⊗ Mn − Tn ⊗ (cos( π n+1 ) ⊕ Cn)k → 0 as n → ∞. Now, q is Lipschitz on the unit ball of Mn2 ≃ PnB(H)Pn ⊗ Mn with Lipschitz constant independent of n. From this we see that (5.1) kq(Tn ⊗ Mn) − q((Tn ⊗ (cos( π n+1 ) ⊕ Cn))k −−−−→ n→∞ 0. Assume for simplicity that kq(T )k = 1. Since Tn and we have S ∗OT−−−−→ T , so does cos( π n+1 )Tn, (5.2) kq(cos( π n+1 )Tn)k > 1 − ε 2 for n sufficiently large. Combining (5.1) and (5.2), we have, for n sufficiently large, kq(Tn ⊗ Mn)k > kq(Tn ⊗ (cos( π n+1 ) ⊕ Cn))k − ε 2 = kq((Tn ⊗ cos( π n+1 )) ⊕ (Tn ⊗ Cn))k − = kq((Tn ⊗ cos( π n+1 )) ⊕ q(Tn ⊗ Cn)k − ≥ kq(cos( π n+1 )Tn)k − ε 2 = 1 − ε. > 1 − ε 2 − ε 2 ε 2 ε 2 Since ε was arbitrary, this shows that ⊕nπn is an isomorphism. (cid:3) Our final proof of Theorem 5.1 will require a few preliminary results. Though it is a little more work on the outset, this approach is what we will generalize to prove Theorem 7.4, at which point we will appreciate the preliminaries established here. We begin by recalling a few definitions. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 15 Definition 5.5. A separable C∗-algebra A is quasidiagonal (QD) if there exists a sequence of completely positive contractive maps φn : A → Mkn that are asymp- totically multiplicative and asymptotically isometric. It is easy to see that a separable RFD C∗-algebra is QD since it embeds faithfully into some product Qn Mkn . Definition 5.6. A set of operators Ω ⊆ B(H) is called quasidiagonal (QD) if there exists an increasing sequence of finite-rank projections P1 ≤ P2 ≤ ... converging strongly to I such that kPnT − T Pnk → 0 for every T ∈ Ω. A representation π of a C∗-algebra A on H is called quasidiagonal (QD) if π(A) is a QD set of operators on H. Note that π being a QD representation of A is not equivalent to π(A) being a QD C∗-algebra (see [10, Remark 7.5.3]). The issue arises from a Fredholm index obstruction, which can be avoided by taking essential representations, i.e., those in which the image contains no nonzero compact operators. In fact, Voiculescu has shown in [52] that a separable C∗-algebra A is QD if and only if every faithful essential representation of A (on a separable Hilbert space) is quasidiagonal. Definition 5.7. An operator T on H is called quasitriangular if there exists an increasing sequence of finite-rank projections P1 ≤ P2 ≤ ... converging strongly to I such that kT Pn − PnT Pnk → 0; it is called bi-quasitriangular if T and T ∗ are quasitriangular. A QD operator is automatically bi-quasitriangular, and so if π is a QD rep- resentation of a C∗-algebra A, then every π(a) ∈ π(A) is bi-quasitriangular. In particular, if π is an essential representation of a separable RFD C∗-algebra, then every π(a) ∈ π(A) is bi-quasitriangular. Our first step is an observation which utilizes Apostol, Foias, and Voiculescu's characterization of the norm closure of nilpotents in B(H) [2, Theorem 2.7] (which is also a key ingredient in the proof of Herrero's result [30, Corollary 4.8]). Proposition 5.8. If π is a faithful representation of a projective C∗-algebra A on a separable Hilbert space H, then for any a ∈ A, π(a) is the norm limit of nilpotent operators in B(H) if and only if its spectrum is connected and contains 0. In particular, if T ∈ B(H) is a universal contraction, then there exists a sequence (Tn)n≥1 of nilpotent operators in B(H) that converge to T in norm. Moreover, these can be chosen to be contractions satisfying T n n = 0 for each n ≥ 1. Proof. It will suffice to prove the unital case. Let π be a faithful representation of A, which is also essential by Proposition 4.5. By [2, Theorem 2.7], we know that an operator T ∈ B(H) is the norm limit of nilpotent operators if and only if T is bi-quasitriangular, σ(T ) and σe(T ) are connected, and 0 ∈ σe(T ). Because A is projective, it is RFD. Since π is essential, every element π(a) ∈ π(A) is automatically bi-quasitriangular, and σ(π(a)) = σe(π(a)). Hence we have that π(a) is the norm limit of nilpotent operators in B(H) if and only if σ(π(a)) is connected and contains zero. In the case of A, note that the spectrum of the generator must be D, and any sequence of nilpotent operators converging in norm to T can be easily adjusted to satisfy the extra criteria. (cid:3) 16 KRISTIN COURTNEY AND DAVID SHERMAN Of course, Proposition 5.3 says we can actually arrange it so that each Tn is also block-diagonal. However, this will not be of any help when we generalize Theorem 5.1 in Section 7.1, and Proposition 5.8 is really all we need. The following theorem gives an "exactness"-type result, which says that any faithful representation of A on a separable Hilbert space "asymptotically factorizes" through the family of universal C∗-algebras generated by nilpotent contractions of increasing finite orders. Theorem 5.9. Let π : A → B(H) be a faithful representation of A on a separable Hilbert space H, and let φn : A → An := C∗ uhxn : kxnk ≤ 1, xn n = 0i be the unital -homomorphisms induced by mapping x 7→ xn for each n ≥ 1. Then there exists a sequence of -homomorphisms such that ψn ◦ φn converges to π pointwise in norm. ψn : An → B(H) Proof. By Proposition 5.8 there exists a sequence (Tn) ⊆ B(H) that converges in norm to π(x) and satisfies kTnk ≤ 1 and T n n = 0 for each n ≥ 1. For each n ≥ 1, let ψn : An → C∗(Tn, I) ⊆ B(H) be the canonical surjective unital -homomorphsim induced by mapping xn 7→ Tn. Then, for any a ∈ A, kψn ◦ φn(a) − ak → 0. (cid:3) Remark 5.10. By Remark 5.4, the identity map on A does not factor through the family {An}n≥1, which means we cannot hope to extend this to a "nuclearity" result. With Theorem 5.9, we can give our final proof of Theorem 5.1. Proof. The family {φn} from Theorem 5.9 separates the elements of A. By [49, Theorem 5], for each n ≥ 1 the C∗-algebra An is projective and hence also RFD by Proposition 4.4. Let {ρn,k}k∈N be a separating family of finite-dimensional representations of An. Then the compositions {ρn,k ◦ φn}n,k∈N form a separating family of finite-dimensional representations of A, and, moreover, ρn,k ◦ φn(x) is nilpotent for each n, k ∈ N where x denotes the generator of A. (cid:3) We conclude this section with a slight generalization of Theorem 5.1. Theorem 5.11. For each λ ∈ D, A has a separating family of finite-dimensional representations that send the generator to contractive matrices with spectrum {λ}. The proof due to Herrero can be adapted to show this by replacing [30, Corollary 4.8] with the remark in [30] just after it. In the remainder of this section we sketch how our second proof can also be adapted to prove Theorem 5.11. As it turns out, by [38, Theorem 10], the universal C∗-algebra Aλ,n = C∗ uhxn : kxnk ≤ 1, (xn − λ)n = 0i is RFD for each n ≥ 1. So all we need is another "exactness" result, like Theorem 5.9, which says that every faithful representation (π, H) of A asymptotically fac- torizes through the Aλ,n. To define the maps ψn as in Theorem 5.9, it was crucial THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 17 that a universal contraction T is the norm limit of nilpotent contractions Tn, each with appropriate nilpotency order. For this more general setting, we need T to be the norm limit of contractions Tn satisfying (Tn − λI)n = 0 for each n ≥ 1, which we show in Lemma 5.12 below. The rest of the argument for Theorem 5.11 will then run like the third proof of Theorem 5.1. Lemma 5.12. Let H be a separable Hilbert space, T ∈ B(H) be a universal con- traction operator, and λ ∈ D. Then there exists a sequence of contractive operators (Tn) ⊆ B(H) that converges in norm to T and satisfies (Tn − λI)n = 0 for each n ≥ 1. Proof. Since σ(T − λI) = ¯D − λ is connected and contains 0, by Proposition 5.8, there is a sequence (Nn) ⊆ B(H) of nilpotents so that Nn → T − λI in norm and N n n = 0 for each n ≥ 1. Then Nn + λI → T in norm; however, each Nn + λI may not be a contraction. To remedy this, we define Tn := cnNn + λI for each n ≥ 1 where cn =( 1 : kNn + λIk ≤ 1 : kNn + λIk > 1 . 1−λ kNn+λIk−λ Then (Tn − λI)n = (cnNn)n = 0 for each n ≥ 1. Since kNn + λIk → kT k = 1, it follows that cn → 1 and hence that Tn → T in norm. Moreover, if kNn + λIk ≤ 1, then Tn = cnNn + λI = Nn + λI is a contraction. If kNn + λIk > 1, then 1 > cn, and we have that kTnk = kcnNn + λIk = kcn(Nn + λI) + (1 − cn)λIk ≤ cnkNn + λIk + λ1 − cn = cnkNn + λIk + λ − λcn = cn(kNn + λIk − λ) + λ = (1 − λ) + λ = 1. Thus (Tn) is the desired sequence. (cid:3) Remark 5.13. We cannot replace λ ∈ D from Theorem 5.11 with λ ∈ C\D in either of the preceding results. Of course, if λ > 1, then no contraction Y ∈ B(H) can satisfy (Y − λI)n = 0. And if Y ∈ B(H) is any contraction with (Y − λI)n = 0 for some n ≥ 1 and some λ ∈ T, then Y = λI. Indeed, by [38, Lemma 4] we know that Y is unitarily equivalent to an upper triangular array (yij)i,j≥1 with yii = λ for i ≥ 1. But, since kY k ≤ 1, it follows that yij = 0 for i 6= j, and Y = λI. In other words, for any λ ∈ T and n ≥ 1, C∗ uhx : kxk ≤ 1, (x − λ)n = 0i ≃ C via the map x 7→ λ. This argument shows that kN +λIk > 1 for any λ ∈ T and any nonzero nilpotent operator N . (Just take y = N + λI, and run the previous argument as a contradic- tion.) But this fails if we assume only that N is quasinilpotent, i.e., σ(N ) = {0}. For example, let V be the Volterra integration operator, and N = (I + V )−1 − I. We know from [29, Problem 190] that kN + Ik = k(I + V )−1k = 1, and σ(N ) = {0}. However, N 6= 0. 6. A and Connes' Embedding Problem As we mentioned in the introduction, A often behaves much like C∗(F2), in particular when it comes to the role of C∗(F2) in Kirchberg's characterizations of Connes' Embedding Problem. In this section, we elucidate the relation of A to Connes' Embedding Problem and consider some related questions. Before we begin, we must first establish some relevant background. 18 KRISTIN COURTNEY AND DAVID SHERMAN 6.1. The WEP, LLP, and some results of Kirchberg. For two C∗-algebras A and B, A ⊗max B and A ⊗min B refer to the completion of the algebraic tensor product A ⊙ B of A and B with respect to the maximal and minimal C∗-norms, respectively. The embedding A ⊙ B → A ⊗min B extends uniquely to a surjective -homomorphism A ⊗max B → A ⊗min B. See [10, Chapter 3] for definitions and other relevant properties, and see [5, Section 2] for a short discussion on universal C∗-algebras and tensor products. Though it is readily verifiable that, for any C∗-algebras A, B and C with A ⊆ B, A ⊗min C ⊆ B ⊗min C, this containment can easily fail for the maximal tensor product because subalgebras tend to have more representations. In [34], Lance characterized embeddings of C∗- algebras that are always preserved under the maximal tensor product in terms of completely positive extensions (see [10, Proposition 3.6.6] for a proof). We say a linear map is positive if it sends positive elements to positive elements and completely positive if this remains true even after amplification with matrix algebras. If a completely positive map is contractive, we call it cpc, and if it is moreover unital, we call it ucp. We give Lance's result as a proposition/definition hybrid. Proposition 6.1. Let A and B be unital C∗-algebras and 1B ∈ A ⊂ B an inclu- sion. The inclusion is called relatively weakly injective if the following equivalent conditions hold. (1) There exists a ucp map φ : B → πu(A)′′ so that φ(a) = πu(a) for all a ∈ A, where (πu, Hu) is the universal representation of A from the Gelfand- Naimark Theorem. (2) For any C∗-algebra C, there is a natural inclusion A ⊗max C ⊆ B ⊗max C. The map φ : B → πu(A)′′ is called a weak expectation. If a C∗-algebra A embeds relatively weakly injectively into B(H) for any Hilbert space H, then it is said to have Lance's Weak Expectation Property (WEP). Proposition 6.2. If A is a projective C∗-algebra and B surjects onto A, then A embeds relatively weakly injectively into B. Indeed, the identity map idA : A → A lifts to a -homomorphism ι : A → B, which must be injective since idA is. The weak expectation B → πu(A) is just the induced surjective -homomorphism B → πu(A). Let A and B be C∗-algebras and J a closed two-sided ideal in B with quotient map π : B → B/J. A cpc map φ : A → B/J is liftable if there is a cpc map φ : A → B such that π ◦ φ = φ. We say a cpc map φ : A → B/J is locally liftable if for any finite dimensional operator system S ⊆ A there is a cpc map φ : S → B such that π ◦ φ = φS. A unital C∗ algebra A has the (local) lifting property or (L)LP if any ucp5 map from A into a quotient C∗-algebra is locally liftable. A non-unital C∗-algebra has the (L)LP if and only if its unitization does. 5Yes, we did unceremoniously shift from cpc maps to ucp maps, but it turns out to be sufficient to restrict ourselves to ucp maps. Using the fact that the set of liftable cpc maps from a separable operator system into a C ∗-quotient is closed ([4, Theorem 6]), one can argue (as in [10, Lemma 13.1.2]) that any ucp map with a (local) cpc lift has a (local) ucp lift. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 19 We can view the LP as projectivity in the category of unital C∗-algebras with (unital) cpc maps. Then the LLP can be thought of as a "localized ucp projectivity" and the WEP as "weak injectivity." Proposition 6.3. [31, Corollaries 2.6 and 3.3] The LLP and WEP pass to relatively weakly injective subalgebras. A key class of examples comes from Kirchberg [32, Lemma 3.3]. Lemma 6.4. For any free group F, C∗(F) has the LLP. Moreover, C∗(F) has the LP if F has countably many generators. With this, we can readily prove the following proposition, which greatly simplifies checking whether or not a given C∗-algebra has the LLP. This is a mild rephrasing of [10, Corollary 13.1.4], and we include a brief argument for the reader's convenience. Proposition 6.5. A unital C∗-algebra A has the LLP if and only if the identity on A is locally liftable. To see why it suffices to lift the identity of A, choose a free group F such that we have a surjection π : C∗(F) → A. Let B be a C∗-algebra with two-sided closed ideal J and φ : A → B/J a ucp map. Suppose S ⊆ A is a finite dimensional operator system, and let ρ : S → C∗(F) be the lift of idAS. Then φ ◦ π : C∗(F) → B/J is a ucp map, which lifts to a ucp map ψ : C∗(F) → B. Our desired lift of φS is ψ ◦ ρ. B ψ φ φ◦π B/J C∗(F) ρ S ⊆ π A Given the above proposition, the following is surely known to the experts. (The special case for A follows immediately from [42, Proposition 2.2].) Proposition 6.6. Any projective C∗-algebra has the LLP. In particular, A has the LLP. Next we give Kirchberg's deep and elegant characterization of Connes' Embed- ding Problem. Below is an augmented and abridged version of [31, Proposition 8.1] (omitting some of the equivalent conditions and adding one that quickly follows). Theorem 6.7 (Kirchberg). The following conjectures are equivalent. (1) Every finite von Neumann algebra with separable predual is embeddable into an ultrapower of the hyperfinite II1-factor. (2) C∗(F) ⊗max C∗(F) ≃ C∗(F) ⊗min C∗(F) canonically (where F is any non- abelian free group). (3) C∗(F) ⊗max C∗(F) is RFD. (4) The LLP implies the WEP. (5) C∗(F) has the WEP. 20 KRISTIN COURTNEY AND DAVID SHERMAN Item (1) is traditionally referred to as Connes' Embedding Problem (CEP). Item (3) is not included in [31]; however, it is equivalent to (2) by virtue of the following proposition, which has often been alluded to in the literature. We briefly sketch an argument here; for a complete proof, see [15, Proposition A.0.1]. Proposition 6.8. Given two RFD C∗-algebras A and B, A ⊗max B is RFD if and only if A ⊗max B = A ⊗min B canonically. By taking tensor products of finite dimensional representations, one can readily verify that the minimal tensor product of two RFD C∗-algebras is again RFD. For the other implication, it will suffice to show that any finite dimensional represen- tation of the maximal tensor product factors through the minimal tensor product. To that end, let A and B be C∗-algebras and π : A ⊗max B → Mn be a finite di- mensional representation with restrictions πA : A → Mn and πn : B → Mn, i.e., πA and πB have commuting ranges and πA⊙B = πA × πB. Then we have the induced -homomorphism πA ⊗ πB : A ⊗min B → πA(A) ⊗min πB(B) ≃ πA(A) ⊗max πB(B). On the other hand, since πA(A) and πB(B) commute, the natural embeddings of πA(A) and πB(B) into Mn induce a representation of πA(A) ⊗max πB(B) such that πA(a)⊗πB(b) 7→ πA(a)πB(b) = π(a⊗b) for each a ∈ A and b ∈ B. The composition of this representation with πA ⊗ πB gives a representation of A ⊗min B that agrees with π on A ⊙ B. 6.2. A versus C∗(F2). Let us first show that A is "smaller" than C∗(F2). Lemma 6.9. We have that A is a quotient of C∗(F2), but not vice versa. Proof. Let x be the canonical generator of A, with x1 + ix2 its decomposition into real and imaginary parts. Then A = C∗(x, 1) = C∗(x1, x2, 1) = C∗(eix1 , eix2), since the spectrum of eixj is contained in a proper arc of the circle, and a continuous branch of −i log(·) takes eixj back to xj . Any C∗-algebra generated by two unitaries is a quotient of C∗(F2). To establish the other statement, it suffices to recall that C∗(F2) is not singly generated (as a unital C∗-algebra). For C∗(F2) surjects onto C(T2), since the latter is generated by two unitaries, and C(T2) is not singly generated because T2 is not homeomorphic to a compact subset of the plane. (cid:3) Although C∗(F2) and A are not isomorphic, they play the same universal role for CEP. This has been observed before, specifically by Pisier in [43, Proposition 16.13], where he shows that CEP has an affirmative answer if and only if A has the WEP (compare with Theorem 6.7(5)). It turns out that this can be linked to a strong structural relationship between the two algebras. Theorem 6.10. A and C∗(F2) embed relatively weakly injectively into one another. Proof. We will use Proposition 6.1(1); in this case it will turn out that the required ucp maps can be taken to be -homomorphisms. From Lemma 6.9, there exists a surjection π : C∗(F2) → A. So, A embeds relatively weakly injectively into C∗(F2) by Proposition 6.2. For the other relatively weakly injective embedding, let πu : C∗(F2) → B(Hu) be the universal representation of C∗(F2), and let U1, U2 ∈ B(Hu) be the images of the generators under πu. Let A1, A2 ∈ B(Hu) be self-adjoint elements such that Uj = eiAj and THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 21 C∗(U1, U2) ⊆ C∗(A1, A2) = C∗( A1 α + i A2 α ) ⊆ C∗(U1, U2)′′, where α = kA1 + iA2k. Let x = x1 + ix2 be the canonical generator of A as before. Then eiαxj are unitaries in A and C∗(eiαx1 , eiαx2) ⊆ C∗(x1, x2) = A. By universality, there exist surjective unital -homomorphisms, φ : C∗(U1, U2) → C∗(eiαx1 , eiαx2) and ψ : A → C∗( A1 α + i A2 α ) with φ(Uj) = eiαxj and ψ(xj ) = Aj α . Then ψ ◦ φ(Uj) = ψ(eiαxj ) = eiα( 1 α Aj) = eiAj = Uj. It follows that ψ ◦ φ = idC ∗(U1,U2), and this completes the proof. (cid:3) We now leverage relative weak injectivity to get the following characterizations of CEP. Items (1) and (3) were deduced differently by Pisier [43, Proposition 16.13] and Blecher-Duncan [6, Section 6.1], respectively. Theorem 6.11. The following are equivalent to CEP. (1) A has the WEP. (2) Projectivity implies the WEP. (3) A ⊗max A ≃ A ⊗min A canonically. (4) A ⊗max A is RFD. (5) A⊗max A has a separating family of finite-dimensional representations that map the two generators x ⊗ 1 and 1 ⊗ x to nilpotent matrices. Proof. (2) ⇒ (1) ⇐⇒ CEP ⇒ (2): Condition (2) above implies (1) above by the fact that A is projective (Proposition 4.1). By Theorem 6.10 and Proposition 6.3, A has the WEP if and only if C∗(F2) does, which is item (5) in Theorem 6.7. Item (4) in Theorem 6.7 implies (2) above because projective algebras have the LLP (Proposition 6.6). (3) ⇐⇒ (4) ⇐⇒ CEP: The equivalence of (3) and (4) follows from Proposition 6.8. With Theorem 6.10 and a few applications of Lance's characterization of relatively weakly injective embeddings (Proposition 6.1), we have the embeddings and A ⊗max A ⊂ C∗(F2) ⊗max C∗(F2) C∗(F2) ⊗max C∗(F2) ⊂ A ⊗max A. Since residual finite dimensionality passes to subalgebras, we have that (4) is equiv- alent to C∗(F2) ⊗max C∗(F2) being RFD (item (3) of Theorem 6.7). (4) ⇐⇒ (5): The reverse direction is obvious, so we assume (4) and show (5). Since (3) and (4) have already been shown to be equivalent, it suffices to prove the claim in (5) for A ⊗min A, and this essentially amounts to the argument that the minimal tensor product of two RFD C∗-algebras is RFD. By Theorem 5.1, for i = 1, 2, there exist separating families of finite-dimensional representations {σ(i) n maps the generator to a nilpotent matrix. For n, m ≥ 1, let πn,m = σ(1) m denote the -homomorphism on A ⊗min A that maps a ⊗ b to σ(1) m (b). Then {πn,m}m,n∈N form the desired family of representations. (cid:3) n }n∈N of A where each σ(i) n ⊗ σ(2) n (a) ⊗ σ(2) 22 KRISTIN COURTNEY AND DAVID SHERMAN Remark 6.12. To highlight the difficulty in proving part (3) of Theorem 6.11, con- sider the following. Because residual finite dimensionality is preserved by exten- sions, A ⊗max A is RFD if and only if A0 ⊗max A0 is. Moreover, by Proposition 6.8 A0 ⊗max A0 is RFD if and only if it is canonically isomorphic to A0 ⊗min A0. How- ever, recall that A0 is contractible, and contractibility is preserved under taking tensor products (just take the homotopy between the zero map and the identity on the contractible C∗-algebra and tensor it with the identity on the other C∗-algebra). In particular, this means that both A0 ⊗min A0 and A0 ⊗max A0 are contractible, which makes them very difficult to distinguish with topological invariants. The authors would like to thank Hannes Thiel for clarifying some points on this. Remark 6.13. Using their universality, one can identify A⊗max A with the universal unital C∗-algebra generated by a pair of doubly commuting contractions (i.e., the generators commute and commute with each other's adjoints). In [13] it was asked whether the universal C∗-algebra generated by a pair of commuting contractions, which can be identified with the maximal C∗-algebra for the bidisk algebra [7, Example 2.3], is RFD. An affirmative answer would amount to a noncommutative -polynomial analogue (in the spirit of (3.2)) to Ando's von Neumann Inequality from [1]. It does not seem to be known whether this question is easier than the question for A ⊗max A. The universal (unital) C∗-algebra generated by a pair of unrelated contractions is a full (unital) free product of RFD C∗-algebras, which is RFD by [19, Theorem 3.2]. Remark 6.14. Though the question of whether or not A ⊗max A is RFD proves to be quite challenging, the proof of [10, Proposition 7.4.5] can be readily adapted to show that A ⊗max A is quasidiagonal. 7. C∗-algebras generated by multiple universal contractions We conclude with a few remarks on C∗-algebras generated by two or more uni- versal contractions. B(H)m an m-row contraction if kP SiS∗ 7.1. Universal row contractions. We call an m-tuple of operators (S1, ..., Sm) ∈ i k ≤ 1. (Note that the Si are not assumed to be commuting.) Such tuples have received considerable attention in multivari- ate operator theory, especially with regard to generalizing the commutant lifting theorem and von Neumann's inequality (e.g., [11], [20], [21], [44], and [45]). In this section, we generalize Theorems 3.1 and 5.1 to universal C∗-algebras generated by m-row contractions6: Rm := C∗ m uy1, ..., ym :(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 ≤ 1+ . yiy∗ i(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Some of the techniques we have developed for A = R1 can be generalized to prove the same results for Rm. In particular, where before we had von Neumann-type inequalities for noncommutative -polynomials in one variable, in this section we obtain a noncommutative -polynomial analogue to Popescu's von Neumann in- equality for row contractions [45, Theorem 2.1]. 6Note that the row contraction relation automatically enforces a norm condition on the gen- erators. As we mentioned in Section 4, this will guarantee that these universal C ∗-algebras exist (and are separably acting). THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 23 Proposition 7.1. Fix 1 < m < ∞ and let {y1, ..., ym} denote the generators of Rm. Then for any 1 ≤ k < m, any subset {yi1, ..., yik } ⊆ {y1, ..., ym} forms a jik = 1, and, moreover, universal k-row contraction. each yi is a universal contraction. j=1 yji y∗ In particular, kPk Proof. Since k m 0 ≤ Xj=1 yjiy∗ ji ≤ yjy∗ j , Xj=1 it follows that the {yi1, ..., yik } form a k-row contraction. For any k-row contraction (Z1, ..., Zk), the restriction to C∗(yi1 , ..., yik ) of the map Rm → C∗(Z1, ..., Zk, 1) sending yij → Zj for 1 ≤ j ≤ k and the rest of the generators to 0 is a surjection. (cid:3) By [36, Theorem 5.1], Rm is projective for 1 ≤ m < ∞, just like A. From this it follows that Rm is RFD (see [13, Example 3] for a more direct proof). It turns out we can say moreover that the noncommutative -polynomials in the generators of Rm form a dense subset of elements that actually attain their norms under some finite-dimensional representation (just as with A). Theorem 7.2. Fix 1 ≤ m < ∞, and let q be a noncommutative -polynomial in m variables of degree d. Then there exists a representation π of Rm on a (2m)d+1- dimensional Hilbert space such that kπ(q(y1, ..., ym))k = kq(y1, ..., ym)k. Proof. To prove this, we will need to change little from the proof of Theorem 3.1. The key there was that compressions of π0(x) are still contractions, which means they induce representations of A. Hence, the argument can be adapted for any universal C∗-algebra C∗hGRi with a finite set of generators and with relations that are preserved under compressions. All we have to prove is that Rm is such an algebra. To that end, suppose Y1, ..., Ym ∈ B(H) form a row contraction, and let P ∈ B(H) be a projection. Since P ≤ IH, we have that YiP Y ∗ for each 1 ≤ i ≤ m. Then i ≤ YiY ∗ i m (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 P YiP Y ∗ YiP Y ∗ YiY ∗ ≤ 1. (cid:3) i P(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) m ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 i (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) m ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 i (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Now we turn our attention back to nilpotents. In Section 5, we gave three arguments showing that the universal contraction can be faithfully represented as a countable direct sum of nilpotent matrices. The third can be generalized to show the same for Rm, for 1 ≤ m < ∞. The main idea of the proof is to establish an "exactness"-type result like Theorem 5.9, which says that any faithful representation of Rm "asymptotically factorizes" though the family of universal C∗-algebras generated by order n nilpotents that form row contractions: Rm,n = C∗ u(cid:28)y1,n, ..., ym,n : kyi,nk ≤ 1, yn i,n = 0, i = 1, ..., m, i=1 yi,ny∗ i,nk ≤ 1 (cid:29) Just as in the proof of Theorem 5.9, we will first need to know that any faithful representation of a universal row contraction is the norm limit of row contractions whose entries are each nilpotent. To that end, we have the following lemma., kPm Lemma 7.3. Let 1 ≤ m < ∞ and π : Rm → B(H) a faithful representation on a separable Hilbert space. Then, for each 1 ≤ j ≤ m, there exists a sequence 24 KRISTIN COURTNEY AND DAVID SHERMAN (Tj,n)n≥1 in B(H) that converges in norm to π(yj ), and for each n ≥ 1, T n and (T1,n, ..., Tm,n) forms a row contraction. j,n = 0 Proof. By Propositions 7.1 and 5.8, each π(yj ) is the norm limit of a sequence of nilpotent contractions (Nj,n)n≥1, which we can assume satisfy N n j,n = 0 for all j,nk−1/2} (where we n ≥ 1. Now for each n ≥ 1, we let αn = min{1, kPm assume αn = 1 if Nj,n = 0 for all 1 ≤ j ≤ m), and define j=1 Nj,nN ∗ Tj,n := αnNj,n for each 1 ≤ j ≤ m. We immediately have T n j,n = 0 for all n ≥ 1 and 1 ≤ j ≤ m. Moreover, since Nj,n → π(yj ) for each 1 ≤ j ≤ m, it follows from Proposition 7.1 that αn → 1, and hence Tj,n → π(yj) for each 1 ≤ j ≤ m. Finally we check that (T1,n, ..., Tm,n) forms a row contraction for each n ≥ 1. If j,nk > 1, then j,n = αnN n j=1 Nj,nN ∗ m m (N1,n, ..., Nm,n) is a row contraction, then αn = 1. If kPm −1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) j,n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 j,n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) j,n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) α2 nNj,nN ∗ Nj,nN ∗ Tj,nT ∗ m m Xj=1 Nj,nN ∗ j,n(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = 1. (cid:3) Theorem 7.4. For 1 ≤ m < ∞, Rm has a separating family of finite-dimensional representations that map each generator to a nilpotent. In particular, there exist nilpotent matrices Nk1, ..., Nkm ∈ Mjk , k ∈ N, such that (⊕kNk1 , ..., ⊕kNkm ) is a universal m-row contraction. Proof. Let π be a faithful representation of Rm on a separable Hilbert space H, and let (Ti,n)n≥1 be the sequences of nilpotents guaranteed by Lemma 7.3 for each 1 ≤ i ≤ m. By universality, for each n ≥ 1, we can define a surjective -homomorphsim ψn from kPm Rm,n = C∗ u(cid:28)y1,n, ..., ym,n : kyi,nk ≤ 1, yn i,n = 0, i = 1, ..., m, i=1 yi,ny∗ i,nk ≤ 1 (cid:29) to C∗(1, T1,n, ..., Tm,n) by sending yi,n 7→ Ti,n for each i = 1, ..., m. Let φn : Rm → Rm,n denote the surjection induced by sending generators to generators. Then, for each i = 1, ..., m, kψn ◦ φn(yi) − π(yi)k → 0, which means that ψn ◦ φn converge pointwise in norm to π. In particular, the {φn} separate the points of Rm. By [37, Theorem 2.3], each Rm,n is projective and hence RFD by Proposition 4.4. So, composing the {φn} with the separating families of finite-dimensional representations of the Rm,n's gives us the desired family of representations of Rm. The matrices in the statement of the theorem come from taking the direct sum (cid:3) of this family of representations. Remark 7.5. Since the generators of Rm are all universal contractions, Example 5.2 also shows that Theorem 7.2 will not hold if we also want the generators to be mapped to nilpotents. We conclude this section by rephrasing the results in terms of a Popescu-von Neumann inequality for noncommutative -polynomials on row contractions. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 25 Corollary 7.6. Let q be a noncommutative -polynomial in m variables of degree d and (T1, ..., Tm) a row contraction. Then kq(T1, ..., Tm)k ≤ sup{kq(A1, ..., Am)k : (A1, ..., Am) ∈ (Mm n )≤1, n ≥ 1} = max{kq(A1, ..., Am)k : (A1, ..., Am) ∈ (Mm = sup{kq(N1, ..., Nm)k : (N1, ..., Nm) ∈ (Mm (2m)d+1 )≤1} n )≤1, N n i = 0, 1 ≤ i ≤ m, n ≥ 1}. 7.2. Universal Pythagorean C∗-algebras. An m-tuple of operators (S1, ..., Sm) in B(H)m satisfying the identity m S∗ i Si = 1 Xi=1 is often called a column isometry. Such operators have recently been utilized in [9] to construct several interesting unitary representations of Thompson's groups F and T (and their n-ary versions Fn and Tn). We adopt the language of [9] in calling Pm = C∗ ua1, ..., am : universal Pythagorean C∗-algebras. a∗ i ai = 1+ , m Xi=1 Below, we adapt Choi's technique from [12, Theorem 7] to show that each Pm is RFD for m > 1. Note that this is not true for m = 1. The universal C∗-algebra of a single isometry is not RFD because every finite-dimensional isometry is unitary. Theorem 7.7. For each m > 1, Pm is RFD. Proof. First, we faithfully and nondegenerately represent Pm on a separable Hilbert space H, denoting the generators by A1, ..., Am. Let (Pn) be an increasing sequence of projections, each of rank n, such that Pn converges strongly to IH. For each n ≥ 1 and 1 ≤ i ≤ m, let B(n) , and note that Cn ≤ Pn for each n ≥ 1. Define operators A(n) i=1 B(n)∗ i B(n) 1 , ..., A(n) i m ∈ B(PnH ⊕ PnH) by i = PnAiPn and Cn =Pm 2 =(cid:18) (Pn − Cn)1/2 B(n) 0 2 A(n) 1 =(cid:18)B(n) 1 0 0 Pn(cid:19) , A(n) for each 2 < i ≤ m. For each n, we compute 0(cid:19) , and A(n) i =(cid:18)B(n) i 0 0 0(cid:19) 0 0(cid:19) i A(n) A(n)∗ i m Xi=1 1 0 0 1 B(n) =(cid:18)B(n)∗ =(cid:18)Pn Pn(cid:19) +(cid:18)B(n)∗ Pn(cid:19) = IPnH⊕PnH. 0 0 2 B(n) 2 + Pn − Cn 0 0 0(cid:19) +(cid:18)Pm i=3 B(n)∗ i B(n) 0 i 1 , ..., A(n) So, (A(n) M2n. For each n ≥ 1, let πn : C∗(IH, A1, ..., Am) → C∗(Pn ⊕ Pn, A(n) B(PnH ⊕ PnH) be the representation induced by Ai 7→ A(n) It remains to show that π = ⊕nπn is an isometry. m ) satisfies the column isometry identity in B(PnH ⊕ PnH) ≃ m ) ⊂ for each 1 ≤ i ≤ m. 1 , ..., A(n) i We now consider PnH ⊕ PnH as a subspace of H ⊕ H in the obvious way, thereby as an operator on H ⊕ H (acting as zero on (PnH ⊕ PnH)⊥). viewing each A(n) i 26 KRISTIN COURTNEY AND DAVID SHERMAN Since the Pn converge -strongly to IH, we have that A(n) to A1 ⊕ 1H and A(n) noncommutative -polynomial q in two variables, converges -strongly converges -strongly to Ai ⊕ 0 for 2 ≤ i ≤ m. So, for any 1 i q(A(n) 1 , ..., A(n) m ) S ∗OT−−−−→ q(A1 ⊕ IH, A2 ⊕ 0, ..., Am ⊕ 0). Note that (A1 ⊕ IH, A2 ⊕ 0, ..., Am ⊕ 0) is also a universal m-column isometry. Indeed, it clearly satisfies the column isometry identity, and, since kq(A1 ⊕ IH, A2 ⊕ 0, ..., Am ⊕ 0)k = kq(A1, ..., Am)⊕ q(IH, 0, ..., 0)k ≥ kq(A1, ..., Am)k for any noncommutative -polynomial q in two variables, the natural surjection induced by mapping A1 7→ A(n) i ⊕ 0 for 2 ≤ i ≤ m is isometric. Now, let q be a noncommutative -polynomial in m variables and ǫ > 0. Assume 1 ⊕ 1H and Ai 7→ A(n) kq(A1, ..., Am)k = 1. Then kq(A1 ⊕ IH, A2 ⊕ 0, ..., Am ⊕ 0)k = 1 and for all sufficiently large n. Hence, kq(A(n) 1 , ..., A(n) m )k ≥ 1 − ǫ, kπ(q(A1, ..., Am))k ≥ kπn(q(A1, ..., Am))k = kq(A(n) 1 , ..., A(n) m )k ≥ 1 − ǫ for all sufficiently large n. Since ǫ was arbitrary, π is isometric on the set of noncommutative -polynomials on A1, ..., Am, which, by continuity, means it is isometric on C∗(IH, A1, ..., Am) ≃ Pm. (cid:3) References [1] T. Ando, On a pair of commutative contractions, Acta Sci. Math. (Szeged) 24 (1963), 88 -- 90. [2] C. Apostol, C. Foias, and D. Voiculescu, On the norm-closure of nilpotents, II, Rev. Roum. Math. Pures et Appl. 19 (1974), 549 -- 577. [3] W. Arveson, Subalgebras of C ∗-algebras, II, Acta Math. 128 (1972), 271 -- 308. [4] W. Arveson, Notes on extensions of C ∗-algebras, Duke Math. J. 44 (1977), 329 -- 355. [5] B. Blackadar, Shape theory for C ∗-algebras, Math. Scand. 56 (1985), 249 -- 275. [6] D. P. Blecher and B. Duncan, Nuclearity-related properties for nonselfadjoint operator alge- bras, J. Operator Theory 65 (2011), 47 -- 70. [7] D. P. Blecher and V. I. Paulsen, Explicit construction of universal operator algebras and applications to polynomial factorization, Proc. Amer. Math. Soc. 112 (1991), 839 -- 850. [8] B. Brenken and Z. Niu, The C ∗-algebra of a partial isometry, Proc. Amer. Math. Soc. 140 (2012), 199 -- 206. [9] A. Brothier and V. Jones, Pythagorean representations of Thompson's groups, preprint, 2018 arXiv: 1807.06215. [10] N. Brown and N. Ozawa, C ∗-Algebras and Finite-Dimensional Approximations, Graduate Studies in Mathematics 88, American Mathematical Society, Providence, 2008. [11] J. Bunce, Models for n-tuples of non-commuting operators, J. Funct. Anal. 57 (1984), 21 -- 30. [12] M. Choi, The full C ∗-algebra of the free group on two generators, Pacific J. Math. 87 (1980), 41 -- 48. [13] R. Clouatre and C. Ramsey, Residually finite-dimensional operator algebras, preprint, 2018 arXiv:1806.00038. [14] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. Math. (2) 104 (1976), 73 -- 115. [15] K. Courtney, C ∗-algebras and their finite-dimensional representations, PhD thesis (2018), University of Virginia, Charlottesville, United States. [16] K. Courtney and T. Shulman, Elements of C ∗-algebras attaining their norm in a finite- dimensional representation, Canad. J. Math. to appear. arXiv:1707.01949. [17] R. E. Curto and D. A. Herrero, On closures of joint similarity orbits, Integral Equations Operator Theory 8 (1985), 489 -- 556. [18] J. Dixmier, C ∗-algebras, North Holland Publ. Co., Amsterdam, 1977. THE UNIVERSAL C ∗-ALGEBRA OF A CONTRACTION 27 [19] R. Exel and T. Loring, Finite-dimensional representations of free product C ∗-algebras, Int. J. Math, 3 (1992), 469 -- 476. [20] A. Frazho, Models for non-commuting operators, J. Funct. Anal. 48 (1982), 1 -- 11. [21] A. Frazho, Complements to models for noncommuting operators, J. Funct. Anal. 59 (1984), 445 -- 461. [22] T. Fritz, T. Netzer, and A. Thom, Can you compute the operator norm? Proc. Amer. Math. Soc. 142 (2014), 4265 -- 4276. [23] U. Haagerup and P. de la Harpe, The numerical radius of a nilpotent operator on a Hilbert space, Proc. Amer. Math. Soc. 115 (1992), 371 -- 379. [24] D. Hadwin, An operator-valued spectrum, Indiana University Math. J. 26 (1977), 329 -- 340. [25] D. Hadwin, An asymptotic double commutant theorem for C ∗-algebras, Trans. Amer. Math. Soc. 244 (1978), 273 -- 297. [26] D. Hadwin, Nonseparable approximate equivalence, Trans. Amer. Math. Soc. 266 (1981), 203 -- 231. [27] D. Hadwin, A lifting characterization of RFD C ∗-algebras, Math. Scand. 115 (2014), 85 -- 95. [28] D. Hadwin and T. Shulman, Variations of projectivity for C ∗-algebras, preprint, arXiv:1709.01379. [29] P. R. Halmos, A Hilbert Space Problem Book, Second Edition, Springer-Verlag, New York, 1982. [30] D. A. Herrero, Quasidiagonality, similarity and approximation by nilpotent operators, Indiana Univ. Math. J. 30 (1981), 199 -- 232. [31] E. Kirchberg, On non-semisplit extensions, tensor products and exactness of group C ∗- algebras, Invent. Math. 112 (1993), 449 -- 489. [32] E. Kirchberg, Commutants of unitaries in UHF algebras and functorial properties of exact- ness, J. Reine Angew. Math. 452 (1994), 39 -- 78. [33] C. S. Kubrusly, An Introduction to Models and Decompositions in Operator Theory, Birkhauser Boston, Inc., Boston, MA, 1997. [34] E. C. Lance, On nuclear C ∗-algebras, J. Funct. Anal. 12 (1973), 157 -- 176. [35] T. Loring, Lifting Solutions to Perturbing Problems in C ∗-Algebras, Fields Institutue Mono- graphs 8, American Mathematical Society, Providence, 1997. [36] T. Loring and T. Shulman, Non-commutative semialgebraic sets and associated lifting prob- lems, Trans. Amer. Math. Soc. 364 (2012), 721 -- 744. [37] T. Loring and T. Shulman, Non-commutative semialgebraic sets in nilpotent variables, New York J. Math. 14 (2012), 361 -- 372. [38] T. Loring and T. Shulman, Lifting algebraic contractions in C ∗-algebras, Oper. Theory Adv. Appl. 233 (2014), 85-92. [39] B. B. Morrel and P. S. Muhly, Centered operators, Studia Math. 51 (1974), 251 -- 263. [40] J. von Neumann, Eine Spektraltheorie fur allgemeine Operatoren eines unitaren Raumes, Math. Nachr. 4 (1951), 258 -- 281. [41] T. Oikhberg, A note on universal operators, in: Ordered Structures and Applications, pp. 339-347, Trends Math., Birkhauser/Springer, Cham, 2016. [42] N. Ozawa, On the lifting property for universal C ∗-algebras of operator spaces, J. Operator Theory, 46 (2001), 579-591. [43] G. Pisier, Introduction to Operator Space Theory, London Mathematical Society Lecture Note Series 294, Cambridge University Press, 2003. [44] G. Popescu, Isometric dilations for infinite sequences of noncommuting operators, Trans. Amer. Math. Soc. 316 (1989), 523 -- 536. [45] G. Popescu, Von Neumann inequality for (B(H)n )1, Math. Scand. 68 (1991), 292 -- 304. [46] H. Radjavi and P. Rosenthal, Invariant Subspaces, Springer-Verlag, New York, 1973. [47] G.-C. Rota, Note on the invariant subspaces of linear operators, Rend. Circ. Mat. Palermo (2) 8 (1959), 182 -- 184. [48] G.-C. Rota, On models for linear operators, Comm. Pure Appl. Math. 13 (1960), 469 -- 472. [49] T. Shulman, Lifting of nilpotent contractions, Bull. London Math. Soc. 40 (2008), 1002 -- 1006. [50] D. Topping, UHF algebras are singly generated, Math. Scand. 22 (1968), 224 -- 226. [51] D. Voiculescu, A non-commutative Weyl-von Neumann theorem, Rev. Roum. Math. Pures et Appl. 21 (1976), 97 -- 113. [52] D. Voiculescu, A note on quasi-diagonal C ∗-algebras and homotopy, Duke Math. J. 62 (1991), 267 -- 271. 28 KRISTIN COURTNEY AND DAVID SHERMAN Department of Mathematics, University of Virginia, 141 Cabell Drive, Kerchof Hall, P.O. Box 400137, Charlottesville, VA 22904-4137 USA Current address: Mathematical Institute, WWU Munster, Einsteinstr. 62, 48149 Munster, Germany E-mail address: [email protected] Department of Mathematics, University of Virginia, 141 Cabell Drive, Kerchof Hall, P.O. Box 400137, Charlottesville, VA 22904-4137 USA E-mail address: [email protected]
1708.03258	1	1708	2017-08-10T15:33:54	E-theory Spectra for graded C*-algebras	[ "math.OA", "math.AT" ]	This paper brings together C-algebras and algebraic topology in terms of viewing a C-algebraic invariant in terms of a topological spectrum. E-theory, E(A,B), is a bivariant functor in the sense that is a cohomology functor in the first variable and a homology functor in the second variable but underlying goes from the category of separable C-algebras and -homomorphisms to the category of abelian groups and group homomorphisms. Here we create a generalisation of a orthogonal spectrum to quasi-topological spaces for E-theory. This includes a rich product structure in the context of graded separable C*-algebras.	math.OA	math	E-theory Spectra for graded C ∗-algebras Sarah L. Browne May 14, 2018 Abstract This paper brings together C ∗-algebras and algebraic topology in terms of viewing a C ∗-algebraic invariant in terms of a topological spectrum. E-theory, E(A, B), is a bivariant functor in the sense that is a cohomology functor in the first variable and a homology functor in the second variable but underlying goes from the category of separable C ∗-algebras and ∗-homomorphisms to the category of abelian groups and group homomorphisms. Here we create a generalisation of a orthogonal spectrum to quasi-topological spaces for E-theory . This includes a rich product structure in the context of graded separable C ∗-algebras. Introduction Separable C ∗-algebras are analytical objects used in non-commutative geometry. E-theory, E(A, B), is a invariant of C ∗-algebras and is a bivariant functor from the category of separable C ∗-algebras and ∗-homomorphisms to the category of abelian groups and group homomor- phisms. Here we will only consider real graded C ∗-algebras since then we have a useful tool for application in differential geometry. E-theory is a cohomology theory in its first variable and a homology theory in its second variable. It was first defined by Higson [8] in a categor- ical manner and Connes and Higson [4] gave a concrete description using homotopy classes of certain morphisms of C ∗-algebras. This will form section 1. Spectrum is a topological notion used to represent stable homotopy theories. Here we re- quire the notion of a quasi-spectrum, which is considered as a spectrum over quasi-topological spaces. We incorporate an action of the orthogonal group and create an orthogonal quasi- spectrum. This has a rich product structure as we will see. The main construction encodes the Bott periodicity of E-theory proving in the complex graded case by Guentner-Higson [6] and in the real case by Browne [2], as an Ω-quasi- spectrum coming from weak equivalences. Additionally we also encode the product structure of E-theory. This then will give us a stable homotopy theory. It is worth noting that the spectrum will be dependent on C ∗-algebras just like the K-theory spectrum for C ∗-categories is dependent on C ∗-categories defined by Paul D. Mitchener [11]. The author likes to think of this as a "local" spectrum and it is suited for possible applications to positive scalar curvature by generalising work of Weiss-Williams [14] 1 Further we include a way of connecting K-theory spectrum and K-homology spectrum using the relations of E-theory with both K-theory and K-homology in the final section of the paper. Acknowledgements The author would like to thank their PhD supervisor, Paul Mitchener for all the valuable meetings and discussions during her PhD. The author would also like to thank the EPSRC for funding her PhD which made this possible. Lastly the author wishes to thank many mathematicians for conversations and particularly Jamie Gabe for his insight during a visit at the University of Southampton. 1 E-theory This section details the analytic necessities which we need to inlude in our definition of the quasi-spectrum for E-theory. A complex C ∗-algebra is a complex Banach space with involution ∗ satisfying for all a ∈ A, the C ∗-identity a∗a = a2. A real C ∗-algebra is a real Banach space with an involution, satisfying th C ∗-identity and additonally that the element 1 + a∗a is invertible in A for all a ∈ A. A C ∗-algebra is called separable if its underlying topological space has a countable dense subset and all of our C ∗-algebras will be separable. A map f : A → B of real C ∗-algebras is called a ∗-homomorphism and is an algebra homomorphism and satisfies f (a∗) = (f (a))∗ for all a ∈ A. A grading on a C ∗- algebra A is an automorphism δA : A → A such that δ2 A = 1. We can also think of a grading by allocating an even and odd notion on A. That is we have A = Aeven ⊕ Aodd, where Aeven = {a ∈ A δA(a) = a} and Aodd = {a ∈ A δA(a) = −a}. Then additionally we define the degree of an element a ∈ A by: deg(a) =(0, 1, if a ∈ Aeven if a ∈ Aodd. An important example of a graded real C ∗-algebra is the algebra of continuous real-valued functions vanishing at infinity, S = C0(R), under the supremum norm and grading given by δ(f )(x) = f (−x) for all x ∈ R. Denote this algebra by S. Also we can consider the suspension of a C ∗-algebra ΣA = {f : [0, 1] → A f (0) = f (1) = 0}. If A is graded then ΣA has grading coming from A. Also we can consider the n-fold suspen- sion, ΣnA = Σn−1ΣA. Now we should note that a graded ∗-homomorphism is a ∗-homomorphism that preserves the grading. That is, if an element is even then its image is even and if an element is odd then its image is odd. Furthermore we can define a different type of morphism between C ∗-algebras:- 2 Definition 1.1. Let A and B be real graded C ∗-algebras with gradings δA and δB respec- tively. A graded asymptotic morphism ϕ : A 99K B is a family of functions {ϕt}t∈[1,∞) such that: 1. the map t 7→ ϕt(a), from [1, ∞) to B is continuous and bounded for each a ∈ A, 2. limt→∞ ϕt(ab) − ϕt(a)ϕt(b) = 0, for each a, b ∈ A, 3. limt→∞ ϕt(a + λb) − ϕt(a) − λϕt(b) = 0, for each a, b ∈ A, λ ∈ R, 4. limt→∞ ϕt(a∗) − ϕt(a)∗ = 0, for each a ∈ A, 5. limt→∞ ϕt(δA(a)) − δB(ϕt(a)) = 0 for each a ∈ A. Denote the set of these by Asyg(A, B). It will be useful to consider the graded tensor product: Definition 1.2. Let A and B be graded C ∗-algebras with gradings δA and δB respectively. Then define Ab⊗B to be the completion of the algebraic tensor product of A and B in the ϕ(ai)ψ(bi) norm ai ⊗ bi = sup ϕ,ψ where ϕ : A 99K C, ψ : A 99K C are graded asymptotic morphisms to a common graded C ∗-algebra C. We equip Ab⊗B with involution, multiplication and grading defined by: 1. (ab⊗b)∗ = (−1)deg(a)deg(b)a∗ ⊗ b∗ 2. (ab⊗b)(cb⊗d) = (−1)deg(b)deg(c)(ac ⊗ bd) 3. γ(ab⊗b) = α(a) ⊗ β(b) Extending by linearity gives Ab⊗B. We have the following result, Lemma 4.5 in [7], for asymptotic morphisms: Lemma 1.3. Let ϕ : A1 99K A2 and ψ : B1 99K B2 be (graded) asymptotic morphisms, then the compositions Xi Xi and A1b⊗B1 A1b⊗B1 ϕ b⊗1 −−→ A2b⊗B1 1 b⊗ψ−−→ A1b⊗B2 1 b⊗ψ −−→ A2b⊗B2, ϕ b⊗1−−→ A2b⊗B2, are equal. The 1 symbolises the relevant identity morphism. Now E-theory is built out of homotopy classes so we need the notion of a homotopy of graded asymptotic morphism:- Definition 1.4. A homotopy of graded asymptotic morphisms ϕt, ψt : A 99K B is a graded asymptotic homomorphism θt : A 99K C([0, 1], B) such that θt(a)(0) = ϕt(a) and θt(a)(1) = ψt(a). 3 Denote the set of homootopy classes of asymptotic morphisms from A to B by JA, BK. We require the next remark to impose a particular grading on compact operators. Remark 1.5. Let H be a Hilbert space equipped with the orthogonal decomposition H = H0 ⊕ H1, where H0 denotes the even elements and H1 denotes the odd elements. Then the C ∗-algebra K(H) of compact operators on such a Hilbert space is graded. For this grading, we consider 2 × 2 matrices of operators where the diagonal matrices are even and the off diagonal ones are odd. That is we have a grading defined by β : K(H) → K(H) β(T ) =(T −T if T is even if T is odd. Then the graded E-theory groups are defined by: En g (A, B) = JSb⊗Ab⊗K(H), ΣnBb⊗K(H)K, where ΣnB for n ≥ 0 is the n-fold suspension of a C ∗-algebra and it's grading comes from B, and where K(H) denotes the compact operators on some real graded Hilbert space with grading as detailed in Remark 1.5. If we have a graded ∗-homomorphisms ϕ, ψ : A → B, we can also form homotopy classes of these, by setting ϕ = ϕt, ψ = ψt for all t ∈ [1, ∞). Denote the set of homotopy classes of ∗-homomorphisms from A to B by [A, B]. It will become important to consider the graded ∗-homomorphism defined in [6] and also detailed in [2]. The idea is to restrict to the set of continuous functions on the interval [−R, R], denoted by SR and use functional calculus and define ∆ : S → Sb⊗S, f (XRb⊗1 + 1b⊗XR) ∈ SRb⊗SR, and thereafter obtain a graded ∗-homomorphism. Now in E-theory we have a Bott periodicity result that can be found in [2]. Let F = R or C. Before we state the result we note that Fn,0 is the Clifford algebra on n generators e1, e2, . . . en such that e2 i = 1 and eiej = −ejei for all i, j. Then we have the following: Theorem 1.6. There is a ∗-homomorphim b : S → Σb⊗F1,0 inducing an isomorphism Also for Section 4, we have the following definitions coming from the E-theory definition. E(A, B) ∼= E(A, ΣBb⊗F1,0). The K-theory groups are given by: and the K-homology is given by Kn(A) = [S, ΣnAb⊗K(H)] ∼= En(R, ΣnA), Khom(A) = En(A, R). 4 2 Quasi-topological spaces In this section we generalise the notion of orthogonal spectra to quasi-topological spaces. In order to do this, we have to define the notion of a quasi-continuous group action and further prove we have a symmetric monoidal structure on the category of quasi-orthogonal sequences which we also define. Firstly we recall some notions of quasi-topological spaces by Spanier [12] which we will need as it is not known if we can put a standard topology on the set of asymptotic morphisms. Definition 2.1. A quasi-topology on a set X, is a collection of sets of maps from C to X for each compact Hausdorff space C, written Q(C, X), called quasi-continuous and satisfying: • any constant map C → X belongs to Q(C, X), • if f : C1 → C2 is a map of compact Hausdorff spaces and g ∈ Q(C2, X) then gf ∈ Q(C1, X), • for a disjoint union C = C1 ∐ C2 of closed compact Hausdorff spaces, a map g : C → X is contained in Q(C, X) if and only if gCi ∈ Q(Ci, X) for i = 1, 2, • for every f : C1 → C2 surjective map of compact Hausdorff spaces, then a map h : C2 → X is quasi-continuous if h ◦ f is quasi-continuous. A quasi-topological space is a set X endowed with a quasi-topology as described above. If X is a topological space we can obtain a quasi-topology on X by considering Q(C, X) as the set of continuous maps from C to X in the topology of X. A map of quasi-topological spaces f : X → Y is called quasi-continuous if g ∈ Q(C, X) implies that the composite f g ∈ Q(C, Y ). Also by the definition of quasi-continuous maps, a composite of quasi-continuous maps is also quasi-continuous. A quasi-homeomorphism f : X → Y between quasi-topological spaces is a quasi-continuous bijection with a quasi- continuous inverse g : Y → X. Definition 2.2. Let X, Y be quasi-topological spaces and f, g : X → Y be quasi-continuous maps. Then a homotopy is a quasi-continuous map such that for all x ∈ X, H(x)(0) = f (x) and H(x)(1) = g(x). H : X → C([0, 1], Y ), The suspension and loop space of a quasi-topological space X are defined similarly to the case of topological spaces by ΣtopX = S1 ∧ X, and ΩX = {µ : S1 → X µ is quasi-continuous and basepoint preserving} and we consider the circle S1 with the quasi-topology that comes from the standard topology on R2. That is, our quasi-continuous maps are the continuous maps from every compact Hausdorff space in to S1 in the topology from R2. Now we check that Σtop and Ω are 5 adjoints in the category where objects are quasi-topological spaces and arrows are quasi- continuous maps. In order to do this, we consider an abstract result and obtain it as a corollary. Proposition 2.3. Let X, Y and Z be quasi-topological spaces. Then F (X ∧ Y, Z) and F (X, F (Y, Z)), are quasi-homeomorphic. Proof. We define α : F (X ∧ Y, Z) → F (X, F (Y, Z)) by ((α(f ))(x))(y) = f (x ∧ y), where f ∈ F (X ∧ Y, Z), x ∈ X and y ∈ Y . Then α is quasi-continuous since f is quasi- continuous. Define β : F (X, F (Y, Z)) → F (X ∧ Y, Z) by (β(g))(x ∧ y) = (g(x))(y), where g ∈ F (X, F (Y, Z)), x ∈ X and y ∈ Y . So β is quasi-continous since g is quasi- continuous. Finally α and β are inverses, so we obtain a quasi-homeomorphism. Corollary 2.4. Σtop and Ω are adjoints in the category of quasi-topological spaces. That is F (ΣtopX, Y ) and F (X, ΩY ) are quasi-homeomorphic. Proof. This follows from Proposition 2.3 since and F (ΣtopX, Y ) = F (X ∧ S1, Y ), F (X, ΩY ) = F (X, F (S1, Y )). Let A, B be C ∗-algebras. Further by work of Dardalat-Meyer [5] we can define a quasi topology on Asy(A, B), the set of asymptotic morphisms from A to B. We define the set of quasi-continuous maps from a compact Hausdorff space Y to Asy(A, B) to be the Asy(A, C(Y, B)), mentioned in [5]. That is, more precisely we have Definition 2.5. For a compact Hausdorff space Y , a map h : Y → Asy(A, B) is quasi- continuous when for each t ∈ [1, ∞) the map eht(a) : Y → B defined by is continuous. eht(a)(y) = h(y)t(a), 6 Now we check that this is a quasi-topology. Proposition 2.6. The set of asymptotic morphisms from A to B, Asy(A, B), is a quasi- topological space when equipped with the above quasi-topology. Proof. We must check the axioms. Let c : Y → Asy(A, B) be constant. Then for y ∈ Y , c(y) = ft : A 99K B for a fixed f . Then we need to show that c is quasi-continuous. That is a ∈ A, y ∈ Y by to show for each map c, the map ect(a) : Y → B is continuous. Then we define this map for ect(a)(y) = c(y)t(a) = ft(a), which is continuous as a function of Y and hence c is continuous. Now let f : Y1 → Y2 be a map of compact Hausdorff spaces and let g : Y2 → Asy(A, B) be quasi-continuous. Then we want to show that gf : Y1 → Asy(A, B) is quasi-continuous. That is, we need to show thatfgf t(a) : Y1 → B is continuous. Now f is continuous and since g is quasi-continuous, we have eachegt(a) : Y2 → B is continuous and Now, egt(a)(y) = g(y)t(a). fgf t(a)(y) =egt(a)f (y), Let Y = Y1 ∐ Y2 of compact Hausdorff spaces. Then we need to show that g : Y → Asy(A, B) is quasi-continuous if and only if gYi is quasi-continuous for i = 1, 2. Suppose which is continuous in Y so is fgf t(a) is continuous, yielding that gf is quasi-continuous. g : Y → Asy(A, B) is quasi-continuous. Then map egt(a) : Y → B defined by continuous function is continuous. Suppose that gYi is quasi-continuous, then (egYi)t(a) : Yi → B is continuous. Then by properties of continuous functions we know that if the restrictions are continuous then the map of the disjoint uniion will be continuous. is continuous. Now by properties of continuous functions we know that the restriction of a egt(a)(y) = g(y)t(a), Finally, we need to check for every surjective map f : Y1 → Y2 of compact Hausdorff spaces that g : Y2 → Asy(A, B) is quasi-continuous if gf : Y1 → Asy(A, B) is quasi-continuous. Then for a particular map f and by the above argument is continuous. Let f : Y1 → Y2 be the identity, then fgf t(a)(y) =egt(a)f (y), fgf t(a)(y) =egt(a)f (y) =egt(a), on Asy(A, B) defined as above. and henceegt(a) is continuous and the result follows. So we do indeed have a quasi-topology For a quasi-topological space X, let X+ denote the space with a basepoint. Proposition 2.7. The quasi-topological spaces ΩAsy(A, B) and Asy(A, ΣB) are quasi-homeomorphic. 7 Proof. By the definition of a quasi-topology, we know that ΩAsy(A, B) : = Q(S1, Asy(A, B))+. Then by the definition 2.5, we have that Q(Y, Asy(A, B))+ = Asy(A, C(Y, B)+), and then that ΩAsy(A, B) = Q(S1, Asy(A, B))+ = Asy(A, C(S1, B)+) = Asy(A, ΣB). The above results hold in the case of graded asymptotic morphisms. 2.1 Group actions The following definition makes sense since a topological group can be viewed as a quasi- topological group. Definition 2.8. Let G be a topological group acting on a quasi-topological space X. Then the group action is called quasi-continuous if the map G × X → X is quasi-continuous. If this is the case we say that the set X is a quasi G-space. Definition 2.9. A map f : X → Y of quasi G-spaces is a quasi G-map if it is G-equivariant. That is, for all g ∈ G, we have Now we consider basepoint preserving group actions. f (gx) = g(f (x)). Proposition 2.10. Let G and H be groups. Let X be a quasi G-space, Y a quasi H-space where the group actions preserve the basepoints of both X and Y . Then there are basepoint preserving actions of G × H on X × Y , X ∨ Y and X ∧ Y . These actions are defined in the obvious way. We now need the notion of a balanced smash product. Definition 2.11. Let X be a right quasi G-space and Y a left quasi G-space, then we can from the balanced smash product X ∧G Y , which is the quotient space X ∧ Y / ∼G where (xg ∧ y) ∼G (x ∧ g−1y) ⇔ (x ∧ y) ∼G (xg ∧ g−1y), for all g ∈ G. 8 Let the equivalence class of x ∧ y be denoted by x ∧G y. Now using these we can construct a left quasi G-space. Let G be a topological group and H a subgroup. Then let X be a based left quasi H- space where G acts by preserving the basepoint. Let G+ = G ∐ {∗}, then we can construct the right quasi G-space denoted G+ ∧H X using the above equivalence classes. Additionally we can actually define a left quasi G-action on this space by the following map: (f, g ∧H x) 7→ f g ∧H x, for all f ∈ G. To prove this is well-defined action it is a formality of using the fact that H is a subgroup of G. Let X and Y be based quasi G-spaces. Then let QG(X, Y ) denote the set of basepoint preserving quasi G-maps. Then we have the following result: Proposition 2.12. Let H be a subgroup of a group G. Let X be a left quasi H-space and Y a left quasi G-space. There there is a natural bijection QH(X, Y ) ←→ QG(G+ ∧H X, Y ). Proof. We define α : QH(X, Y ) → QG(G+ ∧H X, Y ). Let f ∈ QH (X, Y ) and g ∧H x ∈ G+ ∧H X, then we define α(f )(g ∧H x) = gf (x), and we then need to check that α is well-defined and also G-equivariant. Since (g ∧H x) ∼ (gh ∧H h−1x), then α(f )(gh ∧H h−1x) = ghf (h−1x) = ghh−1f (x) since f is a quasi H-map = gf (x) = α(f )(g ∧H x). Then α(f ) is G-equivariant since α(f )(g′g ∧H x) = g′gf (x) = g′(gf (x)) = g′α(f )(g ∧H x), for all g′ ∈ G. Now define β : QG(G+ ∧H X, Y ) → QH(X, Y ). Let k ∈ QG(G+ ∧H X, Y ), x ∈ X, then where e denotes the identity in G. Then β is H-equivariant since β(k)(x) = k(e ∧H x), β(k)(hx) = k(e ∧H hx) = k(h−1h ∧H hx) = k(h ∧H x) by equivalence relations = hk(e ∧H x) as k is a H-map = hβ(k)x. Then is is clear that α and β are inverse maps, and both are natural in X and Y , so the result follows. 9 2.2 Quasi-Orthogonal sequences Let O be the category of finite dimensional real Euclidean inner product spaces and linear isometric isomorphisms where we have objects to be the set and morphisms are obj(O) = {Rn n = 0, 1, . . .} O(A, B) =(O(n), ∅, if A = B = Rn otherwise. It should be noted that this is a small category since the collection of objects is a set. Let T denote the category of quasi-topological spaces with basepoints and quasi-continuous maps. So obj(T ) is the collection of quasi-topological spaces with basepoints and the mor- phisms T (X, Y ) are the set of basepoint preserving quasi-continuous maps from X to Y . Then we can obtain the product category T × T where obj(T × T ) are pairs (X, Y ) of quasi-topological spaces with basepoints, and morphisms are (T × T )((X, Y )(Z, W )) = {(f, g) f ∈ T (X, Z), g ∈ T (Y, W )}. Proposition 2.13. The smash product ∧ : T × T → T of quasi-topological spaces is a functor. The following definition of a quasi-orthogonal sequence is going to form part of the definition of a orthogonal quasi-spectrum. Definition 2.14. Let O and T be the categories defined above. Then we define the category of quasi orthogonal sequences formed as the functor category T O with objects obj(T O) = {functors X : O → T Xn := X(Rn)}, together with a left quasi-continuous basepoint preserving action of O(n) on each Xn for all n ≥ 0, and morphisms T O (X, Y ) = {ϕ : X → Y ϕ is a natural transformation}, and such that a natural transformation is formed of sets of quasi-continuous basepoint pre- serving maps ϕn : Xn → Yn that are O(n)-equivariant for n ≥ 0, or equivalently that the map ϕn commutes with the group action of O(n) on Xn and Yn. A useful example of such a functor category will be the unit sequence coming from the orthogonal sequence defined below. Consider a based topological space K, then define the orthogonal sequence with n-space: (GpK)n =(O(n)+ ∧ K, {∗}, if n = p otherwise 10 Then the unit sequence is when we just have the topological space S0, given by the sequence G0S0 = {S0, ∗, ∗, . . .}. We also consider quasi-biorthogonal sequences since they will help us in defining our smash product structure. The category of quasi-biorthogonal sequences is defined to be the category with objects obj(T O×O ) = {X : O × O → T X is a functor}, together with a quasi-continuous basepoint preserving left-action of O(m) × O(n), and T O×O(X, Y ) = {ψ : X → Y ψ is a natural transformation}, formed of sets of quasi-continuous basepoint preserving maps ψm,n : Xm,n → Ym,n, where Xm,n := X(Rm, Rn), that are O(m) × O(n)-equivariant for all m, n ≥ 0. Using this we can define the external smash product of two quasi-orthogonal sequence X and Y . Definition 2.15. Define the external smash product X∧Y to be the quasi-biorthogonal sequence given by the composition O × O X×Y−−−→ T × T ∧−→ T , defined by (X∧Y )m,n = (X∧Y )(Rm, Rn) = X(Rm) ∧ Y (Rn) = Xm ∧ Yn. Then by Proposition 2.10, the quasi-topological space Xn ∧ Ym has a quasi-O(n) × O(m)- action. For a general quasi-orthogonal sequence X we can define a quasi-biorthogonal sequence X ◦ ⊕ by: (X ◦ ⊕)m,n = (X ◦ ⊕)(Rm, Rn) = X(Rm+n) = Xm+n. Now we can construct the tensor product of quasi-orthogonal sequences since the category T is complete and cocomplete. Definition 2.16. For quasi orthogonal sequence X and Y we define the tensor product of X and Y to be the quasi-orthogonal sequence (X ⊗ Y )n = _p+q=n O(n)+ ∧O(p)×O(q) Xp ∧ Yq, where we define the O(n)-action on (X ⊗ Y )n by acting on each wedge summand. Then we can combine the external smash product and tensor product of quasi-orthogonal sequences as a natural bijection: Proposition 2.17. For quasi-orthogonal sequences X, Y and Z, there is a natural bijection T O×O (X∧Y, Z ◦ ⊕) ←→ T O(X ⊗ Y, Z). 11 Proof. Let f : X∧Y → Z◦⊕ be a natural transformation in the category of quasi-biorthogonal sequences. Then fp,q : Xp ∧ Yq → Z ◦ ⊕ is quasi O(p) × O(q)-equivariant and then by propo- sition 2.12, this corresponds to a quasi O(n)-equivariant map, with n = p + q f p,q : O(n)+ ∧O(p)×O(q) Xp ∧ Yq → Zn. Now fixing n and letting p and q vary, this allows us to obtain a quasi O(n)-equivariant map f n = _p+q=n O(n)+ ∧O(p)×O(q) Xp ∧ Yq → Zn, which is a quasi-continuous basepoint preserving O(n)-equivariant map in T O from X ⊗ Y to Z. Now we construct a map the other way. Let g ∈ T O (X ⊗ Y, Z). Then g is a wedge summand of basepoint preserving quasi-continuous O(n)-equivariant maps gn : _p+q=n O(n)+ ∧O(p)×O(q) Xp ∧ Yq → Zn, for all n ≥ 0. Also, we can write that gn =Wp+q=n gp,q, where gp,q : O(n)+ ∧O(p)×O(q) Xp ∧ Yq → Zn, and by proposition 2.12, we obtain a basepoint preserving quasi-continuous O(p) × O(q)- equivariant map as required. Let G0S0 be the unit quasi-orthogonal sequence (G0S0)n = (S0, ∗, ∗, . . .). For details of the following see Chapter 4 of [3]. Proposition 2.18. The category of quasi-orthogonal sequences forms a symmetric monoidal category (T O , ⊗, G0S0). Let S = (S0, S1, S2 . . .) be the quasi-orthogonal sequence defined in terms of quasi- topological spaces. For a proof of the subsequent result see Proposition 4.3.13 in [3]. Proposition 2.19. The orthogonal sequence of quasi-topological spaces S = (S0, S1, S2 . . .) is a commutative monoid in the symmetric monoidal category (T O, ⊗, G0S0). 3 E-theory orthogonal quasi-spectra This section brings together ideas from the previous section, since we will define the notion of an orthogonal quasi-spectrum which is a quasi-orthogonal sequence with added structure. We will show that we have an orthogonal quasi-spectrum representing the graded E-theory groups and thereafter show we have a smash product. 12 3.1 Quasi-Spectra We begin by defining concepts we have seen before in terms of quasi-topological spaces. A quasi-spectrum is a sequence of based quasi-topological spaces X0, X1, . . . with structure maps ǫ : Xm → ΩXm+1 that are quasi-continuous. An Ω-quasi-spectrum is a quasi-spectrum where for all natural numbers m the structure maps ǫ : Xm → ΩXm+1 are weak equivalences. Then we can define an orthogonal quasi-spectrum: Definition 3.1. An orthogonal quasi-spectrum is • a sequence of based quasi-topological spaces X0, X1, . . . • a basepoint preserving quasi-continuous left action of O(m) on each Xm for all m, and • a collection of based structure maps σ = σm : Xm ∧ S1 → Xm+1 that are quasi- continuous, such that for each m, n ≥ 0, the iterated map m : Xm ∧ Sn → Xm+1 ∧ Sn−1 → . . . → Xm+n, σn is quasi-continuous and O(m) × O(n)-equivariant. In the same manner, we have that a morphism of orthogonal quasi-spectrum f : X → Y is a collection of quasi-O(m)-equivariant maps fm : Xm → Ym for all m, which satisfy the following commutative diagram: Xm ∧ S1 f ∧idS1 / Ym ∧ S1 σm σm Xm+1 fm+1 / Ym+1, or alternatively that the following diagram commutes: Xm ǫm fm / Ym ǫm ΩXm+1 Ωfm+1 / ΩYm+1. It is easily seen that any orthogonal spectrum is an orthogonal quasi-spectrum. By Corol- lary 2.4, the structure maps in the definition of quasi-spectrum can be defined in terms of loop spaces. Notice that an orthogonal quasi-spectrum is a quasi-orthogonal sequence with more structure. Proposition 3.2. The category of right S-modules, mod-S is naturally equivalent to the category of orthogonal quasi-spectrum. 13 / / / / Proof. Consider the multiplication map ν : X ⊗ S → S for a right S-module X. Then by Proposition 2.17 we have a set of O(m) × O(n)-equivariant maps νn m : Xm ∧ Sn → Xm+n, for m, n ≥ 0 with unit quasi-homeomorphism ν0 that the structure maps are then defined by νm. m. Now this action is associative so it follows Conversely, consider the set of structure maps σn p : Xp ∧ Sp → Xn+p, for a spectrum X and p, n ≥ 0, with unit quasi-homeomorphism σ0 p. Then we have a multiplicative map ν : X ⊗ S → X defining a right S-module. Since these constructions are inverses, we have a natural equivalence of these two categories. Hence we can obtain a tensor product of orthogonal quasi-spectrum since we have a tensor product in the category of right S-modules. Definition 3.3. Let X be an orthogonal quasi-spectrum with spaces Xn. For each integer k ∈ Z we define the k-th stable homotopy group πk(X) to be the direct limit πk(X) = lim−→ n πk+nXn, under the maps ǫ∗ : πk+nXn → πk+n+1Xn+1 induced from the structure maps ǫ : Ωk+nXn → Ωk+n+1Xn+1. 3.2 Graded E-theory Spectra Let Asyg(A, B) denote the set of graded asymptotic morphisms from A to B with the quasi- topology as defined in Definition 2.5. Proposition 3.4. The map of quasi-topological spaces defined by is quasi-continuous for all xt ∈ Asyg(A, B). f : Asyg(A, B) → Asyg(Db⊗A, Db⊗B) f (xt) = idDb⊗xt, Proof. Since gradings follow immediately, we consider ungraded asymptotic morphisms through- out the proof. To prove a map of quasi-topological spaces is quasi-continuous, wwe need to check that for a quasi-continuous map g : Y → Asy(A, B) where Y is a compact Haus- g : Y → Asy(A, B) where Y is a compact Hausdorff space is quasi-continuous. Then by defi- dorff space, that the composition f g : Y → Asy(Db⊗A, Db⊗B) is quasi-continuous. Suppose nition 2.5 we know that g is quasi-continuous when for each t ∈ [1, ∞) the mapegt(a) : Y → B defined by egt(a)(y) = g(y)t(a) 14 is continuous. Then we define f g : Y → Asy(Db⊗A, Db⊗B) by for each t ∈ [1, ∞) f (g(y)t)(a) = idDb⊗g(y)t(a) = idDb⊗egt(a)(y) but since g is quasi-continuous and that it follows from the definition of quasi-topology on the set of asymptotic morphisms that f g is quasi-continuous. f (g(y)t)(a) =ff gt(a)(y), Definition 3.5. Let K = K(H). Define X(A, B) to be the sequence of based quasi- topological spaces where m ≥ 0. Define maps ǫm : Xm → ΩXm+1: Xm = Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) by: / ΩAsyg(Sb⊗Ab⊗K, Bb⊗Fm+1,0b⊗K) Asyg(Sb⊗Ab⊗K, Σ(Bb⊗Fm+1,0)b⊗K) ∼= ǫ(xt) = (bb⊗idB b⊗Fm,0 b⊗K) ◦ (idSb⊗xt) ◦ (∆b⊗idA b⊗K), rem 1.6. Alternatively, we also have a map σm : Xm ∧ S1 → Xm+1 defined by for all xt ∈ Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) and the Bott map b ∈ Homg(S, ΣF1,0) from Theo- with xt ∈ Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) and s ∈ S1. Definition 3.6. We define a quasi-continuous action of the group O(m) on the space Xm as follows. First we consider the alternative definition of the Clifford algebra Fm,0 as Cliff(V ). Recall that for V an m-dimensional Euclidean vector space, Cliff(V ) = G(V )/ ∼ where G(V ) is the algebra generated by V subject to the equivalence relation ∼ defined by σm(xt, s) = ǫm(xt)(s), v2 = v2 · 1 for all v ∈ V . We write ab for the product of two elements a, b ∈ Cliff(V ). If V = Rm, then we have a natural group action (H, v) 7→ Hv where H ∈ O(m), v ∈ V . Then we can define a group action of O(m) on G(V ) by H(v1 . . . vk) 7→ H(v1) . . . H(vk) and H(1) = 1 for all H ∈ O(m). Then this gives a group action of O(m) on Cliff(V ) since H(v2) = H(v)H(v) = (H(v))2 = H(v)2 · 1 = v · 1 since H is orthogonal. 15 / So then we get a group action λ : O(m) × Fm,0 → Fm,0, by λ(H, (e1, e2, . . . em)) = H(e1)H(e2) . . . H(em), where H ∈ O(m), e1, e2, . . . em are the generators of the algebra Fm,0. Then we define by λ∗ : O(m) × Bb⊗Fm,0b⊗K(H) → Bb⊗Fm,0b⊗K(H) λ∗(H, bb⊗xb⊗p) = bb⊗λ(H, x)b⊗p with H ∈ O(m), b ∈ B, x ∈ Fm,0 and p ∈ K(H). Then we finally define a group action of O(m) on Xm by λ∗∗ : O(m) × Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) −→ Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)), λ∗∗(H, αt)(x) = λ∗(H, αt(x)), where we have αt ∈ Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)), x ∈ Sb⊗Ab⊗K(H) and H ∈ O(m). Then it follows that this action is O(m)-equivariant. The following is true by Proposition 3.4. Proposition 3.7. The action in the previous definition is a basepoint preserving quasi- continuous action of O(m) on Xm. Proposition 3.8. The map ǫm : Xm → ΩXm+1 is quasi-continuous and hence the map σm : Xm ∧ S1 → Xm+1 is quasi-continuous. Proof. Since both b⊗K(H) bb⊗idB b⊗Fm,0 and ∆b⊗idA b⊗K(H) are ∗-homomorphisms, these two maps are continuous. So it suffices to check the map maps yields a quasi-continuous map. By proposition 3.4, with D = S it follows that the (idSb⊗xt) is quasi-continuous since the a composition of continuous and quasi-continuous map (idSb⊗xt) is quasi-continuous. Hence the map ǫm is quasi-continuous. Since σm is defined in terms of ǫm it is also quasi-continuous. We define the iterated map σn m : Xm ∧ Sn → Xm+n by σn m(xt, s1, s2, . . . sn) = ǫn(xt)(s1)(s2) . . . (sn), . . . ∧ S1. where xt ∈ Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) and s1, s2, . . . sn is contained in S1 ∧ S1 ∧ 16 Proposition 3.9. The iterated map σn and O(m) × O(n)-equivariant. m : Xm ∧ Sn → Xm+n defined above quasi-continuous Proof. By other results it suffices to check that the map is O(m) × O(n)-equivariant. Firstly it is clear that Xm ∧ Sn and Xm+n are quasi O(m) × O(n)-spaces. Let i : O(m) × O(n) → O(m + n) be the inclusion map. defined earlier. J(xt) = (idBb⊗Jb⊗idK(H)) ◦ xt Now O(m + n) acts on Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm+n,0b⊗K(H)) by for all J ∈ O(m + n) and xt : Sb⊗Ab⊗K(H) → Bb⊗Fm+n,0b⊗K(H). Here J acts of Fm+n,0 as Then O(m) × O(n) acts on Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) ∧ Sn by, for all H ∈ O(m), K ∈ O(n), xt ∈ Asyg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) and s ∈ Sn. Then (H, K)(xt, s) = ((idBb⊗Hb⊗idK(H)) ◦ xt, Ks), we need to show that for σ = σn m : Xm ∧ Sn → Xm+n, that σ((H, K)(xt, s)) = i(H, K)σ(xt, s), that is, That is to show, by definition of σ that, σ((Hxt, Ks) = i(H, K)σ(xt, s). ǫ(Hxt)(Ks) = i(H, K)ǫ(xt)(s). Then Let bn = bb⊗ . . .b⊗b be the n-fold graded tensor product of the Bott map, b : S → ΣF1,0. we have bb⊗ . . .b⊗b(λ)(s) ∈ Fn,0 for λ ∈ S n and s ∈ Sn. Then for K ∈ O(n), bn = bb⊗ . . .b⊗b : S n → Σnb⊗Fn,0, (bb⊗ . . .b⊗b)(λ)(Ks) = K(bb⊗ . . .b⊗b)(λ)(s). Then by permuting copies of Σ and extending by linearity we have an action of the orthogonal group. Reconsidering the left hand side yields ǫ(Hxt)(Ks) = i(H, K)ǫ(xt)(s), and the right hand side yields ǫ(Hxt)(Ks) = ((bnb⊗idB b⊗Fm,0 i(H, K)ǫ(xt)(s) = i(H, K)(bb⊗idB b⊗Fm,0 b⊗K(H)) ◦ (idSb⊗Hxt) ◦ (∆b⊗idA b⊗K(H))(Ks)), b⊗K(H)) ◦ (idSb⊗xt) ◦ (∆b⊗idA b⊗K(H)). Then 17 ǫ(Hxt)(Ks) = ((bnb⊗idB b⊗Fm,0 = i(H, 1)((bnb⊗idB b⊗Fm,0 = i(H, 1)i(1, K)(bnb⊗idB b⊗Fm,0 = i(H, K)(bnb⊗idB b⊗Fm,0 b⊗K(H)) ◦ (idSb⊗Hxt) ◦ (∆b⊗idA b⊗K(H)))(Ks) b⊗K(H)) ◦ (idSb⊗xt) ◦ (∆b⊗idA b⊗K(H))(Ks)) b⊗K(H)) ◦ (idSb⊗xt) ◦ (∆b⊗idA b⊗K(H))(s) b⊗K(H)) ◦ (idSb⊗xt) ◦ (∆b⊗idA b⊗K(H))(s) = i(H, K)ǫ(xt)(s) Then the result follows. The proof of the following result follows from the above propositions, namely Proposi- tion 3.7, Proposition 3.8 and Proposition 3.9. Proposition 3.10. The spectrum X(A, B) is an orthogonal quasi-spectrum. Proposition 3.11. If G0, G1, G2, . . . is a sequence of groups with isomorphisms θn : Gn → Gn+1 for n ≥ 0, then Proof. We first need to construct a commutative diagram. Gn = G0. lim−→ n Gn θn . G0 δ "❊❊❊❊❊❊❊❊ <②②②②②②②② ψ Gn+1 As θn is an isomorphism for all n ≥ 0, we have inverses, so δ = (θ0)−1 . . . (θn−1)−1 and ψ = (θ0)−1 . . . (θn−1)−1(θn)−1 and hence the diagram commutes. Now we check that G0 is unique. Suppose we have a group H such that we have a group homomorphism f : G0 → H which fits into the following diagram Gn θn µ1 . G0 f / H µ2 δ #❋❋❋❋❋❋❋❋❋ ;①①①①①①①① ψ Gn+1 Then define f = µ1δ−1 so our diagram commutes. Suppose that we have another group homomorphism g : G0 → H fitting into the diagram. Then by commutativity we have gδ = µ1, so g = µ1δ−1 = f so f is unique. 18 " < # $ $ / ; : : Proposition 3.12. The direct limit lim−→n Proof. This result follows from Proposition 3.11 where Eg(A, Σk+nBb⊗Fn,0) is Eg(A, ΣkB). and using Proposition 1.6. Gn = Eg(A, Σk+nBb⊗Fn,0) Proposition 3.13. For all positive integers k, πkX(A, B) = Eg(A, ΣkB). Proof. Since X is an orthogonal quasi-spectrum we have that πkX(A, B) = lim−→ n πk+nEn. Then lim−→ n πk+nXn = lim−→ n n = lim−→ π0Ωk+nAsyg(Sb⊗Ab⊗K(H), Bb⊗Fn,0b⊗K(H)) π0Asyg(Sb⊗Ab⊗K(H), Σk+nBb⊗Fn,0b⊗K(H)) JSb⊗Ab⊗K(H), Σk+nBb⊗Fn,0b⊗K(H)K Eg(A, Σk+nBb⊗Fn,0) = lim−→ = Eg(A, ΣkB) by Proposition 3.12. = lim−→ n n Proposition 3.14. The orthogonal quasi-spectrum X(A, B) is an Ω-quasi-spectrum. Proof. We just need to check that the structure map ǫ : En → ΩEn+1 is a weak equivalence. That is the map πkEn → πkΩEn+1 is an isomorphism for all k. Now this gives us the map: which is an isomorphism for all k by Theorem 1.6. Eg(A, Σk(Bb⊗Fn,0)) → Eg(A, Σk+1(Bb⊗Fn+1,0)), Theorem 3.15. Let A, B and C be graded C ∗-algebras. Then there is a natural map of orthogonal quasi-spectra µm,n : X(A, B) ∧ X(B, C) → X(A, C), defined by (α ∧ β)t 7→ (βr(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K), where α ∈ Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) and β ∈ Asyg(Sb⊗Bb⊗K, Cb⊗Fn,0b⊗K). In addition the product is associative up to homotopy. 19 Proof. The product gives a natural O(m) × O(n)-equivariant map: Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) ∧ Asyg(Sb⊗Bb⊗K, Cb⊗Fn,0b⊗K) −→ Asyg(Sb⊗Ab⊗K, Cb⊗Fm+n,0b⊗K), given by permuting the m, n and m + n copies of F1,0. Now compatibility with the structure maps follows from the naturality of the structure maps and also since we have the following two diagrams: and Xm(A, B) ∧ Xn(B, C) µm,n / Xm+n(A, C) ǫ∧id ǫ ΩXm+1(A, B) ∧ Xn(B, C) µm+1,n / ΩXm+n+1(A, C), Xm(A, B) ∧ Xn(B, C) µm,n / Xm+n(A, C) id∧ǫ ǫ Xm(A, B) ∧ ΩXn+1(B, C) µm,n+1 / ΩXm+n+1(A, C), where id denote the obvious identities, and the ǫ's denote the required structure maps. These diagrams commute since, µm+1,n(ǫ ∧ id)(α ∧ β)t = µm+1,n(ǫ(α) ∧ β)t b⊗K) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K)i) ◦ (∆b⊗idA b⊗K) b⊗K) ◦ (idS b⊗Sb⊗αt) ◦ (∆b⊗idS b⊗A b⊗K) ◦ (∆b⊗idA b⊗K) = (βr(t)b⊗idΣFm+1,0) ◦ (idSb⊗ǫ(αt)) ◦ (∆b⊗idA b⊗K) = (βr(t)b⊗idΣFm+1,0) ◦ (idSb⊗h(bb⊗idB b⊗Fm,0 = (idΣF1,0b⊗βr(t)b⊗idFm,0) ◦ (bb⊗idS b⊗B b⊗Fm,0 b⊗K) ◦ (idSb⊗βr(t)b⊗idFm,0) ◦ (idS b⊗Sb⊗αt) = (bb⊗idC b⊗Fm+n,0 ◦ (∆b⊗idS b⊗A b⊗K) ◦ (∆b⊗idA b⊗K) by Lemma 1.3, = (bb⊗idC b⊗Fm+n,0 b⊗K) ◦ (idSb⊗ (cid:2)(βr(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K)(cid:3)) ◦ (∆b⊗idA b⊗K) = ǫ((βr(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K)) = ǫ(µm,n)(α ∧ β)t, 20 / / / / and µm,n+1(id ∧g ǫ)(α ∧g β)t = µm,n+1(α ∧ ǫ(β))t = (ǫ(β)r(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K) = h(bb⊗idC b⊗Fn,0 = (bb⊗idC b⊗Fm+n,0 = (bb⊗idC b⊗Fm+n,0 = (bb⊗idC b⊗Fm+n,0 = ǫ((βr(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K)) b⊗K) ◦ (idSb⊗βr(t) ◦ (∆b⊗idB b⊗K)ib⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K) b⊗K) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K) b⊗K) ◦ (idSb⊗βr(t)b⊗idFm,0) ◦ (∆b⊗idB b⊗Fm,0 b⊗K) ◦ (idSb⊗βr(t)b⊗idFm,0) ◦ (idS b⊗Sb⊗αt) ◦ (∆b⊗idS b⊗A b⊗K) ◦ (∆b⊗idA b⊗K) by Lemma 1.3, b⊗K) ◦ (idSb⊗(cid:2)(βr(t)b⊗idFm,0) ◦ (idSb⊗αt) ◦ (∆b⊗idA b⊗K)(cid:3)) ◦ (∆b⊗idA b⊗K) = ǫ(µm,n)(α ∧ β)t, for all α ∧ β ∈ Em(A, B) ∧ En(B, C). Now we check that our product is associative up to homotopy. Let α ∈ Asyg(Sb⊗Ab⊗K, Bb⊗Fm,0b⊗K) and β ∈ Asyg(Sb⊗Bb⊗K, Cb⊗Fn,0b⊗K) and γ ∈ Asyg(Sb⊗Cb⊗K, Db⊗Fp,0b⊗K). Then take the homotopy classes of these elements and we ob- tain E-theory groups, and we know that the E-theory product is associative. 4 Connecting graded K and E-theory spectra This section connects together graded K-theory and E-theory spectra. In particular we form a smash product in terms of these two spectra and consequently combine K-theory and K- homology in a smash product. Here we obtain a connection between K-theory, K-homology and E-theory spectra. Let Hom(A, B) denote the set of graded ∗-homomorphisms from A to B. Then recall that and the K-homology is given by Kn(A) = [S, ΣnAb⊗K(H)] ∼= En g (F, A), Khom(A) = En(A, F). 4.1 A topology on graded ∗-homomorphisms Let Homg(A, B) denote the set of graded ∗-homomorphisms from A to B. We equip Homg(A, B) with the compact open topology as detailed below. Definition 4.1. A basis for a topology on a set A is a collection of subsets A of A such that A is a union of sets from A and such that if A1, A2 are in A then their intersection is a union of sets from A. A subbasis is a collection of subsets B of A where the set A of all finite intersections of sets in B is a basis. 21 Definition 4.2. Let A and B be graded C ∗-algebras. The compact open topology on the set of graded ∗-homomorphisms from A to B, Homg(A, B), is generated by subsets of the following form, B(K, U) = {f ∈ Homg(A, B) f (K) ⊂ U}, where K is compact in A and U is open in B. Here generated means that the sets defined form a subbasis for the open sets when we think of a topology. This then generates a basis for the topology. For simplicity of notation let Homg(A, B) denote the space of graded ∗-homomorphisms from A to B equipped with the compact open topology. Denote the loop space of this space by ΩHomg(A, B). Note that the basepoints for both of these spaces is just the zero ∗-homomorphism, which will we denote by 0. The compact open topology is the choice for our topology since it gives us the correct path components for our loop space and it also allows us to have continuity of particular maps as we will see soon. Now let us consider the generators of the compact open topology on the spaces ΩHomg(A, B) and Homg(A, ΣB). Now we have a basis for Homg(A, B), so we just extend this for ΩHomg(A, B), and it is not to hard to see that a basis for the loop space is the set generated by B(K ′, V ) such that K ′ ⊆ [0, 1] compact and V ⊆ Homg(A, B) open. Combining these, we obtain the following definition. Definition 4.3. The compact open topology on ΩHomg(A, B) is generated by sets of the form B(K ′, B(K, U)), where K ⊆ A compact, K ′ ⊆ [0, 1] compact and U ⊆ B open. The compact open topology on Homg(A, ΣB) is generated by sets of the form B(K, B(K ′, U)) where K ⊆ A compact, K ′ ⊆ [0, 1] compact and U ⊆ B open. Before we check we have the continuity of maps in the following proof, it is worth noting that it is sufficient to check that a map is continuous under a topology by considering a subbasis. That is, to check a map of topological spaces is continuous we just need to check that continuity holds at the level of generating sets for a basis of a topology. For details, see [13], Application 3.2.5. Proposition 4.4. The spaces ΩHomg(A, B) and Homg(A, ΣB) are homeomorphic. Proof. We consider ungraded ∗-homomorphisms since the grading property is immediate. Define f : ΩHom(A, B) → Hom(A, ΣB) as follows. Let µ ∈ ΩHom(A, B) based at 0, then define f (µ)(a)(s) = µ(s)(a), for all a ∈ A and s ∈ [0, 1]. Now define g : Hom(A, ΣB) → ΩHom(A, B) as follows. Let τ ∈ Hom(A, ΣB), define g(τ )(s)(a) = τ (a)(s). Both f and g are well defined since µ and τ are ∗-homomorphisms. We need to show that f ◦ g = id, and g ◦ f = id where id stands for the natural identities. 22 Let ϕ ∈ Hom(A, ΣB), then for all a ∈ A, s ∈ [0, 1]. f g(ϕ)(a)(s) = g(ϕ)(s)(a) = ϕ(a)(s). Similarly, let ψ ∈ ΩHom(A, B), then for all a ∈ A, s ∈ [0, 1], gf (ψ)(s)(a) = f (ψ)(a)(s) = ψ(s)(a). Then f ◦ g = id and g ◦ f = id as required. Now we check that f and g are continuous. By the above discussion, it suffices to check that: and f −1[B(K ′, B(K, U))] = B(K, B(K ′, U)) g−1[B(K, B(K ′, U))] = B(K ′, B(K, U)), for all K ⊆ A compact, K ′ ⊆ [0, 1] compact and U ⊆ B open. Let y ∈ B(K, B(K ′, U)), then f −1(y) = {x f (x) = y}. Now let x ∈ f −1(y), then we know f (x)(s)(a) = x(a)(s) = y(s)(a), so x must be contained in B(K ′, B(K, U)), and similarly we can check the converse, so f is continuous. Similarly we can prove that g is continuous. 4.2 K-theory spectra Now we can define the K-theory spectrum. Definition 4.5. Let K = K(H). Define K(A) to be the sequence of based topological spaces where m ≥ 0. Define maps ǫn : Kn → ΩKn+1: Kn = Homg(Sb⊗Fb⊗K, Ab⊗Fn,0b⊗K) Homg(Sb⊗Fb⊗K, Ab⊗Fn,0b⊗K) ∼= / ΩHomg(Sb⊗Fb⊗K, Ab⊗Fn+1,0b⊗K) Homg(Sb⊗Ab⊗K, Σ(Ab⊗Fn+1,0)b⊗K) for all xt ∈ Asyg(Sb⊗Ab⊗K, Bb⊗Fn,0b⊗K) and the Bott map b ∈ Homg(S, ΣF1,0). b⊗K) ◦ (idSb⊗xt) ◦ (∆b⊗idF b⊗K), ǫ(xt) = (bb⊗idA b⊗Fn,0 by: We now give an alternative definition for the spectrum of graded K-theory in terms of asymptotic morphisms. 23 / Definition 4.6. Let K = K(H). Define K′(A) to be the orthogonal quasi-spectrum with the sequence of based quasi-topological spaces where n ≥ 0. The structure maps ǫ : K ′ K ′ n = Asyg(Sb⊗Fb⊗K, Ab⊗Fn,0b⊗K) n → ΩK ′ n+1: Asyg(Sb⊗Fb⊗K, Ab⊗Fn,0b⊗K) / ΩAsyg(Sb⊗Fb⊗K, Ab⊗Fn+1,0b⊗K) Asyg(Sb⊗Fb⊗K, Σ(Ab⊗Fn+1,0)b⊗K) for all xt ∈ Asyg(Sb⊗Ab⊗K, Bb⊗Fn,0b⊗K) and the Bott map b ∈ Homg(S, ΣF1,0). b⊗K) ◦ (idSb⊗xt) ◦ (∆b⊗idF b⊗K), ǫ(xt) = (bb⊗idA b⊗Fn,0 are defined by: ∼= We now notice that Definition 4.5 and Definition 4.6 are orthogonal and orthogonal quasi- spectra for the same reason that 3.5 forms one and consequently the following result comes from the stable homotopy groups coming from these spectra. Proposition 4.7. The map of spectrum f : K(A) → K′(A), defined by f (ϕ) = ϕ for all ϕ ∈ K(A) is a weak equivalence. Proof. Consider the map f ′ : Homg(S, A) → Asyg(S, A) then this induces the map f ′ JS, AK. Now the map, ∗ : [S, A] → induces an isomorphism at the level of π0, and therefore the map Homg(S, Ab⊗Fn+1,0) → Asyg(S, Ab⊗Fn+1,0), Homg(S, Ab⊗F1,0) → Asyg(S, Ab⊗F1,0), induces an isomorphism at the level of πn. Therefore we have a weak equivalence. Then we can also consider the map f above and the same applies, since we obtain this map by tensoring with the suspension and the complex numbers. Corollary 4.8. The map of spectrum f : K(A) → K′(A) has a natural inverse g : K′ → K at the level of stable homotopy groups. Theorem 4.9. Let A and B be C ∗-algebras. Then there is a natural map of orthogonal quasi-spectra ν ′ m,n : K(A) ∧ E(A, B) → K′(B), defined by (α ∧ βt)t 7→ (βtb⊗idFm,0) ◦ (idSb⊗α) ◦ (∆b⊗idF b⊗K(H)), where α ∈ Homg(Sb⊗Ab⊗K(H), Bb⊗Fm,0b⊗K(H)) and β ∈ Asyg(Sb⊗Bb⊗K(H), Cb⊗Fn,0b⊗K(H)). 24 / Proof. Since the composition of a ∗-homomorphism and an asymptotic morphism is an asymptotic morphism it is clear that α ∧ β is an asymptotic morphism and lies in the required spectra. By the above theorem and Corollary 4.8, we obtain Corollary 4.10. There is a natural map of spectra νm,n : K(A) ∧ E(A, B) → K(B), with the above criteria. Now we finalise this section by combining the graded K-theory spectrum and K-homology spectrum noting that Khom(A) = E(A, F). Theorem 4.11. There is a canonical map of orthogonal quasi-spectra. The map S is natural in the variable B in the obvious sense and natural in the variable A, in the sense that if we have a ∗-homomorphism f : A → A′ then we have the following commutative diagram S : K(Ab⊗B) ∧ Khom(A) → K′(B) S / K′(B) K′(B) f∗∧id id∧f ∗ K(Ab⊗B) ∧ Khom(A) K(Ab⊗B) ∧ Khom(A′) K(A′b⊗B) ∧ Khom(A′) f∗(α)(λ) = (fb⊗idB b⊗Fm,0 f ∗(βt)(a) = βt(fb⊗idS b⊗K(H))(a), S / / K′(B) b⊗K(H))(α(λ)), where f∗ and f ∗ are defined by: and Proof. Writing Khom(A) = E(A, F), we can extend the definition of S to a composition of maps, in order to obtain the following diagram: with α ∈ K(Ab⊗B), β ∈ Khom(A′), a ∈ Sb⊗Ab⊗K and λ ∈ Sb⊗Fb⊗K. / K(Ab⊗B) ∧ E(Ab⊗B, B) id∧ b⊗B / ν ′ m,n / K′(B) K′(B) ν ′ m,n / K′(B) id∧f ∗ K(Ab⊗B) ∧ E(A, F) K(Ab⊗B) ∧ E(A′, F) K(A′b⊗B) ∧ E(A′, F) f∗∧id id∧ b⊗B / / K(A′b⊗B) ∧ E(A′ ⊗ B, B) 25 / O O / O O / Then we have K(Ab⊗B) ∧ E(Ab⊗B, B) → K′(B). ν ′ = ν ′ m,n(id ∧b⊗B)(id ∧ f ∗)(α ∧ βt) = ν ′ m,n(α ∧ f ∗(βt)b⊗idB) m,n(id ∧b⊗B)(α ∧ f ∗(βt)) = (f ∗(βt)b⊗idB b⊗Fm,0) ◦ (idSb⊗α) ◦ (∆b⊗idF b⊗K(H)) = ([βt ◦ (fb⊗idS b⊗K(H))]b⊗idB b⊗Fm,0) ◦ (idSb⊗α) ◦ (∆b⊗idF b⊗K(H)) b⊗K(H)) ◦ (idSb⊗α) ◦ (∆b⊗idF b⊗K(H)) = (βtb⊗idB b⊗Fm,0) ◦ (fb⊗idS b⊗B b⊗Fm,0 b⊗K(H)) ◦ α] ◦ (∆b⊗idF b⊗K(H)) = (βtb⊗idB b⊗Fm,0) ◦ (idSb⊗[(fb⊗idB b⊗Fm,0 = (βtb⊗idB b⊗Fm,0) ◦ (idSb⊗f∗(α)) ◦ (∆b⊗idF b⊗K(H)) m,n(f∗(α) ∧ (βtb⊗idB)) m,n(id ∧b⊗B)(f∗ ∧ id)(α ∧ βt). m,n(id ∧b⊗B)(f∗(α) ∧ βt) = v′ = v′ = v′ Corollary 4.12. There is a canonical map S : K(A ⊗ B) ∧ Khom(A) → K(B) of orthogonal quasi-spectra. The map S is natural in the variable B in the obvious sense and natural in the variable A, in the sense that if we have a ∗-homomorphism f : A → A′ then we have the following commutative diagram K(A ⊗ B) ∧ Khom(A) S / K(B) id∧f ∗ K(A ⊗ B) ∧ Khom(A′) K(B) f∗∧id K(A′ ⊗ B) ∧ Khom(A′) S / / K(B). References [1] G. Bredon, Topology and Geometry, Springer GTM 139, 1993. [2] S. L. Browne, A Bott periodicity proof for real graded C ∗-algebras, arXiv:1611.09887. [3] S. L. Browne, E-theory spectra, PhD thesis, University of Sheffield, available at http://etheses.whiterose.ac.uk/id/eprint/17812. 26 / O O [4] A. Connes and N. Higson, D´eformations, morphismes asymptotiques et K-th´eorie bi- variante, C. R. Acad. Sci. Paris S´er. I Math., 311, 1990, 101-106. [5] M. Dadarlat and R. Meyer, E-theory for C ∗-algebras over topological spaces, J. Funct. Anal., 263(1):216-247, 2012. [6] E. Guentner and N. Higson, Group C ∗-algebras and K-theory, Noncommutative geom- etry, Lecture Notes in Math., 1831, 137–251, Springer, Berlin, 2004. [7] E. Guentner, N. Higson and J. Trout, Equivariant E-Theory for C ∗-algebras, Memoirs of the American Mathematical Society, 2000. [8] N. Higson, Categories of fractions and excision in KK-theory, J. Pure Appl. Algebra, 65, 1990, 119-138. [9] N. Higson and G. Kasparov, Operator K-theory for groups which act properly and iso- metrically on Hilbert space, Electronic Research Announcements of the American Math- ematical Society, 3 (1997), pages 131-142. [10] M. Hovey, B. Shipley and J. Smith, Symmetric spectra, J. Amer. Math. Soc., volume 13, pages 149–208, 2000. [11] P. D. Mitchener, Symmetric K-theory spectra of C ∗-categories, K-theory, 24 (2001), pages 157-201. [12] E. Spanier, Quasi-topologies, Duke Math. J., 20:1–14, 1963. [13] W. A. Sutherland, Introduction to Metric and Topological Spaces, Oxford Science Pub- lications, 1975. [14] M. Weiss and B. Williams. Assembly. In Novikov conjectures, index theorems and rigid- ity, Vol. 2 (Oberwolfach, 1993), volume 227 of London Math. Soc. Lecture Note Ser., pages 332-352. Cambridge Univ. Press, Cambridge, 1995. 27
1811.01922	3	1811	2018-11-28T18:58:53	Quantum extensions of ordinary maps	[ "math.OA", "math-ph", "math.GN", "math-ph" ]	We define a loop to be quantum nullhomotopic if and only if it admits a nonempty quantum set of extensions to the unit disk. We show that the canonical loop in the unit circle is not quantum nullhomotopic, but that every loop in the real projective plane is quantum nullhomotopic. Furthermore, we apply Kuiper's theorem to show that the canonical loop admits a continuous family of extensions to the unit disk that is indexed by an infinite quantum space. We obtain these results using a purely topological condition that we show to be equivalent to the existence of a quantum family of extensions of a given map.	math.OA	math	QUANTUM EXTENSIONS OF ORDINARY MAPS ANDRE KORNELL Department of Mathematics University of California, Davis Abstract. We define a loop to be quantum nullhomotopic if and only if it admits a nonempty quantum set of extensions to the unit disk. We show that the canonical loop in the unit circle is not quantum nullhomotopic, but that every loop in the real projective plane is quantum nullhomotopic. Furthermore, we apply Kuiper's theorem to show that the canonical loop admits a continuous family of extensions to the unit disk that is indexed by an infinite quantum space. We obtain these results using a purely topological condition that we show to be equivalent to the existence of a quantum family of extensions of a given map. Quantum families of extensions. Noncommutative mathematics draws an analogy be- tween unital C-algebras and compact Hausdorff spaces. Commutative unital C-algebras correspond bijectively to compact Hausdorff spaces, up to isomorphism of the former and up to homeomorphism of the latter, so noncommutative unital C-algebras are viewed as "quantum compact Hausdorff spaces". Specifically, each compact Hausdorff space X corre- sponds to the commutative C-algebra C(X) of continuous complex-valued functions on X, and conversely, each noncommutative unital C-algebra is imagined to have the same form, but for some fictitious space X, whose local topology is quantum in the tautological sense that some real-valued functions on this space fail to commute, in the manner of incompatible observables. This correspondence between compact Hausdorff spaces and commutative unital C- algebras extends to a correspondence between continuous maps and unital ∗-homomorphisms. Each map f from a compact Hausdorff space X to a compact Hausdorff space Y induces a unital ∗-homomorphism from C(Y ) to C(X), by precompostion. We thus have a con- travariant equivalence of categories. Each unital ∗-homomorphism between noncommuta- tive unital C-algebras is thus viewed as a continuous function between the corresponding quantum spaces, but in the opposite direction. This "noncommutative metaphor" can be extended further and further [6]. For the purposes of the this paper, we only recall that C(X ⊔ Y ) ∼= C(X) ⊕ C(Y ) and C(X × Y ) ∼= C(X) ⊗ C(Y ), for compact Hausdorff spaces X and Y , so we take the direct sum and the minimal tensor product of C-algebras to be the appropriate generalizations of disjoint union and Cartesian product of compact Hausdorff spaces. The minimal tensor product is preferred over other notions because it distributes over the direct sum. Let S and T be compact Hausdorff spaces, and f an ordinary map from S to T . If S is a subspace of a larger compact Hausdorff space S, then an extension of f to S is of course a E-mail address: [email protected]. 1 map f : S → T making the following diagram commute: S j S f f T We write j : S ֒→ S for the inclusion map. A family of extensions indexed by a compact Hausdorff space Y is instead a continuous function f of two variables making making the following diagram commute: S × Y j×id f S × Y f ×id T × Y T proj1 The existence of a nonempty family of extensions indexed by an ordinary compact Haus- dorff space is of course equivalent to the existence of a single extension. We will show that there is no such equivalence when we consider nonempty families of extensions indexed by a quantum compact Hausdorff space. To make this claim precise, we first apply the functor C to the diagram above, to obtain a diagram in the category of unital C-algebras and unital ∗-homomorphisms: C( S) ⊗ C(Y ) j ⋆⊗id C(S) ⊗ C(Y ) f ⋆⊗id C(T ) ⊗ C(Y ) C(T ) proj⋆ 1 Allowing the index space Y to be quantum, we replace C(Y ), with an arbitrary unital C- algebra B. We say that the quantum space is nonempty just in case the C-algebra B is nonzero. Thus, we ask whether there exists a nonzero unital C-algebra B, and a unital normal ∗-homomorphism φ : C(T ) → C( S) ⊗ B making the following diagram commute: C( S) ⊗ B j ⋆⊗id φ C(S) ⊗ B f ⋆⊗id C(T ) ⊗ B C(T ) a⊗1←[a The core result of the paper provides a purely topological condition equivalent to the existence of a family of extensions indexed by a quantum compact Hausdorff space. Theorem 1.6. Let B be a unital C-algebra. Let ı : T → Hom(C(T ), B) be the map taking each point of T to evaluation at that point. There exists a unital ∗-homomorphism C(T ) → C( S) ⊗ B making the above diagram commute if and only if the ordinary map ı ◦ f extends to S: S f j S T f ı 2 Hom(C(T ), B) Thus, the functor Hom(C( − ), B) is analogous to the quantum monad of Abramsky, Barbosa, de Silva, and Zapata [1]. Quantum nullhomotopic loops. Recall that a loop f : S1 → T is nullhomotopic if and only if f extends to a map on the unit disk D2. As an application of the above theorem, we investigate the corresponding quantum notion. Setting S = S1 and S = D2, we ask whether a map f : S1 → T admits a family of extensions indexed by a nonempty quantum compact Hausdorff space. This notion of quantum nullhomotopy degenerates if we allow arbitrary quantum compact Hausdorff spaces as index spaces. As a consequence of Kuiper's theorem [8] we show that the canonical loop S1 → S1 admits a quantum family of extensions to the unit disk: Theorem 2.6. Let H be an infinite-dimensional Hilbert space, and let L(H) be the unital C-algebra of bounded operators on H. There is a unital ∗-homomorphism φ such that the following diagram commutes: C(D2) ⊗ L(H) j ⋆⊗id φ C(S1) ⊗ L(H) a⊗1←[a C(S1) It follows that every loop in every compact Hausdorff space admits a quantum family of extensions to the unit disk indexed by the quantum compact Hausdorff space corresponding to L(H). Therefore, we make the following definition: Definition 2.1. A loop f : S1 → T is quantum nullhomotopic just in case there is a nonzero finite-dimensional Hilbert space H and unital ∗-homomorphism φ such that the following diagram commutes. C(D2) ⊗ L(H) j ⋆⊗id φ C(S1) ⊗ L(H) f ⋆⊗id C(T ) ⊗ L(H) C(T ) a⊗1←[a Equivalently, we may replace the arbitrary nonzero matrix algebra L(H) with an arbitrary nonzero finite-dimensional C-alegebra B in this definition. This notion of quantum nullhomotopy does not degenerate: Proposition 2.4. The identity function S1 → S1 is not quantum nullhomotopic as a loop in S1. Furthermore, it is distinct from the ordinary notion of nullhomotopy of loops: Corollary 3.3. Every loop in RP2 is quantum nullhomotopic. Corollary 3.5. There is a loop in S1 ∨ S1 that is quantum nullhomotopic, but not nullho- motopic in the ordinary sense. 3 Quantum sets. If X is a locally compact Hausdorff space, then C0(X) = {f ∈ (X, C) lim x→∞ f (x) = 0} is a C-algebra, which is unital iff X is compact. Thus, C-algebras are commonly viewed to be the quantum generalization of locally compact Hausdorff spaces. In particular, quantum sets should correspond to a class of C-algebras. The noncommutative dictionary does not yet include a widely accepted quantum general- ization of sets. If the C-algebra of compact operators is taken to correspond to a quantum set, then in fact there does exist a nonempty quantum set of extensions of the canonical loop to the unit disk. The unital ∗-homomorphism φ of theorem 2.6 can be viewed as a morphism, in the sense of Woronowicz [6], from C(S1) to C(D2) ⊗ L0(H), and thus, as a quantum family of extensions that is indexed by the quantum set corresponding the compact operator algebra L0(H). The notion of quantum set preferred by the author excludes those quantum locally com- pact Hausdorff spaces that correspond to infinite-dimensional compact operator C-algebras. Instead, quantum sets are taken to be those quantum locally compact Hausdorff spaces that correspond to c0-direct sums of finite matrix C-algebras. This notion of discreteness first appeared in the context of quantum Gelfand duality [13] [5] [15] [4]. It was later promoted by the author [7]. Decomposing the C-algebra of a quantum set into matrix algebras, we find that a loop is quantum nullhomotopic if and only if it admits a quantum set of extensions to the unit disk, that is, a quantum family of extensions indexed by a discrete quantum space. Quantum pseudotelepathy. Quantum families of functions can be interpreted as quantum strategies for games in which two players, traditionally named Alice and Bob, cooperate against a Referee without communicating with one another. In some instances, Alice and Bob have a winning strategy that utilizes quantum entanglement, despite having no winning strategy classically. The availability of a quantum strategy is equivalent to the existence of a quantum set of functions of one or another kind [2, proposition 1] [9] [1] [11] [7, 1.2]. In this context, the domain and codomain spaces are taken to be finite. However, it is also possible to interpret the quantum extension problem considered in this paper as an idealized game of the same kind. In the example of theorem 2.6, the Referee sends Alice and Bob elements of D2, and they respond with elements of S1. Alice and Bob lose if their responses disagree, provided that the Referee sent them both the same element of D2, or if either Alice or Bob fails to mirror the Referee's move, provided that the Referee sent them an element of S1. I speculate that this idealized game is not so far removed from the physical world as one might imagine. The players could perhaps exchange continuous values as the momenta of particles, and the Referee could be certain of sending the same value to both Alice and Bob by using a mechanism such as pair creation. The strategies implemented by Alice and Bob would necessarily be continuous, as a physical limitation. Preliminary computations suggest that it is possible to formulate some of the results of the present paper in terms of discrete games. However, these games retain some infinitary ele- ment, e. g., the number of possible moves is infinite, or the length of play is unbounded. Fur- thermore, the quantum advantage becomes probabilistic, rather than deterministic. Thus, the significance of such reformulations is uncertain. 4 Acknowledgements. I thank Neil Ross and Peter Selinger for organizing the Quantum Physics and Logic conference in Halifax, an enriching experience that led me to consider this quantum extension problem. I thank Greg Kuperberg for advising me to emphasize the topological aspect of the problem, in favor of the quantum communication aspect. I thank Rui Soares Barbosa and Sam Staton for giving me the opportunity to present this research at the Oxford Advanced Seminar on Informatic Structures. 1. quantum extensions Let S and T be compact Hausdorff spaces, and let H be a nonzero finite-dimensional Hilbert space. In this section, we demonstrate a one-to-one correspondence between the unital ∗-homomorphisms from C(T ) to C(S) ⊗ L(H), and the continuous functions from S to Hom(C(T ), L(H)). There are several reasonable ways to gloss the term C(S) ⊗ L(H) that turn out to be equivalent, and likewise for Hom(C(T ), L(H)). The C-algebras C(S) and L(H) are both nuclear, so there is a unique cross-norm on their algebraic tensor product. In fact, their algebraic tensor product is already isomorphic to the C-algebra of dim(H) × dim(H) matrices over C(S), so the term C(S)⊗L(H) can refer equivalently to the maximum tensor product, the minimum tensor product, or the algebraic tensor product of the two C- algebras. The space Hom(C(T ), L(H)) consists of unital ∗-homomorphisms from C(T ) to L(H). This set is typically equipped with the point-norm topology, which is characterized by the condition that a net (φα) converges to φ∞ if and only if for each function b in C(T ), the net (φα(b)) converges to φ∞(b) in the operator norm topology. We instead equip Hom(C(T ), L(H)) with the kaonization of the compact-open topology, which we will later show to be equivalent to the point-norm topology. We do so to reason in Steenrod's conve- nient category of compactly generated spaces [14]. We summarize a few of its basic aspects. Steenrod's convenient category is a subcategory of the category of topological spaces and continuous functions that includes all locally compact spaces and all metrizable spaces. It is convenient primarily in the sense that it is Cartesian closed: for all objects X, Y , and Z of the category, the space of continuous functions C(Y, Z) is itself an object of the category, and there is a natural one-to-one correspondence between the continuous function from X × Y to Z and the continuous functions from X to C(Y, Z). However, the category is inconvenient in the sense that the topology of the category-theoretic product X × Y is sometimes different from the usual product topology. The objects of Steenrod's category are the so-called compactly generated Hausdorff spaces. These spaces are characterized by the criterion that any set whose intersection with every compact set is closed must itself be closed. Kaonization is the functor from Hausdorff spaces to compactly generated Hausdorff spaces that adds closed sets according to this criterion; it is typically denoted by the lower-case letter k. The category-theoretic product of two compactly generated Hausdorff spaces X and Y is the kaonization of the usual topological product. The morphisms of Steenrod's category are ordinary continuous functions, and the set of continuous functions C(Y, Z) is canonically equipped with the kaonization of the If X is locally compact, then the product topology compact-open topology on this set. on X × Y is already compactly generated. If Y is compact, and Z is metrizable, then the compact-open topology on C(Y, Z) is also already compactly generated; it is simply the uniform topology. In particular, in Steenrod's category, C(S) = C(S, C) and C(T ) = C(T, C) are canonically equipped with their usual norm topologies. 5 With this overview of the compactly generated spaces, we are ready to formally state and prove the claimed one-to-one correspondence. Definition 1.1. Let A and B be unital C-algebras. We write Hom(A, B) for the set of all unital ∗-homomorphisms from A to B, equipped with the kaonization of the compact-open topology. The space Hom(A, B) is a subspace of the space C(A, B), itself equipped with the kaoniza- tion of its compact open topology. It easy to see that for closed subsets, the kaonization of the subspace topology coincides with the subpace topology of the kaonization. Lemma 1.2. We have a homeomorphism Hom(A, C(S) ⊗ B) ∼= C(S, Hom(A, B)), natural in the compact Hausdorff space S, and the unital C-algebras A and B. Implicitly, the morphisms of compact Hausdorff spaces are continuous functions, and the morphisms of unital C-algebras are unital ∗-homomorphisms. Proof. The compact-open topology on C(S, B) is the topology of uniform convergence. It is a metrizable topology, so it is compactly generated; thus, the topology on C(S, B) in Steenrod's convenient category is just the usual one. There is a well known isomorphism between the C-algebras C(S) ⊗ B and C(S, B) [10, theorem 6.4.17], which is natural in S and B, so it is sufficient to establish that Hom(A, C(S, B)) ∼= C(S, Hom(A, B)). We obtain this natural homeomorphism as a restriction of the following composition: C(A, C(S, B)) ∼= C(A × S, B) ∼= C(S, C(A, B)) Both natural homeomorphisms are instances of the Cartesian closedness of Steenrod's conve- nient category. Each continuous function π from A to C(S, B) corresponds to a continuous function π′ from S to C(A, B), which is defined by π′(s)(a) = π(a)(s) for all a ∈ A, for all s ∈ S. Thus, if π is a unital ∗-homomorphism, then for each s ∈ S, π′(s) is a uni- tal ∗-homomorphism, and vica versa, because the algebraic structure of C(S, B) is defined pointwise. We conclude that the natural homeomorphism C(A, C(S, B)) ∼= C(S, C(A, B)) restricts to a natural homeomorphism Hom(A, C(S, B)) ∼= C(S, Hom(A, B)). (cid:3) For reference, we record that each unital ∗-homomorphism φ : A → C(S) ⊗ B corresponds to the continuous function s 7→ (evals ⊗ idB)(φ( · )). Lemma 1.3. Let V and W be Banach spaces. Write L(V, W )1 for the set of linear operators from V to W of norm at most 1. The point-norm topology on L(V, W )1 is equal to the compact-open topology. The point-norm topology is generated by subbasis elements of the form {t kt(v)−wk < ǫ} for v ∈ V , w ∈ W , and ǫ > 0. The compact-open topology is of course generated by subbasis elements of the form {t t(K) ⊆ U} for K ⊆ V compact, and U ⊆ W open. Both Banach spaces are equipped with their norm topologies. Proof. It is well known that convergence in the compact-open topology is equivalent to uniform convergence on compact subsets of the domain, whenever the codomain is a metric space. The point-norm topology is just the topology of pointwise convergence. Convergence of the former kind clearly implies convergence of the latter kind, so it remains only to show the converse. 6 Let (tα) be a net in L(V, W )1 converging pointwise to a linear operator t∞, itself au- tomatically of norm at most 1. Let K be a compact subset of V , and let ǫ > 0. Being compact, the set V is totally bounded, so we can cover K by a finitely many open balls Bǫ(v1), Bǫ(v2), . . . , Bǫ(vn) of radius ǫ, with centers v1, v2, . . . , vn ∈ V . Consider α sufficiently large so that ktα(v1) − t(v1)k < ǫ, ktα(v2) − tα(v2)k < ǫ, . . ., and ktα(vn) − tα(vn)k < ǫ. Every element v of K is within distance ǫ of some vi, so we calculate: ktα(v) − t(v)k ≤ ktα(v) − tα(vi)k + ktα(vi) − t∞(vi)k + kt∞(vi) − t∞(v)k ≤ ktαk · kv − vik + ǫ + kt∞k · kv − vik ≤ 3ǫ Thus, (tα) converges uniformly to t∞ on K. Therefore, pointwise convergence implies uniform convergence on compact subsets. (cid:3) Proposition 1.4. We have a bijection, natural in the compact Hausdorff space S, and the unital C-algebras A and B, Hom(A, C(S) ⊗ B) ∼= C(S, Hom(A, B)), where Hom(A, B) denotes the set of all unital ∗-homomorphisms from A to B, equipped with the point-norm topology. It takes each each unital ∗-homomorphism φ : A → C(S) ⊗ B to the function s 7→ (evals ⊗ id)(φ( · )). Proof. This proposition is just lemma 1.2, with a different topology on Hom(A, B). The point-norm topology on Hom(A, B) is equal to the compact-open topology by lemma 1.3. Thus, the topology on Hom(A, B) in lemma 1.2 is equivalently the kaonization of the point- norm topology. Since S is a compact Hausdorff space, the set of continuous functions from S to Hom(A, B) is the same if we equip Hom(A, B) with just the point-norm topology, without kaonizing it. Thus, the proposition follows. (cid:3) Lemma 1.5. Let T be a compact Hausdorff space. The bijection of proposition 1.4 takes the canonical unital ∗-homomorphism π : C(T ) → C(T ) ⊗ B, defined by π(a) = a ⊗ 1 to the continuous function ı : T → Hom(C(T ), B), defined by ı(t)(a) = a(t) · 1. Proof. Let p be the function from T to Hom(C(T ), B) that corresponds to π under the bijection of proposition 1.4. For each t ∈ T , and all a ∈ C(T ), we compute: p(t)(a) = (evalt ⊗ id)(π(a)) = (evalt ⊗ id)(a ⊗ 1) = a(t) ⊗ 1 = a(t) · 1. (cid:3) Theorem 1.6. Let S and T be compact Hausdorff spaces, and let B be a unital C ∗-algebra. Let S be a closed subset of S, with inclusion function j : S ֒→ S. Let f be a continuous function from S to T . Use the notation of proposition 1.4 and lemma 1.5. The following are equivalent: (1) The function ı ◦ f : S → Hom(C(T ), B) extends to a continuous function f : S → Hom(C(T ), B). S j S f T f ı 7 Hom(C(T ), B) (2) There is a unital ∗-homomorphism φ : C(T ) → C( S) ⊗ B such that (j⋆ ⊗ id) ◦ φ = (f ⋆ ⊗ id) ◦ π. C( S) ⊗ B j ⋆⊗id φ C(S) ⊗ B f ⋆⊗id C(T ) ⊗ B C(T ) π Proof. Apply proposition 1.4, with A = C(T ) to obtain the following commutative diagram: Hom(C(T ), C( S) ⊗ B)) (j ⋆⊗id)◦ Hom(C(T ), C(S) ⊗ B)) (f ⋆⊗id)◦ Hom(C(T ), C(T ) ⊗ B)) ∼= ∼= ∼= C( S, Hom(C(T ), B)) ◦j C(S, Hom(C(T ), B)) ◦f C(T, Hom(C(T ), B)) The homomorphism π is an element of Hom(C(T ), C(T ) ⊗ B)), and the map ı is an element of C(T, Hom(C(T ), B)). By lemma 1.5, the two elements correspond to each other under the indicated bijection. Condition (1) is equivalently that there exists an element f of C( S, Hom(C(T ), B)) whose image in C(S, Hom(C(T ), B)) is equal to the image of ı in that set. Condition (2) is equivalently that there exists an element φ of Hom(C(T ), C( S) ⊗ B)) whose image in Hom(C(T ), C(S) ⊗ B)) us equal to the image of π in that set. The commutative diagram demonstrates that the two conditions are equivalent, establishing the theorem. (cid:3) We close this section by observing that the space Hom(C(T ), Mn(C)) is necessarily com- pact for every n. Lemma 1.7. Let A be a unital C-algebra, and let n be a positive integer. The space Hom(A, Mn(C)) of unital ∗-homomorphisms from A to Mn(C), equipped with the compact- open topology, or equivalently, with the point-norm topology, is compact. If A is separable, then Hom(A, Mn(C)) has a countable basis for the point-norm topology. Proof. The two topologies are equivalent as a special case of lemma 1.3. To observe com- pactness, we put the set Hom(A, Mn(C)) through the following isomorphisms in Steenrod's convenient category: C(A, Mn(C)) ∼= C(A, C × · · · × C) ∼= C(A, C) × · · · × C(A, C) The set Hom(A, Mn(C)) is closed in C(A, Mn(C)). A unital ∗-homomorphism has norm at most 1, so Hom(A, Mn(C)) corresponds to a closed subset of L(A, C)1 × · · · × L(A, C)1. By lemma 1.3, the compact-open topology on L(A, C)1 is just the point-norm topology, or equivalently, the weak topology. By Alaoglu's theorem, the unit ball L(A, C)1 is compact. Thus, the topology on L(A, C)1 as a subspace of C(A, C) in Steenrod's convenient category is just the weak* topology. Therefore the space Hom(A, Mn(C)) is compact. Furthermore, if A is separable, then the unit ball L(A, C)1 is metrizable [12, exercise 2.5.3], implying the same 8 for Hom(A, Mn(C)). Of course, for compact Hausdorff spaces metrizability is equivalent to the existence of a countable basis. (cid:3) 2. quantum loops We write S1 for the unit circle, D2 for the unit disk, and j : S1 ֒→ D2 for the inclusion map. Definition 2.1. Let T be a compact Hausdorff space. A loop in T is a function f : S1 → T . A loop f is quantum nullhomotopic just in case there is a nonzero finite-dimensional Hilbert space H and a unital ∗-homomorphism φ : C(T ) → C(D2) ⊗ L(H) such that (j⋆ ⊗ id)φ(a) = f ⋆(a) ⊗ 1H for all a ∈ C(T ). C(D2) ⊗ L(H) j ⋆⊗id φ C(S1) ⊗ L(H) f ⋆⊗id C(T ) ⊗ L(H) C(T ) a⊗1←[a Equivalently, f is nullhomotopic if and only if there exists an positive integer n, and a unital ∗-homomorphism φ : C(T ) → Mn(C(D2)) such that applying j⋆ to every element of the matrix φ(a) yields the matrix f ⋆(a) · In for all a ∈ C(T ). Corollary 2.2. Let T be a compact Hausdorff space. A loop f in T is quantum nullhomotopic if and only if there is a positive integer n such that the loop ı⋆(f ) = ı◦f in Hom(C(T ), Mn(C)) is nullhomotopic in the ordinary sense, where ı is defined by ı(t)(a) = a(t) · In. Proof. Apply theorem 1.6, with S = S1 and S = D2. (cid:3) Proposition 2.3. Let g be a continuous function from a compact Hausdorff space T1 to a compact Hausdorff space T2. If a loop f : S1 → T1 is quantum nullhomotopic in T , then the loop g ◦ f : S1 → T2 is quantum nullhomotopic in T2. Proof. Apply corollary 2.2. Let n be a positive integer such that the loop ı⋆(f ) : S1 → Hom(C(T1), Mn(C)) is nullhomotopic in the ordinary sense. We show that the loop ı⋆(g ◦ f ) : S1 → Hom(C(T2), Mn(C)) is a continuous image of the loop ı⋆(f ). ı⋆(f ) S1 Hom(C(T1), Mn(C)) ı⋆(g◦f ) ◦g⋆ Hom(C(T2), Mn(C)) For each point s on the circle S1, and all a2 ∈ C(T2), we calculate that ı⋆(g ◦ f )(s)(a2) = a2(g(f (s))) · In = g⋆(a2)(f (s)) · In = ı⋆(f )(s)(g⋆(a2)) = (ı⋆(f )(s) ◦ g⋆)(a2). Thus, the homomorphism ı⋆(g◦f )(s) is obtained by composing ı⋆(f )(s) with g⋆. This is a con- tinuous function from Hom(C(T1), Mn(C)) to Hom(C(T2), Mn(C)), since Hom(C( − ), Mn(C)) is a functor from the category of compact Hausdorff spaces and continuous functions to Steenrod's convenient category. Being a continuous image of a nullhomotopic loop, the loop ı⋆(g ◦ f ) is also nullhomotopic. Appealing a second time to corollary 2.2, we conclude that g ◦ f is quantum nullhomotopic. (cid:3) 9 Proposition 2.4. The identity function f : S1 → S1 is not quantum nullhomotopic as a loop in S1. Proof. Let f : S1 → S1 be the identity function, and suppose that f is quantum nullho- motopic as a loop in S1. By corollary 2.2, there is a positive integer n such that the loop ı⋆(f ) : S1 → Hom(C(S1), Mn(C)) is nullhomotopic. We will compose ı⋆(f ) with two other continuous functions to obtain a loop which we know is not nullhomotopic, arriving at a contradiction. First, let z be the inclusion of S1 into C; it is a unitary operator in C(S1). Evaluation at z is a function from Hom(C(S1), Mn(C)) to the space U(n) of unitary n × n matrices. This function evalz is continuous because all evaluation functions are continuous in Steenrod's convenient category. Second, the determinant is a continuous function from U(n) to S1. Thus, we obtain a nullhomotopic loop det ◦ evalz ◦ f in S1. ı⋆(f ) S1 Hom(C(S1, Mn(C)) evalz U(n) det S1 For each point s on the circle S1 we calculate that (det ◦ evalz ◦ ı⋆(f ))(s) = det(z(f (s)) · In) = det(s · In) = sn. It is a basic fact that this loop is not nullhomotopic. We have reached a contradiction. Therefore, f is not quantum nullhomotopic. (cid:3) Lemma 2.5. Let B a be a unital C-algebra, and let z : S1 ֒→ C be the inclusion map. Evaluation at z is a homeomorphism Hom(C(S1), B) → U(B), where U(B) denotes the set of unitary operators in B, equipped with the norm topology. Proof. Recall that the C-algebra C(S1) is isomorphic to the group C-algebra C∗(Z) [3, proposition VII.1.1]. The inclusion function z : S1 ֒→ C is mapped to the generator u1 of C∗(Z). The C-algebra C∗(Z) is evidently the universal C-algebra for a single unitary operator: for any C-algebra B, there is a bijective correspondence between the unital ∗-homomorphisms from C∗(Z) to B, and the unitary operators in B. Each such unital ∗- homomorphism sends the generator u1 to the corresponding unitary in B. Therefore, we have a bijection evalz from Hom(C(S1), B) to U(B), defined by ρ 7→ ρ(z). The bijection evalz is continuous, because all evaluation functions are continuous in Steen- rod's convenient category. To show that it is a homeomorphism, let (ρλ) be a net in Hom(C(S1), B) whose image (evalz(ρλ)) converges to evalz(ρ) for some homomorphism ρ in Hom(C(S1), B). In other words, (ρλ(z)) converges to ρ(z). The operations that make up the ∗-algebra structure of B are all continuous in the norm topology, so in fact, (ρλ(a)) converges to ρ(a) for all a in the unital ∗-subalgebra A0 of C(S1) that is generated by z. Let a be any element of C(S1). The algebra A0 is dense in C(S1) by the Stone-Weierstrass theorem, so we can choose a sequence (an) in A0 such that kan − ak ≤ 1/n for each positive 10 integer n. We now estimate that for each positive integer n, kρ(a) − ρλ(a)k ≤ kρ(a) − ρ(an)k + kρ(an) − ρλ(an)k + kρλ(an) − ρλ(a)k ≤ kρk · kan − ak + kρ(an) − ρλ(an)k + kρλk · kan − ak ≤ = 1 n 2 n + kρ(an) − ρλ(an)k + 1 n + kρ(an) − ρλ(an)k The term kρ(an) − ρλ(an)k converges to 0 as λ goes to infinity, so lim supλ kρ(a) − ρλ(a)k ≤ 2/n. We have this estimate for each positive integer n, so limλ kρ(a) − ρλ(a)k = 0. In other words, ρλ converges to ρ in the point-norm topology. Therefore, eval−1 : U(B) → z Hom(C(S1), B) is norm-(point-norm) continuous, and thus, by lemma 1.3, it is norm- (compact-open) continuous. It is an elementary fact about compactly generated spaces that kaonization does not introduce or eliminate compact subset. In this case, if a set K ⊆ Hom(C(S1), B) if compact in the kaonized compact-open topology, then it is compact in the compact-open topology itself, and vice versa. Thus, we have shown that evalz induces a one-to-one correspondence between the compact subsets of Hom(C(S1), B) and the compact subsets of U(B). Since both spaces are compactly generated, evalz must be a homeomorphism. (cid:3) Theorem 2.6. Let H be a separable infinite-dimensional Hilbert space. There is a unital ∗-homomorphism φ : C(S1) → C(D2) ⊗ L(H) such that (j⋆ ⊗ id)(φ(a)) = a ⊗ 1H for all a ∈ C(S1), where j : S1 ֒→ D2 is the inclusion map. C(D2) ⊗ L(H) j ⋆⊗id φ C(S1) ⊗ L(H) a⊗1←[a C(S1) The proof of this theorem is a more elaborate variation on the proof of proposition 2.4. We write U(H) for the set of unitary operators in L(H), equipped with the norm topology. Proof. We apply theorem 1.6, with S = S1, T = S1, S = D2, B = L(H), and f = idS1. Specifically, we show that the function ı : S1 → Hom(C(S1), L(H)), defined by ı(s)(a) = a(s) · 1H for all a ∈ C(S1) and s ∈ S1, extends to D2. Composing ı with the homeomorphism evalz of lemma 2.5, we obtain the loop evalz ◦ ı in U(H). By Kuiper's theorem [8, theorem (3)], U(H) is contractible, so the loop evalz ◦ ı is nullhomotopic. Since evalz is a homeomorphism by lemma 2.5, the loop ı must also be nullhomotopic. Applying, theorem 1.6, we reach the desired conclusion. (cid:3) 3. every loop in RP2 is quantum nullhomotopic We define the circle S1 to be a subspace of C, and the sphere S2 to be a subspace of C × R. Explicitly, the circle S1 consists of complex numbers α such that α2 = 1, and the sphere S2 consists of pairs (α, t) in C × R such that α2 + t2 = 1. The sphere S2 is homeomorphic to CP1; we explicitly describe one such homeomorphism that is convenient in this context. 11 Write R1(M2(C)) for the space of reflection matrices of negative determinant, i. e., for the space of 2 × 2 complex matrices b such that b∗ = b, b2 = 1, and det(b) = −1. There is a nice homeomorphism h : S2 → R1(M2(C)) given by the formula h(α, t) = (cid:18) t α α −t(cid:19) . It satisfies h(−x) = −h(x) for all points x ∈ S2. The elements of R1(M2(C)) are exactly the matrices with spectrum equal to {1, −1}. Thus, for any matrix b in R1(M2(C)), the matrices (1 + b)/2 and (1 − b)/2 are rank-one projections, orthogonal to each other. We obtain a pair of homeomorphisms x 7→ (1 + h(x))/2, and x 7→ (1 − h(x))/2, from the space S2 to the space P1(M2(C)) of rank-one projections, with the property that their values at every point x of the sphere S2 are orthogonal. Lemma 3.1. Let T be a compact Hausdorff space. The space Hom(C(T ), M2(C)) is a quotient of S2 × T × T : S2 × T × T q1 (S2 × T × T )/Z2 q2 Hom(C(T ), M2(C)) All three spaces are compact Hausdorff spaces. The space (S2 × T × T )/Z2 is the quotient of S2 × T × T by the involution (x, t1, t2) 7→ (−x, t2, t1), and q1 is the corresponding quotient map. The quotient map q2 takes each orbit {(x, t1, t2), (−x, t2, t1)} to the homomorphism (∗) evalt1( · ) I2 + h(x) 2 + evalt2( · ) I2 − h(x) 2 . The quotient map q2 identifies distinct orbits {(x, t1, t2), (−x, t2, t1)} and {(x′, t′ if and only if t1 = t2 = t′ 1 = t′ 2. 1, t′ 2), (−x′, t′ 2, t′ 1)} Proof. The first space S2 × T × T is a compact Hausdorff space because it is a product of compact Hausdorff spaces. The second space (S2 × T × T )/Z2 is a compact Hausdorff space because it is the quotient of a compact Hausdorff space by the action of a finite group. The third space Hom(C(T ), M2(C)) is a compact Hausdorff space by lemma 1.7. The expression (∗) clearly names a morphism in Steenrod's convenient category, i. e., a continuous function from S2 × T × T to C(C(T, C), M2(C)). It is immediately apparent from the expression that this continuous function is invariant under the action of Z2. Therefore, it factors through the quotient map q1, via a continuous function q2. For all points x in S2, the matrices (1 + h(x))/2 and (1 − h(x))/2 are rank-one projections that sum to the identity matrix. Furthermore, for all points t1 and t2 in T the evaluation functions evalt1 and evalt2 are unital ∗-homomorphisms. It follows that for all points (x, t1, t2) in the product space S2 × T × T , the expression (∗) names a unital ∗-homomorphism. Therefore, the range of q2 is a subset of Hom(C(T ), M2(C)). We collect a couple of basic facts. First, the range of any element ρ in Hom(C(T ), M2(C)) is a commutative unital C-subalgebra of M2(C). In particular, it is isomorphic to either C or C2. Second, if ρ = (q2 ◦ q1)(x, t1, t2) for some (x, t1, t2) ∈ S2 ×T ×T , then the dimension of ρ(C(T )) depends on whether the points t1 and t2 are equal or distinct. If they are equal, then ρ = evalt1(·)I2, so ρ(C(T )) ∼= C. If t1 and t2 are distinct, then the distinct characters evalt1 and evalt2 both factor through ρ : C(T ) → ρ(C(T )), so ρ(C(T )) cannot be one-dimensional, and thus, ρ(C(T )) ∼= C2. 12 1, t′ 1 = t′ 1( · )I2, so t1 = t′ Let ρ be any element of Hom(C(T ), M2(C)). First, assume that ρ(C(T )) ∼= C. It follows that ρ is the composition of a unital ∗-homomorphism γ : C(T ) → C with the canonical inclusion C ֒→ M2(C). By Gelfand duality, any such homomorphism γ is equal to evalt for some point t. Thus, ρ = evalt( · )I2 = q2(q1((1, 0), t, t)). Now assume that ρ(C(T )) ∼= C2. It follows that we can write ρ as γ1( · )b1 + γ2( · )b2, for pairwise orthogonal projections b1 and b2 in ρ(C(T )), and unital ∗-homomorphisms γ1 and γ2 from C(T ) to C. By Gelfand duality, there exist points t1 and t2 such that γ1 = evalt1 and γ2 = evalt2. There also exists a point x in the sphere S2 such that b1 = (1 + h(x))/2, and consequently, such that b2 = (1 − h(x))/2. Thus, ρ = evalt1( · )((1 + h(x))/2) + evalt2( · )((1 − h(x))/2) = q2(q1(x, t1, t2)). Therefore, q2 is surjective. As any continuous surjective function between compact Hausdorff spaces it must be a quotient map. 2), (−x′, t′ 2, then q(w) = evalt1( · )I2 = evalt′ 1)} be orbits in (S2 × T × 2, t′ 1( · )I2 = q2(w′). Conversely, T )/Z2. If t1 = t2 = t′ assume that q2(w) = q2(w′), and write ρ = q2(w) = q2(w′). The range ρ(C(T )) is isomorphic to either C or C2. If it is isomorphic to C, then t1 = t2, t′ 2, and evalt1( · )I2 = q2(w) = q2(w′) = evalt′ Let w = {(x, t1, t2), (−x, t2, t1)} and w′ = {(x′, t′ 1. Thus, t1 = t2 = t′ 1 = t′ 2. 1 = t′ Therefore, assume that the range ρ(C(T )) is isomorphic to C2. The two minimal projec- tions in q2(w)(C(T )) are (1 + h(x))/2 or (1 − h(x))/2, and similarly the two minimal projec- tions in q2(w′)(C(T )) are (1 + h(x′))/2 or (1 − h(x′))/2. Since the C-algebras q2(w)(C(T )) and q2(w′)(C(T )) are equal, the projection (1 + h(x))/2 must be equal to either (1 + h(x′))/2 or to (1 − h(x′))/2. In other words, h(x) must be equal to either h(x′) or h(−x′), so w = w′. Therefore, q2(w) = q2(w′) if and only if w = w′ or t1 = t2 = t′ (cid:3) 1 = t′ 2. If ψ is a path in S2, and φ1 and φ2 are paths in T , we write [ψ, φ1, φ2] for the homotopy class of the path (ψ, φ1, φ2) defined by τ 7→ (ψ(τ ), φ1(τ ), φ2(τ )). Theorem 3.2. Let T be a compact Hausdorff space with distinguished point t0. Write q for the composition of the quotient maps q1 and q2 in lemma 3.1, and write q∗ : π1(S2 × T × T, ((1, 0), t0, t0)) → π1(Hom(C(T ), M2(C)), evalt0) for the induced homomorphism of fundamental groups. For every path ϕ : [0, 1] → T , begin- ning and ending at t0, we have that q∗[cnst(1,0), ϕ, cnstt0] = q∗[cnst(1,0), cnstt0, ϕ]. Proof. Let σ be the path in S2, beginning at (1, 0) and ending at (−1, 0), defined by σ(τ ) = (exp(πiτ ), 0). The path q ◦ (σ, cnstt0, cnstt0) is the constant path at evalt0, by lemma 3.1. It follows that q∗[σ, cnstt0, cnstt0] is the identity element of π1(Hom(C(T ), M2(C)), evalt0). Let ϕ : [0, 1] → T be any path in T beginning and ending at t0. Appealing to lemma 3.1 in the first step, we calculate that q∗[cnst(1,0), ϕ, cnstt0] = q∗[cnst(−1,0), cnstt0, ϕ] = q∗[cnst(−1,0), cnstt0, ϕ] · q∗[σ, cnstt0, cnstt0] = q∗([cnst(−1,0), cnstt0, ϕ] · [σ, cnstt0, cnstt0]) = q∗[σ, cnstt0, ϕ] = q∗([σ, cnstt0, cnstt0] · [cnst(1,0), cnstt0, ϕ]) = q∗[σ, cnstt0, cnstt0] · q∗[cnst(1,0), cnstt0, ϕ] = q∗[cnst(1,0), cnstt0, ϕ] 13 The embedding ı : T ֌ Hom(C(T ), M2(C)) clearly factors through q: (cid:3) S2 × T × T cnst(1,0)×∆ q T ı Hom(C(T ), M2(C)) The map diagonal map ∆ : T → T × T is defined by ∆(t) = (t, t). Corollary 3.3. Every loop in RP2 is quantum nullhomotopic. Proof. Let t0 be any point of T = RP2, and let ϕ : [0, 1] → T be any path that begins and ends at t0. ı∗[ϕ] = q∗((cnst(1,0) × ∆)∗[ϕ]) = q∗[cnst(1,0), ϕ, ϕ] = q∗([cnst(1,0), ϕ, cnstt0] · [cnst(1,0), cnstt0, ϕ]) = q∗[cnst(1,0), ϕ, cnstt0] · q∗[cnst(1,0), cnstt0, ϕ] = q∗[cnst(1,0), ϕ, cnstt0] · q∗[cnst(1,0), ϕ, cnstt0] = q∗([cnst(1,0), ϕ, cnstt0] · [cnst(1,0), ϕ, cnstt0]) (cid:3) Thus, ı∗[φ] is the identity element of π1(Hom(C(T ), M2(C)), evalt0). Therefore, for every loop f : S1 → T , the loop ι ◦ f is nullhomotopic. By corollary 2.2, we conclude that every loop in T = RP2 is quantum nullhomotopic. Proposition 3.4. Let T be a compact Hausdorff space with distinguished point t0. The group ı∗(π1(T, t0)) is commutative. Proof. Let ı∗[ϕ1] and ı∗[ϕ2] be elements of ı∗(π1(T, t0)). Appealing to theorem 3.2, we cal- culate that q∗[cnst(1,0), ϕ1, cnstt0] · q∗[cnst(1,0), ϕ2, cnstt0] = q∗[cnst(1,0), ϕ1, cnstt0] · q∗[cnst(1,0), cnstt0, ϕ2] = q∗[cnst(1,0), ϕ1, ϕ2] = q∗[cnst(1,0), cnstt0, ϕ2] · q∗[cnst(1,0), ϕ1, cnstt0] = q∗[cnst(1,0), ϕ2, cnstt0] · q∗[cnst(1,0), ϕ1, cnstt0] Similarly q∗[cnst(1,0), cnstt0, ϕ1] and q∗[cnst(1,0), cnstt0, ϕ2] commute. Since we can write ı∗[ϕ1] = q∗[cnst(1,0), cnstt0, ϕ1] · q∗[cnst(1,0), ϕ1, cnstt0], and likewise for ı∗[ϕ2], it follows that ı∗[ϕ1] and ı∗[ϕ2] commute. Therefore, the group ı∗(π1(T, t0)) is commutative. (cid:3) Corollary 3.5. There is a loop in (S1, 1) ∨ (S1, 1) that is quantum nullhomotopic, but not nullhomotopic in the ordinary sense. Proof. The fundamental group of T = S1 ∨S1 is the free group on two generators. Therefore, the commutator of its two generators yields a loop f in S1 ∨ S1 that is not nullhomotopic. However, the image of this loop f in Hom(C(T ), M2(C)) must be nullhomotopic by propo- sition 3.4. It follows that f is quantum nullhomotopic by corollary 2.2. (cid:3) 14 References [1] S. Abramsky, R. S. Barbosa, N. de Silva, and O. Zapata, The Quantum Monad on Relational Structures, Proc. MFCS 2017 (2017). [2] P. J. Cameron, A. Montanaro, M. W. Newman, S. Severini, and A. Winter, On the quantum chromatic number of a graph, Electron. J. Combin. 14 (2007), no. 1. [3] K. R. Davidson, C-algebras by Example, Fields Institute Monographs 6 (1996). [4] K. De Commer, P. Kasprzak, A. Skalski, and P. So ltan, Quantum actions on discrete quantum spaces and a generalization of Clifford's theory of representations (2016), available at arXiv:1611.10341. [5] E. Effros and Z.-J. Ruan, Discrete Quantum Groups I. The Haar Measure, Int. J. Math 5 (1994). [6] J. M. Garcia-Bondia, J. C. Varilly, and H. Figueroa, Elements of Noncommutative Geometry, Birkauser, 2000. [7] A. Kornell, Quantum sets, conditionally accepted for publication in J. Math. Phys. [8] N. H. Kuiper, The homotopy type of the unitary group of Hilbert space, Topology 3 (1965). [9] L. Mancinska and D. E. Roberson, Quantum homomorphisms, J. Combin. Theory, Series B 118 (2016). [10] G. J. Murphy, C-algebras and Operator Theory, Academic Press, 1990. [11] B. Musto, D. J. Reutter, and D. Verdon, A compositional approach to quantum functions, to appear in J. Math. Phys. (2017), available at arXiv:1711.07945. [12] G. K. Pedersen, Analysis NOW, Springer Science+Business Media, 1995. [13] P. Podle´s and S. L. Woronowicz, Quantum deformation of Lorentz Group, Comm. Math. Phys. 130 (1990), no. 2. [14] N. E. Steenrod, A convenient category of topological spaces, Michigan Math J. 14 (1967), no. 2. [15] A. Van Daele, Discrete quantum groups, J. Algebra 180 (1996). 15
1508.07904	3	1508	2017-07-19T20:08:35	Part II, Free Actions of Compact Groups on C*-Algebras	[ "math.OA", "math.QA" ]	We study a simple subclass of free actions of non-Abelian groups on unital C*-algebras, namely cleft actions. These are characterized by the fact that the associated noncommutative vector bundles are trivial. In particular, we provide a complete classification theory for these actions and describe its relations to classical principal bundles.	math.OA	math	Part II, Free Actions of Compact Groups on C∗-Algebras Kay Schwieger ∗ Stefan Wagner † Abstract We study a simple class of free actions of non-Abelian groups on unital C∗-alge- bras, namely cleft actions. These are characterized by the fact that the associated noncommutative vector bundles are trivial. In particular, we provide a complete classification theory for these actions and describe its relations to classical principal bundles. Keyword: Weakly cleft action, C∗-algebra, factor system, cocycle action MSC2010: 46L85, 37B05 (primary), 55R10, 16D70 (secondary). 1 Introduction In this presentation we investigate a special class of group actions on unital C∗-algebras. The experience with group actions on topological spaces shows that free actions are easier to understand and to classify. In this context a free action of a group G on a space P is typically regarded as a topological principal G-bundle over the quotient P/G. For instance, locally trivial bundles are, up to equivalence, characterized by the Čech cohomology H 1(X, G). It is therefore reasonable to expect that free actions on noncommutative spaces are easier to understand and classify, too. In the first part of this series [18] we investigated free actions of compact Abelian groups on unital C∗-algebras. This classification relies on the fact that the corresponding iso- typic components are Morita self-equivalence over the fixed point algebra. For non- Abelian compact groups the bimodule structure is more subtle. For this reason the current article concentrates on a simple class of free actions of non-Abelian groups, ∗University of Helsinki, [email protected] †Blekinge Tekniska Högskola, [email protected] 1 namely cleft actions. Regarded as noncommutative principal bundles, these actions are characterized by the fact that all associated noncommutative vector bundles are trivial. In particular, for such actions all Chern classes (cf. [2]) vanish. Although this prop- erty looks very limiting, in fact, many noncommutative phenomena already show up here. Therefore, cleft actions may be used as toy models of noncommutative principal bundles, e. g. for Chern-Simmons actions (see [23, 12]). The earliest classification result for free actions of compact, but not necessarily Abelian groups goes back to Wassermann [21]. He showed that free and ergodic actions, i. e., actions with full multiplicity and trivial fixed point space, are, up to equivalence, char- acterized by unitary 2-cocycles on the dual of the group. For finite groups Davydov [3] presented an alternative classification using classical group cohomology. Although free and ergodic actions only correspond to principal bundles over the singleton base space, these result yet show that noncommutative geometry admits more interesting examples than the classical theory. The prototypes of free and ergodic actions are the quantum 2-tori Aθ, θ ∈ T, equipped with the gauge action of T2. Varying θ ∈ T, these are in fact all non-equivalent free ergodic actions of the 2-torus T2. Non-ergodic examples can then be obtained by taking continuous bundles of ergodic actions. A prominent example of this type is the gauge action of the 2-torus T2 on the Heisenberg group C∗-algebra A (see [7]). The algebra A can be written as C∗ -- bundle over T, where for each θ ∈ T the fiber is the noncommutative 2-tori Aθ. The action of T2 on the algebra A is the fiberwise gauge action. Echterhoff, Nest, and Oyono-Oyono [7] proposed such bundles as noncommutative principal torus bundles. Concerning non-Abelian groups, the classification results of Wassermann can easily be extended to bundles of free ergodic actions. Such bundle actions are up to equivalence determined by the continuous family of unitary 2-cocycles corresponding to the ergodic actions in each fiber. Beyond the concept of bundles of ergodic actions there are multiple directions to in- troduce further noncommutativity. Any C∗-algebra obtained by forming a bundle of ergodic actions over a compact base space X always contains C(X) in its center. In order to explore noncommutative principal bundles over a noncommutative base space, this requirement must be abolished. Without this restriction new examples are immedi- ately available even for a classical base space. For instance, given a coaction γ of G on X, the crossed product C(X) ⋊γ G with the dual action of G is cleft (cf. Example 4.9) but C(X) is central only if the coaction is inner, i. e., if it is trivial up to a 1-cocycle (cf. Example 5.15). A second direction to explore is to not only consider actions of classical groups but of quantum groups. An algebraic approach to cleft actions of quantum groups (alias Hopf algebras) is already established in the theory of Hopf-Galois extensions (see e. g. [6, 17]). There cleft actions are free actions with a convolution invertible cleaving map. Doi [5] has shown that cleft actions can be written as a twisted crossed product and provided 2 a classification for these crossed products. Also the work of Vaes and Vainerman [20] should be mentioned, who studied cleft actions of locally compact Hopf von Neumann algebras. An essential part of our presentation will lift the algebraic constructions to the C∗-algebraic framework. In this article we restricts ourselves to classical compact groups for sake of a simple presentation and because most essential problems are already present in this restricted context. The article is structured as follows. After this introduction and some preliminaries, we recall the basic decomposition of a C∗-dynamical system into (generalized) isotypic components in Section 3. In Section 4 we introduce cleft C∗-dynamical systems and a weaker notion, called weakly cleft, and discuss their relations. Section 5 presents the characterization of weakly cleft systems in terms of factor systems, alias cocycle actions. Moreover, we discuss some relations between the type of the dynamical system and the form of its factor system. In particular, we classify all cleft topological (classical) principal bundles. Finally, in Section 6 we show that factor systems and weakly cleft C∗-dynamical systems are, up to equivalence, indeed in 1-1-correspondence by explicitly constructing the dynamical system. 2 Preliminaries and Notations Once and for the rest of the paper we fix a compact group G. All integrals over G are taken with respect to the Haar probability measure. By a representation of G we always mean a finite-dimensional unitary representation. For a representation (π, V ) we write dπ for its dimension, for the dual representation we write (¯π, ¯V ). The set of equivalence classes of irreducible representations will be denoted by G. All our constructions behave naturally with respect to intertwiners and hence, for sake of brevity, we do not distinguish between a representation of G and its equivalence class. Let A be a unital C∗-algebra. For the unit of A we write 1A or simply 1. For an element u ∈ A we denote by Ad[u] : A → A the map x 7→ uxu∗. All tensor products of C∗-algebras are taken with respect to the minimal tensor product. We will frequently deal with multiple tensor products of unital C∗-algebras A, B, and C. If there is no ambiguity, we regard A, B, and C as subalgebras of A ⊗ B ⊗ C and extend maps on A, B, or C canonically by tensoring with the identity map. For sake of clarity we may occasionally use the leg numbering notation, e. g., for x ∈ A ⊗ C we write x13 to denote the corresponding element in A ⊗ B ⊗ C. Our main focus in this paper will be on C∗-dynamical systems, by which we mean triples (A, G, α) consisting of a unital C∗-algebra A together with a group of ∗-automorphisms αg : A → A, g ∈ G, such that for each x ∈ A the map g 7→ αg(x) is continuous. We typically write B = AG for the fixed point algebra of the dynamical system and we 3 denote by P0 the associated the conditional expectation P0(x) := RG αg(x) dg, x ∈ A. More general, for an irreducible representation π ∈ G we denote by Pπ : A → A the G-equivariant projection onto the isotypic component A(π) := Pπ(A), which is given by Pπ(x) := dπZG Tr(π∗ g) αg(x) dg, x ∈ A. Two C∗-dynamical system (A, G, α) and (A′, G, α′) are called equivalent if there is an isomorphism ϕ : A → A′ with ϕ ◦ αg = α′ g ◦ ϕ for all g ∈ G. For the necessary background on modules of C∗-algebras we recommend [1], here we only briefly recall some relevant definition. For a C∗-algebra B a right pre-Hilbert B-module is a right B-module H equipped with a sesquilinear map h·, ·iB : H×H → B that satisfies the usual axioms of a definite inner product with B-linearity in the second component.1 We may define a norm on H by putting kxkH = khx, xiBk1/2. If H is complete with respect to this norm then H is called a right Hilbert B-module. A linear operator T : H → H on a right Hilbert B-module H is called adjointable if there is an operator T ⋆ : H → H satisfying hT x, yiB = hx, T ⋆yiB for all x, y ∈ H. Adjointable operators are automatically bounded but the converse does not hold. The set L(H) of all adjointable operators on a right Hilbert B-module is a C∗-algebra. A correspondence over B, or a right Hilbert B- bimodule, is a B-bimodule H equipped with a B-valued inner product h·, ·iB which turns it into a right Hilbert B-module such that the left action of B on H is via adjointable operators. For two correspondences H, K over the same algebra B we denote by H ⊗B K their tensor product, which is again a correspondence over B. The elementary tensors x ⊗ y (x ∈ H, y ∈ K) are total in H ⊗B K. The inner product on H ⊗B K is given by hx1 ⊗ y1, x2 ⊗ y2iB = hy1, hx1, x2iB . y2iB for all x1, x2 ∈ H1 and y1, y2 ∈ K. 3 Decomposition of C∗-Dynamical Systems As a background for later discussions we first would like to recall the general decom- position of C∗-dynamical systems (A, G, α). In analogy to the GNS-construction the conditional expectation P0 onto the fixed point space B := AG allows to equip A with the definite B-valued inner product hx, yiB := P0(x∗y) = ZG αg(x∗y) dg for x, y ∈ A. We write L2(A) for the right Hilbert B-module obtained by taking the completion of A with respect to the corresponding norm. The C∗-algebra A admits a 1 In the literature the notion is usually relaxed even further to pre-C∗-algebras and non-definite inner products. But we do not need this more general framework. 4 faithful representation on L2(A) given by λ : A → L(cid:0)L2(A)(cid:1), λ(x)y := x · y. This allows to identify A with the subalgebra λ(A) ⊆ L(cid:0)L2(A)(cid:1) and we implicitly do so unless confusion arise. For each g ∈ G we have a unitary operator on L2(A) given by Ugx := αg(x) for x ∈ A ⊆ L2(A). The map g 7→ Ug is a strongly continuous representation of G that implements the automorphisms αg, g ∈ G, in the sense that λ(cid:0)αg(x)(cid:1) = Ug λ(x) U ⋆ g , x ∈ A. As every representation of G, the algebra A can be decomposed into its isotypic compo- nents with respect to the actions. More precisely, the sum of the isotypic components is A(π) is a dense ∗-subalgebra of A. Moreover, the isotypic components are mutually orthogonal, closed, linear subspaces of L2(A) with direct and Palg π∈ G L2(A) = Mπ∈ G A(π). One aspect of this presentation will be the multiplication of A. It is worth noting that the multiplication is determined by the action of A(π) ⊆ A on A(ρ) ⊆ L2(A) for all pairs of irreducible representations π and ρ. Instead of dealing with isotypic components, it will be more convenient to consider, for a representation π, the generalized isotypic component A2(π) := (cid:8)x ∈ A ⊗ L(V )(cid:12)(cid:12) πg · αg(x) = x ∀g ∈ G(cid:9). Obviously, A2(π) is a B-bimodule for the usual left and right multiplication. In addition, we may equip A2(π) with the B-valued inner product hx, yiB := 1 dπ (idA ⊗ Tr)(x∗y) for x, y ∈ A2(π). Then the space A2(π) is a correspondence over B (see e. g. [4] for completeness of the norm). If π is irreducible, the map x 7→ (idA ⊗ Tr)(x) gives an isomorphism of correspondences from A2(π) to the dual isotypic component A(¯π) with inverse given by y 7→ dπRG αg(y) ⊗ πg dg. In the following we will frequently use this identification. The multiplication between isotypic components is well captured by family of maps mπ,ρ : A2(π) ⊗B A2(ρ) −→ A2(π ⊗ ρ) ⊆ A ⊗ L(V ) ⊗ L(W ) mπ,ρ(x ⊗ y) := x12 · y13 for pairs of representations (π, V ) and (ρ, W ) of G. For an irreducible representation σ ∈ G and a representation π of G let us denote by Pσ,π : L(Vπ) → L(Vσ) the map 5 Pσ,π(x) := Pk=1 v∗ kxvk with an orthonormal basis of intertwiners v1, . . . , vm : Vσ → Vπ. The map Pσ,π does not depend on the choice of intertwiners. Then for irreducible representations π, ρ ∈ G and elements x ∈ A2(π) and y ∈ A2(ρ) the operator λ(x) takes the form λ(x)y = Mσ∈ G Pσ,π⊗ρ(cid:0)mπ,ρ(x ⊗ y)(cid:1). (1) Since the sum of the isotypic components is norm dense in A, the C∗-algebra λ(A) is generated by the set of these operators λ(x) with x ∈ A2(π), π ∈ G. In particular, the multiplication of A can be recovered from the maps mπ,ρ for π, ρ ∈ G in this way. 4 Cleft and Weakly Cleft C∗-Dynamical Systems In principle, C∗-dynamical systems (A, G, α) with a given fixed point algebra B can be classified in a functorial way in terms of the module structure of the generalized isotypic components and their multiplicative relation (see [11]). In this presentation we will focus on the class of cleft actions. Definition 4.1. A C∗-dynamical system (A, G, α) is called cleft if for every irreducible representation (π, V ) of G the set A2(π) ⊆ A ⊗ L(V ) contains a unitary element. Cleft C∗-dynamical systems are precisely the so called semidual actions discussed in [22]. For such a C∗-dynamical system (A, G, α) it follows along the same lines as in [22, Thm. 10] that the crossed product A ⋊ G is isomorphic to AG ⊗ K, generalizing Green's Theorem (cf. [8, Cor. 15] and [7]). In the algebraic theory of actions of Hopf algebras these actions are up to equivalence given by twisted crossed products (cf. [6, 5], see also [20]). Most arguments in our discussion rely on the following weaker hypothesis only and establishing the results in a slightly wider framework has some technical advantages later on. Definition 4.2. A C∗-dynamical system (A, G, α) is called weakly cleft if for every irreducible representation (π, V ) of G the set A2(π) ⊆ A ⊗ L(V ) contains an element s such that s∗s = 1 and ss∗x = x for all x ∈ A2(π). We call such an element s a non-degenerate isometry. Unfortunately, we cannot present an examples of a weakly cleft but not cleft action. In fact, in simple examples our results show that weakly cleft dynamical systems are automatically cleft (see Lemma 4.6 and 5.9). We do not yet know whether this holds in general. Most proofs of this article only rely on the weakly cleft assumption but can 6 be simplified for cleft systems. The advantage of dealing with weakly cleft actions is that this property can be characterized by the right Hilbert module structure of the (generalized) isotypic components. More precisely, the following lemma shows that the C∗-system (A, G, α) is weakly cleft if and only if each isotypic component A(π), π ∈ G, is a free right Hilbert B-module of rank d2 π. In the ergodic case, B = C1, this is the same as saying that the Hilbert space A(π) has its maximal dimension d2 π, i. e., A(π) has full multiplicity. Lemma 4.3. For an element s ∈ A2(π) the following statements are equivalent: (a) s is a non-degenerated isometry. (b) The map ϕ : B ⊗ L(Vπ) → A2(π), ϕ(x) := sx is an isomorphism of right Hilbert B-modules. Proof. For one implication notice that for an isometry s the map ϕ(x) = sx is an isometry for the right inner product. If s is non-degenerate, ϕ admits the inverse ϕ−1(y) = s∗y. Together this proves the implication from (a) to (b). For the converse implication, suppose that ϕ is an isomorphism. Then the selfadjoint element p := s∗s ∈ B ⊗ L(V ) is a projection, since it satisfies hp, xiB = hϕ(p), ϕ(x)iB = 1 dπ (id ⊗ Tr)(s∗ss∗sx) = hp2, xiB for all x ∈ B ⊗ L(V ). Injectivity of ϕ then implies that s is in fact an isometry because ϕ(p) = ss∗s = s = ϕ(1). It follows that the inverse of ϕ is given by ϕ−1(x) = s∗x and hence we have x = ϕ(cid:0)ϕ−1(x)(cid:1) = ss∗x for all x ∈ A2(π). To distinguish cleft and weakly cleft, let us introduce a third property, which is of great independent interest. Let (A, G, α) be a C∗-dynamical system. For a representation (π, V ) of G we write A2(π)A2(π)∗ for the linear subspace generated by products xy∗ of elements x, y ∈ A2(π), and we put C(π) := {x ∈ A ⊗ L(V ) (cid:0)αg ⊗ Ad[πg](cid:1)(x) = x ∀g ∈ G}. Definition 4.4. A C∗-dynamical system (A, G, α) is called free if for all representations π of G we have A2(π)A2(π)∗ = C(π). Free actions have attained special interest in the literature, see, e. g., [13, 16, 15, 4]. For the definition given here and the relation to noncommutative principal bundles, we refer to [18, Section 3]. It should be noted that A2(π)A2(π)∗ is always a ∗-ideal in C(π). Therefore, the dynamical system is free if and only if the closure of A2(π)A2(π)∗ contains the unit 1A⊗L(V ). 7 Lemma 4.5. A C∗-dynamical system is cleft if and only if it is weakly cleft and free. Proof. If (A, G, α) is cleft, it is obviously weakly cleft. Since for every representation (π, V ) of G we find a unitary u ∈ A2(π), the set A2(π)A2(π)∗ contains uu∗ = 1A⊗L(V ) and hence the dynamical system is free. To show the converse, suppose that (A, G, α) is weakly cleft and free and let (π, V ) be a representation of G. Then we find a non- degenerated isometry s ∈ A2(π) and obtain A2(π)A2(π)∗ = (cid:0)s · B ⊗ L(V )(cid:1) ·(cid:0)s ⊗ B ⊗ L(V )(cid:1)∗ = s · B ⊗ L(V ) · s∗. This set is closed and, since the dynamical system is free, it contains the unit. It follows that s must have full range, i. e., s is a unitary element and hence the system is cleft. Lemma 4.6. For a compact Abelian group G, every weakly cleft C∗-dynamical system is cleft. Proof. For an Abelian groups generalized isotypic components and the isotypic compo- nent of the dual representation literally coincide. If the C∗-dynamical system is weakly cleft, then for each π ∈ G we find an isometry s ∈ A(¯π). Hence for the dual representa- tion ¯π the set A(π)A(π)∗ = A(¯π)∗A(¯π) contain the element s∗s = 1. Example 4.7. We start with the most basic example as a prototype. Let B be a unital C∗-algebra and denote by C(G) the C∗-algebra of continuous functions on G. We consider the C∗-dynamical system(cid:0)B⊗C(G), G, id ⊗r(cid:1) where the action on C(G) is given by the right translation (rgf )(h) := f (hg) for f ∈ C(G) and h ∈ G. Clearly, the fixed point space of this system is B = B ⊗ 1G. We may identify elements in a tensor product with C(G) with functions on G in the usual way. Then, for an irreducible representation (π, V ) of G, the correspondence A2(π) over B is given by A2(π) = {f : G → B ⊗ L(V ) continuous πgf (g) = f (e) ∀g ∈ G}. The function u(g) := π∗ the system is cleft. g then obviously is a unitary element in A2(π), which shows that If B is commutative, i. e., B = C(X) for some compact Hausdorff space X, then the dynamical system of Example 4.7 can be be understood as a trivial principal G-bundle over the space X. The next example shows that also non-trivial principal bundles may give rise to cleft actions. Moreover, in the later Corollary 5.10 we will provide a characterization of cleft topological principal bundles. Example 4.8. Fix n ∈ N and denote by Cn := {ζ ∈ C ζ n = 1} the group of n-th roots of unity. We consider the C∗-algebra A := C(T) of continuous function on the circle with the action of Cn given by rotations, i. e., for ζ ∈ Cn and f ∈ C(T) put (αζf )(z) := f (ζ · z), z ∈ T. 8 For an irreducible representation of Cn, i. e., an element k ∈ Z/nZ, we have A2(−k) = A(k) = {f ∈ C(T) f (ζ · z) = ζ k · f (z) ∀z ∈ T}. The action is cleft, because an invertible element in A2(−k) is, for instance, given by the function f (z) := zk, z ∈ T. This dynamical system can be understood as the non-trivial principal Cn-bundle corresponding to the n-fold covering p : T → T, p(z) := zn. In fact, since all vector bundles over T are trivial (cf. [19, Section 18]), every principal bundles over T with arbitrary compact structure group G gives rise to a cleft C∗-dynamical system. Example 4.9. Consider a compact group G and a closed normal subgroup N and suppose that the action of N on C(G) by right translations is cleft (e. g. suppose that G/N is finite). Furthermore, let δ : A → M(cid:0)A ⊗ C ∗(G)) be a coaction of G on a unital C∗-algebra A. We recall that a twist over G/N is a unitary corepresentation W ∈ M(A ⊗ C ∗(G/N)) of G/N such that (id ⊗q) ◦ δ = Ad[W ] and (δ ⊗ id)(W ) = W13, where q denotes the natural ∗-homomorphism q : C ∗(G) → C ∗(G/N). Such a twist gives rise to an ideal IW of the crossed product A ⋊δ G, called twisting ideal, which is invariant under the canonical action of N on A ⋊δ G (see [14]). The twisted crossed product A ⋊δ,W G := (A ⋊δ G)/IW then carries an action of N, which we denote by α, that is, we obtain a C∗-dynamical system (A ⋊δ,W G, N, α). Since the natural embedding of C(G) into A ⋊δ G is G-equivariant, the factorized ho- momorphism kG : C(G) → A ⋊δ,W G is N-equivariant. Finally, a few moments thought shows that being cleft is preserved under equivariant ∗-homomorphisms and, therefore, it follows that the above C∗-dynamical system is cleft. 5 Factor Systems Let (A, G, α) be a C∗-dynamical system with fixed point algebra B. Suppose for the moment that the system is weakly cleft, that is, for each irreducible representation π ∈ G we find a non-degenerate isometry sπ ∈ A2(π) ⊆ A ⊗ L(Vπ). For the trivial representation, denoted by 1, we pick s1 := 1B. We may extend this family of isome- tries to non-irreducible representations by decomposing each representation (π, V ) into a direct sum of irreducible representations (σk, Vk) with intertwiners vk : Vk → V for k. It is easily checked that this provides a non-degenerate isometry in A2(π) and that the construction does not depend on the choice of intertwiners. each 1 ≤ k ≤ m and put sπ := Pm k=1 vksσkv∗ By Lemma 4.3, for each π ∈ G the space A2(π) is isomorphic to B ⊗ L(Vπ) as a right Hilbert B-module, but in general not as a left B-module. In order to describe the left 9 action of B consider the map γπ : B → B ⊗ L(Vπ), γπ(b) := s∗ π(b ⊗ 1π)sπ. (2) The correspondence A2(π) is then isomorphic to the vector space B ⊗ L(Vπ) equipped with the usual right multiplication by B, the usual right B-valued inner product, and the left multiplication given by b . x := γπ(b) x, for all x ∈ B ⊗ L(Vπ) and b ∈ B. Lemma 5.1. The map γπ is a unital ∗-homomorphism. Proof. Obviously, γπ(b∗) = γπ(b)∗ for all b ∈ B and γπ is unital, since sπ is an isometry. For each b ∈ B the element (b ⊗ 1)sπ lies in A2(π). The non-degeneracy of sπ then implies for all b1, b2 ∈ B that γπ(b1)γπ(b2) = s∗ π(b1 ⊗ 1)(b2 ⊗ 1)sπ = γπ(b1b2). π(b1 ⊗ 1)sπs∗ π(b2 ⊗ 1)sπ = s∗ Remark 5.2. We would like to point out that, since sπ is non-degenerated, the map γπ : B → B ⊗ L(Vπ) is uniquely determined by the property (b ⊗ 1)sπ = sπγπ(b) for all b ∈ B. In the following we will frequently deal with tensor products in which precisely one factor is B. In this cases we will allow more flexibility for the position of the factor B, that is, we reshuffle the tensor factor is such a way that B is at a convenient position, usually the first factor, and the other factors are kept in order. Coming back to the dynamical system, the multiplicative structure among the sets A2(π) for different representations can be phrased in terms of the elements sπ, too. For two irreducible representations π, ρ ∈ G consider the multiplication map mπ,ρ : A2(π) ⊗B A2(ρ) −→ A2(π ⊗ ρ) ⊆ A ⊗ L(Vπ) ⊗ L(Vρ), mπ,ρ(x ⊗ y) := x12 y13. (3) This is a module map for the right action of B ⊗ L(Vπ) ⊗ L(Vρ) on domain and codomain. Therefore, it is uniquely determined by the element mπ,ρ(sπ, sρ). For this element there is a unique element ω(π, ρ) ∈ B ⊗ L(Vπ ⊗ Vρ) with mπ,ρ(sπ, sρ) = sπ⊗ρ · ω(π, ρ). In fact, ω(π, ρ) is the isometry given by With this element the multiplication map can be written as ω(π, ρ) = s∗ π⊗ρ (sπ)12 (sρ)13. mπ,ρ(sπx ⊗ sρy) = sπ⊗ρ · ω(π, ρ) · (idπ ⊗γρ)(x) · (1π ⊗ y) (4) (5) 10 for all x ∈ B ⊗ L(Vπ), y ∈ B ⊗ L(Vρ). We want to classify and characterize weakly cleft C∗-dynamical systems in terms of the ∗-homomorphisms γπ and the isometries ω(π, ρ). For this purpose, we fix a group G and a C∗-algebra B and we consider pairs (γ, ω) consisting of two families γ = (γπ)π∈ G and ω = (cid:0)ω(π, ρ)(cid:1)π,ρ∈ G where 1. for each π ∈ G, we have a unital ∗-homomorphism γπ : B → B ⊗ L(Vπ) and 2. for each π, ρ ∈ G, we have an isometry ω(π, ρ) ∈ B ⊗ L(Vπ ⊗ Vρ). Definition 5.3. 1. The pair (γ, ω) is called a factor system for (G, B) if ω(1, 1) = 1B and the family satisfies ω(π, ρ) · (idπ ⊗γρ)(cid:0)γπ(b)(cid:1) = γπ⊗ρ(b) · ω(π, ρ), (cid:0)1π ⊗ ω(ρ, σ)(cid:1) · (idπ⊗ρ ⊗γσ)(cid:0)ω(π, ρ)(cid:1)∗ = ω(π, ρ ⊗ σ)∗ · ω(π ⊗ ρ, σ) for all π, ρ, σ ∈ G and b ∈ B. (6) (7) 2. Two factor systems (γ, ω) and (γ′, ω′) for (G, B) are called conjugated if there is a family v = (vπ)π∈ G of unitaries vπ ∈ B ⊗ L(Vπ) such that γ′ π = Ad[v∗ π] ◦ γπ, vπ⊗ρ · ω′(π, ρ) = ω(π, ρ) · (idπ ⊗γρ)(vπ) · vρ. Equations (6) and (7) are twisted versions of the equations for a coaction and for a 2-cocycle. For this reason, we refer to condition (6) as the coaction condition and to condition (7) as the cocycle condition. Immediate examples of factor systems are accordingly given a by coaction with a trivial cocycle or by a 2-cocycle with a trivial coaction (see Remark 5.4). Remark 5.4. 1. The normalization condition ω(1, 1) = 1B can always be achieved by passing to a normalized, conjugated system (in the straightforwardly generalized sense). Together with the coaction and cocycle condition the normalization implies γ1 = idB and ω(π, 1) = 1 = ω(1, π) for all π ∈ G. 2. It is worth mention that we may rephrase things in terms of the group C∗-algebra C ∗(G), more precisely its multiplier algebra M C ∗(G). The family (γπ)π∈ G may be equivalently written as a single unital ∗-homomorphism γ : B → B ⊗ M C ∗(G) and the family (ω(π, ρ))π,ρ∈ G as an isometry ω ∈ B ⊗ M C ∗(G × G). Then Equa- tions (6) and (7) can be casted in the form ω ·(cid:0)(id ⊗γ) ◦ γ(cid:1)(b) = (δ ◦ γ)(b) · ω, (1 ⊗ ω) · (id ⊗γ)(ω∗) = (id ⊗δ)(ω∗) · (δ ⊗ id)(ω), 11 where δ : M C ∗(G) → M C ∗(G × G) denotes the usual comultiplication. This formulation is used for extending the theory to compact quantum groups. If ω is unitary, the pair (γ, ω) is sometimes called a cocycle action (see e. g. [20]). The following statement summarizes that the construction from the beginning of this section indeed provides examples of factor systems. Lemma 5.5. Let (A, G, α) be a weakly cleft C∗-dynamical system and let B denote its fixed point algebra. Furthermore, let s = (sπ)π∈ G be a family of non-degenerate isometries sπ ∈ A2(π) with s1 = 1B. If γ = (γπ)π∈ G and ω = (cid:0)ω(π, ρ)(cid:1)π,ρ∈ G are the associated families of homomorphisms and unitaries given by Equations (2) and (4), respectively, then the pair (γ, ω) is a factor system. Proof. In order to show the coaction condition (6), we recall that we have sπγπ(b) = (b ⊗ 1)sπ for every b ∈ B (see Remark 5.2). Successively applying this relation and its starred version then yields Equation (6), i. e., in B ⊗ L(Vπ) ⊗ L(Vρ) we obtain ω(π, ρ) · γρ(cid:0)γπ(b)(cid:1) = s∗ π⊗ρsπsργρ(cid:0)γπ(b)(cid:1) = s∗ = γπ⊗ρ(b)s∗ π⊗ρsπsρ = γπ⊗ρ(b) · ω(π, ρ) π⊗ρbsπsρ for all π, ρ ∈ G and b ∈ B. In order to verify the cocycle condition (7) let us first consider its left hand side ω(ρ, σ) · γσ(cid:0)ω(π, ρ)(cid:1)∗ = s∗ ρ⊗σsρsσ · s∗ σ(s∗ π⊗ρsπsρ)∗sσ = s∗ ρ⊗σsρsσs∗ σs∗ ρs∗ πsπ⊗ρsσ The rightmost product s∗ degeneracy of sσ allows us to cancel the factor sσs∗ in L(Vπ) ⊗ A2(ρ), which allows us to cancel even further to obtain πsπ⊗ρsσ lies in L(Vπ ⊗ Vρ) ⊗ A2(σ). Therefore, the non- πsπ⊗ρ lies σ. Similarly, the element s∗ ρs∗ ω(ρ, σ) · γσ(cid:0)ω(π, ρ)(cid:1)∗ = s∗ ρ⊗σs∗ πsπ⊗ρsσ. For the right hand side of the cocycle condition, the non-degeneracy of sπ⊗ρ⊗σ likewise implies ω(π, ρ ⊗ σ)∗ · ω(π ⊗ ρ, σ) = s∗ ρ⊗σs∗ πsπ⊗ρ⊗σs∗ π⊗ρ⊗σsπ⊗ρsσ = s∗ ρ⊗σs∗ πsπ⊗ρsσ. Comparing with the simplification of the left side then yields the cocycle condition for all π, ρ, σ ∈ G. The next result states that the weakly cleft C∗-dynamical systems are uniquely deter- mined by their factor systems up to equivalence. 12 Theorem 5.6. Let (A, G, α) and (A′, G, α′) be weakly cleft C∗-dynamical systems with the same fixed point algebra B and let (γ, ω) and (γ′, ω′) be associated factor systems, respectively. Then the following statements are equivalent: (a) The dynamical systems (A, G, α) and (A′, G, α′) are equivalent. (b) The factor systems (γ, ω) and (γ′, ω′) are conjugated. Proof. As a distinction we add a prime to all notions referring to (A′, G, α′). 1. To prove that (a) implies (b) it suffice to show that for the same dynamical sys- tem (A, G, α) different choices of non-degenerate isometries sπ ∈ A2(π), π ∈ G, π, π ∈ G, be lead to conjugated factor systems. For this purpose let sπ and s′ two such choices and let us denote by (γ, ω) and (γ′, ω′) the associated factor sys- tems, respectively. Consider first a fixed representation π ∈ G. By Lemma 4.3 there are unique elements vπ, v′ πv′ π. Uniqueness implies vπv′ πvπ and, since sπ is a isometry, we also have π = v∗ πsπvπ = (s′ πvπ = v∗ v∗ π. For the ∗-homomorphisms of the factor systems we therefore obtain π) = 1. Hence vπ and v′ π = 1 = v′ π)∗(s′ π ∈ B ⊗ L(Vπ) with s′ π = sπvπ and sπ = s′ π are unitaries with v′ πs∗ γ′ π(b) = (s′ π)∗(b ⊗ 1)s′ π = v∗ πs∗ π(b ⊗ 1)sπvπ = v∗ πγπ(b)vπ, for every π ∈ G and b ∈ B. For the isometries of the factor systems we may use the non-degeneracy of sρ and the fact that vπsρvρ ∈ L(Vπ) ⊗ A2(ρ) to conclude for all π, ρ ∈ G: vπ⊗ρ ω′(π, ρ) = vπ⊗ρ(sπ⊗ρvπ⊗ρ)∗sπvπsρvρ = s∗ π⊗ρsπ(sρs∗ ρ)vπsρvρ = s∗ π⊗ρsπsργρ(vπ)vρ = ω(π, ρ) γρ(vπ) vρ. 2. For the converse implication, (b) ⇒ (a), let sπ ∈ A2(π) and s′ 2(π), π ∈ G, by non-degenerate isometries with associated factor systems (γ, ω) and (γ′, ω′), respectively. Furthermore, let vπ, π ∈ G, be a family of unitaries realizing the conjugation of the factor systems as in Definition 5.3. For every representation π of G the map π ∈ A′ ϕπ : A′ 2(π) → A2(π), s′ πx 7→ sπvπx for all x ∈ B ⊗ L(Vπ) is well-defined by Lemma 4.3. Moreover, it is straightforward to check that ϕπ is a unitary map between the two right Hilbert B-modules. Since L2(A) = Lπ∈ GA2(π) and likewise for A′, taking direct sums yields a unitary map Furthermore, the maps ϕπ, π ∈ G, intertwine with the multiplication maps, that is, for all π, ρ ∈ G we have V : L2(A′) → L2(A), V := Mπ∈ G ϕπ. mπ,ρ(cid:0)ϕπ(sπx) ⊗ ϕρ(sρy)(cid:1) = ϕπ⊗ρ(cid:0)m′ π,ρ(sπx ⊗ sπy)(cid:1) 13 for all x ∈ B ⊗ L(Vπ) and y ∈ B ⊗ L(Vρ). Together with Equation (1) this shows that the homomorphism Φ(x) := V xV ⋆ maps A′ ⊆ L(cid:0)L2(A′)(cid:1) into A ⊆ L(cid:0)L2(A)(cid:1) and hence may be restricted to an injective ∗-homomorphism Φ : A′ → A. Exchanging the role of sπ and s′ π shows that Φ is in fact a ∗-isomorphism. Obviously, we have ϕπ(xπ∗ g) = ϕπ(x)π∗ g for all x ∈ A2(π) and g ∈ G. It follows that V intertwines the G-action on L2(A) and L2(A′), that is, we have V Ug = U ′ gV for all g ∈ G. Consequently, Φ intertwines αg = Ad[Ug] on A and α′ g = Ad[U ′ g] on A′. The following lemma rephrases freeness in terms of the multiplication maps defined in (3). As a consequence we find that cleft dynamical systems are characterized factor systems where the isometries ω(π, ρ) are in fact unitaries. Lemma 5.7. A C∗-dynamical system (A, G, α) is free if and only if for all π ∈ G the multiplication map mπ,¯π : A2(π) ⊗B A2(¯π) → A2(π ⊗ ¯π), mπ,¯π(x ⊗ ¯y) := x12 · ¯y13 has dense range or, equivalently, is surjective. Proof. First we note that mπ,¯π is an isometry of correspondences over B and hence it is surjective if and only if it has dense range. Let us fix a finite-dimensional representation (π, V ) of G and denote by d its dimension. For sake of a convenient notation we fix a basis of V and write elements x ∈ A ⊗ L(V ) as matrices x = (xi,j)1≤i,j≤d with entries in A. Likewise we write elements of A ⊗ L( ¯V ) and A ⊗ L(V ⊗ ¯V ) as matrices with respect to the dual basis on ¯V and the product basis, respectively. A straightforward computation shows that the transpose map A2(¯π) → A2(π)∗, x = (xi,j)i,j 7→ xt := (xj,i)i,j is a linear bijection. Moreover, similar computations show that the map ϕ : A2(π ⊗ ¯π) → C(π) ⊗ L(V ), ϕ(x)(i,j),(k,ℓ) := x(i,k),(j,ℓ) is a linear bijection. Now consider the composition ψ := ϕ ◦ mπ,¯π, which takes the concrete form ψ(x ⊗ ¯y)(i,j),(k,ℓ) = xi,j ¯yk,ℓ for x ∈ A2(π) and ¯y ∈ A2(¯π). Since A2(π) and A2(¯π) are right L(V )- and L( ¯V )-modules, respectively, the range of ψ is a bimodule for 1C(π) ⊗ L(V ). It follows that the range of ψ is of the form J ⊗ L(V ) for some subspace J ⊆ C(π). Furthermore, for elements x, y ∈ A2(π) we may put ¯y := (y∗)t and find (idC(π) ⊗ Tr)(cid:0)ψ(x ⊗ ¯y)(cid:1) = (cid:18) d Xℓ=1 xi,ℓ y∗ ℓ,j(cid:19)i,j = xy∗, which shows that J = A2(π)A2(π)∗. We conclude that the multiplication map mπ,¯π is surjective if and only if ψ has full range if and only if C(π) = J = A2(π)A2(π)∗. 14 Theorem 5.8. For a weakly cleft C∗-dynamical system (A, G, α) the following state- ments are equivalent: (a) The dynamical system is cleft or, equivalently, free. (b) For some factor system (γ, ω) -- and hence for all factor systems -- all elements ω(π, ρ) for π, ρ ∈ G are unitary. Proof. By Lemma 4.5, for a cleft system we may choose unitary elements sπ in A2(π). π⊗ρsπsρ (π, ρ ∈ G) of the corresponding Then it follows that the elements ω(π, ρ) = s∗ factor system are unitary, too, which proves one implication. For the converse implica- tion we take advantage of Lemma 5.7. Indeed, suppose that we have non-degenerated isometries sπ ∈ A2(π) for each π ∈ G such that the elements ω(π, ρ) = s∗ π⊗ρsπsρ of the corresponding factor system are unitaries for all π, ρ ∈ G. Then A2(π ⊗ ρ) is given by A2(π ⊗ ρ) = sπ⊗ρ · B ⊗ L(Vπ ⊗ Vρ) = sπ⊗ρ ω(π, ρ) · B ⊗ L(Vπ ⊗ Vρ) = sπsρ · B ⊗ L(Vπ ⊗ Vρ) = sπ ·(cid:2)1B ⊗ L(Vπ) ⊗ 1ρ(cid:3) · sρ ·(cid:2)B ⊗ 1π ⊗ L(Vρ)(cid:3) Since sπ ·(cid:0)1B ⊗ L(Vπ)(cid:1)⊆ A2(π) and sρ ·(cid:0)B ⊗ L(Vρ)(cid:1) = A2(ρ), we conclude that mπ,ρ is surjective and hence the dynamical system is free. Corollary 5.9. Every weakly cleft C∗-dynamical system (A, G, α) with commutative or finite-dimensional fixed point algebra B is cleft. Proof. Under the hypothesis on B an isometry ω ∈ B ⊗ L(V ) for finite-dimensional V is automatically unitary. In Section 6 we will discuss whether the converse of Theorem 5.6 holds, that is, whether every factor system gives rise to a C∗-dynamical system. We postpone this problem for the moment and continue to investigate the relation between the form of the fac- tor system and the type of the dynamical system. First, let us characterize classical systems. Corollary 5.10. Let (A, G, α) be a cleft C∗-dynamical system with a commutative fixed point algebra B and let (γ, ω) be an arbitrary associated factor system. Then A is com- mutative if and only if γπ(b) = b ⊗ 1π, ω(ρ, π) = σ(cid:0)ω(π, ρ)(cid:1) for all π, ρ ∈ G and b ∈ B, where σ denotes the tensor flip of L(Vπ) ⊗ L(Vρ). and 15 Proof. First suppose that A is commutative. Let sπ ∈ A2(π), π ∈ G, be a family of non-degenerated isometries and denote by (γ, u) the associated factor system. Since A is commutative, every element b ⊗ 1 commutes with sπ in A ⊗ L(Vπ). It follows that γπ(b) = s∗ π(b ⊗ 1)sπ = b ⊗ 1 for all b ∈ B. Likewise the elements (sπ)12 and (sρ)13 commute in A ⊗ L(Vπ) ⊗ L(Vρ). It follows that σ(cid:0)ω(ρ, π)(cid:1) = σ(cid:0)s∗ ρ⊗πsρsπ(cid:1) = s∗ π⊗ρ sπ sρ = ω(π, ρ). Conversely, let sπ ∈ A2(π), π ∈ G, be a family of non-degenerate isometries and suppose the asserted condition on the corresponding factor system holds. Consider the family of multiplication maps mπ,ρ : A2(π) ⊗B A2(ρ) → A2(π ⊗ ρ), mπ,ρ(x ⊗ y) = x12 y13. Looking at Equation (5), we have for all x ∈ B ⊗ L(Vπ) and y ∈ B ⊗ L(Vρ): σ(cid:0)mρ,π(sρy ⊗ sπx)(cid:1) = σ(cid:0)sρ⊗π ω(ρ, π) y12 x13(cid:1) = sπ⊗ρ ω(π, ρ) x12 y13 = mπ,ρ(sπx ⊗ sρy), because x12 and y13 commute in B ⊗ L(Vπ) ⊗ L(Vρ). From the general construction of Section 3 we may then deduce that A is commutative. Remark 5.11. 1. Corollary 5.10 provides a classification of cleft topological princi- pal bundles. 2. Similar arguments as in the proof of Corollary 5.10 show the following more general statement for a weakly cleft system: The center of B is contained in the center of A if and only if for every factor system (γ, ω) the homomorphisms act trivially on the center of B, i. e., γπ(z) = z ⊗ 1 for every central element z ∈ B and π ∈ G. For a given group G and a given fixed point algebra B there clearly seems to be a trivial dynamical system, namely (B ⊗C(G), G, id ⊗r). From our discussions in Example 4.7 we immediately derive from Theorem 5.6 that for a weakly cleft dynamical system (A, G, α) the following statements are equivalent: (a) The dynamical system (A, G, α) is equivalent to (cid:0)B ⊗ C(G), G, id ⊗r(cid:1). (b) Every factor system of (A, G, α) is conjugated to the factor system (γ, ω) given by γπ(b) := b ⊗ 1 ω(π, ρ) := 1, π, ρ ∈ G, b ∈ B A next simple class of C∗-dynamical systems are those which are essentially ergodic actions, that is, the dynamical system (A, G, α) arises from an ergodic dynamical system (A0, G, α0) by tensoring with B. 16 Corollary 5.12. For a C∗-dynamical system (A, G, α) the following statements are equivalent: (a) (A, G, α) is isomorphic to (B ⊗ A0, G, idB ⊗α0) with an ergodic cleft C∗-dynamical system (A0, G, α0). (b) (A, G, α) is (weakly) cleft and admits a factor system of the form γπ(b) = b ⊗ 1, ω(π, ρ) ∈ 1B ⊗ L(Vπ ⊗ Vρ), π, ρ ∈ G, b ∈ B. Proof. For the system (B ⊗ A0, G, idB ⊗α0) the generalized isotypic component of π ∈ G is of the form A2(π) = B ⊗ A(0) 2 (π), where A(0) Since α0 is a cleft action, for each π ∈ G we find a unitary element s(0) obtain a unitary sπ ∈ A2(π) by putting sπ := 1B ⊗ s(0) system defined by Equations (2) and (4) is of the asserted form. 2 (π) ⊆ A0 ⊗ L(Vπ) denotes the generalized isotypic component of (A0, G, α0). 2 (π) and π . Then the corresponding factor π ∈ A(0) Conversely, suppose that a factor system (γ, ω) for (A, G, α) has the asserted form. Then ω(π, ρ) is a unitary for each π, ρ ∈ G and the cocycle condition (7) states that ω = (cid:0)ω(π, ρ)(cid:1)π,ρ forms a unitary 2-cocycle for G in the usual sense. By [22] this 2- cocycle provides an ergodic cleft C∗-dynamical system (A0, G, α0) whose factor system is given by ω together with the trivial homomorphisms. (This may also be derived from the construction presented in Section 6.) By the arguments in the first part of the proof the dynamical system (B ⊗ A0, G, idB ⊗α0) then has the factor system (γ, ω) and, by Theorem 5.6, it is isomorphic to (A, G, α). Example 5.13. Let B = L(H) for some Hilbert space H. We claim that up to iso- morphism every cleft dynamical system with fixed point algebra L(H) is of the form (L(H) ⊗ A0, G, id ⊗α0) with an ergodic cleft action (A0, G, α0). This can be proved directly by looking at minimal projections p ∈ L(H) and the corresponding restricted dy- namical system on the algebra pAp. Alternatively, we may consider an arbitrary factor system (γ, ω) for (A, G, α). Then for each representation π ∈ G the ∗-homomorphism γπ : L(H) → L(H) ⊗ L(Vπ) is necessarily of the form γπ(b) = v∗ π(b ⊗ 1)vπ, b ∈ B, for some unitary vπ ∈ L(H) ⊗ L(Vπ). Passing from (γ, ω) to a conjugated factor system, we may without loss of generality assume γπ(b) = b ⊗ 1 for all b ∈ B. Then the coaction condition states that for all π, ρ ∈ G the element ω(π, ρ) ∈ L(H ⊗ Vπ ⊗ Vρ) commutes with L(H) ⊗ 1π⊗ρ and hence lies in 1 ⊗ L(Vπ ⊗ Vρ). Finally, Corollary 5.12 proves the claim. 17 Corollary 5.14. For a C∗-dynamical system (A, G, α) and a unital C∗-subalgebra B0 ⊆ B the following statements are equivalent: (a) There is an α-invariant C∗-subalgebra B0 ⊆ A0 ⊆ A such that the restricted system (A0, G, αA0 ) is cleft with fixed point algebra B0. (b) There is a factor system (γ, ω) of A with unitary elements ω(π, ρ) ∈ B0 ⊗L(Vπ ⊗Vρ) and γπ(B0) ⊆ B0 ⊗ L(Vπ) for all π, ρ ∈ G. In this case A0 can be chosen to have the factor system (cid:0)γπB0, ω(π, ρ)(cid:1)π,ρ∈ G and A is generated by A0 and B. To have a concise statement we claim equivalence here but at this point we prove only one implication. The converse implication will be deduced later as Corollary 6.10. Proof of (a) ⇒ (b). For an α-invariant subalgebra A0 ⊆ A its generalized isotypic com- ponent of π ∈ G, denoted by A(0) 2 (π), is contained in the generalized isotypic com- If the action on A0 is cleft, then for each π ∈ G we find a ponent A2(π) of A. 2 (π). The elements sπ, π ∈ G, give rise to a factor system unitary element sπ ∈ A(0) for A0 and to a factor system for A. Both factor systems share the same unitaries π⊗ρsπsρ ∈ B0 ⊗ L(Vπ ⊗ Vρ) for all π, ρ ∈ G. The ∗-homomorphisms of the ω(π, ρ) = s∗ π(b ⊗ 1)sπ for π ∈ G, that is, they only two factor systems are both given by γπ(b) = s∗ differ by their domains B0 and B, respectively, and the corresponding codomains. Example 5.15. We would like to present a simple example that is not a tensor product with a free ergodic action. For this purpose consider the group G = SU2 and the commutative C∗-algebra C(X) for an arbitrary compact space X on which we fix a non-trivial continuous reflection h : X → X, h ◦ h = idX. The group SU2 has up to equivalence for every dimension precisely one irreducible representation. So we may identify G with the natural numbers starting with V0 denoting the trivial representation and V1 = C2 as the standard representation of SU2. As a factor system (γ, ω) we choose the coaction γn : C(X) → C(X) ⊗ L(Vn), n ∈ N, given by γn(b) := (b ⊗ 1 (b ◦ h) ⊗ 1 if n is odd if n is even, accompanied by the trivial 2-cocycle ω(n, m) := 1 for n, m ∈ N. The coaction condition for γ = (γn)n∈N is easily verified, e. g., using the Clebsch-Gordan formula for SU2. Sup- pose for a moment that there exists a C∗-dynamical system (A, G, α) with the factor system (γ, ω). (This will be proven in Section 6.) It follows from Corollary 5.14 that there is an α-invariant unital subalgebra A0 ⊆ A such that the restricted system is equiv- alent to the trivial system(cid:0)C(SU2), G, r(cid:1) and the algebra A is generated by A0 = C(SU2) and B = C(X). However, these two algebras satisfy non-trivial commutation relations. 18 Namely, we may collect all odd and even isotypic components and split A into its odd and even part, that is, we call x ∈ A even if α−1(x) = x and odd if α−1(x) = (−x). Recall that for each (π, V ) ∈ G the isotypic component A(π), as correspondence over B, is isomorphic to B ⊗ L(V ) with the usual right right multiplication and the left mul- tiplication b . x = γπ(b)x for x ∈ A(π) and b ∈ B. Since γπ maps into the center of B ⊗ L(V ), we obtain for π odd that b . x = ((b ◦ h) ⊗ 1)x = x . (b ◦ h) and for π even likewise b . x = x . b. Hence, for every x ∈ A0 = C(SU2) and b ∈ B = C(X) we have bx = xb bx = x(b ◦ h) if x is even, if x is odd. Theses relations determine the C∗-algebra A and the action of G on A uniquely, that is, A is the unique C∗-algebra generated by C(SU2) and C(X) subject to the above relations. The algebra A is a particular case of the a twisted tensor product (see [9, 10]). Remark 5.16. It is a well-known fact from homotopy theory that there exists, up to isomorphy, exactly one principal SU2-bundle over the 3-sphere S3, namely the trivial principal SU2-bundle S3 × SU2. It is used as a toy model for Chern-Simons theory devel- oped by Witten [23] to derive a 3-dimensional quantum field theory in order to give an intrinsic definition of the Jones polynomial and its generalizations dealing with knots in three dimensional space. Example 5.15 shows that the noncommutative setting provides more possibilities of constructing "noncommutative principal SU2-bundles" over S3. 6 Construction of Cleft Systems In the previous section we have seen that factor systems provide an invariant of weakly cleft C∗-dynamical systems. In the following we will see that they actually provide a full classification. That is, we will show that for every factor system (γ, ω) for (B, G) there actually is a weakly cleft C∗-dynamical system (A, G, α) with factor system (γ, ω). We would like to mention that the construction principle is not limited to weakly cleft actions and can be carried out more abstractly for general actions of compact quantum groups on C∗-algebras (see [11]). However, our restricted setting allows some simplifications. We will split the construction into three steps. In the first step we will show how a factor system gives rise to a multiplication and hence an algebra. The second step will concern the construction of an involution by exploiting the Hilbert module structure. In the last step we proceed along the lines sketched in Section 3 to finally construct the C∗-algebra and the dynamical system. 19 6.1 Associativity and Factor Systems Once and for all let us fix a compact group G and a unital C∗-algebra B, and let (γ, ω) be a pair consisting of a family γ = (γπ)π∈ G of unital ∗-homomorphisms γπ : B → B ⊗ L(Vπ) and a family of isometries ω(π, ρ) ∈ B ⊗ L(Vπ ⊗ Vρ) for all π, ρ ∈ G. For the moment we do not assume that (γ, ω) is a factor system. We extend both families naturally to arbitrary representations of G, that is, we decompose given representations (π, V ), (ρ, W ) into direct sums of irreducible components and define the corresponding maps γπ and elements ω(π, ρ) componentwise. Equivalently, γπ : B → B ⊗ L(V ) and ω(π, ρ) ∈ B ⊗ L(V ⊗ W ) are the unique ∗-homomorphisms and unitary elements, respectively, such that γπ(b)v = vγσ(b) and ω(π, ρ)(v ⊗ w) = (v ⊗ w)ω(σ, σ′) (8) for all isometric intertwiners v : Vσ → V and w : Vσ′ → W with irreducible representa- tions σ, σ′ ∈ G and all b ∈ B. For each representation (π, V ) of G we consider the correspondence B⊗L(V ) over B given by the usual right multiplication, the B-valued inner product hx, yiB := 1 (id ⊗ Tr)(x∗y), dπ and the left multiplication b . x := γπ(b)x for all b ∈ B and x, y ∈ B ⊗ L(V ). We write kxk2 := hx, xi1/2 for the corresponding norm of x ∈ B ⊗ L(V ) and for an irreducible representation σ ∈ G, we write Pσ,π for the map Pσ,π : B ⊗ L(V ) → B ⊗ L(Vσ) given by Pσ,π(x) := Pm kxvk, where v1, . . . , vm : Vσ → V is an orthonormal basis of intertwiners. The map does not depend on the choice of the orthonormal basis. Moreover, the map Pσ,π is adjointable with adjoint given by P ⋆ k for all z ∈ B ⊗ L(Vσ). k=1 v∗ B σ,π(z) = dπ k=1 vkyv∗ dσ Pm For each pair of representations (π, V ) and (ρ, W ) of G we define a linear map by mπ,ρ : [B ⊗ L(V )] ⊗B [B ⊗ L(W )] −→ B ⊗ L(V ⊗ W ) mπ,ρ(x ⊗ y) := ω(π, ρ) · (idπ ⊗γρ)(x) · (1π ⊗ y). We refer to mπ,ρ as multiplication map. For sake of a concise notation we will frequently drop the subindex and simply write m if the respective domains are not substantial and clearly determined by the context. Moreover, we amplify in the canonical way. With this simplifications the map mπ,ρ takes the form mπ,ρ(x ⊗ y) = ω(π, ρ) γρ(x)y for all x ∈ B ⊗ L(V ) and y ∈ B ⊗ L(W ). 20 Lemma 6.1. The map mπ,ρ is an adjointable isometry. Proof. Recall that for any C∗-algebra A, equipped with the natural right inner product hx, yiA := x∗y (x, y ∈ A), and for any isometry s ∈ A the map A → A, a 7→ sa is an adjointable isometry. In particular, it follows that the map z 7→ ω(π, ρ) z on B⊗L(V ⊗W ) is an adjointable isometry. Therefore, the assertion follows from the fact that the tensor product [B⊗L(V )]⊗B [B⊗L(W )] is canonically isometrically isomorphic to B⊗L(V ⊗W ) via the isomorphism x ⊗ y 7→ γρ(x)y for all x ∈ B ⊗ L(V ) and y ∈ B ⊗ L(W ). Lemma 6.2. Suppose (γ, ω) is a factor system. Then, for all representations π, ρ of G the map mπ,ρ is a B-bimodule map. Moreover, we have (idπ ⊗mρ,σ) ◦ (m⋆ π,ρ ⊗ idσ) = m⋆ π,ρ⊗σ ◦ mπ⊗ρ,σ. (9) for all representations π, ρ, σ of G. In particular, the family of maps mπ,ρ is associative, i. e., mπ,ρ⊗σ ◦ (idπ ⊗mρ,σ) = mπ⊗ρ,σ ◦ (mπ,ρ ⊗ idσ). Proof. The right module property of mπ,ρ is obviously satisfied even without assuming that (γ, ω) is a factor system. Furthermore, by the coaction condition of the factor system, Equation (6), we obtain m(cid:0)b . (x ⊗ y)(cid:1) = m(cid:0)γπ(b)x ⊗ y(cid:1) = ω(π, ρ) γρ(cid:0)γπ(b)x(cid:1) y (6)= γπ⊗ρ(b) ω(π, ρ) γρ(x)y = b . m(x ⊗ y) for all b ∈ B, x ∈ B ⊗ L(V ), and y ∈ B ⊗ L(W ). This shows that mπ,ρ is indeed a B-bimodule map. In particular, the left and right hand side of (9) are well-defined maps on the correspondence [B ⊗ L(Vπ ⊗ Vρ)] ⊗B [B ⊗ L(Vσ)]. In order to verify (9) for representations π, ρ, σ of G, we canonically identify the tensor product of the three correspondences B ⊗ L(Vπ), B ⊗ L(Vρ), and B ⊗ L(Vσ) over B with the correspondence B ⊗ L(Vπ ⊗ Vρ ⊗ Vσ) via the isometric isomorphism x ⊗ y ⊗ z 7→ γσ(cid:0)γρ(x)y(cid:1)z. Then the cocycle condition, Equation (7), implies that (idπ ⊗mρ,σ)(cid:0)(m⋆ π,ρ ⊗ idσ)(x)(cid:1) = ω(ρ, σ) γσ(cid:0)ω(π, ρ)(cid:1)∗ (7)= ω(π, ρ ⊗ σ)∗ ω(π ⊗ ρ, σ) x = m⋆ x, π,ρ⊗σ(cid:0)mπ⊗ρ,σ(x)(cid:1) for all x ∈ B ⊗ L(Vπ ⊗ Vρ ⊗ Vσ), which proves Equation (9). Multiplying (9) with (mπ,ρ ⊗ idσ) from the right and mπ,ρ⊗σ from the left yields mπ,ρ⊗σ(idπ ⊗mρ,σ) = mπ,ρ⊗σ ◦ m⋆ π,ρ⊗σ ◦ mπ⊗ρ,σ ◦ (mπ,ρ ⊗ idσ). Since the left hand side of this equation and the term S := mπ⊗ρ,σ ◦ (mπ,ρ ⊗ idσ) on the right hand side are both isometries, the orthogonal projection mπ,ρ⊗σ ◦ m⋆ π,ρ⊗σ acts trivially on the rage of S. Therefore, we may cancel mπ,ρ⊗σ ◦ m⋆ π,ρ⊗σ, which shows associativity. 21 Remark 6.3. The normalization condition ω(1, 1) = 1B of Definition 5.3 is equivalent to the fact that for the trivial representation the multiplication map m recovers the B-bimodule structure, that is, for a representation (π, V ) of G we have m1,π(b ⊗ x) = b . x = γπ(b)x, mπ,1(x ⊗ b) = x . b = x(b ⊗ 1π) for all elements x ∈ B ⊗ L(V ) and b ∈ B. In particular, in this case m1,1 coincides with the usual multiplication of B. In order to define an algebra, we consider the algebraic direct sum of the correspondences B ⊗ L(Vπ) for π ∈ G: B ⊗ L(Vπ). A := Mπ∈ G The left and right action of B are given componentwise and the B-valued inner product is hx, yiB = Pσ∈ Ghxσ, yσiB, where xσ and yσ denote the components of x and y, respec- tively. On A we define the product of x ∈ B ⊗ L(Vπ) and y ∈ B ⊗ L(Vρ) with π, ρ ∈ G by x • y := Xσ∈ G Pσ,π⊗ρ(cid:0)m(x ⊗ y)(cid:1). Bilinear extension then yields a bilinear map (x, y) 7→ x • y on A. For x ∈ B or y ∈ B this product coincides with the left or right action of B, respectively. Due to the intertwining relations (8) the family of maps mπ,ρ behaves nicely with respect to intertwiners. It is straightforward to check that for all representations π, ρ of G and every intertwiner v : Vσ → Vπ we have m(vxv∗ ⊗ y) = (v ⊗ 1ρ) m(x ⊗ y) (v ⊗ 1ρ)∗, m(v∗zv ⊗ y) = (v ⊗ 1ρ)∗m(z ⊗ y)(v ⊗ 1ρ) (10) for all x ∈ B ⊗ L(Vπ) and y ∈ B ⊗ L(Vρ), and similar equations hold for adjoining inter- twiners in the right tensor factor. Using this relations, the associativity of Lemma 6.2 and some algebra straightforwardly yields the following: Lemma 6.4. Suppose (γ, ω) is a factor system. Then the product • on A is associative, that is, (A, •) is an algebra. 6.2 Constructing the Involution Throughout the remainder of the section let us assume that (γ, ω) is a factor system. In order to define an involution on the algebra A, let consider a fixed π ∈ G. We put pπ := P ∗ π-th multiple of the orthogonal projection onto the 1,π⊗¯π(1B), that is, pπ is the d2 22 fixed point space of the representation π ⊗ ¯π. For every x ∈ B ⊗ L(Vπ) we define an element in B ⊗ L( ¯Vπ) by J(x) := (L⋆ x ◦ m⋆ π,¯π)(pπ), where we denote by Lx : B ⊗ L( ¯Vπ) → [B ⊗ L(Vπ)] ⊗B [B ⊗ L( ¯Vπ)] the adjointable map given by Lxy := x ⊗ y. Extending this antilinearly to all summands provides an antilinear map J : A → A. On the subalgebra B ⊆ A (i. e., for π = 1) we immediately find J(b) = b∗ for all b ∈ B. For the other summands, however, J does not coincide with the usual involution on B ⊗ L(Vπ) in general. Theorem 6.5. For all x, y, z ∈ A we have hJ(x) • y, ziB = hy, x • ziB. Proof. Let x ∈ B ⊗ L(Vπ), y ∈ B ⊗ L(Vρ), and z ∈ B ⊗ L(Vσ) with π, ρ, σ ∈ G. Since (γ, ω) is a factor system, the multiplication maps satisfy (9) of Lemma 6.2. Then, writing z := P ⋆ σ,¯π⊗ρ(z) for short, we obtain hJ(x) • y, ziB = hm¯π,ρ(cid:0)J(x) ⊗ y(cid:1), ziB = hm¯π,ρ(cid:0)L⋆ π,¯π ⊗ idρ)(pπ ⊗ y), ziB π,¯π(pπ) ⊗ y(cid:1), ziB xm⋆ π,¯π⊗ρmπ⊗¯π,ρ(pπ ⊗ y), ziB x(idπ ⊗m¯π,ρ)(m⋆ xm⋆ = hL⋆ (9)= hL⋆ = hmπ⊗¯π,ρ(pπ ⊗ y), mπ,¯π⊗ρ(x ⊗ z)iB. (8)= h(P ⋆ id,π⊗¯π ⊗ idρ)(y), (idπ ⊗P ⋆ σ,¯π⊗ρ)m(x ⊗ z)iB. We fix an isometric intertwiner w : C → Vπ⊗¯π and we choose an arbitrary orthonormal bases of intertwiners v1, . . . , vm : Vσ → ¯Vπ ⊗ Vρ. Then uk := q dπdρ (1π ⊗ vk)∗(w ⊗ 1ρ) for 1 ≤ k ≤ m form an orthonormal basis of intertwiners from ρ to π ⊗ σ. Orthonormality follows from the fact that for every T ∈ L( ¯Vπ) we have w∗(1π ⊗ T )w = 1 Tr(T ). Com- dπ pleteness holds because the dimension of the intertwiner space from ρ to π ⊗ σ coincides with the dimension of the intertwiner space from σ to ρ ⊗ ¯π. With this intertwiners we find Pρ,π⊗σ = (Pid,π⊗¯π ⊗ idρ)(idπ ⊗P ⋆ σ,¯π⊗ρ). Continuing the above computation, we obtain dσ hJ(x) • y, ziB = hy, Pρ,π⊗σm(x ⊗ z)iB = hy, x • ziB. The assertion for arbitrary x, y, z ∈ A then follows by extending (anti-)linearly. Corollary 6.6. Let P0 : A → B denote the orthogonal projection onto the direct sum- mand B ⊆ A corresponding to the trivial representation. Then for all x, y ∈ A we have hx, yiB = P0(cid:0)J(x) • y(cid:1). Proof. The element 1B is a unit for the multiplication of A and fixed by the orthogonal projection P0. Hence by Theorem 6.5 we obtain hx, yiB = h1B, J(x) • yiB = (cid:10)1B, P0(cid:0)J(x) • y(cid:1)(cid:11)B = P0(cid:0)J(x) • y(cid:1). 23 Remark 6.7. We would like to mention that by Corollary 6.6 the element J(x) can be equivalently be characterized as the unique element of A satisfying hJ(x), yiB = P0(x • y) for all y ∈ A. Corollary 6.8. A is an involutive algebra, i. e., for all x, y ∈ A we have J(cid:0)J(x)(cid:1) = x, J(x • y) = J(y) • J(x). Proof. By applying Theorem 6.5 twice we get and (cid:10)J(cid:0)J(x)(cid:1), z(cid:11)B = h1B, J(x) • ziB = hx, ziB hJ(x • y), ziB = h1B, x • y • ziB = hJ(x), y • ziB = hJ(y) • J(x), ziB for all z ∈ A. Since the inner product separates points, this yields the assertion. 6.3 Construction of Free Actions Having the algebra A and the involution on A in hands, the construction of the C∗-alge- bra and the cleft action follows the outline presented in Section 3. We consider A as a right pre-Hilbert B-module. The inner product is positive definite but, unless G is finite, A is not closed with respect to the induced norm kxk2 := khx, xiBk1/2 x ∈ A. op = kP0(cid:0)J(x) • x(cid:1)k1/2 op , We denote by ¯A the completion with respect to this norm. Equivalently, the correspon- dence ¯A over B is the direct sum of the previously discussed correspondences B ⊗ L(Vπ) with π ∈ G. Theorem 6.5 allows us to extend each left multiplication by x ∈ A to an adjointable map on the completion ¯A, which we again denote by λ[x] : ¯A → ¯A. Then the map λ : A → L( ¯A), x 7→ λ[x] is an representation of the ∗-algebra A by adjointable operators on ¯A. The vector 1B ∈ ¯A is clearly cyclic and separating for the subalgebra λ(A). In particular, λ is faithful and hence isometric on B ⊆ A. We denote by A the C∗-algebra generated by the range of λ. To simplify the notation we identify the algebra (A, •) with the subalgebra λ(A) ⊆ A. For sake of clarity, we extend the notation for the ∗-algebra A to the C∗-algebra L( ¯A), that is, for elements x, y ∈ L( ¯A) we write x • y for their product and for x ∈ A we write J(x) for its adjoint. Since for x ∈ A we have kλ[x]k2 ≥ kλ[x]1Bk2 2, we may regard the C∗-algebra A as a subset of ¯A, so that A ⊆ A ⊆ ¯A. 2 = kxk2 24 On each direct summand B ⊗ L(Vρ) ⊆ ¯A, π ∈ G, we have a continuous action of G by the usual right multiplication Ug(x) := x (1B ⊗ π∗ g) for g ∈ G and x ∈ B ⊗ L(Vπ). For each g ∈ G the map Ug is unitary on B ⊗ L(Vρ) with respect to the B-valued inner product. Taking direct sums and continuous extension, we obtain a strongly continuous unitary representation g 7→ Ug ∈ L( ¯A) of the group G on the correspondence ¯A. We denote by α = (αg)g∈G the associated automorphism group on L( ¯A), i. e., we put αg(x) := Ug • x • U ⋆ g , x ∈ L( ¯A). For each element x ∈ A ⊆ A we have αg(x) = Ug(x), that is, on the algebra A ⊆ ¯A the actions α = (αg)g∈G and U = (Ug)g∈G coincide. Since A is dense in A, it follows that g 7→ αg(x) is continuous for every x ∈ A. Summarizing, we have constructed a C∗-dynamical system (A, G, α). Theorem 6.9. The C∗-dynamical system (A, G, α) is weakly cleft with fixed point algebra B and factor system (γ, ω). Proof. In the first part of the proof we will show that for the action U = (Ug)g∈G on ¯A the isotypic component of π ∈ G is given by the direct summand B ⊗ L( ¯Vπ). This verifies in particular that B is the fixed point space. By Theorem 6.5 it follows that the canonical B-valued inner product on A coincides with the inner product on the larger space ¯A, that is, ZG αg(cid:0)J(x) • y(cid:1) dg = P0(cid:0)J(x) • y) 6.5 = hx, yiB. for all x, y ∈ A. Using Lemma 4.3 we conclude that the dynamical system (A, G, α) is weakly cleft. In the second part of the proof we confirm that (γ, ω) is indeed a factor system of (A, G, α). 1. We consider the action U = (Ug)g∈G on the correspondence ¯A over B. Obviously, all elements of B ⊆ ¯A are fixed by the action. For an element x ∈ B ⊗ L(Vπ) in a di- g dg vanishes because π is irreducible. Taking linear combinations and continuous extensions rect summand of ¯A with id 6= π ∈ G the integral RG Ug(x) dg = RG xπ∗ then shows that RG Ug(x) dg = P0(x) for all x ∈ ¯A. Adapting the arguments, we obtain for every π ∈ G and x ∈ ¯A dπZG Tr(π∗ g)Ug(x) dg = P¯π(x), where P¯π : ¯A → B ⊗ L( ¯Vπ) denotes the orthogonal projection onto the direct summand B ⊗ L( ¯Vπ) ⊆ ¯A. We conclude that B ⊗ L( ¯Vπ) is the isotypic component of π in ¯A and hence also in A ⊆ ¯A. 25 2. It remains to verify that (γ, ω) is a factor system of (A, G, α). According to the first part of the proof the isotypic component of ¯π ∈ G in A is given by B ⊗ L(Vπ). As a right B-module B ⊗ L(Vπ) is generated by the element 1B ⊗ 1π of norm 1. The natural isomorphism of the isotypic and the generalized isotypic component together with Lemma 4.3 therefore provides us with the non-degenerate isometry sπ ∈ A2(π) ⊆ A ⊗ L(Vπ) given by sπ = dπZG 1B ⊗ π∗ g ⊗ πg dg = (1B ⊗ F ), where F ∈ L(V ) ⊗ L(V ) denotes the tensor flip. This family s = (sπ)π∈ G of isometries then gives rise to a factor system (γ, ω) by Equations (2) and (4). For each π ∈ G the homomorphism γπ is uniquely determined by the left action of B on sπ. For convenience we also write • for the multiplication on all A ⊗ L(Vπ), π ∈ G. Then the left action of B on sπ is given by (b ⊗ 1π) • sπ = dπZG(cid:0)b • (1B ⊗ π∗ g)(cid:1) ⊗ πg = (cid:0)γπ(b) ⊗ 1π(cid:1)(1B ⊗ F ) = sπ(1B ⊗ F )(cid:0)γπ(b) ⊗ 1π(cid:1)(1B ⊗ F ) g)(cid:1) ⊗ πg = dπZG(cid:0)γπ(b)(1B ⊗ π∗ = sπ • γπ(b) for every b ∈ B. Consequently, γπ = γπ. With some algebra involving the multipli- cation •, which we leave to the reader, it is then straightforward to show that we also have sπ • sρ = sπ⊗ρ • ω(π, ρ) for all π, ρ ∈ G. This proves ω(π, ρ) = s∗ π⊗ρ • sπ • sρ = ω(π, ρ). Summarizing, we have shown that (γ, ω) is indeed the factor system associated with the isometries sπ, π ∈ G, which finishes the proof. This concrete representation theorem of weakly cleft actions finally allows us as a corol- lary to complete the proof of Corollary 5.14, which we asserted in Section 5. For conve- nience we repeat the relevant statement as a reminder in a slightly more general form. Corollary 6.10. Let (A, G, α) be a C∗-dynamical system and B0 ⊆ B a unital C∗-sub- algebra of fixed points. Furthermore, let (γ, ω) be a factor system such that γπ(B0) ⊆ B0 ⊗ L(Vπ), ω(π, ρ) ∈ B0 ⊗ L(Vπ ⊗ Vρ), for all π, ρ ∈ G. Then there is an α-invariant subalgebra A0 ⊆ A such that the restricted dynamical system (A0, G, αA0 ) is weakly cleft and A is generated by A0 and B. Proof. By Theorem 6.9 and Theorem 5.6 we may assume that A admits an α-invariant dense subalgebra of the form A = Lπ∈ G B ⊗ L(Vπ) with the multiplication x • y = Xσ∈ G Pσ(cid:0)ω(π, ρ) γρ(x) y(cid:1) 26 and the action αg(x) = x(1B ⊗ π∗ π, ρ ∈ G. By the hypotheses on (γ, ω) the subset g) for all x ∈ B ⊗ L(Vπ), y ∈ B ⊗ L(Vρ), g ∈ G, and A0 := Mπ∈ G B0 ⊗ L(Vπ) is an α-invariant ∗-subalgebra of A (cf. also Remark 6.7). Let A0 ⊆ A be the closure of A0 with respect to the operator norm. By Theorem 6.9, the dynamical system restricted to A0 is weakly cleft with the restriction of (γ, ω) to B0 as factor system. The space A is linearly generated by all products x • b = (b ⊗ x) with x ∈ 1B ⊗ L(Vπ), π ∈ G, and b ∈ B. Since A0 contains all elements x ∈ 1B ⊗ L(Vπ), π ∈ G, the algebra A is generated by A0 and B. Taking closures, A is generated by A0 and B. Acknowledgment We would like to acknowledge the Center of Excellence in Analysis and Dynamics Re- search (Academy of Finland, decision no. 271983 and no. 1138810) for supporting this research. References [1] B. Blackadar. Operator Algebras, volume 112, Operator Algebras and Non- Communtative Geometry of Encyclopaedia Math. Sci. Springer, 2006. [2] T. Brzeziński and P. M. Hajac. The Chern-Galois character. C. R. Math. Acad. Sci. Paris, 338(2):113 -- 116, 2004. [3] A. A. Davydov. Galois algebras and monoidal functors between categories of repre- sentations of finite groups. Journal of Algebra, 244(1):273 -- 301, 2001. [4] K. De Commer and M. Yamashita. A construction of finite index C∗-algebra in- clusions from free actions of compact quantum groups. Publ. Res. Inst. Math. Sci., 49(4):709 -- 735, 2013. [5] Y. Doi. Equivalent crossed products for a Hopf algebra. Comm. Algebra, 17(12):3053 -- 3085, 1989. [6] Y. Doi and M. Takeuchi. Cleft comodule algebras for a bialgebra. Comm. Algebra, 14(5):801 -- 817, 1986. 27 [7] S. Echterhoff, R. Nest, and H. Oyono-Oyono. Principal non-commutative torus bundles. Proc. Lond. Math. Soc., 99(1):1 -- 31, 2009. [8] P. Green. C∗-algebras of transformation groups with smooth orbit space. Pacific J. Math., 72(1):71 -- 97, 1977. [9] G. G. Kasparov. The operator K-functor and extensions of C∗-algebras. Mathemat- ics of the USSR-Izvestiya, 16(3), 1981. [10] R. Meyer, S. Roy, and S. L. Woronowicz. Quantum group-twisted tensor products of C∗-algebras. Internat. J. Math., 25(02):1450019, 2014. [11] S. Neshveyev. Duality theory for nonergodic actions. Münster J. Math., 2013. to appear. [12] O. Pfante. A Chern-Simons action for noncommutative spaces in general with the example SUq(2). J. Noncommut. Geom., 8(3):611 -- 654, 2014. [13] C. N. Phillips. Equivariant K-Theory and Freeness of Group Actions on C∗-Algebras, volume 1274 of Lecture Notes in Math. Springer, 1987. [14] J. Phillips and I. Raeburn. Twisted crossed products by coactions. J. Aust. Math. Soc., 56(3):320 -- 344, June 1994. [15] N. C. Phillips. Freeness of actions of finite groups on C∗-algebras. In Operator structures and dynamical systems, volume 503 of Contemp. Math., pages 217 -- 257. Amer. Math. Soc., 2009. [16] M. A. Rieffel. Proper actions of groups on C∗-algebras. In H. Araki and R. V. Kadison, editors, Mappings of Operator Algebras, volume 84 of Progr. Math, pages 141 -- 182. Birkhäuser, 1991. [17] P. Schauenburg. Hopf-Galois and bi-Galois extensions. In G. Janelidze, B. Pareigis, and W. Tholen, editors, Galois Theory, Hopf Algebras, and Semiabelian Categories, volume 43 of Fields Inst. Commun., pages 469 -- 515. Amer. Math. Soc., 2004. [18] K. Schwieger and S. Wagner. Part I, Free actions of compact Abelian groups on C∗-algebras. Adv. Math., 317:224 -- 266, 2017. [19] N. Steenrod. The Topology of Fibre Bundles. Princeton Univ. Press, 1st edition, 1999. [20] S. Vaes and L. Vainerman. Extensions of locally compact quantum groups and the bicrossed product construction. Adv. Math., 175(1):1 -- 101, 2003. 28 [21] A. Wassermann. Ergodic actions of compact groups an operator algebras. II. Clas- sification of full multiplicity ergodic actions. Canad. J. Math., 40(6):1482 -- 1527, 1988. [22] A. Wassermann. Ergodic actions of compact groups on operator algebras. I. General theory. Ann. of Math. (2), 130(2):273 -- 319, 1989. [23] E. Witten. Quantum field theory and the Jones polynomial. Comm. Math. Phys, 121(3):351 -- 399, 1989. 29
1210.2964	1	1210	2012-10-10T16:03:26	Tensorial Function Theory: From Berezin transforms to Taylor's Taylor series and back	[ "math.OA", "math.CV", "math.FA", "math.RA" ]	Let $H^{\infty}(E)$ be the Hardy algebra of a $W^{}$-correspondence $E$ over a $W^{}$-algebra $M$. Then the ultraweakly continuous completely contractive representations of $H^{\infty}(E)$ are parametrized by certain sets $\mathcal{AC}(\sigma)$ indexed by $NRep(M)$ - the normal *-representations $\sigma$ of $M$. Each set $\mathcal{AC}(\sigma)$ has analytic structure, and each element $F\in H^{\infty}(E)$ gives rise to an analytic operator-valued function $\hat{F}_{\sigma}$ on $\mathcal{AC}(\sigma)$ that we call the $\sigma$-Berezin transform of $F$. The sets ${\mathcal{AC}(\sigma)}_{\sigma\in\Sigma}$ and the family of functions ${\hat{F}_{\sigma}}_{\sigma\in\Sigma}$ exhibit "matricial structure" that was introduced by Joeseph Taylor in his work on noncommutative spectral theory in the early 1970s. Such structure has been exploited more recently in other areas of free analysis and in the theory of linear matrix inequalities. Our objective here is to determine the extent to which the matricial structure characterizes the Berezin transforms.	math.OA	math	TENSORIAL FUNCTION THEORY: FROM BEREZIN TRANSFORMS TO TAYLOR'S TAYLOR SERIES AND BACK PAUL S. MUHLY AND BARUCH SOLEL Abstract. Let H ∞(E) be the Hardy algebra of a W ∗-correspond- ence E over a W ∗-algebra M . Then the ultraweakly continuous completely contractive representations of H ∞(E) are parametrized by certain sets AC(σ) indexed by N Rep(M ) - the normal ∗-repres- entations σ of M . Each set AC(σ) has analytic structure, and each element F ∈ H ∞(E) gives rise to an analytic operator-valued function bFσ on AC(σ) that we call the σ-Berezin transform of F . The sets {AC(σ)}σ∈Σ and the family of functions {bFσ}σ∈Σ exhibit "matricial structure" that was introduced by Joeseph Taylor in his work on noncommutative spectral theory in the early 1970s. Such structure has been exploited more recently in other areas of free analysis and in the theory of linear matrix inequalities. Our objective here is to determine the extent to which the matricial structure characterizes the Berezin transforms. 1. Introduction Our purpose in this paper is to explore relations among three subjects: (1) the theory of Berezin transforms that arise from the representation theory of tensor algebras and Hardy algebras of W ∗-correspondences, (2) infinite dimensional holomorphy, and (3) the theory of free holo- morphic functions initiated by Joseph Taylor in [Tay72, Tay73]. In this introduction we indicate the connections we have in mind and provide a bit of context. Details and a fuller account, including relevant defini- tions of terms left undefined here, will be given in subsequent sections. Suppose first that M is a W ∗-algebra and that E is a W ∗-correspond- ence over M.1 With these ingredients one may build two operator algebras, T+(E) and H ∞(E), that are generated by a copy of M and The research of both authors was supported in part by a US-Israel Binational Science Foundation grant. The second author was also supported by the Technion V.P.R. Fund. 1We shall assume throughout that M has a separable predual and that E is countably generated. 1 TENSORIAL FUNCTION THEORY 2 the creation operators {Tξ ξ ∈ E} acting on the full Fock space F (E) = M ⊕ E ⊕ E⊗2 ⊕ E⊗3 ⊕ · · · : T+(E), the tensor algebra of E, is the norm closed algebra generated by these objects and H ∞(E), the Hardy algebra of E, is its ultraweak closure. One may think of T+(E) as a generalization of the disc algebra A(D), and H ∞(E) may be viewed as a generalization of the classical Hardy space, H ∞(D), consisting of the bounded analytic functions on D. Indeed, when M = C = E, then T+(E) is naturally completely isometrically isomorphic to A(D) and H ∞(E) is naturally completely isometrically isomorphic and weak- ∗ homeomorphic to H ∞(D). Another important example to keep in mind is that which arises when M = C and E = Cd, d ≥ 2. In this case, T+(E) is naturally completely isometrically isomorphic to Gelu Popescu's noncommutative disc algebra Ad [Pop96] and H ∞(E) has come to be called the algebra of noncommutative analytic Toeplitz operators. The terminology is due to Davidson and Pitts [DP98], and much of the initial theory of these algebras is due to them. Other interesting finite dimensional settings can be constructed from graphs or quivers. (See [Muh97, MS99, KK06] for examples.) It is worthwhile to emphasize that even if one were interested only in these finite dimensional examples, it is useful to work with the gen- eral theory. One reason is that the general theory is invariant under Morita equivalence [MS00, MS11a] and, among many things, Morita theory allows one to study every T+(E) and H ∞(E) in terms of an- alytic crossed products - generalizations of twisted polynomial rings - that have played such a prominent role in the theory of non-self-adjoint operator algebras. The point is that although T+(E) and H ∞(E) look very much like algebras of functions of several (noncommutative) vari- ables, as we shall see, they behave much more like algebras of functions of one variable than one might expect. In [MS98, Theorem 2.9], we showed that every completely contractive representation of T+(E) is given by a pair, (σ, z), where σ is a normal representation of M on a Hilbert space Hσ and z : E ⊗σ Hσ → Hσ is an operator of norm at most 1 that intertwines σE ◦ ϕ and σ, where σE is the representation of L(E) that is induced by σ in the sense of Rieffel [Rie74a] and where ϕ denotes the left action of M on E. We denote the representation associated to (σ, z) by σ × z. (In general, if A is a not-necessarily-self-adjoint algebra and if π and ρ are two representations of A by bounded operators on Hilbert spaces Hπ and Hρ, respectively, then we shall write I(π, σ) for the collection of all operators C from Hπ to Hρ such that Cπ(a) = ρ(a)C for all a ∈ A TENSORIAL FUNCTION THEORY 3 and we call I(π, ρ) the intertwiner space or the space of intertwiners from π to ρ.) For each normal representation σ of M, we endow I(σE ◦ ϕ, σ) with the operator norm and we write D(0, 1, σ) for the open unit ball in I(σE ◦ ϕ, σ). Then, when σ is fixed, each F ∈ T+(E) gives rise to a B(Hσ)-valued function bFσ defined on the closed unit ball D(0, 1, σ) by the formula (1.1) bFσ(z) := σ × z(F ). all normal representations of M on separable Hilbert space, NRep(M), the σ-Berezin transform of F .2 The collection of all σ-Berezin trans- a continuous function on D(0, 1, σ) with values in B(Hσ), where both Because of formulas that we derived in [MS09, Theorem 13], we call bFσ forms, {bFσ}σ∈N Rep(M ), obtained by letting σ range over the collection of is called the Berezin transform of F and will be written simply as bF . It is easy to see that for F ∈ T+(E), each σ-Berezin transform bFσ is spaces are given the norm operator topology. Further, bFσ is holomor- phic in the sense of Frechet [HP74, 112 and 778] when restricted to D(0, 1, σ). A similar sort of representation exists for elements F ∈ H ∞(E). How- ever, for these F , bFσ makes good sense only on the set of zs in D(0, 1, σ) such that σ × z extends from T+(E) to an ultraweakly continuous rep- resentation of H ∞(E) in B(Hσ). We denote the set of such points by AC(σ) and for reasons spelled out in [MS11b] we call them the abso- lutely continuous points in D(0, 1, σ). It turns out that D(0, 1, σ) ⊆ AC(σ) [MS04, Corollary 2.14]. Thus, for F ∈ H ∞(E), bFσ makes sense It is, in fact, bounded and holomorphic The Frechet power series of bFσ is easy to calculate and has a remarkably simple expression in terms of the tensorial "Fourier series" in which F may be expressed using the gauge automorphism group built from the number operator on the full Fock space F (E). We call power series with this special form tensorial power series (Definition 2.9). It is natural to inquire about the structure of such power series, in general, and one soon sees that much of standard elementary theory of complex as a function on D(0, 1, σ). with respect to the norm topologies on D(0, 1, σ) and B(Hσ). 2We are indebted to Lew Coburn for calling our attention to the connection be- tween our formulas and the classical Berezin transform associated with the Hardy space on the open unit disc in the complex plane. We note, too, that our termi- nology agrees with that of Gelu Popescu [Pop08] in those settings where his theory and ours overlap. TENSORIAL FUNCTION THEORY 4 analysis on the open unit disc can be recapitulated in the more general setting we are describing. When M = C and E = Cd, this has been done already by Popescu in [Pop06]. is an algebra under pointwise multiplication that may be identified naturally with a quotient of H ∞(E) by an ultraweakly closed ideal For a given σ, we write B(σ) for {bFσ F ∈ H ∞(E)}. Then B(σ) and the map F → bFσ is a complete quotient map. If σ is a faithful normal representation of M of infinite multiplicity, then the map is a completely isometric isomorphism from H ∞(E) onto B(σ) [MS08, Lemma 3.8]. In general, however, B(σ) is a proper quotient of H ∞(E). Indeed, when M = C, and E = Cd, with d ≥ 2, then each normal representation of M is, of course, (unitarily equivalent to) a multiple of the representation σ1 of C on the one-dimensional Hilbert space, C. The σ1-Berezin transforms {bFσ1 F ∈ H ∞(E)} are the multipliers of the Drury-Arveson space and so form a commutative algebra. On the other hand, B(∞σ1) is isomorphic to the algebra of noncommutative analytic Toeplitz operators in the fashion just described. For 2 ≤ n < ∞, B(nσ1) is a completion of the algebra of d generic n × n matrices. So, as soon as n ≥ 2, it is noncommutative. But also, by virtue of the polynomial identities these algebras satisfy, it is easy to see that when n 6= m, B(nσ1) ≇ B(mσ1), and that no finite dimensional representation of C yields a faithful representation of either H ∞(E) or of T+(E) in terms of Berezin transforms. Thus there arise very natural questions: How much is lost when forming the σ-Berezin transform of an element in H ∞(E)? How might one reconstruct an F from its finite dimensional σ-Berezin transforms? What extra information is required? While we are still far from giving definitive answers to these questions, we believe that what we accomplish here is a helpful start. There is an important feature of the discs D(0, 1, σ) that plays a central roll in our theory: For any two normal representations of M, σ and τ , (1.2) D(0, 1, σ) ⊕ D(0, 1, τ ) ⊆ D(0, 1, σ ⊕ τ ). z12 The meaning of this inclusion is easy to understand when one realizes that I((σ ⊕ τ )E ◦ ϕ, σ ⊕ τ ) may be viewed as a set of operator matrices (cid:20)z11 z22(cid:21) acting as operators from HσE ◦ϕ ⊕ Hτ E◦ϕ to Hσ ⊕ Hτ , where z21 z11 ∈ I(σE ◦ ϕ, σ), z12 ∈ I(τ E ◦ ϕ, σ), z21 ∈ I(σE ◦ ϕ, τ ), and z22 ∈ I(τ E ◦ ϕ, τ ). So D(0, 1, σ) ⊕ D(0, 1, τ ) is just the collection of those matrices (cid:20)z11 z22(cid:21) ∈ D(0, 1, σ ⊕ τ ) in which the off-diagonal entries z12 z21 vanish. TENSORIAL FUNCTION THEORY 5 Definition 1.1. A family of sets {U(σ)}σ∈N Rep(M ), with U(σ) ⊆ I(σE◦ ϕ, σ), satisfying U(σ)⊕ U(τ ) ⊆ U(σ ⊕ τ ) is called a matricial family of sets. Matricial families, in particular, families of discs {D(0, 1, σ)}σ∈N Rep(M ), enjoy properties that are very similar to the properties of the domains that Taylor first considered in [Tay72, Section 6], when he introduced a notion of localization for free algebras. They are very closely related to the fully matricial sets of Dan Voiculescu [Voi04, Voi10], which are matricial sets in our terminology but also satisfy additional conditions which we do not use here. In contexts when E is finite dimensional, matricial sets are essentially the noncommutative sets that Bill Helton, Igor Klep, Scott McCullough and others study in the setting of lin- ear matrix inequalities [HKM11b, HKM11a, HKMS09]. They are also closely connected to the noncommutative sets in the work of Dmitry Kaliuzhnyi-Verbovetskyi and Victor Vinnikov devoted to noncommu- tative function theory [KVV09, KV10, KVV]. There is another property that the discs also enjoy, viz. for any con- traction t ∈ I(τ, σ), the inclusion, (1.3) tD(0, 1, τ )(IE ⊗ t∗) ⊆ D(0, 1, σ), holds. This shows that the discs are matricially convex in the sense of operator space theory (see [ER00]). While matricial convexity does not play a role in our immediate considerations, it already has proved useful elsewhere. satisfies the equation The Berezin transform, bF = {bFσ}σ∈N Rep(M ), of an element F ∈ H ∞(E) (1.4) bFσ⊕τ (z⊕ w) = bFσ(z)⊕bFτ (w), z⊕ w ∈ D(0, 1, σ)⊕ D(0, 1, τ ). This, too, is a critical feature of the functions in Taylor's theory and in the other places just cited. In fact, the Berezin transforms have an additional property that we will use repeatedly: Definition 1.2. Suppose {U(σ)}σ∈N Rep(M ) is a matricial family of sets and suppose that for each σ ∈ NRep(M), fσ : U(σ) → B(Hσ) is a function. We say that f := {fσ}σ∈N Rep(M ) is a matricial family of functions in case (1.5) for every z ∈ U(σ), every w ∈ U(τ ) and every C ∈ I(σ, τ ) such that (1.6) Cfσ(z) = fτ (w)C Cz = w(IE ⊗ C). TENSORIAL FUNCTION THEORY 6 When the family {U(σ)}σ∈N Rep(M ) is {D(0, 1, σ)}σ∈N Rep(M ), and f = {fσ}σ∈N Rep(M ) is a Berezin transform, then it is easy to see that the assumptions on an operator C : Hσ → Hτ that C ∈ I(σ, τ ) and satisfies equation (1.6) express the fact that C lies in I(σ × z, τ × w). But then, equation (1.5) is immediate. It is simply a manifestation of the structure of the commutant of the representation (σ ⊕ τ ) × (z⊕ w). In this setting also, the defining hypothesis for a matricial family can be written simply as I(σ × z, τ × w) ⊆ I(fσ(z), fτ (w)), (1.7) for all σ, τ ∈ NRep(M), z ∈ AC(σ), and w ∈ AC(τ ). Consequently, we sometimes say that a matricial family respects intertwiners. Observe that if a family respects intertwiners, then it automatically satisfies equations like (1.4). What is surprising is the following converse - a nonlinear double com- mutant theorem of sorts - which extends the double commutant the- orem for induced representations of Hardy algebras [MS04, Corollary 3.10]. We shall prove a more refined statement in Theorem 4.4. Theorem 1.3. If f = {fσ}σ∈N Rep(M ) is a matricial family of functions, with fσ defined on AC(σ) and mapping to B(Hσ), then there is an F ∈ H ∞(E) such that f is the Berezin transform of F , i.e., fσ = bFσ for every σ. Three features of Theorem 1.3 deserve comment: (1) the domain of each fσ is AC(σ), (2) all the representations in NRep(M) are used, and (3) no hypotheses on the nature of the functions fσ are made. They are not assumed to be analytic or continuous; they are not even assumed to be bounded. If we relax (1) and assume only that the functions are defined on D(0, 1, σ), then the theorem breaks down: In Example 4.5 we will exhibit an unbounded family f = {fσ}σ∈N Rep(M ), where fσ is defined only on D(0, 1, σ), that satisfies (1.7). However, we shall also prove that if the fσ : D(0, 1, σ) → B(Hσ) are bounded uniformly in σ ∈ NRep(M) and if they respect intertwiners, then f is a Berezin transform (Theorem 4.8). With regard to (2), it is helpful to reflect on the fact that NRep(M) is really a W ∗-category in the sense of [GLR85]. In fact, it is essentially Riefel's category Normod − M [Rie74b]. We do not need much of the theory of such categories to achieve our goals here, but that theory has guided our thinking. The objects of NRep(M) are the normal repre- sentations of M on separable Hilbert space. (We have abused notation a bit and have simply written σ ∈ NRep(M), when σ is a normal rep- resentation of M.) The set of morphisms from σ to τ is the intertwiner TENSORIAL FUNCTION THEORY 7 space I(σ, τ ). We want to consider subcategories Σ of NRep(M) with the property that if σ and τ are in Σ, then so is σ ⊕ τ . Such a category is an additive subcategory of NRep(M). In the literature, particularly that dealing with the setting where M = C and E = Cd, it is im- portant to extract as much information as is possible from the finite dimensional representations of M, and, of course, the finite dimensional representations of M determine an additive subcategory of NRep(M). Although, we do not know how to modify the hypotheses in Theorem 1.3 so that we can restrict to every additive subcategory, we can prove that if Σ is a full additive subcategory of NRep(M) that consists of faithful representations of M, if f = {fσ}σ∈Σ is a uniformly bounded matricial family of functions, with fσ : D(0, 1, σ) → B(Hσ), then each of the fσs is Frechet holomorphic in D(0, 1, σ) and the Frechet-Taylor expansion of fσ can be expressed explicitly in terms of tensors from the tensor powers of E in the same way that the Taylor series for Berezin transforms can be expressed, i.e., in terms of our tensorial power series. But to do this, we must first use ideas from Kaliuzhnyi-Verbovetskyi and Vinnikov [KVV09, KV10, KVV] to show that the fσs have Taylor series in the sense of Joseph Taylor [Tay73] and then use a variant of our duality theorem [MS04, Theorem 3.6] to show that Taylor's Taylor coefficients are tensors of the desired type. Thus, we will show that there are strong interconnections among three notions of holomorphy and their associated power series: Frechet holomorphy for maps be- tween Banach spaces, Taylor's notion of free holomorphy, and tensorial holomorphy. We were led to these by studying Berezin transforms of elements of H ∞(E). Although Theorem 4.8 shows that a matricial family of functions f = {fσ}σ∈N Rep(M ), with fσ : D(0, 1, σ) → B(Hσ), is a Berezin transform if and only if f is uniformly bounded in σ, the proof very much depends upon properties of NRep(M) that are not shared by all subcategories. The best we can say at this stage is that if Σ is a full additive subcategory of NRep(M) consisting of faithful representations, if f = {fσ}σ∈Σ is a family of functions with fσ : D(0, 1, σ) → B(Hσ), and if supσ∈Σ supz∈D(0,1,σ) kfσ(z)k < ∞, then f is a family of tensorial power series if and only if f is a matricial family (Theorem 5.1). Additional information appears to be needed to conclude that such an f is a Berezin transform. Section 2 is dedicated to recapping a number of facts we need from the theory of tensor algebras and Hardy algebras. Tensorial power series are also introduced and some of their properties developed. Section 3 is devoted to developing general properties of matricial sets and functions. Section 4 is devoted to proving a refined version of Theorem 1.3. The TENSORIAL FUNCTION THEORY 8 connection between Taylor's Taylor series and our Tensorial Taylor series is made in Section 5. Finally in Section 6 we extend our work in Section 5 to cover maps between correspondence duals. Acknowledgement. We are pleased to thank Victor Vinnikov for en- lightening conversations we had with him over matricial function the- ory. His work in this area, especially that with Dmitry Kaliuzhnyi- Verbovetskyi and Mihai Popa, has been a source of inspiration to us. 2. Preliminaries Throughout, M will be a fixed W ∗-algebra with separable predual. However, we shall not fix a representation in advance, but instead we shall want to access the entire category of normal representations of M on separable Hilbert space, NRep(M). We shall always assume that such representations are unital. Strictly speaking, we shall not want to identify unitarily equivalent representations. But when M = C, there is no harm in doing so. Also, throughout this paper, Σ will denote an additive subcategory of NRep(M). This means that whenever σ and τ are in Σ so is σ⊕τ , where Hσ⊕τ := Hσ⊕Hτ and σ⊕τ (a) := σ(a)⊕τ (a). In particular, we shall be interested in additive subcategory generated by a single representation σ. It consists of all the finite multiples of σ, which we write nσ. When we want to consider an infinite multiple of σ, we will be explicit about this and denote it by ∞σ. Only at certain points will we need to assume that Σ is a full subcategory of NRep(M), meaning that when σ and τ are in Σ, then all the intertwiners between σ and τ are also in Σ. We will be explicit about where this assumption is used. We will, however, always assume that if σ lies in Σ and if τ = mσ, then the natural injection ιk of Hσ into Hτ , that identifies Hσ with the kth summand of Hτ , is a morphism in Σ, and so is its adjoint, ι∗ k, which is the projection of Hτ onto Hσ. We shall fix a W ∗-correspondence E over M. Thus, E is a self-dual right Hilbert module over M that is endowed with a left action given by a faithful normal representation ϕ of M in the algebra of all continuous linear right module maps on E, L(E). (Recall that because E is a self- dual right Hilbert module over M, every bounded right module map on E is automatically adjointable.) As with Hilbert space representations, we shall assume normal homomorphisms between W ∗-algebras are uni- tal and, in particular, we shall assume that ϕ is unital. Like M, E is a dual space. We shall refer to the weak-∗ topology on a W ∗-algebra or on a W ∗-correspondence as the ultraweak topology [MS11b, Paragraph 2.2]. TENSORIAL FUNCTION THEORY 9 The k-fold tensor powers of E, balanced over M (via ϕ), will be denoted E⊗k. Recall that E⊗k is the self-dual completion of the C ∗-tensor power and so is a W ∗-correspondence over M. The left action of M on E⊗k will be denoted by ϕk, i.e., ϕk(a)(ξ1 ⊗ ξ2 ⊗ · · · ⊗ ξk) := (ϕ(a)ξ1) ⊗ ξ2 ⊗ · · · ⊗ ξk. In particular, ϕ0 is just left multiplication of M on M and ϕ1 = ϕ [MS11b, Paragraph 2.7]. There are a number of spaces and algebras that we will want to build from the tensor powers E⊗k. First, there is the Fock space F (E) := M ⊕ E ⊕ E⊗2 ⊕ E⊗3 ⊕ · · · Here we interpret the sum as the self-dual completion of the Hilbert C ∗-module direct sum. It is a natural W ∗- correspondence over M, with the left action ϕ∞ := ϕ0 ⊕ ϕ1 ⊕ ϕ2 ⊕ · · · [MS11b, Paragraph 2.7]. For ξ ∈ E, we write Tξ for the creation operator determined by ξ. By definition, Tξη := ξ ⊗ η, η ∈ F (E). It is easy to check that Tξ is bounded, with norm dominated by kξk. Since we are assuming ϕ is injective and unital, it follows that kTξk = kξk. The norm-closed subalgebra of L(F (E)) generated by ϕ∞(M) and {Tξ}ξ∈E will be denoted T+(E) and called the tensor algebra of E. The Hardy algebra of E is defined to be the ultraweak closure of T+(E) in L(F (E)). The (nonclosed) subalgebra of T+(E) generated by ϕ∞(M) and{Tξ}ξ∈E will be denoted by T0(E) and called the algebraic tensor algebra of E. In the purely algebraic setting it would simply be called the tensor algebra of E.3 Before getting too far along with the theory, it will be helpful to have an example that we can follow as the theory develops. Example 2.1. (The Basic Example) In this example we let M = C and we let E = Cd. The left and right actions coincide and the inner product is the usual one, except we choose it to be conjugate linear in the right hand variable. We interpret Cd as ℓ2(N), when d = ∞. The representations of M on Hilbert space, in this case, are all multiples of the identity representation σ1 which represents C on C via multiplication, i.e., we may view the objects in NRep(M) as {nσ1 n ∈ N ∪ {∞}}. For m, n ∈ N ∪ {∞}, the intertwiner space I(mσ1, nσ1) may be identified with the n × m matrices, where when either m or n is infinite, we interpret the n × m matrices as bounded linear operators. To keep the notation simple, when it is convenient we shall write Mn(C) for B(ℓ2(N)) when n = ∞. An additive subcategory of NRep(M) is determined by any additive subsemigroup of N ∪ {∞}. 3Strictly speaking, because our coefficient algebra M may be noncommutative, in the algebra literature T0(E) would be called the tensor M -ring determined by the bimodule E. (See, e.g., [Coh06, P. 134].) TENSORIAL FUNCTION THEORY 10 It is customary to identify F (Cd) with ℓ2(F+ The Fock space F (E) is usually called the full Fock space of the Hilbert d ), where F+ space Cd. d denotes the free semigroup on d generators. The identification is carried out explicitly by letting {ei 1 ≤ i ≤ d} denote the standard basis for Cd and then sending the decomposable tensor ei1 ⊗ ei2 ⊗ · · · ⊗ eik ∈ (Cd)⊗k ⊆ F (Cd) to δw, where δw is the point mass at the word w = i1i2 · · · ik. The vacuum vector Ω in F (Cd) is sent to the empty word ∅ in F+ d . When this identification is made, then the creation operator Tei is identified with the Si, where Si is the isometry defined by the equation Siδw = δiw, w ∈ F+ d . Also under this identification, the algebraic tensor algebra T0(Cd) is identified with the free algebra on d generators, ChX1, X2,· · · , Xdi, realized as convolution operators on ℓ2(F+ d ) through the correspondence Xi ↔ Si. Of course, T+(Cd) may be viewed as the norm closed, unital algebra generated by the Si, and, as we mentioned in the introduction, in this guise it is known as Popescu's noncommutative disc algebra Its ultraweak closure, H ∞(E), which in this case coincides [Pop96]. with its weak operator closure, is the noncommutative analytic Toeplitz algebra. Definition 2.2. The full Cartesian product Πk≥0E⊗k will be called the formal tensor series algebra determined by M and E and will be denoted by T+((E)). Elements in T+((E)) will be denoted by formal sums, θ ∼Pk≥0 θk, θk ∈ E⊗k. It is easy to see that T+((E)) is, indeed, an algebra under coordinate- wise addition and scalar multiplication, with the product given by the formula where θ ∼ Pk≥0 θk, η ∼ Pk≥0 ηk, and ζ ∼ Pk ζk are related by the equation θ ∗ η = ζ, ζk := Xk=l+m θl ⊗ ηm, k, l, m ≥ 0. Further, we may view T0(E), T+(E) and H ∞(E) as subsets of T+((E)) using the "Fourier coefficient operators" calculated with respect to the gauge automorphism group acting on L(F (E)) [MS11b, Paragraph 2.9]. That is, if Pn is the projection of F (E) onto E⊗n and if Wt := eintPn, ∞Xn=0 TENSORIAL FUNCTION THEORY 11 then {Wt}t∈R is a one-parameter (2π-periodic) unitary group in L(F (E)) such that if {γt}t∈R is defined by the formula γt = Ad(Wt), then {γt}t∈R is an ultraweakly continuous group of ∗-automorphisms of L(F (E)), called the gauge automorphism group of L(F (E)). This group leaves each of the algebras T0(E), T+(E), and H ∞(E) invariant and so do each of the Fourier coefficient operators {Φj}j∈Z defined on L(F (E)) by the formula (2.1) Φj(F ) := 1 2π 2π 0 e−intγt(F ) dt, F ∈ L(F (E)), where the integral converges in the ultraweak topology. Further, if F ∈ H ∞(E), then Φj(F ) is of the form Tθj , where θj ∈ E⊗j. This is clear if F ∈ T0(E). For a general element F ∈ H ∞(E), the result follows from the facts that T0(E) is ultraweakly dense in H ∞(E) and Φj is ultraweakly continuous. In fact, if Σk is defined by the formula (2.2) Σk(F ) := Xj<k (1 − j k )Φj(F ), F ∈ L(F (E)), then for F ∈ L(F (E)), limk→∞ Σk(F ) = F , in the ultra- weak topology. It follows that for F ∈ H ∞(E), F determines and is uniquely determined by its sequence of Fourier coefficients θ = {θj}j≥0, where Tθj := Φj(F ). The correspondence E determines an endo-functor ΦE from NRep(M) into NRep(M). For an object σ in NRep(M), ΦE(σ) = σE ◦ ϕ, where σE is the representation of L(E) that is induced from σ in the sense of Rieffel [Rie74a]: For X ∈ L(E), σE(X) = X ⊗ IHσ acting on E ⊗σ Hσ. It is [Rie74a, Theorem 5.3] that ensures that for T ∈ I(σ, τ ), ΦE(T ) should be IE ⊗ T ∈ I(ΦE(σ), ΦE(τ )). Following the notation in [MS11b, Paragraph 2.6] and elsewhere in our work, we set Eσ∗ = I(σE ◦ ϕ, σ) for each σ ∈ NRep(M). Also, for z0 ∈ Eσ∗ and 0 ≤ R ≤ ∞, we write D(z0, R, σ, E) for the open disc or ball in Eσ∗ of radius R centered at z0. Usually, E will be dropped from the notation, since it will be understood from context. The space Eσ∗ is a bimodule over σ(M)′, where for a, b ∈ σ(M)′ and ξ ∈ Eσ∗, a · ξ · b := aξ(IE ⊗ b). In fact, it is a left Hilbert module over σ(M)′, where the inner product is given by the formula hξ, ηi = ξη∗ and as such, Eσ∗ becomes a correspondence over σ(M)′. Although the correspondence properties of Eσ∗ have figured heavily elsewhere in our work, they do not play much of a role here. Only the bimodule (In Section 6 we will properties are of consequence in this section. TENSORIAL FUNCTION THEORY 12 use properties of the correspondence Eσ, which is a right Hilbert W ∗- module over σ(M)′.) Once σ is fixed, each z ∈ Eσ∗ determines a homomorphism, denoted σ×z, from T0(E) to B(Hσ), which is defined on the generators of T0(E) via the formulas σ × z(ϕ∞(a)) = σ(a), a ∈ M, σ × z(Tξ)(h) = z(ξ ⊗ h), ξ ∈ E, h ∈ Hσ. The fact that these equations define a representation of T0(E) on B(Hσ) follows from the universal properties of tensor algebras once it is checked that the map ξ → σ × z(Tξ) is a bimodule map in the sense that σ × z(Ta·ξ·b) = σ(a)(σ × z(Tξ))σ(b), a, b ∈ M, ξ ∈ E. And conversely, every representation ρ of T0(E) on Hilbert space such that the restriction of ρ to ϕ∞(M) in T0(E) is a normal representation of M, say σ, and the restriction of ρ to the subspace {Tξ}ξ∈E yields an ultraweakly continuous bimodule map of E, must be of the form σ × z, where z is given by the formula z(ξ ⊗ h) = ρ(Tξ)h. As we indicated in the introduction, it is of fundamental importance for our theory, that σ × z extends from T0(E) to a completely contractive representation of T+(E) on B(Hσ) if and only if z lies in the norm closed disc D(0, 1, σ) [MS11b, Theorem 2.9]. Moreover, for kzk < 1, σ × z extends further to be an ultraweakly continuous, completely contractive representation of H ∞(E) on Hσ [MS04, Corollary 2.14]. We write AC(σ, E), or simply AC(σ), when E is understood, for the collection of all z ∈ D(0, 1, σ) such that σ × z extends to an ultraweakly continuous, completely contractive representation of H ∞(E) on Hσ. We then call z an absolutely continuous point in D(0, 1, σ) and we say that σ × z is an absolutely continuous representation. The terminology derives from the fact that one can define an H ∞-functional calculus in the sense of Sz.-Nagy and Foiaş for a contraction operator if and only if its minimal unitary dilation is absolutely continuous. (See [MS11b], where this perspective is developed at length.) Remark 2.3. There is now a bit of ambiguity that needs to be clarified. When we write I(σ × z, τ × w) we shall always mean the set of all operators C : Hσ → Hτ such that C ∈ I(σ, τ ) and Cz = w(IE ⊗ C). It is straightforward to see that this happens if and only if C ∈ I(σ, τ ) and Cσ × z(F ) = τ × w(F )C for all F ∈ T0(E). Since T0(E) is norm dense in T+(E) and since σ × z and τ × w extend to continuous TENSORIAL FUNCTION THEORY 13 representations of T+(E) when kzk,kwk ≤ 1, it follows that I(σ×z, τ × w) = {C : Hσ → Hτ Cσ × z(F ) = τ × w(F )C for all F ∈ T+(E)}, when kzk,kwk ≤ 1. Likewise, since AC(σ) is the set of all z ∈ Eσ∗ such that σ×z extends from T+(E) to an ultraweakly continuous completely contractive representation of H ∞(E), we find that I(σ × z, τ × w) = {C : Hσ → Hτ Cσ × z(F ) = τ × w(F )C for all F ∈ H ∞(E)} when z ∈ AC(σ) and w ∈ AC(τ ). With the notation we have established, it follows that T0(E) may be represented as a space polynomial-like, B(Hσ)-valued functions on Eσ∗ via the formula: (2.3) bFσ(z) := σ × z(F ), F ∈ T0(E), z ∈ Eσ∗. Since our primary objective is to understand the nature of the functions bFσ, it will be helpful to work carefully through some examples that illuminate their definition. Consider first the case when F = Tξ1Tξ2 = Tξ1⊗ξ2. Then for h ∈ Hσ, we have (2.4) (σ × z)(Tξ1Tξ2)(h) = (σ × z)(Tξ1)(σ × z)(Tξ2)(h) = z(ξ1 ⊗ z(ξ2 ⊗ h)) = z(IE ⊗ z)(ξ1 ⊗ ξ2 ⊗ h). More generally, we find that (σ× z)(Tξ1⊗ξ2⊗···⊗ξk)(h) = z(IE ⊗ z)· · · (IE⊗k−1 ⊗ z)(ξ1⊗ξ2⊗· · ·⊗ξk⊗h). Observe that z(IE ⊗ z)· · · (IE⊗k−1 ⊗ z) is a bounded linear operator from E⊗k ⊗ Hσ to Hσ, with norm dominated by kzkk. Further, z(IE ⊗ z)· · · (IE⊗k−1⊗z) intertwines σE⊗k◦ϕk and σ. Thus z(IE⊗z)· · · (IE⊗k−1⊗ z) is an element of I(σE⊗k ◦ ϕk, σ) = (E⊗k)σ∗. In [MS04, Page 363] z(IE ⊗ z)· · · (IE⊗k−1 ⊗ z) is called the kth generalized power of z. We shall denote it here by z(k) and we shall denote the function z → z(k), from Eσ∗ to (E⊗k)σ∗, by Zk, i.e., Zk(z) := z(k) = z(IE⊗z)· · · (IE⊗k−1⊗z). By convention, we set Z0(z) ≡ IHσ . Example 2.4. In the context of our basic example, Example 2.1, E⊗σ Hσ is Cd ⊗ Hσ, which we shall view simply as columns of length d of vectors from Hσ. We may do this, since there is no consequential balancing going on. After all, σ is a representation of C. It follows that Eσ∗ := I(σE ◦ ϕ, σ) may be identified with all d-tuples of operators in B(Hσ) arranged as a row. That is, we may and shall write Eσ∗ = {z = (Z1, Z2,· · · , Zd) Zi ∈ B(Hσ)}. So, if F ∈ T0(E) and if we identify T0(E) with the free algebra on d generators, as we did in our basic TENSORIAL FUNCTION THEORY 14 becomes example, we may write F = Pw∈F+ bFσ(z) = Xw∈F+ d d awZw, awXw, and then equation (2.3) where Zw = Zi1Zi2 · · · Zik, w = i1i2 · · · ik. In particular, if σ happens to be an n-dimensional representation with n < ∞, then Eσ∗ is the collection of all d-tuples of n × n matrices and equation (2.3) yields the representation of the free algebra on d generators on the algebra of d generic n × n matrices. The function Zk in this setting assigns to z = (Z1, Z2,· · · , Zd) the tuple (Zw)w=k of length dk that is ordered lexicographically by the words of length k. In the following lemma and, indeed, throughout this paper, we identify E⊗(k+l) with E⊗k ⊗ E⊗l; E⊗0 := M. Lemma 2.5. The functions Zk, from Eσ∗ to (E⊗k)σ∗, satisfy the equa- tion Zk+l(z) = Zk(z)(IE⊗k ⊗ Zl(z)), z ∈ Eσ∗. Further, each Zk is Frechet differentiable and the Frechet derivative of Zk, denoted DZk, is given by the formula (2.5) (DZk)(z)[ζ] = k−1Xl=0 Zl(z)(IE⊗l ⊗ ζ)(IE⊗(l+1) ⊗ Zk−l−1(z)) for all z, ζ ∈ Eσ∗. The proof is a straightforward calculation based on the definition of Zk and will be omitted. Observe that for each z ∈ Eσ∗, DZk(z) is a bounded linear operator from Eσ∗ to (E⊗k)σ∗ whose norm is dominated by kkzkk−1. If θ ∈ E⊗k, we shall write Lθ for the linear operator from Hσ to E⊗k⊗σ Hσ that maps h to θ⊗ h. Evidently, kLθk ≤ kθk, with equality holding if σ is faithful. The calculations that led from equation (2.4) through Lemma 2.5 immediately yield Theorem 2.6. Suppose F =Pn k=0 Tθk ∈ T0(E), with θk ∈ E⊗k. Then for σ ∈ NRep(M) and z ∈ Eσ∗, the operator bFσ(z) in B(Hσ) is given by the formula (2.6) bFσ(z) = nXk=0 Zk(z)Lθk. TENSORIAL FUNCTION THEORY 15 Further, as a function of z ∈ Eσ∗, bFσ is Frechet differentiable and its Frechet derivative, mapping Eσ∗ to B(Hσ), is given by the formula (2.7) DbFσ(z)[ζ] = nXk=0 DZk(z)[ζ]Lθk. Of course, we would like to extend our function theory beyond the realm of the algebraic tensor algebra. For this purpose, we follow Popescu [Pop06] (who followed Hadamard, who followed Cauchy) and introduce the following definition. Definition 2.7. For θ ∼Pk≥0 θk in T+((E)), we define R(θ) to be 1 R(θ) = (lim sup k→∞ kθkk k )−1, and we call R(θ) the radius of convergence of the formal series θ. Of course, R(θ) is a non-negative real number or +∞. Proposition 2.8. Suppose θ ∼ Pk≥0 θk in T+((E)) and R = R(θ) is not zero. Then for each σ ∈ NRep(M), for each z0 ∈ Eσ∗, and for each z ∈ D(z0, R, σ), the series ∞Xk=0 (2.8) Zk(z − z0)Lθk converges in the norm of B(Hσ). The convergence is uniform on any subdisk D(z0, ρ, σ), ρ < R; and it is uniform in σ, as well. The resulting function, fσ, is analytic as a map from D(z0, R, σ) to B(Hσ) and the Frechet derivative of fσ is given by the formula Dfσ(z)[ζ] = ∞Xk=1 DZk(z − z0)[ζ]Lθk, z ∈ D(z0, R, σ), ζ ∈ Eσ∗. 1 Proof. The uniform convergence of (2.8) follows a standard argument: R(θ) = lim supk kθkk1/k, there Fix ρ, ρ′ such that 0 < ρ < ρ′ < R. Since is an m such that, for all k ≥ m, kθkk1/k < 1/ρ′. For such k, for all σ and for all z ∈ D(z0, ρ, σ), kZk(z − z0)Lθkk ≤ kz − z0kkkθkk ≤ ( ρ ρ′ )k. Since ρ < ρ′, we see that (2.8) converges uniformly on D(z0, ρ, σ). The remaining assertions can be proved by similar elementary estimates. Alternatively, one can appeal to [HP74, Theorems 3.17.1 and 3.18.1]. (cid:3) Definition 2.9. The series (2.8) is called the tensorial power series determined by the formal tensor series θ, the point z0, and the repre- sentation σ. TENSORIAL FUNCTION THEORY 16 Remark 2.10. Following the line of thought developed in elementary texts on complex function theory, one might expect that if z ∈ Eσ∗ has norm larger than R(θ), then the series (2.8) diverges. However, the situation is quite a bit more complicated than in one complex variable. For example, if M = E = C and if dim Hσ ≥ 2, then we may identify Eσ∗ with B(Hσ), and if z is any nilpotent element in Eσ∗, the series (2.8) converges no matter what R(θ) is. The situation is even more complicated: There are series θ with finite R(θ) such that for "most" σ the series (2.8) converges on all of Eσ∗, i.e., almost all fσ are entire. A particularly instructive example, due to Luminet [Lum86, Exam- ple 2.9], is constructed as follows. Let Sk(X1, X2,· · · , Xk) denote the standard identity in k noncommuting variables: Sk(X1, X2,· · · , Xk) = Xs∈Sk sign(s)Xs(1)Xs(2) · · · Xs(k), where Sk denotes the permutation group on k letters and where sign(s) is 1 if s is even and −1 if s is odd. Then the set of these identities determines an element θ in T+((C2)) by identifying T+((C2)) with the free formal series in two noncommuting variables X1 and X2 and setting θ(X1, X2) =Xk≥2 Sk(X1, X1X2,· · · , X1X k−1 2 ). When this series is written as a series of words in X1 and X2, θ(X1, X2) = Xw∈F+ 2 λwXw, one sees that for each positive integer d there is a word w of length at least d so that λw = 1. It follows that R(θ) < ∞. However, if σ is any finite dimensional representation of C = M, the series (2.8) converges throughout Eσ∗. Indeed, for any given z there are only finitely many nonzero terms in the series. On the other hand, we will be able to show later that given R′ > R(θ), there is a σ ∈ NRep(M) and an element z ∈ Eσ∗ with kzk = R′ such that the series (2.8) diverges. The next proposition establishes a bridge between the theory developed in [MS04, MS05, MS08, MS09] and the focus of this paper. First, recall that if F ∈ H ∞(E), then the Fourier coefficients Φj(F ) are of the form Tθj for θj ∈ E⊗j. Further, the series Pj≥0 Tθj is Cesaro summable to F (2.2). On the other hand, each Fourier coefficient operator Φj is contractive and so the norm of each θj is dominated by kFk. Thus the radius of convergence of the series θ ∼ Pj≥0 θj is at least one. Since TENSORIAL FUNCTION THEORY 17 σ × z(Tθ) = Zj(z)Lθ for θ ∈ E⊗j and z ∈ D(0, 1, σ), the validity of the following proposition is immediate. Proposition 2.11. For F ∈ H ∞(E) and σ ∈ Σ, the σ-Berezin trans- form bFσ, defined on AC(σ) by the formula, bFσ(z) := σ × z(F ), admits the tensorial power series expansion, (2.9) ∞Xj=0 bFσ(z) = Zj(z)Lθj , where Tθj = Φj(F ). The series converges in norm for z ∈ D(0, 1, σ), uniformly in σ and on any sub-disk, D(0, ρ, σ), ρ < 1. Moreover, the function bFσ(z) is bounded by kFk throughout AC(σ). Remark 2.12. In view of Proposition 2.11, one might wonder if every bounded analytic function on D(0, 1, σ), with tensorial power series P∞ j=0 Zj(z)Lθj , comes from a function in H ∞(E). Thanks to a very detailed study by Arveson, such is not the case. He shows that in one of the simplest settings, when M = C, E = Cd (d ≥ 2), and σ is one- dimensional, there are bounded analytic functions on D(0, 1, σ) (which is the unit ball in Cd) which are not of the form bFσ(z) for any element F in H ∞(Cd) (see [Arv98, Theorem 3.3]). 3. Matricial Families and Functions 0 0 z21 0 z12 Suppose σ and τ are two normal representations of our W ∗-algebra M on Hilbert spaces Hσ and Hτ , respectively. We have noted that if one writes σ ⊕ τ matricially as (σ ⊕ τ )(·) = (cid:20)σ(·) τ (·)(cid:21), then E(σ⊕τ )∗ = I((σ ⊕ τ )E ◦ ϕ, (σ ⊕ τ )) may be written as matrices of oper- ators(cid:20)z11 z22(cid:21), viewed as operators from HσE ◦ϕ ⊕ Hτ E ◦ϕ to Hσ ⊕ Hτ , where z11 ∈ I(σE◦ϕ, σ), z12 ∈ I(τ E◦ϕ, σ), z21 ∈ I(σE◦ϕ, τ ), and where z22 ∈ I(τ E ◦ ϕ, τ ). In particular, note that all the matrices of the form (cid:20)z11 z22(cid:21), with z11 ∈ I(σE ◦ ϕ, σ) and z22 ∈ I(τ E ◦ ϕ, τ ) are contained 0 E(σ⊕τ )∗. We abbreviate this fact by writing Eσ∗ ⊕ Eτ ∗ ⊆ E(σ⊕τ )∗. We will have occasion later to think about higher order matrices. That is, if σ1, σ2,· · · , σn are n (not-necessarily distinct) representations in NRep(M), then E(σ1⊕σ2⊕···⊕σn)∗ can be viewed as n× n matrices whose i, j-entries lie in I(σE j ◦ ϕ, σi). When this is done, we will be able to write P⊕ 1≤i≤n Eσi∗ ⊆ E(σ1⊕σ2⊕···⊕σn)∗. TENSORIAL FUNCTION THEORY 18 Remark 3.1. In Definition 1.1, we defined the notion of a matricial family of sets indexed by all of NRep(M). Evidently, that notion can be "relativized" to additive subcategories. So if Σ is an additive subcategory of NRep(M), then a family of sets U = {U(σ)}σ∈Σ that satisfies the following two properties will be called a matricial E, Σ- family of sets. If Σ or E are clear from context, we shall simply call U a matricial family of sets. sums, i.e., (1) Each U(σ) is contained in Eσ∗. (2) The family {U(σ)}σ∈Σ is closed with respect to taking direct U(σ) ⊕ U(τ ) :=(cid:20)U(σ) U(τ )(cid:21) ⊆ U(σ ⊕ τ ). 0 0 If each of the members of U is described by a common property, such as being open or a domain, then we shall adjust the terminology ap- propriately, e.g., by saying that the family is a matricial E, Σ-family of open sets or domains. We shall say that U unitarily invariant, if each U(σ) is unitarily invariant in the sense that for each unitary operator u ∈ σ(M)′ and each z ∈ U(σ), u−1·z·u := u−1z(IE⊗u) lies in U(σ). Re- latedly, U is called matricially convex, if for any σ and τ in Σ and for all v ∈ I(σ, τ ) such that vv∗ = IHτ , v ·U(σ)· v∗ = vU(σ)(IE ⊗ v∗) ⊆ U(τ ). Example 3.2. We already have noted that {D(0, 1, σ)}σ∈N Rep(M ) is a matricially convex, matricial family of sets. However, the best we can say, in general, is that {AC(σ)}σ∈N Rep(M ) is unitarily invariant. For a more general class of discs, suppose that {ζσ}σ∈Σ is a family of vectors with ζσ ∈ Eσ∗ such that ζσ⊕τ = ζσ ⊕ ζτ for all σ, τ ∈ Σ. We shall call ζ := {ζσ}σ∈Σ an additive field of vectors over Σ. Given such a field ζ and an R, 0 ≤ R ≤ ∞, we shall call D(ζ, R) := {D(ζσ, R, σ)}σ∈Σ the matricial disc determined by the field ζ. If ζσ = 0 for all σ ∈ Σ, we shall simply write D(0, R), calling it the matricial disc {D(0, R, σ)}σ∈Σ we already have defined. It is clear that D(ζ, R) is a matricial set when ζ is an additive field of vectors over Σ, but in general D(ζ, R) won't be matricially convex. However, it will be matricially convex when ζ = {ζσ}σ∈Σ is a central additive field in the sense that ζτ (IE ⊗ C) = Cζσ for all C ∈ I(σ, τ ), σ, τ ∈ Σ, as is evident from the definition. Note that for ζ to be a central additive field requires more than each ζσ being central in the sense of [MS08, Definition 4.11], which simply means that ζσ(IE ⊗ C) = Cζσ for all C ∈ I(σ, σ) = σ(M)′. However, if each ζσ is central in that sense, then D(ζ, R) is a unitarily invariant matricial set. Note, too, that if Σ is the subcategory consisting of all multiples of a single representation σ, together with all the spaces TENSORIAL FUNCTION THEORY 19 I(nσ, mσ), n, m ∈ N, and if ζσ is a central vector in Eσ∗, then {ζmσ}m∈N is a central additive field on Σ, where ζmσ is the m-fold direct sum of copies of ζσ. To see these examples in the simplest, most concrete setting, consider the following addition to our basic example, Example 2.1. Example 3.3. Let M = C and let E = Cd. Then since every repre- sentation σ of C is simply a multiple of the basic, 1-dimensional rep- resentation σ1, i.e., σ = nσ1 for a suitable positive integer n or ∞, it follows that Eσ∗ = {z = (Z1, Z2,· · · Zd) Zi ∈ Mn(C)}, where we inter- pret M∞(C) as B(ℓ2(N)). The disc D(0, R, σ) is {z = (Z1, Z2,· · · Zd) kP ZiZ ∗ i k 1 2 < R}, and a z = (Z1, Z2,· · · , Zd) ∈ D(0, 1, σ) is in AC(σ), when σ is a finite multiple of σ1, if and only if z is completely non- coisometric [MS11b, Corollary 5.7] in the sense that kz(k)∗hk → 0 for every h ∈ Cn [MS04, Remark 7.2]. (No such simple characterization of AC(∞σ1) is known.) For a z ∈ Eσ∗ and unitary u ∈ σ(M)′ = Mn(C), u−1 · z · u = u−1z(IE ⊗ u) = (u−1Z1u, u−1Z2u,· · · , u−1Zdu). Thus z is central if and only if each Zi is a scalar multiple of the identity. Fur- ther, as we have seen, an additive subcategory Σ of NRep(M) is simply determined by an additive subsemigroup of N∪ {∞} and a central ad- ditive field {ζσ}σ∈Σ is a family of central elements such that for each i, 1 ≤ i ≤ d, the scalar that forms the scalar multiple of the identity in the ith element of ζσ is independent of σ. Remark 3.4. The notion of a matricial family of functions indexed by NRep(M), defined in Definition 1.2, can be "relativized" to additive subcategories, too. So, if {U(σ)}σ∈Σ is a matricial E, Σ-family, where Σ is an additive subcategory of NRep(M), then a family of functions f = {fσ}σ∈Σ, with fσ : U(σ) → B(Hσ), σ ∈ Σ, is an E, Σ-matricial family of functions, or simply a matricial family of functions, in case f respects intertwiners in the sense that for every z ∈ U(σ), every w ∈ U(τ ), I(σ×z, τ×w) ⊆ I(fσ(z), fτ (w)). (Thanks to our convention established in Remark 2.3, I(σ×z, τ ×w) ⊆ I(fσ(z), fτ (w)) if and only if every C ∈ I(σ, τ ) that satisfies the equation Cz = w(IE ⊗ C) also satisfies the equation Cfσ(z) = fτ (w)C.) We shall say that f is continuous, or holomorphic, etc., in case each fσ has the indicated property. Proposition 3.5. Let Σ be an additive subcategory of NRep(M) and let D(ζ, R) be the matricial disc determined by a central additive field ζ on Σ, where 0 < R ≤ ∞. If θ ∈ T+((E)) has radius of convergence TENSORIAL FUNCTION THEORY 20 at least R, then the collection {fσ}σ∈Σ of tensorial power series deter- mined by θ, fσ(z) := Pk≥0 Zk(z − ζσ)Lθk, forms a matricial family of holomorphic functions on D(ζ, R). Proof. We saw in Proposition 2.8 that each fσ is a holomorphic B(Hσ)- valued function defined throughout D(ζσ, R, σ). We therefore need only check that {fσ}σ∈Σ preserves intertwiners. For this, it suffices to check that for each k, the kth terms of {fσ}σ∈Σ preserve intertwiners. If C ∈ I(σ × z, τ × w), then by definition C ∈ I(σ, τ ), so C ∈ I(Z0(z − ζσ)Lθ0,Z0(w−ζτ )Lθ0) because Z0(z−ζσ) is identically IHσ in z, similarly for Z0(w − ζτ ), and because when we view Lθ0 as a map from Hσ to M ⊗σ(M ) Hσ we identify θ0 with σ(θ0), when we identify Hσ with M ⊗σ(M ) Hσ, as is customary; similarly for Lθ0 and τ (θ0). To handle I(Z1(z− ζσ)Lθ1,Z1(w− ζτ )Lθ1), observe first that Lθ1C = (IE ⊗ C)Lθ1 whether Lθ1 is viewed as a map from Hσ to E ⊗σ Hσ or from Hτ to E ⊗τ Hτ . Further, Cζσ = ζτ (IE ⊗ C) by definition of a central family and Cz = w(IE ⊗ C) by the hypothesis that C ∈ I(σ × z, τ × w). Thus C ∈ I(Z1(z − ζσ)Lθ1,Z1(w − ζτ )Lθ1). The general case I(Zk(z− ζσ)Lθk,Zk(w− ζτ )Lθk) is handled by noting that it suffices to check the intertwining condition when θk is a decomposable tensor, say θk = ξ1⊗ξ2⊗· · ·⊗ξk and noting that in this case, Zk(z−ζσ)Lθk = Z1(z− ζσ)Lξ1Z1(z − ζσ)Lξ2 · · ·Z1(z − ζσ)Lξk and similarly Zk(w − ζτ )Lθk = Z1(w − ζτ )Lξ1Z1(w − ζτ )Lξ2 · · ·Z1(w − ζτ )Lξk . For these expressions, it is obvious that C intertwines, by virtue of the fact that C ∈ I(Z1(z− ζσ)Lθ1,Z1(w − ζτ )Lθ1). A concept that is closely related to the notion of a matricial family of functions is given in (cid:3) Definition 3.6. Suppose E and F are two W ∗-correspondences over the same W ∗-algebra, M and suppose Σ is an additive subcategory of NRep(M). If {U(σ)}σ∈Σ is an E, Σ-matricial family of sets and if {V(σ)}σ∈Σ is an F, Σ-matricial family of sets, then we call a family of maps {fσ}σ∈Σ, where fσ maps U(σ) to V(σ), an E, F, Σ-matricial family of maps, or for short, a matricial family of maps, in case Cfσ(z) = fτ (w)(IF ⊗ C) for every C : Hσ → Hτ in I(σ, τ ) such that Cz = w(IE ⊗ C), for all z ∈ U(σ) and all w ∈ V(τ ). To say the same thing more succinctly, {fσ}σ∈Σ is a matricial family of maps in case (3.1) I(σ × z, τ × w) ⊆ I(σ × fσ(z), τ × fτ (w)), z ∈ U(σ), w ∈ U(τ ). TENSORIAL FUNCTION THEORY 21 To help forestall confusion that may develop, we emphasize that we will be consistent in our distinction between functions and maps: ma- tricial families of functions map matricial families of sets to algebras of operators; matricial families of maps are families of maps between matricial families of sets. As we shall see later, matricial families of maps are connected to homomorphisms between Hardy algebras. 4. A special Generator In ring theory, a module G is a generator for the category of left mod- ules over a ring R, RM, in case every M ∈ RM is the image of a homomorphism from the algebraic direct sum of a suitable number of copies of G [AF92, Page 193]. In [Rie74b], Rieffel defines a generator for NRep(M) in a similar fashion, but allows infinite Hilbert space direct sums. In [Rie74b, Proposition 1.3], he proves that a representa- tion σ is a generator for NRep(M) in this extended sense if and only if σ is faithful. Here we shall develop a useful analogue of the notion of a generator for the category of ultraweakly continuous, completely contractive representations of H ∞(E). Definition 4.1. We shall say that a generator π for NRep(M) is an infinite generator in case it is an infinite multiple of a generator for NRep(M), i.e., an infinite multiple of a faithful normal representation of M. We shall say that σ0 is a special generator for NRep(M) if σ0 = πF (E) ◦ ϕ∞ for an infinite generator for NRep(M). Remark 4.2. Of course σ0 and π are equivalent in NRep(M) if π is an infinite generator. However, we want to consider additive subcategories of NRep(M) that are not necessarily closed under forming infinite di- rect sums. Consequently, it is important for our considerations to make a distinction between σ0 and π. If σ0 = πF (E) ◦ ϕ∞, acting on Hσ0 = F (E)⊗π Hπ, is a special generator for NRep(M), and if s0 is defined by the formula s0(ξ ⊗ h) = Tξh, ξ ∈ E, h ∈ F (E) ⊗π Hπ, then σ0 × s0 is an induced representation of H ∞(E) in the sense of [MS99]. In [MS11b, Proposition 2.3] we show that σ0 × s0 is unique up to unitary equivalence in the sense that if π′ has the same properties as 0 is constructed from π′ in a similar fashion to σ0×s0, then π and if σ′ σ′ 0⊗ s′ 0 is unitarily equivalent to σ0× s0. Further, if σ× z is any induced representation of H ∞(E), then there is a subspace K of Hπ that reduces π such that σ× z is unitarily equivalent to σ0 × s0F (E)⊗π K. Observe that by construction σ0×s0 is absolutely continuous, so s0 ∈ AC(σ0). In 0⊗s′ TENSORIAL FUNCTION THEORY 22 fact, σ0×s0 is a generator for the category of all ultraweakly continuous completely contractive representations of H ∞(E). This assertion is the content of the following theorem, which is a summary of Theorems 4.7 and 4.11 of [MS11b]. Theorem 4.3. Suppose σ ∈ NRep(M) and z ∈ D(0, 1, σ). Then z lies in AC(σ) if and only if Hσ is the closed linear span of the ranges of the operators in I(σ0 × s0, σ × z). In this event, the stronger equation holds: Hσ =[{Ran(C) C ∈ I(σ0 × s0, σ × z)}. With this theorem at our disposal, we are able to prove the following theorem that has Theorem 1.3 as an immediate corollary. Theorem 4.4. Suppose Σ is a full additive subcategory of NRep(M) that contains a special generator σ0 for NRep(M). If f = {fσ}σ∈Σ is a family of functions such that each fσ maps AC(σ) to B(Hσ), then f is a Berezin transform (restricted to {AC(σ)}σ∈Σ) if and only if f is an E, Σ-matricial family. Proof. We already have noted Berezin transforms are matricial fam- ilies of functions on {AC(σ)}σ∈N Rep(M ). So certainly, their restric- tions to {AC(σ)}σ∈Σ are matricial E, Σ-families. For the converse, suppose that {fσ}σ∈Σ is a matricial family of functions defined on {AC(σ)}σ∈Σ. Since Σ is assumed to be full, for every σ and τ in Σ, IΣ(σ, τ ) = IN Rep(M )(σ, τ ) where the subscripts indicate the category under consideration. It follows that for every z ∈ AC(σ) and for ev- ery w ∈ AC(τ ), IΣ(σ × z, τ × w) = IN Rep(M )(σ × z, τ × w). So our hypotheses guarantee that for every C ∈ IN Rep(M )(σ0 × s0, σ0 × s0), Cfσ0(s0) = fσ0(s0)C. That is, fσ0(s0) lies in the double commutant of σ0 × s0(H ∞(E)). However, σ0 × s0 is the restriction of πF (E) to H ∞(E), where σ0 = πF (E) ◦ ϕ∞, and πF (E)(H ∞(E)) is its own double commutant by [MS04, Corollary 3.10]. Thus there is an F ∈ H ∞(E) so that fσ0(s0) = bFσ0(s0). (4.1) If σ is an arbitrary representation in Σ and if z ∈ AC(σ), then for every C ∈ IN Rep(M )(σ0 × s0, σ × z), fσ(z)C = Cfσ0(s0) because Σ is full and {fσ}σ∈Σ preserves intertwiners by hypothesis. However, by (4.1), Cfσ0(s0) = CbFσ0(s0). Hence we have fσ(z)C = CbFσ0(s0) = bFσ(z)C, TENSORIAL FUNCTION THEORY 23 (cid:3) where the second equality is justified by Remark 2.3. Since the ranges of the C in IN Rep(M )(σ0 × s0, σ × z) cover all of Hσ, by Theorem 4.3, we conclude that fσ(z) = bFσ(z). We digress momentarily to provide an example promised in the intro- duction that shows that Theorems 4.4 and 1.3 can fail if the hypothesis that the matricial function f = {fσ}σ∈Σ in question is defined only on {D(0, 1, σ)}σ∈Σ and not on the collection of larger sets, {AC(σ)}σ∈Σ. Example 4.5. We let M = C = E. Then NRep(M) may be identified with {nσ1}0<n≤∞, where, recall, σ1 is the one-dimensional representa- tion of C on C. The disc D(0, 1, σ) is just the collection of all operators of norm less than 1 in B(Hσ). We set fσ(z) = (IHσ − z)−1, i.e., fσ is just the resolvent operator restricted to D(0, 1, σ). Then it is immedi- ate that f = {fσ}σ∈N Rep(M ) is a matricial family of functions defined on {D(0, 1, σ)}σ∈N Rep(M ). However, since none of the fσs is bounded, f is not a Berezin transform of an element in H ∞(E) ≃ H ∞(T). This example should be compared with Theorem 4.8. We would like to use Theorem 4.4 to obtain information about which matricial families of functions come from tensorial power series that have a given radius of convergence. First, however, we take up an issue that was left hanging after Proposition 2.8. Proposition 4.6. Suppose θ ∼P θk ∈ T+((E)) has a finite radius of convergence R = R(θ). If R′ > R, then there a σ ∈ NRep(M) and z ∈ Eσ∗, with kzk = R′, such that the tensorial power series Pk Zk(·)Lθk diverges at z; indeed, Pk kZk(z)Lθkk = ∞. Proof. Choose ρ, with R < ρ < R′, let σ be σ0 and set z = R′s0. Since 1 ρ < 1 R, there are infinitely many ks for which kθkk1/k > 1 ρ. On the other hand, Zk(z)Lθk = Zk(R′s0)Lθk = R′k(Tθk ⊗π IK0). Consequently, kZk(z)Lθkk = R′kkθkk. So for each k satisfying kθkk1/k > 1 ρ, we have kZk(z)Lθkk > ( R′ ρ )k > 1. Since there are infinitely many such ks, the series Pk kZk(z)Lθkk diverges to ∞. Theorem 4.7. Suppose Σ is a full additive subcategory of NRep(M) containing a special generator for NRep(M). If f = {fσ}σ∈Σ is a fam- ily of functions, with fσ mapping D(0, R, σ) to B(Hσ), then there is a formal tensor series θ with R(θ) ≥ R such that f is the family of tenso- rial power series determined by θ, {Pk≥0 Zk(z)Lθk z ∈ D(0, R, σ)}σ∈Σ, if and only if f is an E, Σ-matricial family of functions. (cid:3) TENSORIAL FUNCTION THEORY 24 Proof. Arguments we have used before show that if the fσs admit ten- sorial power series expansions of the indicated kind, then the family preserves intertwiners. Consequently, we shall concentrate on the con- verse assertion. So assume that {fσ}σ∈Σ preserves intertwiners. Since Σ is assumed to be full, we may and will drop the subscripts Σ and NRep(M) on all intertwining spaces. Also, we will let σ0 = πF (E)◦ ϕ∞, for a suitable infinite generator π. The key to our analysis is to focus on fσ0 in order to bring properties of H ∞(E) into play and then to use the intertwining property of the family {fσ}σ∈Σ to propagate to the other discs D(0, R, σ). Fix 0 < r < R and consider fσ0(rs0). Recall that σ0 × s0(H ∞(E)) = πF (E)(H ∞(E)). Consequently, every C in the com- mutant of πF (E)(H ∞(E)) lies in I(σ0×s0, σ0×s0) = I(σ0×rs0, σ0×rs0) and, thus commutes with fσ0(rs0). Since πF (E)(H ∞(E)) equals its own double commutant, there is an element Fr ∈ H ∞(E) such that fσ0(rs0) = πF (E)(Fr) = cFrσ0(s0). Now take a z ∈ D(0, r, σ). Then k 1 r zk < 1 and so σ × 1 r z is absolutely continuous. We conclude, by Theorem 4.3, that (4.2) _{Ran(C) C ∈ I(σ0 × s0, σ × 1 r z)} = Hσ. Also, for every C ∈ I(σ0 × s0, σ × 1 r z) = I(σ0 × rs0, σ × z), Cfσ0(rs0) = fσ(z)C by hypothesis. Since fσ0(rs0) =cFrσ0(s0), we see that fσ(z)C = CcFrσ0(s0). But by Remark 2.3, CcFr σ0(s0) =cFrσ( 1 kzk < r < R. r z)C, so we conclude from (4.2) that , (4.3) 1 r z), fσ(z) =cFr σ( We need to remove the dependence of Fr on r. So if 0 < r < r1 < z). r2 < R and if kzk ≤ r, we obtain the equation cFr1 σ( 1 z) = cFr2 σ( 1 In particular, r1 r2 We now would like to apply the "Fourier coefficient maps" Φk to equa- tion (4.4). To give this its proper meaning, note that, whenever X ∈ H ∞(E) and 0 < t < 1, (4.5) \Φk(X)σ0 (ts0) = tkΦk(X) ⊗ IHπ. This is easy to verify by first taking X = Tξ for ξ ∈ E⊗m, and then using linearity and ultra-weak continuity. Thus, applying Φk to equation (4.4) r r1 cFr1 σ0( s0) = cFr2 σ0( r r2 s0). TENSORIAL FUNCTION THEORY 25 (4.4), we obtain (4.6) rk rk 1 Φk(Fr1) ⊗ IHπ = rk Φk(Fr2) ⊗ IHπ. rk 2 for all k ≥ 0, which implies that r 7→ 1 rk Φk(Fr) is constant in r, 0 < r < R, for every k. Consequently, since the image of Φk is {Tθk θk ∈ E⊗k}, there is a θk ∈ E⊗k, independent of r, so that Tθk ⊗ IHπ = 1 rk Φk(Fr) ⊗ IHπ, 0 < r < R. Φk(Fr) = rkTθk = Trkθk, Canceling "⊗IHπ ", as we may, we conclude that (4.7) Now fix 0 < r < R and z ∈ Eσ∗ with kzk < r. For 0 ≤ k, let ξk = rkθk ∈ E⊗k. Then with Fr in place of F and 1 r z in place of z, we find that 0 < r < R. 1 r z) =Xk≥0 1 rkZk(z)Lξk cFrσ( 1 andP 1 rkkZk(z)Lξkk < ∞. By (4.3), we conclude that fσ(z) =cFr σ( 1 rkZk(z)Lξk = Pk≥0 Zk(z)Lθk. Thus PkZk(z)Lθkk < ∞ and Pk≥0 fσ(z) =Pk≥0 Zk(z)Lθk. By Theorem 4.6, R(θ) ≥ R. It remains to show that the series θ ∼P θk is uniquely determined by {fσ}σ∈Σ. In fact, it is uniquely determined by fσ0. Suppose θ′ ∼P θ′ is another series with R(θ′) ≥ R and suppose Zk(z)Lθk = fσ0(z) =Xk≥0 Zk(z)Lθ′ Xk≥0 k k r z) = Zk(ts0)Lθk =Xk≥0 for all z ∈ D(0, R, σ0). However, as we have seen from our analysis that yielded (4.5), Zk(ts0)Lθk = tkTθk ⊗ IHπ . Thus we conclude that for 0 ≤ t < min{1, R(θ)}, tkTθk ⊗ IHπ =Xk≥0 Xk≥0 where all the series converge in the operator norm on B(F (E)⊗π Hπ). Since the map X 7→ πF (E)(X), X ∈ L(F (E)), is a faithful normal representation of L(F (E)), Xk≥0 tkTθk =Xk≥0 =Xk≥0 Zk(ts0)Lθ′ k ⊗ IHπ , tkTθ′ tkTθ′ k k as norm-convergent series in H ∞(E). So, if we apply Φk to both sides, we conclude that tkTθk = tkTθ′ k for every k and θ = θ′. (cid:3) for every k. Hence θk = θ′ k TENSORIAL FUNCTION THEORY 26 represents a bounded holomorphic function fσ on D(0, 1, σ) for a par- As we noted in Remark 2.12, it can happen that a seriesPk≥0 Zk(z)Lθk ticular σ and R(θ) ≥ 1, θ ∼P θk, but fσ is not a σ-Berezin transform. However, the following proposition shows that if fσ is a member of a matricial family of functions f = {fσ}σ∈Σ that are uniformly bounded in σ, then f is a Berezin transform. Theorem 4.8. Suppose Σ is an additive full subcategory of NRep(M) that contains a special generator for NRep(M) and suppose f = {fσ}σ∈Σ is an E, Σ-matricial family of functions, with fσ defined on D(0, 1, σ) and mapping to B(Hσ). Then f is a Berezin transform restricted to {D(0, 1, σ)}σ∈Σ if and only if (4.8) sup{kfσ(z)k σ ∈ Σ, z ∈ D(0, 1, σ)} < ∞. Proof. If there is an F ∈ H ∞(E) such that fσ = bFσ for all σ, then certainly kfσ(z)k ≤ kFk for all σ and z. So we shall attend to the converse and suppose sup{kfσ(z)k σ ∈ NRepO(M), z ∈ D(0, 1, σ)} = A < ∞. If 0 < r < 1, then as we saw in the proof of Theorem 4.7, there is an Fr ∈ H ∞(E) such that fσ0(rs0) = πF (E)(Fr) and fσ(z) =cFr σ( 1 r z) for all kzk < r and all σ (see (4.3)). Also, it follows from Theorem 4.7 that f is a tensorial power series {Pk≥0 Zk(·)Lθk z ∈ D(0, 1, σ)} where the series θ ∼ Pk≥0 θk in T+((E)) has R(θ) ≥ 1. For 0 < t < r < 1, we thus have s0) = fσ0(ts0) =Xk≥0 cFrσ0( Zk(ts0)Lθk =Xk≥0 tkTθk ⊗ IHπ t r and it follows that for every integer m ≥ 0, for every 0 < r < 1. Note that, for every 0 < r < 1, Fr = πF (E)(Fr) = fσ0(rs0) ≤ A. Thus {Fr} is a bounded set. If rn ր 1 and if F is an ultraweak limit point of {Frn}, say F = limα Frnα for an appropriate subnet of {rn}, then for every m ≥ 0 we have Φm(πF (E)(F )) = lim Φm(πF (E)(Frnα )) = lim Φm(dFrnα σ0 (s0)) = lim rm nαTξm⊗π IHπ = Tξm⊗π IHπ. It follows that Φm(F ) = Tθm and, using Therefore t r s0)) = tmTθk ⊗ IHπ. Φm(cFrσ0( Φm(cFrσ0(s0)) = rmTθk ⊗ IHπ TENSORIAL FUNCTION THEORY 27 Proposition 2.11, we have, for every σ ∈ Σ and every z ∈ D(0, 1, σ), fσ(z) =Xk≥0 Zk(z)Lθk = bFσ(z). (cid:3) Thus fσ = bFσ. In the following theorem we want to consider two W ∗-correspondences over M, E and F , and relate maps between their families of absolutely continuous representations to homomorphisms between their Hardy al- gebras. Theorem 4.9. Suppose Σ is a full additive subcategory of NRep(M) that contains a special generator for NRep(M). Suppose also that E and F are two W ∗-correspondences over M and that f = {fσ}σ∈Σ is a family of maps, with fσ : AC(σ, E) → AC(σ, F ). Then f is an E, F, Σ-matricial family of maps if and only if there is an ultraweakly continuous homomorphism α : H ∞(F ) → H ∞(E) such that for every z ∈ AC(σ, E) and every Y ∈ H ∞(F ), (4.9) Proof. Suppose that f = {fσ}σ∈Σ is a matricial family of maps and let Y be an element of H ∞(F ). For each σ ∈ Σ, define gσ : AC(σ, E) → B(Hσ) by gσ(z) = bY (fσ(z)). If z ∈ AC(σ, E), if w ∈ AC(τ, E), and if C ∈ I(σ× z, τ × w), then since f is assumed to be a matricial family of maps, C ∈ I(σ× fσ(z), τ × fτ (w)). By Theorem 4.4, Cgσ(z) = gτ (w)C. Thus, by Theorem 4.4 again, there is an operator α(Y ) in H ∞(E) such that [α(Y )(z) = gσ(z) for all z ∈ AC(σ, E). Thus equation 4.9 is satisfied. Note that α(Y ) is uniquely determined by virtue of the uniqueness assertion in Theorem 4.7. Since [α(Y )(z) = bY (fσ(z)). (4.10) [α(Y )(z) = gσ(z) = bY (fσ(z)), it is clear that α is a homomorphism. It remains to prove that α is ultraweakly continuous. For that purpose, let σ be the special generator σ0 and let z = s0 in (4.10), to conclude that α(Y ) ⊗ IHπ = bY (fσ0(s0)), from which it follows immediately that α is ultraweakly continuous. For the converse, suppose the family f implements a homomorphism α via equation (4.10). To show that f is an E, F, Σ-matricial family of maps, we must show that the family preserves intertwiners in the sense of equation (3.1). So let C ∈ I(σ × z, τ × w) and apply Theorem 4.4 to conclude that CbY (fσ(z)) = C [α(Y )(z) = [α(Y )(w)C = bY (fτ (w)) TENSORIAL FUNCTION THEORY 28 ξ = fτ (w)Lτ (cid:3) ξ h := ξ ⊗ h and if we define Lτ ξ C = fτ (w)(IE ⊗ C)Lσ for all Y ∈ H ∞(F ). In particular, this holds for Y = Tξ for every ξ ∈ F . So, if we write Lσ ξ for the map from Hσ → F ⊗σ Hσ defined by the equation Lσ ξ similarly, then for ξ and bY (fτ (w)) = fτ (w)Lτ Y = Tξ we conclude that bY (fσ(z)) = fσ(z)Lσ ξ . Thus Cfσ(z)Lσ ξ . Since this equation holds for all ξ ∈ F , we conclude that Cfσ(z) = fτ (w)(IE ⊗ C). Thus C ∈ I(fσ(z), fτ (w)). Remark 4.10. Recall that an element z ∈ Eσ∗ lies in the center of Eσ∗ in case b · z = z · b for all b ∈ σ(M)′ [MS08, Definition 4.11], in which case we write z ∈ Z(Eσ∗). Thus z ∈ Z(Eσ∗) if and only if σ(M)′ ⊆ I(σ × z, σ × z). It follows that if {fσ}σ∈N Rep(M ) is a matricial family of maps, with fσ : AC(σ, E) → AC(σ, F ), then fσ(AC(σ, E) ∩ Z(Eσ∗)) ⊆ AC(σ, F ) ∩ Z(F σ∗). 5. Function Theory without a generator In this section we shall focus on additive subcategories Σ of NRep(M) that do not necessarily contain a special generator and address the problem of deciding which matricial families of functions {fσ}σ∈Σ have tensorial power series representations as in Proposition 2.8. In partic- ular, we shall prove the following theorem which complements and is something of a converse to Proposition 3.5. Theorem 5.1. Suppose Σ is an additive subcategory of NRep(M) and that D(ζ, R) is a matricial disc determined by an additive field ζ = {ζσ}σ∈Σ on Σ and R, 0 < R ≤ ∞. Suppose, also, that f = {fσ}σ∈Σ is a matricial family of functions defined on D(ζ, R) that is locally uniformly bounded on D(ζ, R) in the sense that for each r < R, sup σ∈Σ z∈D(ζσ,r)kfσ(z)k < ∞. sup Then: (1) Each fσ is Frechet analytic on D(ζσ, R). (2) If the subcategory is full, if the additive field ζ is also central, and if each σ ∈ Σ is faithful, then f is a family of tensorial power series {Pk≥0 Z k(z − ζσ)Lθk z ∈ D(ζk, R, σ)}σ∈Σ, where θ ∼Pk≥0 θk has R(θ) ≥ R. To achieve this goal, we use the matrix analysis initiated by Tay- lor in [Tay72], and developed further in the work of Voiculescu in [Voi04, Voi10], in the work of Helton and his collaborators [HKMS09, TENSORIAL FUNCTION THEORY 29 HKM11a, HKM11b] and especially in the investigations of Kaliuzhnyi- Verbovetskyi and Vinnikov [KVV09, KV10, KVV06]. Indeed, the first half of Theorem 5.1 is proved by showing that Taylor's matrix analysis leads to one of Taylor's Taylor series for each fσ. The existence of such a series implies that fσ is Frechet analytic. The second half of the the- orem involves showing that each of Taylor's Taylor coefficients comes from an element of a suitable tensor power of E⊗k. Throughout this section, Σ will denote a fixed additive subcategory of NRep(M). No other assumptions will be placed on Σ, except as explicitly stated in added hypotheses in the statements of results. In particular, we do not assume that Σ is full, nor do we assume that Σ contains a special generator. The matricial sets we will consider will primarily be matricial discs D(ζ, R) = {D(ζσ, R, σ)}σ∈Σ, where ζ is an additive field of vectors on Σ (see Example 3.2). Note, in particular, that the assumption that ζ is an additive field on Σ guarantees that ζmσ = ζσ ⊕ ζσ ⊕ · · · ⊕ ζσ (m-summands). The following result can be found in [KV] for the case when M = C. Lemma 5.2. Let f = {fσ}σ∈Σ be a matricial family of functions de- fined on a matricial E, Σ-family {U(σ)}σ∈Σ where Σ is an additive subcategory of NRep(M). Suppose σ, τ ∈ Σ, z ∈ U(σ), w ∈ U(σ) and u ∈ I(τ E ◦ ϕ, σ) are such that (cid:18) z u 0 w (cid:19) ∈ U(σ ⊕ τ ). Then there is an operator ∆fσ,τ (z, w)(u) ∈ B(Hτ , Hσ) such that (1) fσ⊕τ ((cid:18) z u 0 w (cid:19)) =(cid:18) fσ(z) ∆fσ,τ (z, w)(u) fτ (w) 0 (cid:19) . (2) If b ∈ σ(M)′ is such that z(IE⊗b) = bz and such that(cid:18) z 0 w (cid:19) ∈ U(σ⊕τ ), then fσ(z)b = bfσ(z) and ∆fσ,τ (z, w)(bu) = b∆fσ,τ (z, w)(u). In particular ∆fσ,τ (z, w)(tu) = t∆fσ,τ (z, w)(u) for t ∈ C such that bu (cid:18) z 0 w (cid:19) ∈ U(σ ⊕ τ ). tu (3) If b ∈ τ (M)′ is such that w(IE ⊗ b) = bw and such that (cid:18) z u(IE ⊗ b) w 0 (cid:19) ∈ U(σ ⊕ τ ), then fτ (w)b = bfτ (w) and ∆fσ,τ (z, w)(u(IE⊗b)) = ∆fσ,τ (z, w)(u)b. TENSORIAL FUNCTION THEORY 30 (4) If σ = τ , z ∈ U(σ) and b ∈ σ(M)′ such that (cid:18) z 0 z + bu (cid:19) ∈ bu U(σ ⊕ σ), then fσ(z + bu) − fσ(z) = ∆fσ,σ(z, z + bu)(bu) and, if t ∈ C, fσ(z + tu) − fσ(z) = t∆fσ,σ(z, z + tu)(u). Remark 5.3. It should be emphasized that in (1), ∆fσ,τ (z, w)(u) repre- sents an operator in B(Hτ , Hσ), whose full dependence on z, w and u has still to be determined. Among other things, part (2) of the lemma proves that ∆fσ,τ (z, w)(u) extends to be homogeneous in u of degree one, but the additivity in u will be proved later in Lemma 5.9. It should also be emphasized that to prove that ∆fσ,τ (z, w)(u) extends to be ho- mogeneous in u the full force of the assumption that f is matricial is not used: We may assume that σ = τ and that the intertwiners in- volved are operator matrices whose entries are scalar multiples of the identity (on Hσ). These are present in any of the categories we use, as was stated at the outset of Section 2. Proof. Write Iσ for the identity operator on Hσ. Then (5.1) and (5.2) (5.3) and (5.4) (cid:19) =(cid:18) Iσ 0 (cid:18) z u 0 w (cid:19)(cid:18) IE ⊗ Iσ 0 (cid:19) z (cid:0) 0 Iτ (cid:1)(cid:18) z u 0 w (cid:19) = w(cid:0) 0 IE ⊗ Iτ (cid:1) . fσ⊕τ ((cid:18) z u 0 w (cid:19))(cid:18) Iσ 0 (cid:19) fσ(z) (cid:0) 0 Iτ (cid:1) fσ⊕τ ((cid:18) z u 0 w (cid:19)) = fτ (w)(cid:0) 0 Iτ (cid:1) . a21 a22(cid:21) , then (5.3) shows that a11 = 0 w(cid:21)) as(cid:20)a11 a12 0 (cid:19) =(cid:18) Iσ Since f preserves intertwiners, we may write So, if we write fσ⊕τ ((cid:20)z u fσ(z) and a21 = 0, while (5.4) also shows that a21 = 0 as well as a22 = fτ (w). The remaining entry, a12, is taken as the definition of ∆fσ,τ (z, w)(u). This proves (1). TENSORIAL FUNCTION THEORY 31 For (2) note that the equality (cid:18) b 0 Iτ (cid:19)(cid:18) z u 0 w (cid:19) =(cid:18) z 0 w (cid:19)(cid:18) IE ⊗ b bu 0 0 0 IE ⊗ Iτ (cid:19) implies (cid:18) b 0 Iτ (cid:19)(cid:18) fσ(z) ∆fσ,τ (z, w)(u) fτ (w) 0 0 (cid:19) =(cid:18) fσ(z) ∆fσ,τ (z, w)(bu) fτ (w) 0 (cid:19)(cid:18) b 0 Iτ (cid:19) , 0 proving (2). The proof of (3) is similar using the equality (cid:18) Iσ 0 b (cid:19)(cid:18) z u(IE ⊗ b) w 0 0 (cid:19) =(cid:18) z u 0 w (cid:19)(cid:18) IE ⊗ Iσ 0 0 IE ⊗ b (cid:19) . For (4), simply note that (cid:0) −Iσ Iτ (cid:1)(cid:18) z 0 z + bu (cid:19) = z(cid:0) −IE ⊗ Iσ bu IE ⊗ Iτ (cid:1) (cid:3) and use the properties of f . Even though the linearity in u of the operator ∆fσ,τ (z, w)(u) has still to be shown, ∆fσ,τ (z, w) can be viewed as a noncommutative difference operator with values in B(Hτ , Hσ). Definition 5.4. With the hypotheses and notation as in Lemma 5.2, we call ∆fσ,τ (z, w) the Taylor difference operator determined by f and the points z and w. If σ = τ and z = w, we call ∆fσ,σ(z, z) the Taylor derivative of f at z and denote it ∆fσ(z, z) or ∆fσ(z). Even though its linearity in u has still to be shown, we can also define non commutative difference operators of higher order. We will need these first. So to this end, note that by applying part (1) of Lemma 5.2 repeatedly one finds that for every σ0, . . . , σn in Σ, for every zi ∈ U(σi) (0 ≤ i ≤ n) and for every uj ∈ I(σE j ◦ ϕ, σj−1) (1 ≤ j ≤ n), the matrix u1 z0 (5.5) fσ0⊕σ1⊕···⊕σn(   z1 . . . 0 ... ... 0 · · · 0 ... 0 · · · . . . . . . zn−1 un zn 0   ) 0 . . . . . . . . . · · · TENSORIAL FUNCTION THEORY 32 has a block upper triangular form with fσ0(z0), . . . , fσn(zn) on the main diagonal, assuming of course that the argument is in U(σ0 ⊕ σ1 ⊕ · · ·⊕ σn), i.e., (5.6) fσ0⊕σ1⊕···⊕σn(   = z0 u1 z1 . . . 0 ... ... 0 · · ·  fσ0(z0)  0 ... ... 0 0 . . . . . . . . . · · · 0 ... 0 · · · . . . . . . zn−1 un zn a02 fσ1(z1) a12 . . . a01 0   ) · · · . . . . . . a0n a1n ... fσn−1(zn−1) an−1n fσn(zn) · · · · · ·   Definition 5.5. The function of u1, u2, · · · , un defined by the 0, n entry of the right hand side of Equation 5.6 will be called the nth- order Taylor difference operator determined by z0, z1,· · · ,zn, and will be denoted ∆nfσ0,σ1,··· ,σn(z0, . . . , zn). If z0 = z1 = · · · = zn = z, we call ∆nfσ,σ,··· ,σ(z, z,· · · , z) := ∆nfσ(z) the nth-order Taylor derivative of fσ at z. The arguments for the proofs of parts (2) and (3) of Lemma 5.2 can be adapted easily to prove the following lemma. We omit the details. Lemma 5.6. Suppose {zi} and {uj} are as above. Fix 1 ≤ k ≤ n + 1 and choose b ∈ σk−1(M)′ so that bzk−1 = zk−1(IE ⊗ b) (in particular, b can be in C), then: (1) If 1 < k ≤ n and if both ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , uk−1, buk, . . . , un) and ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , uk−1(IE⊗b), uk, . . . , un) are well defined, in the sense that the argument matrices in the expression 5.5 lie in U(σ0 ⊕ σ1 ⊕ · · · ⊕ σn), then ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , uk−1, buk, . . . , un) = ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , uk−1(IE ⊗ b), uk, . . . , un). TENSORIAL FUNCTION THEORY 33 (2) If k = 1 and if ∆nfσ0,σ1,··· ,σnf (z0, . . . , zn)(bu1, . . . , un) is well defined, then ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(bu1, . . . , un) = b∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un) . (3) If k = n + 1 and if ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un(IE ⊗ b)) is well defined, then ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un(IE ⊗ b)) = ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un)b. Remark 5.7. Again we want to note that the argument for Lemma 5.6 shows that if we take the σis to be one and the same σ, so that the matrix   z0 0 ... ... 0 u1 z1 . . . · · · 0 . . . . . . . . . · · · 0 ... 0 · · · . . . . . . zn−1 un zn 0   acts from E ⊗σ Hmσ to Hmσ, then the only matrices necessary to show that ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un) extends to be homogeneous in each of the uis are block matrices whose entries are scalar multiples of the identity on Hσ. Lemma 5.8. Given σ0, . . . , σn in Σ, zi ∈ U(σi) (0 ≤ i ≤ n) and uj ∈ I(σE j ◦ ϕ, σj−1) (1 ≤ j ≤ n), then for 0 ≤ i < j ≤ n, the i, j entry, aij of the right hand side of Equation (5.6) is ai,j = ∆jfσ0,σ1,··· ,σj (z0, . . . , zj)(u1, . . . , uj). Proof. The proof proceeds by induction. For n = 1, the assertion is Lemma 5.2(1). So assume it holds for n and write σ = ⊕n+1 i=0 σi. We apply Lemma 5.2(1) repeatedly. If we partition σ as σ = (⊕n i=0σi) ⊕ σn+1, Lemma 5.2 and the induction hypothesis prove the lemma for all j ≤ n and for i = j = n + 1. To obtain the formula for ai,n+1, simply write σ as σ = (⊕i−2 k=i−1σk), then apply Lemma 5.2 and the induction assumption. (cid:3) k=0σk) ⊕ (⊕n+1 The argument used in the proof of the following lemma was shown to us by Victor Vinnikov. It will appear in his joint work with D. S. Kaliuzhnyi-Verbovetskyi [KVV]. TENSORIAL FUNCTION THEORY 34 Lemma 5.9. Let f = {fσ}σ∈Σ be a matricial family of functions de- fined on a matricial E, Σ-family {U(σ)}σ∈Σ where Σ is an additive sub- category of NRep(M). Suppose σ, τ ∈ Σ, z ∈ U(σ), w ∈ U(σ) and ui ∈ w (cid:19) ∈ U(σ⊕ τ ), I(τ E ◦ ϕ, σ) for i = 1, 2 are such that A :=(cid:18) z u1 + u2 B :=    ∈ U(2σ⊕τ ).   ∈ U(2σ⊕τ ) and C :=  0 z 0 u2 0 z u1 0 0 w z 0 u1 0 z u2 0 0 w Then ∆fσ,τ (z, w)(u1 + u2) = ∆fσ,τ (z, w)(u1) + ∆f (z, w)(u2). Proof. Considering f2σ⊕τ (B) and f2σ⊕τ (C) and using Lemma 5.2(1) with 2σ ⊕ τ split as σ ⊕ (σ ⊕ τ ) we find that these matrices have the following from: and f2σ⊕τ (B) =  f2σ⊕τ (C) =  fσ(z) x y 0 0 fσ(z) 0 0 fσ(z) ∆fσ,τ (z, w)(u2) 0 u fτ (w) v fσ(z) ∆fσ,τ (z, w)(u1) 0 fτ (w)       . for some x, y, u, and v. Writing S for the 3 × 3 permutation matrix associated with the transposition (1, 2), we see that C = SBS−1 = SBS. Thus f (C) = Sf (B)S and, therefore, y = ∆fσ,τ (z, w)(u1) and f2σ⊕τ (B) =  f (z) 0 0 x ∆fσ,τ (z, w)(u1) f (z) ∆fσ,τ (z, w)(u2) 0 fτ (w) Now write D =(cid:18) I fσ⊕τ (A)D. Since fσ⊕τ (A) = (cid:18) fσ(z) ∆fσ,τ (z, w)(u1 + u2) 0 0 I (cid:19). Then DB = AD and, thus Df2σ⊕τ (B) = (cid:19), we may fτ (w) I 0 0 equate the (1, 3) entries to obtain the result. (cid:3) Remark 5.10. Again, in Lemma 5.9, if we set τ = σ, we see that the only matrices used in the proof are matrices that are block scalar matrices. Combining Lemma 5.9 with Lemma 5.2(2), we conclude that ∆fσ,τ (z, w)(u) extends to be linear in u. We can now deduce the multilinearity of ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(·, . . . ,·) (with z0, . . . , zn fixed) from the linearity of ∆fσ,τ (z, w)(u) in u. In order TENSORIAL FUNCTION THEORY 35 to avoid specifying the conditions on the variables as in the statement of the lemma above, we restrict ourselves to matricial discs. Corollary 5.11. Let f = {fσ}σ∈Σ be a matricial family of functions, defined in a matricial disc D(ζ, r). Then for every σ0, . . . , σn in Σ and every zi ∈ D(ζσi, r, σi), (0 ≤ i ≤ n), the function ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un), n ◦ ϕ, σn−1). j ◦ ϕ, σj−1) (1 ≤ j ≤ n) with norm sufficiently 1 ◦ ϕ, σ0) × · · · × defined for uj ∈ I(σE small can be extended to a multilinear map on I(σE I(σE Proof. For n = 1, this is shown in Lemma 5.9. For the general case, fix 1 ≤ j ≤ n and write z0 u1   0 ... ... 0 z1 . . . · · · 0 . . . . . . . . . · · · 0 ... 0 · · · . . . . . . zn−1 un zn 0   =(cid:18) X Z 0 Y (cid:19) where X is a j×j block and Z is a j×(n−j) block with uj in the bottom left corner and all other entries 0. Applying fσ0⊕σ1⊕···⊕σn to this matrix we get, on one hand, a matrix with ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un) in the upper right corner and, on the other hand, a 2 × 2 block matrix that we shall simply write as ∆f (X, Y )(Z) in the (1, 2) corner. Using Lemma 5.9, we know that ∆f (X, Y )(Z) is linear in Z and, thus, linear in uj. Therefore ∆nfσ0,σ1,··· ,σn(z0, . . . , zn)(u1, . . . , un), being a corner, is linear in uj. This completes the proof. (cid:3) The following theorem is an analogue of Taylor's Taylor Theorem with remainder [Tay73, Proposition 4.2]. Theorem 5.12. Let f = {fσ}σ∈Σ be a matricial family of functions defined on a matricial disc D(ζ, r). Fix σ ∈ Σ and choose z and w in D(ζσ, r, σ) so that the matrix (5.7)   z . . . z w 0 . . . . . . . . . · · · 0 ... ... 0 · · · · · · . . . . . . z 0 0 ... 0 w z + w   TENSORIAL FUNCTION THEORY 36 lies in D(ζ(n+1)σ, r, (n + 1)σ). Then fσ(z+w) = n−1Xk=0 ∆kfσ(z)(w, . . . , w)+∆nfσ,σ,··· ,σ(z, . . . , z, z+w)(w, . . . , w). Proof. Denote the matrix (5.7) by A and note that, A   IE ⊗ Iσ ... ... ... IE ⊗ Iσ   =   Iσ ... ... ... Iσ   (z + w). Applying the intertwining property of f(n+1)σ and Lemma 5.8 completes the proof. (cid:3) We would like to pass to the limit as n → ∞, in Theorem 5.12, but to do this, we must impose an additional hypothesis on f , viz. f must be locally uniformly bounded in the sense of Theorem 5.1. The following theorem should be compared with Theorem 4.8. There, we assumed that the additive subcategory Σ is full and contains a special generator for NRep(M). Here we make no assumptions on Σ other than it is an additive subcategory of NRep(M). Theorem 5.13. Let f = {fσ}σ∈Σ be a matricial family of functions de- fined on a matricial disc D(ζ, r) and suppose that f is locally uniformly bounded in the sense of Theorem 5.1. Then: (1) Each fσ is Frechet differentiable in z, z ∈ D(ζσ, r, σ), and f ′ σ(z)(w) = ∆f (z)(w). (2) Each fσ may be expanded on D(ζσ, r, σ) as (5.8) fσ(ζσ + z) = ∞Xk=0 ∆kfσ(ζσ)(z, . . . , z), where the series converges absolutely and uniformly on every disc D(0, r0, σ) with r0 < r. Proof. To prove (1), let r0 < r and let M be greater than sup σ∈Σ sup z∈D(ζσ,r0)kfσ(z)k, TENSORIAL FUNCTION THEORY 37 which is finite by assumption. Fix z ∈ D(ζσ, r0, σ) and note that for w with norm sufficiently small (say, w ≤ s), the matrix (5.9) lies in D(ζ3σ, r0, 3σ) = D(ζσ ⊕ ζσ ⊕ ζσ, r0, 3σ). But this implies in particular, that   z w 0 z 0 0 z + w 0 w   (cid:20)z w z(cid:21) 0 lies in D(ζ2σ, r0, 2σ). Our boundedness assumption, then implies that for kwk ≤ s, k∆fσ(z)(w)k ≤ k(cid:20)fσ(z) ∆fσ(z)(w) fσ(z) 0 (cid:21) k = kf2σ((cid:20)z w z(cid:21))k ≤ M, 0 i.e., ∆fσ(z)(·) extends to be a bounded linear operator. Further, the same sort of argument shows that when we apply f3σ to the matrix (5.9) for kwk ≤ s, we have ∆2fσ,σ,σ(z, z, z+w)(w, w) ≤ M, by definition of ∆2f·,·,·(·,·,·)(·,·). But by Theorem 5.12, fσ(z+w)−fσ(z)−∆fσ(z)(w) = ∆2fσ,σ,σ(z, z, z+w)(w, w). So, if we write w0 for w w, then fσ(z+w)− s )2∆2fσ,σ,σ(z, z, z + w)(w0, w0) ≤ w2M fσ(z) − ∆fσ(z)(w) = ( w , which proves that s2 s 1 wfσ(z + w) − fσ(z) − ∆fσ(z)(w) −→ 0 as w → 0. Thus (1) is proved. To prove (2), fix 0 < r0 < r1 < r, write q := r1/r0 and let r2 := r1√1+q2 < r0. Then for every n and every z ∈ D(ζσ, r2, σ),  q(z − ζσ)  ζσ w :=  0 ... ... 0 ζσ . . . · · · · · · . . . . . . ζσ 0 0 . . . . . . . . . · · · 0 ... 0 q(z − ζσ) z  lies in D(ζ(n+1)σ, r1, (n + 1)σ) ⊆ D(ζ(n+1)σ, r, (n + 1)σ). Thus ∆nfσ,σ,··· ,σ(ζσ, . . . , ζσ, z)(q(z − ζσ), . . . , q(z − ζσ)) ≤ f (w) ≤ M. From Lemma 5.8, we conclude that ∆nfσ,σ,··· ,σ(ζσ, . . . , ζσ, z)(z − ζσ, . . . , z − ζσ) ≤ q−nM −→n→∞ 0. So Theorem 5.12 proves (5.8) for z ∈ D(ζσ, r2, σ). TENSORIAL FUNCTION THEORY 38 To pass to z in the larger disc D(ζσ, r0, σ), observe that for such z u :=   ζσ 0 ... ... 0 q(z − ζσ) ζσ . . . · · · · · · . . . . . . ζσ 0 0 . . . . . . . . . · · · 0 ... 0 q(z − ζσ) ζσ   lies in the disc D(ζ(n+1)σ, r1, (n + 1)σ) ⊆ D(ζ(n+1)σ, r, (n + 1)σ). Conse- quently ∆nfσ,σ,··· ,σ(ζσ, . . . , ζσ)(z − ζσ, . . . , z − ζσ) = q−n∆nfσ,σ,··· ,σ(ζσ, . . . , ζσ)(q(z − ζσ), . . . , q(z − ζσ)) ≤ q−nM. It follows that the series in Equation (5.8) converges absolutely and uniformly on D(0, r0). To see that the sum is fσ(ζσ + z), fix z ∈ D(0, r0, σ) and consider the function h(λ) := fσ(ζσ +λz). By part (1) h is analytic in a disc centered at the origin in the complex plane having radius bigger than 1. On the other hand we may also form g(λ) := ∞Xk=0 ∆kfσ,σ,··· ,σ(ζσ, . . . , ζσ)(λz, . . . , λz) = ∞Xk=0 λk∆kfσ,σ,··· ,σ(ζσ, . . . , ζσ)(z, . . . , z), which is also analytic in a disc centered at the origin of radius bigger than 1. By what we have just shown, these two Banach space-valued functions agree on a neighborhood of the origin in the complex plane. Therefore, they agree on the intersection of their domains of definition ([HP74, Theorem 3.11.5]), which includes 1, i.e., Equation (5.8) is valid throughout the disc D(0, r0, σ). (cid:3) Remark 5.14. As a special case of Theorem 5.13, we obtain a formula that was inspired by [KV]. If σ is a normal representation of M, then the subcategory Σ generated by σ is just the collection of finite mul- tiples of σ. The collection of morphisms from mσ to nσ is just the n× m matrices over σ(M)′. A ζ0 ∈ Eσ∗ generates an additive field over Σ simply by setting ζkσ = ζ0 ⊕ ζ0 ⊕ · · · ⊕ ζ0 (k summands). Then, if f = {fmσ}m∈N is a locally bounded, matricial function on the matricial disc D(ζ, r), ζ = {ζmσ}m∈N, then Theorem 5.8 implies that for every TENSORIAL FUNCTION THEORY 39 m ∈ N and every z in D(ζmσ, r, mσ), ∞Xk=0 fmσ(z) = ∆kfmσ(ζmσ)(z − ζmσ, . . . , z − ζmσ). Suppose f = {fσ}σ∈Σ is a matricial function defined on a matricial disc D(ζ, r) and suppose that f is locally uniformly bounded as in Theorems (5.1) and (5.8). Then we have just seen that each fσ is Frechet differ- entiable throughout D(ζσ, r, σ). By [HP74, Theorem 26.3.10] fσ can be expanded in a unique power series about each point z ∈ D(ζσ, r, σ), and the series converges at least in the largest open ball centered at z con- tained in D(ζσ, r, σ). The terms of the power series are built from the higher order Frechet derivatives of fσ. Recall from the general theory of differentiable functions on Banach spaces (applied to our setting) that the nth order Frechet derivative of fσ is a B(Hσ)-valued function, de- noted Dnfσ, that is defined on D(ζσ, r, σ)×Eσ∗×Eσ∗×· · ·×Eσ∗ and has the following properties: For each z, Dnfσ(z)(u1, . . . , un) is a bounded, symmetric, multilinear function of (u1, . . . , un) ∈ Eσ∗×Eσ∗×· · ·×Eσ∗, where the norm kDnfσ(z)(·, . . . ,·)k is locally bounded in z ∈ D(0, r, σ) ([HP74, Theorem 26.3.5]). If we write Dnfσ(z)(u) for Dnfσ(z)(u, . . . , u), then Dnfσ(z)(u) is homogeneous of degree n in u and for each a ∈ D(ζσ, r, σ) there is an r′, depending on a, such that (5.10) fσ(z + u) = ∞Xn=0 1 n! Dnfσ(z)(u), with the convergence uniform for z − a < r′ and u < r′ ([HP74, Theorem 3.17.1]). When z = ζσ we find that (5.11) fσ(ζσ + u) = ∞Xn=0 1 n! Dnfσ(ζσ)(u) Since each of the summands in equations (5.8) and (5.11) is homoge- neous, we conclude that ∆kfσ(ζσ)(z) = 1 k! Dkfσ(ζσ)(z) for all k and z for which the left hand side is well defined. We may therefore summarize our analysis as follows. Corollary 5.15. Let Σ be an additive subcategory of NRep(M), let f = {fσ}σ∈Σ be a matricial family of functions defined on a matricial disc D(ζ, r), and suppose that f is locally uniformly bounded on D(ζ, r). TENSORIAL FUNCTION THEORY 40 Then, for every σ ∈ Σ, every k ≥ 0, and every z ∈ D(0, r, σ), Taylor's Taylor derivatives and the Frechet derivatives of fσ are related by the equation (5.12) ∆kfσ(ζσ)(z, . . . , z) = 1 k! Dkf (ζσ)(z), and so ∆kfσ(ζσ)(z) is the restriction to the diagonal of D(ζσ, r, σ)k of a bounded, symmetric, k-linear map on Eσ∗ × Eσ∗ · · · × Eσ∗. Conse- quently, we may write fσ(ζσ + z) = = ∞Xk=0 ∞Xk=0 ∆kfσ(ζσ)(z) 1 k! Dkf (ζσ)(z) Remark 5.16. At this point we pause to take stock of what we have proved, and with what hypotheses. First of all, we have proved the first assertion of Theorem 5.1. Further, we identified the higher Frechet derivatives of f with Taylor's Taylor derivatives (Equation (5.12)). We did this with the minimal hypotheses that we have placed on our addi- tive categories, vis., that they contain the natural injections and pro- jections for finite direct sums. To obtain the second assertion of Theorem 5.1, we will need the as- sumptions stated there. Even though in some of the results to follow we can we can get by with slightly weaker assumptions, for the remain- der of this section, we shall assume: Σ is full; every σ in Σ is faithful; and ζ is central. The assumption that ζ is central allows us to assume that ζσ = 0 for each σ. That is, we can translate the disc D(ζ, R) to D(0, R) whenever ζ is central. Lemma 5.17. Suppose Σ is a full additive subcategory of NRep(M), and suppose that ζ = {ζσ}σ∈Σ is a central additive field on Σ. Suppose, also, that f = {fσ}σ∈Σ is a matricial function define on the matricial disc D(ζ, R), 0 < R ≤ ∞, and define g = {gσ}σ∈Σ by setting gσ(z) := fσ(z+ζσ). Then g is a matricial function on D(0, R) and for all positive integers k, ∆kfσ(ζσ)(z − ζσ) = ∆kgσ(0)(z − ζσ). Proof. The fact that g is matricial is immediate from the observation that when ζ is central, I(σ × (z + ζσ), τ × (w + ζτ )) = I(σ × z, τ × w). TENSORIAL FUNCTION THEORY 41 The equation for Taylor's derivatives is immediate from their definition, Definition 5.5. (cid:3) We want to show that this symmetric multilinear map is the restriction of a completely bounded map defined on the tensor product Eσ∗ ⊗ Eσ∗ · · · ⊗ Eσ∗, balanced over σ(M)′. To accomplish this, we need the following technical lemma. Lemma 5.18. Suppose f = {fσ}σ∈Σ is a matricial function defined on a matricial disc centered at the origin D(0, r) = {D(0, r, σ}σ∈Σ. Suppose that p and k are positive integers and that for 1 ≤ j ≤ k, integers l(j) and m(j) are defined and satisfy 1 ≤ l(1) < m(1), l(2) < m(2), l(3) < · · · < m(k) ≤ p. Suppose also that u1, . . . , uk are in Eσ∗ and let U be the p × p matrix over Eσ∗ (viewed as an element of Epσ∗) whose (l(j), m(j)) entry is uj, for all j, 1 ≤ j ≤ k, and whose other entries are all zero. Then A := fpσ(U), viewed as a p × p matrix over B(Hσ), will have non zero entries only in positions (l(j), m(j + s)) where m(j) = l(j + 1), m(j + 1) = l(j + 2), . . . , m(j + s − 1) = l(j + s). In these positions we will have Al(j),m(j+s) = ∆s+1fσ(0)(uj, . . . , uj+s). Proof. Suppose first that, for every j, m(j) = l(j) + 1. Then the non zero entries of U are all on the first diagonal above the main one and the result follows easily from Lemma 5.8 (noting that ∆rf (0)(z1, . . . , zr) = 0 if one of the zi is 0). For the general case, let θ be a permutation on {1, . . . , p} such that θ(l(1)) = 1, θ(m(j)) = θ(l(j)) + 1 for all j and θ(l(j + 1)) = θ(m(j)) + m(j) − l(j + 1) for all 1 ≤ j ≤ p − 1. Such θ always exists (but, in general, is not unique). Write S for the permutation matrix associated with θ and consider U′ = SU(I ⊗ S−1). This matrix will have non zero entries only on the first diagonal above the main diagonal and, thus, f (U′) = SAS−1 will be of the form described above and it will follow that A satisfies the assertion of the lemma. (cid:3) Recall that if V1, V2,· · · , Vk, and X are operator spaces and if ϕ : V1 × V2 × · · · × Vk → X is a multilinear map, then one writes ϕ(n) for the multilinear map ϕ(n) : Mn(V1) × · · · × Mn(Vk) → Mn(X) defined by ϕ(n)((α1 ⊗ v1),· · · , (αk ⊗ vk)) = α1α2 · · · αk ⊗ ϕ(v1, v2, . . . , vk) TENSORIAL FUNCTION THEORY 42 for αi ∈ Mn(C) and vi ∈ Vi [ER00, Section 9.1]. Lemma 5.19. Let f = {fσ}σ∈Σ be a matricial family of functions defined on a matricial E, Σ-family of discs {D(0, r, σ)}σ∈Σ, where Σ is a full additive subcategory of NRep(M). For σ ∈ Σ, write ϕ : Eσ∗ × Eσ∗ · · · × Eσ∗ → B(Hσ) for the map ϕ(u1, . . . , uk) = ∆kfσ(0)(u1, . . . , uk), ui ∈ Eσ∗, i = 1, 2,· · · , k. Then, when Mn(Eσ∗) is identified with Enσ∗, we have ϕ(n)(U1, . . . , Uk) = ∆kfσ(0)(U1, . . . , Uk) (5.13) for Ui ∈ Enσ∗. Proof. Since the functions on both sides of (5.13) are k-linear, it suf- fices to prove the lemma for matrices Ui that have only one non zero entry. If U is such a matrix and the only non zero entry is u, which lies in the (i, j) position, we write U = εi,j ⊗ u. So, we write Uj = εr(j),s(j)⊗uj and, using the definition of ϕ(n), we have ϕ(n)(U1, . . . , Uk) = εr(1),s(1) · · · εr(k),s(k) ⊗ ∆kfσ(0)(u1, . . . , uk). Thus ϕ(n)(U1, . . . , Uk) = εr(1),s(k) ⊗ ∆kfσ(0)(u1, . . . , uk) provided s(j) = r(j + 1) for all 1 ≤ j ≤ k−1, and it equals 0 otherwise. In order to compute the right hand side of (5.13), we form the matrix B :=   0 U1 0 . . . 0 ... ... 0 · · · 0 . . . . . . . . . · · · 0 ... 0 · · · . . . . . . 0 Uk 0 0   , write A for the (k + 1) × (k + 1) matrix f(k+1)nσ(B) and note that ∆kfnσ(0)(U1, . . . , Uk) is the n × n block of A in the (1, k + 1) position. If we now view A as a matrix of size n(k+1)×n(k+1) (over B(Hσ)), we see that A satisfies the assumptions of Lemma 5.18, with p = n(k + 1), l(1) = r(1), m(1) = n + s(1), l(2) = n + r(2) etc. It follows from that lemma that the only non zero entry in the upper-right n× n block can be in the (l(1), m(k)) position and this will be non zero only if m(1) = l(2), m(2) = l(3), . . . , m(k − 1) = l(k). Using our notation here, this entry will be non zero if and only if s(j) = r(j + 1) for all 1 ≤ j ≤ k − 1. If this is the case, then by Lemma 5.18, this entry will be ∆kfσ(0)(u1, . . . , uk). This completes the proof. (cid:3) TENSORIAL FUNCTION THEORY 43 Proposition 5.20. Let f = {fσ}σ∈Σ be a matricial family of functions defined on a matricial disc D(0, r) = {D(0, r, σ)}σ∈Σ where Σ is a full, additive subcategory of NRep(M). Suppose f is uniformly bounded in norm by M so that, for every σ ∈ Σ, and for every z ∈ D(0, r, σ), fσ(z) ≤ M. Then, for every k, the map (5.14) is a k-linear map, balanced over σ(M)′, and is completely bounded, with ∆kfσ(0)(·,· · · ,·) : Eσ∗ × Eσ∗ · · · × Eσ∗ → B(Hσ) ∆kfσ(0)(·,· · · ,·)cb ≤ M rk . Proof. With the notation preceding Lemma 5.19, we have ∆kfσ(0)(·,· · · ,·)cb = sup{∆kfσ(0)(n) : n ≥ 1}. Note that in [ER00] this norm is denoted · mb but we follow the notation in [BLM04] and in other places in the literature. Using Lemma 5.19, it suffices to show that ∆kfσ(0)(·,· · · ,·) ≤ M rk . For this, consider ui ∈ D(0, 1, σ) and write u′  0 ) ≤ M. i := rui ∈ D(0, r, σ). Then u′ 1 0 . . . k) ≤ f(k+1)σ( ∆kfσ(0)(u′ 1, . . . , u′   0 ... ... 0 · · · · · · . . . . . . 0 0 0 . . . . . . . . . · · · 0 ... 0 u′ k 0  By the k-linearity of the map, we get ∆kfσ(0)(u1, . . . , uk) ≤ M rk , proving the complete norm estimate. The only thing left to prove is the fact that the map is balanced, but this follows from Lemma 5.6. (cid:3) Recall that, in the discussion preceding Lemma 2.5, Zk is defined by the formula, Zk(z) := z(k) = z(IE ⊗ z)· · · (IE⊗k−1 ⊗ z), for z ∈ Eσ∗. For the statement of the following theorem it will be convenient to use the natural extension of Zk to Eσ∗ × Eσ∗ × · · · × Eσ∗ and write Zk(u1, . . . , uk) := u1(IE ⊗ u2)· · · (IE⊗k−1 ⊗ uk), for u1, . . . , uk in Eσ∗. Theorem 5.21. Let f = {fσ}σ∈Σ be a matricial family of functions defined on a matricial disc D(0, r) over a full, additive subcategory Σ of faithful representations in NRep(M). Suppose that f is defined and bounded uniformly M on D(0, r). Then, for every k, there is a unique θk ∈ E⊗k, with θk ≤ M (5.15) ∆kfσ(0)(u1, . . . , uk)h = Zk(u1, . . . , uk)(θk ⊗ h) rk , such that TENSORIAL FUNCTION THEORY 44 for every u1, . . . , uk in Eσ∗ and every h ∈ Hσ. Proof. Consider the map ψ : Eσ × Eσ × · · · × Eσ → B(H) defined by ψ(η1, . . . , ηk) = ∆kfσ(0)(η∗ k, . . . , η∗ 1)∗. k, . . . , η∗ k, . . . , η∗ 1(I ⊗ b∗))∗ = (c∗∆kf (0)(η∗ For simplicity, we write here N for the von Neumann algebra σ(M)′. Using Proposition 5.20, we see that this is a k-linear map, balanced over σ(M)′, with norm not exceeding M rk . Applying [BMP00, Theorem 2.3], we find that it induces a linear, completely bounded map Ψ : Eσ ⊗hN Eσ ⊗hN · · · ⊗hN Eσ → B(Hσ), where ⊗hN is the module Haagerup tensor product, and Ψ ≤ M rk . But Eσ ⊗hN Eσ ⊗hN · · · ⊗hN Eσ = Eσ ⊗C ∗ Eσ ⊗C ∗ · · · ⊗C ∗ Eσ where ⊗C ∗ is the internal tensor product of C ∗-correspondences (see [Ble97, Theorem 4.3] ). For b, c ∈ N and η1, . . . , ηk ∈ Eσ, we have Ψ(b·η1⊗η2⊗· · ·⊗ηk·c) = Ψ((IEσ ⊗b)η1⊗η2⊗ · · ·⊗ ηkc) = ∆kfσ(0)(c∗η∗ 1)(I ⊗ b∗))∗ = bΨ(η1⊗η2⊗· · ·⊗ηk)c (using Lemma 5.6). Thus, Ψ is a bimodule map. Write F for the C ∗-correspondence Eσ ⊗C ∗ Eσ ⊗C ∗ · · · ⊗C ∗ Eσ (over N = σ(M)′). Using the terminology of [MS98], we say that (Ψ, ι) is a completely bounded covariant representation of F where ι is the identity representation of σ(M)′ on Hσ. It follows from [MS98, Lemma 3.5] that there is a bounded map Ψ : F ⊗ι Hσ → Hσ such that Ψ(ϕF (·) ⊗ IH) = ι(·) Ψ and Ψ ≤ M rk . Now note that (Eσ)⊗k is the self-dual completion of F and, using Remark 1.8 in [Vis11], we have F ⊗ι Hσ = (Eσ)⊗k⊗ι Hσ. Thus we can view Ψ as a map from (Eσ)⊗k⊗ι Hσ into Hσ satisfying Ψ(ϕ(·)⊗IH) = ι(·) Ψ. Applying [MS04, Theorem 3.6 and Lemma 3.7], there is an element θk ∈ E⊗k that corresponds to Ψ via the isomorphism ((Eσ)⊗k)ι ∼= E⊗k. More precisely, we have, using Equation (3.1) in [MS04], for every η1, η2, . . . , ηk in Eσ and every h ∈ Hσ, (5.16) L∗ θk(IE⊗(k−1)⊗ηk)· · · (IE⊗η2)η1h = Ψ(ηk⊗ηk−1⊗· · ·⊗η1⊗h) = = ∆kfσ(0)(η∗ Taking adjoints and writing ui for η∗ k)∗h. 1, . . . , η∗ i , we obtain the desired result. (cid:3) The following corollary is now immediate, by Theorem 5.13 and The- orem 5.21. Corollary 5.22. Let f = {fσ}σ∈Σ be a uniformly bounded matri- cial family of functions defined on a matricial disc D(0, r), where Σ is a full additive subcategory of faithful representations in NRep(M). Then there is a uniquely determined series θ ∼ Pk≥0 θk in T+((E)) TENSORIAL FUNCTION THEORY 45 with R(θ) ≥ r such that f is the family of tensorial power series {Pk≥0 Zk(z)Lθk z ∈ D(0, r, σ)}σ∈Σ . Remark 5.23. Corollary 5.22 and Lemma 5.17 immediately yield the second assertion of Theorem 5.1 6. Series for matricial families of maps In the previous section we studied matricial families of functions in contexts where special generators are not present in the category under consideration. In this section we focus on matricial families of maps (as in Theorem 4.9). Many of the results proved for functions extend to the setting of such families with only minor changes necessary. Indeed, the formulas that go into defining ∆kfσ when f is a map are minor variants of the formulas that enter into the definitions of ∆kfσ when f is a function. One has only to replace equations like Cfσ(z) = fτ (w)C with Cfσ(z) = fτ (w)(IF ⊗ C). To illustrate, recall that to say f is a matricial family of maps means that for each pair, σ and τ in Σ, for each z ∈ D(0,r,σ) , for each w ∈ D(0, r, τ ), and for each C ∈ I(σ, τ ) such that Cz = w(IE ⊗ C) we have So, if (cid:20)z u Cfσ(z) = fτ (w)(IF ⊗ C). 0 w(cid:21) lies in D(0, r, σ ⊕ τ ) and if fσ⊕τ ((cid:20)z u 0 w(cid:21)) = (cid:20)a11 a21 a21 a22(cid:21), 0 w (cid:19)(cid:18) IE ⊗ Iσ (cid:19) , a21 a22(cid:19)(cid:18) IF ⊗ Iσ (cid:19) , (cid:18) Iσ 0 (cid:19) z =(cid:18) z u (cid:18)Iz 0(cid:19) fσ(z) =(cid:18)a11 a12 0 0 then because we must have (6.1) which implies that a11 = fσ(z) and a21 = 0. Formula (6.1) is essen- tially formula (5.3), and the other formulas in the analysis of matricial functions have similar modifications for matricial maps. In particu- lar, one can proceed to define ∆fσ,τ (z, w)(u) as a12, and prove that ∆fσ,τ (z, w)(u) is linear in u. With matricial maps, however, ∆fσ,τ (z, w)(·) is a map from Eσ∗ to F σ∗. Once the distinction between matricial maps and functions is recog- nized, the entire body of results that begins with part (1) of Definition 5.2, ends with Remark 5.16, and does not inolve the bimodule properties of Eσ∗ (as a bimodule over σ(M)′), goes through mutatis mutandis for matricial maps. In particular, the series expansions of functions found in Corollary 5.15 make sense and remain valid for maps. However, the TENSORIAL FUNCTION THEORY 46 notion of a "tensorial power series of maps" has not been defined and the series expansion found in assertion (2) of Theorem 5.1 does not make sense in the setting of maps. So our principal goal, Theorem 6.1, is to exhibit the appropriate replacement and to give a definition of tensorial power series of maps. We shall assume our category Σ ⊆ NRep(M) is additive, full, and has the property that every representation σ ∈ Σ is faithful. Although matricial maps can be defined on arbitrary matricial sets, for simplicity we shall restrict ourselves to matricial discs centered at the origin. Thus we will consider matricial maps f = {fσ}σ∈Σ defined on an E, Σ- matricial disc D(0, r), for some r. We shall assume that f maps into an F, Σ-matricial disc. This is tantamount to assuming that f is uniformly bounded in σ. We will write MLM (E, F ) for the maps in L(E, F ) that are bimod- ule maps. That is, T ∈ L(E, F ) lies in MLM (E, F ) if and only if T (ϕE(a)ξb) = ϕF (a)T (ξ)b, for all a, b ∈ M. Our goal is to prove the following theorem that complements part (2) of Theorem 5.1. Theorem 6.1. Let E and F be two W ∗-correspondences over the same W ∗-algebra, M, and suppose Σ is a full additive subcategory of NRep(M) whose objects are all faithful representations of M. If f = {fσ}σ∈Σ is a matricial family of maps, mapping an E, Σ-disc D(0, r, E) to an F, Σ-disc D(0, R, F ), then there is a uniquely defined sequence of bimodule maps {Dkf}∞ k=0, where for each k, Dkf lies in MLM (F, E⊗k), such that for every z ∈ D(0, r, σ), fσ(z) = fσ(0) +Xk≥1 (6.2) Zk(z)(Dkf ⊗ IHσ ). Theorem 6.1 immediately suggests the following definition for a tenso- rial power series of maps between two correspondence duals. Definition 6.2. A map f defined from an open disc D(ζ0, r, σ) in Eσ∗ to F σ∗ is said to have a tensorial power series expansion on D(ζ0, r, σ) in case there is a sequence {Θk}k≥0, with Θk in MLM (F, E⊗k), such that f (z) =Xk≥0 Zk(z − ζ0)(Θk ⊗ IHσ ) for all z ∈ D(ζ0, r, σ). Under our standing hypotheses on Σ, the derivatives ∆kfσ(0) are really k-linear bimodule maps on (Eσ∗)k mapping to F σ∗, when we view Eσ∗ as a bimodule over σ(M)′, and they are balanced over σ(M)′. The principal obstacle to proving Theorem 6.1 turns out to be isolating the TENSORIAL FUNCTION THEORY 47 dependence of ∆kfσ(0) on σ. This, in turn requires a careful study of the dualities involved. Our analysis therefore rests on the following two lemmas. Note that the first is couched in terms of maps from Eσto F σ instead of maps from Eσ∗ to F σ∗. The reason is that as we noted on page 12, Eσ and F σ are right correspondences over σ(M)′ and this allows us to apply our Duality Theorem, [MS04, Theorem 3.6], to iden- tify E and F with the second duals, Eσ,ι and F σ,ι, respectively, where ι is the identity representation of σ(M)′ on Hσ. This identification will prove central to what follows. Lemma 6.3. If E and F are two W ∗ − correspondences over M and if σ ∈ NRep(M) is a faithful normal representation, then: (1) Every τ ∈σ(M )′ Lσ(M )′(Eσ, F σ) induces is a unique bounded map τ∗ : F → E such that for every η ∈ Eσ and every θ ∈ F , (6.3) η∗Lτ∗(θ) = τ (η)∗Lθ, where Lθ : Hσ → F ⊗σ Hσ is given by Lθh = θ ⊗ h. (2) The map τ∗ lies in MLM (F, E) and the correspondence τ 7→ τ∗ is contravariant and surjective, i.e., (τ1τ2)∗ = τ2∗τ1∗ and (σ(M )′Lσ(M )′(Eσ, F σ))∗ =M LM (F, E). Proof. Recall that Eσ := I(σ, σE ◦ ϕ) is a (right) W ∗-correspondence over σ(M)′. Recall, too, that if ι denotes the identity representation of σ(M)′ on Hσ, then the map WE : E → Eσ,ι such that WE(ξ)∗(η⊗ h) = L∗ ξ(ηh), where ξ ∈ E, η ∈ Eσ and h ∈ Hσ, and where Lξ : Hσ → E⊗Hσ is given by Lξh := ξ ⊗ h, is a correspondence isomorphism [MS04, Theorem 3.6]. Similarly, one has a correspondence isomorphism WF : F → F σ,ι. Also, note that, given g ∈ F σ,ι (so that g : Hσ → F σ ⊗ι Hσ and g(bh) = (ϕι(b) ⊗ IHσ)gh, where b ∈ σ(M) and ϕι(·) is the left action of σ(M) on F σ,ι) and τ ∈σ(M )′ Lσ(M )′(Eσ, F σ), we have that (τ ∗ ⊗ IHσ )g ∈ Eσ,ι. Now define τ∗(θ) = W −1 E ((τ ∗ ⊗ IHσ)WF (θ)). It follows easily that τ∗ ∈M LM (F, E). Also, (τ ∗⊗I)WF (θ) = WE(τ∗(θ)) and, for every η ∈ Eσ and h ∈ Hσ, WE(τ∗(θ))∗(η⊗ h) = WF (θ)(τ ∗(η)⊗ h). Using the definitions of WE and WF , we find that L∗ τ∗(θ)(ηh) = L∗ θ(τ (η)h). By taking adjoints, we obtain Equation (6.3). For the uniqueness statement in (1), suppose that ξ ∈ E satisfies η∗Lξ = τ (η)∗Lθ for all η ∈ Eσ. Then, for every η ∈ Eσ and h ∈ Hσ, ξ(ηh) = L∗ τ∗(θ)(ηh) and, since the images of all η ∈ Eσ span E⊗Hσ, we L∗ find that L∗ τ∗(θ) = L∗ ξ on E ⊗ Hσ, which in turn implies that τ∗(θ) = ξ. TENSORIAL FUNCTION THEORY 48 The fact that the map τ 7→ τ∗ is contravariant follows easily from the definition. The fact that it is surjective follows from duality. (cid:3) Lemma 6.4. Suppose E and F are two W ∗-correspondences over the same W ∗-algebra M. For i = 1, 2, let σ1 and σ2 be two faithful represen- tations in NRep(M) and let τi be a map in σi(M )′Lσi(M )′(Eσi, F σi). If for every c, d ∈ I(σ1, σ2) we have τ1((I ⊗ c∗)ηd) = (I ⊗ c∗)τ2(η)d for ev- ery η ∈ Eσ2, then τ1 = τ2. Consequently, under our standing hypothe- ses that Σ is additive, full, and composed of faithful representations, if {τσ}σ∈Σ is a family of maps, with τσ ∈σ(M )′ Lσ(M )′(Eσ, F σ), and if for every σ1, σ2 ∈ Σ and every c, d ∈ I(σ1, σ2) we have τσ1((I ⊗ c∗)ηd) = (I ⊗ c∗)τσ2(η)d for every η ∈ Eσ2, then the maps τσ∗, σ ∈ Σ, obtained from Lemma 6.3 are independent of σ. Proof. First note that (IE ⊗ c∗)ηd lies in Eσ1 for every c, d ∈ I(σ1, σ2) and every η ∈ Eσ2. If we write ηc,d for (IE ⊗ c∗)ηd, then by the assumption, τ1(ηc,d) = (I ⊗ c∗)τ2(η)d. So, for every θ ∈ F , we have η∗ c,dLτ1∗(θ) = τ1(ηc,d)∗Lθ = ((I ⊗ c∗)τ2(η)d)∗Lθ = d∗τ2(η)∗Lθ(I ⊗ c) = d∗η∗Lτ2∗(θ)(I ⊗ c) = d∗η∗(I ⊗ c)Lτ2∗(θ) = η∗ (6.4) for every η ∈ Eσ2 and c, d ∈ I(σ1, σ2). Since both σ1 and σ2 are normal faithful representations of M, W{d(Hσ1) : d ∈ I(σ1, σ2)} = Hσ2 and W{c∗(Hσ2) : c ∈ I(σ1, σ2)} = Hσ1. It follows from [MS04, Lemma 3.5] that W{η(Hσ2) : η ∈ Eσ2} = E ⊗σ2 Hσ2 and, therefore, that _{ηc,d(Hσ1) : η ∈ Eσ2, c, d ∈ I(σ1, σ2)} = E ⊗σ1 Hσ1. c,dLτ2∗(θ). Thus L∗ τ1∗(θ)ηc,d = L∗ τ2∗(θ)ηc,d (6.5) Combining these equations above we conclude that τ1∗ = τ2∗. (cid:3) Corollary 6.5. Under the hypotheses of Theorem 6.1, the matricial map f has a well-defined Taylor derivative ∆fσ(0)(·) that is a map from Eσ∗ to F σ∗ in σ(M )′Lσ(M )′(Eσ∗, F σ∗). Its transposed map η 7→ (∆fσ(0)(η∗))∗ lies in σ(M )′Lσ(M )′(Eσ, F σ) and has the following depen- dence on σ: There is a map Df ∈M LM (F, E) such that for every σ ∈ Σ and every z ∈ Eσ∗, ∆fσ(0)(z) = z ◦ (Df ⊗ IHσ). Proof. As we indicated at the beginning of this section, the existence and linearity of ∆fσ(0)(·) is proved using the arguments of Lemma 5.2(1) and Lemma 5.9. The fact that ∆fσ(0)(·) is a bimodule map uses the arguments of Lemma 5.2(2) and (3). (Note that we are entitled to apply these because we are assuming Σ is full.) Thus the transposed TENSORIAL FUNCTION THEORY 49 map η 7→ (∆fσ(0)(η)∗)∗ lies in σ(M )′Lσ(M )′(Eσ, F σ), also, and it fol- lows from Lemma 6.3 and Lemma 6.4 that there is an element Df of MLM (F, E) such that, for every σ ∈ Σ, every η ∈ Eσ and every θ ∈ F , η∗LDf (θ) = ∆fσ(0)(η∗)Lθ. Writing z in place of η∗ and applying the two sides of this equality to h ∈ Hσ, we obtain the equation ∆fσ(0)(z)(θ ⊗ h) = z(Df (θ) ⊗ h) = z(Df ⊗ IH)(θ ⊗ h) and the result follows. (cid:3) A similar analysis allows us to identify the dependence of ∆kfσ(0) on σ, k > 1. Lemma 6.6. Under the hypotheses of Theorem 6.1, we conclude that for all k ∈ N and σ ∈ Σ, the Taylor derivative ∆kfσ(0)(·, . . . ,·) is well defined and is a completely bounded k-linear, bimodule map from Eσ∗ ×· · ·× Eσ∗ to F σ∗ that is balanced over σ(M)′. Moreover, there is a uniquely determined map Dkf in MLM (F, E⊗k) such that for every σ ∈ Σ and every z ∈ Eσ∗, ∆kfσ(0)(z) = Zk(z) ◦ (Dkf ⊗ IHσ). Proof. The proof of the existance of ∆kfσ(0)(·, . . . ,·) uses the same arguments as the analogous result for matricial families of functions, as we already have mentioned. To get the bimodule properties of ∆kfσ(0)(·, . . . ,·), note that as in the proof of Theorem 5.21, ∆kfσ(0)(·, . . . ,·) induces a bimodule map Ψ : Eσ ⊗C ∗ Eσ ⊗C ∗ · · · ⊗C ∗ Eσ → F σ. Using Ψ in place of τ in Lemma 6.3, we obtain a map Ψ∗ ∈M LM (F, E⊗k). Note that we do not know that Ψ induces a bimodule map on (Eσ)⊗k, but all we needed in the proof of Lemma 6.3 is to know that the map Ψ ⊗ IHσ : (Eσ)⊗k ⊗ι Hσ → F ⊗ι Hσ is well defined and inter- twines the actions of σ(M)′. This holds here since (Eσ)⊗k ⊗ι H = Eσ ⊗C ∗ Eσ ⊗C ∗ · · · ⊗C ∗ Eσ ⊗ι H, thanks to an observation of Viselter [Vis11, Remark 1.8]. So we do indeed obtain a map Ψ∗ ∈M LM (F, E⊗k) such that, for every θ ∈ F and every z ∈ Eσ∗, Zk(z)LΨ∗(θ) = ∆kfσ(0)(z)Lθ. The map Dkf that we want is Ψ∗. (cid:3) The proof of Theorem 6.1 is essentially complete. All that is necessary is to observe that an analogue of the expansion (5.8) holds also for ma- tricial families of maps. Since we have identified the Taylor derivatives with the Dkf , the proof is complete. We conclude by showing how the Schur class automorphisms we con- sidered in [MS08] fit into theory we have developed here. TENSORIAL FUNCTION THEORY 50 Example 6.7. For a central element γ ∈ D(Z(Eσ)), the map gγ : D(Eσ∗) → D(Eσ∗) is defined by gγ(z) = ∆γ(I − zγ)−1(γ∗ − z)∆−1 γ∗ where ∆γ = (IH − γ∗γ)1/2 ∈ Z(σ(M)′) = σ(Z(M)) and ∆γ∗ = (IE⊗H − γγ∗)1/2 ∈ B(E ⊗σ Hσ). In [MS08, Lemma 4.20] it is shown that there is a completely isometric automorphism αγ of H ∞(E) such that \αγ(X)(z) = bX(gγ(z)) for z ∈ D(0,1,σ). Using Theorem 4.9 we see that gγ is a matricial family of maps (with E = F ) and, thus, Corollary 6.1 applies and we can write (6.6) gγ(z) = gγ(0) +Xk≥1 Zk(z)(Dkgγ ⊗ IHσ ). On the other hand, using the expansion (I − zγ)−1 =Pk≥0(zγ)k, one can write gγ as (6.7) gγ(z) = ∆γγ∗∆−1 γ∗ +Xk≥1 ∆γzγzγ · · · γz∆γ∗ where, in the k-th term, z appears k times. At first glance, it might not be evident that the terms in the expansion (6.7) can be written in the form of the terms in (6.6). To deal with the zeroth term, simply note that ∆γγ∗ = γ∗∆γ∗ so that ∆γγ∗∆−1 γ∗ = γ∗ = I − γγ∗ ∈ (ϕE(M)⊗ I)′ γ∗ = gγ(0). For the first term, note that ∆2 (since γ ∈ Eσ) but also, for a ∈ σ(M)′, γ∗(I ⊗a) = aγ (since γ is in the center of Eσ) and, thus, ∆2 γ∗ ∈ (IE ⊗ σ(M)′)′ = L(E) ⊗ IH. It follows that one can write ∆γ∗ = X⊗IH for an X ∈ L(E)∩ϕE(M)′. Note also that ∆γ is in Z(σ(M)) = σ(Z(M)) and (identifying ∆γ with σ−1(∆γ)), the first term in (6.7) can be written ∆γz∆γ∗ = z(ϕE(∆γ) ⊗ IHσ )(X ⊗ IHσ) = z((ϕE(∆γ)X) ⊗ IHσ) so that, writing Dgγ = ϕE(∆γ)X ∈ L(E) ∩ ϕE(M)′ =M LM (E), we have the first term of (6.7) written in the form of (6.6). (Note that ϕE(∆γ) ∈ ϕE(M)′ since ∆γ ∈ Z(M)). For k ≥ 2 a similar computation is possible. It is a little less straightfor- ward than the computation of Dgγ, but the case when k = 2 illustrates amply what to do. With γ as above, define a map Y : E⊗H → E⊗2⊗H by Y ζh = (IE ⊗ ζ)γh for every ζ ∈ Eσ and h ∈ Hσ. To see that this map is well defined and bounded, compute Pi(I ⊗ ζi)γhi2 = Pi,jh(I ⊗ ζi)γhi, (I ⊗ ζj)γhji = Pi,jhγ∗(I ⊗ ζ ∗ j ζi)γhi, hji. Since γ is in the center of Eσ, the last expression equals Pi,jhγ∗γζ ∗ j ζihi, hji. As ζj ∈ Eσ, this is equal toPi,jhζ ∗ j (ϕE(γ∗γ)⊗ IH )ζihi, hji = (ϕE(γ∗γ)⊗ TENSORIAL FUNCTION THEORY 51 IH)1/2Pi ζihi2. Thus Y ≤ γ. It is also easy to check that Y in- tertwines IE ⊗ b and IE⊗2 ⊗ b for every b ∈ σ(M)′. Thus, there is some Y0 ∈M LM (E, E⊗2) such that Y = Y0⊗IH . For z ∈ Eσ∗ and ζ ∈ Eσ, we have zζ ∈ σ(M)′ and, since γ is central, we conclude that for h ∈ Hσ, γzζh = (I ⊗ zζ)γh = (I ⊗ z)(Y0 ⊗ I)ζh. Thus γz = (IE ⊗ z)(Y0 ⊗ IHσ). We now compute: ∆γzγz∆γ∗ = zγz(ϕE(∆γ) ⊗ IH )(X ⊗ IH) = Z2(z)(Y0ϕE(∆γ)X ⊗ IH), which shows that D2gγ = Y0ϕE(∆γ)X. References [AF92] Frank W. Anderson and Kent R. Fuller, Rings and categories of mod- ules, second ed., Graduate Texts in Mathematics, vol. 13, Springer- Verlag, New York, 1992. MR 1245487 (94i:16001) [Arv98] William Arveson, Subalgebras of C ∗-algebras. III. Multivariable oper- ator theory, Acta Math. 181 (1998), no. 2, 159 -- 228. MR 1668582 (2000e:47013) David P. Blecher, A new approach to Hilbert C ∗-modules, Math. Ann. 307 (1997), no. 2, 253 -- 290. MR 1428873 (98d:46063) [Ble97] [BLM04] David P. Blecher and Christian Le Merdy, Operator algebras and their modules -- an operator space approach, London Mathematical Society Monographs. New Series, vol. 30, The Clarendon Press Oxford Univer- sity Press, Oxford, 2004, Oxford Science Publications. MR 2111973 (2006a:46070) [Coh06] [BMP00] David P. Blecher, Paul S. Muhly, and Vern I. Paulsen, Categories of operator modules (Morita equivalence and projective modules), Mem. Amer. Math. Soc. 143 (2000), no. 681, viii+94. MR 1645699 (2000j:46132) P. M. Cohn, Free ideal rings and localization in general rings, New Mathematical Monographs, vol. 3, Cambridge University Press, Cam- bridge, 2006. MR 2246388 (2007k:16020) Kenneth R. Davidson and David R. Pitts, The algebraic structure of non-commutative analytic Toeplitz algebras, Math. Ann. 311 (1998), no. 2, 275 -- 303. MR 1625750 (2001c:47082) Edward G. Effros and Zhong-Jin Ruan, Operator spaces, London Math- ematical Society Monographs. New Series, vol. 23, The Clarendon Press Oxford University Press, New York, 2000. MR 1793753 (2002a:46082) [GLR85] P. Ghez, R. Lima, and J. E. Roberts, W ∗-categories, Pacific J. Math. [DP98] [ER00] 120 (1985), no. 1, 79 -- 109. MR 808930 (87g:46091) [HKM11a] J. William Helton, Igor Klep, and Scott McCullough, Analytic mappings between noncommutative pencil balls, J. Math. Anal. Appl. 376 (2011), no. 2, 407 -- 428. MR 2747767 (2012b:46135) [HKM11b] , Proper analytic free maps, J. Funct. Anal. 260 (2011), no. 5, 1476 -- 1490. MR 2749435 TENSORIAL FUNCTION THEORY 52 [HP74] [KK06] [HKMS09] J. William Helton, Igor Klep, Scott McCullough, and Nick Slinglend, Noncommutative ball maps, J. Funct. Anal. 257 (2009), no. 1, 47 -- 87. MR 2523335 (2011b:47037) Einar Hille and Ralph S. Phillips, Functional analysis and semi-groups, American Mathematical Society, Providence, R. I., 1974, Third printing of the revised edition of 1957, American Mathematical Society Collo- quium Publications, Vol. XXXI. MR 0423094 (54 #11077) Elias Katsoulis and David W. Kribs, The C ∗-envelope of the tensor algebra of a directed graph, Integral Equations Operator Theory 56 (2006), no. 3, 401 -- 414. MR 2270844 (2007m:46088) D. S. Kaliuzhnyi-Verbovetskyi, Noncommutative functions and fixed point theorems, Slides from 2012 joint Mathematics meeting. D. S. Kaliuzhnyi-Verbovetskyi and V. Vinnikov, Noncommutative ra- tional functions, their difference-differential calculus and realizations, ArXiv e-prints (2010). D. S. Kaliuzhnyi-Verbovetskyi and V. Vinnikov, Foundations of non- commutative function theory, In Preparation. [KV] [KV10] [KVV] [KVV06] Dmitry S. Kalyuzhnyi-Verbovetskiı and Victor Vinnikov, Non- commutative positive kernels and their matrix evaluations, Proc. Amer. Math. Soc. 134 (2006), no. 3, 805 -- 816 (electronic). MR 2180898 (2006f:47019) [KVV09] Dmitry S. Kaliuzhnyi-Verbovetskyi and Victor Vinnikov, Singularities of rational functions and minimal factorizations: the noncommutative and the commutative setting, Linear Algebra Appl. 430 (2009), no. 4, 869 -- 889. MR 2489365 (2010a:47032) [Lum86] Denis Luminet, A functional calculus for Banach PI-algebras, Pacific J. [MS98] [MS99] [MS00] [MS04] [MS05] [MS08] [MS09] Math. 125 (1986), no. 1, 127 -- 160. MR 860755 (88c:46060) Paul S. Muhly and Baruch Solel, Tensor algebras over C ∗- correspondences: representations, dilations, and C ∗-envelopes, J. Funct. Anal. 158 (1998), no. 2, 389 -- 457. MR 1648483 (99j:46066) , Tensor algebras, induced representations, and the Wold de- composition, Canad. J. Math. 51 (1999), no. 4, 850 -- 880. MR 1701345 (2000i:46052) , On the Morita equivalence of tensor algebras, Proc. London Math. Soc. (3) 81 (2000), no. 1, 113 -- 168. MR 1757049 (2001g:46128) , Hardy algebras, W ∗-correspondences and interpolation theory, Math. Ann. 330 (2004), no. 2, 353 -- 415. MR 2089431 (2006a:46073) , Hardy algebras associated with W ∗-correspondences (point eval- uation and Schur class functions), Operator theory, systems theory and scattering theory: multidimensional generalizations, Oper. Theory Adv. Appl., vol. 157, Birkhäuser, Basel, 2005, pp. 221 -- 241. MR 2129649 (2006d:46071) , Schur class operator functions and automorphisms of Hardy algebras, Doc. Math. 13 (2008), 365 -- 411. MR 2520475 (2010g:46098) , The Poisson kernel for Hardy algebras, Complex Anal. Oper. Theory 3 (2009), no. 1, 221 -- 242. MR 2481905 (2011a:47166) [MS11a] , Morita transforms of tensor algebras, New York J. Math. 17A (2011), 87 -- 100. MR 2782729 TENSORIAL FUNCTION THEORY 53 [MS11b] [Muh97] [Pop96] [Pop06] [Pop08] , Representations of Hardy algebras: absolute continuity, in- tertwiners, and superharmonic operators, Integral Equations Operator Theory 70 (2011), no. 2, 151 -- 203. MR 2794388 (2012g:47218) Paul S. Muhly, A finite-dimensional introduction to operator algebra, Operator algebras and applications (Samos, 1996), NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., vol. 495, Kluwer Acad. Publ., Dordrecht, 1997, pp. 313 -- 354. MR 1462686 (98h:46062) Gelu Popescu, Non-commutative disc algebras and their representations, Proc. Amer. Math. Soc. 124 (1996), no. 7, 2137 -- 2148. MR 1343719 (96k:47077) Anal. 241 (2006), no. 1, 268 -- 333. MR 2264252 (2007h:47015) , Free holomorphic functions on the unit ball of B(H)n, J. Funct. , Noncommutative Berezin transforms and multivariable oper- ator model theory, J. Funct. Anal. 254 (2008), no. 4, 1003 -- 1057. MR 2381202 (2009j:47015) [Rie74a] Marc A. Rieffel, Induced representations of C ∗-algebras, Advances in Math. 13 (1974), 176 -- 257. MR 0353003 (50 #5489) [Rie74b] , Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure [Tay72] [Tay73] [Vis11] [Voi04] [Voi10] Appl. Algebra 5 (1974), 51 -- 96. MR 0367670 (51 #3912) Joseph L. Taylor, A general framework for a multi-operator functional calculus, Advances in Math. 9 (1972), 183 -- 252. MR 0328625 (48 #6967) , Functions of several noncommuting variables, Bull. Amer. Math. Soc. 79 (1973), 1 -- 34. MR 0315446 (47 #3995) Ami Viselter, Covariant representations of subproduct systems, Proc. Lond. Math. Soc. (3) 102 (2011), no. 4, 767 -- 800. MR 2793449 (2012d:46170) Dan Voiculescu, Free analysis questions. I. Duality transform for the coalgebra of ∂X : B, Int. Math. Res. Not. (2004), no. 16, 793 -- 822. MR 2036956 (2005a:46140) Dan-Virgil Voiculescu, Free analysis questions II: the Grassmannian completion and the series expansions at the origin, J. Reine Angew. Math. 645 (2010), 155 -- 236. MR 2673426 (2012b:46144) Department of Mathematics, University of Iowa, Iowa City, IA 52242 E-mail address: [email protected] Department of Mathematics, Technion, 32000 Haifa, Israel E-mail address: [email protected]
1210.4670	1	1210	2012-10-17T08:55:44	Completeness of $n$--tuple of projections in $C^*$--algebras	[ "math.OA", "math.FA" ]	Let $(P_1,...,P_n)$ be an $n$--tuple of projections in a unital $C^*$--algebra $\aa$. We say $\pn$ is complete in $\aa$ if $\aa$ is the linear direct sum of the closed subspaces $P_1\aa,...,P_n\aa$. In this paper, we give some necessary and sufficient conditions for the completeness of $\pn$ and discuss the perturbation problem and topology of the set of all complete $n$--tuple of projections in $\aa$. Some interesting and significant results are obtained in this paper.	math.OA	math	Completeness of n -- tuple of projections in C∗ -- algebras Shanwen Hu∗ and Yifeng Xue† Department of mathematics and Research Center for Operator Algebras East China Normal University, Shanghai 200241, P.R. China Abstract Let (P1,··· , Pn) be an n -- tuple of projections in a unital C ∗ -- algebra A. We say (P1,··· , Pn) is complete in A if A is the linear direct sum of the closed subspaces P1A,··· , PnA. In this paper, we give some necessary and sufficient conditions for the completeness of (P1,··· , Pn) and discuss the perturbation problem and topology of the set of all complete n -- tuple of projections in A. Some interesting and significant results are obtained in this paper. 2010 Mathematics Subject Classification: 46L05, 47B65 Key words: projection, idempotent, complete n -- tuple of projections 0 Introduction Throughout the paper, we always assume that H is a complex Hilbert space with inner product < ·,· >, B(H) is the C ∗ -- algebra of all bounded linear operators on H and A is a C ∗ -- algebra with the unit 1. Let A+ denote the set of all positive elements in A. It is well -- known that A has a faithful representation (ψ, Hψ) with ψ(1) = I (cf. [8, Theorem 1.6.17] or [17, Theorem 1.5.36]). For T ∈ B(H), let Ran(T ) (resp. Ker(T )) denote the range (resp. kernel) of T . Let V1, V2 be closed subspaces in H such that H = V1 ∔ V2 = V ⊥ 1 ∔ V2, that is, V1 and V2 is in generic position (cf. [6]). Let Pi be projection of H onto Vi, i = 1, 2. Then H = Ran(P1) ∔ Ran(P2) = Ran(I − P1) ∔ Ran(P2). In this case, Halmos gave very useful matrix representations of P1 and P2 in [6]. Following Halmos' work, Sunder investigated in [14] the n -- tuple closed subspaces (V1,··· , Vn) in H which satisfying the condition H = V1 ∔··· ∔ Vn. If let Pi be the projection of H onto Vi, i = 1,··· , n, then the condition H = V1 ∔ ··· ∔ Vn is equivalent to H = Ran(P1) ∔ ··· ∔ Ran(Pn). Now the question yields: when does the relation H = Ran(P1) ∔ ··· ∔ Ran(Pn) hold for an n -- tuple of projections (P1,··· , Pn)? When n = 2, Buckholdtz proved in [3] that Ran(P1) ∔ Ran(P2) = H iff P1 − P2 is invertible in B(H) iff I − P1P2 is invertible in B(H) and iff P1+P2−P1P2 is invertible in B(H). More information about two projections can be found in [2]. Koliha and Rakocevi´c generalized Buckholdtz's ∗E-mail: [email protected] †E-mail: [email protected] 1 work to the set of C ∗ -- algebras and rings. They gave some equivalent conditions for decomposition R = P R ∔ Q R or R = R P ∔ R Q in [9] and [10] for idempotent elements P and Q in a unital ring R. They also characterized the Fredhomness of the difference of projections on H in [11]. For n ≥ 3, the question remains unknown so far. But there are some works concerning with this problem. For example, the estimation of the spectrum of the finite sum of projections on H is given in [1] and the C ∗ -- algebra generated by certain projections is investigated in [13] and [15], etc.. Let Pn(A) denote the set of n -- tuple (n ≥ 2) of non -- trivial projections in A and put PCn(A) = {(P1,··· , Pn) ∈ Pn(A) P1A ∔ ··· ∔ PnA = A}. It is worth to note that if A = B(H) and (P1,··· , Pn) ∈ Pn(B(H)), then (P1,··· , Pn) ∈ PCn(B(H)) if and only if Ran(P1) ∔ ··· ∔ Ran(Pn) = H (see Theorem 1.2 below). In this paper, we will investigate the set PCn(A) for n ≥ 3. The paper consists of four sections. In Section 1, we give some necessary and sufficient conditions that make (P1,··· , Pn) ∈ Pn(A) be in PCn(A). In Section 2, using some equivalent conditions for (P1,··· , Pn) ∈ PCn(A) obtained in §1, we obtain an explicit expression of Pi1 ∨ ··· ∨ Pik for {i1,··· , ik} ⊂ {1,··· , n}. We discuss the perturbation problems for (P1,··· , Pn) ∈ PCn(A) in Section 3. We find an interesting result: if (P1,··· , Pn) ∈ Pi invertible in A, then kPiA−1Pjk < (cid:2)(n − 1)kA−1kkAk2(cid:3)−1, Pn(A) with A = i 6= j implies PiA−1Pj = 0, i 6= j, i, j = 1,··· , n in this section. We show in this section that for given ǫ ∈ (0, 1), if (P1,··· , Pn) ∈ Pn(A) satisfies condition kPiPjk < ǫ, then there exists an n -- tuple of mutually orthogonal projections (P ′ n) ∈ Pn(A) such that kPi−P ′ ik < 2(n−1)ǫ, i = 1,··· , n, which improves a conventional estimate: ik < (12)n−1n!ǫ, i = 1,··· , n (cf. [8]). In the final section, we will study the kPi − P ′ topological properties and equivalent relations on PCn(A). 1,··· , P ′ n Pi=1 1 Equivalent conditions for complete n -- tuples of projec- tions in C∗ -- algebras Let GL(A) (resp. U (A)) denote the group of all invertible (resp. unitary) elements in A. Let Mk(A) denote matrix algebra of all k × k matrices over A. For any a ∈ A, we set aA = {ax x ∈ A} ⊂ A. Definition 1.1. An n -- tuple of projections (P1,··· , Pn) in A is called complete in A, if (P1,··· , Pn) ∈ PCn(A). Theorem 1.2. Let (P1,··· , Pn) ∈ Pn(A). Then the following statements are equiv- alent: (1) (P1,··· , Pn) is complete in A. (2) Hψ = Ran(ψ(P1)) ∔ ··· ∔ Ran(ψ(Pn)) for any faithful representation (ψ, Hψ) of A with ψ(1) = I. 2 (3) Hψ = Ran(ψ(P1)) ∔ ··· ∔ Ran(ψ(Pn)) for some faithful representation (ψ, Hψ) of A with ψ(1) = I. (4) Pj6=i Pj + λPi ∈ GL(A), i = 1, 2,··· , n and ∀ λ ∈ [1 − n, 0). Pj(cid:1) + Pi ∈ GL(A) for 1 ≤ i ≤ n and all λ ∈ C\{0}. (5) λ(cid:0)Pj6=i Pi=1 Pi ∈ GL(A) and PiA−1Pi = Pi, i = 1,··· , n. Pi=1 Pi ∈ GL(A) and PiA−1Pj = 0, i 6= j, i, j = 1,··· , n. (7) A = (6) A = n n (8) there is an n-tuple of idempotent operators (E1,··· , En) in A such that EiPi = n Ei, PiEi = Pi, i = 1,··· , n and EiEj = 0, i 6= j, i, j = 1,··· , n, Ei = 1. Pi=1 In order to show Theorem 1.2, we need following lemmas. Lemma 1.3. Let B, C ∈ A+\{0} and suppose that λB + C is invertible in A for every λ ∈ R\{0}. Then there is a non -- trivial orthogonal projection P ∈ A such that B = (B + C)1/2P (B + C)1/2, C = (B + C)1/2(1 − P )(B + C)1/2. Proof. Put D = B + C and Dλ = λB + C, ∀ λ ∈ R\{0}. Then D ≥ 0, D and Dλ are all invertible in A, ∀ λ ∈ R\{0}. Put B1 = D−1/2BD−1/2, C1 = D−1/2CD−1/2. Then B1 + C1 = 1 and D−1/2DλD−1/2 = λB1 + C1 = λ + (1 − λ)C1 = (1 − λ)(λ(1 − λ)−1 + C1) λ 1 − λ is invertible in A for any λ ∈ R\{0, 1}. Since λ 7→ is a homeomorphism from R\{0, 1} onto R\{−1, 0}, it follows that σ(C1) ⊂ {0, 1}. Note that B1 and C1 are all non -- zero. So σ(C1) = {0, 1} = σ(B1) and hence P = B1 is a non -- zero projection in A and B = D1/2P D1/2, C = D1/2(1 − P )D1/2. Lemma 1.4. Let B, C ∈ A+\{0}. Then the following statements are equivalent: (1) for any non -- zero real number λ, λB + C is invertible in A. (2) B + C is invertible in A and B(B + C)−1B = B. (3) B + C is invertible in A and B(B + C)−1C = 0. (4) B + C is invertible in A and for any B′, C ′ ∈ A+ with B′ ≤ B and C ′ ≤ C, B′(B + C)−1C ′ = 0. 3 Proof. (1)⇒(2) By Lemma 1.3, there is a non -- zero projection P in A such that B = D1/2P D1/2, C = D1/2(1 − P )D1/2, where D = B + C ∈ GL(A). So B(B + C)−1B = D1/2P D1/2D−1D1/2P D1/2 = B. The assertion (2) ⇔ (3) follows from B(B + C)−1B = B(B + C)−1(B + C − C) = B − B(B + C)−1C. (3) ⇒ (4) For any C ′ with 0 ≤ C ′ ≤ C, 0 ≤ B(B + C)−1C ′(B + C)−1B ≤ B(B + C)−1C(B + C)−1B = 0, we have B(B +C)−1C ′ = B(B +C)−1C ′1/2C ′1/2 = 0. This implies C ′(B +C)−1B = 0. In the same way, we get that for any B′ with 0 ≤ B′ ≤ B, C ′(B + C)−1B′ = 0. (4)⇒(3) is obvious. (2)⇒(1) Set X = (B + C)−1/2B and Y = (B + C)−1/2C. Then X, Y ∈ A and X ∗X = B, X + Y = (B + C)1/2. Thus, for any λ ∈ R\{0}, X + λY = (B + C)−1/2(B + λC) (X + λY )∗(X + λY ) = ((1 − λ)X + λ(B + C)1/2)∗((1 − λ)X + λ(B + C)1/2) = (1 − λ)2B + 2λ(1 − λ)B + λ2(B + C) = B + λ2C and consequently, (X + λY )∗(X + λY ) ≥ B + C if λ > 1 and (X + λY )∗(X + λY ) ≥ λ2(B + C) when λ < 1. This indicates that (X + λY )∗(X + λY ) is invertible in A. Noting that B + C ≥ k(B + C)−1k−1 · 1, we have, for any λ ∈ R\{0}, (B + λC)2 = (X + λY )∗(B + C)(X + λY ) ≥ k(B + C)−1k−1(X + λY )∗(X + λY ). Therefore, B + λC is invertible in A, ∀ λ ∈ R\{0}. Pibi. n Pi=1 ... . Then Put I = . . . , X = 1 0   0   P1 0 ... 0   Now we begin to prove Theorem 1.2. (1)⇒(6) Statement (1) implies that there are b1,··· , bn ∈ A such that 1 =     and Y =  b1 0 ··· ... . . .  bn 0 ··· Pi=1 I = XY = XY Y ∗X ∗ ≤ kY k2XX ∗ = kY k2 ··· Pn 0 ··· ... . . . 0 ···   0 . . . 0 ... 0 0 n Pi   4 and so that A = n Pi=1 Pi is invertible in A. Therefore, from A = P1A ∔ ··· ∔ PnA and Pi = P1A−1Pi + ··· + PiA−1Pi + ··· + PnA−1Pi = 0 + ··· + 0 } {z i−1 , +Pi + 0 + ··· + 0 } {z n−i i = 1,··· , n, we get that Pi = PiA−1Pi, i = 1,··· , n. (2)⇒(3) is obvious. (3)⇒(4) Set Qi = ψ(Pi), i = 1,··· , n. From Hψ = Ran(Q1) ∔ ··· ∔ Ran(Qn), Pi=1 we obtain idempotent operators F1,··· , Fn in B(Hψ) such that Fi = I, FiFj = 0, i 6= j and FiHψ = QiHψ, i, j = 1,··· , n. So FiQi = Qi, QiFi = Fi and FjQi = 0, i 6= j, 1 ≤ i, j ≤ n. Using these relations, it is easy to check that n n Xi=1 Xi=1 n (cid:0) (cid:0) n λiQi(cid:1)(cid:0) λ−1 i F ∗ Xi=1 i Fi(cid:1)(cid:0) n Xi=1 λ−1 i F ∗ i Fi(cid:1) = λiQi(cid:1) = n Xi=1 Xi=1 n Fi = I, F ∗ i = I, for any non -- zero complex number λi, i = 1,··· , n. Particularly, for any λ ∈ [1−n, 0), Qj(cid:1) + Qi(cid:1)−1 = λ−1Xj6=i (cid:0)λ(cid:0)Xj6=i Qj(cid:1) + Qi is invertible ψ(A), 1 ≤ i ≤ n by [17, Corollary 1.5.8] F ∗ j Fj + F ∗ i Fi in B(Hψ). Thus, λ(cid:0)Pj6=i and so that λ(cid:0)Pj6=i (4)⇒(5) Put Ai(λ) = Pj6=i (Ai(λ))2 ≤ 2(cid:0)Xj6=i Pj(cid:1) + Pi ∈ GL(A) since ψ is faithful and ψ(1) = I. Pj + λPi, i = 1,··· , n, λ ∈ R\{0}, then Pj(cid:1)2 + 2λ2Pi ≤ 2(n − 1)Xj6=i ≤ 2 max{n − 1, λ2}(P1 + ··· + Pn). Pj + 2λ2Pi So Ai(λ) is invertible in A, ∀ λ ∈ [1− n, 0) means that A = P1 +··· + Pn is invertible in A. Consequently, Ai(λ) is invertible in A when λ > 0, ∀ 1 ≤ i ≤ n. Now we show that Ai(λ) is invertible in A for i = 1,··· , n and λ < 1 − n. Put A1i = PiAPi, A2i = PiA(1 − Pi), A4i = (1 − Pi)A(1 − Pi), i = 1,··· , n. A4i(cid:21), i = 1,··· , n. Noting that Express Ai(λ) as the form Ai(λ) = (cid:20)A1i + (λ − 1)Pi A2i A4i is invertible in (1 − Pi)A(1 − Pi) (A ≥ kA−1k−1 · 1, A4i ≥ kA−1k−1(1 − Pi)) and A∗ 2i Ai(λ)(cid:20) Pi −A−1 4i A∗ 0 2i 1 − Pi(cid:21) = (cid:20)A1i − A2iA−1 4i A∗ 0 A4i(cid:21) , 2i + (λ − 1)Pi A2i 5 we get that Ai(λ) is invertible iff A1i − A2iA−1 i = 1,··· , n. Since A1i ≤ nPi, it follows that 4i A∗ 2i + (λ − 1)Pi is invertible in PiAPi, −A1i + A2iA−1 4i A∗ 2i − (λ − 1)Pi ≥ (1 − n − λ)Pi + A2iA−1 4i A∗ 2i ≥ (1 − n − λ)Pi n n Pi=1 when λ < 1 − n, i = 1,··· , n. Therefore, Ai(λ) is invertible in A for λ < 1 − n and i = 1,··· , n. Applying Lemma 1.4 to Pj6=i (5)⇒(6) and (6)⇒ (7) easily. (7)⇒(8) Set Ei = PiA−1, i = 1··· , n. Then Ei is an idempotent elements in A Ei = 1 and PiEi = Ei, Pj and Pi, 1 ≤ i ≤ n, we can get the implications (8)⇒(1) Let E1,··· , En be idempotent elements in A such that EiEj = δijEi, Ei = 1 and EiPi = Pi, PiEi = Ei, i, j = 1,··· , n. Then EiA = PiA, i = 1,··· , n and EiEj = 0, i 6= j, i, j = 1,··· , n. It is obvious that EiPi = Pi, i = 1,··· , n. Pi=1 and A = E1A ∔ ··· EnA = P1A ∔ ··· ∔ PnA. Pi=1 resentation of A with ψ(1) = I. Put E′ E′ iE′ i = I and Ran(E′ Hψ = Ran(Q1) ∔ ··· ∔ Ran(Qn). Remark 1.5. (1) Statement (3) in Theorem 1.2 can not be replaced by "for any (8)⇒(2) Let E1,··· , En be idempotent elements in A such that EiEj = δijEi, Ei = 1 and EiPi = Pi, PiEi = Ei, i, j = 1,··· , n. Let (ψ, Hψ) be any faithful rep- i = ψ(Ei) and Qi = ψ(Pi), i = 1,··· , n. Then i) = Ran(Qi), i, j = 1,··· , n. Consequently, j = δijE′ i, Pi=1 E′ n n 4 Li=1 H and put A = B(H (4)),  , P2 =    , P3 =     0 0 I I I 0 0 .   I 1 ≤ i ≤ n, Pi − Pj6=i Pj is invertible". For example, let H (4) = I I 0 0 P1 =   Clearly, Pi − Pj6=i is, (P1, P2, P3) is not complete in A. Pi ∈ GL(A), then Pi6=j ∈ Pn(A), if Pi=1 Pj is invertible, 1 ≤ i ≤ 3, but P2 + P3 − 2P1 is not invertible, that (2) According to the proof of (3)⇒(4) of Theorem 1.2, we see that for (P1,··· , Pn) Pi − λPi ∈ GL(A), ∀ 1 ≤ i ≤ n and λ > n − 1. n Corollary 1.6 ([3, Theorem 1]). Let P1, P2 be non -- trivial projections in B(H). Then H = Ran(P1) ∔ Ran(P2) iff P1 − P2 is invertible in B(H). Proof. By Theorem 1.2, H = Ran(P1) ∔ Ran(P2) implies that P1 − P2 ∈ GL(B(H)). 6 Conversely, if P1 − P2 ∈ GL(B(H)), then from 2(P1 + P2) ≥ (P1 − P2)2, we get that P1 + P2 ∈ GL(B(H)) and so that P1 − λP2, P2 − λP1 ∈ GL(B(H)), ∀ λ > 1 by Remark 1.5 (2). Thus, for any λ ∈ (0, 1], P1 − λP2 and P2 − λP1 are all invertible in B(H). Consequently, H = Ran(P1) ∔ Ran(P2) by Theorem 1.2. 2 Some representations concerning the complete n -- tuple of projections We first statement two lemmas which are frequently used in this section and the later sections. Lemma 2.1. Let B ∈ A+ such that 0 ∈ σ(B) is an isolated point. Then there is a unique element B† ∈ A+ such that BB†B = B, B†BB† = B†, BB† = B†B. Proof. The assertion follows from Proposition 3.5.8, Proposition 3.5.3 and Lemma 3.5.1 of [17]. Remark 2.2. The element B† in Lemma 2.1 is called the Moore -- Penrose inverse of B. When 0 6∈ σ(B), B† , B−1. The detailed information can be found in [17]. The following lemma comes from [17, Lemma 3.5.5] and [4, Lemma 1]: Lemma 2.3. Let P ∈ A be an idempotent element. Then (1) P + P ∗ − 1 ∈ GL(A). (2) R = P (P + P ∗ − 1)−1 is a projection in A satisfying P R = R and RP = P . Moreover, if R′ ∈ A is a projection such that P R′ = R′ and R′P = P , then R′ = R. Pi. By Theorem 1.2, A ∈ GL(A) Let (P1,··· , Pn) ∈ PCn(A) and put A = n and Ei = PiA−1, 1 ≤ i ≤ n are idempotent elements satisfying conditions Pi=1 EiEj = 0, i 6= j, EiPi = Pi, PiEi = Ei, i = 1,··· , n, and n Xi=1 Ei = 1. i + Ei− 1)−1, 1 ≤ i ≤ n. So the C ∗ -- algebra C ∗(P1,··· , Pn) By Lemma 2.3, Pi = Ei(E∗ generated by P1,··· , Pn is equal to the C ∗ -- algebra C ∗(E1,··· , En) generated by E1,··· , En. Put Qi = A−1/2PiA−1/2, i = 1,··· , n. Then QiQj = δijQi by Theorem 1.2, i, j = 1,··· , n and Qi = 1. Thus, n Pi=1 Pi = A1/2QiA1/2 and Ei = PiA−1 = A1/2QiA−1/2, i = 1,··· , n. (2.1) 7 Proposition 2.4. Let (P1,··· , Pn) ∈ PCn(A) with A = i = 1,··· , n, (cid:0) n Pi=1 λiPi(cid:1)−1 = A−1(cid:0) n Pi=1 λ−1 Proof. Keeping the symbols as above. We have i Pi(cid:1)A−1. Pi=1 n n Xi=1 (cid:0) λiPi(cid:1)−1 = A−1/2(cid:0) n Xi=1 λ−1 i Qi(cid:1)A−1/2 = A−1(cid:0) n Pi=1 Pi. Then for any λi 6= 0, λiQi(cid:1)A1/2. Thus, n λiPi = A1/2(cid:0) n Pi=1 i Pi(cid:1)A−1. Pr=1 Xi=1 λ−1 k k Now for i1, i2,··· , ik ∈ {1, 2,··· , n} with i1 < i2 < ··· < ik, put A0 = Pr=1 Qir . Then A0, Q0 ∈ A and Q0 is a projection. From (2.1), A0 = A1/2Q0A1/2. Q0 = Thus, σ(A0)\{0} = σ(Q0AQ0)\{0} (cf. [17, Proposition 1.4.14]). Since Q0AQ0 is invertible in Q0AQ0, it follows that 0 ∈ σ(Q0AQ0) is an isolated point and so that 0 ∈ σ(A0) is also an isolated point. So we can define Pi1 ∨···∨Pik to be the projection A† 0A0 ∈ A by Lemma 2.1. This definition is reasonable: if P ∈ A is a projection such that P ≥ Pir , r = 1,··· , k, then P A0 = A0 and Pir and hence P A0A† 0 = A0A† 0 = (1 − A† 0, i.e., P ≥ Pi1 ∨ ··· ∨ Pik ; Since A0 ≥ Pir , we have 0A0) 0A0) ≥ (1 − A† 0A0)Pir (1 − A† 0A0)A0(1 − A† and consequently, Pir (1 − A† Proposition 2.5. Let (P1,··· , Pn) ∈ PCn(A) with A = above and {j1,··· , jl} = {1,··· , n}\{i1,··· , ik} with j1 < ··· < jl. Then 0A0) = 0, that is, Pir ≤ Pi1 ∨ ··· ∨ Pik , i = 1,··· , k. Pi. Let i1,··· , ik be as Pi=1 n k k Pi1 ∨ ··· ∨ Pik = A1/2(cid:2)(cid:0) Xr=1 = (cid:0) Xr=1 Pir(cid:1)(cid:2)(cid:0) Qir(cid:1)A(cid:0) Xr=1 Xr=1 Pir(cid:1)2 + Qir(cid:1)(cid:3)−1A1/2 Xt=1 Pjt(cid:3)−1(cid:0) k k l (2.2) (2.3) k Xr=1 Pir(cid:1). Proof. Using the symbols Pi, Qi, Ei as above. According to (2.1), k k k k Xr=1 Xr=1 Pir = A1/2(cid:0) Thus (cid:0) k Eir(cid:1)(cid:0) k Pir(cid:1) = Pr=1 Pr=1 Pr=1 Xr=1 Eir(cid:1)Pi1 ∨ ··· ∨ Pik = Pi1 ∨ ··· ∨ Pik , Pi1 ∨ ··· ∨ Pik(cid:0) (cid:0) Xr=1 Eir = A1/2(cid:0) Eir = (cid:0) k Pr=1 Qir(cid:1)A1/2, Pr=1 Xr=1 Pir and k k k Pir(cid:1)A−1. Then we have Xr=1 Eir(cid:1) = Xr=1 k k Qir(cid:1)A−1/2. Eir , according to the definition of Pi1 ∨ ··· ∨ Pik . 8 Since k Pr=1 Eir is an idempotent element in A, it follows from Lemma 2.3 that k k Xr=1 Xr=1 Pi1 ∨ ··· ∨ Pik = (cid:0) Eir(cid:1)(cid:2) Noting that (cid:0) k Qir(cid:1)A(cid:0) k Qir(cid:1) ∈ GL(cid:0)(cid:0) k Pr=1 Pr=1 Pr=1 is invertible in (cid:0) k Qjt(cid:1)A(cid:0) k Qjt(cid:1) and Pt=1 Pt=1 Xr=1 Xr=1 ir + Eir ) − 1 = A−1/2(cid:2)(cid:0) Xr=1 = A−1/2(cid:2)(cid:0) Qir(cid:1)A + A(cid:0) Xr=1 Qir(cid:1)A(cid:0) (E∗ k k k k (E∗ ir + Eir ) − 1(cid:3)−1 ∈ A. Qir(cid:1)A(cid:0) k Qir(cid:1)(cid:1); (cid:0) l Pr=1 Pt=1 (2.4) Qjt(cid:1) Qjt(cid:1)A(cid:0) k Pt=1 k Xr=1 Qir(cid:1)−(cid:0) Qir(cid:1) − A(cid:3)A−1/2 Xt=1 Qjt(cid:1)A(cid:0) Xt=1 l l Qjt(cid:1)(cid:3)A−1/2, we obtain that k Xr=1 (cid:2) (E∗ ir + Eir ) − 1(cid:3)−1 Xr=1 = A1/2(cid:2)(cid:2)(cid:0) k Qir(cid:1)A(cid:0) k Xr=1 Qir(cid:1)(cid:3)−1−(cid:2)(cid:0) l Xt=1 Qjt(cid:1)A(cid:0) l Xt=1 Qjt(cid:1)(cid:3)−1(cid:3)A1/2. Pjt = A1/2(cid:0) l Pt=1 Qjt(cid:1)A1/2 and(cid:0) k Pr=1 Pir(cid:1)2 k Note that Combining this with (2.4), we can get (2.2). l Pir = A1/2(cid:0) k Qir(cid:1)A1/2, Pr=1 Pr=1 Pt=1 = A1/2(cid:0) k Qir(cid:1)A(cid:0) k Qir(cid:1)A1/2. Therefore, Pr=1 Pr=1 Xt=1 Xr=1 Xr=1 Xr=1 Pjt(cid:3)−1(cid:0) (cid:0) Pir(cid:1) Pir(cid:1)(cid:2)(cid:0) Pir(cid:1)2 + Xr=1 Xr=1 Xr=1 = A1/2(cid:0) Qir(cid:1)(cid:0)(cid:2)(cid:0) Qir(cid:1)A(cid:0) Qir(cid:1)(cid:3)−1 + = Pi1 ∨ ··· ∨ Pik k k k k k k l l Xt=1 Qjt(cid:1)(cid:0) k Xr=1 Qir(cid:1)A1/2 by (2.2). 3 Perturbations of a complete n -- tuple of projections Recall that for any non -- zero operator C ∈ B(H), the reduced minimum modulus γ(C) is given by γ(C) = {kCxk x ∈ (Ker(C))⊥, kxk = 1} (cf. [17, Remark 1.2.10]). We list some properties of the reduced minimum modulus as our lemma as follows. Lemma 3.1 (cf. [17]). Let C be in B(H)\{0}, Then 9 (1) kCxk ≥ γ(C)kxk, ∀ x ∈ (Ker(C))⊥. (2) γ(C) = inf{λ λ ∈ σ(C)\{0}}, where C = (C ∗C)1/2. (3) γ(C) > 0 iff Ran(C) is closed iff 0 is an isolated point of σ(C) if 0 ∈ σ(C). (4) γ(C) = kC −1k−1 when C is invertible. (5) γ(C) ≥ kBk−1 when CBC = C for B ∈ B(H)\{0}. For a ∈ A+, put β(a) = inf{λ λ ∈ σ(a)\{0}}. Combining Lemma 3.1 with the n H , H and H1 = n Li=1 T 2 T1T2 1 T 2 T2T1 2 ··· ··· TnT1 T2T2 faithful representation of A, we can obtain Corollary 3.2. Let a ∈ A+. Then (1) β(a) > 0 if and only if 0 ∈ σ(a) is isolated when a 6∈ GL(A). (2) β(c) ≥ kck−1 when aca = a for some c ∈ A+\{0}. Let E be a C ∗ -- subalgebra of B(H) with the unit I. Let (T1,··· , Tn) be an n -- tuple Ran(Ti) ⊂ Ker(Ti) ⊂ H. Since H = Ran(Ti) ⊕ Ker(Ti), i = 1,··· , n, it of positive operators in E with Ran(Ti) closed, i = 1,··· , n. Put H0 = Li=1 follows that H0 ⊕ H1 = H. Put Tij = TiTj(cid:12)(cid:12)Ran(Tj ), i, j = 1,··· , n and set T =   ∈ Mn(E), T =   ··· T1n ··· T2n    ∈ B(H0), ··· ··· ··· Tnn (3.1) Clearly, H1 ⊂ Ker(T ) and it is easy to check that Ker(T ) = H1 when Ker( T ) = {0}. Thus, in this case, T can be expressed as T = (cid:20) T 0 0(cid:21) with respect to the orthogonal decomposition H = H0 ⊕ H1 and consequently, σ(T ) = σ( T ) ∪ {0}. Lemma 3.3. Let (T1,··· , Tn) be an n -- tuple of positive operators in E with Ran(Ti) closed, i = 1,··· , n. Let H0, H1, H be as above and T, T be given in (3.1). Suppose that T is invertible in B(H0). Then (1) σ( T ) = σ(cid:0) n Pi=1 (2) 0 is an isolated point in σ(cid:0) n Pi=1 (3) {T1a1,··· , Tnan} is linearly independent for any a1,··· , an ∈ E with Tiai 6= 0, ··· T1Tn ··· T2Tn ··· ··· T 2 ··· n Ti(cid:1) if 0 ∈ σ(cid:0) n Pi=1 T11 T12 T21 T22 ··· ··· Tn1 Tn2 T 2 i (cid:1)\{0}. 0 Ti(cid:1). n Li=1  i = 1,··· , n. 10 ∈ Mn(E). Then ZZ ∗ =   T1 0 ... 0 Proof. (1) Put Z =   Z ∗Z = T . Thus, σ(cid:0) n Pi=1 ··· Tn ··· 0 ... . . . ··· 0 i (cid:1)\{0} = σ(T )\{0} = σ( T ). (2) According to (1), 0 is an isolated point of σ(cid:0) n Pi=1 in E. So by Lemma 2.1, there is G ∈ E+ such that T 2 T 2 i n Pi=1   0 . . .   0 and T 2 i (cid:1) if n Pi=1 T 2 i is not invertible n T 2 n n n n T 2 n T 2 (cid:0) T 2 T 2 T 2 T 2 Xi=1 Xi=1 Xi=1 Xi=1 Xi=1 i , G(cid:0) i (cid:1)G = G, (cid:0) Xi=1 i (cid:1). i (cid:1)G(cid:0) i (cid:1) = i (cid:1)G ∈ E. Then P0 is a projection with Ran(P0) = Ker(cid:0) n Put P0 = I −(cid:0) n Pi=1 Pi=1 i (cid:1) = Ker(cid:0) n Noting that Ker(cid:0) n Ti(cid:1) = Pi=1 Pi=1 Ti=1 Pi=1 T 2 1≤i≤n kTik) i ≤ ( max i (cid:1). T 2 i ∈ GL((I −P0)E(I −P0)) Ti is invertible in i (cid:1)G = G(cid:0) with the inverse G and Ti(cid:1) when 0 ∈ σ(cid:0) n (I − P0)E(I − P0). Thus, 0 is an isolated point of σ(cid:0) n Pi=1 Pi=1 i ∈ E+ such that TiT † (3) By Lemma 3.1 (3) and Lemma 2.1, there is T † Ti, we get that Ker(Ti), Pi=1 n Pi=1 n Pi=1 T 2 T 2 n n n T † i TiT † i = T † i , T † i Ti = TiT † i , i = 1,··· , n. Thus, Ran(Ti) = Ran(TiT † n Let a1,··· , an ∈ E with Tiai 6= 0, i = 1,··· , n such that Pi=1 Ti(cid:1). i Ti = Ti, i ), i = 1,··· , n. λiTiai = 0 for some λ1,··· , λn ∈ C. For any ξ ∈ H, put x = since T is invertible. Thus, λiTiT † n Li=1 λiTiT † i aiξ ∈ H0. Then T x = 0 and x = 0 i aiξ = 0, ∀ ξ ∈ H and hence λi = 0, i = 1,··· , n. The following result duo to Levy and Dedplanques is very useful in Matrix Theory: Lemma 3.4 (cf. [7]). Suppose complex n × n self -- adjoint matrix C = [cij ]n×n is strictly diagonally dominant, that is, Pj6=icij < cii, i = 1,··· , n. Then C is invertible and positive. Proposition 3.5. Let T1,··· , Tn ∈ A+. Assume that (1) γ = min{β(T1),··· , β(Tn)} > 0 and (2) there exists ρ ∈ (0, γ] such that η = max{kTiTjk i 6= j, i, j = 1,··· , n} < (n − 1)−1ρ2. Then for any δ ∈ [η, (n − 1)−1ρ2), we have (1) σ(cid:0) n Pi=1 i (cid:1)\{0} ⊂ [ρ2 − (n − 1)δ, ρ2 + (n − 1)δ]. T 2 11 Li=1 ··· S1n ··· S2n ··· ··· ··· Snn Ti(cid:1). Ti(cid:1) if 0 ∈ σ(cid:0) n Pi=1 (2) 0 is an isolated point of σ(cid:0) n Pi=1 (3) (cid:0) n Ti(cid:1)A = T1A ∔ ··· ∔ TnA. Pi=1 Proof. (1) Let (ψ, Hψ) be a faithful representation of A with ψ(1) = I. We may assume that H = Hψ and E = ψ(A). Put Si = ψ(Ti), Sij = SiSj(cid:12)(cid:12)Ran(Sj ), i, j = 1,··· , n. Then max{kSiSjk 1 ≤ i 6= j ≤ n} = η and γ(Si) = β(Ti) by Lemma 3.1, 1 ≤ i ≤ n. Set H0 = Ran(Si) and n S =   S11 S12 S21 S22 ··· ··· Sn1 Sn2   Then for any λ < ρ2 − (n− 1)δ, we have Pj6=ikSijk ≤ (n− 1)η < ρ2 − λ. It follows from ρ2 − λ −kS12k ··· −kS1nk ··· −kS2nk −kS21k ··· ··· ··· ρ2 − λ −kSn1k −kSn2k ···   ∈ B(H0), S0 = Lemma 3.4 that S0 is positive and invertible. Therefore the quadratic form ρ2 − λ ···   , f (x1, x2,··· , xn) = n Xi=1 (ρ2 − λ)x2 i − 2 X1≤i<j≤n kSijkxixj is positive definite and hence there exists α > 0 such that for any (x1,··· , xn) ∈ Rn, f (x1,··· , xn) ≥ α(x2 1 + ··· + x2 n). So for any ξ = i = 1,··· , n, by Lemma 3.1 and n Li=1 ξi ∈ H0, kSiξik ≥ γ(Si)kξik ≥ ρkξik, ξi ∈ Ran(Si) = (Ker(Si))⊥, < ( S − λI)ξ, ξ > = ≥ n Xi=1 Xi=1 n n λkξik2 + X1≤i<j≤n Xi kSiξik2 − (ρ2 − λ)kξik2 − 2 X1≤i<j≤n Xi=1 k = f (kξ1k,··· ,kξkk) ≥ α kξik2. kSijkkξikkξjk (< Sijξj, ξi > + < S∗ ijξi, ξj >) Therefore, S − λI is invertible. Similarly, for any λ > ρ2 + (n − 1)δ, we can obtain that λI − S is invertible. So σ( S) ⊂ [ρ2 − (n − 1)δ, ρ2 + (n − 1)δ] ⊂ (0, ρ2 + (n − 1)δ] and consequently, i (cid:1)\{0} = σ(cid:0) n σ(cid:0) n Pi=1 Pi=1 Ti(cid:1) = σ(cid:0) n (2) Since σ(cid:0) n Pi=1 Pi=1 i(cid:1)\{0} ⊂ [ρ2 − (n − 1)δ, ρ2 + (n − 1)δ] by Lemma 3.3. Si(cid:1), the assertion follows from Lemma 3.3 (2). T 2 S2 12 (3) By (2) and Lemma 2.1, (cid:0) n Pi=1 Ti(cid:1)† ∈ A exists. Set E = (cid:0) n Pi=1 Ti(cid:1)(cid:0) n Pi=1 Ti(cid:1)†. Ti(cid:1)A ⊂ T1A + ··· + TnA for E(cid:0) n Pi=1 Obviously, EA = (cid:0) n Ti(cid:1) = Pi=1 Pi=1 Ti, we get that (1− E)Ti(1− E) ≤ (1− E)(cid:0) n Ti(cid:1)(1− E) = 0, i.e., Pi=1 Ti = ETi, i = 1,··· , n. So TiA ⊂ EA, i = 1,··· , n and hence From Ti ≤ Pi=1 Ti. n n T1A + ··· + TnA ⊂ EA = (cid:0) n Xi=1 Ti(cid:1)A ⊂ T1A + ··· + TnA. Since for any a1,··· , an ∈ A with Tiai 6= 0, i = 1,··· , n, {S1ψ(ai),··· , Snψ(an)} is linearly independent in E, we have {T1a1,··· , Tnan} is linearly independent in A. Therefore, (cid:0) n Pi=1 Ti(cid:1)A = EA = T1A ∔ ··· ∔ TnA. Let P1, P2 be projections on H. Buckholtz shows in [3] that Ran(P1) ∔ Ran(P2) = H iff kP1 + P2 − Ik < 1. For (P1,··· , Pn) ∈ Pn(A), we have Corollary 3.6. Let (P1,··· , Pn) ∈ Pn(A) with (cid:13)(cid:13) (P1,··· , Pn) is complete in A. Proof. For any i 6= j, Pi=1 n Pi − 1(cid:13)(cid:13) < (n − 1)−2. Then kPiPjk2 = kPiPjPik ≤ (cid:13)(cid:13)Pi(cid:0)Xk6=i Pk − 1(cid:1)Pi(cid:13)(cid:13) ≤ (cid:13)(cid:13) = (cid:13)(cid:13)Pi(cid:0) Pk(cid:1)Pi(cid:13)(cid:13) Xk=1 Xk=1 n n Pk − 1(cid:13)(cid:13) < 1 (n − 1)2 . Thus kPiPjk < (n − 1)−1. Noting that ρ = min{β(P1),··· , β(Pn)} = 1, η = max{kPiPjk 1 ≤ i < j ≤ n} < 1 n − 1 , we have (cid:0) n Pi(cid:1)A = P1A ∔ ··· ∔ PnA by Proposition 3.5. Pi=1 Pi − 1(cid:13)(cid:13) < (n − 1)−2, we have From (cid:13)(cid:13) Pi=1 A = P1A ∔ ··· ∔ PnA. Thus, (P1,··· , Pn) is complete in A. Pi=1 n n Pi is invertible in A and so that Combing Corollary 3.6 with Theorem 1.2 (3), we have Corollary 3.7. Let P1,··· , Pn be non -- trivial projections in B(H) with (cid:13)(cid:13) Pi−I(cid:13)(cid:13) < (n − 1)−2. Then H = Ran(P1) ∔ ··· ∔ Ran(Pn). Let (P1,··· , Pn) ∈ Pn(A). A well -- known statement says: "for any ǫ > 0, there is δ > 0 such that if kPiPjk < δ, i 6= j, i, j = 1,··· , n, then there are mutually orthogonal projections P ′ ik < ǫ, i = 1,··· , n". It may n ∈ A with kPi − P ′ 1,··· , P ′ Pi=1 n 13 be the first time appeared in Glimm's paper [5]. By using the induction on n, he gave its proof. But how δ depends on ǫ is not given. Lemma 2.5.6 of [8] states this statement and the author gives a slightly different proof. We can find from the proof of [8, Lemma 2.5.6] that the relation between δ and ǫ is δ ≤ (12)(n−1)n! ǫ . The next corollary will give a new proof of this statement with the relation δ = ǫ for ǫ ∈ (0, 1). n ǫ 2(n − 1) Corollary 3.8. Let (P1,··· , Pn) ∈ Pn(A). Given ǫ ∈ (0, 1). If P1,··· , Pn satisfy condition: kPiPjk < δ = , 1 ≤ i < j ≤ n, then there are mutually orthogonal projections P ′ 2(n − 1) 1,··· , P ′ Pi=1 n ∈ A such that kPi − P ′ Pi. Noting that γ = min{β(P1),··· , β(Pn)} = 1, kPiPjk < ik < ǫ, i = 1,··· , n. Proof. Set A = , n − 1 1 ≤ i < j ≤ n and taking ρ = 1, we have σ(A)\{0} ⊂ [1 − (n − 1)δ, 1 + (n − 1)δ] by Proposition 3.5 (1). So the positive element A† exists by Lemma 2.1. Set P = A†A = AA† ∈ A. From AA†A = A and A†AA† = A†, we get that Pi ≤ P , i = 1,··· , n and AP = P A = A, A†P = P A† = A†. So A ∈ GL(PAP ) with the inverse A† ∈ PAP . Let σP AP (A†) stand for the spectrum of A† in PAP . Then 1 σP AP (A†) = σ(A†)\{0} = {λ−1 λ ∈ σ(A)\{0}} ⊂ [(1 + (n − 1)δ)−1, (1 − (n − 1)δ)−1], (3.2) Now by Proposition 3.5, PA = AA = P1A ∔ ··· ∔ PnA. Thus, by using Pi ≤ P , i = 1,··· , n, we have PAP = P1(PAP ) ∔ ··· ∔ Pn(PAP ) and then PiA†Pj = δijPi, i = (A†)1/2Pi(A†)1/2 ∈ A, i = 1,··· , n. Then i, j = 1,··· , n by Theorem 1.2. Put P ′ 1,··· , P ′ P ′ n are mutually orthogonal projections and moreover, for 1 ≤ i ≤ n, kP ′ i − Pik ≤ k(A†)1/2Pi(A†)1/2 − Pi(A†)1/2k + kPi(A†)1/2 − Pik (3.3) Note that 0 < (n − 1)δ < 1/2. Applying Spectrum Mapping Theorem to (3.2), ≤ (k(A†)1/2k + 1)k(A†)1/2 − Pk. we get that k(A†)1/2k) ≤ (1 − (n − 1)δ)−1/2 < √2 and kP − (A†)1/2k ≤ (1 − (n − 1)δ)−1/2 − 1 < 2 1 + √2 (n − 1)δ. Thus kP ′ i − Pik < 2(n − 1)δ = ǫ by (3.3). Applying Theorem 1.2 and Corollary 3.8 to an n -- tuple of linear independent unit vectors, we have: Corollary 3.9. Let (α1,··· , αn) be an n -- tuple of linear independent unit vectors in Hilbert space H. (1) There is an invertible, positive operator K in B(H) and an n -- tuple of mutually orthogonal unit vectors (γ1,··· , γn) in H such that γi = Kαi, i = 1,··· , n. 14 (2) Given ǫ ∈ (0, 1). If < αi, αj > < , 1 ≤ i < j ≤ n, then there exists an n -- tuple of mutually orthogonal unit vectors (β1,··· , βn) in H such that kαi − βjk < 2ǫ, i = 1,··· , n. 2(n − 1) ǫ n Proof. Set H1 = span{α1,··· , αn} and Piξ =< ξ, αi > αi, ∀ ξ ∈ H1, i = 1,··· , n. Then (P1,··· , Pn) ∈ Pn(B(H1)) and Ran(P1) ∔ ··· ∔ Ran(Pn) = H1. Pi is invertible in B(H1) and PiA−1 By Theorem 1.2, A0 = 1,··· , n. Put K = A−1/2 αi, i = 1,··· , n, where P0 is the 0 projection of H onto H ⊥ 1 . It is easy to check that K is invertible and positive in B(H) with γi = Kαi, i = 1,··· , n and (γ1,··· , γn) is an n -- tuple of mutually orthogonal unit vectors. This proves (1). Pi=1 + P0 and γi = A−1/2 0 Pj = δijPi, i, j = 0 (2) Note that kPiPjk = < αi, αj > < ik < ǫ, i = 1,··· , n. Put β′ Corollary 3.8, there are mutually orthogonal projections P ′ kPi − P ′ mutually orthogonal and kαi − β′ Then < βi, βj >= δijβi, i, j = 1,··· , n and kαi − βik ≤ kαi − β′ ik < ǫ, i = 1,··· , n. Set βi = kβ′ ik + 1 − kβ′ ik < 2ǫ, i = P ′ ǫ 2(n − 1) , 1 ≤ i < j ≤ n. Thus, by n ∈ A such that 1,··· , β′ n are i, i = 1,··· , n. i αi, i = 1,··· , n. Then β′ ik−1β′ 1,··· , P ′ for i = 1,··· , n. Now we give a simple characterization of the completeness of a given n -- tuple of n projections in C ∗ -- algebra A as follows. Theorem 3.10. Let P1,··· , Pn be projections in A. Then (P1,··· , Pn) is complete Pi is invertible in A and kPiA−1Pjk < (cid:2)(n − 1)kA−1kkAk2(cid:3)−1, Pi=1 if and only if A = ∀ i 6= j, i, j = 1,··· , n. Proof. If (P1,··· , Pn) is complete, then by Theorem 1.2, A is invertible in A and PiA−1Pj = 0, ∀ i 6= j, i, j = 1,··· , n. Now we prove the converse. Put Ti = A−1/2PiA−1/2, 1 ≤ i ≤ n. Then Ti = 1. Since Ti(A1/2PiA1/2)Ti = Ti, we have β(Ti) ≥ kA1/2PiA1/2k−1 ≥ kAk−1, i = 1,··· , n by Corollary 3.2. Put ρ = kAk−1. Then Pi=1 n kTiTjk ≤ kA−1kkPiA−1Pjk < (cid:2)(n − 1)kAk2(cid:3)−1 = ρ2 n − 1 , i 6= j, i, j = 1,··· , n. Thus by Proposition 3.5 (3), A = T1A ∔ ··· ∔ TnA. Note that TiA = A−1/2(PiA), i = 1,··· , n. So P1A ∔ ··· ∔ PnA = A1/2A = A, i.e., (P1,··· , Pn) ∈ PCn(A). Corollary 3.11. Let (P1,··· , Pn) ∈ PCn(A) and let (P ′ that kPi − P ′ then (P ′ ik < (cid:2)4n2(n− 1)kA−1k2(nkA−1k + 1)(cid:3)−1, i = 1,··· , n, where A = n) ∈ Pn(A). Assume Pi, 1,··· , P ′ Pi=1 n 1,··· , P ′ n) ∈ PCn(A). 15 Proof. Set B = n Pi=1 i . Since nkA−1k ≥ kAkkA−1k ≥ 1, it follows that kA − Bk < P ′ 1 2kA−1k . Thus B is invertible in A with kB−1k ≤ kA−1k 1 − kA−1kkA − Bk < 2kA−1k, kB−1 − A−1k < 2kA−1k2kA − Bk. Note that PiA−1Pj = 0, i 6= j, i, j = 1,··· , n, we have kP ′ i B−1P ′ jk ≤ kP ′ i (B−1 − A−1)P ′ jk + k(P ′ i − Pi)A−1P ′ jk + kPiA−1(Pj − P ′ j)k ≤ 2kA−1k2kA − Bk + kA−1kkPi − P ′ < < 1 1 ik + kA−1kkPj − P ′ jk 2n2(n − 1)kA−1k (n − 1)kB−1kkBk2 . So (P ′ 1,··· , P ′ n) is complete in A by Theorem 3.10. 4 Some equivalent relations and topological properties on PCn(A) n Let A be a C ∗ -- algebra with the unit 1 and let GL0(A) (resp. U0(A)) be the connected component of 1 in GL(A) (resp. in U (A)). Set Xi=1 PIn(A) = (cid:8)(P1,··· , Pn) ∈ Pn(A) Xi=1 POn(A) = (cid:8)(P1,··· , Pn) ∈ Pn(A) Definition 4.1. Let (P1,··· , Pn) and (P ′ (1) We say (P1,··· , Pn) is equivalent to (P ′ Pi ∈ GL(A)(cid:9) Pi = 1, PiPj = 0, i 6= j, i, j = 1,··· , n(cid:9). 1,··· , P ′ n), denoted by (P1,··· , Pn) ∼ n) be in PCn(A). 1,··· , P ′ n (P ′ 1,··· , P ′ n), if there are U1,··· , Un ∈ A such that Pi = U ∗ i Ui, P ′ i = UiU ∗ i . (2) (P1,··· , Pn) and (P ′ (P1,··· , Pn) ∼u (P ′ i = 1,··· , n. 1,··· , P ′ 1,··· , P ′ n) are called to be unitarily equivalent, denoted by n), if there is U ∈ U (A) such that U PiU ∗ = P ′ i , (3) (P1,··· , Pn) and (P ′ (P1,··· , Pn) ∼h (P ′ PCn(A) such that F (0) = (P1,··· , Pn) and F (1) = (P ′ n) are called homotopically equivalent, denoted by n), if there exists a continuous mapping F : [0, 1] → 1,··· , P ′ 1,··· , P ′ 1,··· , P ′ n). It is well -- know that 1,··· , P ′ n) ⇒ (P1,··· , Pn) ∼ (P ′ 1,··· , P ′ n) (P1,··· , Pn) ∼h (P ′ and if U (A) is path -- connected, (P1,··· , Pn) ∼u (P ′ 1,··· , P ′ n) ⇒ (P1,··· , Pn) ∼h (P ′ 1,··· , P ′ n). 16 n Lemma 4.2. Let (P1,··· , Pn) be in PCn(A) and C be a positive and invertible element in A with PiC 2Pi = Pi, i = 1,··· , n. Then (CP1C,··· , CPnC) ∈ PCn(A) and (P1,··· , Pn) ∼h (CP1C,··· , CPnC) in PCn(A). Proof. From (CPiC)2 = CPiC 2PiC = CPiC, 1 ≤ i ≤ n, we have (CP1C,··· , CPnC) Pi=1 Pi ∈ GL(A) and PiA−1Pi = ∈ Pn(A). (P1,··· , Pn) ∈ PCn(A) implies that A = Pi, 1 ≤ i ≤ n by Theorem 1.2. So (CPiC)(cid:16) n (CPiC)(cid:17)−1 Pi=1 (CPiC) = CPiA−1PiC and hence (CP1C,··· , CPnC) ∈ PCn(A) by Theorem 1.2. Put Ai(t) = C tPiC t and Bi(t) = C −tPiC −t, ∀ t ∈ [0, 1], i = 1,··· , n. Then Qi(t) , Ai(t)Bi(t) = C tPiC −t is idempotent and Ai(t)Bi(t)Ai(t) = Ai(t), ∀ t ∈ [0, 1], i = 1,··· , n. Thus Ai(t)A = Qi(t)A, ∀ t ∈ [0, 1], i = 1,··· , n. By Lemma 2.3, Pi(t) , Qi(t)(Qi(t)+(Qi(t))∗−1)−1 is a projection in A satisfying Qi(t)Pi(t) = Pi(t) and Pi(t)Qi(t) = Qi(t), ∀ t ∈ [0, 1], i = 1,··· , n. Clearly, Ai(t)A = Qi(t)A = Pi(t)A, ∀ t ∈ [0, 1] and t 7→ Pi(t) is a continuous mapping from [0, 1] into A, i = 1,··· , n. Thus, from (C tP1C t)A ∔ ··· ∔ (C tPnC t)A = A, ∀ t ∈ [0, 1], we get that F (t) = (P1(t),··· , Pn(t)) ∈ PCn(A), ∀ t ∈ [0, 1]. Note that F : [0, 1] → PCn(A) is continuous with F (0) = (P1,··· , Pn). Note that Ai(1) = CPiC is a projection with Ai(1)Qi(1) = CPiCCPiC −1 = Qi(1) and Qi(1)Ai(1) = Ai(1), i = 1,··· , n. So Pi(1) = Ai(1), i = 1,··· , n and F (1) = (CP1C,··· , CPnC). The assertion follows. For (P1,··· , Pn) ∈ PCn(A), A = Pi ∈ GL(A) and Qi = A−1/2PiA−1/2 is a projection with QiQj = 0, i 6= j, i, j = 1,··· , n (see Theorem 1.2), that is, (Q1,··· , Qn) ∈ POn(A). Since C = A−1/2 satisfies the condition given in Lemma 4.2, we have the following: n Pi=1 Corollary 4.3. Let (P1,··· , Pn) ∈ PCn(A) and let (Q1,··· , Qn) be as above. Then (P1,··· , Pn) ∼h (Q1,··· , Qn) in PCn(A). Theorem 4.4. Let (P1,··· , Pn) and (P ′ statements are equivalent: n) ∈ PCn(A). Then the following 1,··· , P ′ 1,··· , P ′ n). (1) (P1,··· , Pn) ∼ (P ′ (2) there is D ∈ GL(A) such that for 1 ≤ i ≤ n, PiDD∗Pi = Pi and P ′ (3) there is (S1,··· , Sn) ∈ PCn(A) such that i = D∗PiD. (P1,··· , Pn) ∼u (S1,··· , Sn) ∼h (P ′ 1,··· , P ′ n). 17 n n Pi, A′ = i = 1,··· , n. Put A = Proof. The implication (3)⇒(1) is obvious. We now prove the implications (1)⇒(2) and (2)⇒(3) as follows. Pi=1 Xi=1 W ∗W = A′−1/2(cid:0) Xi=1 = A′−1/2(cid:0) i Ui = Pi, UiU ∗ (1)⇒ (2) Let Ui ∈ A be partial isometries such that U ∗ i and W = A−1/2(cid:0) n Pi=1 PiU ∗ Xi=1 i(cid:1)A′−1/2 i UiPi(cid:1)A−1(cid:0) i P ′ Xi=1 i(cid:1)A′−1/2 = A′−1/2(cid:0) i P ′ P ′ i UiPiU ∗ i = P ′ i , i(cid:1)A′−1/2. Then i P ′ i(cid:1)A′−1/2 = 1. Pi=1 PiU ∗ P ′ P ′ P ′ n n n n Similarly, W W ∗ = 1. Thus, W ∈ U (A). Set D = A−1/2W A′1/2 ∈ GL(A). Then, for 1 ≤ i ≤ n, D∗PiD = (cid:0) n Xi=1 P ′ i UiPi(cid:1)A−1PiA−1(cid:0) n Xi=1 PiU ∗ i(cid:1) = P ′ i P ′ i UiPiU ∗ i P ′ i = P ′ i and PiDD∗Pi = Pi follows from (D∗PiD)2 = D∗PiD. (2)⇒(3) Put U = (DD∗)−1/2D. Then U ∈ U (A). Set C = U ∗(DD∗)1/2U and Si = U ∗PiU , 1 ≤ i ≤ n. Then (S1,··· , Sn) ∈ PCn(A) with (S1,··· , Sn)∼u (P1,··· , Pn) and (P ′ n) = (CS1C,··· , CSnC). Since SiC 2Si = U ∗PiDD∗PiU = Si, i = 1,··· , n, it follows from Lemma 4.2 that 1,··· , P ′ n) ∼h (S1,··· , Sn) in PCn(A). 1,··· , P ′ (P ′ Proposition 4.5. For Pn(A), PCn(A), PIn(A) and POn(A), we have (1) PIn(A) is open in Pn(A). (2) PCn(A) is a clopen subset of PIn(A). (3) POn(A) is a strong deformation retract of PCn(A). (4) PCn(A) is locally connected. Thus every connected component of PCn(A) is path -- connected. (5) (P1,··· , Pn), (P ′ 1,··· , P ′ n) ∈ PCn(A) are in the same connected component iff there is D ∈ GL0(A) such that Pi = D∗P ′ i D, i = 1,··· , n. Proof. (1) Since h(P1,··· , Pn) = and GL(A) is open in A, it follows that PIn(A) = h−1(GL(A)) is open in Pn(A). Pi is a continuous mapping from Pn(A) into A n Pi=1 (2) Define F : PIn(A) → R by F (P1,··· , Pn) = X1≤i<j≤n (n − 1)(cid:13)(cid:13) n Xi=1 2(cid:13)(cid:13)(cid:0) Pi(cid:13)(cid:13) n Xi=1 Pi(cid:1)−1(cid:13)(cid:13)(cid:13)(cid:13)Pi(cid:0) n Xi=1 Pi(cid:1)−1Pjk. 18 Clearly, F is continuous on PIn(A). By means of Theorem 3.10, we get that PCn(A) = F −1((−1, 1)) is open in PIn(A) and PCn(A) = F −1({0}) is closed in PIn(A). (3) Define the continuous mapping r : PCn(A) → POn(A) by r(P1,··· , Pn) = (cid:0)A−1/2P1A−1/2,··· , A−1/2PnA−1/2(cid:1), A = n Xi=1 Pi. by Theorem 1.2. Clearly, r(P1,··· , Pn) = (P1,··· , Pn) when (P1,··· , Pn) ∈ POn(A). This means that POn(A) is a retract of PCn(A). For any t ∈ [0, 1] and i = 1,··· , n, put Hi(P1,··· , Pn, t) = A−t/2PiAt/2(A−t/2PiAt/2 + At/2PiA−t/2 − 1)−1. Similar to the proof of Lemma 4.2, we have H(P1,··· , Pn, t) = (H1(P1,··· , Pn, t),··· , Hn(P1,··· , Pn, t)) is a continuous mapping from PCn(A) × [0, 1] to PCn(A) with H(P1,··· , Pn, 0) = (P1,··· , Pn) and H(P1,··· , Pn, 1) = r(P1,··· , Pn). Furthermore, when (P1,··· , Pn) ∈ POn(A), A = 1. In this case, H(P1,··· , Pn, t) = (P1,··· , Pn), ∀ t ∈ [0, 1]. There- fore, POn(A) is a strong deformation retract of PCn(A). (4) Let (P1,··· , Pn) ∈ PCn(A). Then by Corollary 3.11, there is δ ∈ (0, 1/2) such that for any (R1,··· , Rn) ∈ Pn(A) with kPi − Rik < δ, 1 ≤ i ≤ n, we have (R1,··· , Rn) ∈ PCn(A). Let (R1,··· , Rn) ∈ PCn(A) with kPj − Rjk < δ, i = 1,··· , n. put Pi(t) = Pi, Ri(t) = Ri and ai(t) = (1 − t)Pi + tRi, ∀ t ∈ [0, 1], i = 1,··· , n. Then Pi, Ri, ai are self -- adjoint elements in C([0, 1],A) = B and kPi − aik = max t∈[0,1]kPi − ai(t)k < δ, i = 1,··· , n. It follows from [17, Lemm 6.5.9 (1)] that there exists a projection fi ∈ C ∗(ai) (the C ∗ -- subalgebra of B generated by ai) such that kPi−fik ≤ kPi−aik < δ, i = 1,··· , n. Thus, kPi − fi(t)k < δ, i = 1,··· , n and consequently, F (t) = (f1(t),··· , fn(t)) is a continuous mapping of [0, 1] into PCn(A). Since ai(0) = Pi, ai(1) = Ri and fi(t) ∈ C ∗(ai(t)), ∀ t ∈ [0, 1], we have f (0) = (P1,··· , Pn) and f (1) = (R1,··· , Rn). This means that PCn(A) is locally path -- connected. (5) There is a continuous path P (t) = (P1(t),··· , Pn(t)) ∈ PCn(A), ∀ t ∈ [0, 1] such that P (0) = (P1,··· , Pn) and P (1) = (P ′ n). By [12, Corollary 5.2.9.], there is a continuous mapping t 7→ Ui(t) of [0, 1] into U (A) with Ui(0) = 1 such that Pi(t) = Ui(t)P1U ∗ 1,··· , P ′ i (t), ∀ t ∈ [0, 1] and i = 1,··· , n. Set i (t)Pi(t)(cid:17)(cid:16) n Xi=1 i (t)(cid:17)1/2 Pi(cid:17)−1/2(cid:16) n Xi=1 W (t)(cid:16) n Pi=1 W (t) = (cid:16) n Xi=1 Pi(cid:17)−1/2 PiU ∗ and D(t) = (cid:16) n Pi=1 , ∀ t ∈ [0, 1]. Then W (t) ∈ U (A) with W (0) = 1, D(t) ∈ GL(A) with D(0) = 1 and W (t), D(t) are all continuous on Ui(t)PiU ∗ Ui(t)PiU ∗ i (t)(cid:17)−1/2 19 [0, 1] with D∗(t)PiD(t) = Pi(t) (see the proof of (1)⇒(2) in Theorem 4.4), ∀ t ∈ [0, 1] and i = 1,··· , n. Put D = D(1). Then D ∈ GL0(A) and D∗PiD = P ′ i , i = 1,··· , n. i , i = 1,··· , n. Then i U ∗ = (DD∗)1/2Pi(DD∗)1/2, Conversely, if there is D ∈ GL0(A) such that D∗PiD = P ′ U = (DD∗)−1/2D ∈ U0(A) and PiDD∗Pi = Pi, U P ′ 1U ∗,··· , U P ′ i = 1,··· , n. Thus, (P ′ n) ∼h (U P ′ 1,··· , P ′ nU ∗) and ((DD∗)1/2P1(DD∗)1/2,··· , (DD∗)1/2Pn(DD∗)1/2) ∼h (P1,··· , Pn) by Lemma 4.2. Consequently, (P ′ n) ∼h (P1,··· , Pn). As ending of this section, we consider following examples: 1,··· , P ′ n n−1 Pi=1 Pi=1 Example 4.6. Let A = Mk(C), k ≥ 2. Define a mapping ρ : PCn(A) → Nn−1 by ρ(P1,··· , Pn) = (Tr(P1),··· , Tr(Pn−1)), where 2 ≤ n ≤ k and Tr(·) is the canonical trace on A. By Theorem 1.2, (P1,··· , Pn) ∈ PCn(A) means that A = Pi ∈ GL(A) and (A−1/2P1A−1/2,··· , A−1/2PnA−1/2) ∈ POn(A). Put Qi = A−1/2PiA−1/2, i = 1,··· , n. Since (P1,··· , Pn) ∼h (Q1,··· , Qn) by Corollary 4.3, it follows that Tr(Pi) = Tr(Qi), i = 1,··· , n and Tr(A) = k. Thus Tr(Pn) = k − Pi. Note that U (A) is path -- connected. So, for (P1,··· , Pn), (P ′ n) ∈ PCn(A), 1,··· , P ′ n) are in the same connected component if and only if 1,··· , P ′ n). (P1,··· , Pn) and (P ′ ρ(P1,··· , Pn) = ρ(P ′ The above shows that PCk(A) is connected and PCn(A) is not connected when k ≥ 3 and 2 ≤ n ≤ k − 1. Example 4.7. Let H be a separable complex Hilbert space and K(H) be the C ∗ -- algebra of all compact operators in B(H). Let A = B(H)/K(H) be the Calkin algebra and π : B(H) → A be the quotient mapping. Then PCn(A) is path -- connected. In fact, if (P1,··· , Pn), (P ′ 1,··· , Q′ 1,··· , Q′ n) ∈ PCn(A), then we can find (Q1,··· , Qn), n) ∈ POn(A) such that (P1,··· , Pn) ∼h (Q1,··· , Qn) and (P ′ (Q′ n) ∼h (Q′ n) by Corollary 4.3. Since B(H) is of real rank zero, it follows from [17, Corollary B.2.2] or [16, Lemma 3.2] that there are projections R1,··· , Rn and R′ 1,··· , R′ n in B(H) such that π(Ri) = Qi, π(R′ 1,··· , P ′ 1,··· , P ′ 1,··· , P ′ i) = Q′ RiRj = δijRi, R′ iR′ j = δijR′ i, i, j = 1,··· , n, n i, i = 1,··· , n and Xi=1 Xi=1 Ri = R′ n i = I. n Pi=1 1,··· , R′ Vi. Then V ∈ U (B(H)) and V RiV ∗ = R′ Note that R1,··· , Rn, R′ tors V1,··· , Vn in B(H) such that V ∗ V = U (A). Then (U Q1U ∗,··· , U QnU ∗) = (Q′ is path -- connected, we have (Q1,··· , Qn) ∼h (Q′ (P1,··· , Pn) ∼h (P ′ n 6∈ K(H). So there are partial isometry opera- i, i = 1,··· , n. Put i, i = 1,··· , n. Put U = π(V ) ∈ 1,··· , Qn) in POn(A). Since U (B(H)) n) in PCn(A). Finally, n). This means that PCn(A) is path -- connected. i Vi = Ri, ViV ∗ 1,··· , Q′ 1,··· , P ′ i = R′ 20 References [1] P.E. Bjørstad and J. Mandel, On the spectra of sums of orthogonal projections with applications to parallel computing, BIT, 31 (1991), 76 -- 88. [2] A. Bottcher and I.M. Spitkovsky, A gentle guide to the basics of two projections theory, Linear Algebra Appl., 432 (2010), 1412 -- 1459. [3] D. Buckholtz, Hilber space idempotents and involutions, Proc. Amer. Math. Soc., 128 (2000), 1415 -- 1418. [4] G. Chen and Y. Xue, The expression of generalized inverse of the perturbed operators under type I perturbation in Hilbert spaces, Linear Algebra Appl., 285 (1998), 1 -- 6. [5] J. Glimm, On a certain class of operator algebras, Trans. Amer. Math. Soc., 95 (1960), 318 -- 340. [6] P. Halmos, Two subspace, Tran. Amer. Math. Soc., 144 (1969), 381 -- 389. [7] R.A. Horn and R. Johnson, Matrix Analysis, Cambridge University Press, 1986. [8] H. Lin, An Intruduction to the Classification of Amenable C ∗-Algebras, World Scintific, 2001. [9] J.J. Koliha and V. Rakocevi´c, Invertibility of the sum of idempotents, Linear and Multilinear Algebra, 50 (2002), 285 -- 292. [10] J.J. Koliha and V. Rakocevi´c, Invertibility of the difference of idempotents, Lin- ear and Multilinear Algebra 51 (2003), 97 -- 110. [11] J.J. Koliha and V. Rakocevi´c, Fredholm properties of the difference of orthogonal projections in a Hilbert space, Integr. Equ. Oper. Theory, 52 (2005), 125 -- 134. [12] N.E. Wegge -- Olsen, K -- Theory and C*-Algebras, A Friendly Approach, OUP, 1993. [13] T. Shulman, On universal C ∗-algebras generated by n projections with scalar sum, Amer. Math. Soc. 137(1) (2009), 115 -- 122. [14] V. S. Sunder, N subspaces, Can. J. Math., XL (1) (1988), 38 -- 54. [15] N.L. Vasilevski, C ∗ -- algebras generated by orthogonal projections and their ap- plications, Integr. Equ. Oper. Theory, 31 (1998), 113 -- 132. [16] Y. Xue, The reduced minimum modulus in C ∗ -- algebras, Integr. Equ. Oper. Theory, 59 (2007), 269 -- 280. [17] Y. Xue, Stable Perturbations of Operators and Related Topics, World Scientific, 2012. 21
1302.5922	1	1302	2013-02-24T16:33:25	Factors from trees	[ "math.OA" ]	We construct factors of type $\tn$ for $n\in\NN, n\geq 2$ from group actions on homogeneous trees and their boundaries. Our result is a discrete analogue of a result of R.J Spatzier, where the hyperfinite factor of type $\tone$ is constructed from a group action on the boundary of the universal cover of a manifold.	math.OA	math	FACTORS FROM TREES JACQUI RAMAGGE AND GUYAN ROBERTSON Abstract. We construct factors of type III1/n for n ∈ N, n ≥ 2 from group actions on homogeneous trees and their boundaries. Our result is a discrete analogue of a result of R.J Spatzier, namely [S, Proposition 1], where the hyperfinite factor of type III1 is con- structed from a group action on the boundary of the universal cover of a manifold. 1. Introduction Let Γ be a group acting simply transitively on the vertices of a ho- mogeneous tree T of degree n + 1 < ∞. Then, by [FTN, Ch. I, Theo- rem 6.3], Γ ∼= Z2 ∗ · · · ∗ Z2 ∗ Z ∗ · · · ∗ Z where there are s factors of Z2, t factors of Z, and s + 2t = n + 1. Thus Γ has a presentation Γ = (cid:10)a1, . . . , as+t : a2 i = 1 for i ∈ {1, . . . , s}(cid:11) , we can identify the Cayley graph of Γ constructed via right multipli- cation with T and the action of Γ on T is equivalent to the natural action of Γ on its Cayley graph via left multiplication. We can associate a natural boundary to T , namely the set Ω of semi- infinite reduced words in the generators of Γ. The action of Γ on T induces an action of Γ on Ω. For each x ∈ Γ, let Ωx = {ω ∈ Ω : ω = x · · · } be the set of semi-infinite reduced words beginning with x. The set {Ωx}x∈Γ is a set of basic open sets for a compact Hasudorff topology on Ω. Denote by x the length of a reduced expression for x. Let V m = {x ∈ Γ : x = m} and define Nm = V m. Then Ω is the disjoint union of the Nm sets Ωx for x ∈ V m. We can also endow Ω with the structure of a measure space. Ω has a unique distinguished Borel probability measure ν such that ν (Ωx) = 1 n + 1 (cid:18) 1 n(cid:19)x−1 1 2 JACQUI RAMAGGE AND GUYAN ROBERTSON for every nontrivial x ∈ Γ. The sets Ωx, x ∈ Γ generate the Borel σ-algebra. This measure ν on Ω is quasi-invariant under the action of Γ, so that Γ acts on the measure space (Ω, ν) and enabling us to extend the action of Γ to an action on L∞(Ω, ν) via g · f (ω) = f (g−1 · ω) for all g ∈ Γ, f ∈ L∞(Ω, ν), and ω ∈ Ω. We may therefore consider the von Neumann algebra L∞(Ω, ν) ⋊ Γ which we shall write as L∞(Ω) ⋊ Γ for brevity. 2. The Factors We note that the action of Γ on Ω is free since if gω = ω for some g ∈ Γ and ω ∈ Ω then we must have either ω = ggg · · · or ω = g−1g−1g−1 · · · and ν(cid:0)ggg · · · , g−1g−1g−1 · · ·(cid:1) = 0. The action of Γ on Ω is also ergodic by the proof of [PS, Propo- sition 3.9], so that L∞(Ω) ⋊ Γ is a factor. Establishing the type of the factor is not quite as straightforward. We begin by recalling some classical definitions. Definition 2.1. Given a group Γ acting on a measure space Ω, we define the full group, [Γ], of Γ by [Γ] = {T ∈ Aut(Ω) : T ω ∈ Γω for almost every ω ∈ Ω} . The set [Γ]0 of measure preserving maps in [Γ] is then given by [Γ]0 = {T ∈ [Γ] : T ◦ν = ν} Definition 2.2. Let G be a countable group of automorphisms of the measure space (Ω, ν). Following W. Krieger, define the ratio set r(G) to be the subset of [0, ∞) such that if λ ≥ 0 then λ ∈ r(G) if and only if for every ǫ > 0 and Borel set E with ν(E) > 0, there exists a g ∈ G and a Borel set F such that ν(F ) > 0, F ∪ gF ⊆ E and for all ω ∈ F . dν◦g dν (cid:12)(cid:12)(cid:12)(cid:12) (ω) − λ(cid:12)(cid:12)(cid:12)(cid:12) < ǫ Remark 2.3. The ratio set r(G) depends only on the quasi-equivalence class of the measure ν, see [HO, §I-3, Lemma 14]. It also depends only on the full group in the sense that [H] = [G] ⇒ r(H) = r(G). FACTORS FROM TREES 3 The following result will be applied in the special case where G = Γ. However, since the simple transitivity of the action doesn't play a role in the proof, we can state it in greater generality. Proposition 2.4. Let G be a countable subgroup of Aut(T ) ≤ Aut(Ω). Suppose there exist an element g ∈ G such that d(ge, e) = 1 and a subgroup K of [G]0 whose action on Ω is ergodic. Then r(G) = (cid:8)nk : k ∈ Z} ∪ {0(cid:9) . Proof. By Remark 2.3, it is sufficient to prove the statement for some group H such that [H] = [G]. In particular, since [G] = [hG, Ki] for any subgroup K of [G]0, we may assume without loss of generality that K ≤ G. By [FTN, Chapter 2, part 1)], for each g ∈ G and ω ∈ Ω we have dν◦g dν (ω) ∈ (cid:8)nk : k ∈ Z} ∪ {0(cid:9) . Since G acts ergodically on Ω, r(G) \ {0} is a group. It is therefore enough to show that n ∈ r(G). Write x = ge and note that νx = ν◦g−1. By [FTN, Chapter 2, part1)] we have (1) dνx dν (ω) = n, for all ω ∈ Ωx e . Let E ⊆ Ω be a Borel set with ν(E) > 0. By the ergodicity of K, there exist k1, k2 ∈ K such that the set F = {ω ∈ E : k1ω ∈ Ωx e and k2g−1k1ω ∈ E} has positive measure. Finally, let t = k2g−1k1 ∈ G. By construction, F ∪ tF ⊆ E. More- over, since K is measure-preserving, dν◦t dν (ω) = dν◦g−1 dν (k1ω) = dνx dν (k1ω) = n for all ω ∈ F by (1), since k1 ∈ Ωx e . This proves n ∈ r(G), as required. (cid:3) Corollary 2.5. If, in addition to the hypotheses for Proposition 2.4, the action of G is free, then L∞(Ω) ⋊ G is a factor of type III1/n. Proof. Having determined the ratio set, this is immediate from [C1, Corollaire 3.3.4]. (cid:3) Thus, if we can find a countable subgroup K ≤ [Γ]0 whose action on Ω is ergodic we will have shown that L∞(Ω) ⋊ Γ is a factor of type III1/n. To this end, we prove the following sufficiency condition for ergodicity. 4 JACQUI RAMAGGE AND GUYAN ROBERTSON Lemma 2.6. Let K be group which acts on Ω. If K acts transitively on the collection of sets {Ωx : x ∈ Γ, x = m} for each natural number m, then K acts ergodically on Ω. Proof. Suppose that X0 ⊆ Ω is a Borel set which is invariant under K and such that ν(X0) > 0. We show that this necessarily implies ν(Ω \ X0) = 0, thus establishing the ergodicity of the action. Define a new measure µ on Ω by µ(X) = ν(X ∩ X0) for each Borel set X ⊆ Ω. Now, for each g ∈ K, µ(gX) = ν(gX ∩ X0) = ν(X ∩ g−1X0) ≤ ν(X ∩ X0) + ν(X ∩ (g−1X0 \ X0)) = ν(X ∩ X0) = µ(X), and therefore µ is K-invariant. Since K acts transitively on the basic open sets Ωx associated to words x of length m this implies that µ(Ωx) = µ(Ωy) whenever x = y. Since Ω is the union of Nm disjoint sets Ωx, x ∈ V m, each of which has equal measure with respect to µ, we deduce that µ(Ωx) = c Nm for each x ∈ V m, where c = µ(X0) = ν(X0) > 0. Thus µ(Ωx) = cν(Ωx) for every x ∈ Γ. Since the sets Ωx, x ∈ Γ generate the Borel σ-algebra, we deduce that µ(X) = cν(X) for each Borel set X. Therefore ν(Ω \ X0) = c−1µ(Ω \ X0) = c−1ν((Ω \ X0) ∩ X0) = 0, thus proving ergodicity. (cid:3) In the last of our technical results, we give a constructive proof of the existence of a countable ergodic subgroup of [Γ]0. Lemma 2.7. There is a countable ergodic group K ≤ Aut(Ω) such that K ≤ [Γ]0. Proof. Let x, y ∈ V m. We construct a measure preserving automor- phism kx,y of Ω such that (1) kx,y is almost everywhere a bijection from Ωx onto Ωy, (2) kx,y is the identity on Ω \ (Ωx ∪ Ωy). FACTORS FROM TREES 5 It then follows from Lemma 2.6 that the group K = hkx,y : {x, y} ⊆ V m, m ∈ Ni acts ergodically on Ω and the construction will show explicitly that K ≤ [Γ]0. Fix x, y ∈ V m and suppose that we have reduced expressions x = x1 . . . xm, and y = y1 . . . ym. Define kx,y to be left multiplication by yx−1 on each of the sets Ωxz where z = 1 and z /∈ {x−1 m }. Then kx,y is a measure preserving bijection from each such set onto Ωyz. If ym = xm then kx,y is now well defined everywhere on Ωx. m , y−1 Suppose now that ym 6= xm. Then kx,y is defined on the set Ωx \Ωxy−1 , which it maps bijectively onto Ωy \ Ωyx−1 m . Now define kx,y to be left m z where z = 1 multiplication by yx−1 and z /∈ {xm, ym}. Then kx,y is a measure preserving bijection of each such Ωxy−1 m ymx−1 on each of the sets Ωxy−1 m z onto Ωyx−1 m z. m on the set Ωx \ Ωxy−1 Thus we have extended the domain of kx,y so that it is now defined m ym. m ymx−1 on the m xm, which it maps bijectively onto Ωy \ Ωyx−1 Next define kx,y to be left muliplication by yx−1 m ymx−1 sets Ωxy−1 m xmz where z = 1 and z /∈ {x−1 m , y−1 m }. Continue in this way. At the jth step kx,y is a measure preserving bijection from Ωx \ Xj onto Ωy \ Yj where ν(Xj) → 0 as j → ∞ so that eventually kx,y is defined almost everywhere on Ω. Finally, define kx,y(xy−1 m xmy−1 m xmy−1 m xm . . .) = yx−1 m ymx−1 m ymx−1 m ym . . . thus defining kx,y everywhere on Ω in such a way that its action is pointwise approximable by Γ almost everywhere. Hence K = hkx,y : {x, y} ⊆ V m, m ∈ Ni is a countable group with an ergodic measure-preserving action on Ω and K ≤ [Γ]0. (cid:3) We are now in a position to prove our main result. Theorem 2.8. The von Neumann algebra L∞(Ω) ⋊ Γ is the hyperfinite factor of type III1/n. Proof. By applying Corollary 2.5 with G = Γ, g ∈ Γ any generator of Γ, and K as in Lemma 2.7 we conclude that L∞(Ω) ⋊ Γ is a factor of type III1/n. To see that the factor is hyperfinite simply note that the action of Γ is amenable as a result of [A, Theorem 5.1]. We refer to [C2, 6 JACQUI RAMAGGE AND GUYAN ROBERTSON Theorem 4.4.1] for the uniqueness of the hyperfinite factor of type III1/n. (cid:3) Remark 2.9. Taking different measures on Ω should yield hyperfinite factors of type IIIλ for any 0 < λ < 1. We have concentrated on the geometrically interesting case. Remark 2.10. In [Sp1], Spielberg constructs IIIλ factor states on the algebra O2. The reduced C ∗-algebra C(Ω) ⋊r Γ is a Cuntz-Krieger algebra OA by [Sp2]. What we have done is construct a type III1/n factor state on some of these algebras OA. Remark 2.11. From [C2, p. 476], we know that if Γ = Q ⋊ Q∗ acts naturally on Qp, then the crossed product L∞(Qp)⋊Γ is the hyperfinite factor of type III1/p. This may be proved geometrically as above by regarding the the boundary of the homogeneous tree of degree p + 1 as the one point compactification of Qp as in [CKW]. References [A] S. Adams. Boundary amenability for word hyperbolic groups and an ap- plication to smooth dynamics of simple groups. Preprint 1993. [C1] [CKW] D. I. Cartwright, V. Kaimanovich, and W. Woess. Random walks on the affine group of local fields and homogeneous trees. M.S.R.I. Preprint No. 022-94, 1993. A. Connes. Une classification des facteurs de type III. Ann. Scient. Ec. Nrom. Sup., 6 (1973), pp. 133 -- 252. A. Connes. On the classification of von Neumann algebras and their au- tomorphisms. Symposia Mathematica XX, pp. 435 -- 478, Academic Press 1978. [C2] [HO] [FTN] A. Fig`a-Talamanca and C. Nebbia, Harmonic Analysis and Representa- tion Theory for Groups Acting on Homogeneous Trees, Cambridge Uni- versity Press, London 1991. T. Hamachi and M. Osikawa, Ergodic Groups Acting of Automorphisms and Krieger's Theorems, Seminar on Mathematical Sciences No. 3, Keio University, Japan, 1981. C. Pensavalle and T. Steger, Tensor products and anisotropic principal series representations for free groups. Preprint, University of Georgia, June 1993. R. J. Spatzier, An example of an amenable action from geometry, Ergod. Th. & Dynam. Sys. (1987), 7, 289 -- 293. J. Spielberg, Diagonal states on O2, Pac. J. Maths. 144 (1990), 351 -- 382. J. Spielberg, Free product groups, Cuntz-Krieger algebras and covariant maps, Int. J. Maths. (1991), 2, 457 -- 476. [Sp1] [Sp2] [PS] [S] FACTORS FROM TREES 7 Mathematics Department, University of Newcastle, Callaghan, NSW 2308, Australia E-mail address: [email protected] Mathematics Department, University of Newcastle, Callaghan, NSW 2308, Australia E-mail address: [email protected]
1201.0792	1	1201	2012-01-04T00:36:49	Amenability for Fell bundles over groupoids	[ "math.OA" ]	We establish conditions under which the universal and reduced norms coincide for a Fell bundle over a groupoid. Specifically, we prove that the full and reduced C*-algebras of any Fell bundle over a measurewise amenable groupoid coincide, and also that for a groupoid G whose orbit space is T_0, the full and reduced algebras of a Fell bundle over G coincide if the full and reduced algebras of the restriction of the bundle to each isotropy group coincide.	math.OA	math	AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS AIDAN SIMS AND DANA P. WILLIAMS Abstract. We establish conditions under which the universal and reduced norms coincide for a Fell bundle over a groupoid. Specifically, we prove that the full and reduced C ∗-algebras of any Fell bundle over a measurewise amenable groupoid coincide, and also that for a groupoid G whose orbit space is T0, the full and reduced algebras of a Fell bundle over G coincide if the full and reduced algebras of the restriction of the bundle to each isotropy group coincide. Contents Introduction 1. Fell bundles 2. Amenable groupoids 3. The disintegation theorem revisited 4. Fell bundles over amenable groupoids 5. Measurewise amenable groupoids 6. Fibrewise-amenable Fell bundles Appendix A. Nondegenerate Borel ∗-functors References 1 2 4 4 5 8 9 11 12 Introduction If G is an amenable group, then the reduced crossed product and full crossed product for any action of G on a C∗-algebra coincide. This result was proved for discrete groups by Zeller-Meyer in [20] and in general by Takai in [18]. Since the C∗- algebra of a Fell bundle over a groupoid G is a very general sort of crossed product by G, it is reasonable to expect the universal norm and reduced norm to coincide on Γc(G; B) when G is suitably amenable. Immediately the situation is complicated because amenability for groupoids is not as clear cut as it is for groups. There are three reasonable notions of amenability for a second countable locally compact Hausdorff groupoid: (topological) amenability, measurewise amenability and, for lack of a better term, "metric amenability" by which we simply mean that the reduced norm and universal norm on Cc(G) coincide. Amenability implies measurewise amenability which in turn implies metric amenabil- ity. While there are situations where the converses hold, it is unknown if they hold in general. Our main result here, Theorem 1, is that if G is measurewise amenable as defined in [1], then the reduced norm and universal norm on Γc(G; B) coincide Date: 22 December 2011. 2000 Mathematics Subject Classification. 46L55. Key words and phrases. Groupoid; Fell bundle; amenable; reduced C ∗-algebra. 1 2 AIDAN SIMS AND DANA P. WILLIAMS for any Fell bundle B over G. This result subsumes the usual result for group dynamical systems and the result for groupoid dynamical systems; for a discus- sion of this, see [16, Examples 10 and 11]. The result for groupoid systems is also asserted in [1, Proposition 6.1.10] where they cite [15, Theorem 3.6]. Since it is usually hard to determine if a groupoid found in the wild is amenable in any given one the three flavors mentioned above, we also prove in Theorem 4 that groupoids which act nicely on their unit spaces in the sense that that G\G(0) is T0 and whose stability groups are all amenable are themselves measurewise amenable. This result may be known to experts, but seems worth advertising. We also show that if G is a groupoid whose orbit space is T0 and if B is a Fell bundle over G such that the full and reduced C∗-algebras of the restriction of B to each isotropy group in G coincide, then the full and reduced C∗-algebras of the whole bundle coincide. This is a formally stronger result than the combination of Theorem 4 and Theorem 1: there are many examples of Fell bundles over non-amenable groups whose full and reduced C∗-algebras coincide (see, for example, [3]). We start with very short sections on Fell bundles and amenable groupoids to clarify our definitions and point to the relevant literature. In Section 3 we point out a simple strengthening of the disintegration theorem for Fell bundles (from [10]) which is needed here. For readability, the details are shifted to Appendix A. In Section 4 we prove our main theorem. In Section 6 we show that groupoids with T0 orbit space and amenable stability groups are measurewise amenable. In Section 6 we prove that bundles over groupoids with T0 orbit space whose restrictions to isotropy groups are metrically amenable are themselves metrically amenable. Since we appeal to the disintegration theorem for Fell bundles, we require sep- arability for our results. In particular, all the groupoids and spaces that appear will be assumed to be second countable, locally compact and Hausdorff. Except when it is clearly not the case, for example B(H) and other multiplier algebras, all the algebras and Banach spaces that appear are separable. We also assume that our Fell bundles are always saturated. The underlying Banach bundles are only required to be upper semicontinuous. 1. Fell bundles We will refer to [10, §1] for details of the definition of a Fell bundle p : B → G over a groupoid as well as of the construction of the associated C∗-algebra C∗(G, B). (The examples in [10, §2] would be very helpful supplementary reading.) Roughly speaking, a Fell bundle p : B → G is an upper-semicontinuous Banach bundle endowed with a partial multiplication compatible with p such that the fibres A(u) over units u are C∗-algebras and such that each fibre B(x) is an A(r(x)) -- A(s(x))- imprimitivity bimodule with respect to the inner products and actions induced by the multiplication on B. In particular, when x and y are composable, multiplication in B implements isomorphisms B(x) ⊗A(s(x)) B(y) ∼= B(xy). The space Γc(G; B) of continuous sections of B then carries a natural convolution and involution. The C∗-algebra C∗(G, B) is the completion of Γc(G; B) with respect to the universal norm for representations which are continuous with respect to the inductive-limit topology on G. Regarding our notation: as above, we use a roman letter, B(x), for the fibre over x together with its Banach space structure, but we will use both A(u) and B(u) for the fibre over a unit u so as to distinguish its dual roles. The Fell bundle AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 3 axioms imply that A := Γ0(G(0); B) is a C∗-algebra which is called the C∗-algebra of B over G(0); in particular it is a C0(G(0))-algebra. So for u ∈ G(0) we write A(u) for the fibre over u when we are thinking of it as a C∗-algebra, and we write B(u) when we are thinking of it instead as an A(u) -- A(u)-imprimitivity bimod- ule. We assume that our Fell bundles are separable, so in addition to G being second countable, we assume that the Banach space Γ0(G; B) is separable. By axiom, our Fell bundles are saturated in that B(x)B(y) = B(xy), where B(x)B(y) denotes span{ bxby : bx ∈ B(x), by ∈ B(y) }. If F is a locally closed subset1 of G(0), then we abuse notation slightly and write Γc(F ; B) in place of Γc(F ; BF ) (as we have already done for A = Γ0(G(0); B) above). If we let G(F ) := GF = { x ∈ G : s(x) ∈ F and r(x) ∈ F } be the restriction of G to F , then G(F ) is a locally compact groupoid with Haar system {λu}u∈F . As above, we write C∗(G(F ), B) in place of C∗(G(F ), BG(F )). Recall the definition of the reduced norm on Γc(G; B) from [16]. If π is a representation of A = Γ0(G(0); B), then using [19, Example F.25] and the discussion preceding [11, Definition 7.9], we can assume that there is a Borel Hilbert bumdle G(0) ∗ H , a finite Radon measure µ on G(0) and representations πu of A on H(u), factoring through Au, such that π =Z ⊕ G(0) πu dµ(u). For u ∈ G(0), we frequently regard the πu as representations of A(u). Even if π is nondegenerate, we can only assume that µ-almost all of the πu are nondegenerate. Indeed, we could have πu = 0 for a null set of u. The formula (Ind π)(f )(g ⊗ h) = (f ∗ g) ⊗ h for f, g ∈ Γc(G; B) and h ∈ L2(G(0) ∗ H , µ) determines a representation Ind π of Γc(G; B) on the completion of Γc(G; B) ⊙ L2(G(0) ∗ H , µ) with respect to the inner product (f ⊗ h g ⊗ k) =(cid:0)π(g∗ ∗ f )h k(cid:1) =ZG(0)ZG(cid:0)πu(cid:0)g(x−1)∗f (x−1)(cid:1)h(u) k(u)(cid:1) dλu(x) dµ(u) =ZG(0)ZG(cid:0)πu(cid:0)g(x)∗f (x)(cid:1)h(u) k(u)(cid:1) dλu(x) dµ(u). (1) The reduced norm on Γc(G; B) is given by kf kr := sup{ k(Ind π)(f )k : π is a representation of A }. Since ker(Ind π) depends only on ker π (by, for example, [12, Corollary 2.73]) this definition of k · kr agrees with other definitions in the literature -- for example Exel's in [3] and Moutuou and Tu's in [8]. So C∗ r (G, B) is the quotient of C∗(G, B) by r (G,B) of Ind π for any faithful representation π of the C∗-algebra the kernel IC ∗ A = Γ0(G(0); B) of B over G(0). 1Recall that a subset of a locally compact Hausdorff space is locally compact if and only if it is locally closed and is locally closed if and only if it is the intersection of a closed set and an open set [19, Lemmas 1.25 and 1.26]. 4 AIDAN SIMS AND DANA P. WILLIAMS 2. Amenable groupoids Let G be a second-countable locally compact groupoid with Haar system λ. Re- nault [14, p. 92] originally defined G to be topologically amenable, or just amenable, if there is a net {fi} ⊂ Cc(G) such that (a) the functions u 7→ fi ∗ f ∗ (b) fi ∗ f ∗ i → 1 uniformly on compact subsets of G. i (u) are uniformly bounded on C0(G(0)), and Later, in the extensive treatment by Anatharaman-Delaroche and Renault, an a priori different definition was given: [1, Definition 2.2.8]; however, [1, Proposi- tion 2.2.13(iv)] and its proof show that the two notions of amenability are equiva- lent. It is not hard to see, using standard criteria such as [19, Proposition A.17], that a group is amenable as a groupoid if and only if it is amenable as a group. measure on G (that is, ν(·) = RG(0) λu(·) dµ(u)). Let µ be a quasi-invariant measure on G(0), and let ν := µ ◦ λ be the induced In [1, Definition 3.2.8], µ is called amenable if there exists a suitably invariant mean on L∞(G, ν). The pair (G, λ) is measurewise amenable if every quasi-invariant measure µ is amenable [1, Definition 3.2.8]. Since L∞(G, ν) depends only on the equivalence class of ν, if µ′ is equivalent to µ and µ is amenable, then so is µ′. Since [1] considers only σ-finite measures, to demonstrate that (G, λ) is measurewise amenable, it suffices to show that every finite quasi-invariant measure µ is amenable. It follows from [1, Theorem 2.2.17] and [1, Theoerem 3.2.16] that amenability and measurewise amenability, respectively, are preserved under groupoid equiva- lence. Theorem 17 of [17] implies that metric amenability is preserved as well. In particular, none of the three flavors of amenability of G depend on the choice of Haar system λ. In this note, we will use the characterization of amenability of (G, λ, µ) given in [1, Proposition 3.2.14(v)]. If G is amenable then it is measurewise amenable by [1, Proposition 3.3.5]. If G is measurewise amenable then it is metrically amenable by [1, Proposition 6.1.8]. 3. The disintegation theorem revisited Our main tool here is the disintegration theorem from [10]. Fix a nondegenerate representation L of C∗(G, B). Then [10, Theorem 4.13] implies that there are a quasi-invariant measure µ on G(0), a Borel Hilbert bundle G(0) ∗ H , and a Borel ∗-functor b 7→(cid:0)r(b), π(b), s(b)(cid:1) (see [10, Definition 4.5]) from B into End(G(0) ∗ H ) such that L is equivalent to the integrated form of the associated strict representa- tion (µ, G(0) ∗ H , π) of B. For h, k ∈ L2(G(0) ∗ H , µ) and f ∈ Γc(G; B), we then have (cid:0)L(f )h k(cid:1) =ZG(cid:16)π(f (x))h(s(x)) (cid:12)(cid:12) k(r(x)(cid:17)∆(x)− 1 2 dν(x). Regrettably, the authors of [10] neglected to point out that the Borel ∗-functor associated to L constructed in [10, Theorem 4.13] is nondegenerate in the sense that for all x ∈ G, (2) π(B(x))H(s(x)) = span{ π(b)v : b ∈ B(x) and v ∈ H(s(x)) } = H(r(x)). We outline why this is true in Appendix A, and at the same time, we tidy up some details of the proof of the disintegration theorem itself. AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 5 4. Fell bundles over amenable groupoids Our first main theorem says that every Fell bundle over a measurewise amenable groupoid is metrically amenable. Theorem 1. Let G be a second-countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0) . Suppose that p : B → G is a separable Fell bundle over G. If G is measurewise amenable, then the reduced norm on Γc(G; B) is equal to the universal norm, so C∗ r (G, B) = C∗(G, B). Our proof follows the lines of Renault's proof of the corresponding result for groupoid C∗-algebras, suitably modified for the bundle context. Before getting into the proof, we need to do a little set-up. We will continue with the following notation for the remainder of the section. Fix a ∈ IC ∗ r (G,B) and let L be a nondegenerate representation of C∗(G, B). As in Section 3, we may assume that L is the integrated form of a strict representation (cid:0)µ, G(0) ∗ H , π(cid:1) of B which is nondegenerate in the sense that (2) holds for all x. Fix a unit vector h in L2(G(0) ∗ H , µ), and let ωh be the associated vector state. To prove Theorem 1, it suffices to see that ωh(a) = 0. Let G ∗ Hr be the pullback of G(0) ∗ H over the range map. We may describe it as follows. Let (hj)∞ j=1 be a special orthorgonal fundamental sequence for G(0) ∗ H as in [19, Proposition F.6]. For each j, let hj(x) = hj(r(x)) ∈ H(r(x)). Then j=1. fundamental sequence (hj)∞ Let ν = µ ◦ λ be the measure on G induced by µ, and recall that ν−1 denotes G ∗ Hr is isomorphic to the Borel Hilbert bundle built from `x∈G H(r(x)) with the measure ν−1(f ) = RG f (x−1)dν(x). Since µ is quasi-invariant, ν and ν−1 are equivalent measures. By passing to an equivalent measure, we may assume that the Radon-Nikodym derivative ∆ = dν/dν−1 is multiplicative from G to (0, ∞) -- there is a nice proof of this in [9, Theorem 3.15].2 Lemma 2. Define U : Γc(G; B) ⊙ L2(G(0) ∗ H , µ) → L2(G ∗ Hr, ν−1) by U (f ⊗ h)(x) = π(cid:0)f (x)(cid:1)h(cid:0)s(x)(cid:1). Then U is isometric and extends to a unitary, also de- noted by U , from HInd πµ onto L2(G ∗ Hr, ν−1). Furthermore, U intertwines the regular representation Ind πµ with the representation Mπ of C∗(G, B) on L2(G ∗ H , ν−1) given on f ∈ Γc(G; B) by (3) (cid:0)Mπ(f )ξ η(cid:1) =ZGZG(cid:16)π(f (xy))ξ(y−1)(cid:12)(cid:12) η(x)(cid:17) dλs(x)(y) dν−1(x). Proof. That π is a Borel ∗-functor, f is a continuous section and (cid:0)U (f ⊗ h)(x) hj(x)(cid:1) =(cid:0)π(f (x))h(s(x)) hj(r(x)(cid:1) ∞ = Xk=1(cid:0)h(s(x)) hk(s(x))(cid:1)(cid:0)π(f (x))hk(s(x)) hj(r(x))(cid:1), imply that x 7→ (cid:0)U (f ⊗ h)(x) hj(x)(cid:1) is Borel. Thus U (g ⊗ h) ∈ B(G ∗ Hr). The representation πµ comes from a Borel ∗-functor defined on all of B, so the 2The proof in [9] unfortunatly remains unpublished, but it is based on Hahn's [4, Corol- lary 3.14]. 6 AIDAN SIMS AND DANA P. WILLIAMS formula (1) for the inner product on HInd πµ becomes (cid:0)f ⊗ h g ⊗ k(cid:1) =ZG(0)ZG(cid:0)π(cid:0)g(x)∗f (x)(cid:1)h(u) k(u)(cid:1) dλu(x) dµ(u) =ZG(cid:0)π(f (x))h(s(x)) π(g(x))k(s(x))(cid:1) dν−1(x) =ZG(cid:0)U (f ⊗ h)(x) U (g ⊗ k)(x)(cid:1) dν−1(x). Hence U (f ⊗ h) ∈ L2(G ∗ H , ν−1) and U is an isometry. Since π is nondegenerate, an argument like that of [19, Lemma F.17], shows that the range of U is dense. Hence, U is a unitary as claimed. For the last assertion, recall that Ind πµ(f ) acts by convolution. Thus (cid:0)Mπ(f )ξ η(cid:1) =ZG(cid:0)Mπ(f )ξ(x) η(x)(cid:1) dν−1(x) =ZGZG(cid:0)π(f (y))ξ(y−1x) η(x)(cid:1) dλr(x)(y) dν−1(x) =ZGZG(cid:0)π(f (xy))ξ(y−1) η(x)(cid:1) dλs(x)(y) dν−1(x). (cid:3) To prove Theorem 1, we invoke measurewise amenability in the form of [1, Propo- sition 3.2.14(v)]. So we fix a sequence (fn)∞ n=1 of Borel functions on G such that (a) u 7→RG fn(x)2 dλu(x) is bounded on G(0), n ∗ fn(u) ≤ 1 for all u ∈ G(0) and n ∗ fn → 1 in the weak-∗ topology on L∞(G, ν). (b) f ∗ (c) f ∗ To keep notation compact, we denote f ∗ n ∗ fn by en, so that for y ∈ G, en(y) =ZG fn(x−1)fn(x−1y) dλr(y)(x), Proof of Theorem 1. Recall the notation fixed at the beginning of the section: in particular, L is the integrated form of a nondegenerate strict representation of C∗(G, B) on a Hilbert bundle G(0) ∗ G, h is a unit vector in L2(G(0) ∗ H , µ) and ωh is the associated vector state. We claim that ωh(g) ≤ k(Ind πµ)(g)k for all g ∈ Γc(G; B). Fix g ∈ Γc(G; B). Then Define a sequence (αn)∞ n=1 of complex numbers by ωh(g) =(cid:0)L(g)h h(cid:1) =ZG(cid:0)π(g(y))h(s(y)) h(r(y))(cid:1)∆(y)− 1 αn :=ZG =ZG(0)ZGZG en(y)(cid:0)π(g(y))h(s(y)) h(r(y))(cid:1)∆(y)− 1 fn(x−1)fn(x−1y)(cid:0)π(g(y))h(s(y)) h(r(y))(cid:1)∆(y)− 1 2 dν(y) 2 2 dν(y). (4) dλu(x) dλu(y) dµ(u). (It is tempting to write ωh(eng) for αn, but the en are assumed only to be Borel, so the pointwise products eng may not belong to Γc(G; B).) By assumption on the en, the αn converge to ωh(g). So it suffices to show that (5) αn ≤ k(Ind πµ)(g)k for all n ∈ N. AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 7 Fix n ∈ N. Define hn : G → H by hn(x) = ∆(x) Then for each j, the function x 7→(cid:0)hn(x) hj(x)(cid:1) = ∆(x) 1 2 fn(x−1)h(cid:0)r(x)(cid:1). 2 fn(x−1)(cid:0)h(r(x)) hj(r(x))(cid:1) 1 is Borel, so hn ∈ L2(G ∗ H , ν−1). Starting from (4), we apply Fubini's theorem, substitute xy for x, and then use first that ν = ∆ν−1 and then that ∆ is multi- plicative to calculate: 2 dλs(x)(y) dν(x) αn =ZGZG fn(x−1)fn(y)(cid:0)π(g(xy))h(s(y)) h(r(x))(cid:1)∆(xy)− 1 =ZGZG fn(x−1)fn(y)(cid:0)π(g(xy))h(s(y)) h(r(x))(cid:1) =ZGZG(cid:0)π(g(xy))hn(y−1) hn(x)(cid:1) dλs(x)(y) dν−1(x) =(cid:0)Mπ(g)hn hn(cid:1). khnk2 =ZG khn(x)k2 dν−1(x) =ZG 2 ∆(x) dλs(x)(y) dν−1(x) ∆(xy)− 1 We have and since ν = ∆ν−1 and en(u) ≤ 1 for all u, it follows that fn(x−1)2kh(r(x))k2∆(x) dν−1(x), khnk2 =ZG(0) en(u)kh(u)k2 dµ(u) ≤ khk2 = 1. Hence the Cauchy-Schwarz inequality gives (5). Thus ωh(g) ≤ k(Ind πµ)(g)k for all g ∈ Γc(G; B) as claimed. Since Γc(G; B) is dense in C∗(G, B), it follows that ωh(a) ≤ k(Ind πµ)(a)k for all a ∈ C∗(G, B). In particular, if a ∈ IC ∗ required. r (G,B), then (Ind πµ)(a) = 0, and hence ωh(a) = 0 as (cid:3) Example 3. Recall from [2] that given a row-finite k-graph Λ with no sources, a Λ- system of C∗-correspondences consists of an assignment v 7→ Av of C∗-algebras to vertices and an assignment λ 7→ Xλ of an Ar(λ) -- As(λ) correspondence to each path λ, together with isomorphisms χµ,ν : Xµ ⊗As(µ) Xν → Xµν for each composable pair µ, ν ∈ Λ, all subject to an appropriate associativity condition on the χµ,ν (see [2, Definitions 3.1.1 and 3.1.2] for details). Suppose that X is such a system, and suppose that each Xλ is nondegenerate as a left Ar(λ)-module, and full as a right Hilbert As(λ)-module, and that the left action of Ar(λ) is by compact operators. By [2, Theorem 4.3.1], the construction of Sections 4.1 and 4.2 of the same paper associates to X a saturated Fell bundle EX over the k-graph groupoid GΛ of [7]. Moreover, [2, Theorem 4.3.6] says that the C∗-algebra C∗(A, X, χ) of the Λ-system is isomorphic to the reduced C∗-algebra C∗ r (GΛ, EX ) of the Fell bundle. Theorem 5.5 of [7] says that GΛ is amenable, and hence also measurewise r (GΛ, EX ) = C∗(GΛ, EX ); in par- amenable. Hence our Theorem 1 implies that C∗ ticular C∗(A, X, χ) ∼= C∗(GΛ, EX ). 8 AIDAN SIMS AND DANA P. WILLIAMS Since 1-graphs are precisely the path-categories E∗ of countable directed graphs E, and since an E∗-system of correspondences can be constructed from any assign- ment of C∗-algebras Av to vertices v, and Ar(e) -- As(e) C∗-correspondences Xe to edges e (see [2, Remark 3.1.5]), Example 3 provides a substantial library of examples of our result 5. Measurewise amenable groupoids Our initial motivation for proving Theorem 1 was to show that if G has T0 orbit space and amenable stability groups then the full and reduced C∗-algebras of any Fell bundle over G coincide: roughly, since C∗(G, B) is a C0(G\G(0))-algebra, representations will factor through restrictions to orbit groupoids G([u]), each of which is amenable because is is equivalent to the amenable stability group G(u) := {x ∈ G : r(x) = u = s(x)} (see section 6 for details). However, the following argument shows that the result follows directly from Theorem 1. We thank Jean Renault for pointing us in the direction of [1, Proposition 5.3.4]. Theorem 4. Suppose that G is a second countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0) . Suppose that the orbit space G\G(0) is T0 and that each stability group G(u) is amenable. Then G is measurewise amenable. Our proof requires some straightforward observations as well as some nontrivial results from [1]. Lemma 5. Suppose that µ is a quasi-invariant finite measure on G(0) and that F ⊂ G(0) is a locally compact G-invariant subset such that µ(G(0) \ F ) = 0. Then (G, λ, µ) is amenable if and only if (G(F ), λF , µF ) is amenable. Proof. Recall that (G, λ, µ) is amenable if there is an invariant mean on L∞(G, ν) where ν = µ◦λ. Since µF ◦λF = νG(F ), we have L∞(G, ν) ∼= L∞(G(F ), µF ◦λF ). In particular, an invariant mean on L∞(G) gives an invariant mean on L∞(G(F )) and vice versa. (cid:3) Lemma 6. Suppose that G\G(0) is T0. Then, as a Borel space, G\G(0) is countably separated and each orbit [u] is locally closed in G and hence locally compact. Proof. Since subsets of a locally compact Hausdorff space are locally compact if and only if they are locally closed (see [19, Lemma 1.26]), the lemma is an immediate consequence of the Mackey-Glimm-Ramsay dichotomy [13, Theorem 2.1]. (cid:3) Proof of Theorem 4. Suppose that µ is a finite quasi-invariant measure on G(0). It suffices to show that (G, λ, µ) is amenable. Let p : G → G\G(0) be the orbit map, and let µ be the forward image µ(f ) = µ(f ◦ p) of µ under p. By Lemma 6, G\G(0) is countably separated as a Borel space. Hence we can disintegrate µ -- as, for example, in [19, Theorem I.5] -- so that for each orbit [u] there is a probability measure ρ[u] on G(0) supported on [u] such that µ =ZG\G(0) ρ[u] dµ([u]). It follows from [1, Proposition 5.3.4] that ρ[u] is quasi-invariant for almost all [u] and that (G, λ, µ) is amenable if each (G, λ, ρ[u]) is. Since ρ[u](G(0) \ [u]) = 0, Lemma 5 implies that it is enough to see that each (G([u]), λ[u], µ[u]) is amenable. Since [u] is locally compact, G([u]) is a locally compact transitive groupoid equivalent to G(u), AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 9 which is assumed to be amenable. Hence [1, Theorem 2.2.13] implies that G([u]) is amenable, and therefore also measurewise amenable by [1, Proposition 3.3.5]. (cid:3) 6. Fibrewise-amenable Fell bundles In the preceding section, we showed that if G\G(0) is T0 and each stability group is amenable, then G is measurewise amenable. In particular, if p : B → G is a bundle over such a groupoid, then its full and reduced algebras coincide. In this section, we show that it suffices that G\G(0) is T0 and that for each u ∈ G(0), the full and reduced algebras of the restriction of B to the isotropy group G(u) coincide. To see that this is a strictly stronger theorem, and also that the hypothesis is genuinely checkable, we refer the reader to the results, for example, of [3]. Theorem 7. Let G be a second-countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0) , and let p : B → G be a separable Fell bundle over G. Suppose that the orbit space G\G(0) is T0 and that for each unit u, the full and reduced cross-sectional algebras C∗(G(u), B) and C∗ r (G(u), B) coincide. Then the full and reduced norms on Γc(G; B) are equal and hence C∗ r (G, B) = C∗(G, B). To prove the theorem, we first use the equivalence theorem of [16] to see that the full and reduced C∗-algebras of a Fell bundle over transitive groupoid coincide whenever the full and reduced algebras of its restriction to any isotropy group coincide. Lemma 8. Let G be a second-countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0) , and let p : B → G be a separable Fell bundle over G. Suppose that G is transitive. Then the following are equivalent. (a) For some unit u, the full and reduced cross-section algebras C∗(G(u), B) and C∗ r (G(u), B) coincide. (b) For every unit u, the full and reduced cross-section algebras C∗(G(u), B) (c) The full and reduced norms on Γc(G; B) are equal and hence C∗ r (G, B) = and C∗ r (G(u), B) coincide. C∗(G, B). Proof. Fix u ∈ G(0). Then Gu := s−1(u) is a (G, G(u))-equivalence, and as in [5, Theorem 1], E := p−1(Gu) implements an equivalence between B and p−1(G(u)). Consequently, [16, Theorem 14] implies that the natural surjection of C∗(G, B) onto C∗ r (G, B) is an isomorphism if and only if the kernel Ir of the natural map of C∗(G(u), B) onto C∗ r (G(u), B) is trivial. Since u ∈ G(0) was arbitrary, the result follows. (cid:3) To finish off our proof of Theorem 7, we need the following special case of [6, Theorem 3.7]. As above, let p : B → G be a separable Fell bundle over G with associated C∗-algebra A = Γ0(G(0); B). Recall from [6, Proposition 2.2] that G acts on Prim A which we identify with { (u, P ) : u ∈ G(0) and P ∈ Prim A(u) }. Let U be an open G-invariant subset of G(0) with complement F . Then { (u, P ) ∈ Prim A : u ∈ F } is a closed invariant subset of Prim(A), and corresponds to the G-invariant ideal { a ∈ A : a(u) = 0 for all u ∈ F } of A. By [6, Proposition 3.3], the corresponding bundle BI is the one with fibres BI (x) =(B(x) {0} if x ∈ G(U ) if u ∈ G(F ), 10 AIDAN SIMS AND DANA P. WILLIAMS so we can identify it with the bundle BG(U) over G(U ). Moreover, BI is the complementary bundle BI (x) =({0} B(x) if u ∈ G(U ) if u ∈ G(F ), which we may identify with the bundle BG(F ) over G(F ). Thus, as a special case of Theorem 3.7 of [6], we obtain the following result. Lemma 9. Let G be a second-countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0) , and let p : B → G be a separable Fell bundle over G. Suppose that U is a G-invariant open subset of G(0) with complement F . There is a short exact sequence of C∗-algebras 0 / C∗(G(U ), B) ι / C∗(G, B) q / C∗(G(F ), B) / 0, where ι is induced by inclusion and q by restriction on sections. As an application of Lemma 9, recall3 that there is a nondegenerate map M : C0(G(0)) → M (C∗(G, B)) given on sections by M (φ)f (x) = φ(cid:0)r(x)(cid:1)f (x). Suppose that the orbit space G\G(0) is Hausdorff. Then we may identify C0(G\G(0)) with the subalgebra of Cb(G(0)) consisting of functions which are constant on orbits and vanish at infinity on the orbit space. We extend M to Cb(G(0)) and restrict to C0(G\G(0)) to obtain a nondegenerate map of C0(G\G(0)) into the center of M (C∗(G, B)), making C∗(G, B) into a C0(G\G(0))-algebra. As usual, if u ∈ G(0), we let [u] be the corresponding orbit in G\G(0). Corollary 10. Let G be a second-countable locally compact Hausdorff groupoid with Haar system {λu}u∈G(0), and let p : B → G be a separable Fell bundle over G. If G\G(0) is Hausdorff, then C∗(G, B) is a C0(G\G(0))-algebra with fibres C∗(G, B)([u]) ∼= C∗(G([u]), B). Proof. Recall that C∗(G, B)([u]) is the quotient of C∗(G, B) by the ideal J[u] = span{ φ · a : φ ∈ C0(G\G(0)), φ([u]) = 0 and a ∈ C∗(G, B) }. Using Lemma 9, we can identify J[u] with C∗(G(U ), B), where U = G(0) \ [u], and C∗(G, B)/J[u] with C∗(G([u]), B) as claimed. (cid:3) Proof of Theorem 7. Fix in irreducible representation π of C∗(G; B) and an ele- ment f ∈ Γc(G; B). It suffices to show that kπ(f )k ≤ kf kC ∗ r (G;B). By [13, Theorem 2.1], the orbit space G\G(0) is locally Hausdorff and every orbit [u] is locally closed in G(0). Since G\G(0) is second countable, [19, Lemma 6.3] im- plies that there is a countable ordinal γ and a nested open cover { Un : 0 ≤ n ≤ γ } of G\G(0) such that U0 = ∅, Uγ = G\G(0) and Un+1 \ Un is Hausdorff (and dense) in (G\G(0)) \ Un. For n ≤ γ, let Vn := { u ∈ G(0) : [u] ∈ Un }. Then each Vn is an open invariant subset of G(0). Using Lemma 9, we can identify C∗(G(Vn), B) with an ideal in C∗(G, B). In fact, { C∗(G(Vn), B) }n≤γ is a composition series of ideals in C∗(G, B). By [19, Lemma 8.13], there exists 0 < n ≤ γ such that π lives on the subquotient C∗(G(Vn), B)/C∗(G(Vn−1), B); that is, π is the canon- ical lift ¯ρ of an irreducible representation ρ of the ideal C∗(G(Vn), B) such that 3A proof can be constructed along the lines of [10, Proposition 4.2]. / / / / AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 11 ker ρ ⊃ C∗(G(Vn−1), B). Lemma 9 implies that C∗(G(Vn), B)/C∗(G(Vn−1), B) ∼= C∗(G(Vn \ Vn−1), B). By construction, G(Vn \ Vn−1) has Hausdorff orbit space Un \ Un−1. Hence C∗(G(Vn \ Vn−1), B) is a C0(Un \ Un−1)-algebra and ρ fac- tors through a fibre C∗(G(Vn \ Vn−1), B)([u]) ∼= C∗(G([u]), B) for some u ∈ Vn by [19, Proposition C.5]. Since, by assumption, C∗(G[u]), B) = C∗ r (G([u]), B), we have ker(Ind π[u]) ⊂ ker ρ where π[u] factors through a faithful representation of the quotient AVn ([u]) of the C∗-algebra A(Vn) of C∗(G(Vn), B) corresponding to the closed set [u] ⊂ Vn. (Note that A(Vn) is the ideal of A corresponding to Vn ⊂ G(0).) The kernel of π = ¯ρ is then contained in the kernel of the canoni- cal lift of Ind π[u] to C∗(G, B). It is not hard to check that the canonical lift of Ind π[u] to C∗(G, B) is Ind ¯π[u] where ¯π[u] is the canonical lift of π[u] to A. Hence kπ(f )k = k ¯ρ(f )k ≤ k Ind ¯π[u](f )k ≤ kf kC ∗ (cid:3) r (G,B). Appendix A. Nondegenerate Borel ∗-functors Let p : B → G be a Fell bundle over a second-countable locally compact Haus- dorff groupoid, and let G(0) ∗ H be a Borel Hilbert bundle. A Borel ∗-functor π from B to End(G(0) ∗ H ) is a map π : b 7→(cid:0)r(b), π(b), s(b)(cid:1) such that π(b) ∈ B(cid:0)H(s(b)), H(r(b))(cid:1) for all b and such that π respects adjoints and the partial linear and multiplicative structure of B (see [10, Definition 4.5]). Following [10, §4], a strict representation of B is a triple (µ, G(0) ∗ H , π) consisting of a quasi-invariant measure µ on G(0), a Borel Hilbert bundle G(0) ∗ H and a Borel ∗-functor π. It is common practice to use π and π interchangeably, and we will drop the caret henceforth. A strict representation determines a bounded representation via integration (see [10, Proposition 4.10]); indeed, a Borel ∗-functor defined on p−1(GF ) for any µ-conull set F ⊂ G(0) is sufficient. Nevertheless, it is convenient to have π defined everywhere. The purpose of this section is to point out that the disintegration theorem [10, Theorem 4.13] for Fell bundles can be strengthened to assert that π can be taken to be nondegenerate as defined in §3. At the same time, we correct an error in the construction of π in [10]. In the proof of [10, Theorem 4.13], starting from a pre-representation L of B on a dense subspace H0 of a Hilbert space H, the authors showed that for any orthonormal basis {ζi : i ∈ N} for span{L(f )ξ : f ∈ Γc(G; B), ξ ∈ H0}, setting H′ 00 := span{ζi : i ∈ N}, there is a saturated Borel µ-conull set F ⊂ G(0) and a Borel Hilbert bundle F ∗ H whose fibres H(u) are Hilbert-space completions of Γc(Gu; B) ⊙ H′ 00 (see [10, Lemma 5.18] and [10, Lemma 5.20]). For f ∈ Γc(G; B) and h ∈ H′ 00, the class of f ⊗ h in H(u) is denoted by f ⊗u h. The space H(u) may be trivial for some u. For each z ∈ GF , b ∈ B(z) and f ∈ Γc(G; B), let π(b)f denote a section satisfying π(b)f (x) = ∆(z) 1 2 bf (z−1x) for x ∈ Gr(b). The Borel ∗-functor in the disintegration of L constructed in [10] is defined by π(b)(f ⊗s(b) ζi) = π(b)f ⊗r(b) ζi. Since F is saturated, G is the disjoint union of GF and GG(0)\F . Since the latter is ν-null, we can extend π to all of G by defining it as we please on p−1(GG(0)\F ), and this will not affect the integrated representation. To ensure that π is still a 12 AIDAN SIMS AND DANA P. WILLIAMS genuine Borel ∗-functor, one sets H(u) := {0} for each u /∈ F and π(b) := 0 for b /∈ p−1(GF ). (In [10], the authors mistakenly let (G(0) \ F ) ∗ H be a constant field and let π(b) be the identity operator, but such a π is not a ∗-functor since as it doesn't preserve the partial linear structure.) We claim that π is nondegenerate in the sense that (2) holds for all z ∈ G. It holds trivially for z 6∈ GF , so fix z ∈ GF , and let u := r(z). We start with two observations. (A) If fi → f in the inductive limit topology on Γc(Gu; B) then fi ⊗u ζk → f ⊗u ζk in H(u). To see this, observe that equation (5.19) of [10] is bounded by Kkf k∞kgk∞ where K is constant depending only on supp f and supp g. (B) If { ei } is an approximate identity in A(u), and, for each i, eig represents any section in Γc(G; B) such that (eig)(x) = eig(x) for x ∈ Gu, then eig → g in the inductive limit topology on Γc(Gu; B). This follows from a compactness argument using that A(u) acts nondegenerately on B(x). By (B), to establish (2) for z, it suffices to see that each eig ⊗r(z) ζk belongs to π(B(z)H(s(z)). Fix b1, . . . bn ∈ B(z) such that bjb∗ j ∼ ei. Xj Then by (A), we have and this suffices. π(bj)(π(b∗ j )g ⊗ ζk) ∼ eig ⊗ ζk, Xj Remark 11. Just as ∗-functors are automatically bounded (see [10, Remark 4.6]), there is a sense in which the Borel ∗-functor appearing in any strict representation (µ, G(0) ∗ H , π) is essentially nondegenerate. We claim that (6) for all x ∈ G. π(cid:0)B(x)(cid:1)H(cid:0)(s(x)(cid:1) = π(cid:0)A(r(x)(cid:1)H(cid:0)(r(x)(cid:1) The right-hand side of (6) is the essential space of the representation πr(x) of A(cid:0)r(x)(cid:1) determined by π, so (2) holds whenever πr(x) is nondegenerate. So if the representation πµ of A = Γ0(G(0); B) determined by π is nondegenerate, then πu is nondegenerate for µ-almost all u, so (2) holds on a ν-conull subset of G (where, as usual, ν = µ ◦ λ). To establish (6), we use that B is saturated: one the one hand, while on the other hand, π(cid:0)B(x)(cid:1)H(cid:0)s(x)(cid:1) = π(cid:0)A(r(x)(cid:1)(cid:1)π(cid:0)B(x)(cid:1)H(cid:0)s(x)(cid:1) ⊂ π(cid:0)A(r(x)(cid:1)(cid:1)H(cid:0)r(x)(cid:1), π(cid:0)B(x)(cid:1)H(cid:0)s(x)(cid:1) ⊃ π(cid:0)B(x)(cid:1)π(cid:0)B(x∗)(cid:1)H(cid:0)r(x)(cid:1) = π(cid:0)A(cid:0)r(x)(cid:1)(cid:1)H(cid:0)r(x)(cid:1). References [1] Claire Anantharaman-Delaroche and Jean Renault, Amenable groupoids, Monographies de L'Enseignement Math´ematique [Monographs of L'Enseignement Math´ematique], vol. 36, L'Enseignement Math´ematique, Geneva, 2000. With a foreword by Georges Skandalis and Appendix B by E. Germain. [2] Valentin Deaconu, Alex Kumjian, David Pask, and Aidan Sims, Graphs of C ∗- correspondences and Fell bundles, Indiana U. Math. J. 59 (2011), 1687 -- 1735. [3] Ruy Exel, Amenability for Fell bundles, J. reine angew. Math. 492 (1997), 41 -- 73. [4] Peter Hahn, Haar measure for measure groupoids, Trans. Amer. Math. Soc. 242 (1978), 1 -- 33. AMENABILITY FOR FELL BUNDLES OVER GROUPOIDS 13 [5] Marius Ionescu and Dana P. Williams, A classic Morita equivalence result for Fell bundle C ∗-algebras, Math. Scand. 108 (2011), 251 -- 263. [6] , Remarks on the ideal structure of Fell bundle C ∗-algebras, Houston J. Math. (2012), in press. (arXiv:math.OA.0912.1124). [7] Alex Kumjian and David Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20. [8] El-Kaıoum M. Moutuou and Jean-Louis Tu, Equivalence of fell systems and their reduced C ∗-algebras, preprint, 2011. (arXiv:math.OA.1101.1235v1). [9] Paul S. Muhly, Coordinates in operator algebra, CMBS Conference Lecture Notes (Texas Christian University 1990), 1999. In continuous preparation. [10] Paul S. Muhly and Dana P. Williams, Equivalence and disintegration theorems for Fell bun- dles and their C ∗-algebras, Dissertationes Math. (Rozprawy Mat.) 456 (2008), 1 -- 57. [11] , Renault's equivalence theorem for groupoid crossed products, NYJM Monographs, vol. 3, State University of New York University at Albany, Albany, NY, 2008. Available at http://nyjm.albany.edu:8000/m/2008/3.htm. [12] Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Mathematical Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. [13] Arlan Ramsay, The Mackey-Glimm dichotomy for foliations and other Polish groupoids, J. Funct. Anal. 94 (1990), 358 -- 374. [14] Jean Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer-Verlag, New York, 1980. [15] , The ideal structure of groupoid crossed product C ∗-algebras, J. Operator Theory 25 (1991), 3 -- 36. [16] Aidan Sims and Dana P. Williams, An equivalence theorem for reduced Fell bundle C ∗- algebras, preprint, 2011. (arXiv:math.OA.1111.5753v1). [17] , Renault's equivalence theorem for reduced groupoid C ∗-algebras, J. Operator Theory (2012), in press. (arXiv:math.OA.1002.3093). [18] Hiroshi Takai, On a duality for crossed products of C ∗-algebras, J. Funct. Anal. 19 (1975), 25 -- 39. [19] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Monographs, vol. 134, American Mathematical Society, Providence, RI, 2007. [20] Georges Zeller-Meier, Produits crois´es d'une C ∗-alg`ebre par un groupe d'automorphismes, J. Math. Pures Appl. (9) 47 (1968), 101 -- 239. School of Mathematics and Applied Statistics, University of Wollongong, NSW 2522, Australia E-mail address: [email protected] Department of Mathematics, Dartmouth College, Hanover, NH 03755-3551 E-mail address: [email protected]
1311.1193	1	1311	2013-11-05T20:53:26	Noncommutative solenoids and their projective modules	[ "math.OA" ]	Let p be prime. A noncommutative p-solenoid is the C-algebra of Z[1/p] x Z[1/p] twisted by a multiplier of that group, where Z[1/p] is the additive subgroup of the field Q of rational numbers whose denominators are powers of p. In this paper, we survey our classification of these C-algebras up to -isomorphism in terms of the multipliers on Z[1/p], using techniques from noncommutative topology. Our work relies in part on writing these C-algebras as direct limits of rotation algebras, i.e. twisted group C*-algebras of the group Z^2 thereby providing a mean for computing the K-theory of the noncommutative solenoids, as well as the range of the trace on the K_0 groups. We also establish a necessary and sufficient condition for the simplicity of the noncommutative solenoids. Then, using the computation of the trace on K_0, we discuss two different ways of constructing projective modules over the noncommutative solenoids.	math.OA	math	NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES FR´ED´ERIC LATR´EMOLI`ERE AND JUDITH A. PACKER pi × Zh 1 pi twisted by a multiplier of that group, where Zh 1 Abstract. Let p be prime. A noncommutative p-solenoid is the C ∗-algebra of Z h 1 pi is the additive subgroup of the field Q of rational numbers whose denominators are powers of p. In this paper, we survey our classification of these C ∗-algebras up to pi, using techniques -isomorphism in terms of the multipliers on Zh 1 from noncommutative topology. Our work relies in part on writing these C ∗- algebras as direct limits of rotation algebras, i.e. twisted group C-algebras of the group Z2, thereby providing a mean for computing the K-theory of the noncommutative solenoids, as well as the range of the trace on the K0 groups. We also establish a necessary and sufficient condition for the simplicity of the noncommutative solenoids. Then, using the computation of the trace on K0, we discuss two different ways of constructing projective modules over the noncommutative solenoids. pi × Zh 1 3 1 0 2 v o N 5 ] . A O h t a m [ 1 v 3 9 1 1 . 1 1 3 1 : v i X r a 1. Introduction Twisted group algebras and transformation group C∗-algebras have been studied since the early 1960's [8] and provide a rich source of examples and problems in C- algebra theory. Much progress has been made in studying such C∗-algebras when the groups involved are finitely generated (or compactly generated, in the case of Lie groups). Even when G = Zn, these C∗-algebras give a rich class of examples which have driven much development in C-algebra theory, including the foundation of noncommutative geometry by Connes [2], the extensive study of the geometry of quantum tori by Rieffel [14, 16, 17, 18], the expansion of the classification problem from AF to AT algebras by G. Elliott and D. Evans [4], and many more (L. Baggett and A. Kleppner [1], and S. Echterhoff and J. Rosenberg [3]). In this paper, we present our work on twisted group C∗-algebras of the Cartesian square of the discrete group Zh 1 pi of p-adic rationals, i.e. the additive subgroup of Q whose elements have denominators given by powers of a fixed p ∈ N, p ≥ 1. The Pontryagin duals of these groups are the p-solenoid, thereby motivating our terminology in calling these C∗-algebras noncommutative solenoids. We review our computation of the K-groups of these C∗-algebras, derived in their full technicality in [10], and which in and of itself involves an intriguing problem in the theory of Abelian group extensions. We were also able to compute the range of the trace Date: March 31, 2013. 1991 Mathematics Subject Classification. Primary 46L40, 46L80; Secondary 46L08, 19K14. Key words and phrases. C-algebras; solenoids; projective modules; p-adic analysis. 1 2 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER on the K0-groups, and use this knowledge to classify these C∗-algebras up to ∗- isomorphism, in [10], and these facts are summarized in a brief survey of [10] in the first two sections of this paper. This paper is concerns with the open problem of classifying noncommutative solenoids up to Morita equivalence. We demonstrate a method of constructing an equivalence bimodule between two noncommutative solenoids using methods due to M. Rieffel [16], and will note how this method has relationships to the theory of wavelet frames. These new matters occupy the last three sections of this paper. Acknowledgments. The authors gratefully acknowledge helpful conversations with Jerry Kaminker and Jack Spielberg. 2. Noncommutative Solenoids This section and the next provide a survey of the main results proven in [10] concerning the computation of the K-theory of noncommutative solenoids and its application to their classification up to -isomorphism. An interesting connection between the K-theory of noncommutative solenoids and the p-adic integers is un- earthed, and in particular, we prove that the range of the K0 functor on the class of all noncommutative solenoids is fully described by all Abelian extensions of the group of p-adic rationals by Z. These interesting matters are the subject of the next section, whereas we start in this section with the basic objects of our study. We shall fix, for this section and the next, an arbitrary p ∈ N with p > 1. Our story starts with the following groups: Definition 2.1. Let p ∈ N, p > 1. The group Zh 1 inductive limit of the sequence of groups: pi of p-adic rationals is the Z z7→pz −−−−→ Z z7→pz −−−−→ Z z7→pz −−−−→ Z z7→pz −−−−→ · · · which is explicitly given as the group: (2.1) endowed with the discrete topology. p(cid:21) =(cid:26) z pk ∈ Q : z ∈ Z, k ∈ N(cid:27) Z(cid:20) 1 From the description of Zh 1 pi as an injective limit, we obtain the following Proposition 2.2. Let p ∈ N, p > 1. The Pontryagin dual of the group Zh 1 pi is the result by functoriality of the Pontryagin duality. We denote by T the unit circle {z ∈ C : z = 1} in the field C of complex numbers. p-solenoid group, given by: endowed with the induced topology from the injection Sp ֒→ TN. The dual pairing between QN and SN is given by: n+1 = zn(cid:9) , Sp =(cid:8)(zn)n∈N ∈ TN : ∀n ∈ N zp pk , (zn)n∈N(cid:29) = zq (cid:28) q pi and (zn)n∈N ∈ Sp. k, where q pk ∈ Zh 1 We study in [10] the following C-algebras. NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 3 Definition 2.3. A noncommutative solenoid is a C-algebra of the form C∗(cid:18)Z(cid:20) 1 p(cid:21) × Z(cid:20) 1 p(cid:21), σ(cid:19) , where p is a natural number greater or equal to 2 and σ is a multiplier of the group The first matter to attend in the study of these C-algebras is to describe all are T-valued unless otherwise specified, with T the unit circle in C. Note that the pi × Zh 1 pi. Zh 1 the multipliers of the group Zh 1 group Zh 1 Using [8], we compute in [10] the group H 2(cid:16)Zh 1 pi up to equivalence, as follows: tipliers of Zh 1 pi up to equivalence, where our multipliers pi has no nontrivial multiplier, so our noncommutative solenoids are the pi, T(cid:17) of T-valued mul- pi × Zh 1 pi × Zh 1 pi × Zh 1 natural object to consider. Theorem 2.4. [10, Theorem 2.3] Let p ∈ N, p > 1. Let: Ξp = {(αn) : α0 ∈ [0, 1) ∧ (∀n ∈ N ∃k ∈ {0, . . . , N − 1} pαn+1 = αn + k)} which is a group for the pointwise addition modulo one. There exists a group iso- and α = ρ(σ), and if f is a multiplier of class σ, then f is cohomologous to: morphism ρ : H 2(cid:16)Zh 1 Ψα :(cid:18)(cid:18) q1 pk1 , pi × Zh 1 pk2(cid:19) ,(cid:18) q3 pi, T(cid:17) → Ξp such that if σ ∈ H 2(cid:16)Zh 1 pk4(cid:19)(cid:19) 7−→ exp(cid:0)2iπα(k1+k4)q1q4(cid:1) . pk3 q4 q2 , pi × Zh 1 pi, T(cid:17) For any p ∈ N, p > 1, the groups Ξp and Sp are obviously isomorphic as topo- logical groups; yet it is easier to perform our computations in the additive group Ξp pi, T(cid:17) is isomorphic in what follows. Thus, as a topological group, H 2(cid:16)Zh 1 pi × Zh 1 to Sp. Moreover, we observe that a corollary of Theorem (2.4) is that Ψα and Ψβ are cohomologous if and only if α = β ∈ Ξp. The proof of Theorem (2.4) involves the standard calculations for cohomology classes of multipliers on discrete Abelian groups, due to A. Kleppner, generalizing results of Backhouse and Bradley. With this understanding of the multipliers of Zh 1 pi, we thus propose pi, σ(cid:17). Let us start by to classify the noncommutative solenoids C∗(cid:16)Zh 1 is the C∗-completion of the involutive Banach algebra (cid:0)ℓ1 (Γ) , ∗σ, ·∗(cid:1), where the recalling [20] that for any multiplier σ of a discrete group Γ, the C-algebra C∗ (Γ, σ) pi × Zh 1 pi × Zh 1 twisted convolution ∗σ is given for any f1, f2 ∈ ℓ1 (Γ) by f1(γ1)f2(γ − γ1)σ(γ1, γ − γ1), f1 ∗σ f2 : γ ∈ Γ 7−→ Xγ1∈Γ while the adjoint operation is given by: f ∗ 1 : γ ∈ Γ 7−→ σ(γ, −γ)f1(−γ). The C-algebra C∗ (Γ, σ) is then shown to be the universal C-algebra generated by a family (Wγ)γ∈Γ of unitaries such that WγWδ = σ(γ, δ)Wγδ for any γ, δ ∈ Γ [20]. We shall henceforth refer to these generating unitaries as the canonical unitaries of C∗ (Γ, σ). 4 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER One checks easily that if σ and η are two cohomologous multipliers of the discrete group Γ, then C∗ (Γ, σ) and C∗ (Γ, η) are -isomorphic [20]. Thus, by Theorem Noncommutative solenoids, defined in Definition (2.3) as twisted group algebras Notation 2.5. For any p ∈ N, p > 1 and for any α ∈ Ξp, the C-algebra the form Ψα with α ∈ Ξp. With this in mind, we introduce the following notation: pi × Zh 1 pi of (2.4), we shall henceforth restrict our attention to multipliers of Zh 1 C∗(cid:16)Zh 1 pi, Ψα(cid:17), with Ψα defined in Theorem (2.4), is denoted by A S pi × Zh 1 pi, also have a presentation as transformation group C∗-algebras, in pi × Zh 1 of Zh 1 action of Zh 1 pi on Sp defined for all q The C-crossed-product C(Sp) ⋊θα Zh 1 pk ∈ Zh 1 pi is -isomorphic to A S ((zn)n∈N) =(cid:16)e(2iπα(k+n)q)zn(cid:17)n∈N pi and for all (zn)n∈N ∈ Sp by: Proposition 2.6. [10, Proposition 3.3] Let p ∈ N, p > 1 and α ∈ Ξp. Let θα be the a manner similar to the situation with rotation C-algebras: Whichever way one decides to study them, there are longstanding methods in place to determine whether or not these C∗ algebras are simple (see for instance []). For now, we concentrate on methods from the theory of twisted group C∗-algebras. α . θα q pk . α . Theorem-Definition 2.6.1. [13, Theorem 1.5] The symmetrizer group of a mul- tiplier σ : Γ × Γ → T of a discrete group Γ is given by Sσ =(cid:8)γ ∈ Γ : ∀g ∈ Γ σ(γ, g)σ(g, γ)−1 = 1(cid:9) . The C-algebra C∗(Γ, σ) is simple if, and only if the symmetrizer group Sσ is reduced to the identity of Γ. In [10], we thus characterize when the symmetrizer group of the multipliers of Zh 1 pi × Zh 1 pi given by Theorem (2.4) is non-trivial: Theorem 2.7. [10, Theorem 2.12] Let p ∈ N, p > 1. Let α ∈ Ξp. Denote by Ψα the multiplier defined in Theorem (2.4). The following are equivalent: (1) the symmetrizer group SΨα is non-trivial, (2) the sequence α has finite range, i.e. the set {αj : j ∈ N} is finite, (3) there exists k ∈ N such that (pk − 1)α0 ∈ Z, (4) the sequence α is periodic, (5) there exists a positive integer b ∈ N such that: SΨα = bZ(cid:20) 1 p(cid:21) × Z(cid:20) 1 p(cid:21) =(cid:26)(br1, br2), (r1, r2) ∈ Z(cid:20) 1 p(cid:21) × Z(cid:20) 1 p(cid:21)(cid:27) . Theorem (2.8), when applied to noncommutative solenoids via Theorem (2.7), allows us to conclude: Theorem 2.8. [10, Theorem 3.5] Let p ∈ N, p > 1 and α ∈ Ξp. Then the following are equivalent: (1) the noncommutative solenoid A S (2) the set {αj : j ∈ N} is infinite, α is simple, NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 5 (3) for every k ∈ N, we have (pk − 1)α0 6∈ Z. In particular, if α ∈ Ξp is chosen with at least one irrational entry, then by definition of Ξp, all entries of α are irrational, and by Theorem (2.8), the noncom- mutative solenoid A S α is simple. The reader may observe that, even if α ∈ Ξp only has rational entries, the noncommutative solenoid may yet be simple -- as long as α has infinite range. We called this situation the aperiodic rational case in [10]. Example 2.9 (Aperiodic rational case). Let p = 7, and consider α ∈ Ξ7 given by α =(cid:18) 2 7 , 2 49 , 2 343 , 2 2401 , · · ·(cid:19) =(cid:18) 2 7n(cid:19)n∈N . Note that αj ∈ Q for all j ∈ N, yet Theorem (2.8) tells us that the noncommutative solenoid A S α is simple! The following is an example where the symmetrizer subgroup is non-trivial, so that the corresponding C∗-algebra is not simple. Example 2.10 (Periodic rational case). Let p = 5, and consider α ∈ Ξ5 given by α =(cid:18) 1 62 , 25 62 , 5 62 , 1 62 , · · ·(cid:19) . Theorem (2.7) shows that the symmetrizer group of the multiplier Ψα of(cid:0)Z(cid:2) 1 5(cid:3)(cid:1)2 given by Theorem (2.4) is: Sα =(cid:26)(cid:18) 62j1 5k , 62j2 5k (cid:19) ∈ Q : j1, j2 ∈ Z, k ∈ N(cid:27) . Hence the noncommutative solenoid A S α is not simple by Theorem (2.8). We conclude this section with the following result about the existence of traces on noncommutative solenoids, which follows from [7], since the Pontryagin dual α for any α ∈ Ξp via the dual Sp × Sp of Zh 1 action: pi × Zh 1 pi acts ergodically on A S Theorem 2.11. [10, Theorem 3.8] Let p ∈ N, p > 1 and α ∈ Ξp. There exists at least one tracial state on the noncommutative solenoid A S α . Moreover, this tracial state is unique if, and only if α is not periodic. Moreover, since noncommutative solenoids carry an ergodic action of the com- pact groups Sp, if one chooses any continuous length function on Sp, then one may employ the results found in [18] to equip noncommutative solenoids with quantum compact metric spaces structures and, for instance, use [19] and [9] to obtain vari- ous results on continuity for the quantum Gromov-Hausdorff distance of the family of noncommutative solenoids as the multiplier and the length functions are left to vary. In this paper, we shall focus our attention on the noncommutative topology of our noncommutative solenoids, rather than their metric properties. In [10, Theorem 3.17], we provide a full description of noncommutative solenoids as bundles of matrix algebras over the space S 2 p , while in contrast, in [10, Propo- sition 3.16], we note that for α with at least (and thus all) irrational entry, the noncommutative solenoid A S α is an inductive limit of circle algebras (i.e. AT), with real rank zero. Both these results follow from writing noncommutative solenoids as inductive limits of quantum tori, which is the starting point for our next section. 6 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER 3. Classification of the noncommutative Solenoids Noncommutative solenoids are classified by their K-theory; more precisely by their K0 groups and the range of the traces on K0. The main content of our paper [10] is the computation of the K-theory of noncommutative solenoids and its application to their classification up to -isomorphism. The starting point of this computation is the identification of noncommutative solenoids as inductive limits of sequences of noncommutative tori. A noncommuta- tive torus is a twisted group C-algebra for Zd, with d ∈ N, d > 1 [14]. In particular, for d = 2, we have the following description of noncommutative tori. Any multiplier of Z2 is cohomologous to one of the form: σθ :(cid:18)(cid:18)z1 z2(cid:19) ,(cid:18)y1 y2(cid:19)(cid:19) 7−→ exp(2iπθz1y2) for some θ ∈ [0, 1). Consequently, for a given θ ∈ [0, 1), the C-algebra C∗(cid:0)Z2, σθ(cid:1) is the universal C-algebra generated by two unitaries U, V such that: U V = e2iπθV U . We will employ the following notation throughout this paper: Notation 3.1. The noncommutative torus C∗(cid:0)Z2, σθ(cid:1), for θ ∈ [0, 1), is denoted by Aθ. Moreover, the two canonical generators of Aθ (i.e. the unitaries corresponding to (1, 0), (0, 1) ∈ Z2), are denoted by Uθ and Vθ, so that UθVθ = e2iπθVθUθ. For any θ ∈ [0, 1), the noncommutative torus Aθ is -isomorphic to the crossed- product C-algebra for the action of Z on the circle T generated by the rotation of angle 2iπθ, and thus Aθ is also known as the rotation algebra for the rotation of angle θ -- a name by which it was originally known. pi × Zh 1 Zh 1 The following question naturally arises: since Aθ is a twisted Z2 algebra, and pi can be realized as a direct limit group built from embeddings of Z2 into itself, is it possible to build our noncommutative solenoids A S α as a direct limits of rotation algebras? The answer is positive, and this observation provides much structural information regarding noncommutative solenoids. Theorem 3.2. [10, Theorem 3.7] Let p ∈ N, p > 1 and α ∈ Ξp. For all n ∈ N, let ϕn be the unique -morphism from Aα2n into Aα2n+2 given by: (Uα2n 7−→ U p Vα2n 7−→ V p α2n+2 α2n+2 Then: Aα0 ϕ0−→ Aα2 ϕ1−→ Aα4 ϕ2−→ · · · converges to the noncommutative solenoid A S is the family of canonical unitary generators of A S tion algebra Aα2n embeds in A S α . Moreover, if (Wr1,r2)(r1,r2)∈Z[ 1 p ]×Z[ 1 p ] α , then, for all n ∈ N, the rota- α via the unique extension of the map: Uα2n 7−→ W( 1 Vα2n 7−→ W(0, 1 pn ,0) pn ).  to a -morphism, given by the universal property of rotation algebras; one checks that this embeddings, indeed, commute with the maps ϕn. NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 7 Our choice of terminology for noncommutative solenoids is inspired, in part, by Theorem (3.2), and the well established terminology of noncommutative torus for rotation algebras. Moreover, as we shall now see, our study of noncommutative solenoids is firmly set within the framework of noncommutative topology. The main result from our paper [10] under survey in this section and the pre- vious one is the computation of the K-theory of noncommutative solenoid and its application to their classification. An interesting connection between our work on which in turn are classified by means of the group of p-adic integers, emerges as a consequence of our computation. We shall present this result now, starting with noncommutative solenoid and classifications of Abelian extensions of Zh 1 some reminders about the p-adic integers and Abelian extensions of Zh 1 pi by Z, pi, and refer to [10] for the involved proof leading to it. Theorem-Definition 3.2.1. Let p ∈ N, p > 1. The set: Zp =(cid:8)(Jk)k∈N : J0 = 0 and ∀k ∈ N Jk+1 ≡ Jk mod pk(cid:9) is a group for the operation defined as: (Jk)k∈N + (Kk)K∈N = ((Jk + Kk) mod pk)k∈N for any (Jk)k∈N, (Kk)k∈N ∈ Zp. This group is the group of p-adic integers. following manner: One may define the group of p-adic integer simply as the set of sequences valued in {0, . . . , p − 1} with the appropriate operation, but our choice of definition will make our exposition clearer. We note that we have a natural embedding of Z as a subgroup of Zp by sending z ∈ Z to the sequence (z mod pk)k∈N. We shall henceforth identify Z with its image in Zp when no confusion may arise. Theorem 3.3. [10] Let p ∈ N, p > 1 and let J = (Jk)k∈N ∈ Zp. Define the map pi, i.e. a map We can associate, to any p-adic integer, a Schur multiplier of Zh 1 pi × Zh 1 pi → Z which satisfies the (additive) 2-cocycle identity, in the ξj : Zh 1 ξJ : Zh 1 pi × Zh 1 pi → Z by setting, for any q1 pk2(cid:19) = ξJ(cid:18) q1 where all fractions are written in their reduced form, i.e. such that the exponent of p at the denominator is minimal (this form is unique). Then: if k2 > k1, if k1 > k2, if k1 = k2, with q − q1 − q2 q pr (Jk1 − Jr) pi: pk2 ∈ Zh 1 pk1 (Jk2 − Jk1 ) pk2 (Jk1 − Jk2 ) pk1 , q2 pr = q1 pk1 + q2 pk2 , • For any J, K ∈ Zp, the Schur multipliers ξJ and ξK are cohomologous if, • ξJ is a Schur multiplier of Zh 1 • Any Schur multiplier of Zh 1 pi [10, Lemma 3.11]. pi is cohomologous to ξJ for some J ∈ Zp [10, and only if J − K ∈ Z [10, Theorem 3.14]. Theorem 3.16]. q2 , pk1 In particular, Ext(cid:16)Zh 1 pi, Z(cid:17) is isomorphic to Zp /Z . 8 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER Schur multipliers provide us with a mean to describe and classify Abelian ex- tensions of Zh 1 pi by Zp. Our interest in Theorem (3.3) lies in the remarkable observation that the K0 groups of noncommutative solenoids are exactly given by these extensions: be the Schur multiplier of Zh 1 with underlying set Z × Zh 1 for all (z1, r1), (z2, r2) ∈ Z × Zh 1 of Zh 1 Then: Theorem 3.4. [10, Theorem 3.12] Let p ∈ N, p > 1 and let α = (αk)k∈N ∈ Ξp. For any k ∈ N, define Jk = pkαk − α0, and note that by construction, J ∈ Zp. Let ξJ pi defined in Theorem (3.3), and let QJ be the group pi and operation: (z1, r1) ⊞ (z2, r2) = (z1 + z2 + ξJ (r1, r2) , r1 + r2) 1 Furthermore, we have: α (cid:1), char- and, moreover, all tracial states of A S acterized by: pi. By construction, QJ is an Abelian extension pi by Z given by the Schur multiplier ξJ . K0(cid:0)A S α (cid:1) = QJ α lift to a single trace τ on K0(cid:0)A S τ : (1, 0) 7→ 1 and (cid:18)0, α (cid:1) = Z(cid:20) 1 K1(cid:0)A S pk(cid:19) 7→ αk. p(cid:21) × Z(cid:20) 1 p(cid:21). We observe, in particular, that given any Abelian extension of Zh 1 pi by Z, one pi of the form ξJ for some can find, by Theorem (3.3), a Schur multiplier of Zh 1 α =(cid:16) α0+Jk only Abelian extensions of Zh 1 pi by Z are given as K0 groups of noncommutative J ∈ Zp, and, up to an arbitrary choice of α0 ∈ [0, 1), one may form the sequence , and check that α ∈ Ξp; thus all possible Abelian extensions, and solenoids. With this observation, the K0 groups of noncommutative solenoids are uniquely described by a p-adic integer modulo an integer, and the information contained in the pair (K0(A S α ), τ ) of the K0 group of a noncommutative solenoid and its trace, is contained in the pair (J, α0) with J ∈ Zp /Z as defined in Theorem (3.4). pk (cid:17)k∈N Remark 3.5. For any p ∈ N, p > 1 and α ∈ Ξp, the range of the unique trace τ on K0(A S α ), as described by Theorem (3.4), is the subgroup Z ⊕ ⊕k∈NαnZ. Let γ = z + z1αn1 + . . . zkαnk be an arbitrary element of this set, where, to fix notations, we assume n1 < . . . < nk. Then, since αn+1 ≡ pαn mod 1 for any n ∈ N, we conclude that we can rewrite γ simply as z′ + yαnk , for some z′, y ∈ Z. Thus the range of our trace on K0(A S α ) is given by: τ(cid:0)K0(cid:0)A S α (cid:1)(cid:1) = {z + yαk : z, y ∈ Z, k ∈ N} . We thus have a complete characterization of the K-theory of noncommutative solenoids. This noncommutative topological invariant, in turn, contains enough information to fully classify noncommutative solenoids in term of their defining NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 9 multipliers. We refer to [10, Theorem 4.2] for the complete statement of this clas- sification; to keep our notations at a minimum, we shall state the corollary of [10, Theorem 4.2] when working with p prime: Theorem 3.6. [10, Corollary 4.3] Let p, q be two prime numbers and let α ∈ Ξp and β ∈ Ξq. Then the following are equivalent: (1) The noncommutative solenoids A S (2) p = q and a truncated subsequence of α is a truncated subsequence of β or β are -isomorphic, α and A S (1 − βk)k∈N. Theorem (3.6) is given in greater generality in [10, Theorem 4.2], where p, q are not assumed prime; the second assertion of the Theorem must however be phrased in a more convoluted manner: essentially, p and q must have the same set of prime factors, and there is an embedding of Ξp and Ξq in a larger group Ξ, whose elements are still sequences in [0, 1), such that the images of α and β for these embeddings are sub-sequences of a single element of Ξ. We conclude this section with an element of the computation of the K0 groups exists a Rieffel-Powers projection in Aα2k whose image in A S α for the embedding given by Theorem (3.2) has K0 class the element γ, whose trace is thus naturally given by Theorem (3.4). Much work is needed, however, to identify the range of pk(cid:17) ∈ K0(cid:0)A S in Theorem (3.4). Given γ = (cid:16)0, 1 K0 as the set of all Abelian extensions of Zh 1 turn, by Zp /Z , as we have shown in this section. α (cid:1), if α0 is irrational, then there pi by Z, and parametrize these, in We now turn to the question of the structure of the category of modules over noncommutative solenoids. In the next two sections, we show how to apply some constructions of equivalence bimodules to the case of noncommutative solenoids as a first step toward solving the still open problem of Morita equivalence for noncommutative solenoids. 4. Forming projective modules over noncommutative solenoids from the inside out Projective modules for rotation algebras and higher dimensional noncommuta- tive tori were studied by M. Rieffel ([16]). F. Luef has extended this work to build modules with a dense subspace of functions coming from modulation spaces (e.g., Feichtinger's algebra) with nice properties ([11], [12]). One approach to building projective modules over noncommutative solenoids is to build the projective mod- ules from the "inside out". We first make some straightforward observations in this direction. We recall that, by Notation (2.5), for any p ∈ N, p > 1, and for any α ∈ Ξp, where Ξp is defined in Theorem (2.4), the C-algebra C∗(cid:16)Zh 1 pi × Zh 1 pi, Ψα(cid:17), where the multiplier Ψα was defined in Theorem (2.4), is denoted by A S α . In this section, we will work with p a prime number. Last, we also recall that by Notation (3.1), the rotation algebra for the rotation of angle θ ∈ [0, 1) is denoted by Aθ, while its canonical unitary generators are denoted by Uθ and Vθ, so that UθVθ = e2iπθVθUθ. Theorem (3.4) describes the K0 groups of noncommutative solenoids, and, among other conclusions, state that there always exists a unique trace on the K0 of any 10 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER noncommutative solenoid, lifted from any tracial state on the C-algebra itself. With this in mind, we state: Proposition 4.1. Let p be a prime number, and fix α ∈ Ξp, with α0 6∈ Q. Let γ = z + qαN for some z, q ∈ Z and N ∈ N, with γ > 0. Then there is a left projective module over A S α whose K0 class has trace γ, or equivalently, whose K0 class is given by(cid:16)z, q pi. pk(cid:17) ∈ Z × Zh 1 Proof. By Remark (3.5), γ is the image of some class in K0(A S α ) for the trace on this group. Now, since αN +1 = pαN + j for some j ∈ Z by definition of Ξα, we may as well assume N is even. As K0(A S α ) is the inductive limit of K0(Aαk )k∈2N by Theorem (3.2), γ is the trace of an element of K0(AαN )), where AαN is identified as a subalgebra of A S α (again using Theorem (3.2). By [14], there is a projection Pγ in AαN whose K0 class has trace γ, and it is then easy to check that the left projective module P A S (cid:3) α fulfills our proposition. α over A S So, for example, with the notations of the proof of Proposition (4.1), if Pγ is a α with trace γ ∈ (0, 1), one can construct the equivalence projection in Aαn ⊂ A S bimodule A S α − A S α Pγ − Pγ A S α Pγ. From this realization, not much about the structure of Pγ A S α Pγ can be seen, although it is possible to write this C∗-algebra as a direct limit of rotation algebras. Let us now discuss this matter. Suppose we have two directed sequences of C∗-algebras: ϕ2−−−−→ · · · ϕ0−−−−→ A1 ϕ1−−−−→ A2 A0 and B0 ψ0−−−−→ B1 ψ1−−−−→ B2 ψ2−−−−→ · · · Suppose further that for each n ∈ N there is an equivalence bimodule Xn between An and Bn An − Xn − Bn, and that the (Xn)n∈N form a directed system, in the following sense: there exists a direct system of module monomorphisms X0 i0−−−−→ X1 i1−−−−→ X2 i2−−−−→ · · · satisfying, for all f, g ∈ Xn and b ∈ Bn: hin(f ), in(g)iBn+1 = ψn(hf, giBn ) and in(f · b) = in(f ) · ψn(b), with analogous but symmetric equalities holding for the Xn viewed as left-An mod- ules. Now let A be the direct limit of (An)n∈N, B be the direct limit of (Bn)n∈N and X be the direct limit of (Xn)n∈N (completed in the natural C∗-module norm). Then X is an A − B bimodule. If one further assumes that the algebra of adjointable operators on X viewed as a A − B bimodule, L(X ), can be obtained via an appro- priate limiting process from the sequence of adjointable operators {L(Xn)}∞ n=1 ( NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 11 where each Xn is a An − Bn bimodule), then in addition one has that X is a strong Morita equivalence bimodule between A and B. So suppose that γ ∈ (0, 1) is as in the statement of Proposition (4.1), for some α ∈ Ξp not equal to zero, and suppose that we know that there is a positive integer N and a projection Pγ in AαN whose K0 class has trace γ. Again, without loss of generality, we assume that N is even. Then setting An = AαN +2n , Xn = AαN +2n Pγ, and Bn = PγAαN +2n Pγ, all of the conditions in the above paragraphs hold a priori, since A S α is a direct limit of the AαN +2n, so that certainly B = Pγ A S α Pγ is a direct limit of the PγAαN +2nPγ, and X = A S α Pγ can be expressed as a direct limit of the Xn = AαN +2nPγ, again by construction, with the desired conditions on the adjointable operators satisfied by construction. It would be interesting to see how far this set-up could be extended to more general directed systems of Morita equivalence bimodules over directed systems of C∗-algebras, but we leave this project to a future endeavor. We discuss very simple examples, to show how the directed system of bimodules is constructed. Example 4.2. Fix an irrational α0 ∈ [0, 1), let p = 2, and consider α ∈ Ξ2 given by α = (α0, α1 = , α2 = , · · · , αn = αn 2n , · · · , ), α0 2 α0 4 Consider Pα0 ∈ Aα0 ⊂ Aα1 a projection of trace α0 = 2α1. The bimodule is equivalent to Rieffel's bimodule Aα0 − Aα0 · Pα0 − Pα0 Aα0 Pα0 Aα0 − Cc(R) − A 1 α0 = B0. Let β0 = 1 α0 a bimodule . Rieffel's theory, specifically Theorem 1.1 of [15], again shows there is is the same as Aα2 − Aα2 · Pα0 − Pα0 Aα2 Pα0 Aα2 − Aα2 · P4α2 − P4α2 Aα2 P4α2 which is equivalent to Rieffel's bimodule Aα2 − Cc(R × F4) − C(T × F4) ⋊τ1 Z = B1, where F4 = Z/4Z, and the action of Z on T × F4 is given by multiples of ( β2 for β2 = 1 α2 , [1]F4]), i.e. multiples of (β0, [1]F4). , i.e. multiples of ( 1 α0 4 , [1]F4), At the nth stage, using Theorem 1.1 of [15] again, we see that is the same as Aα2n − Aα2n · Pα0 − Pα0 Aα2n Pα0 Aα2n − Aα2n · P2nαn − P2nαn Aαn P2nαn which is equivalent to Aαn − Cc(R × F4n ) − C(T × F4n ) ⋊τn Z = Bn, 12 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER where the action of Z on T × F4n is given by multiples of ( β2n 1 4n , [1]F4n ), for β2n = , [1]F4n ), i.e. multiples of (β0, [1]F4n ), for F4n = , i.e. multiples of ( 1 α0 = 4n α0 α2n Z/4nZ. From calculating the embeddings, we see that for α = (α0, α0 2 , · · · , α 2n , · · · ) ∈ Ξ2, we have that AS α is strongly Morita equivalent to a direct limit B of the Bn. The structure of B is not clear in this description, although each Bn is seen to be a variant of a rotation algebra. As expected, one calculates tr(K0(AS α)) = α0 · tr(K0(B)). 5. Forming projective modules over noncommutative solenoids using p-adic fields Under certain conditions, one can construct equivalence bimodules for A S α (α ∈ Ξp,p prime) by using a construction of M. Rieffel [16]. The idea is to first embed pi as a co-compact 'lattice' in a larger group M , and the quotient group M /Γ will be exactly the solenoid Sp. We thank Jerry Kaminker and Jack Spielberg for telling us about this trick. pi × Zh 1 Γ = Zh 1 We start with a brief description of the field of p-adic numbers, with p prime. Algebraically, the field Qp is the field of fraction of the ring of p-adic integers Zp -- we introduce Zp as a group, though there is a natural multiplication on Zp turning it into a ring. A more analytic approach is to consider Qp as the completion of the field Q for the p-adic metric dp, defined by dp(r, r′) = r − r′p for any r, r′ ∈ Q, where · p is the p-adic norm defined by: r =(p−n if r 6= 0 and where r = pn a 0 if r = 0. b with a, b are both relatively prime with p, If we endow Q with the metric dp, then series of the form: ajpj ∞Xj=k will converge, for any k ∈ Z and aj ∈ {0, . . . , p − 1} for all j = k, . . .. This is the p-adic expansion of a p-adic number. One may easily check that addition and multiplication on Q are uniformly continuous for dp and thus extend uniquely to Qp to give it the structure of a field. Moreover, one may check that the group Zp of p- adic integer defined in Section 3 embeds in Qp as the group of p-adic numbers of the j=0 ajpj with aj ∈ {0, . . . , p − 1} for all j ∈ N. Now, with this embedding, one could also check that Zp is indeed a subring of Qp whose field of fractions is Qp (i.e. Qp is the smallest field containing Zp as a subring) and thus, both constructions described in this section agree. Last, the quotient of the (additive) group Qp by its subgroup Zp is the Prufer p-group Z(p∞) = {z ∈ T : ∃n ∈ N z(pn) = 1}. formP∞ 5.1. Embedding Z( 1 completion of Q and Zh 1 p ) as a lattice in a self-dual group. Since Qp is a metric pi is a subgroup of Q, we shall identify, in this section, NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 13 Zh 1 pi as a subgroup of Qp with no further mention. We now define a few group homomorphisms to construct a short exact sequence at the core of our construction. Let ω : R → Sp be the standard "winding line" defined for any t ∈ R by: ω(t) =(cid:16)e2πit, e2πi t p , e2πi t p2 , · · · , e2πi t pn , · · ·(cid:17) . in {0, . . . , p − 1}. We define the sequence ζ(γ) by setting for all j ∈ N: j=k ajpj for a (unique) family (aj)j=k,... of elements Let γ ∈ Qp and write γ =P∞ with the convention thatPk We thus may define the map ζj(γ) = e 2πi(cid:16)Pj m=k am pj−m+k (cid:17) j · · · is zero if k < j. Π :(Qp × R −→ Sp γ 7−→ Π(γ, t) = ζj (γ) · ω(t). then one checks that the following is an exact sequence: ι−−−−→ Qp × R Π−−−−→ Sp −−−−→ 1 pi −→ Qp × R 7−→ ι(r) = (r, −r), If we set r ι :(Zh 1 pi 1 −−−−→ Zh 1 1 −−−−→ Zh 1 pi × Zh 1 different terms in Zh 1 family of different embeddings of Zh 1 It follows that there is an exact sequence Indeed, we will show later that it is possible to perturb the embeddings of the pi −−−−→ [Qp × R] × [Qp × R] −−−−→ Sp × Sp −−−−→ 1. pi by elements of Qp \ {0} and R \ {0} to obtain a pi × Zh 1 pi × Zh 1 pi into [Qp × R]2. We now observe that M = Qp × R is self-dual. We shall use the following standard notation: Notation 5.1. The Pontryagin dual of a locally compact group G is denoted by bG. The dual pairing between a group and its dual is denoted by h·, ·i : G × bG → T. Let us show that M ∼= cM . To every x ∈ Qp, we can associate the character χx : q ∈ Qp 7→ e2iπi{x·q} where {x · q}p is the fractional part of the product x · q in Qp, i.e. it is the sum of the terms involving the negative powers of p in the p-adic expansion of x · q. The map x ∈ Qp 7→ χx ∈ cQp is an isomorphism of topological group. Similarly, every character of R is of the form χr : t ∈ R 7→ e2iπrt for some r ∈ R. Therefore every character of M is given by χ(x,r) : (q, t) ∈ Qp × R 7−→ χx(q)χr(t) for some (x, r) ∈ Qp × R (see [6]) for further details on characters of specific locally compact abelian groups). It is possible to check that the map (x, r) 7→ χ(x,r) is a group isomorphism between M and cM , so that M = Qp × R is indeed self-dual. 14 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER 5.2. The Heisenberg representation and the Heisenberg equivalence bi- prime number, and now let M = [Qp × R]. We have shown in the previous section that M is self-dual, since both Qp and R are self-dual. Now suppose there is an pi where p is some module of Rieffel. In this section, we write Γ = Zh 1 embedding ι : Γ → M ×cM . Let the image ι(Γ) be denoted by D. In the case we are considering, D is a discrete co-compact subgroup of M ×cM . Following the method of M. Rieffel [16], the Heisenberg multiplier η : (M × cM ) × (M × cM ) → T is pi × Zh 1 η((m, s), (n, t)) = hm, ti, (m, s), (n, t) ∈ M ×cM . (We note we use the Greek letter 'η' rather than Rieffel's 'β', because we have used 'β' elsewhere. Following Rieffel, the symmetrized version of η is denoted by the letter ρ, and is the multiplier defined by: defined by: L2(M ), defined as π, where ρ((m, s), (n, t)) = η((m, s), (n, t))η((n, t), (m, s)), (m, s), (n, t) ∈ M ×cM . M. Rieffel [16] has shown that CC (M ) can be given the structure of a left C∗(D, η) module, as follows. One first constructs an η-representation of M ×cM on π(m,s)(f )(n) = hn, si f (n + m), (m, s) ∈ M ×cM , n ∈ M . When the representation π is restricted to D, we still have a projective η-representation of D, on L2(M ), and its integrated form gives CC (M ) the structure of a left C∗(D, η) module, i.e. for Φ ∈ CC (D, η), f ∈ CC (M ), Φ((d, χ))π(d,χ)(f )(n) Φ((d, χ)) hn, χi f (n + d). π(Φ) · f (n) = X(d,χ)∈D = X(d,χ)∈D f (n)π(d,χ)(g)(n)dn = ZM There is also a CC (D, η) valued inner product defined on CC (M ) given by: Moreover, Rieffel has shown that setting hf, giCC (D,η) =ZM D⊥ = {(n, t) ∈ M ×cM : ∀(m, s) ∈ D ρ((m, s), (n, t)) = 1}, CC (M ) has the structure of a right C∗(D⊥, η) module. Here the right module structure is given for all f ∈ Cc(M ), Ω ∈ Cc(D⊥) and n ∈ M by: f (n)hn, χi g(n + d)dn. f · Ω(n) = X(c,ξ)∈D⊥ π∗ (c,ξ)(f )(n)Ω(c, ξ), and the CC (D⊥, η)-valued inner product is given by hf, giCC (D⊥,η)(c, ξ) =ZM f (n)π(c,ξ)(g)(n)dn =ZM where f, g ∈ CC (M ), Ω ∈ CC (D⊥, η), and (c, ξ) ∈ D⊥. f (n) hn, ξi g(n + c)dn, Moreover, Rieffel shows in [16, Theorem 2.12] that C∗(D, η) and C∗(D⊥, η) are strongly Morita equivalent, with the equivalence bimodule being the completion of CC (M ) in the norm defined by the above inner products. NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 15 In order to construct explicit bimodules, we first define the multiplier η more precisely, and then discuss different embeddings of Zh 1 pi × Zh 1 In the case examined here, the Heisenberg multiplier η : [Qp×R]2×[Qp×R]2 → T pi into M ×cM . is given by: Definition 5.2. The Heisenberg multiplier η : [Qp × R]2 × [Qp × R]2 → T is defined by η[((q1, r1), (q2, r2)), ((q3, r3), (q4, r4))] = e2πir1r4 e2πi{q1q4}p , where {q1q4}p is the fractional part of the product q1 · q4, i.e. the sum of the terms involving the negative powers of p in the p-adic expansion of q1q4. pi × Zh 1 The following embeddings of Zh 1 Definition 5.3. For θ ∈ R, θ 6= 0, we define ιθ : Zh 1 different embeddings of Zh 1 pi: We start by observing that for r1, r2, r3, r4 ∈ Zh 1 pi in [Qp × R]2 will prove interesting: pi → [Qp × R]2 by pi into [Qp × R]2 and their influence on the We examine the structure of the multiplier η more precisely and then discuss different equivalence bimodules they allow us to construct. pi × Zh 1 pi × Zh 1 ιθ(r1, r2) = [(r1, θ · r1), (r2, r2)]. η(ιθ(r1, r2)), ιθ(r3, r4)) = e2πi{r1r4}p e2πiθr1r4 = e2πir1r4e2πiθr1r4 = e2πi(θ+1)r1r4. (Here we used the fact that for ri, rj ∈ Z( 1 One checks that setting Dθ = ιθ(cid:16)Zh 1 exactly -isomorphic to the noncommutative solenoid A S pi(cid:17), the C∗-algebra C∗(Dθ, η) is α for p ), {rirj }p ≡ rirj modulo Z.) θ + 1 pi × Zh 1 pn , · · ·(cid:19) =(cid:18) θ + 1 pi × Zh 1 pn (cid:19)n∈N . pi as the discrete subgroup D inside θ =(cid:26)(cid:16)r1, − θ(cid:17) , (r2, −r2)i ,h(cid:16)r3, − θ(cid:17) , (r2, −r2) : r1, r2 ∈ Z(cid:20) 1 p(cid:21)(cid:27) . θ(cid:17) , (r4, −r4)i(cid:17) = e−2πi( 1 θ , η) is also a non-commutative solenoid A S r3 r1 θ +1)r1r4. β where β = α =(cid:18)θ + 1, θ + 1 p , · · · , For this particular embedding of Zh 1 M ×cM , we calculate that D⊥ r1 Moreover, η(cid:16)h(cid:16)r1, − pnθ(cid:17)n∈N . Note that for (cid:16)1 − θ+1 It is evident that C∗(D⊥ and α =(cid:18)θ + 1, θ + 1 p , · · · , θ + 1 pn , · · ·(cid:19) , β =(cid:18)1 − θ + 1 pnθ (cid:19)n∈N , 16 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER we have θ · τ(cid:0)K0(cid:0)A S β (cid:1)(cid:1) α (cid:1)(cid:1) = τ(cid:0)K0(cid:0)A S with the notations of Theorem (3.4). Thus in this case we do see the desired relationship mentioned in Section 4: the range of the trace on the K0 groups of the two C∗-algebras are related via multiplication by a positive constant. We can now generalize our construction above as follows. Definition 5.4. For any x ∈ Qp \ {0}, and any θ ∈ R \ {0}, there is an embedding defined for all r1, r2 ∈ Zh 1 ιx,θ : Z(cid:20) 1 pi by p(cid:21) × Z(cid:20) 1 p(cid:21) → [Qp × R]2 ιx,θ(r1, r2) = [(x · r1, θ · r1), (r2, r2)]. Then, we shall prove that for all α ∈ Ξp there exists x ∈ Qp \ {0} and θ ∈ R \ {0} such that, by setting Dx,θ = ιx,θ(cid:18)Z(cid:20) 1 p(cid:21) × Z(cid:20) 1 p(cid:21)(cid:19) the twisted group C-algebra C∗(D, η) is -isomorphic to A S α . As a first step, we prove: Lemma 5.5. Let p be prime, and let M = Qp × R. Let (x, θ) ∈ [Qp \ {0}]× [R\ {0}], and define ιx,θ : Zh 1 pi × Zh 1 pi → [Qp × R]2 ∼= M ×cM by: Then ιx,θ(r1, r2) = [(x · r1, θ · r1), (r2, r2)] for all r1, r2 ∈ Z(cid:20) 1 p(cid:21). Let η denote the Heisenberg cocycle defined on [M ×cM ]2 and let p(cid:21) × Z(cid:20) 1 p(cid:21)(cid:19) . θ(cid:19)(cid:21) : t1, t2 ∈ Z(cid:20) 1 p(cid:21)(cid:27) . D = ιx,θ(cid:18)Z(cid:20) 1 x,θ =(cid:26)(cid:20)(t1, −t1),(cid:18)x−1t2, − x,θ =(cid:26)[(q1, s1), (q2, s2)] : ∀r1, r2 ∈ Z(cid:20) 1 =(cid:26)[(q1, s1), (q2, s2)] : ∀r1, r2 ∈ Z(cid:20) 1 =(cid:26)[(q1, s1), (q2, s2)] : ∀r1, r2 ∈ Z(cid:20) 1 Proof. By definition, D⊥ D⊥ t2 Now if r2 = 0, and r1 = pn, for any n ∈ Z, this implies p(cid:21) ρ([ιx,θ(r1, r2)], [(q1, s1), (q2, s2)]) = 1(cid:27) p(cid:21) ρ([(x · r1, θ · r1), (r2, r2)], [(q1, s1), (q2, s2)]) = 1(cid:27) p(cid:21) e2πiθr1s2e2πi{x·r1q2}p e2πis1r2e2πi{q1r2}p = 1(cid:27) . ∀n ∈ Z e2πiθpns2 e2πi{x·pnq2}p = 1, so that if we choose s2 = − t2 Likewise, if we take r1 = 0, and r2 = pn, for any n ∈ Z, we need (q1, s1) such that pi ⊆ R, we need q2 = x−1t2. θ for some t2 ∈ Zh 1 ∀n ∈ Z e2πis1pn e2πi{q1pn}p = 1. NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 17 Again fixing q1 = t1 ∈ Zh 1 pi, this forces s1 = −t1. Thus D⊥ x,θ =(cid:26)(cid:20)(t1, −t1),(cid:18)x−1t2, − t2 θ(cid:19)(cid:21) : t1, t2 ∈ Z(cid:20) 1 p(cid:21)(cid:27) , as we desired to show. (cid:3) One thus sees that the two C∗-algebras C∗(Dx,θ, η) and C∗(D⊥ x,θ, η) are strongly Morita equivalent (but not isomorphic, in general), and also the proof of this lemma shows that C∗(D⊥ x,θ, η) is a noncommutative solenoid. We can use Lemma (5.5) to prove the following Theorem: Theorem 5.6. Let p be prime, and let α = (αi)i∈N ∈ Ξp, with α0 ∈ (0, 1). Then there exists (x, θ) ∈ [Qp \{0}]×[R\{0}] with C∗(Dx,θ, η) isomorphic to the noncom- mutative solenoid A S of Rieffel produces an equivalence bimodule between A S algebra B, and B is itself isomorphic to a noncommutative solenoid. α , where Dx,θ = ιx,θ(cid:16)Zh 1 pi(cid:17). Moreover, the method α and another unital C∗- pi × Zh 1 Proof. By definition of Ξp, for all j ∈ N there exists bj ∈ {0, . . . , p − 1} such that j=0 bjpj ∈ Zp ⊂ Qp. Let θ = α0, and now consider for this specific x and this specific θ the C∗-algebra C∗(Dx,θ, η). By Definition (5.4), ιx,θ(r1, r2) = [(x · r1, θ · r1), (r2, r2)], pαj+1 = αj + bj. We construct an element of the p-adic integers, x =P∞ for r1, r2 ∈ Zh 1 η (ιx,θ(r1, r2), ιx,θ(r3, r4)) = η ([(x · r1, θ · r1) , (r2, r2)] , [(x · r3, θ · r3) , (r4, r4)]) pi. Then = e2πiθr1r4 e2πi{xr1r4}p , r1, r2, r3, r4 ∈ Z(cid:20) 1 p(cid:21), and, setting ri = ji pki , 1 ≤ i ≤ 4, and setting θ = α0, we obtain η(cid:18)ιx,α0(cid:18) j1 pk1 , j2 pk2(cid:19) , ιx,α0(cid:18) j3 for all j1 j4 pk1 +k4 e 2πi{x j1 j4 pk1+k4 }p We now note that the relation pαj+1 = αj + bj, bj ∈ {0, 1, · · · , p − 1} allows us to prove inductively that ∀n ≥ 1 αn = j1 pk1 , , , , j4 pk3 2πiα0 j4 pk4 j3 pk3 j2 pk2 pk4(cid:19)(cid:19) = e p(cid:21). ∈ Z(cid:20) 1 α0 +Pn−1 pi × Zh 1 pk4(cid:19)(cid:19) = e2πi(α(k1 +k4)j1j4) pk4 ∈ Zh 1 j=0 bjpj pn pk1 and j4 2πi α0j1 j4 = e j4 . By Theorem (2.4), the multiplier Ψα on Zh 1 Ψα(cid:18)(cid:18) j1 pk1 , j2 pk2(cid:19) ,(cid:18) j3 pk3 , pi is defined by: A p-adic calculation now shows that for j1 Zp, we have {x j1j4 bjpj) · pk1 +k4 }p = (Pk1+k4−1 2πi{x j1 j4 j=0 pk1 +k4 e }p = e2πi(Pk1 +k4 −1 j=0 bj pj j1j4)/pk1 +k2 . pk1 +k4 e2πi(Pk1 +k4 −1 j=0 bj pj j1j4)/pk1 +k2 . pi and x =P∞ j=0 bjpj ∈ j1j4 pk1 +k2 modulo Z, so that 18 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH A. PACKER We thus obtain η(ιx,θ(r1, r2), ιx,θ(r3, r4)) = Ψα((r1, r2), (r3, r4)) for all r1, r2, r3, r4 ∈ Zh 1 pi,as desired. To prove the final statement of the Theorem, we use Lemma 5.5. We have α is isomorphic to C∗(Dx,θ, η), and the discussion prior to the statement of x,θ, η) = B. shown A S Lemma 5.5 shows that C∗(Dx,θ, η) is strongly Morita equivalent to C∗(D⊥ But the proof of Lemma 5.5 gives that D⊥ C∗(D⊥ x,θ is isomorphic to Zh 1 x,θ, η) = B is a noncommutative solenoid, as we desired to show. pi × Zh 1 pi, so that (cid:3) Remark 5.7. It remains an open question to give necessary and sufficient conditions under which two noncommutative solenoids A S β would be strongly Morita equivalent, although it is evident that a necessary that the range of the trace on K0 of one of the C∗-algebras should be a constant multiple of the range of the trace on the K0 group of the other. By changing the value of θ to be α0 + j, j ∈ Z, and adjusting the value of x ∈ Qp accordingly, one can use the method of Theorem α and A S 5.6 to construct a variety of embeddings ιx,θ of Zh 1 pi × Zh 1 pi into [Qp × R]2 that provide lattices Dx,θ such that C∗(Dx,θ, η) and A S α are ∗-isomorphic, but such that the strongly Morita equivalent solenoids C∗(D⊥ x,θ, η) vary in structure. This might lend some insight into classifying the noncommutative solenoids up to strong Morita equivalence, as might a study between the relationship between the two different methods of building equivalence modules described in Sections 4 and 5. has used the Heisenberg equivalence bimodule construction of Rieffel to construct different families of Gabor frames in modulation spaces of L2(Rn) for modulation and translation by Zn ([11], [12]). It is of interest to see how far this analogy can Remark 5.8. In the case where the lattice Z2n embeds into Rn × cRn, F. Luef pi acting on be taken when studying modulation and translation operators of Zh 1 L2(Qp × R), and we are working on this problem at present. References [1] L. Baggett and A. Kleppner, Multiplier representations of abelian groups, J. Functional Anal- ysis 14 (1973), 299-324. [2] A. Connes, C* -- alg`ebres et g´eom´etrie differentielle, C. R. de l'academie des Sciences de Paris (1980), no. series A-B, 290. [3] S. Echterhoff and J. Rosenberg, Fine structure of the Mackey machine for actions of abelian groups with constant Mackey obstruction, Pacific J. Math. 170 (1995), 17-52. [4] G. Elliott and D. Evans, Structure of the irrational rotation C ∗-algebras, Annals of Mathe- matics 138 (1993), 477 -- 501. [5] L. Fuchs, Infinite Abelian Groups, Volume I, Academic Press, New York and London, 1970. [6] E. Hewitt and K. Ross, Abstract Harmonic Analysis, Volume II, Springer-Verlag Berlin, 1970. [7] R. Hoegh-Krohn, M. B. Landstad, and E. Stormer, Compact ergodic groups of automor- phisms, Annals of Mathematics 114 (1981), 75 -- 86. [8] A. Kleppner, Multipliers on Abelian groups, Mathematishen Annalen 158 (1965), 11 -- 34. [9] F. Latr´emoli`ere, Approximation of the quantum tori by finite quantum tori for the quantum gromov-hausdorff distance, Journal of Funct. Anal. 223 (2005), 365 -- 395, math.OA/0310214. [10] F. Latr´emoli`ere and J. Packer, Noncommutative solenoids, Submitted (2011), 30 pages, ArXiv: 1110.6227. NONCOMMUTATIVE SOLENOIDS AND THEIR PROJECTIVE MODULES 19 [11] F. Luef, Projective modules over noncommutative tori and multi-window Gabor frames for modulation spaces, J. Funct. Anal. 257 (2009), 1921 -- 1946. [12] , Projections in noncommutative tori and Gabor frames, Proc. Amer. Math. Soc. 139 (2011), 571 -- 582. [13] J. Packer and I. Raeburn, On the structure of twisted group C ∗-algebras, Trans. Amer. Math. Soc. 334 (1992), no. 2, 685 -- 718. [14] M. A. Rieffel, C-algebras associated with irrational rotations, Pacific Journal of Mathematics 93 (1981), 415 -- 429. [15] [16] [17] [18] [19] , The cancellation theorem for the projective modules over irrational rotation C ∗- algebras, Proc. London Math. Soc. 47 (1983), 285 -- 302. , Projective modules over higher-dimensional non-commutative tori, Can. J. Math. XL (1988), no. 2, 257 -- 338. , Non-commutative tori -- a case study of non-commutative differentiable manifolds, Contemporary Math 105 (1990), 191 -- 211. , Metrics on states from actions of compact groups, Documenta Mathematica 3 (1998), 215 -- 229, math.OA/9807084. , Gromov-Hausdorff distance for quantum metric spaces, Mem. Amer. Math. Soc. 168 (March 2004), no. 796, math.OA/0011063. [20] G. Zeller-Meier, Produits crois´es d'une C-alg`ebre par un groupe d' Automorphismes, J. Math. pures et appl. 47 (1968), no. 2, 101 -- 239. Department of Mathematics, University of Denver, 80208 E-mail address: [email protected] Department of Mathematics, Campus Box 395, University of Colorado, Boulder, CO, 80309-0395 E-mail address: [email protected]
1511.01193	3	1511	2016-05-07T01:49:16	A short note on Cuntz splice from a viewpoint of continuous orbit equivalence of topological Markov shifts	[ "math.OA" ]	Let $A$ be an $N\times N$ irreducible matrix with entries in $\{0,1\}$. We present an easy way to find an $(N+3)\times (N+3)$ irreducible matrix $\bar{A}$ with entries in $\{0,1\}$ such that their Cuntz--Krieger algebras ${\mathcal{O}}_A$ and ${\mathcal{O}}_{\bar{A}}$ are isomorphic and $ \det(1 -A) = - \det(1-\bar{A}). $ As a consequence, we know that two Cuntz--Krieger algebras ${\mathcal{O}}_A$ and ${\mathcal{O}}_B$ are isomorphic if and only if the one-sided topological Markov shift $(X_A, \sigma_A)$ is continuously orbit equivalent to $(X_B, \sigma_B)$ or $(X_{\bar{B}}, \sigma_{\bar{B}}).$	math.OA	math	A short note on Cuntz splice from a viewpoint of continuous orbit equivalence of topological Markov shifts 6 1 0 2 y a M 7 ] Kengo Matsumoto Department of Mathematics Joetsu University of Education Joetsu, 943-8512, JAPAN Abstract Let A be an N ×N irreducible matrix with entries in {0, 1}. We present an easy way to find an (N + 3) × (N + 3) irreducible matrix ¯A with entries in {0, 1} such that their Cuntz -- Krieger algebras OA and O ¯A are isomorphic and det(1 − A) = −det(1 − ¯A). As a consequence, we know that two Cuntz -- Krieger algebras OA and OB are isomorphic if and only if the one-sided topological Markov shift (XA, σA) is continuously orbit equivalent to either (XB, σB) or (X ¯B, σ ¯B). For an N × N irreducible matrix A with entries in {0, 1}, let us denote by G(A) the abelian group ZN /(1 − At)ZN and by uA the position of the class [(1, . . . , 1)] of the vector (1, . . . , 1) in the group G(A). Throughout this short note, matrices are all assumed to be irreducible and not any permutation matrices. J. Cuntz in [3] has shown that the pair (K0(OA), [1]) of the K0-group K0(OA) of the Cuntz -- Krieger algebra OA and the class [1] of the unit in K0(OA) is isomorphic to (G(A), uA). In [12], M. Rørdam has shown that (G(A), uA) is a complete invariant of the isomorphism class of OA (see [6] for N ≤ 3). For an N × N irreducible matrix A = [A(i, j)]N i,j=1 with entries in {0, 1}, the (N + 2) × (N + 2) irreducible matrix A− defined by . A O h t a m [ 3 v 3 9 1 1 0 . 1 1 5 1 : v i X r a A− = A(N − 1, 1) A(1, 1) ... A(N, 1) 0 0   . . . A(1, N ) A(1, N − 1) 0 0 ... ... . . . A(N − 1, N − 1) A(N − 1, N ) 0 0 1 0 . . . . . . 1 1 1 1 . . . A(N, N − 1) A(N, N ) ... 0 0 ... 1 0   is called the Cuntz splice for A, which has been first introduced in [4] by J. Cuntz, related to classification problem for Cuntz -- Krieger algebras. In [4], he had used the notation A− instead of the above A−. The crucial property of the Cuntz splice is that G(A−) is isomorphic to G(A) and det(1 − A−) = −det(1 − A). The Cuntz splice 1 1 0 0 1 1 1 0 0 1 1 1 0 0 1 1     1 1 1 ] is denoted by 2−. for the matrix [ 1 1 In the proof of the above Rørdam's result [12, Theorem 6.5], J. Cuntz's theorem [12, Theorem 7.2] is used which says that O2 ∼= O2− implies OA ⊗ K ∼= OA− ⊗ K for all irreducible non-permutation matrices A. Since Rørdam has proved O2 ∼= O2− ([12, Lemma 6.4]), the result OA ⊗ K ∼= OA− ⊗ K holds for all irreducible non-permutation matrices A. By using this result, Rørdam has also obtained that the group G(A) is a complete invariant of the stable isomorphism class of OA. Let us denote by BF(A) the abelian group G(At) = ZN /(1 − A)ZN , which is called the Bowen -- Franks group for N × N matrix A ([1]). Although BF(A) is isomorphic to G(A) as a group, there is no canonical isomorphism between them. Related to classification theory of symbolic dynamical systems, J. Franks has shown that the pair (BF(A), sgn(det(1−A))) is a complete invariant of the flow equivalence class of the two-sided topological Markov shift ( ¯XA, ¯σA) by using Bown -- Franks's result [1] for the group BF(A) and Parry -- Sullivan's result [11] for the determinant det(1 − A). Combining this with the Rørdam's result for the stable isomorphism classes of the Cuntz -- Krieger algebras, OA is stably isomorphic to OB if and only if ( ¯XA, ¯σA) is flow equivalent to either ( ¯XB, ¯σB) or ( ¯XB−, ¯σB−). In [9], the author has introduced a notion of continuous orbit equivalence in one- sided topological Markov shifts to classify Cuntz -- Krieger algebras from a view point of topological dynamical system. In [10], H. Matui and the author have shown that the triple (G(A), uA, sgn(det(1 − A))) is a complete invariant of the continuous orbit equivalence class of the right one-sided topological Markov shift (XA, σA). This result is rephrased by using the above mentioned Rørdam's result for isomorphism classes of the Cuntz -- Krieger algebras such that the pair (OA, sgn(det(1 − A))) is a complete invariant of the continuous orbit equivalence class of the one-sided topological Markov shift (XA, σA). The C ∗-algebra OA− is not necessarily isomorphic to OA, whereas they are stably isomorphic, because the position uA− in G(A−) generally is different from the position uA in G(A). We note that the group G(A) determines the absolute value det(1 − A). If G(A) is infinite, Ker(1 − A) is not trivial so that det(1 − A) = 0. If G(A) is finite, it forms a finite direct sum Z/m1Z ⊕ · · · ⊕ Z/mrZ for some m1, . . . , mr ∈ N so that det(1 − A) = m1 · · · mr (cf. [4], [5], [12]). By [10, Lemma 3.7], we know that there is a matrix A′ with entries in {0, 1} such that the triple (G(A), uA, sgn(det(1 − A))) is isomorphic to (G(A′), uA′, −sgn(det(1 − A′))), which means that there exists an isomorphism Φ : G(A) → G(A′) such that Φ(uA) = uA′ and sgn(det(1 − A)) = −sgn(det(1 − A′)). Following the given proof of [10, Lemma 3.7], the construction of the matrix A′ seems to be slightly complicated and the matrix size of A′ becomes much bigger than that of A. It is not an easy task to present the matrix A′ for the given matrix A in a concrete way. In this short note, we directly present an (N + 3) × (N + 3) matrix ¯A with entries in {0, 1} such that (G(A), uA, sgn(det(1−A))) is isomorphic to (G( ¯A), u ¯A, −sgn(det(1− ¯A))). The matrix ¯A is constructed such that if A is an irreducible non-permutation matrix, so is ¯A. 2 A(1, 1) ... A(N − 1, 1) 0 A(N, 1) A(1, 1) ... A(N − 1, 1) 0 A(N, 1) 0 0     ... . . . A(1, N ) A(1, N − 1) 0 ... . . . A(N − 1, N − 1) A(N − 1, N ) 0 1 . . . . . . 0 A(N, N − 1) A(N, N ) 0 0 ...   ... ... . . . A(1, N ) A(1, N − 1) 0 0 0 ... ... . . . A(N − 1, N − 1) A(N − 1, N ) 0 0 0 1 0 0 . . . . . . 0 1 0 1 1 1 . . . . . . 0 1 1 A(N, N − 1) A(N, N ) 0 0 0 0 0 0 ...   . (1) We define A◦ = and ¯A = (A◦)− = The operation A → A◦ is nothing but an expansion defined by Parry -- Sullivan in [11], and preserves their determinant: det(1 − A) = det(1 − A◦). The following figure is a graphical expression of the matrix ¯A from A. vN vN vN +1 Figure 1: vN +2 vN +3 We provide two lemmas. The first one is seen in [1]. The second one is seen in [4] and [12] in a different form. Lemma 1 ([1, Theorem 1.3]). The map ηA : (x1, . . . , xN −1, xN , xN +1) ∈ ZN +1 → (x1, . . . , xN −1, xN + xN +1) ∈ ZN induces an isomorphism ¯ηA from G(A◦) to G(A) such that ¯ηA([(1, . . . , 1, 0)]) = uA. Lemma 2 (cf. [4, Proposition 2], [12, Proposition 7.1]). The map ξA : (x1, . . . , xN ) ∈ ZN → (x1, . . . , xN , 0, 0) ∈ ZN +2 induces an isomorphism ¯ξA from G(A) to G(A−) such that ¯ξA([(1, . . . , 1, 0)]) = uA−. 3 Proof. For y = (y1, . . . , yN ) ∈ ZN , put We then have ξA(z) = = (1 − At −) z1 ... zN z =        z1 ... zN 0 0 x1 ... xN 0 0     = (1 − At)  y1 ... yN .   . y1 ... yN 0 −yN     = (1 − At −) z1 ... zN zN +1 zN +2     −)ZN +2 so that ξA : ZN → ZN +2 induces a Hence we have ξA((1 − At)ZN ) ⊂ (1 − At homomorphism from G(A) to G(A−) denoted by ¯ξA. Suppose that [ξ(x1, . . . , xN )] = 0 in G(A−) so that for some (z1, . . . , zN +2) ∈ ZN +2. It then follows that zN +1 = 0, zN +2 = −zN so that x1 ... xN     = (1 − At)  z1 ... zN   . This implies [(x1, . . . , xN )] = 0 in G(A) and hence ¯ξA is injective. For (x1, . . . , xN , xN +1, xN +2) ∈ ZN +2, we have x1 ... xN xN +1 xN +2     = x1 ... xN −1 xN − xN +2 0 0     0 ... 0 xN +2 xN +1 xN +2       x1 ... xN −1 0 0 + = xN − xN +2   + (1 − At −) 0 ... 0 −zN +2 −zN +1   .   This implies that [(x1, . . . , xN , xN +1, xN +2)] = ¯ξA([(x1, . . . , xN −1, xN − xN −2)]) in G(A−). Therefore ¯ξA : G(A) → G(A−) is surjective and hence an isomorphism. In particular, we see that [(1, . . . , 1, 1, 1)] = ¯ξA([(1, . . . , 1, 0)]) in G(A−). We have the following theorem by the preceding two lemmas. Theorem 3. For an N × N matrix A with entries in {0, 1}, let ¯A be the (N + 3) × (N + 3) matrix with entries in {0, 1} defined in (1). Then there exists an isomorphism Φ : G(A) → G( ¯A) such that Φ(uA) = u ¯A and the matrices A, ¯A satisfy det(1 − A) = −det(1 − ¯A). If A is an irreducible non-permutation matrix, so is ¯A. 4 Proof. Define Φ : G(A) → G( ¯A) by Φ = ¯ξA◦ ◦¯η−1 Since det(1 − ¯A) = −det(1 − A◦) = −det(1 − A), we see the desired assertion. A so that Φ(uA) = ¯ξA◦([(1, . . . , 1, 0)]) = u ¯A. Let P be an N × N permutation matrix coming from a permutation of the set {1, 2, . . . , N }. Since there exists a natural isomorphism ΦP : G(A) −→ G(P AP −1) such that ΦP (uA) = uP AP −1 and det(1−A) = det(1−P AP −1), the triplet (G(A), uA, det(1−A)) does not depend on the choice of the vertex vN in the directed graph of the matrix A. We have some corollaries. Corollary 4. Let A be an irreducible non-permutation matrix with entries in {0, 1}. Then OA is isomorphic to O ¯A and det(1 − A) = −det(1 − ¯A). Let ¯1 denote the matrix 0 1 0 0 1 0 1 0 0 1 1 1 0 0 1 1     which is the matrix ¯A for the 1 × 1 matrix A = [1]. By the above theorem, we have Corollary 5. (K0(O¯1), u¯1) = (Z, 1). Hence the simple purely infinite C ∗-algebra O¯1 has the same K-theory as the C ∗- algebra O1 = C(S1) of the continuous functions on the unit circle S1 with the positions of their units, whereas (K0(O1−), u1− ) = (Z, 0) for the matrix 1− =h 1 1 0 0 1 1i by [6] (cf. [4, p. 150]). 1 1 1 The following corollary has been shown in [10]. Its proof is now easy by using [12]. Corollary 6 ([10, Lemma 3.7]). Let F be a finitely generated abelian group and u an element of F . Let s = 0 when F is infinite and s = −1 or 1 when F is finite. Then there exists an irreducible non-permutation matrix A such that (F, u, s) = (G(A), uA, sgn(det(1 − A)). Proof. By [12, Proposition 6.7 (i)], we know that there exists an irreducible non-permutation matrix A such that (F, u) = (G(A), uA). If s = sgn(det(1−A)), the matrix A is the desired one, otherwise ¯A is the desired one. Let A and B be two irreducible non-permutation matrices with entries in {0, 1}. The one-sided topological Markov shifts (XA, σA) and (XB, σB) are said to be flip continuously orbit equivalent if (XA, σA) is continuously orbit equivalent to either (XB, σB) or (X ¯B, σ ¯B). Similarly two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are said to be flip flow equivalent if ( ¯XA, ¯σA) is flow equivalent to either ( ¯XB, ¯σB), or ( ¯X ¯B, ¯σ ¯B). We thus have the following corollaries. Corollary 7. Let A, B be irreducible and not any permutation matrices with entries in {0, 1}. (i) OA is isomorphic to OB if and only if the one-sided topological Markov shifts (XA, σA) and (XB, σB) are flip continuously orbit equivalent. 5 (ii) OA is stably isomorphic to OB if and only if the two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flip flow equivalent. Let us denote by [OA] the isomorphism class of the Cuntz -- Krieger algebra OA as a C ∗-algebra. Since (G(A), uA) is isomorphic to (G( ¯A), u ¯A), we have [OA] = [O ¯A]. We regard the sign sgn(det(1 − A)) of det(1 − A) as the orientation of the class [OA]. Then we can say that the pair ([OA], sgn(det(1 − A))) is a complete invariant of the continuous orbit equivalence class of the one-sided topological Markov shift (XA, σA). In the rest of this short note, we present another square matrix A of size N + 3 i,j=1 of size N such that OA is isomorphic to O A and from a square matrix A = [A(i, j)]N det(1 − A) = −det(1 − A). Define (N + 3) × (N + 3) matrix A by setting A(N − 1, 1) A = A(1, 1) ... 0 A(N, 1) 0 0   ... . . . A(1, N ) A(1, N − 1) 0 0 0 ... ... . . . A(N − 1, N − 1) A(N − 1, N ) 0 0 0 1 0 0 . . . 0 1 0 . . . . . . 1 0 1 0 1 1 . . . A(N, N − 1) A(N, N ) 0 0 0 ... 0 0 0 ...   . (2) The difference between the previous matrix ¯A in (1) and the above matrix A is the only Its graphical expression of the matrix A from A is the ((N + 2), (N + 2))-component. following figure. vN vN vN +1 Figure 2: vN +2 vN +3 By virtue of [6], we know the following proposition. Proposition 8. The Cuntz -- Krieger algebras O ¯A and O A are isomorphic, and det(1− ¯A) = det(1 − A). Proof. Let us denote by ¯Ai the ith row vector of the matrix ¯A of size N + 3. We put Ei i the row vector of size N + 3 such that Ei = (0, . . . , 0, 1, 0, . . . , 0) where the ith component is one, and the other components are zero. Then we have ¯AN +2 = EN +1 + ¯AN +3. Since 6 the (N + 2)th row AN +2 of A is AN +2 = EN +1 + EN +3, and the other rows of A are the same as those of ¯A, the matrix A is obtained from ¯A by the primitive transfer ¯A =⇒ EN +1+ ¯AN +3→ AN +2 A in the sense of [6, Definition 3.5]. We obtain that O ¯A is isomorphic to O A by [6, Theorem 3.7], and det(1 − ¯A) = det(1 − A) by [6, Theorem 8.4]. Before ending this short note, we refer to differences among the three matrices A−, ¯A, A from a view point of dynamical system. As (G(A−), det(1 − A−)) = (G( ¯A), det(1 − ¯A)) = (G( A), det(1 − A)), there is a possibility that their two sided topological Markov shifts ( ¯XA−, ¯σA−), ( ¯X ¯A, ¯σ ¯A), ( ¯X A, ¯σ A) are topologically conjugate. We however know that they are not topologically conjugate to each other in general by the following example. Denote by pn(¯σA) the cardinal number of the n-periodic points {x ∈ ¯XA ¯σn A(x) = x} of the topological Markov shift ( ¯XA, ¯σA). The zeta function ζA(z) for ( ¯XA, ¯σA) is defined by ζA(z) = exp ∞ Xn=1 pn(¯σA) n zn! (cf.[8]). It is well-known that the formula ζA(z) = the matrices A−, ¯A, A for [ 1 1 1 1 ] respectively. It is direct to see that 1 det(1−zA) holds ([2]). Let us denote by 2−, ¯2, 2 ζ2−(z) = 1 1 − 4z + 3z2 + 2z3 − z4 , ζ¯2(z) = 1 1 − 3z + 4z3 − z4 , ζ2(z) = 1 1 − 3z + z2 + z3 + z4 . The zeta function is invariant under topological conjugacy so that ( ¯X2−, ¯σ2−), ( ¯X¯2, ¯σ¯2), ( ¯X2, ¯σ2) are not topologically conjugate to each other. This paper is a revised version of the paper entitled "Continuous orbit equivalence of topological Markov shifts and Cuntz splice" arXiv:1511.01193v2 [math.OA]. Acknowledgment. This work was supported by JSPS KAKENHI Grant Number 15K04896. References [1] R. Bowen and J. Franks, Homology for zero-dimensional nonwandering sets, Ann. Math. 106(1977), pp. 73 -- 92. [2] R. Bowen and O. E. Lanford III, Zeta functions of the shift transformation, Trans. Amer. Math. Soc. 112(1964), pp. 55 -- 66. [3] J. Cuntz, A class of C ∗-algebras and topological Markov chains II: reducible chains and the Ext- functor for C ∗-algebras, Invent. Math. 63(1980), pp. 25 -- 40. [4] J. Cuntz, The classification problem for the C ∗-algebra OA, Geometric methods in operator algebras, Pitman Research Notes in Mathematics Series 123(1986), pp. 145 -- 151. 7 [5] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56(1980), pp. 251 -- 268. [6] M. Enomoto, M. Fujii and Y. Watatani, K0-groups and classifications of Cuntz -- Krieger algebras, Math. Japon. 26(1981), pp. 443 -- 460. [7] J. Franks, Flow equivalence of subshifts of finite type, Ergodic Theory Dynam. Systems 4(1984), pp. 53 -- 66. [8] D. Lind and B. Marcus, An introduction to symbolic dynamics and coding, Cambridge University Press, Cambridge, 1995. [9] K. Matsumoto, Orbit equivalence of topological Markov shifts and Cuntz -- Krieger algebras, Pacific J. Math. 246(2010), pp. 199 -- 225. [10] K. Matsumoto and H. Matui, Continuous orbit equivalence of topological Markov shifts and Cuntz -- Krieger algebras, Kyoto J. Math. 54(2014), pp. 863 -- 878. [11] W. Parry and D. Sullivan, A topological invariant for flows on one-dimensional spaces, Topology 14(1975), pp. 297 -- 299. [12] M. Rørdam, Classification of Cuntz -- Krieger algebras, K-theory 9(1995), pp. 31 -- 58. 8
1210.4533	2	1210	2013-03-04T17:21:05	The Cuntz semigroup and stability of close C*-algebras	[ "math.OA" ]	We prove that separable C-algebras which are completely close in a natural uniform sense have isomorphic Cuntz semigroups, continuing a line of research developed by Kadison - Kastler, Christensen, and Khoshkam. This result has several applications: we are able to prove that the property of stability is preserved by close C-algebras provided that one algebra has stable rank one; close C-algebras must have affinely homeomorphic spaces of lower-semicontinuous quasitraces; strict comparison is preserved by sufficient closeness of C-algebras. We also examine C-algebras which have a positive answer to Kadison's Similarity Problem, as these algebras are completely close whenever they are close. A sample consequence is that sufficiently close C-algebras have isomorphic Cuntz semigroups when one algebra absorbs the Jiang-Su algebra tensorially.	math.OA	math	THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS FRANCESC PERERA, ANDREW TOMS, STUART WHITE, AND WILHELM WINTER Abstract. We prove that separable C ∗-algebras which are completely close in a natural uniform sense have isomorphic Cuntz semigroups, continuing a line of research developed by Kadison-Kastler, Christensen, and Khoshkam. This result has several applications: we are able to prove that the property of stability is preserved by close C ∗-algebras provided that one algebra has stable rank one; close C ∗-algebras must have affinely homeomorphic spaces of lower-semicontinuous quasitraces; strict comparison is preserved by sufficient closeness of C ∗-algebras. We also examine C ∗-algebras which have a positive answer to Kadison's Similarity Problem, as these algebras are completely close whenever they are close. A sample consequence is that sufficiently close C ∗-algebras have isomorphic Cuntz semigroups when one algebra absorbs the Jiang- Su algebra tensorially. 1. introduction In 1972 Kadison and Kastler introduced a metric d on the C ∗-subalgebras of a given C ∗-algebra by equipping the unit balls of the subalgebras with the Hausdorff metric (in norm) ([27]). They conjectured that sufficiently close C ∗-subalgebras of B(H) should be isomorphic, and this conjecture was recently established by Christensen, Sinclair, Smith and the last two named authors ([17]) when one C ∗-algebra is separable and nuclear. The one-sided version of this result -- that a sufficiently close near inclusion of a nuclear separable C ∗-algebra into another C ∗-algebra gives rise to a true inclusion -- was later proved by Hirshberg, Kirchberg, and the third named author ([23]). These results and others (see [16], [10]) have given new momentum to the perturbation theory of operator algebras. The foundational paper [27] was concerned with structural properties of close algebras, showing that the type decomposition of a von Neumann al- gebra transfers to nearby algebras. We continue this theme here asking Date: October 9, 2018. Research partially supported by EPSRC (grants No. EP/G014019/1 and No. EP/I019227/1), by the DFG (SFB 878), by NSF (DMS-0969246), by the DGI MICIIN (grant No. MTM2011-28992-C02-01), and by the Comissionat per Universitats i Recerca de la Generalitat de Catalunya. Andrew Toms is partially supported by the 2011 AMS Centennial Fellowship. 1 2 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER "Which properties or invariants of C ∗-algebras are preserved by small per- turbations?" With the proof of the Kadison-Kastler conjecture the answer for nuclear separable C ∗-algebras is, "All of them." Here we consider gen- eral separable C ∗-algebras where already, there are some results. Sufficiently close C ∗-algebras have isomorphic lattices of ideals ([34]) and algebras whose stabilizations are sufficiently close have isomorphic K-theories ([29]). This was extended to the Elliott invariant consisting of K-theory, traces, and their natural pairing, in [16]. A natural next step is to consider the Cuntz semigroup of (equivalence classes of) positive elements (in the stabilisation) of a C ∗-algebra, due both to its exceptional sensitivity in determining non- isomorphism ([42]), classification results using the semigroup ([39]) and the host of C ∗-algebraic properties that can be formulated as order-theoretic properties of the semigroup: for example there is strong evidence to suggest that the behaviour of the Cuntz semigroup characterises important algebraic regularity properties of simple separable nuclear C ∗-algebras ([31, 44, 45]). We prove that algebras whose stabilizations are sufficiently close do indeed have isomorphic Cuntz semigroups, a surprising fact given the sensitivity of a Cuntz class to perturbations of its representing positive element. This is in stark contrast with the case of Murray-von Neumann equivalence classes of projections, where classes are stable under perturbations of the represent- ing projection of size strictly less than one. The bridge between these two situations is that we can arrange for the representing positive element of a Cuntz class to be almost a projection in trace. We exploit this fact through the introduction of what we call very rapidly increasing sequences of positive contractions, increasing sequences where each element almost acts as a unit on its predecessor. The Kadison-Kastler metric d is equivalent to a complete version dcb (given by applying d to the stabilisations) if and only if Kadison's Similar- ity Problem has a positive solution [16, 9]; the latter is known to hold in considerable generality, for instance in the case of Z-stable algebras ([25]). We show how this result, and a number of other similarity results for C ∗- algebras, can be put in a common framework using Christensen's property Dk ([12]), and, building on [16], make a more careful study of automatic complete closeness and its relation to property Dk. We prove that if an algebra A has Dk for some k, then d(A ⊗ K, B ⊗ K) ≤ C(k)d(A, B), where C(k) is a constant independent of A and B; as a consequence sufficiently close C ∗-algebras have isomorphic Cuntz semigroups provided one algebra is Z-stable. Stability is perhaps the most basic property one could study in pertur- bation theory, yet proving its permanence under small perturbations has seen very little progress. We take a significant step here by proving that stability is indeed preserved provided that one of the algebras considered has stable rank one. The proof is an application of our permanence re- sult for the Cuntz semigroup. Another application is our proof that stably THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 3 close C ∗-algebras have affinely homeomorphic spaces of lower semicontin- uous 2-quasitraces. This extends and improves an earlier results from [16] showing that the affine isomorphism between the trace spaces of stably close C ∗-algebras obtained in [16] is weak∗-weak∗ continuous. The paper is organized as follows: Section 2 contains the preliminaries on the Cuntz semigroup and the Kadison-Kastler metric; Section 3 establishes the permanence of the Cuntz semigroup under complete closeness; Section 4 discusses property Dk and proves our permanence result for stability; Section 5 proves permanence for quasitraces. 2. Preliminaries Throughout the paper we write A+ for the positive elements of a C ∗- 1 for the positive contractions in algebra A, A1 for the unit ball of A and A+ A. In the next two subsections we review the definition and basic properties of the Cuntz semigroup. A complete account can be found in the survey [4]. Sn Mn(A)+. 2.1. The Cuntz semigroup. Let A be a C ∗-algebra. Let us consider on (A ⊗ K)+ the relation a - b if vnbv∗ n → a for some sequence (vn) in A ⊗ K. Let us write a ∼ b if a - b and b - a. In this case we say that a is Cuntz equivalent to b. Let Cu(A) denote the set (A ⊗ K)+/ ∼ of Cuntz equivalence classes. We use hai to denote the class of a in Cu(A). It is clear that hai ≤ hbi ⇔ a - b defines an order on Cu(A). We also endow Cu(A) with an addition operation by setting hai + hbi := ha′ + b′i, where a′ and b′ are orthogonal and Cuntz equivalent to a and b respectively (the choice of a′ and b′ does not affect the Cuntz class of their sum). The semigroup W (A) is then the subsemigroup of Cu(A) of Cuntz classes with a representative in Alternatively, Cu(A) can be defined to consist of equivalence classes of countably generated Hilbert modules over A [18]. The equivalence relation boils down to isomorphism in the case that A has stable rank one, but is rather more complicated in general and as we do not require the precise definition of this relation in the sequel, we omit it. We note, however, that the identification of these two approaches to Cu(A) is achieved by associating the element hai to the class of the Hilbert module aℓ2(A). 2.2. The category Cu. The semigroup Cu(A) is an object in a category of ordered Abelian monoids denoted by Cu introduced in [18] with additional properties. Before stating them, we require the notion of order-theoretic compact containment. Let T be a pre-ordered set with x, y ∈ T . We say that x is compactly contained in y -- denoted by x ≪ y -- if for any increasing sequence (yn) in T with supremum y, we have x ≤ yn0 for some n0 ∈ N. An object S of Cu enjoys the following properties (see [18, 4]), which we use repeatedly in the sequel. In particular the existence of suprema in property 4 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER P3 is a crucial in our construction of a map between the Cuntz semigroups of stably close C ∗-algebras. P1 S contains a zero element; P2 the order on S is compatible with addition: x1 + x2 ≤ y1 + y2 whenever xi ≤ yi, i ∈ {1, 2}; P3 every countable upward directed set in S has a supremum; P4 for each x ∈ S, the set x≪ = {y ∈ S y ≪ x} is upward directed with respect to both ≤ and ≪, and contains a sequence (xn) such that xn ≪ xn+1 for every n ∈ N and supn xn = x; P5 the operation of passing to the supremum of a countable upward directed set and the relation ≪ are compatible with addition: if S1 and S2 are countable upward directed sets in S, then S1 + S2 is upward directed and sup(S1 + S2) = sup S1 + sup S2, and if xi ≪ yi for i ∈ {1, 2}, then x1 + x2 ≪ y1 + y2 . We say that a sequence (xn) in S ∈ Cu is rapidly increasing if xn ≪ xn+1 for all n. We take the scale Σ(Cu(A)) to be the subset of Cu(A) obtained as supremums of increasing sequences from A+. For objects S and T from Cu, the map φ : S → T is a morphism in the category Cu if M1 φ is order preserving; M2 φ is additive and maps 0 to 0; M3 φ preserves the suprema of increasing sequences; M4 φ preserves the relation ≪. 2.3. The Kadison-Kastler metric. Let us recall the definition of the met- ric on the collection of all C ∗-subalgebras of a C ∗-algebra introduced in [27]. Definition 2.1. Let A, B be C ∗-subalgebras of a C ∗-algebra C. Define a metric d on all such pairs as follows: d(A, B) < γ if and only if for each x in the unit ball of A or B, there is y in the unit ball of the other algebra such that kx − yk < γ. In this definition, we typically take C = B(H) for a Hilbert space H. The complete, or stabilised version, of the Kadison-Kastler metric is defined by dcb(A, B) = d(A ⊗ K, B ⊗ K) inside C ⊗ K (here K is the compact operators on ℓ2(N)); the notion dcb is used for this metric as dcb(A, B) ≤ γ is equivalent to the condition that d(Mn(A), Mn(B)) ≤ γ for every n. We repeatedly use the standard fact that if d(A, B) < γ, then given a positive contraction a ∈ A+ 1 , there exists a positive contraction b ∈ B+ 1 with ka − bk < 2α. One way of seeing this is to use the hypothesis d(A, B) < γ to approximate a1/2 by some c ∈ B1 with ka1/2 − ck < γ. Then take b = cc∗ so that ka − bk ≤ ka1/2(a1/2 − c)k + k(a1/2 − c∗)ck < 2γ. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 5 There is also a one-sided version of closeness introduced by Christensen in [13], which is referred to as a γ-near inclusion: Definition 2.2. Let A, B be C ∗-subalgebras of a C ∗-algebra C and let γ > 0. Write A ⊆γ B if for every x in the unit ball of B, there is y ∈ B such that kx− yk ≤ γ (note that y need not be in the unit ball of B). Write A ⊂γ B if there exists γ′ < γ with A ⊆γ ′ B. As with the Kadison-Kastler metric, we also use complete, or stabilised, near inclusions: write A ⊆cb,γ B when A ⊗ Mn ⊆γ B ⊗ Mn for all n, and A ⊂cb,γ B when there exists γ′ < γ with A ⊆cb,γ B. 3. Very rapidly increasing sequences and the Cuntz semigroup We start by noting that, for close C ∗-algebras of real rank zero, an isomor- phism between their Cuntz semigroups can be deduced from existing results in the literature. For a C ∗-algebra A, let V (A) be the Murray and von Neumann semigroup of equivalence classes of projections in S∞ n=1 A ⊗ Mn and write Σ(V (A)) = {[p] ∈ V (A) p = p2 = p∗ ∈ A}. This is a local semigroup in the sense that if p, q, p′ and q′ are projections in A with p′q′ = 0 and p ∼ p′, q ∼ q′, then [p] + [q] = [p′ + q′] ∈ Σ(V (A)). Recall that, if A has real rank zero, then the work of Zhang [46] shows that V (A) has the Riesz refinement property. By definition, this means that when- ever x1, . . . , xn, y1, . . . , ym ∈ V (A) satisfy Pi xi = Pj yj, then there exist zi,j ∈ V (A) with Pj zi,j = xi and Pi zi,j = yj for each i, j. The case m = n = 2 of this can be found as [3, Lemma 2.3], and the same proof works in general. The Cuntz semigroup of a C ∗-algebra of real rank zero is completely determined by its semigroup of projections (see [33] when A additionally has stable rank one and [1] for the general case). We briefly recall how this is done. An interval in V (A) is a non-empty, order hereditary and upward directed subset I of V (A), which is said to be countably generated provided there is an increasing sequence (xn) in V (A) such that I = {x ∈ V (A) x ≤ xn for some n}. The set of countably generated intervals is denoted by Λσ(V (A)), and it has a natural semigroup structure. Namely, if I and J have generating sequences (xn) and (yn) respectively, then I + J is the interval generated by (xn + yn). Given a positive element a in A ⊗ K in a σ-unital C ∗-algebra of real rank zero A, put I(a) = {[p] ∈ V (A) p - a}. The correspondence [a] 7→ I(a) defines an ordered semigroup isomorphism Cu(A) ∼= Λσ(V (A)). Theorem 3.1. Let A and B be σ-unital C ∗-subalgebras of a C ∗-algebra C, with d(A, B) < 1/8. If A has real rank zero, then B also has real rank zero and Cu(A) ∼= Cu(B). Proof. That B has real rank zero follows from [16, Theorem 6.3]. We know from [35, Theorem 2.6] that there is an isomorphism of local semigroups Φ1 : Σ(V (A)) → Σ(V (B)) (with inverse, say, Ψ1). This is defined as Φ1[p] = 6 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER [q], where q is a projection in B such that kp− qk < 1/8. Given p ∈ Mn(A), by [46, Theorem 3.2] we can find projections {pi}i=1,...,n in A such that [p] = Pi[pi]. Now extend Φ1 to Φ : V (A) → V (B) by Φ([p]) = Pi Φ1([pi]). Let us check that Φ is well defined. If [p] = Pi[pi] = Pj[qj] for projections pi and qj in A, then use refinement to find elements aij ∈ V (A) such that [pi] = Pj aij and [qj] = Pi aij. We may also clearly choose projections zij, z′ ij], and such that zij ⊥ zik if j 6= k, and ij ⊥ z′ z′ ij ∈ A such that aij = [zij] = [z′ lj if i 6= l. Then: X Φ1([pi]) = X = X X X Φ1([zij]) Φ1([z′ j i i j ij]) = X j X i Φ1([z′ ij]) = X j Φ1([qj]) . It is clear that Φ is additive and that ΦΣ(V (A)) = Φ1. Using Ψ1, we construct an additive map Ψ : V (B) → V (A), with ΨΣ(V (B)) = Ψ1. Since Ψ1 ◦ Φ1 = idΣ(V (A)), it follows that Ψ ◦ Φ = idV (A). Similarly Φ ◦ Ψ = idV (B). Now, since Cu(A) ∼= Λσ(V (A)) and Cu(B) ∼= Λσ(V (B)), it follows that Cu(A) is isomorphic to Cu(B). (cid:3) We turn now to very rapidly increasing sequences. These provide the key tool we use to transfer information between close algebras at the level of the Cuntz semigroup. n=1 in A+ Definition 3.2. Let A be a C ∗-algebra. We say that a rapidly increasing sequence (an)∞ 1 is very rapidly increasing if given ε > 0 and n ∈ N, there exists m0 ∈ N such that for m ≥ m0, there exists v ∈ A1 with k(vamv∗)an − ank < ε. Say that a very rapidly increasing sequence (an)∞ in (A ⊗ K)1 + represents x ∈ Cu(A) if supnhani = x. n=1 (3.1) The following two functions are used in the sequel to manipulate very rapidly increasing sequences. Given a ∈ A+ and ε > 0, write (a − ε)+ for hε(a), where hε is the continuous function hε(t) = max(0, t − ε). For 0 ≤ β < γ, let gβ,γ be the piecewise linear function on R given by 0, t−β γ−β , β < t < γ; 1,   With this notation, the standard example of a very rapidly increasing se- n=1 for a ∈ A+ quence is given by (g2−(n+1),2−n(a))∞ 1 . This sequence represents hai. In this way every element of the Cuntz semigroup of A is represented by a very rapidly increasing sequence from (A ⊗ K)+ 1 . In the next few lemmas we develop properties of very rapidly increasing sequences, starting with a technical observation. t ≤ β; t ≥ γ. gβ,γ(t) = THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 7 Lemma 3.3. Let A be a C ∗-algebra and let a, b ∈ A+ 1 and v ∈ A1 satisfy kv∗bva − ak ≤ δ for some δ > 0. Suppose that 0 < β < 1 and γ ≥ 0 satisfy γ + δβ−1 < 1, then h(a − β)+i ≤ h(b − γ)+i in Cu(A). Proof. Let p ∈ A∗∗ denote the spectral projection of a for the interval [β, 1]. When p = 0, then (a − β)+ = 0 and the result is trivial, so we may assume that p 6= 0. Then ap is invertible in pA∗∗p with inverse x satisfying kxk ≤ β−1. Compressing (v∗bva− a) by p and multiplying by x, we have kpv∗bvp− pk ≤ δβ−1. Thus kpv∗(b − γ)+vp − pk ≤ k(b − γ)+ − bk + kpv∗bvp − pk ≤ γ + δβ−1, and so pv∗(b − γ)+vp ≥ (cid:0)1 − (γ + δβ−1)(cid:1) p. As p acts as a unit on (a − β)+, we have (a − β)+ = (a − β)1/2 + + p(a − β)1/2 ≤ (cid:0)1 − (γ + δβ−1)(cid:1)−1 = (cid:0)1 − (γ + δβ−1)(cid:1)−1 Thus (a − β)+ - (b − γ)+. (a − β)1/2 (a − β)1/2 + pv∗(b − γ)+vp(a − β)1/2 + v∗(b − γ)+v(a − β)1/2 + . + (cid:3) The next lemma encapsulates the fact that the element of the Cuntz semigroup represented by a very rapidly increasing sequence (an)∞ n=1 of con- tractions depends only on the behaviour of parts of the an with spectrum near 1. n=1 be a very rapidly increasing sequence in A+ Lemma 3.4. Let (an)∞ 1 . Then for each λ < 1, the sequence (h(an − λ)+i)∞ n=1 has the property that for each n ∈ N, there is m0 ∈ N such that for m ≥ m0, we have h(an − λ)+i ≪ h(am − λ)+i. Furthermore sup (3.2) n h(an − λ)+i = sup n hani. Proof. Fix n ∈ N and 0 < ε < λ and take 0 < δ small enough that λ+ε−1δ < 1. As (an)∞ n=1 is very rapidly increasing, there exists m0 such that for m ≥ m0, there exists v ∈ A1 with k(v∗amv)an − ank < δ. Lemma 3.3 gives h(an − ε)+i ≤ h(am − λ)+i, so that h(an− λ)+i ≪ ham− λ)+i as ε < λ. This shows that (h(ar − λ)+i)∞ is upward directed and that r=1 h(an − δ)+i ≤ sup r h(ar − λ)+i, for all n and all ε > 0, from which (3.2) follows. (cid:3) We can modify elements sufficiently far down a very rapidly increasing sequences with contractions so that they almost act as units for positive contractions dominated in the Cuntz semigroup by the sequence. 8 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Lemma 3.5. Let A be a C ∗-algebra. (1) Suppose that a, b ∈ A+ (2) Let (an)∞ 1 satisfy a - b. Then for all ε > 0, there exists n=1 be a very rapidly increasing sequence in A+ v ∈ A with kv∗bva − ak ≤ ε and kv∗bvk ≤ 1. 1 and suppose a ∈ A+ 1 satisfies hai ≪ suphani. Then, for every ε > 0, there exists m0 ∈ N such that for m ≥ m0, there exists v ∈ A1 with k(v∗amv)a − ak < ε. Proof. (1). Fix ε > 0 and find r > 0 so that ka1+r − ak ≤ ε/2. Now ar - b, so there exists w ∈ A with kar − w∗bwk ≤ ε/4. Thus kw∗bwk ≤ 1 + ε/4, and so, writing v = (1+ε/4)−1/2w, we have kv∗bvk ≤ 1 and kw∗bw−v∗bvk ≤ ε/4. As such kar − v∗bvk ≤ ε/2 and so kv∗bva − ak ≤ kv∗bv − arkkak + ka1+r − ak ≤ ε/2 + ε/2 = ε, as claimed. (2). As hai ≪ suphani, there exists some m1 ∈ N with a - am1 ∼ a2 m1. Fix ε > 0 and by part (1), find w ∈ A with k(w∗a2 m1w)a − ak < ε/2 and kw∗a2 n=1 is very rapidly increasing, find some m0 > m1 such that for m ≥ m0 there exists t ∈ A1 with k(t∗amt)am1 − am1k ≤ ε′. Given such m and t, we have m1wk ≤ 1. Now set ε′ = ε/(2kwk) and, as (an)∞ k(w∗am1t∗amtam1w)a − ak ≤ kw∗am1kk(t∗amt)am1 − am1kkwkkak + k(w∗a2 ≤ kwkε′ + ε/2 = ε, m1w)a − ak as kw∗am1k ≤ 1. As such we can take v = tam1w ∈ A1. (cid:3) It follows immediately from part (2) above, that two very rapidly increas- ing sequences representing the same element of the Cuntz semigroup can be intertwined to a single very rapidly increasing sequence. n)∞ n=1, (a′ Proposition 3.6. Let (an)∞ n=1 be very rapidly increasing sequences in a C ∗-algebra A representing the same element x ∈ Cu(A). Then these sequences can be intertwined after telescoping to form a very rapidly increas- ing sequence which also represents x, i.e. there exists m1 < m2 < ··· and n1 < n2 < ··· such that (am1 , a′ n2,··· ) is a very rapidly increasing sequence. n1, am2, a′ Given a rapidly increasing sequence in A+ 1 , we can use the functions gβ,γ from (3.1) to push the spectrum of the elements of the sequence out to 1 and extract a very rapidly rapidly increasing sequence representing the same element of the Cuntz semigroup. Lemma 3.7. Let A be a C ∗-algebra and (an)∞ sequence in A+ sequence (g2−(mn +1),2−mn (an))∞ 1 . Then there exists a sequence (mn)∞ n=1 is very rapidly increasing and n=1 be a rapidly increasing n=1 in N such that the n hg2−(mn +1),2−mn (an)i = sup sup n hani. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 9 In particular, every element of the scale Σ(Cu(A)) can be expressed as a very rapidly increasing sequence of elements from A+ 1 . Proof. We will construct the mn so that an−1 - g2−(mn +1),2−mn (an) and for each 1 ≤ r < n, there exists v ∈ A1 with k(v∗g2−(mn +1),2−mn (an)v)g2−(mr +1),2−mr (ar) − g2−(mr +1),2−mr (ar)k < 2−n. Fix n ∈ N and suppose m1,··· , mn−1 have been constructed with these properties. As (g2−(m+1),2−m(an))∞ m=1 is a very rapidly increasing sequence representing hani and han−1i ≪ hani, there exists fmn such that han−1i ≪ h(g2−(m+1),2−m(an))i for m ≥ fmn. Further, for 1 ≤ r < n, hg2−(mr +1),2−mr (ar)i ≪ hari ≪ sup m h(g2−(m+1),2−m(an)i and so the required mn can be found using Part (2) of Lemma 3.5. The resulting sequence (g2−(mn +1),2−mn (an))∞ n=1 is very rapidly increasing by construction. As an−1 - g2−(mn +1),2−mn (an) - an for all n, we have supnhg2−(mn +1),2−mn (an)i = supnhani. (cid:3) We now consider the situation where we have two close C ∗-algebras acting on the same Hilbert space. The following lemma ensures that we can produce a well defined map between the Cuntz semigroups. Lemma 3.8. Let A, B be C ∗-algebras acting on the same Hilbert space and suppose that a ∈ A+ 1 satisfy ka − bk < 2α for some α < 1/27. Suppose that (an)∞ 1 with hai ≪ suphani. Then, there exists n0 ∈ N with the property that for n ≥ n0 and bn ∈ B+ 1 and b ∈ B+ n=1 is a very rapidly increasing sequence in A+ 1 with kbn − ank < 2α, we have h(b − 18α)+i ≪ h(bn − γ)+i ≪ h(bn − 18α)+i in Cu(B), for all γ with 18α < γ < 2/3. Proof. Fix γ with 2/3 > γ > 18α (which is possible as α < 1/27). By taking ε = 2α−ka− bk in Lemma 3.5 (2), there exists n0 ∈ N such that for n ≥ n0, there exists v ∈ A1 with k(v∗anv)a − ak < 2α − ka − bk. Fix such an n ≥ n0 and v ∈ A1, and take bn ∈ B+ choose some w ∈ B1 with kw − vk < α. We have 1 with kan − bnk < 2α and kw∗bnw − v∗anvk ≤ 2kw − vk + kbn − ank < 4α so that k(w∗bnw)b − bk ≤ k((w∗bnw) − 1)(b − a)k + k(w∗bnw − v∗anv)ak + k(v∗anv)a − ak ≤ kb − ak + 4α + k(v∗anv)a − ak ≤ 6α. 10 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Taking δ = 6α, β = 18α and 2/3 > γ′ > γ > 18α so that γ′ + δβ−1 < 1, Lemma 3.3 gives h(b − 18α)+i ≤ h(bn − γ′)+i ≪ h(bn − γ)+i ≪ h(bn − 18α)+i, as claimed. (cid:3) Proposition 3.9. Let A, B be C ∗-algebras acting on the same Hilbert space with the property that there exists α < 1/27 such that for each a ∈ A1 there exists b ∈ B1 with ka−bk < α. Then there is a well defined, order preserving map Φ : Σ(Cu(A)) → Σ(Cu(B)) given by Φ(suphani) = suph(bn − 18α)+i, n=1 is a very rapidly increasing sequence in A+ 1 and bn ∈ B+ whenever (an)∞ have kan − bnk < 2α for all n ∈ N. Moreover, if d(A, B) < α for α < 1/42, then Φ is a bijection with inverse Ψ : Σ(Cu(B)) → Σ(Cu(A)) obtained from interchanging the roles of A and B in the definition of Φ. 1 n=1 n=1 and (b′ n)∞ Proof. Suppose first that α < 1/27. To see that Φ is well defined, we apply Lemma 3.8 repeatedly. Firstly, given a very rapidly increasing sequence n=1 in A+ (an)∞ 1 representing an element x ∈ Σ(Cu(A)) and a sequence (bn)∞ in B+ 1 with kan − bnk < 2α for all n, Lemma 3.8 shows that the sequence (h(bn − 18α)+i)∞ n=1 is upward directed. Indeed, for each m, take a = am and b = bm in Lemma 3.8 so that h(bm − 18α)+i ≪ h(bn − 18α)+i for all sufficiently large n. As such supnh(bn − 18α)+i exists in Σ(Cu(B)). Secondly, this supremum does not depend on the choice of (bn)∞ n=1. Con- sider two sequences (bn)∞ n=1 satisfying kbn − ank < 2α and kb′ n − ank < 2α for all n. For each n, Lemma 3.8 shows that there exists m0 such that for m ≥ m0, we have h(bn − 18α)+i ≪ h(b′ m − 18α)+i, Thus supnh(bn − 18α)+i = supnh(b′ Thirdly, given two very rapidly increasing sequences (a′ n)∞ n=1 in A+ n=1 and (bn)∞ n)∞ n=1 in B+ nk,kbn − ank < 2α for all n, Lemma 3.8 gives supnh(b′ 1 with kb′ n − n=1 and (an)∞ 18α)+i ≤ supnh(bn − 18α)+i. n=1 represent the same element of Σ(Cu(A)), this shows that the map Φ given in the proposition is well defined. In general, this third observation also shows that Φ is order preserving. ni ≤ supnhani, and sequences (b′ n − 18α)+i ≪ h(bm − 18α)+i. and h(b′ n − 18α)+i. 1 with supnha′ In particular, when (a′ n=1 and (an)∞ n − a′ n)∞ Now suppose that d(A, B) < α < 1/42 and let Ψ : Σ(Cu(B)) → Σ(Cu(A)) be the order preserving map obtained by interchanging the roles of A and B above. Take x ∈ Σ(Cu(A)) and fix a very rapidly increasing sequence (an)∞ 1 with kan−bnk < 2α for all n. For each n, Lemma 3.8 gives m > n with 1 representing x. Fix a sequence (bn)∞ n=1 in B+ n=1 in A+ h(bn − 18α)+i ≪ h(bm − γ)+i ≪ h(bm − 18α)+i, for any γ with 18α < γ < 2/3. Passing to a subsequence if necessary, we can assume this holds for m = n + 1 and hence ((bn − 18α)+)∞ n=1 is a rapidly THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 11 n=1 in B+ ni = sup n=1 in A+ increasing sequence. By Lemma 3.7, there exists a sequence (mn)∞ so that, defining b′ increasing sequence (b′ n=1 in N n = g2−(mn+1),2−mn ((bn − 18α)+), we have a very rapidly 1 with n h(bn − 18α)+i = Φ(x). 1 with kcn − b′ n)∞ n hb′ sup Choose a sequence (cn)∞ nk < 2α for each n so that the definition of Ψ gives Ψ(Φ(x)) = suph(cn − 18α)+i. We now show that x ≤ Ψ(Φ(x)) ≤ x. Fix 0 < β < 1 with α(18 + 24β−1) < 1. This choice can be made as α < 1/42. Fix n ∈ N. As h(bn − 18α)+i ≪ suprhb′ ri, Lemma 3.5 (2) provides m0 ∈ N such that for m ≥ m0, there exists w ∈ B1 with (3.3) Take v ∈ A1 with kv − wk < α. Then (3.4) Combining the estimates (3.3), (3.4) and noting that kw∗b′ w is a contraction, gives mw)(bn − 18α)+ − (bn − 18α)+k < 2α − kan − bnk. mk + 2kv − wk < 4α. k(v∗cmv − w∗b′ mw)k ≤ kcm − b′ k(w∗b′ mw − 1k ≤ 1 as k(v∗cmv)an − ank ≤ k((v∗cmv) − 1)(an − bn)k mw)bnk + k(v∗cmv − w∗b′ + k(w∗b′ + k(w∗b′ < kan − bnk + 4α + 18α + (2α − kan − bnk) = 24α. mw − 1)(bn − (bn − 18α)+)k mw)(bn − 18α)+ − (bn − 18α)+k Taking γ = 18α, δ = 24α, Lemma 3.3 gives h(an − β)+i ≤ h(cm − 18α)+i ≤ Ψ(Φ(x)). k playing the role of a, (b′ As n was arbitrary, supnh(an − β)+i ≤ Ψ(Φ(x)). As β < 1, Lemma 3.4 gives supnh(an − β)+i = supnhani = x so that x ≤ Ψ(Φ(x)). For the reverse inequality, fix k ∈ N and apply Lemma 3.8 (with the roles of A and B reversed, b′ n=1 the role of (an)) and γ = 1/2, so 18α < γ < 2/3) to find some n ∈ N such that h(ck − 18α)+i ≤ h(cn − 1/2)+i. Now, just as in the proof of Lemma 3.8, there is z ∈ B1 with k(z∗bn+1z)bn − bnk ≤ 6α. Let p ∈ B∗∗ be the spectral projection of bn for [18α, 1], so that, just as in the proof of Lemma 3.3, kz∗bn+1zp − pk ≤ 1/3. Fix y ∈ A1 with ky − zk ≤ α. Since p is a unit for (bn − 18α)+, it is a unit for b′ n = g2−(mn+1),2−mn ((bn − 18α)+), giving the estimate n)∞ ky∗an+1ycn − cnk ≤ ky∗an+1ycn − z∗bn+1zcnk + k(z∗bn+1z − 1)(cn − b′ n)k + k(z∗bn+1z)b′ ≤ 4α + 2α + 1/3 = 6α + 1/3. n − b′ nk 12 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Take δ = 6α + 1/3, β = 1/2 and γ = 0, so that γ + β−1δ = 2/3 + 12α < 1. Thus Lemma 3.3, gives and hence h(cn − 1/2)+i ≤ han+1i, h(ck − 18α)+i ≤ han+1i ≤ x. (cid:3) n=1 is a sequence in (B ⊗ K)+ n=1 is a very rapidly increasing sequence in (A ⊗ K)+ 1 with kan − bnk < 2α for all n ∈ N. Taking the supremum over k gives Ψ(Φ(x)) ≤ x. Theorem 3.10. Let A and B be C ∗-algebras acting on the same Hilbert space with dcb(A, B) < α < 1/42. Then (Cu(A), Σ(Cu(A))) is isomor- phic to (Cu(B), Σ(Cu(B))). Moreover, an order preserving isomorphism Φ : Cu(A) → Cu(B) can be defined by Φ(suphani) = suph(bn − 18α)+i, whenever (an)∞ 1 and (bn)∞ Proof. We have d(A⊗ K, B⊗ K) < α < 1/42. By definition Σ(Cu(A⊗ K)) = Cu(A) and Σ(Cu(A ⊗ K)) = Cu(B). By applying Proposition 3.9 to A ⊗ K and B⊗K, we obtain mutually inverse order preserving bijections Φ : Cu(A⊗ K) → Cu(B ⊗ K) and Ψ : Cu(B ⊗ K) → Cu(A⊗ K), given by Φ(suphani) = suph(bn − 18α)+i, whenever (an)∞ n=1 is a very rapidly increasing sequence in (A⊗ K)+ 1 with kan − bnk < 2α for all n=1 in A+ n ∈ N. Given a very rapidly increasing sequence (an)∞ 1 representing an element x ∈ Σ(Cu(A)), we can find a sequence (bn)∞ 1 with kan − bnk < 2α, so that Φ(x) = suph(bn − 18α)+i ∈ Σ(Cu(B)). Since Φ and Φ−1 are order preserving bijections, they also preserve the relation ≪ of compact containment and suprema of countable upward directed sets, as these notions are determined by the order relation ≤. Further, taking an = bn = 0 for all n, shows that Φ(0Cu(A)) = 0Cu(B). Finally, note that Φ preserves addition: given very rapidly increasing sequences (an)∞ n=1 and n=1 in (A ⊗ K)+ 1 representing x and y in Cu(A), the sequence (an ⊕ a′ n)∞ (a′ n) is very rapidly increasing in M2(A ⊗ K) ∼= A ⊗ K. If (bn)∞ n=1 have kan − bnk,ka′ n − b′ nk < 2α for all n, then n=1 is a sequence in (B ⊗ K)+ 1 and (bn)∞ n=1 in B+ n=1, (b′ n)∞ k(an ⊕ a′ n) − (bn ⊕ b′ n)k < 2α, and has ((bn ⊕ b′ n) − 18α)+ = (bn − 18α)+ ⊕ (b′ n − 18α)+. In this way we see that Φ(x + y) = Φ(x) + Φ(y). (cid:3) In particular properties of a C ∗-algebra which are determined by its Cuntz semigroup transfer to completely close C ∗-algebras. One of the most notable of these properties is that of strict comparison. Corollary 3.11. Let A and B be C ∗-algebras acting on the same Hilbert space with dcb(A, B) < 1/42 and suppose that A has strict comparison. Then so too does B. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 13 4. Z-stability and automatic complete closeness Given a C ∗-algebra A ⊂ B(H), [9] shows that the metrics d(A,·) and dcb(A,·) are equivalent if and only if A has a positive answer to Kadison's similarity problem from [26]. The most useful reformulation of the similar- ity property for working with close C ∗-algebras is due to Christensen and Kirchberg. Combining [29] and [14, Theorem 3.1], it follows that a C ∗- algebra A has a positive answer to the similarity problem if and only if A has Christensen's property Dk from [13] for some k. Definition 4.1. Given an operator T ∈ B(H), we write ad(T ) for the derivation ad(T )(x) = xT − T x. A C ∗-algebra A has property Dk for some k > 0 if, for every non-degenerate representation π : A → B(H), the in- equality d(T, π(A)′) ≤ kkad(T )π(A)k (4.1) holds for all T ∈ B(H). A von Neumann algebra A is said to have the property D∗ k if the inequality (4.1) holds for all unital normal representations π on H and all T ∈ B(H). By taking weak∗-limit points, it follows that if A is a weak∗-dense C ∗- subalgebra of a von Neumann algebra M and A has property Dk, then M has property D∗ k. That property Dk converts near containments to completely bounded near containments originates in [13, Theorem 3.1]. The version we give below improves on the bounds γ′ = 6kγ from [13] and γ′ = (1 + γ)2k − 1 from [16, Corollary 2.12]. Proposition 4.2. Suppose that A has property Dk for some k > 0. Then for γ > 0, every near inclusion A ⊆γ B (or A ⊂γ B) with A and B acting non- degenerately on the same Hilbert space, gives rise to a completely bounded near inclusion A ⊆cb,γ ′ B (or A ⊂cb,γ ′), where γ′ = 2kγ. Proof. Suppose A ⊆γ B is a near inclusion of C ∗-algebras acting non- degenerately on H and fix n ∈ N. Let C = C ∗(A, B) and let π : C → B(K) be the universal representation of C. Then π(A)′′ has property D∗ k so that π(A)′′ ⊆cb,2kγ π(B)′′ by [8, Proposition 2.2.4]. By definition, for n ∈ N we have π(A)′′ ⊗ Mn ⊆2kγ π(B)′′ ⊗ Mn. As π is the universal representation of C the Hahn-Banach argument used to deduce [13, equation (3)] from [13, equation (2)] gives A ⊗ Mn ⊆2kγ B ⊗ Mn, as required. The result when we work with strict near inclusions A ⊂γ B follows immediately. (cid:3) C ∗-algebras with no bounded traces (such as stable algebras) where shown to have the similarity property in [20]. Using the property Dk version of this fact, the previous proposition gives automatic complete closeness when one algebra has no bounded traces. The argument below which transfers the absence of bounded traces to a nearby C ∗-algebra essentially goes back to [27, Lemma 9]. We use more recent results in order to get better estimates. 14 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Corollary 4.3. Suppose that A and B are C ∗-algebras which act non- degenerately on the same Hilbert space and satisfy d(A, B) < γ for γ < (2 + 6√2)−1. Suppose that A has no bounded traces (for example if A is sta- ble). Then B has no bounded traces, and therefore A ⊂cb,3γ B, B ⊂cb,3γ A and dcb(A, B) < 6γ. Proof. Suppose d(A, B) < (2 + 6√2)−1 and τ : B → C is a bounded trace. Let π : B → B(H) be the GNS-representation of B corresponding to τ . Then there is a larger Hilbert space H and a representation π : C ∗(A, B) → B( H) such that π is a direct summand of πB. That is, the projection p from H onto H is central in π(B) and π(b) = pπ(b)p for all b ∈ B. Then, by [27, Lemma 5], we have d(π(A)′′, π(B)′′) ≤ d(A, B), and hence there is a projection q ∈ π(A)′′ with kp − qk ≤ γ/√2 by [28, Lemma 1.10(ii)]. If q is an infinite projection in π(A)′′, then as d(A, B) < (2+6√2)−1, one can follow the argument of [16, Lemma 6.1] (using the estimate kp−qk < γ/√2 in place of kp − qk < 2γ) to see that p is infinite in π(B)′′, giving a contradiction. If q is finite, then qπ(A)′′q has a finite trace ρ and ρ ◦ πA defines a bounded trace on A, and again we have a contradiction. Thus B has no bounded traces. Theorem 2.4 of [12] shows that a properly infinite von Neumann algebra has property D∗ 3/2. As such, every C ∗-algebra with no bounded traces has property D3/2. Since A and B both have property D3/2, Proposition 4.2 gives A ⊂cb,3γ B and B ⊂cb,3γ A, whence dcb(A, B) < 6γ. (cid:3) Corollary 4.4. Suppose that A and B are C ∗-algebras which act non- degenerately on the same Hilbert space and satisfy d(A, B) < 1/252 and suppose that A has no bounded traces (for example when A is stable). Then (Cu(A), Σ(Cu(A)) ∼= (Cu(B), Σ(Cu(B)). Proof. Combine Corollary 4.3 with Theorem 3.10 (noting that 6d(A, B) < 1/42). (cid:3) We can use the Cuntz semigroup to show that stability transfers to close C ∗-algebras provided one algebra has stable rank one. To detect stability for a σ-unital C ∗-algebra we use the following criterion from [32, Lemma 5.4] which reformulates the earlier characterisation from [24]. Lemma 4.5. Let A be a σ-unital C ∗-algebra and let c ∈ A be a strictly positive element. Then, A is stable if and only if for every ǫ > 0, there is b ∈ A+ such that (c − ǫ)+ ⊥ b and (c − ǫ)+ - b. Following [40], we say that a C ∗-algebra A has weak cancellation provided Cu(A) satisfies the property that x + z ≪ y + z implies x ≤ y. It was proved in [41, Theorem 4.3] that if A has stable rank one, then W (A) has the property defining weak cancelation. When A has stable rank one, so too does A ⊗ K [38, Theorem 3.6], and so A has weak cancelation. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 15 Lemma 4.6. Let A be a σ-unital C ∗-algebra with weak cancelation. Then A is stable if and only if Cu(A) = Σ(Cu(A)). Proof. If A is stable, then Σ(Cu(A)) = Cu(A). Indeed, given n ∈ N, choose an automorphism θn : K ⊗ K → K ⊗ K with θn(K ⊗ e11) = K ⊗ Mn and let ψ : A → A ⊗ K be an isomorphism. Then (ψ−1 ⊗ idK)(idA ⊗ θn)(ψ ⊗ idK) is an automorphism of A ⊗ K which maps A ⊗ e11 onto A ⊗ Mn. In this way the class of a positive element in A ⊗ Mn lies in the scale Σ(Cu(A)). For x ∈ (A ⊗ K)+ and ε > 0, we have (x − ε)+ ∈ S∞ n=1(A ⊗ Mn), and hence h(x − ε)+i ∈ Σ(Cu(A)). Since the scale is defined to be closed under suprema, it follows that Cu(A) = Σ(Cu(A)). Conversely, let c ∈ A be a strictly positive element so that Σ(Cu(A)) = {x ∈ Cu(A) : x ≤ hci} and let ǫ > 0 be given. The hypothesis ensures that 2hci ≤ hci, and so we can find δ > 0 such that 2h(c − ǫ 4 )+i ≪ h(c − δ)+i. Now write (c − δ)+ = (c − δ)+gǫ/2,ǫ(c) + (c − δ)+(1M (A) − gǫ/2,ǫ(c)), and observe that h(c − δ)+gǫ/2,ǫ(c)i ≤ hgǫ/2,ǫ(c)i = h(c − ǫ 2 )+i ≪ h(c − ǫ 4 )+i. We now have that 2h(c − ǫ 4 )+i ≪ h(c − δ)+i ≤ h(c − ǫ 2 )+i + h(c − δ)+(1M (A) − gǫ/2,ǫ(c))i and so weak cancellation enables us to conclude that h(c − ǫ 4 )+i ≤ h(c − δ)+(1 − gǫ/2,ǫ(c))i. Let b = (c − δ)+(1 − gǫ/2,ǫ(c)). (c − ǫ)+ ≤ (c − ǫ A is stable. It is clear that b ⊥ (c − ǫ)+ and that 4 )+ - b. Thus we may invoke Lemma 4.5 to conclude that (cid:3) Theorem 4.7. Let A and B be σ-unital C ∗-algebras with A stable and d(A, B) < 1/252 and suppose either A or B has stable rank one. Then B is stable. Proof. By Corollary 4.4, we have an isomorphism (Cu(A), Σ(Cu(A)) ∼= (Cu(B), Σ(Cu(B)). Since A is stable Cu(A) = Σ(Cu(A)). Our isomorphism condition now tells us that Cu(B) = Σ(Cu(B)). If B has stable rank one, then it has weak cancelation, whereas if A has stable rank one, A has weak cancelation and, as weak cancelation is a property of the Cuntz semigroup, so too does B. The result now follows from Lemma 4.6. (cid:3) We now turn to the situation in which one C ∗-algebra is Z-stable. In [12], Christensen shows that McDuff II1 factors have property D5/2, and hence via the estimates of [36], have similarity length at most 5. (In fact McDuff factors, and more generally II1 factors with Murray and von Neumann's property Γ have length 3 [15], but at present we do not know how to use 16 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER this fact to obtain better estimates for automatic complete closeness of close factors with property Γ.) In [37, 30, 25] analogous results have been established in a C ∗-setting: in particular Z-stable C ∗-algebras ([25]) and C ∗-algebras of the form A ⊗ B, where B is nuclear and has arbitrarily large unital matrix subalgebras ([37]) have similarity degree (and hence length) at most 5. Here we show how to use the original von Neumann techniques from [12] to show that a class of algebras generalising both these examples have property D5/2 (recapturing the upper bound 5 on the length). A similar result has been obtained independently by Hadwin and Li [22, Corollary 1] working in terms of the similarity degree as opposed to property Dk. Once we have this Dk estimate, Proposition 4.2 applies. In particular we obtain uniform estimates on the cb-distance dcb(A, B) in terms of d(A, B) when A is Z-stable. Given a von Neumann algebra M ⊂ B(H) and x ∈ B(H), write cow M(x) for the weak∗ closed convex hull of {uxu∗ : u ∈ U(M)}. If M is injective, M(x) ∩ M′ is non-empty for all x ∈ B(H). then by Schwartz's property P, cow Note that for a non-degenerately represented C ∗-algebra A ⊂ B(H), we have kad(T )Ak = kad(T )A′′k. We say that an inclusion A ⊂ C of C ∗-algebras is non-degenerate if the inclusion map is non-degenerate. Proposition 4.8. Let C be a C ∗-algebra and A, B ⊂ C be commuting non- degenerate C ∗-subalgebras which generate C. Suppose B is nuclear and has no non-zero finite dimensional representations. Then C has property D5/2, and hence similarity length at most 5. Proof. Suppose C is non-degenerately represented on H and fix x ∈ B(H). The non-degeneracy assumption ensures that A and B are non-degenerately represented on H. Note that C ′′ has no finite type I part as B has no non- zero finite dimensional representations. Let p be the central projection in C ′′ so that C ′′p is type II1 and C ′′(1 − p) is properly infinite. Fix a unital type I∞ subalgebra M0 ⊂ (1 − p)C ′′(1 − p) and let M = (M0 ∪ pB)′′ which M(x)∩ (M∪{p})′. is injective. By Schwartz's property P, there exists y ∈ cow As in [12, Theorems 2.3, 2.4], ky − xk ≤ kad(x)C ′′k and kad(y)C ′′k ≤ kad(x)C ′′k. Write y1 = yp and y2 = y(1 − p). If p 6= 1, then the properly 2}′ and so by [12, Corollary 2.2], infinite algebra M0 lies in C ′′(1− p)∩{y2, y∗ kad(y2)C ′′(1−p)k = 2d(y2, C ′(1 − p)). Take x2 ∈ C ′′(1 − p) with kx2 − y2k = kad(y2)C ′′(1−p)k/2 ≤ kad(x)C ′′k/2. If p 6= 0, then we argue exactly as in the proof of [12, Proposition 2.8] to produce first z1 ∈ A′p with ky1 − z1k ≤ kad(y1)C ′′pk/2 ≤ kad(x)C ′′k/2. Continuing with the proof of [12, Proposition 2.8], as B′′p and A′′p commute, B ′′p(z1) ∩ B′p B ′′p(z1) is contained in A′p and hence there exists x1 ∈ cow cow with kx1 − z1k ≤ kad(z1)B ′′pk ≤ kad(z1 − y1)B ′′pk ≤ 2kz1 − y1k ≤ kad(x)C ′′k. Then ky1 − x1k ≤ ky1 − z1k + kz1 − x1k ≤ 3kad(x)C ′′k/2. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 17 If p = 0, take x1 = 0 and the same inequality holds. The element x1+x2 ∈ C ′ has kx − (x1 + x2)k ≤kx − yk + k(y1 − x1) + (y2 − x2)k ≤kad(x)C ′′k + max(ky1 − x2k,ky2 − x2k) ≤ 5kad(x)C ′′k/2. Therefore C has property D5/2, and so by [36, Remark 4.7] has length at most 5. (cid:3) Corollary 4.9. Let A be a Z-stable C ∗-algebra. Then A has property D5/2 and length at most 5. The main result of [16] is that the similarity property transfers to close C ∗-algebras. This work is carried out with estimates depending on the length and length constant of A, but it is equally possible to carry out this work entirely in terms of property Dk so it can be applied to Z-stable algebras. Our objective is to obtain a version of [16, Corollary 4.6] replacing the hypothesis that A has length at most ℓ and length constant at most K with the formally weaker hypothesis that A has property Dk (if A has the specified length and length constants, then it has property Dk for k = Kℓ/2, conversely if A has property Dk, then it has length at most ⌊2k⌋, but a length constant estimate is not known in this case, see [36, Remark 4.7]). This enables us to use Corollary 4.9 obtain an isomorphism between the Cuntz semigroups of sufficiently close C ∗-algebras when one algebra is Z- stable. To achieve a Dk version of [16, Section 4], we adjust the hypotheses in Lemma 4.1, Theorem 4.2 and Theorem 4.4 of [16] in turn, starting with Lemma 4.1. We begin by isolating a technical observation. Lemma 4.10. Let M be a finite von Neumann algebra with a faithful tracial state acting in standard form on H and let J be the conjugate linear modular conjugation operator inducing an isometric antisomorphism x 7→ JxJ of M onto M′ ∼= Mop. Suppose that S is another von Neumann algebra acting k, then M′ ⊂cb,2kγ nondegenerately on H with M′ ⊂γ S. If M has property D∗ S. Proof. As J is isometric, M ⊂γ J SJ, so that M ⊂cb,2kγ J SJ by Proposition 4.2. Now, for each n ∈ N, let Jn denote the isometric conjugate linear operator of component wise complex conjugation on Cn so that J ⊗ Jn is a conjugate linear isometry on H ⊗ Cn. We can conjugate the near inclusion M ⊗ Mn ⊂2kγ JSJ ⊗ Mn by J ⊗ Jn to obtain M′ ⊗ Mn ⊂2kγ S ⊗ Mn, as required. (cid:3) The next lemma is the modification of [16, Lemma 4.1]. The expression for β below is a slight improvement over that of the original. Lemma 4.11. Let M and N be von Neumann algebras of type II1 faithfully and non-degnerately represented on H with common centre Z which admits 18 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER a faithful state. Suppose d(M, N) = α and M has property D∗ k. If α satisfies 24(12√2k + 4k + 1)α < 1/200, then d(M′, N′) < 2β + 1200kα(1 + β) where β = 96kα(600k + 1). Proof. This amounts to showing that the hypothesis in [16, Lemma 4.1] that M contains an weak∗ dense C ∗-algebra A of length at most ℓ and length constant at most K can be replaced by the statement that M has property D∗ k (and that the specified expressions on β are valid). The hypothesis that M has such a weak∗ dense C ∗-algebra is initially used to see that M has property Dk at the beginning of the lemma and then applied to a unital normal representation to obtain [16, equation (4.5)]. As such property D∗ k suffices for this estimate. The other use of this hypothesis comes on page 385 in the last paragraph of the lemma, to obtain [16, equation (4.28)]. Using the notation of this paragraph, the von Neumann algebra TM is a cutdown of M acting as M⊗IG on H ⊗ G by the projection ei0,i0 from the commutant of M on this space. Since ei0,i0 is unitarily equivalent in this commutant to a projection of the form e ⊗ g0, where e is a projection from the commutant of M on H of full central support and g0 is a minimal projection in B(G), it follows that ei0,i0 has full central support in the commutant of M on H ⊗ G. As such TM is isomorphic to M, so has property D∗ k. Thus Lemma 4.10 can be applied to M ⊂48(600kα+α) T ′ the near inclusion T ′ N2 from [16, equation (4.25)] giving M ⊂cb,96k(600kα+α)) T ′ T ′ N2 . It then follows that TM ⊗ B(ℓ2(Λ)) ⊂96k(600kα+α) TN ⊗ B(ℓ2(Λ)), which is precisely [16, equation (4.28)] with our new estimate for β replacing that of the original. We then deduce that d(M′, N′) ≤ 2β + 1200kα(1 + β) in just the same way that [16, equation (4.30)] is obtained from [16, equation (4.28)]. (cid:3) Now we adjust Theorem 4.2 of [16]. The resulting constant β is obtained by taking α = 11γ in the previous lemma. Note that there is an unfortunate omission in the value of β in Theorem 4.2 of [16] which should be given by taking α = 11γ in Lemma 4.1 of [16], so should be K((1 + 316800kγ + 528γ)ℓ − 1): this has no knock on consequences to Theorem 4.4 of [16] where the correct value of β is used. Lemma 4.12. Let A and B be C ∗-algebras acting on a Hilbert space and suppose that d(A, B) = γ. Suppose A has property Dk and 24(12√2k + 4k + 1)γ < 1/2200. Then d(A′, B′) ≤ 10γ + 2β + 13200kγ(1 + β), where β = 1056k(600kγ + γ). THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 19 Proof. This amounts to replacing the hypothesis that A has length at most ℓ and length constant at most K with the condition that A has property Dk in Theorem 4.2 of [16]. The length hypothesis on A is used to show that certain II1 von Neumann closures of A satisfy [16, Lemma 4.1], but since the weak∗-closure of a C ∗-algebra with property Dk has property D∗ k, Lemma 4.11 can be used in place of [16, Lemma 4.1]. Note that in the proof of [16, Theorem 4.2] the reference to injective von Neumann algebras n=1 Mn has the similarity property). The correct statement is that these algebras have property D∗ (cid:3) having property D1 is incorrect (it is an open question whether Q∞ 1 which is all that is used. Finally we can convert Theorem 4.4 of [16]. Note the typo in the statement 1−2η−kγ . of this theorem, the definition of k should be The same change should be made in Corollary 4.6 of [16]. k k 1−2η−2kγ rather than Proposition 4.13. Let A and B be C ∗-subalgebras of some C ∗-algebra C with d(A, B) < γ and suppose that A has property Dk. Write β = 1056(600kγ + γ) and η = 10γ + 2β + 13200kγ(1 + β) and suppose that (4.2) 24(12√2k + 4k + 1)γ < , 2η + 2kγ < 1. Then dcb(A, B) ≤ 4kγ, where k = 1 2200 k 1 − 2η − 2kγ . . Proof. We check that B has property Dk. This amounts to weakening the hypothesis of [16, Theorem 4.4] in just the same way as the preceeding lemmas. Applying Lemma 4.12 in place of Theorem 4.2 of [16] in the proof of Theorem 4.4 of [16], shows that under the hypotheses of this proposition B has property Dk, where k = k 1 − 2η − 2kγ This is valid as property Dk descends to quotients so, following the proof of [16, Theorem 4.4], the algebra ρ(A) inherits property Dk allowing the use of Lemma 4.12 above in place of [16, Theorem 4.2]. Note that one should take care with issues of degeneracy here. In particular, the representation π of B in the proof of Theorem 4.4 of [16] should be assumed non-degenerate. Proposition 4.2 now shows that B ⊂cb,2k A and A ⊂cb,2k B. Therefore (cid:3) dcb(A, B) ≤ 2 max(2kγ, 2kγ) = 4kγ Corollary 4.14. Let A be a C ∗-algebra generated by two commuting non- degenerate C ∗-subalgebras one of which is nuclear and has no finite dimen- sional irreducible representations. Suppose that A ⊂ B(H) and B is an- other C ∗-subalgebra of B(H) with d(A, B) < γ for γ < 1/6422957. Then dcb(A, B) < 1/42 and (Cu(A), Σ(Cu(A)) ∼= (Cu(B), Σ(Cu(B)). 20 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Proof. By Proposition 4.8, A has property Dk for k = 5/2 so in Proposition 4.13, β = 1585056γ and η = 3203122γ + 52306848000γ2 so that 2η + 2kγ < 1011γ < 1 for γ < 10−11. The bound on γ ensures that (4.2) holds so that Proposition 4.13 applies. Further this bound gives 4kγ 1 − 2η − 2kγ < 1 42 , and so the result follows from Proposition 4.13 and Theorem 3.10. (cid:3) In particular, C ∗-algebras sufficiently close to Z-stable algebras are au- tomatically completely close and have the Cuntz semigroup of a Z-stable algebra. The question of whether the property of Z-stability transfers to sufficiently close subalgebras raised in [17] remains open. Corollary 4.15. Let A be a Z-stable C ∗-algebra and suppose that B is another C ∗-algebra acting on the same Hilbert space as A with d(A, B) < 1/6422957. Then dcb(A, B) < 1/42, (Cu(A), Σ(Cu(A)) ∼= (Cu(B), Σ(Cu(B)). In particular B has the Cuntz semigroup of a Z-stable algebra. 5. Quasitraces In this section we use our isomorphism between the Cuntz semigroups of completely close C ∗-algebras to give an affine homeomorphism between the lower semicontinuous quasitraces on such algebras. This isomorphism is compatible with the affine isomorphism of the trace spaces of close C ∗- algebras constructed in [16, Section 5]. Given a C ∗-algebra A, write T (A) for the cone of lower semicontinuous traces on A and QT2(A) for the cone of lower semicontinuous 2-quasitraces on A. Precisely, a trace τ on A is a linear function τ : A+ → [0,∞] van- ishing at 0 and satisfying the trace identity τ (xx∗) = τ (x∗x) for all x ∈ A. A 2-quasitrace is a function τ : A+ → [0,∞] vanishing at 0 which satis- fies the trace identity and which is linear on commuting elements of A+. Write Ts(A) for the simplex of tracial states on A and QT2,s(A) for the bounded 2-quasitraces on A of norm one. Lower semicontinuous traces and 2-quasitraces on A extend uniquely to lower semicontinuous traces and 2- quasitraces respectively on A ⊗ K (see [7, Remark 2.27(viii)]). In [19, Section 4], Elliott, Robert and Santiago extend earlier work of Blackadar and Handelman, setting out how functionals on Cu(A) arise from elements of QT2(A). Precisely a functional on Cu(A) is a map f : Cu(A) → [0,∞] which is additive, order preserving, has f (0) = 0 and preserves the suprema of increasing sequences. Given τ ∈ QT2(A), the expression dτ (hai) = limn→∞ τ (a1/n) gives a well defined functional on Cu(A), where we abuse notation by using τ to denote the extension of the original lower semicontinuous 2-quasitrace to A ⊗ K. Alternatively, one can define dτ by dτ (hai) = limn→∞ τ (an), where (an)∞ n=1 is any very rapidly increasing se- quence from (A⊗ K)+ 1 representing hai. Conversely, given a functional f on Cu(A), a lower semicontinuous 2-quasitrace on A ⊗ K (and hence on A) is THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 21 0 f (h(a − t)+i)dt. With this notation, the assignments The topology on QT2(A) is specified by saying that a net (τi) in QT2(A) given by τf (a) = R ∞ τ 7→ dτ and f 7→ τf are mutually inverse (see [19, Proposition 4.2]). converges to τ ∈ QT2(A) if and only if lim sup i τi((a − ε)+) ≤ τ (a) ≤ lim inf i τi(a) for all a ∈ A+ and ε > 0. With this topology QT2(A) is a compact Hausdorff space [19, Theorem 4.4] and T (A) is compact in the induced topology [19, Theorem 3.7]. In a similar fashion, the cone of functionals on Cu(A) is topologised by defining λi → λ if and only if lim sup i λi(h(a − ε)+i) ≤ λ(hai) ≤ lim inf i λi(hai) for all a ∈ (A ⊗ K)+ and ε > 0. Theorem 4.4 of [19] shows that the affine map τ 7→ dτ is a homeomorphism between the cone QT2(A) and the cone of functionals on the Cuntz semigroup. Theorem 5.1. (1) Let A, B be C ∗-algebras acting non-degenerately on a Hilbert space with dcb(A, B) < 1/42. Then the isomorphism Φ : (Cu(A), Σ(Cu(A))) → (Cu(B), Σ(Cu(B))) given by Theorem 3.10 induces an affine homeomorphism satisfying (5.1) bΦ : QT2(B) → QT2(A) d Φ(τ )(x) = dτ (Φ(x)) for all x ∈ Cu(A) and τ ∈ QT2(B). (2) Suppose additionally that A and B are unital and dcb(A, B) < γ < 1/2200. Then the map Φ from (1) is compatible with the map Ψ : Ts(B) → Ts(A) given in Lemma 5.4 of [16]. Precisely, for τ ∈ Ts(B), we have bΦ(τ ) ∈ Ts(A) ⊂ QT2(A) and bΦ(τ ) = Ψ(τ ). Proof. The first part of the theorem is a consequence of Theorem 3.10 and [19, Proposition 4.2]: given τ ∈ QT2(B), define bΦ(τ ) to be the lower semi- continuous 2-quasitrace induced by the functional dτ ◦ Φ on Cu(A). It is immediate from the construction that the map bΦ is affine, bijective and the identity (1) holds. To show that bΦ is continuous, we use the homeomorphism between the cone of lower semicontinuous quasi-traces and functionals on the Cuntz semi- group in [19, Theorem 4.4]. Consider a net (τi) in QT2(B) with τi → τ . Fix a ∈ A+, then, dτ (Φ(hai)) ≤ lim inf i dτi(Φ(hai)), as dτi → dτ . Now take ε > 0 and fix a contraction b ∈ (B ⊗ K)+ with Φ(hai) = hbi. As (h(b−1/n)+i)∞ n=1 is very rapidly increasing with supremum 22 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER hbi, there exists n ∈ N with Φ(h(a − ε)+i) ≤ h(b − 1/n)+i. As lim sup dτi(hb − 1/n)+i) ≤ dτ (hbi), it follows that i lim sup i d bΦ(τi)(h(a − ε)+i) ≤ d bΦ(τ )(hai) ≤ lim inf i d bΦ(τ )(hai). Thus d bΦ(τi) → d bΦ(τ ) and so, using the homeomorphism between QT2(A) and functionals on Cu(A), we have bΦ(τi) → bΦ(τ ). Therefore bΦ is continuous, and hence a homeomorphism between QT2(B) and QT2(A). For the second part we first need to review the construction of the map Ψ from [16]. Suppose dcb(A, B) < γ < 1/2200. Write C = C ∗(A, B) and let C ⊂ B(H) be the universal representation of C so that M = A′′ and N = B′′ are isometrically isomorphic to A∗∗ and B∗∗ respectively. Note that the Kaplansky density argument of [27, Lemma 5] gives dcb(M, N) ≤ dcb(A, B). Following the proof of [16, Lemma 5.4] we can find a unitary u ∈ (Z(M) ∪ Z(N))′′ such that Z(uMu∗) = Z(N) and ku − 1Ck ≤ 5γ. We write A1 = uAu∗ and M1 = uMu∗. There is now a projection zfin ∈ Z(M1) = Z(N) which simultaneously decomposes M1 = M1zfin ⊕ M1(1 − zfin) and N = Nzfin ⊕ N(1 − zfin) into the finite and properly infinite parts respectively ([16, Lemma 3.5] or [27]). Given a tracial state τ on B, there is a unique extension τ ′′ to N, which then factors uniquely through the centre valued trace TrNzfin on Nzfin. That is, τ ′′(x) = (φτ ◦ TrNzfin )(xzfin) for some state φτ on Nzfin. The map Ψ in [16] is then given by defining Ψ(τ )(y) = (φτ ◦ TrM1zfin)(uyu∗zfin) for y ∈ A. Now fix τ ∈ Ts(B). For m ∈ N and a ∈ (A⊗Mm)+ 1 , consider the standard very rapidly increasing sequence (g2−(n+1),2−n(a))∞ n=1 which represents hai. Let pn ∈ M ⊗ Mm be the spectral projection for a for [2−(n+1), 1], so that the alternating sequence g2−2,2−1(a), p1, g2−3,2−2(a), p2, g2−4,2−3(a), p3, . . . is very rapidly increasing. Then (5.2) dΨ(τ )(hai) = sup Choose bn ∈ (B ⊗ Mm)+ n (Ψ(τ ))(g2−n ,2−(n+1)(a)) = sup n Ψ(τ )′′(pn). 1 with kg2−(n+1),2−n(a)− bnk ≤ 2γ and projections qn ∈ N ⊗ Mm with kpn − qnk ≤ 2γ (by a standard functional calculus argument, see [11, Lemma 2.1]). Note that dcb(M1, N) ≤ 11γ and the algebras (M1 ⊗ Mm)(zfin ⊗ 1m) and (N1 ⊗ Mm)(zfin ⊗ 1m) have the same centre. Since k(u ⊗ 1m)p(u ⊗ 1m)∗(zfin ⊗ 1m) − q(zfin ⊗ 1m)k < 1/2, Lemma 3.6 of [16] applies to show that (TrM1zfin⊗trm)((u⊗1m)pn(u⊗1m)∗(zfin⊗1m)) = (TrNzfin⊗trm)(q(zfin⊗1m)). This ensures that Ψ(τ )′′(pn) = τ ′′(qn) for all n. As each (qn − 18γ)+ = qn, the sequence (b1 − 18γ)+, q1, (b2 − 18γ)+, q2, (b3 − 18γ)+, q3, . . . THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 23 is upwards directed by Lemma 3.8 and the supremum of this sequence defines Φ(hai). We then have (5.3) τ ′′(qn). dτ (Φ(hai)) = sup n Indeed, dτ (Φ(hai)) is given by sup τ (cn), where (cn)∞ n=1 is any very rapidly increasing sequence in (B⊗K)+ representing Φ(hai). But, working in Cu(N), Proposition 3.6 shows that any such very rapidly increasing sequence (cn)∞ n=1 can be intertwined with the very rapidly increasing sequence (qn)∞ n=1 after telescoping, and this establishes (5.3). Combining (5.2) and (5.3), we have (5.4) dΨ(τ )(hai) = d bΦ(τ )(hai) for all m ∈ N and a ∈ (A ⊗ Mm)+. As functionals on the Cuntz-semigroup preserve suprema, (5.4) holds for all a ∈ (A⊗K)+, whence Ψ(τ ) = bΦ(τ ). (cid:3) The homeomorphism between the lower semicontinuous quasi-traces can be used to establish the weak∗-continuity of the map between the tracial state spaces of close unital C ∗-algebras from [16, Section 5] resolving a point left open there. In particular this shows that the map defined in [16] provides an isomorphism between the Elliott invariants of completely close algebras, as a priori the For any closed two-sided ideal I ✂A, the subcone TI (A) of T (A) consists of those τ ∈ T (A) such that the closed two-sided ideal generated by {x ∈ A+ : τ (x) < ∞} is I. Proposition 3.11 of [19] shows that the relative topology on TI (A) is the topology of pointwise convergence on the positive elements of the Pedersen ideal of I. In particular, Ts(A) ⊂ TA(A). In particular, the induced topology on Ts(A) is just the weak∗-topology. Corollary 5.2. Suppose that A and B are unital C ∗-algebras acting non- degenerately on a Hilbert space with dcb(A, B) < 1/42 and d(A, B) < 1/2200. Then the affine isomorphism Ψ : Ts(B) → Ts(A) between tracial state spaces in [16, Section 5] is a homeomorphism with respect to the weak∗-topologies. We end with two further corollaries of Theorem 5.1. Corollary 5.3. Let A and B be unital C ∗-algebras acting non-degenerately on the same Hilbert space with dcb(A, B) < 1/2200. Suppose every bounded 2-quasitrace on A is a trace, then the same property holds for B. Proof. Given τ ∈ QT2,s(B), its image bΦ(τ ) lies in QT2,s(A) = Ts(A). By Theorem 5.1 (2) (applied with A and B interchanged) as claimed. τ = bΦ−1(bΦ(τ )) = Ψ−1(bΦ(τ )) ∈ Ts(B), (cid:3) The question of whether exactness transfers to (completely) close C ∗- algebras raised in [16] remains open, but we do at least obtain the following corollary. 24 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER Corollary 5.4. Let A and B be unital C ∗-algebras acting non-degenerately on the same Hilbert space with dcb(A, B) < 1/2200 and suppose A is exact. Then every bounded 2-quasitrace on B is a trace. Proof. This is immediate from Haagerup's result that bounded 2-quasitraces on exact C ∗-algebras are traces ([21]) and the previous corollary. (cid:3) We end by noting that the isomorphism between the Cuntz semigroups of completely close algebras in Theorem 3.10 can also be used to directly recapture an isomorphism between the Elliott invariants in significant cases. Let CuT be the functor A 7→ Cu(A ⊗ C(T)) mapping the category of C ∗- algebras into the category Cu introduced in [18] and let Ell be the Elliott invariant functor taking values in the category Inv whose objects are the 4-tuples arising from the Elliott invariant. Let C be the subcategory of separable, unital, simple finite and Z-stable algebras A with QT2(A) = T (A) (for example if A is exact). Then, building on work from [6, 5], Theorem 4.2 of [2] provides functors F : Inv → Cu and G : Cu → Inv such that there are natural equivalences of functors F ◦ EllC ∼= CuTC and G◦ CuTC ∼= EllC (a similar result for simple unital ASH algebras which are not type I and have slow dimension growth can be found in [43]). Note that in Theorem 4.2 there is an implicit nuclearity hypothesis, which is only actually used in order to see QT2(A) = T (A) -- the result holds in the generality stated. Thus if A and B are Z-stable C ∗-algebras with dcb(A, B) sufficiently small, and A is simple, separable, unital finite and has QT2(A) = T (A), then B enjoys all these properties. Further, since tensoring by an abelian algebra does not increase the complete distance between A and B (see [13, Theorem 3.2] for this result in the context of near inclusions -- the same proof works for the metric dcb), CuT(A) ∼= CuT(B) by Theorem 3.10. Thus Ell(A) ∼= Ell(B). Acknowledgements. This work was initiated at the Centre de Recerca Matem`atica (Bellaterra) during the Programme "The Cuntz Semigroup and the Classification of C ∗-algebras" in 2011. The authors would like to thank the CRM for the financial support for this programme and the conducive research environment. References [1] R. Antoine, J. Bosa, and F. Perera. Completions of monoids with applications to the Cuntz semigroup. Internat. J. Math., 22(6):837 -- 861, 2011. [2] R. Antoine, M. Dadarlat, F. Perera and L. Santiago. Recovering the Elliott invariant from the Cuntz semigroup. Trans. Amer. Math. Soc., in press, arXiv:1109.5803v1, [math.OA], 2011. [3] P. Ara and E. Pardo. Refinement monoids with weak comparability and applications to regular rings and C ∗-algebras. Proc. Amer. Math. Soc., 124(3):715 -- 720, 1996. [4] P. Ara, F. Perera, and A. S. Toms. K-theory for operator algebras. Classification of C ∗-algebras. In Aspects of operator algebras and applications, volume 534 of Contemp. Math., pages 1 -- 71. Amer. Math. Soc., Providence, RI, 2011. [5] N. P. Brown, A. S. Toms. Three applications of the Cuntz semigroup. Int. Math. Res. Not. IMRN, no. 19, Art. ID rnm068, 14 pp, 2007. THE CUNTZ SEMIGROUP AND STABILITY OF CLOSE C ∗-ALGEBRAS 25 [6] N. P. Brown, F. Perera, A. S. Toms. The Cuntz semigroup, the Elliott conjecture, and dimension functions on C∗-algebras. J. Reine Angew. Math. 621, 191-211, 2008. [7] E. Blanchard and E. Kirchberg. Non-simple purely infinite C ∗-algebras: the Hausdorff case. J. Funct. Anal., 207:461 -- 513, 2004. [8] J. Cameron, E. Christensen, A. Sinclair, R. R. Smith, S. White, and A. Wiggins. Kadison-Kastler stable factors. arXiv:1209.4116 [math.OA], 2012. [9] J. Cameron, E. Christensen, A. Sinclair, R. R. Smith, S. White, and A. Wiggins. A remark on the similarity and perturbation problems. C. R. Math. Rep. Acad. Sci. Canada, in press, arXiv:1206.5405 [math.OA], 2012. [10] J. Cameron, E. Christensen, A. Sinclair, R. R. Smith, S. White, and A. Wiggins. Type II1 factors satisfying the spatial isomorphism conjecture. Proc. Natl. Acad. Sci. USA.,109(5) 20338-20343, 2012. [11] E. Christensen. Perturbations of type I von Neumann algebras. J. London Math. Soc. (2), 9:395 -- 405, 1975. [12] E. Christensen. Perturbations of operator algebras. II. Indiana Univ. Math. J., 26(5):891 -- 904, 1977. [13] E. Christensen. Near inclusions of C ∗-algebras. Acta Math., 144(3-4):249 -- 265, 1980. [14] E. Christensen. Extensions of derivations. II. Math. Scand., 50(1):111 -- 122, 1982. [15] E. Christensen. Finite von Neumann algebra factors with property Γ. J. Funct. Anal., 186:366-380, 2001. [16] E. Christensen, A. Sinclair, R. R. Smith, and S. White. Perturbations of C ∗-algebraic invariants. Geom. Funct. Anal., 20(2):368 -- 397, 2010. [17] E. Christensen, A. Sinclair, R. R. Smith, S. White, and W. Winter. Perturbations of nuclear C ∗-algebras. Acta Math., 208:93 -- 150, 2012. [18] K. T. Coward, G. A. Elliott, and C. Ivanescu. The Cuntz semigroup as an invariant for C ∗-algebras. J. Reine Angew. Math., 623:161 -- 193, 2008. [19] G. Elliott, L. Robert, and L. Santiago. The cone of lower semicontinuous traces on a C ∗-algebra. Amer. J. Math., 133(4):969 -- 1005, 2011. [20] U. Haagerup. Solution of the similarity problem for cyclic representations of C ∗- algebras. Ann. of Math. (2), 118(2):215 -- 240, 1983. [21] U. Haagerup. Quasi-traces on exact C ∗-algebras are traces. Preprint, 1991. [22] D. Hadwin and W. Li The similarity degree of some C ∗-algebras. arXiv:1210.6015. [math.OA], 2012. [23] I. Hirshberg, E. Kirchberg, and S. White. Decomposable approximations of nuclear C ∗-algebras. Adv. Math., 230:1029 -- 1039, 2012. [24] J. Hjelmborg and M. Rørdam. On stability of C ∗-algebras. J. Funct. Anal., 155(1):153 -- 170, 1998. [25] M. Johanesova and W. Winter. The similarity problem for -stable C ∗-algebras. Bull. London Math. Soc., 44(6):1215 -- 1220, 2012. [26] R. V. Kadison. On the orthogonalization of operator representations. Amer. J. Math., 77:600 -- 620, 1955. [27] R. V. Kadison and D. Kastler. Perturbations of von Neumann algebras. I. Stability of type. Amer. J. Math., 94:38 -- 54, 1972. [28] M. Khoshkam. Perturbations of C ∗-algebras and K-theory. J. Operator Theory, 12(1):89 -- 99, 1984. [29] E. Kirchberg. The derivation problem and the similarity problem are equivalent. J. Operator Theory, 36(1):59 -- 62, 1996. [30] W. Li and J. Shen. A note on approximately divisible C ∗-algebras. arXiv:0804.0465 [math.OA], 2008. [31] H. Matui and Y. Sato. Strict comparison and Z-absorption of nuclear C ∗-algebras. Acta. Math., in press. arXiv:1111.1637v1 [math.OA], 2011. 26 F. PERERA, A. TOMS, S. WHITE, AND W. WINTER [32] E. Ortega, F. Perera, and M. Rørdam. The Corona factorization property, stability, and the Cuntz semigroup of a C ∗-algebra. Int. Math. Res. Not. IMRN, (1):34 -- 66, 2012. [33] F. Perera. The structure of positive elements for C ∗-algebras with real rank zero. Internat. J. Math., 8(3):383 -- 405, 1997. [34] J. Phillips. Perturbations of C ∗-algebras. Indiana Univ. Math. J., 23:1167 -- 1176, 1973/74. [35] J. Phillips and I. Raeburn. Perturbations of AF-algebras. Canad. J. Math., 31(5):1012 -- 1016, 1979. [36] G. Pisier. The similarity degree of an operator algebra. St. Petersburg Math. J., 10(1):103 -- 146, 1999. [37] F. Pop. The similarity problem for tensor products of certain C ∗-algebras. Bull. Aus- tral. Math. Soc., 70(3):385 -- 389, 2004. [38] M. A. Rieffel. Dimension and stable rank in the K-theory of C ∗-algebras. Proc. Lon- don Math. Soc.,, 46:301 -- 333, 1983. [39] L. Robert Classification of inductive limits of 1-dimensional NCCW complexes. Adv. Math., 231(5):2802 -- 2836, 2012. [40] L. Robert and L. Santiago. Classification of C ∗-homomorphisms from C0(0, 1] to a C ∗-algebra. J. Funct. Anal., 258(3):869 -- 892, 2010. [41] M. Rørdam and W. Winter. The Jiang-Su algebra revisited. J. Reine Angew. Math., 642:129 -- 155, 2010. [42] A. Toms. On the classification problem for nuclear C ∗-algebras. Ann. of Math. (2), 167(3):1029 -- 1044, 2008. [43] A. Tikuisis. The Cuntz semigroup of continuous functions into certain simple C ∗- algebras. Int. J. Math. 22, 1-37, 2011. [44] W. Winter. Decomposition rank and Z-stability. Invent. Math., 179(2):229 -- 301, 2010. [45] W. Winter. Nuclear dimension and Z-stability of pure C ∗-algebras. Invent. Math., 187(2):259 -- 342, 2012. [46] S. Zhang. Diagonalizing projections in multiplier algebras and in matrices over a C ∗-algebra. Pacific J. Math., 145(1):181 -- 200, 1990. Francesc Perera, Department of Mathematics, 08193 Bellaterra, Barcelona, Spain. E-mail address: [email protected] Andrew Toms, Department of Mathematics, Purdue University, 150 North University Street, West Lafayette, IN 47907, USA. E-mail address: [email protected] Stuart White, School of Mathematics and Statistics, University of Glasgow, University Gardens, Glasgow Q12 8QW, Scotland. E-mail address: [email protected] Wilhelm Winter, Mathematisches Institut der WWU Munster, Einsteinstrasse 62, 48149, Munster, Germany. E-mail address: [email protected]
1203.1806	2	1203	2012-05-15T12:19:03	Some analysis on amalgamated free products of von Neumann algebras in non-tracial setting	[ "math.OA" ]	Several techniques together with some partial answers are given to the questions of factoriality, type classification and fullness for amalgamated free product von Neumann algebras.	math.OA	math	SOME ANALYSIS ON AMALGAMATED FREE PRODUCTS OF VON NEUMANN ALGEBRAS IN NON-TRACIAL SETTING YOSHIMICHI UEDA Abstract. Several techniques together with some partial answers are given to the questions of factoriality, type classification and fullness for amalgamated free product von Neumann algebras. 1. Introduction It was quite recent that the complete answers were given in [27, 28] to the questions of factoriality, type classification, fullness and Sd- and τ -invariants for arbitrary free product von Neumann algebras. It is natural as a next project to consider the same questions for more general amalgamated free product von Neumann algebras. Such attempts were already made by us [21, 24, 23] almost 10 years ago for amalgamated free products over Cartan subalgebras. However the results there are far from satisfactory as compared to those on plain free prod- uct von Neumann algebras. The aim of this paper is to take a still very first step towards 'satisfactory' answers to those questions for amalgamated free product von Neumann algebras. As simple consequences we will give partial answers at least when amalgamated free products are taken over type I von Neumann algebras, which are improvements of our previous works [21, 24, 23, 25, 26]. The proofs in [27, 28] are divided into analytical and combinatorial parts in essence. Combi- natorial parts are completed by some 'induction arguments', whose essential idea originates in several works due to Dykema, especially [6]. On the other hand, analytical parts are devoted to proving several inequalities involving the Hilbert space norms arising from some states of particular form (instead of so-called free product states themselves), whose essential ideas ap- parently go back to the ICC argument for factoriality of group von Neumann algebras and the so-called 14 ε-argument both due to Murray and von Neumann. However our problems are of the nature of type III von Neumann algebras, and thus the lack of trace causes main difficulties. Hence the key is to overcome such difficulties. Here we will take up such analytical aspects in the general amalgamated free product setup, and indeed improve the analytical results in [27, 28] with new techniques from the recent amazing development on type II1 factors opened by several breakthroughs due to Popa. We hope that the technical facts provided in this paper are sufficient as analytical parts in future 'best-possible' answers to the questions mentioned above at least in the case where amalgamated free products are taken over type I von Neumann subalgebras. The organization of this paper is as follows. Section 2 is preliminaries on amalgamated free product von Neumann algebras. In section 3 we provide a non-tracial version of one of the results in Ioana -- Peterson -- Popa's article [9, Theorem 1.1]. In relation to it we provide a non-tracial adaptation of the so-called intertwining-by-bimidule criterion due to Popa, which may be of independent interest as future reference. In the same section we also generalize our previous results of controlling central sequences [27, Proposition 3.5],[28, Proposition 3.1] to the amalgamated free product setting. In section 4, we give several partial answers to 1 2 Y. UEDA the questions mentioned above by utilizing technologies developed in §3. Those include an answer to the factoriality and non-amenability questions of a given amalgamated free product (M, E) = (M1, E1) ⋆N (M2, E2) when M1 is 'diffuse relative to N ', M2 'non-trivial relative to N ', and N of type I. Standard notation rule here follows our previous papers [27, 28]; for example, the center, the unitary group and the set of projections of a given von Neumann algebra M are denoted by Z(M ), M u and M p, respectively, and also the central support of e ∈ M p in M by cM e . Notations and facts concerning amalgamated free products of von Neumann algebras will be summarized in next section 2. 2. Amalgamated Free Product von Neumann Algebras Let M1 ⊇ N ⊆ M2 be σ-finite von Neumann algebras, and faithful normal conditional expectations E1 : M1 → N , E2 : M2 → N be given. Their amalgamated free product (M, E) = (M1, E1) ⋆N (M2, E2) is a pair of von Neumann algebra M containing M1 ⊇ N ⊆ M2 and faithful normal conditional expectation E : M → N satisfying (i) M = M1 ∨ M2, (ii) 2 )≡ 0, where Λ◦(M ◦ 2 ) denotes the set of all E ↾Mk = Ek (k = 1, 2) and (iii) E ↾Λ◦(M ◦ alternating words in M ◦ 2 := Ker(E2). The construction of such a pair is a bit complicated, but this simple formulation perfectly serves as a working definition. The construction was introduced in the tracial setting in [15] based on the C∗-algebraic one [31]. Its modular theoretical treatment was given in [21], and will be reviewed below. 1 ,M ◦ 1 := Ker(E1) and M ◦ 1 , M ◦ Let χ be a faithful normal semifinite weight on N . Then the modular automorphism σχ◦E , t t ∈ R, is simply computed as σχ◦E t ↾Mk = σχ◦Ek t (k = 1, 2), (2.1) see [21, Theorem 2.6]. This formula together with famous Takesaki's criterion shows that for each k = 1, 2 there is a unique faithful normal conditional expectation EMk : M → Mk characterized by (2.2) EMk ↾Λ◦(M ◦ 1 ,M ◦ 2 )\M ◦ k≡ 0. This fact is easily confirmed in the exactly same way as in [27, Lemma 2.1]. It is clear that E ◦ EMk = E holds. Consider the natural inclusion of the so-called continuous cores: fM := M ⋊σχ◦E R ⊇ fMk := Mk ⋊σχ◦Ek R (k = 1, 2) ⊇ eN := N ⋊σχ R, which is independent of the choice of χ thanks to Connes's Radon-Nikodym cocycle theorem. (2.3) σχ◦E R, eEk := Ek ¯⊗IdB(L2(R)) ↾Mk⋊ The canonical liftings (still being faithful normal conditional expectations) eE : fM → eN , eEk : fMk → eN (k = 1, 2) are constructed by eE := E ¯⊗IdB(L2(R)) ↾M ⋊ Remark that the original E and Ek are recovered as the restrictions of eE and eEk to M and Mk via the canonical embeddings M ֒→ fM and Mk ֒→ fMk, respectively. Here is a simple but important fact [21, Theorem 5.1] that fM1 and fM2 are freely independent with amalgamation over eN with respect to eE, and moreover fM = fM1 ∨ fM2. Consequently the following natural (k = 1, 2) and Tr eN on fM , fMk and = Tr eN ◦ eEk eN , respectively, (see [20, Theorem XII.1.1]) must satisfy TrfM = Tr eN ◦ eE and Tr fMk (see e.g. [14, §4]). (fM , eE) = (fM1, eE1) ⋆ eN (fM2, eE2). The canonical faithful normal semifinite traces Tr fM , TrfMk formula holds: (2.4) R . σχ◦Ek (2.5) NON-TRACIAL AMALGAMATED FREE PRODUCTS 3 Let M ω ⊇ M ω k (k = 1, 2) ⊇ N ω be the ultraproducts of M ⊇ Mk (k = 1, 2) ⊇ N . Here the inclusion relation is guaranteed by the existence of conditional expectations, and E and Ek (k = 1, 2) can be lifted up to Eω : M ω → N ω and Eω k → N ω, respectively. All the necessary facts on ultraproducts of von Neumann algebras are summarized in [27, §§2.2]. Remark that M ω 2 are freely independent with amalgamation over N ω with respect to Eω, see [23, Proposition 4]. However it is hopeless due to [18, Lemma 2.2] that M ω = M ω 1 ∨M ω 2 holds. 1 and M ω k : M ω 3. Technical Results 3.1. A non-tracial adaptation of Popa's intertwining-by-bimodule criterion. Let M be an arbitrary σ-finite (possibly type III) von Neumann algebra, and A, B be its (possibly non- unital) von Neumann subalgebras with units 1A, 1B, respectively. Suppose that B is semifinite with a faithful normal semifinite trace TrB and furthermore that there is a faithful normal conditional expectation EB : 1BM 1B → B. Proposition 3.1. The following are equivalent: (i) There is no net uλ of unitaries in A which satisfies EB(y∗uλx) −→ 0 σ-strongly for any x, y ∈S(cid:8)1AM p p ∈ Bp; TrN (p) < +∞}. (ii) There are a normal (possibly non-unital) ∗-homomorphism ρ : A → Mn(C) ¯⊗B with finite n ∈ N and a non-zero partial isometry w ∈ Mn(C) ¯⊗M such that -- (Trn ¯⊗TrB)(ρ(1A)) < +∞, -- ww∗ ≤ e11 ⊗ 1A and w∗w ≤ ρ(1A), and -- (e11 ⊗ a)w = wρ(a) for all a ∈ A. (iii) There are non-zero projections e ∈ A, f ∈ B, a normal unital ∗-isomorphism θ : eAe → f Bf and a non-zero partial isometry v ∈ M such that e is finite in A and TrB(f ) < +∞, -- the central support cA -- vv∗ ≤ e and v∗v ≤ f , and -- xv = vθ(x) for all x ∈ eAe. Suppose further that M has an almost periodic weight ψ such that both A and B sit inside the centralizer Mψ, ψ ↾B is still semifinite, and the EB is the unique ψ ↾1BM1B -preserving one. Then the w in (ii) and the v in (iii) can be chosen in such a way that there is a common eigenvalue λ of ∆ψ so that (idn ¯⊗ σψ t (v) = λitv for all t ∈ R. t )(w) = λitw and σψ As usual let us write A (cid:22)M B (with EB and TrB) if the above equivalent conditions (i) -- (iii) hold. Remark that no assumption on A is necessary. The proof is of course modeled after Popa's original one for finite von Neumann algebras, but some cares are necessary. Indeed we observed this fact with B finite several years ago, through our attempt to get better understanding of the fundamental articles [16, 17] due to Popa. Houdayer and Vaes informed us that they have also observed it with B finite independently (see [8, Theorem 2.3]), and moreover Vaes corrected our misunderstanding on some argument in [4, §2]. The proof below is just a combination and/or a reformulation of several existing proofs of Popa's criterion [16, Appendix],[17, §2] (also see [2, Appendix F],[29, Appendix C] for its exposition) and its variants [1, §3],[4, §2],[7, §4], etc. The same idea as in e.g. the proof of (1) ⇒ (4) in [29, Proposition C.1] perfectly works for (ii) ⇒ (i). (Note that the proof of (4) ⇒ (1) in [2, Theorem F.12] does not work at this point due to the lack of finite trace. Thus we could not prove (iii) ⇒ (i) directly.) Hence the main parts below are (ii) ⇔ (iii) and (i) ⇒ (ii). Proof of (ii) ⇒ (i): We may assume that ρ(1A) = Pn to [11, Corollary 3.20]. Since (Trn ¯⊗TrB)(ρ(1A)) < +∞, one has w = Pn k=1 ekk ⊗ pk with pk ∈ Bp thanks k=1 e1k ⊗ wk with 4 Y. UEDA One can find a net uλ in Au in such a way that EB(w∗ i,j=1 eij ⊗ EB(w∗ wk = wkpk ∈ S(cid:8)1AM p p ∈ Bp; TrN (p) < +∞}. On contrary, suppose that (i) is not true. i uλwj ) −→ 0 σ-strongly for all i, j, and hence ρ(uλ)(id ¯⊗EB)(w∗w) = Pn i uλwj) −→ 0 σ-strongly. Therefore, k(idn ¯⊗EB)(w∗w)kTrn ¯⊗TrB = kρ(uλ)(idn ¯⊗EB)(w∗w)ρ(1A)kTrn ¯⊗TrB −→ 0, a contradiction to w 6= 0. (cid:3) k=1 ek = cA e1 = cA k=1 e1k ⊗ cA kvk = e1 and vkv∗ i,j=1 eij ⊗ θ(v∗ i=1 viv∗ i,k=1 e1ieik ⊗ viv∗ i = cA Proof of (iii) ⇒ (ii): Since v∗v ∈ θ(eAe)′, one can find a non-zero z ∈ Z(eAe)p = (Z(A)e)p in such a way that the normal ∗-homomorphism x ∈ zAz = (eAe)z 7→ θ(x)v∗v is injective. Since cA e is finite in A, by [12, Proposition 8.2.1] one can find non-zero, mutually orthogonal and equivalent (in A) e1, . . . , en ∈ Ap in such a way that e1 ≤ z and Pn e1 . We have e1vv∗ 6= 0, since θ(e1)v∗v = v∗(e1vv∗)v by the choice of z and e1 ≤ z. Then one gets partial isometries v1 := e1, v2, . . . , vn ∈ A so that v∗ k = ek (k = 2, . . . , n). Since e1Ae1 ⊆ eAe, we can construct a normal ∗-homomorphism ρ : A → Mn(C) ¯⊗B by ρ(a) := Pn i avj ), a ∈ A. Set w := Pn k=1 e1k ⊗ vkv with v in (iii), which defines i vj v = δijv∗e1v and v∗e1v = θ(e1)v∗v 6= 0 as remarked a non-zero partial isometry, since v∗v∗ ek for all k = 2, . . . , n, we have wρ(a) = Pn before. Since Pn i,j,k=1 e1iejk ⊗ i avkv =Pn j avk) =Pn vivθ(v∗ ek avkv = (e11 ⊗ a)w for all a ∈ A. Since ρ(1A) ≤ 1n ⊗ f , one has (Trn ¯⊗TrB)(ρ(1A)) < +∞. (cid:3) Proof of (ii) ⇒ (iii): As in (ii) ⇒ (i) we may and do assume that ρ(1A) = Pn k=1 ekk ⊗ pk with TrB-finite pk ∈ Bp. Note that any union of finite number of TrB-finite projections is again TrB-finite thanks to the Kaplansky formula [12, Theorem 6.1.7]. Thus p = Wn k=1 pk is TrB-finite, and replacing B by pBp (if necessary) we may and do assume that TrB(1B) < +∞. Notice that A must be of the form A = A0 ⊕ Ker(ρ(−)w∗w) with A0 finite, since ρ(A) is finite. Note here that w∗w ∈ ρ(A)′, and thus ρ(−)w∗w is a normal ∗-homomorphism. Let us first assume that A0 has a type II1 direct summand. By [12, Lemma 6.5.6] one can find nonzero, mutually orthogonal and equivalent (in A0) e1, . . . , en ∈ Ap 0 whose sum is the unit of the type II1 direct summand. With the center-valued trace τ : Mn(C) ¯⊗B → C1 ¯⊗Z(B) we have nτ (ρ(e1)) ≤ τ (1⊗ 1B) = nτ (e11 ⊗ 1B), implying that there is a partial isometry v1 ∈ Mn(C) ¯⊗B such that v∗ 1 ≤ e11 ⊗ 1B, we can construct a normal unital ∗-isomorphism θ : eAe → f Bf with e := e1, f := θ(e) in such a way 1 for x ∈ eAe. Since w∗w ∈ ρ(A)′ ∩ ρ(1A)(cid:0)Mn(C) ¯⊗M(cid:1)ρ(1A) and that e11 ⊗ θ(x) = v1ρ(x)v∗ ww∗ ∈(cid:0)Ce11 ¯⊗A(cid:1)′ 1 is a non-zero partial isometry whose left and right support projections are less than e11 ⊗ e and e11 ⊗ f , respectively, and hence wv∗ 1 = e11 ⊗ v for some non-zero partial isometry v ∈ eM f . Then one has e11 ⊗ xv = (e11 ⊗ x)wv∗ 1 = wρ(x)v∗ We next consider the case that A0 is of type I, that is, there is an abelian (in A) e ∈ Ap e = 1A0. With a MASA A between ρ(eAe) ⊕ Cρ(e)⊥ ⊆ Mn(C) ¯⊗B one can choose, by with cA [11, Theorem 3.18], mutually orthogonal and equivalent (in Mn(C) ¯⊗B) projections q1, . . . , qn from A with Pn k=1 qk = 1n ⊗ 1B. Then one immediately observes (by looking at their center- valued traces) that every qk is equivalent to e11 ⊗ 1B in Mn(C) ¯⊗B. Since ρ(e)w∗w 6= 0, some q := qk must satisfy qρ(e)w∗w 6= 0. In this way, we can choose a non-zero partial isometry v1 ∈ Mn(C) ¯⊗B in such a way v∗ 1 ≤ e11 ⊗ 1B, and thus v∗ 1v1 ∈ ρ(eAe)′ 1 v1 = qρ(e)(≤ ρ(e)), v1v∗ 1 6= 0 (since qρ(e)w∗w 6= 0). Then we can construct a unital normal ∗-homomorphism and wv∗ θ : eAe → f Bf with f := θ(e) by e11 ⊗ θ(x) = v1ρ(x)v∗ 1 for x ∈ eAe and a non-zero y ∈ eM f by e11 ⊗ y = wv∗ 1 ) = e11 ⊗ yθ(x) for x ∈ eAe, since v∗ 1v1 ∈ ρ(eAe)′. Hence xy = yθ(x) for x ∈ eAe. Replacing e by suitable z ∈ Z(eAe)p (if necessary) we can make θ injective with keeping both θ(e) = f and ∩ (e11 ⊗ 1A)(cid:0)Mn(C) ¯⊗M(cid:1)(e11 ⊗ 1A), it is easy to see that wv∗ 1 = e11 ⊗ vθ(x) for x ∈ eAe. 1 . Moreover we have e11 ⊗ xy = (e11 ⊗ x)wv∗ 1 ≤ e11 ⊗ 1B. Since v1ρ(e1)v∗ 1 = wρ(x)v∗ 1 = wv∗ 1 (v1ρ(x)v∗ 1v1 = ρ(e1) and v1v∗ 1 = v1v∗ 1 = wv∗ 1 v1ρ(x)v∗ 0 NON-TRACIAL AMALGAMATED FREE PRODUCTS 5 y = eyf . With the polar decomposition y = vy we get vv∗ ≤ e, v∗v ≤ f and xv = vθ(x) for x ∈ eAe. (cid:3) We have two ways for completing the final part of the proof of (i) ⇒ (ii) below; one is the use of Haagerup's Lp-space technologies and the other that of standard forms due to Araki, Connes and Haagerup. Here we use the latter as easy way. In what follows (M y H, JM , P♮ M ) denotes a standard form of M , see [20, Definition IX.1.13]. Proof of (i) ⇒ (ii): Note that EB(y∗uλx) −→ 0 σ-strongly if and only if kEB(y∗uλx)kTrB −→ 0 for any x, y ∈S{1AM p p ∈ Bp; TrB(p) < +∞}. Thus there are ε > 0 and F ⋐S{1AM p p ∈ Bp; TrB(p) < +∞} so that Xx,y∈F kEB(y∗ux)kTrB ≥ ε for all u ∈ Au. (3.1) Bx1⊥ B)1⊥ BM 1⊥ B, and set B := B ⊕ C1⊥ Each x ∈ F has a TrB-finite px ∈ Bp with x = xpx, and p := Wx∈F px must be TrB-finite as remarked in (ii) ⇒ (iii). Thus, replacing B by pBp (if necessary) we may and do assume that TrB is a finite trace, that is, TrB(1B) < +∞. Choose a faithful normal state ϕ0 on 1⊥ B and E B : x ∈ M 7→ EB(1Bx1B) + ϕ0(1⊥ B giving a faithful normal conditional expectation from the whole M onto B. Clearly B is still finite (since we have assumed that TrB is a finite trace), and the B ∈ B 7→ TrB(b) + α ∈ C defines a faithful normal trace (not weight !) Tr B on mapping b + α1⊥ B. Set ϕ := Tr B ◦ E B, a faithful normal positive linear functional on M , and let ξ0 ∈ P♮ M be its unique representing vector. It is standard, by a usual exhaustion argument like e.g. the proof of [19, Theorem IV.5.5], to see that there is a family of vectors {ξi}i∈I in H so that ξ0 is in the family (thus 0 is regarded as a distinguished element in I) and moreover H =P⊕ i∈I [JM BJM ξi]. Therefore, one can construct an isometry U : H → ℓ2(I)⊗ L2( B) satisfying U ξ0 = δ0 ⊗ ΛTr B (1) and U (JM x∗JM ) = (1⊗ J Bx∗J B)U for x ∈ B, where L2( B) is the usual standard Hilbert space constructed out of Tr B, ΛTr B the canonical embedding of B to L2( B) and J B the canonical unitary conjugation on L2( B). By the construction we observe that P := U U ∗ ∈ B(ℓ2(I)) ¯⊗ B and moreover that the pair P(cid:0)B(ℓ2(I)) ¯⊗ B(cid:1)P and PCδ0 ⊗ 1 with the rank 1 projection PCδ0 onto Cδ0 is nothing but a concrete realization, modulo the unitary equivalence by U , of the basic extension hM, Bi and the Jones projection e B associated with E B. Then TrhM, Bi(−) := (TrB(ℓ2(I)) ¯⊗Tr B)(U (−)U ∗) (3.2) with the usual trace TrB(ℓ2(I)) on B(ℓ2(I)) gives a faithful normal semifinite trace on the basic extension hM, Bi. For x ∈ B one has U xe BU ∗ = PCδ0 ⊗ x and hence TrhM, Bi(xe B) = (TrB(ℓ2(I)) ¯⊗Tr B)(U xe BU ∗) = (TrB(ℓ2(I)) ¯⊗Tr B)(PCδ0 ⊗ x) = Tr B(x). Therefore, we get TrhM, Bi(xe By) = TrhM, Bi(e Byxe B) = TrhM, Bi( EB(yx)e B) = ϕ(yx), x, y ∈ M. (3.3) Let d :=Py∈F ye By∗ ∈ hM, Bi+, and then TrhM, Bi(d) = Py∈F ϕ(y∗y) < +∞ by (3.3). In the exactly same way as in the proof of (1) ⇒ (2) of [2, Theorem F.12] we see, by using (3.1), that the σ-weakly closed convex hull C of {u∗du u ∈ Au} does not contain 0. Moreover, it is plain to see that JM 1BJM d = d. Since 1B ∈ Z( B) and hence JM 1BJM ∈ Z(hM, Bi), we conclude that C sits in (1AJM 1BJM )hM, Bi(1AJM 1BJM ). Since d ≥ 0 and TrhM, Bi(d) < +∞, C is embedded, as a closed convex set, into L2(hM, Bi, TrhM, Bi), the usual GNS Hilbert space associated with TrhM, Bi. Hence one can choose a unique minimal point d0 ∈ C with respect to , which in turn falls in (1AJM 1BJM )hM, Bi(1AJM 1BJM )∩A′ the Hilbert space norm k−kTrhM, Bi 6 Y. UEDA aξj = Xn i=1 and satisfies TrhM, Bi(d0) < +∞. Choosing a suitable spectral projection of d0 we get a nonzero projection e ∈ hM, Bi ∩ A′ such that e ≤ 1AJM 1BJM and TrhM, Bi(e) < +∞. The projection e apparently gives an A -- B bimodule K := eH with left and right (unital) actions a · ξ · b := aJM b∗JM ξ for a ∈ A, ξ ∈ K, b ∈ B. The GNS representation of B associated with TrB is simply given by the restriction B y L2(B) := 1BL2( B) with the canonical embedding ΛTrB := ΛTr B ↾B, and moreover the canonical unitary conjugation JB is also just the restriction of J B to L2(B). Thus we get the right B-module embedding U0 := U ↾K: K ֒→ ℓ2(I) ¯⊗L2(B)B (⊆ (1 ⊗ 1B)(ℓ2(I) ¯⊗L2( B)), and U0U ∗ 0 ∈ B(ℓ2(I)) ¯⊗B satisfies (TrB(ℓ2(I)) ¯⊗TrB)(U0U ∗ 0 ) = TrhM, Bi(e) < +∞ by (3.2). By the same reason as in the beginning of the proof of [2, Proposition F.10] or by [29, Lemma A.1] there are n ∈ N and a nonzero z ∈ Z(B)p such that K0 := JM zJMK is still a non-trivial A -- B bimodule and (U0 ↾JM zJM K 0 - Pn ⊗ z in B(ℓ2(I)) ¯⊗B = )(U0 ↾JM zJM K)∗ = (1 ⊗ JBzJB)U0U ∗ (C1 ¯⊗JBBJB)′, where Pn is a rank n projection in B(ℓ2(I)). Choose a partial isometry v ∈ (C1 ¯⊗JBBJB)′ with v∗v = (U0 ↾JM zJM K)(U0 ↾JM zJM K)∗ and vv∗ ≤ Pn ⊗ z, and then we can define a right B-module embedding V : K0 ֒→ Cn ¯⊗L2(B) by V := v(U0 ↾JM zJM K) with a fixed identification Pnℓ2(I) = Cn. The embedding V gives the normal (possibly non-unital) ∗-homomorphism ρ : a ∈ A 7→ V aV ∗ ∈ Mn(C) ¯⊗B. Let δi (1 ≤ i ≤ n) be a standard basis of Cn, and set ξi (1 ≤ i ≤ n). For a ∈ A, write ρ(a) = Pn := V ∗(δi ⊗ ΛTrB (1B)) ∈ K0 i,j=1 eij ⊗ ρ(a)ij with the matrix units eij associated JM ρ(a)∗ 0 (1 ⊗ JBzJB) = (1 ⊗ z)U0U ∗ with the δi, and then ij JM ξi, 1 ≤ j ≤ n. M. Set ξ := Pn (3.4) Consider M := Mn+1(C) ¯⊗M y L2(M) := Mn+1(C) ¯⊗H (by left matrix-multiplication) with the canonical unitary conjugation JM defined by JM(eij ⊗ ξ) := eji ⊗ (JM ξ) for eij ⊗ ξ ∈ L2(M). The natural cone determined by (M y L2(M), JM) is denoted by P♮ k=1 e0k ⊗ ξk ∈ L2(M), and define a normal (possibly non-unital) ∗-homomorphism ρ : A ֒→ M by ρ(a) := e00 ⊗ a + Pn i,j=1 eij ⊗ ρ(a)ij for a ∈ A. Here a standard matrix unit system eij in Mn+1(C) is indexed by 0, 1, . . . , n. By (3.4) one has ρ(a) ξ = JM ρ(a)∗JM ξ for a ∈ A. A standard fact on polar decomposition in standard forms (c.f. [20, Exercise IX.1.2],[1, Lemma 3.1]) guarantees the existence of a vector ξ ∈ P♮ M and a partial isometry w ∈ M satisfying that w ξ = ξ, w∗ w = [M′ ξ], w w∗ = [M′ ξ] and ρ(a) w = w ρ(a) for a ∈ A. Since (e00 ⊗ 1A) ξ = ξ, one has (e00 ⊗ 1A)[M′ ξ] = [M′ ξ], and thus w w∗ ≤ e00 ⊗ 1A. Here (ρ(A) ⊆) Mn(C) ¯⊗M is naturally regarded as a corner of M by the numbering of the matrix units eij's. Then one has, by (3.4) again, JMρ(1A)JM ξ = ξ, and hence JMρ(1A)JM ξ = ξ. By JM ξ = ξ ∈ P♮ M we get ρ(1A)[M′ ξ] = [M′ ξ] so that w∗ w ≤ ρ(1A) ≤Pn k=1 ekk ⊗ 1B. Therefore, w =Pn k=1 e0k ⊗ wk with wk ∈ 1AM 1B. Letting w := (e10 ⊗ 1A) w = Pn k=1 e1k ⊗ wk ∈ Mn(C) ¯⊗M we have w∗w ≤ ρ(1A), ww∗ ≤ e11 ⊗ 1A and (e11 ⊗ a)w = (e10 ⊗ 1A)ρ(a) w = (e10 ⊗ 1A) w ρ(a) = wρ(a) for a ∈ A. We have assumed (by cutting by a projection in B) that TrB(1B) < +∞, and hence (Trn ¯⊗TrB)(ρ(1A)) < +∞ is now trivial. Hence we are done. (cid:3) Proof of the second part of the assertion: Only the proof of (i) ⇒ (ii) needs small modification to prove this. Let us explain this in what follows. The standard form (M y H, JM , P♮ M ) is constructed from ψ so that JM ∆ψJM = ∆−1 ψ . The TrB is given by ψ ↾B. We need an extra argument in relation to the d0 ∈ (1AJM 1BJM )hM, Bi(1AJM 1BJM ) ∩ A′. By the assumption here the modular operator ∆ψ has a diagonalization ∆ψ = Pλ>0 λ eψ ψ ∈ hM, Bi ∩ A′ for all t ∈ R. Hence all the eψ 0 with some λ defines a non-zero element in hM, Bi ∩ A′. Since eψ λ commutes with 1AJM 1BJM and since λ 's fall in hM, Bi ∩ A′. Thus eψ λ and satisfies ∆it λ d1/2 NON-TRACIAL AMALGAMATED FREE PRODUCTS 7 λ ) = TrhM, Bi(d1/2 0 eψ λ d1/2 0 λ d0 eψ ψ ) ξ still falls in P♮ λ . Hence the A -- B bimodule K0 can be chosen as a subspace of eψ TrhM, Bi(eψ ) ≤ TrhM, Bi(d0) < +∞, we may and do assume d0 = λ H. Therefore, the ξ ∈ eψ λ d0 eψ ψ ) ξ = λit ξ for all t ∈ R. Since IMn+1(C) ¯⊗ ∆ψ L2(M) = Mn+1(C) ¯⊗H satisfies that (IMn+1(C) ¯⊗ ∆it is the modular operator of Trn+1 ¯⊗ ψ on M, (IMn+1(C) ¯⊗ ∆it M, see [20, Lemma IX.1.4]. Remark here that JM there is nothing but the one constructed from Trn+1 ¯⊗ ψ. Hence, ψ ) ξ = ξ hold by the uniqueness of polar decomposition (id ¯⊗σψ for every t ∈ R. These modifications are enough to complete the proof. (cid:3) Remark 3.2. Let E B and ϕ = Tr B ◦ E B be as in the proof of (i) ⇒ (ii) above. Let cE B : hM, Bi → M be the dual operator-valued weight associated with E B in the sense of [13, §§1.2]. It is known that the modular operator ∆ϕ and Connes's spacial derivative (d(ϕ◦ cE B ))/(d(Tr B ◦ AdJM ((−)∗))) must coincide, see e.g. the proof of [10, Proposition 2.2]. Moreover ∆ϕ is affiliated with hM, Bi, since ϕ = Tr B ◦ E B. With these two facts one can prove that the modular operator ∆ϕ is the Radon -- Nikodym derivative of ϕ ◦ cE B , i.e., ϕ ◦ cE B = TrhM, Bi,∆ϕ in the sense of [20, Lemma VIII.2.8]. This explains, in full generality, the relationship that was pointed out in [17, Eq.(1.3.1)] in the almost periodic case. t )( w) = λit w and (IMn+1(C) ¯⊗ ∆it 3.2. A non-tracial version of Ioana -- Peterson -- Popa's theorem. Let us investigate an amalgamated free product (M, E) = (M1, E1) ⋆N (M2, E2). Proposition 3.3. Let A be a (unital) von Neumann subalgebra of the centralizer (M1)ϕ of a certain faithful normal state ϕ, and M1 be a (possibly non-unital) dense (in any von Neumann algebra topology) ∗-subalgebra of M1 with E1(M1) ⊆ M1. Suppose that there is a net vλ of unitaries in A such that E1(y∗vλx) −→ 0 σ-strongly for all x, y ∈ M1. Then any unitary u ∈ M with uAu∗ ⊆ M1 must fall in M1. In particular, NM (A) = NM1 (A) and A′ ∩ M = A′ ∩ M1. Here NP (Q) denotes the set of unitaries u ∈ P with uQu∗ = Q for a given unital inclusion P ⊇ Q of von Neumann algebras. This is nothing but a non-tracial version of [9, Theorem 1.1] due to Ioana, Peterson and Popa. Although the proof below is modeled after their proof, we need to overcome some difficulties due to the lack of trace by utilizing modular theoretic technologies. Proof. Let (M y H, JM , P♮ M be the unique repre- senting vector of ϕ ◦ EM1 . Let eM1 be the so-called Jones projection associated with EM1 , i.e., eM1xξ0 = EM1 (x)ξ0 for x ∈ M , and the basic extension hM, M1i is defined to be M ∨{eM1}′′ = 1JM y H. Consider the projection p := [AJM M1JM u∗ξ0] ∈ A′ ∩ (JM M1JM )′ = JM M ′ A′ ∩ hM, M1i. Notice that aJM x∗JM u∗ξ0 = JM x∗JM u∗(uau∗)ξ0 for a ∈ A and x ∈ M1, and moreover that uau∗ ∈ M1 can be approximated in any von Neumann algebra topology, by analytic elements, say yλ, in M1 with respect to the modular action σϕ. Those altogether show that M ) be a standard form of M , and ξ0 ∈ P♮ aJM x∗JM u∗ξ0 = lim λ JM x∗JM u∗yλξ0 = lim λ JM x∗σϕ i/2(yλ)∗JM u∗ξ0 ∈ [JM M1JM u∗ξ0] ↾M1 = σϕ t ϕ◦EM1 thanks to σ t (t ∈ R) and [20, Lemma VIII.3.18 (ii)]. Consequently we get p ≤ [JM M1JM u∗ξ0] = u∗eM1u, which and dEM1 (eM1) = 1 imply kdEM1 (p)k∞ < +∞, where dEM1 : hM, M1i → M denotes the dual operator-valued weight of EM1 . See [13, §§1.2, Lemma In fact, if this is the case, then p ≤ eM1 so 3.1]. We will prove (1 − eM1)p(1 − eM1) = 0. that u∗ξ0 = eM1u∗ξ0 = EM1 (u∗)ξ0, implying u = EM1 (u) ∈ M1 since ξ0 is separating for M y H. Since kdEM1(p)k∞ < +∞ and dEM1 (eM1 ) = 1 as before, any spectral projection f of (1−eM1)p(1−eM1) corresponding to [δ, 1] with arbitrary δ > 0 still satisfies kdEM1 (f )k∞ < +∞. 8 Y. UEDA Therefore, it suffices to prove that any projection f ∈ A′ ∩hM, M1i satisfying both f ≤ 1− eM1 and kdEM1 (f )k∞ < +∞ must be 0. In what follows we denote by A the ∗-subalgebra of M consisting of all analytic elements with respect to σϕ◦EM1 , which is well-known to be dense in any von Neumann algebra topol- ogy. Set ψ := ϕ ◦ EM1 ◦ dEM1 , a faithful normal semifinite weight on hM, M1i, and let hM, M1i y L2(hM, M1i, ψ) be the GNS representation with canonical embedding Λψ : nψ := {x ∈ hM, M1i ψ(x∗x) < +∞} → L2(hM, M1i, ψ) and norm k − kψ associated with the weight t ◦ EM1 for all t ∈ R) and thus ψ. Remark that EM1 (A) ⊆ A (thanks to EM1 ◦ σ span(AeM1A) becomes a dense (in any von Neumann algebra topology) ∗-subalgebra of n∗ ψ∩nψ, and hence Λψ(span(A eM1A)) is dense in L2(hM, M1i, ψ) by [10, Lemma 2.1]. Thus one can choose a sequence Tn ∈ span(A eM1A) in such a way that kΛψ(Tn − f )kψ −→ 0 as n → ∞, where note that f clearly falls in nψ. Since f ≤ 1 − eM1 and σψ t (eM1 ) = eM1 (t ∈ R) [13, Lemma 5.1], we also have kΛψ((1 − eM1 )Tn(1 − eM1) − f )kψ −→ 0 as n → ∞ so that may and do assume that Tn = (1 − eM1)Tn(1 − eM1) for all n. On contrary, suppose f 6= 0, that is, γ := kΛψ(f )kψ (cid:9) 0. Then one can choose T := Tn0 ∈ span(AeM1A) with some n0 in such a way that ϕ◦EM1 t = σϕ kΛψ(T )kψ ≤ 3γ/2, kΛψ(T − f )kψ ≤ γ/5. (3.5) For any v ∈ Au we compute γ2 − ψ(T ∗vT v∗) ≤ ψ(f vf v∗) − ψ(T ∗vT v∗) ≤ ψ((f − T )∗vf v∗) + ψ(T ∗v(f − T )v∗) ≤ kΛψ(f − T )kψkΛψ(vf v∗)kψ + kΛψ(T )kψkΛψ(v(f − T )v∗)kψ ≤ kΛψ(f − T )kψkΛψ(f )kψ + kΛψ(T )kψkΛψ(f − T )kψ ≤ γ2/2, where the first, the third, the fourth and the fifth inequalities follow from f ∈ A′ ∩ hM, M1i, the Cauchy -- Schwarz inequality, v ∈ (M1)ϕ ⊂ hM, M1iψ, and (3.5), respectively. Therefore, γ2 ≤ 2ψ(T ∗vT v∗) holds for all v ∈ Au. Since T = (1 − eM1)T (1 − eM1 ), we can write T =Pm k=1 xkeM1 yk with xk, yk ∈ A ∩ Ker(EM1 ). Thus, for every v ∈ Au we have keM1x∗ γ2 ≤ 2Xm = 2Xm = 2Xm = 2Xm ≤ 2 max k,l=1 ψ(y∗ kEM1 (x∗ k,l=1 ψ(y∗ k,l=1 ϕ ◦ EM1 (y∗ k,l=1 ϕ ◦ EM1 (σ 1≤k≤mkykk∞ max 1≤l≤mkσ kvxleM1ylv∗) kvxl)eM1ylv∗) kEM1 (x∗ kvxl)ylv∗) (yl)v∗y∗ kEM1 (x∗ ϕ◦EM1 i ϕ◦EM1 i (yl)k∞Xm kvxl)) k,l=1 kEM1 (x∗ kvxl)kϕ. Here the third equality is due to dEM1 (eM1) = 1, the fourth one follows from v ∈ (M1)ϕ ⊆ and yl ∈ A with the so-called modular condition, and finally the last inequality is due Mϕ◦EM1 to the Cauchy -- Schwarz inequality. Consequently we have chosen x1, . . . , xm ∈ A ∩ Ker(EM1 ) and a universal constant C > 0 so that γ2 ≤ CXm k,l=1 kEM1(x∗ kvxl)kϕ for all v ∈ Au. (3.6) Set M◦ 1 := M1 ∩ M ◦ 1 . By the assumption on M1 and by the Kaplansky density theorem any 1 can be approximated in any von Neumann algebra topology by a bounded net of element x ∈ M ◦ NON-TRACIAL AMALGAMATED FREE PRODUCTS 9 λ = xλ − E1(xλ) ∈ M◦ 2 )\ M◦ 1 with xλ ∈ M1, xλ −→ x. Thus M1 + span(Λ◦(M◦ elements x◦ 1) is also dense in M in any von Neumann algebra topology so that the Kaplansky density theorem enables us to approximate each xk (= xk − EM1 (xk)) by a net xk,λ in span(Λ◦(M◦ 2 )\ M◦ 1); namely kxk,λk∞ ≤ 2kxkk∞ and xk,λ −→ xk σ-∗-strongly. Then we have, for every v ∈ Au, 1, M ◦ 1, M ◦ kEM1 (x∗ ≤ kEM1(x∗ ≤ kEM1(x∗ kvxl)kϕ ≤ kEM1(x∗ k,λvxl,λ)kϕ + kEM1 (x∗ k,λvxl,λ)kϕ + k(xk − xk,λ)∗vxl)kϕ◦EM1 k − x∗ k,λvxl,λ)kϕ + kσ (xl)k∞kx∗ ϕ◦EM1 i/2 k,λvxl,λ)kϕ kvxl − x∗ + kxk,λv(xl − xl,λ)kϕ◦EM1 k,λkϕ◦EM1 + 2kxkk∞kxl − xl,λkϕ◦EM1 , where we used, in the last line, that xl ∈ A with [20, Lemma VIII.3.18 (ii)] and v ∈ (M1)ϕ. Let ε > 0 be arbitrary chosen. Then some λ (being independent of v's) satisfies that γ2 ≤ ε + CPm 1 is written as azb with a, b ∈ {1} ∪ M◦ 2 whose leftmost and rightmost letters are chosen from M ◦ j , i = 1, 2, j = 1, . . . , m′, and positive constants Cj > 0, j = 1, . . . , m′, so that k,λvxl,λ)kϕ for all v ∈ Au. Since any element in Λ◦(M◦ 1, z an alternating word in M◦ j z(i) 2 , there are finitely many such words a(i) k,l=1 kEM1(x∗ 2 ) \ M◦ j b(i) 1, M ◦ 1, M ◦ γ2 ≤ ε + Xm′ = ε + Xm′ j=1 j=1 CjkEM1 (a(1) Cjka(1) j z(1) j b(1) j va(2) j z(2) j b(2) j )kϕ j EM1 (z(1) j E1(b(1) j va(2) j )z(2) j )b(2) j kϕ for all v ∈ Au, where the equality comes from the free independence of M1, M2 and (2.2). Applying the above estimate of γ2 to the net v = vλ in our hypothesis we get γ2 ≤ ε (at the limit in λ), a contradiction to γ (cid:9) 0, since ε is arbitrary. (cid:3) Remark 3.4. It is worth while to note that the inequality (3.6) is a general fact. Let P ⊇ Q be σ-finite von Neumann algebras with a faithful normal conditional expectation EQ : P → Q and A be a von Neumann subalgebra of the centralizer Qϕ with some faithful normal state ϕ. The middle part of discussion above shows that for each projection f ∈ A′ ∩ hP, Qi satisfying both f ≤ 1 − eQ and kcEQ(f )k∞ < +∞ there are analytic (with respect to σϕ◦EQ) elements x1, . . . , xm ∈ P and a universal constant C > 0 such that k,l=1 kEQ(x∗ for all v ∈ Au. kΛϕ◦EQ◦ EQ ϕ◦EQ◦ EQ ≤ CXm kvxl)kϕ (f )k2 3.3. A result for controlling central sequences in amalgamated free products. Let us investigate central sequences in an amalgamated free product (M, E) = (M1, E1) ⋆N (M2, E2). The next result is an adaptation and/or an improvement of the methods of [27, Proposition 3.5] and [28, Proposition 3.1] to amalgamated free product von Neumann algebras. In this subsection we use the notations and facts summarized in [27, §§2.2]. Proposition 3.5. Suppose that there is a faithful normal state ϕ on M1 satisfying the following conditions: (a) σϕ (b) For every n ∈ N with n ≥ 2 there are unitaries uk = u(n) t (N ) = N for all t ∈ R. k ∈ (M1)ϕ, 0 ≤ k ≤ n − 1, such that E1(u∗ k1 vk2 ) = 0 for all 0 ≤ k1 6= k2 ≤ n − 1, where Eϕ N denotes the unique ϕ-preserving conditional expectation from M1 onto N , whose existence follows from (a) and Takesaki's criterion. k , vk = v(n) k1 uk2) = Eϕ N (v∗ ϕ ∩ M ω, any y ∈ M ◦ 2 and any sequence (tm)m of real numbers we have Then, for any x ∈ (M1)′ with z :=(cid:2)(σ ϕ◦EM1 tm kE2(y∗y)1/2(x − (EM1 )ω(x))k(ϕ◦EM1 )ω ≤ kyx − xzk(ϕ◦EM1 )ω , (y))m(cid:3) ∈ M ω. 10 Y. UEDA Remark here that any bounded sequence (σ (x(m)))m with arbitrary (x(m))m giving an element in M ω gives again an element in M ω, as shown in the proof of [28, Proposition 3.1]. A key fact behind this is that any modular action σψ satisfies ψ ◦ σψ t = ψ for all t ∈ R. In particular, the element z in the statement above makes sense. ϕ◦EM1 tm 2 M ◦ 1 , M ◦ 1 or Eϕ 1 , M ◦ 1 , that span(Λ◦(M ◦ Proof. Write M ▽ N + M ◦ linear span of the following sets of words: 1 := Ker(Eϕ N (x) + (x − Eϕ 1 ··· M ◦ M ▽ } {z 2 ··· M ◦ N ). One can easily see, by using x ∈ M1 7→ E1(x) + (x − E1(x)) ∈ N (x)) ∈ N + M ▽ 1 ) coincides with the 1 ··· M ◦ } {z Define four closed subspaces X1 := (cid:2)Λϕ◦EM1 X3 := (cid:2)Λϕ◦EM1 Denote by Pi, i = 1, 2, 3, 4, the projection from H onto Xi. Remark that 2 ) \ M ◦ 2 ··· M ◦ } {z 2 )(cid:3), 1 )(cid:3), X2 := (cid:2)Λϕ◦EM1 1 ··· M ◦ 2 )(cid:3) in H := L2(M, ϕ ◦ EM1 ), and 2 ··· M ◦ } {z 1 ··· M ◦ 2 M ▽ 1 )(cid:3), X4 := (cid:2)Λϕ◦EM1 (M ◦ 2 ··· M ◦ H = Λϕ◦EM1 (M1) ⊕ X1 ⊕ X2 ⊕ X3 ⊕ X4. 1 , M ◦ , M ◦ clearly 2 M ▽ alternating alternating alternating alternating (M ◦ (M ◦ (M ◦ M ▽ 2 2 2 . (cid:18)IH − X4 Pi(cid:19) Λϕ◦EM1 i=1 x ∈ M. Let n ∈ N with n ≥ 2 be fixed. Define unitary operators Sk = S(n) (vkxv∗ 0, . . . , n − 1) on H by (x) = Λϕ◦EM1 (EM1 (x)), (ukxu∗ (x) := Λϕ◦EM1 SkΛϕ◦EM1 k), (3.7) (k = k , Tk = T (n) k TkΛϕ◦EM1 (x) := Λϕ◦EM1 x ∈ M, in our hypothesis. Here are simple claims. k), with uk = u(n) k , vk = v(n) k ∈ (M1)ϕ ⊆ Mϕ◦EM1 (A) {SkXi}n−1 (B) {TkX2}n−1 k=0 is an orthogonal family of closed subspaces, i = 3, 4. k=0 is an orthogonal family of closed subspaces. The proofs of those are essentially same, but (A) is easier than (B). Thus we prove only (B) here and leave (A) to the reader. By using x 7→ Ei(x) + (x − Ei(x)) ∈ N + M ◦ i (i = 1, 2) again and again we have (vk2 (M ◦ k2 )∗(vk1 (M ◦ 1 ··· M ◦ = vk2 (M ◦ 2 )v∗ 2 ··· (M ◦ 1 ··· M ◦ 2 )v∗ k1 ) 1 )··· M ◦ 2 )v∗ 1 v∗ k2 vk1 M ◦ k1 ⊆ vk2 N v∗ k1 + vk2 Ker(EM1 )v∗ k1 . The desired assertion immediately follows from that vk ∈ (M1)ϕ; in fact, if k1 6= k2, then ϕ ◦ EM1 (vk2 N v∗ ϕ ◦ EM1 (vk2 Ker(EM1 )v∗ k1 ) = ϕ(N v∗ k1 vk2 ) = ϕ(N Eϕ N (v∗ k1 vk2 )) = {0}, k1 ) = ϕ(vk2 EM1 (Ker(EM1 ))v∗ k1 ) = {0}. Let us choose arbitrary x ∈ (M1)′ ϕ ∩ M ω with representative (x(m))m. For each ε > 0 and each n ∈ N with n ≥ 2 one can choose a neighborhood W = Wε,n in β(N) at ω so that (x(m) − vkx(m)v∗ k , vk = v(n) k k)kH < ε are as above. For each kΛϕ◦EM1 (x(m) − ukx(m)u∗ k)kH < ε, kΛϕ◦EM1 for all 0 ≤ k ≤ n − 1 and m ∈ W ∩ N, where the uk = u(n) i = 3, 4 and every m ∈ W ∩ N we have, with the above Sk = S(n) k , kPiΛϕ◦EM1 nXn−1 = nXn−1 k=0 kSkPiΛϕ◦EM1 k=0nkSkPiΛϕ◦EM1 (x(m))k2 (x(m)) − SkPiS∗ (x(m))k2 (x(m))k2 kΛϕ◦EM1 ≤ 1 2 H H H + kSkPiS∗ kΛϕ◦EM1 Ho (x(m))k2 NON-TRACIAL AMALGAMATED FREE PRODUCTS 11 (ukx(m)u∗ k − x(m))k2 H + 2 nXn−1 k=0 kSkPiS∗ kΛϕ◦EM1 (x(m))k2 H = 2 nXn−1 < 2ε2 + kΛϕ◦EM1 k=1 kSkPiS∗ nXn−1 2 2 nkΛϕ◦EM1 k=0 kSkPiS∗ (x(m))k2 ∞/n. ≤ 2ε2 + ≤ 2ε2 + 2k((x(m))mk2 Similarly, using the claim (B) with T (n) k kΛϕ◦EM1 (x(m))k2 H H (by the claim (A)) kP2Λϕ◦EM1 (x(m))k2 k we have instead of S(n) H < 2ε2 + 2k((x(m))mk2 ∞/n for every m ∈ W ∩ N. Since n and ε are arbitrary, for each δ > 0 one can find a neighborhood Wδ in β(N) at ω so that k(P2 + P3 + P4)Λϕ◦EM1 (x(m))kH < δ (3.8) for all m ∈ Wδ ∩ N. In the standard embedding L2(M ω, (ϕ ◦ EM1 )ω) ֒→ Hω we have, by (3.7) and (3.8), [Dϕ ◦ EM1 : Dχ ◦ E]t σχ◦E t (y) [Dϕ ◦ EM1 : Dχ ◦ E]∗ t e−t2/ℓ dt √ℓπ with a fixed faithful normal state χ on N . Clearly yℓ falls in the σ-weak (or σ-strong) closure of span(M1M ◦ 2 M1), since [Dϕ ◦ EM1 : Dχ ◦ E]t = [Dϕ : Dχ ◦ E1]t ∈ M1 by [20, Corollary IX.4.22 (yℓ))m(cid:3) ∈ M ω, which is well-defined as (ii)] and σχ◦E remarked just before the proof. Note that σ each ℓ we have, by (3.7), (3.8) as before and by [20, Lemma VIII.3.18 (ii)], 2 by (2.1). Set zℓ := (cid:2)(σ ϕ◦EM1 tm ϕ◦EM1 −i/2 (y) ∈ M ◦ (yℓ)). For (yℓ)) = σ ϕ◦EM1 −i/2 ϕ◦EM1 tm ϕ◦EM1 tm (σ (σ t = lim (cid:13)(cid:13)(cid:13)Λ(ϕ◦EM1 )ω ((x − (EM1 )ω(x))zℓ) −(cid:2)(Jσ m→ω(cid:13)(cid:13)Jσ (yℓ))∗J(cid:0)Λϕ◦EM1 m∈Wδ ∩N(cid:13)(cid:13)Jσ ≤ sup ϕ◦EM1 −i/2 ϕ◦EM1 −i/2 ϕ◦EM1 tm ϕ◦EM1 tm (σ (σ ϕ◦EM1 −i/2 (σ ϕ◦EM1 tm (yℓ))∗JP1Λϕ◦EM1 (x(m) − EM1 (x(m))) − P1Λϕ◦EM1 (x(m)))m(cid:3)(cid:13)(cid:13)(cid:13)Hω (x(m))(cid:1)(cid:13)(cid:13)H ϕ◦EM1 −i/2 (yℓ)k∞δ (x(m))(cid:13)(cid:13)H < kσ (yℓ))∗J(P2 + P3 + P4)Λϕ◦EM1 (x(m)))m(cid:3)(cid:13)(cid:13)(cid:13)Hω (x(m))(cid:13)(cid:13)H = lim = lim (y(x(m) − EM1 (x(m)))) − yP1Λϕ◦EM1 (cid:13)(cid:13)(cid:13)Λ(ϕ◦EM1 )ω (y(x − (EM1 )ω(x))) −(cid:2)(yP1Λϕ◦EM1 m→ω(cid:13)(cid:13)Λϕ◦EM1 (x(m))(cid:13)(cid:13)H m→ω(cid:13)(cid:13)y(P2 + P3 + P4)Λϕ◦EM1 (x(m))(cid:13)(cid:13)H < kyk∞δ, m∈Wδ ∩N(cid:13)(cid:13)y(P2 + P3 + P4)Λϕ◦EM1 Λ(ϕ◦EM1 )ω (y(x − (EM1 )ω(x))) =(cid:2)(yP1Λϕ◦EM1 ≤ sup (x(m)))m(cid:3) (yEM1 (x(m)) − EM1 (x(m))σ ϕ◦EM1 tm (3.9) (3.10) (y)))m(cid:3). and hence in Hω, since δ is arbitrary. Trivially, in Hω, Λ(ϕ◦EM1 )ω (y(EM1 )ω(x) − (EM1 )ω(x)z) =(cid:2)(Λϕ◦EM1 Set ϕ◦EM1 σ t (y) e−t2/ℓ dt √ℓπ yℓ :=Z +∞ =Z +∞ −∞ −∞ 12 Y. UEDA with the modular conjugation J of M y H = L2(M, ϕ ◦ EM1 ). Hence, for each ℓ, Λ(ϕ◦EM1 )ω ((x − (EM1 )ω(x))zℓ) =(cid:2)(Jσ in Hω, since δ is arbitrary. Note that ϕ◦EM1 −i/2 (σ ϕ◦EM1 tm (yℓ))∗JP1Λϕ◦EM1 (x(m)))m(cid:3) (3.11) On the other hand, yP1Λϕ◦EM1 (xm)) ∈ spanΛϕ◦EM1 (M ◦ 2 M ◦ 1 ··· M ◦ 2 M ▽ 1 ). Λϕ◦EM1 and (yEM1 (x(m)) − EM1 (x(m))σ ∈ spanΛϕ◦EM1 (y)) ϕ◦EM1 tm (M ◦ ⊕ spanΛϕ◦EM1 2 ) ⊕ spanΛϕ◦EM1 (M ◦ (M ◦ 2 ) ⊕ spanΛϕ◦EM1 1 M ◦ 2 M ▽ 1 ) (M ◦ 1 M ◦ 2 M ▽ 1 ) Jσ ϕ◦EM1 −i/2 (σ ϕ◦EM1 tm (x(m)) (yℓ))∗JP1Λϕ◦EM1 1 ··· M ◦ (M ◦ ∈ spanΛϕ◦EM1 ⊆ Λϕ◦EM1 2 M ▽ 1 σ (M1) ⊕ spanΛϕ◦EM1 ϕ◦EM1 tm (M ◦ 1 M ◦ (yℓ)) 2 ··· ). (yℓ) falls in the σ- Here the last fact follows from [20, Lemma VIII.3.18 (ii)] and that σ strong closure of span(M1M ◦ 2 M1). Therefore, we see, by (3.9) -- (3.11), that Λ(ϕ◦EM1 )ω (y(x − (EM1 )ω(x))) is orthogonal to both Λ(ϕ◦EM1 )ω (y(EM1 )ω(x) − (EM1 )ω(x)z) and Λ(ϕ◦EM1 )ω ((x − −tm (y))m(cid:3), both of which (EM1 )ω(x))zℓ). Finally, letting x := (cid:2)(σ −tm (x(m)))m(cid:3), y :=(cid:2)(σ fall in M ω as remarked just before the proof, we have ϕ◦EM1 ϕ◦EM1 ϕ◦EM1 tm (cid:0)Λ(ϕ◦EM1 )ω ((x − (EM1 )ω(x))z)Λ(ϕ◦EM1 )ω (y(x − (EM1 )ω(x)))(cid:1)(ϕ◦EM1 )ω = (ϕ ◦ EM1 )ω((x − (EM1 )ω(x))∗y∗(x − (EM1 )ω(x))z) = (ϕ ◦ EM1 )ω((x − (EM1 )ω(x))∗ y∗(x − (EM1 )ω(x))y) = lim ℓ→∞ = lim ℓ→∞ = lim (ϕ ◦ EM1 )ω((x − (EM1 )ω(x))∗ y∗(x − (EM1 )ω(x))yℓ) (ϕ ◦ EM1 )ω((x − (EM1 )ω(x))∗y∗(x − (EM1 )ω(x))zℓ) ℓ→∞(cid:0)Λ(ϕ◦EM1 )ω ((x − (EM1 )ω(x))zℓΛ(ϕ◦EM1 )ω (y(x − (EM1 )ω(x)))(cid:1)(ϕ◦EM1 )ω = 0. Consequently we get ky(x − (EM1 )ω(x))k(ϕ◦EM1 )ω ≤ kyx − xzk(ϕ◦EM1 )ω . We have, by (3.9), ky(x − (EM1 )ω(x))k2 (ϕ◦EM1 )ω (x(m)))m(cid:3)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:2)(yP1Λϕ◦EM1 m→ωn(E2(y∗y)P1Λϕ◦EM1 (yP1Λϕ◦EM1 = lim m→ω = lim 2 Hω (x(m))yP1Λϕ◦EM1 (x(m)))ϕ◦EM1 (x(m)))ϕ◦EM1 + ((y∗y − E2(y∗y))P1Λϕ◦EM1 in X3 orthogonal to X1 (x(m)) P1Λϕ◦EM1 {z in X1 } (x(m)))ϕ◦EM1 (x(m)) )ϕ◦EM1o } (x(m))P1Λϕ◦EM1 {z (x(m)))m(cid:3)(cid:13)(cid:13)(cid:13) (x(m))P1Λϕ◦EM1 . Hω 2 = lim m→ω (E2(y∗y)P1Λϕ◦EM1 =(cid:13)(cid:13)(cid:13)(cid:2)(E2(y∗y)1/2P1Λϕ◦EM1 NON-TRACIAL AMALGAMATED FREE PRODUCTS As in showing (3.9) one has (cid:2)(E2(y∗y)1/2P1Λϕ◦EM1 and the proof is completed. (x(m)))m(cid:3) = iΛ(ϕ◦EM1 )ω (E2(y∗y)1/2(x − (EM1 )ω(x))), 13 (cid:3) 4. Some Consequences We first formulate that P is 'non-trivial relative to Q' for a given inclusion of von Neumann algebras P ⊇ Q, and then provide some technical facts. Definition 4.1. A (unital) inclusion P ⊇ Q of von Neumann algebras is said to be entirely non-trivial, if no non-zero direct summand of Q is a direct summand of P . Let P ⊇ Q be an inclusion of von Neumann algebras with a faithful normal conditional expectation EQ. If zP = Qz (as set) for some non-zero z ∈ Z(Q)p, then P z = Qz too by taking adjoints, and thus for x ∈ P one has zx = EQ(zx) = zEQ(x) = EQ(x)z = EQ(xz) = xz, implying z ∈ Z(P ). Hence Qz is a direct summand of P . Therefore, P ⊇ Q is entirely non- trivial if and only if P z 6= Qz or equivalently zP 6= Qz for any non-zero projection z ∈ Z(Q), where P z and zP denote the one-sided ideals of all xz and zx, respectively, with x ∈ P . The next simple lemma, especially (3) there, will frequently be used later. again entirely non-trivial. i ≤ e, and then P ei = P v∗ e =Pi∈I xei =Pi∈I EQ(x)ei = EQ(x)cQ e , and therefore P cQ e = QcQ e . i yi)) = 1, where s(x) denotes the support projection of x = x∗. Lemma 4.2. Let P ⊇ Q be an inclusion of von Neumann algebras with a faithful normal conditional expectation EQ : P → Q. (1) The following are equivalent: (i) P ⊇ Q is entirely non-trivial. (ii) P e 6= Qe or equivalently eP 6= eQ for any non-zero projection e ∈ Q. (2) If P ⊇ Q is entirely non-trivial and f ∈ Q a projection with cQ f = 1, then f P f ⊇ f Qf is (3) If P ⊇ Q is entirely non-trivial, then there is a family {yi}i∈I of elements in Ker(EQ) so that Pi∈I s(EQ(y∗ Proof. (1) By the discussion above (i) is equivalent to P z 6= Qz or equivalently zP 6= Qz for any non-zero z ∈ Z(Q)p. Thus (ii) ⇒ (i) is trivial, and it suffices to show (i) ⇒ (ii). Suppose that P e = Qe for some non-zero e ∈ Qp. By a standard exhaustion argument based on the comparison theorem we can choose an orthogonal family {ei}i∈I of projections in Q such that ei - e in Q for all i ∈ I and cQ i vi = ei and viv∗ i vi ⊆ P evi = Qevi ⊆ Qei, implying P ei = Qei ⊆ Q. For x ∈ P one has xcQ (2) By (1) it suffices to prove that eP f 6= eQf for any non-zero e ∈ Qp with e ≤ f . As in (1) one can find an orthogonal family {fi}i∈I of projections in Q such that fi - f in Q for all i ∈ I and Pi∈I fi = cQ f = 1. On contrary, suppose that eP f = eQf for some non-zero e ∈ Qp with e ≤ f . Then one has eP fi = eQfi in the same way as in (1). Hence, as in the above (1) one can justify, by using EQ, the following computation: eP =Pi∈I eP fi =Pi∈I eQfi = eQ, a contradiction to the entire non-triviality of P ⊇ Q thanks to (1). (3) Choose a maximal (with respect to set-inclusion) family {yi}i∈I of elements in Ker(EQ) so that {s(EQ(y∗ i yi)) 6= 1. Set e := 1 −Pi∈I s(EQ(y∗ i yi)) ∈ Qp \ {0}. Since P ⊇ Q is entirely non-trivial, one has P e 6= Qe by (1), and hence can choose x ∈ P with xe 6∈ Q. Hence xe − EQ(xe) 6= 0 and set y := xe − EQ(xe) ∈ Ker(EQ). Clearly, ye = y, and thus EQ(y∗y) = eEQ(y∗y)e, implying s(EQ(y∗y)) ≤ e = 1 −Pi∈I s(EQ(y∗ (cid:3) i yi))}i∈I is an orthogonal family of projections in Q. SupposePi∈I s(EQ(y∗ e = Pi∈I ei. Choose a partial isometry vi ∈ Q with v∗ i yi)), a contradiction to the maximality of {yi}i∈I . 14 Y. UEDA Let (M, E) = (M1, E1) ⋆N (M2, E2) be an amalgamated free product throughout the rest of this section. Theorem 4.3. Assume that there is a faithful normal state ϕ on M1 such that one can find a (possibly non-unital) dense (in any von Neumann algebra topology) ∗-subalgebra M1 of M1 with E1(M1) ⊆ M1 and a net vλ of unitaries in the centralizer (M1)ϕ in such a way that E1(y∗vλx) −→ 0 σ-strongly for all x, y ∈ M1. Assume also that M2 ⊇ N is entirely non-trivial. Then we have: t t (0) ((M1)ϕ)′ ∩ M = ((M1)ϕ)′ ∩ M1. (1) Z(M ) = Z(M1) ∩ Z(M2) ∩ Z(N ). (2) Let χ be an arbitrary faithful normal semifinite weight on N . Then, if a unitary u = Adu for some t ∈ R, then u must fall in N . In particular, = Adu = σχ◦E2 (3) M is semifinite if and only if there is a faithful normal semifinite trace TrN such that in M satisfies σχ◦E T (M ) = {t ∈ R σχ◦E1 both TrN ◦ E1 and TrN ◦ E2 are traces. for some u ∈ N u}. t ϕ◦EM1 (2) One has σ t Proof. (0) is nothing but what Proposition 3.3 says. (4) Z(fM ) = Z(fM1) ∩ Z(fM2) ∩ Z(eN ). (1) Let x ∈ Z(M ) be arbitrary, and then x must be in M1 by (0). For any y ∈ M ◦ 2 one has y(x − E1(x)) + yE1(x) = yx = xy = E1(x)y + (x − E1(x))y, and thus {yE1(x), E1(x)y}, y(x − E1(x)) and (x − E1(x))y are orthogonal with respect to E due to the free independence between M1 and M2. Thus y(x − E1(x)) = 0 so that (by looking at the E-value of the product of its adjoint and itself) we get (x − E1(x))∗E2(y∗y)(x − E1(x)) = 0. Therefore, E2(y∗y)(x − E1(x)) = 0 for all y ∈ M ◦ 2 . By taking its adjoint one can easily see that (x − E1(x))∗ ↾ran(E2(y∗y))≡ 0 so that (x − E1(x))∗s(E2(y∗y)) = 0 for all y ∈ M ◦ 2 . By Lemma 4.2 (3) 2 so that Pi∈I s(E2(y∗ one can find a family {yi}i∈I of elements in M ◦ i yi)) = 1, which implies x = E1(x) ∈ N . The desired assertion is now immediate. = Ad([Dϕ : Dχ◦ E1]t u) by Connes's Radon -- Nikodym cocycle theorem , we have [Dϕ : Dχ ◦ E1]t u ∈ M1 by 1 . For y ∈ M ◦ (y)(u − (y)u = uy = E1(u)y + (u − E1(u))y, and as in (1) we get 2 by (2.1). The same argument as in (1) again and [20, Corollary IX.4.20]. Since (M1)ϕ ⊆ Mϕ◦EM1 (0). In particular, u ∈ M1, since [Dϕ : Dχ ◦ E1]t ∈ M u E1(u)) + σχ◦E (u− E1(u))y = 0, since σχ◦E shows u = E1(u) ∈ N . The T-set computation is straightforward. (3) M is semifinite if and only if there is a 1-parameter unitary group u(t) in M so that σχ◦E = Adu(t), t ∈ R, for a fixed faithful normal state χ on N . See [20, Theorem VIII.3.14]. t Then u(t) ∈ N by (2). By Stone's theorem u(t) = H it with some positive non-singular, self- adjoint H affiliated with N . Since σχ (u(t)) = u(t), H must indeed be affiliated with the centralizer Nχ. Hence, by [20, Lemma VIII.2.8] we can construct a faithful normal semifinite weight χH−1 on N , and by the construction we observe that χH−1 ◦ E = (χ◦ E)H−1 . χH−1 ◦E (−)H it = id. Hence the Moreover, by [20, Lemma VIII.2.11] we have σ t χH−1 is a desired faithful normal semifinite trace on N . (4) By (0) together with the same argument as in [22, Corollary 4] we observe that ((M1)ϕ)′∩ (M ⋊σϕ◦EM1 R) = ((M1)ϕ)′∩(M1 ⋊σϕ R), where (M1)ϕ ⊂ M1 ⊆ M ֒→ M ⋊σϕ◦EM1 R canonically as in §2. It follows that (fM1)′ ∩fM = Z(fM1), where we need Connes's Radon -- Nikodym cocycle theorem together with [20, Theorem X.1.7]. Choose an arbitrary x ∈ Z(fM ). Then x must fall in Z(fM1) ⊆ fM1. For y ∈ M ◦ 2 one has y(x−eE(x))+yeE(x) = yx = xy = eE(x)y+(x−eE(x))y, and thus y(x − eE(x)) = 0 since fM1, fM2 are freely independent with respect to eE as remarked 2 ⊂ fM ◦ (y)E1(u) = σχ◦E t (y) = σχ◦E2 t 2 we have σχ◦E t t (u(t)) = σχ◦E t = H −itσχ◦E t t t (y) ∈ M ◦ NON-TRACIAL AMALGAMATED FREE PRODUCTS 15 2 as in (1). Therefore, using (cid:3) in §2. In particular, we get E2(y∗y)(x − eE(x)) = 0 for all y ∈ M ◦ Lemma 4.2 (3) as in (1) once again we can prove x = eE(x) ∈ eN . Hence we are done. Let us illustrate how the above theorem is useful by giving next two corollaries. The first corollary shows that Proposition 3.1 is useful to confirm the necessary hypothesis of the theorem. The second one does that the theorem is still applicable beyond the case where N is semifinite. Remark that the first one can be viewed as a simultaneous generalization of both [27, Theorem 3.4] and [21, §4]. Corollary 4.4. Assume that M1 is diffuse, N of type I and M2 ⊇ N entirely non-trivial. Let z ∈ Z(N ) be the unique projection so that N z is diffuse and N z⊥ atomic, and assume further that M1cM1 has no type I direct summand when z 6= 0 (i.e., this last assumption is fulfilled if M1 has no type I direct summand). Then all the assertions of Theorem 4.3 holds with a certain faithful normal state ϕ on M1. z t z Proof. Let us fix a faithful normal semifinite trace TrN on N . Write c := cM1 for simplicity. Clearly σTrN ◦E1 (c) = c for all t ∈ R, and thus Takesaki's criterion shows that there is a TrN ◦E1- preserving unique conditional expectation EL : M1 → L := N ∨ {c}′′ = N c ⊕ N c⊥ (⊇ N ). In particular, one observes that E1 ◦ EL = E1 holds. As in the proof of [27, Theorem 3.4] one can choose a faithful normal state ϕ on M1 such that (M1c)ϕ↾M1c has no type I direct summand and (M1c⊥)ϕ↾M1c⊥ is just only diffuse. Then it is clear that (M1c)ϕ↾M1c 6(cid:22)M1c N c with EL ↾M1c and TrN ◦ E1 ↾N c and that (M1c⊥)ϕ↾ M1c⊥ 6(cid:22)M1c⊥ N c⊥ with EL ↾M1c and TrN ◦ E1 ↾N c⊥, since N c⊥ = (N z⊥)c⊥ is a reduced von Neumann algebra of the atomic part N z⊥. Therefore, by the equivalent condition (i) in Proposition 3.1 there are two nets v(1) of unitaries in λ 2v(2) (M1c)ϕ↾M1c and (M1c⊥)ϕ↾M1c⊥ , respectively, so that EL(y∗ λ x2) −→ 0 σ-strongly for all x1, y1 ∈S{M1p p ∈ (N c)p; TrN ◦ E1(p) < +∞} and all x2, y2 ∈S{M1p p ∈ (N c⊥)p; TrN ◦ E1(p) < +∞}. Remark that EL = (EL ↾M1c)⊕ (EL ↾M1c⊥) in M1 = M1c⊕ M1c⊥ and that TrN (p) < +∞ implies both TrN ◦ E1(pc) < +∞ and TrN ◦ E1(pc⊥) < +∞ for p ∈ N p. Thus, letting vλ := v(1) M1c⊥ = (M1)ϕ one has, for all x, y ∈ S{M1p p ∈ N p; TrN (p) < +∞}, EL(y∗vλx) −→ 0 σ-strongly and hence E1(y∗vλx) = E1(EL(y∗vλx)) −→ 0 σ-strongly. Hence we can apply Theorem 4.3 with the above ϕ and M1 := S{pM1p p ∈ N p; TrN (p) < +∞}. Note here that M1 is indeed a ∗-algebra thanks to the Kaplansky formula [12, Theorem 6.1.7] and dense in any von Neumann algebra topology due to the semifiniteness of TrN . (cid:3) and v(2) λ λ x1) −→ 0 and EL(y∗ λ ∈ (M1c)ϕ↾M1c ⊕ (M1c⊥)ϕ↾ λ ⊕ v(2) 1v(1) Corollary 4.5. Assume that (M1, E1) is one of the following: (i) M1 = N ⋊α G and E1 is the canonical conditional expectation from M1 = N ⋊α G onto N , where α : G y N is an infinite discrete group action preserving a faithful normal state ψ on N . (ii) M1 = Q ¯⊗N and E1 = ψ ¯⊗idN , where Q is a diffuse von Neumann algebra with a faithful normal state ψ. Assume also that M2 ⊇ N is entirely non-trivial. Then all the assertions of Theorem 4.3 holds with ϕ = ψ ◦ E1 in (i) and with ϕ = ϕ0 ¯⊗χ in (ii), where Qϕ0 is diffuse (such a state ϕ0 certainly exists) and χ arbitrary. Proof. Case (i): Since ψ is invariant under the action α, the restriction (ψ ¯⊗idB(ℓ2(G))) ↾N ⋊αG gives a faithful normal conditional expectation from Eψ : M1 = N ⋊α G → L(G) = C1 ⋊ G, and it is plain to see that ψ ◦ E1 := τG ◦ Eψ with the canonical tracial state τG on L(G). Clearly L(G) = C1 ⋊ G sits inside (N ⋊α G)ψ◦E1 and is diffuse (see e.g. [6, Proposition 5.1]). With ϕ := ψ ◦ E1 = τG ◦ Eψ and M1 := span{xλg x ∈ N, g ∈ G} one can choose a net vλ from L(G) = C ⋊ G as in Theorem 4.3, since L(G) is diffuse and E1 ↾L(G)=C1⋊G= τG(−)1. 16 Y. UEDA Case (ii): As in the proof of [27, Theorem 2.4] one can choose a faithful normal state ϕ0 on Q in such a way that the centralizer Qϕ0 is diffuse. Set ϕ := ϕ0 ¯⊗χ with a faithful normal state χ on N and M1 := Q ⊙ N = span{x ⊗ y x ∈ Q, y ∈ N}. Then one can choose a net vλ from Qϕ0 (cid:3) ¯⊗C1 as in Theorem 4.3, since Qϕ0 is diffuse. The next lemma seems well-known, but we do give it for the reader's convenience as a reference for the discussions below. Lemma 4.6. Let (P, F ) = (P1, F1) ⋆Q (P2, F2) be an amalgamated free product. If a projec- tion f ∈ Q has cQ f = 1, then (f P f, F ↾f P f ) = (f P1f, F1 ↾f P1f ) ⋆f Qf (f P2f, F2 ↾f P2f ) holds canonically. viv∗ 1 , P ◦ i1 )(vi1 x2v∗ Proof. Clearly f P1f and f P2f are freely independent with respect to F ↾f P f , and hence it suffices to see that those generate f P f as von Neumann algebra. As in the proof of i vi = cQ f = 1 and 2 ) one has f xf = i2 )··· (vin−1 xnf ) σ-strongly, which falls in the σ-strong closure of 1 , P ◦ 2 ), (cid:3) Lemma 4.2 one can find partial isometries {vi}i∈I in Q such that Pi∈I v∗ i ≤ f for all i ∈ I. For any alternating word x = x1 ··· xn ∈ Λ◦(P ◦ Pi1,...,in−1∈I(f x1v∗ the linear span of Λ◦((f P1f )◦, (f P2f )◦)). Since P is the σ-strong closure of Q+spanΛ◦(P ◦ the assertion is immediate. Lemma 4.7. Let P ⊇ Q be an inclusion of σ-finite von Neumann algebras with a faithful normal conditional expectation EQ : P → Q, and assume that Q is commutative. (1) If P has no type I direct summand and a faithful normal semifinite trace TrP on P with TrP ◦ EQ = TrP , then there is a faithful normal state χ on Q so that for each n ∈ N with n ≥ 2 one can find a unitary un ∈ Pχ◦EQ in such a way that EQ(uk n) = 0 for all 1 ≤ k ≤ n − 1, i.e., EQ(uk1 n (2) If P is diffuse and Q is atomic, then there is a faithful normal state ϕ on P such that (a) the centralizer Pϕ contains Q, (b) there are two unitaries u, v ∈ Pϕ so that EQ(uk) = Eϕ Q(vk) = 0 as long as k 6= 0, i.e., EQ(uk1 ∗ uk2 ) = Eϕ Q denotes the unique ϕ-preserving conditional expectation from P onto Q whose existence follows from (a) and Takesaki's criterion. Q(vk1 ∗ vk2 ) = 0 for all k1 6= k2. Here Eϕ n ) = 0 for all 0 ≤ k1 6= k2 ≤ n − 1. ∗ uk2 (3) Let z ∈ Z(P ) be the central support projection of the type I direct summand of P . Assume that P is diffuse and Qz atomic. Then there is a faithful normal state ϕ on the continuous core eP of P such that (a) the centralizer (eP )ϕ contains eQ, where eQ = Q ⋊σχ R ֒→ eP = P ⋊σχ◦EQ R with a faithful (b) for each n ∈ N with n ≥ 2 one can find a unitary un ∈ (eP )ϕ in such a way that normal state or semifinite weight χ on Q, ∗ uk2 ∗ vk2 n (vk1 n n ) = Eϕ n ) = 0 for eQ denotes the unique (uk n) = Eϕ eQ n) = 0 for all 1 ≤ k ≤ n − 1, i.e., eEQ(uk1 eEQ(uk all 0 ≤ k1 6= k2 ≤ n − 1. Here eEQ = (EQ ¯⊗idB(L2(R))) ↾ eP , and Eϕ ϕ-preserving conditional expectation from eP onto eQ as in (2). eQ The same assertion also holds for P ⊇ Q with EQ themselves, if it is further assumed that P is semifinite and EQ preserves a faithful normal semifinite trace TrP on P . Proof. (1) By assumption TrP ↾Q is semifinite, and thus one can choose an orthogonal sequence {qm}m of projections in Q with TrP (qm) < +∞ and Pm∈N qm = 1. Consider the faithful normal state χ := Pm∈N m∈N Qqm = Q(cid:1) so that Pχ◦EQ Clearly the centralizer Pχ◦EQ contains P⊕ 2mTrP (qm) TrP ↾Qqm on Q. (Remark here that Q is commutative.) m∈N qmP qm (cid:0) ⊇ P⊕ 1 NON-TRACIAL AMALGAMATED FREE PRODUCTS 17 n)) = 0. n) = EQ(EA(uk n) = 0 so that EQ(uk m∈N Cqm. Clearly EQ factors as P Since P is diffuse, so are all qmP qm; hence by the proof of [27, Theorem 3.4] there are faith- must be of type II1. Choose a MASA A in Pχ◦EQ that contains Q. By [11, Corollary 3.16], for each n ∈ N with n ≥ 2 there are n orthogonal e0, . . . , en−1 ∈ Ap, all of which are equivalent in Pχ◦EQ , and Pn−1 i=0 ei = 1. Then one can construct a unitary un ∈ Pχ◦EQ such that une0 = e1un, une1 = e2un, . . . , unen−1 = e0un. Let EA : P → A be the χ ◦ EQ-preserving conditional expectation (whose existence follows from Takesaki's criterion), and clearly EQ ◦ EA = EQ. Then, for every 1 ≤ k ≤ n − 1 one has EA(uk (2) Write Q = P⊕ EQ′ ∩P−→ Q′ ∩ P Ψ−→ Q, where Q′ ∩ m∈N qmP qm and EQ′∩P (x) = Pm∈N qmxqm for x ∈ P . Moreover Ψ is of the form P = P⊕ Ψ(Pm∈N xm) =Pm∈N ψm(xm)qm for xm ∈ qmP qm with faithful normal states ψm on qmP qm. ful normal states ϕm on qmP qm with (qmP qm)ϕm diffuse for all m. Define Φ(Pm∈N xm) = Pm∈N ϕm(xm)qm for xm ∈ qmP qm, giving a faithful normal conditional expectation from Q′∩P onto Q. Set ϕ := χ◦ Φ◦ EQ′∩P , a faithful normal state on P , with a faithful normal state χ on Q. Then Q′ ∩ Pϕ =P⊕ m∈N(qmP qm)ϕm, a direct sum of diffuse von Neumann algebras. One can choose, for each m, unitaries um, vm ∈ (qmP qm)ϕm so that ϕm(uk m) = 0 as long as k 6= 0. (See the proof of [27, Theorem 3.7].) Then u :=Pm∈N um, v :=Pm∈N vm are unitaries in Q′ ∩ Pϕ, and moreover EQ(uk) = Ψ(uk) = 0 and Eϕ (3) Consider P = P z ⊕ P z⊥ ⊇ R := Q ∨ {z}′′ = Qz ⊕ Qz⊥ ⊇ Q. Let χ be an arbitrary faithful normal state on Q. As in the proof of Corollary 4.4 one can show that there is a unique faithful normal conditional expectation ER : P → R with EQ ◦ ER = EQ. Then we have Q(vk) = Φ(vk) = 0 as long as k 6= 0. m) = ψm(vk eER ^EQ↾R ⊇ eQ = Q ⋊σχ R, where eER = (ER ¯⊗idB(L2(R))) ↾ eP and ^EQ ↾R = ((EQ ↾R) ¯⊗idB(L2(R))) ↾ eR= eEQ ↾ eR. Since ER = (ER ↾P z) ⊕ (ER ↾P z⊥) in P = P z ⊕ P z⊥, we have, by [20, Theorem X.1.7 (ii)], ⊇ eR = R ⋊σχ◦(EQ↾R ) R eP = P ⋊σχ◦EQ R (cid:16)eP eER ⊇ eR(cid:17) ∼=(cid:16)fP z ^ER↾P z ⊇ fQz(cid:17) ⊕(cid:16)gP z⊥ ^ER↾P z⊥ ⊇ gQz⊥(cid:17), where the continuous cores and the conditional expectations in the right-hand side are defined z⊥,n) = ( ^ER ↾P z⊥)(uk m∈N Cem, and ER ↾P z factors as P z similarly as above. Since gP z⊥ has no type I direct summand by the assumption here and [20, Theorem XII.1.1] and since ^ER ↾P z⊥ preserves the canonical trace on gP z⊥ see e.g. [14, §4], we can apply (1) to the second (cid:16)gP z⊥ ⊇ gQz⊥(cid:17) with ^ER ↾P z⊥ directly, and get a faithful normal state ϕz⊥ on gP z⊥ with ϕz⊥ ◦ ( ^ER ↾P z⊥) = ϕz⊥ such that for each n ∈ N with n ≥ 2 one can find a unitary uz⊥,n ∈ (fP z)ϕz⊥ in such a way that eER(uk z⊥,n) = 0 for all 1 ≤ k ≤ n− 1. Write Qz =P⊕ E(Qz)′∩P z−→ (Qz)′ ∩ P z Ψ−→ Qz, where (Qz)′ ∩ P z = P⊕ m∈N em(P z)em and E(Qz)′∩P z(x) = Pm∈N emxem for x ∈ P z. More- over, Ψ is of the form Ψ(Pm∈N xm) = Pm∈N ψm(xm)em for xm ∈ em(P z)em with faithful normal states ψm on em(P z)em. By the assumption here P z is diffuse and of type I, and thus so are the em(P z)em; hence the centers of those must be diffuse, and so are all the (em(P z)em)ψm . In the same way as in (2), one can find a unitary uz ∈ ((Qz)′ ∩ P z)χz◦Ψ with 'any' faithful normal state χz on Qz in such a way that Ψ(uk z ) = 0 for all k 6= 0. Denote by λ(t) the generators of C ⋊ R in fP z = (P z) ⋊σχz ◦(ER↾P z ) R (←֓ (Qz) ⋊σχz R = fQz canoni- cally), and set ϕz := τ ◦ ( ^ER ↾P z), a faithful normal state on fP z, with a fixed faithful normal tracial state τ := χz ¯⊗τ0 on fQz = Qz ¯⊗λ(R)′′. Note that λ(t)uz = σχz ◦(ER↾P z) uzλ(t) for all t ∈ R. Thus, for any finite sum x = Pk xkλ(tk) ∈ fP z with xk ∈ P z we (uz)λ(t) = t 18 Y. UEDA ϕ(x) := 1 z ) = ER(uk z ) = ( ^ER ↾P z)(uk have ϕz(uzx) = Pk τ (Ψ(uzE(Qz)′∩P z(xk))λ(tk)) = Pk χz(Ψ(uzE(Qz)′∩P z(xk)))τ0(λ(tk)) = Pk χz(Ψ(E(Qz)′∩P z(xk)uz))τ0(λ(tk)) = Pk ϕz(xkuzλ(tk)) = ϕz(xuz). It follows that uz falls in (fP z)ϕz . Clearly eER(uk z ) = 0 for all k 6= 0. Set 2 (ϕz(xz) + ϕz⊥(xz⊥)) for x ∈ eP , and then ϕ becomes a faithful normal state on eP and satisfies ϕ ◦ eER = ϕ, implying the desired condition (a), since eR is commutative. For each n ∈ N with n ≥ 2 we define the unitary un := uz ⊕ uz⊥,n ∈ fP z ⊕ gP z⊥ = eP , and thus eER(uk z⊥,n) = 0 for all 1 ≤ k ≤ n − 1. Hence the desired condition (b) is immediate as in (1) from the fact that eEQ = eEQ ◦ eER and Eϕ eQ ◦ eER (the latter follows from ϕ ◦ eER = ϕ). The final assertion is shown in the exactly same way (but z ) ⊕ ( ^ER ↾P z⊥)(uk n) = ( ^ER ↾P z)(uk easier) as above. z ) = Ψ(uk (cid:3) = Eϕ eQ We will give two applications of Proposition 3.5. The latter is a straightforward generalization of both [27, Theorem 3.7] and [28, Proposition 3.1]. Remark that the former reproves the assertions (1), (4) in Corollary 4.4 without any use of the technologies provided in §§3.1 -- 3.2. Theorem 4.8. Assume that M1 diffuse, N of type I and M2 ⊇ N entirely non-trivial. Let z ∈ Z(N ) be the unique projection such that N z is diffuse and N z⊥ atomic, and assume further that (M1)cM1 has no type I direct summand when z 6= 0 (i.e., this last assumption is fulfilled if M1 has no type I direct summand). Then (fM )ω =(cid:0)fM(cid:1)′ ∩Z(eN )ω. In particular, fM and hence M itself are non-amenable. If M is additionally assumed to be semifinite, then Mω = M ′ ∩ M ω = M ′ ∩ Z(N )ω also holds. =(cid:0)fM(cid:1)′ ∩(cid:0)fM(cid:1)ω z After the completion of the main part of the present work we learned that Houdayer and Vaes have also independently been obtained a similar (but not same) result as above under different assumptions with different (and simpler) methods. See [8, Theorem 5.8]. More on this will be discussed at the end of this section. eN f = 1. Let us first prove: Proof. Note that (eN ⊇ N ) ∼= (N ¯⊗λ(R)′′ ⊇ N ¯⊗C1). Since N is of type I, one can choose an abelian f ∈ N p (⊂ eN p) with c For each x ∈ eN ω with representative (x(m))m one has f xf = [(f x(m)f )m], and for every m there is a unique z(m) ∈ Z(eN ) with f x(m)f = z(m)f . By c f = 1 the mapping x′ ∈ eN ′ 7→ x′f ∈ eN ′f gives a bijective normal ∗-homomorphism (thus k−k∞-preserving), and hence (z(m))m defines z ∈ Z(eN )ω. Consequently we get f xf = zf ∈ Z(eN )ωf . By Lemma 4.6 together with (2.5) we have the identification f = Z(eN )ωf. f(cid:0)eN(cid:1)ω (4.1) eN (cid:0) ]f M f, ^E ↾f Mf(cid:1) =(cid:0)^f M1f , ^E1 ↾f M1f(cid:1) ⋆]f N f (cid:0)^f M2f , ^E2 ↾f M2f(cid:1). Let c ∈ Z(M1) be the central support projection of the type I direct summand of M1. Then e = cf is that of f M1f too, and f N f e = Z(N )f e must be atomic (or 0 if e = 0) by the e ∼= Z(N )e = Z(N )f e is not assumption here. In fact, if this was not the case, then Z(N )cN atomic, and hence zcN z c ≥ zc ≥ ze 6= 0, a contradiction to that M1cM1 has no type I direct summand. Therefore, by Lemma 4.7 (3) we can apply Proposition 3.5 to (4.2) and thus any x ∈(cid:0) ]f M f(cid:1)′ and any y ∈(cid:0)^f M2f(cid:1)◦ e 6= 0, i.e., ze 6= 0, implying cM1 ∩(cid:0) ]f M f(cid:1)ω must satisfy that z ( ^E2 ↾f M2f )(y∗y)(x − ( ^Ef M1f )ω(x)) = 0, (4.2) (4.3) NON-TRACIAL AMALGAMATED FREE PRODUCTS 19 and (4.1) imply that , that = π−1 = (cid:0) ]f M f(cid:1)′ ∩(cid:0) ]f M f(cid:1)ω ∩(cid:0)]f N f(cid:1)ω . (x) ∈ (cid:0)^f M1f(cid:1)ω (cid:0) ]f M f(cid:1)′ . Consequently (cid:0) ]f M f(cid:1)′ and (cid:0)^f M2f(cid:1)ω ∩(cid:0) ]f M f(cid:1)ω =(cid:0) ]f M f(cid:1)′ where Ef M1f is the unique conditional expectation from f M f onto f M1f determined as (2.2). Note that ^f M2f ⊇ ]f N f with ^E2 ↾f M2f contains f M2f ⊇ f N f with E2 ↾f M2f canonically. Hence, by Lemma 4.2 (2), (3) one can find a family {yi}i∈I in (f M2f )◦ in such a way that Pi∈I s(E2(y∗ i yi)) = f (= 1f N f ). Therefore, it follows from (4.3) as in the proof of Theorem 4.3 that x = (cid:0) ^Ef M1f(cid:1)ω ∩(cid:0)^f M1f(cid:1)ω . In the same way as in the proof of Theorem 4.3, we see, by using the above {yi}i∈I again and the free independence between (cid:0)^f M1f(cid:1)ω Choose a faithful normal semifinite trace TrN on N , and fM ⊇ fMk (k = 1, 2) ⊇ eN are realized as fM = M ⋊σTrN ◦E R ⊇ fMk = Mk ⋊σTrN ◦Ek R ⊇ eN = N ⋊σTrN R. Since σTrN ◦E t ∈ R, ffM f ⊇ ffMkf ⊇ f eN f are naturally identified with ]f M f ⊇ ^f Mkf ⊇ ]f N f . Hence (4.4) f =(cid:0)fM(cid:1)′ f ∩ f eN ωf =(cid:0)fM(cid:1)′ f ∩ Z(eN )ωf. Let πf be the normal surjective ∗-homomorphism x ∈(cid:0)fM(cid:1)′ 7→ xf ∈(cid:0)fM ′∩(cid:0)fM(cid:1)ω(cid:1)f = ∩(cid:0)fM(cid:1)ω (cid:0)fM(cid:1)′ f ∩ (f(cid:0)fM(cid:1)ω ∩(cid:0)fM(cid:1)ω f = 1 too). By (4.5) we have (cid:0)fM(cid:1)′ f (cid:0)(cid:0)fM(cid:1)′ f ∩ Z(eN )ωf(cid:1). As in the proof of Lemma 4.2 one can choose partial isometries {vi}i∈I in eN so that Pi∈I v∗ i ≤ f for all i ∈ I. Then, if x = zf ∈ (cid:0)fM(cid:1)′ f ∩ Z(eN )ωf with z ∈ Z(eN )ω, then i2 )xvi2 = Pi1,i2∈I v∗ i2vi2 = Pi1,i2∈I v∗ we have yz = Pi1,i2∈I v∗ ∩Z(eN )ω. Hence(cid:0)fM(cid:1)′ i2 vi2 = zy for y ∈ fM , implying z ∈(cid:0)fM(cid:1)′ Pi1,i2∈I v∗ f ∩Z(eN )ωf = ∩(cid:0)fM(cid:1)ω ∩ Z(eN )ω(cid:1)f . Consequently (cid:0)fM(cid:1)′ (cid:0)(cid:0)fM(cid:1)′ f ∩ Z(eN )ωf(cid:1) = π−1 f (cid:0)(cid:0)fM(cid:1)′ f (cid:0)(cid:0)(cid:0)fM(cid:1)′ ∩ ∩(cid:0)fM(cid:1)ω ∩ Z(eN )ω. Since (cid:0)fM(cid:1)′ = (cid:0)fM(cid:1)′ Z(eN )ω(cid:1)f(cid:1) = (cid:0)fM(cid:1)′ ∩ Z(eN )ω is commutative, it must equal (fM )ω as observed in [23, (8) in page 360]. The final assertion is also shown in the exactly same way as above by using the final assertion in Lemma 4.7 (3), since there is a faithful normal semifinite trace TrN on N so that TrN ◦ Ek (k = 1, 2) are traces again thanks to Corollary 4.4 (3). (cid:3) f ) (c.f. [30, Lemma 4.1 (i)]), which is also injective due to c (cid:0)fM(cid:1)′ f ∩ f(cid:0)fM(cid:1)ω eN f = 1 (and hence t (f ) = f for all i1 (vi1 yv∗ = π−1 fM c and viv∗ i vi = c eN f = 1 i1 x(vi1 yv∗ i2 )vi2 = (4.4) (4.5) i1 vi1 yzv∗ i1 vi1 zyv∗ Remark 4.9. The same type argument as in Theorem 4.3 (3) works for constructing a faithful normal state χ on N with σχ◦E T = Id with T = −2π/ log λ, 0 < λ < 1, when M is known to be a factor of type IIIλ under the same set of assumptions as in Theorem 4.8. Hence the discrete core of such M can also be written as an amalgamated free product von Neumann algebra of the same form as the continuous core, and an analogous formula for its asymptotic centralizer holds. In particular, the discrete core of such a factor of type IIIλ is an ∞-amplification of a non-strongly stable type II1 factor. Further and more detailed discussions related to this aspect will be given elsewhere. Theorem 4.10. If M1 is diffuse, N of atomic type I and M2 ⊇ N entirely non-trivial, then the following hold true: (1) Mω = M ′ ∩ M ω = M ′ ∩Z(N ) (= Z(M )). Hence M does never have no type III0 direct summand (see [5, Theorem 2.12]), and becomes full in the sense of Connes [5] under the separability of preduals. (2) The Connes τ -invariant τ (M ) (see [5]) is determined under the separability of preduals as follows. Let χ be a faithful normal state on N . Then tm −→ 0 in τ (M ) as m → ∞ if and only if there is a unitary w ∈ N so that σχ◦E tm −→ Adw in Aut(M ) as m → ∞. 20 Y. UEDA fi = 1 and σχ Proof. (1) This is proved along the same line as in the proof of Theorem 4.8 by using only Lemma 4.7 (2) instead together with a well-known fact Z(N ) = Z(N )ω due to the assumption that it is atomic. (2) We can write N = P⊕ i∈I B(Hi). Looking at this structure with the given χ we can choose a collection {ei}i∈I of abelian projections in N with Pi∈I ei = 1 such that for each t (fi) = fi (t ∈ R). i ∈ I there is a larger abelian fi ∈ N p so that ei ≤ fi, cN Assume that tm −→ 0 in τ (M ) as m → ∞. Then there is a sequence (um)m of unitaries in M such that Adum ◦ σχ◦E tm −→ id in Aut(M ) as m → ∞. As observed in the proof [28, Proposition 3.1] the (um)m defines a unitary u ∈ M ω, and clearly ufi = fiu for all i ∈ I. Hence fiu defines a unitary in fiM ωfi = (fiM fi)ω, and we denote it by ui for simplicity. Since fiM1fi is still diffuse, looking at fiM1fi ⊇ fiN fi = Z(N )fi one can choose a faithful normal state ϕ on fiM1fi as in Lemma 4.7 (2). Set ϕ(x) := ϕ(fixfi) + χ ◦ E1(f ⊥ i ), x ∈ M1, which becomes a faithful normal positive linear functional on M1. Clearly fi ∈ (M1) ϕ and thus fi[Dχ ◦ E1 : D ϕ]t = [Dχ ◦ E1 ↾fiM1fi : Dϕ]t for all t ∈ R by the uniqueness part of Connes's Radon-Nikodym cocycle theorem. As observed in the proof of [28, Proposition 3.1] again the sequence vm := [Dχ ◦ E1 : D ϕ]tm defines a unitary v ∈ M ω 1 and also the sequence 1 fi = (fiM1fi)ω. Since ϕ ◦ EM1 ↾fiMfi = ϕ ◦ (EM1 ↾fiMfi fivm = vmfi does a unitary vi ∈ fiM ω (y))m(cid:3) = uvz = uiviz for ), we have yuivi = yuv = (cid:2)(yumvm)m(cid:3) = (cid:2)(umvmσ (y))m(cid:3) ∈ (fiM fi)ω = fiM ωfi in the identification y ∈ (fiM2fi)◦ with z = (cid:2)(σ (fiM fi, E ↾fiMfi ) = (fiM1fi, E1 ↾fiM1fi) ⋆fiN fi (fiM2fi, E2 ↾fiM2fi ) provided by Lemma 4.6. By Proposition 3.5 we get (E2 ↾fiM2fi )(y∗y)(uivi − (EM1 ↾fiMfi )ω(uivi)) = 0 for y ∈ (fiM2fi)◦. By using Lemma 4.2 (2), (3) twice as in the proof of Theorem 4.8 we can prove firstly that uivi ∈ (fiM1fi)ω = fiM ω 1 fi), and finally that ui ∈ fiN ωfi = Z(N )ωfi = Z(N )fi. Therefore, u = Pi∈I eiu = Pi∈I eifiu = Pi∈I eiui ∈ N . Letting w := u∗ ∈ N u we have Adw∗ ◦ σχ◦E 1 fi (since vi ∈ fiM ω tm −→ id in Aut(M ) as m → ∞. 1 fi, secondly that ui ∈ fiM ω ϕ◦(EM1↾fiM fi ) tm ϕ◦(EM1↾fiM fi ) tm i xf ⊥ (cid:3) The next proposition shows that Proposition 3.5 is still useful beyond the case where N is of type I or even semifinite. The proof goes along the same line as that of Theorem 4.8 but is easier than it. Hence the proof is left to the reader. Proposition 4.11. Assume that there is a faithful normal state ϕ on M1 satisfying the following conditions: (a) σϕ (b) For every n ∈ N with n ≥ 2 there are unitaries uk = u(n) t (N ) = N for all t ∈ R. k ∈ (M1)ϕ, 0 ≤ k ≤ n − 1, such that E1(u∗ vk2 ) = 0 for all 0 ≤ k1 6= k2 ≤ n − 1, where Eϕ N denotes the unique ϕ-preserving conditional expectation from M1 onto N , whose existence follows from (a) and Takesaki's criterion. k , vk = v(n) uk2) = Eϕ N (v∗ k1 k1 Assume also that M2 ⊇ N is entirely non-trivial. Then M ′ ∩ M ω = M ′ ∩ N ω holds. Moreover, if it is further assumed that N is finite, then Mω = M ′ ∩ M ω = M ′ ∩ Nω. It is easy to confirm that the (M1, E1) in Corollary 4.5 satisfies the assumption of Proposition 4.11. Thus M ′ ∩ M ω = M ′ ∩ N ω holds under the set of assumptions in Corollary 4.5. Assume that M1 is a von Neumann algebra with separable predual and that N is a Cartan subalgebra in M1. It was proved in [21, Lemma 4.2] that if M1 is further assumed to be a non-type I factor, then there are a faithful normal state ϕ on M1 with ϕ◦ E1 = ϕ and a unitary u ∈ (M1)ϕ such that E1(uk) = 0 as long as k 6= 0. The same assertion can indeed be proved even when M1 is further assumed only to have no type I direct summand (i.e., without being a factor). The proof is similar to [21, Lemma 4.2] but tedious based on disintegration. Hence such (M1, E1) satisfies the assumption of Proposition 4.11. NON-TRACIAL AMALGAMATED FREE PRODUCTS 21 Remark 4.12. Almost all the results obtained above have appropriate 'HNN variants' thanks to tricks given in [26]. Here it should be emphasized that our results so far essentially need assumptions for only one free component. The notion of HNN extensions of von Neumann alge- bras as well as their basic properties including their modular theoretic aspects were established in [25]. In closing of this section we discuss one of Houdayer and Vaes's results [8, Theorem 5.8]. This part of the present paper is added after receiving a draft of [8] in order to point out only one consequence obtained from this and that papers without any new idea. Therefore, some facts provided in [8] are necessary below. The original aim of the present work is to provide amalgamated free product counterparts of the results in [27, §3]. One issue to do so is how to formulate a suitable assumption saying that M1 is 'diffuse relative to N ' which corresponds to that M1 is diffuse when N = C1. The requirement for M1 ⊇ N in Theorem 4.3 seems to be one strong form of them without any restriction on N , but it seems not so easy to check it in general. Thus we propose the requirement for M1 ⊇ N in Corollary 4.4 and Theorem 4.8 as such a candidate in the special case when N is of type I. However a more sophisticated one in the special case seems to be that M1 ⊇ N has no trivial corner, which is proposed in [8, §5] by a different motivation. In fact, Houdayer and Vaes [8, Theorem 5.8] give a factoriality and non-amenability result under the set of assumptions that both Mk ⊇ N , k = 1, 2, have no trivial corner and that N is of type I, and establish their primeness result under the same set of assumptions. Here an inclusion P ⊇ Q of von Neumann algebras is said to have no trivial corner if pP p 6= Qp for any non-zero projection p ∈ Q′ ∩ P . Any exact general relationship between theirs and ours is not immediately clear. However the proof of Theorem 4.8 and general properties on inclusions without trivial corner provided in [8, §§5.1] altogether immediately give an improvement of [8, Theorem 5.8], though it is not immediately clear whether the primeness result in [8, Theorem E] holds or not under the new set of assumptions. Theorem 4.13. If M1 ⊇ N has no trivial corner, N is of type I and M2 ⊇ N entirely non- trivial, then the following hold true: (1) Z(M ) = Z(M1) ∩ Z(M2) ∩ Z(N ). (2) Z(fM ) = Z(fM1) ∩ Z(fM2) ∩ Z(eN ). (3) (fM )ω =(cid:0)fM(cid:1)′ =(cid:0)fM(cid:1)′ ∩(cid:0)fM(cid:1)ω In particular, (3) explains that M does never become amenable. ∩ Z(eN )ω. Proof. It is trivial that (3) ⇒ (2) ⇒ (1), see e.g. the proof of [25, Theorem 5.2] for (3) ⇒ (2) and [20, Theorem X.II.1.1] for (2) ⇒ (1). Thus it suffices to prove only (3). The line of the proof below is exactly identical to that of Theorem 4.8, and thus we keep the notations there. In fact, only one modification is sufficient. By [8, Lemma 5.2, Proposition 5.5] the inclusion ^f M1f ⊇ ]f N f also has no trivial corner. Then it suffices to prove the exactly same assertion as in Lemma 4.7 (1) with replacing the assumption that P has no type I direct summand by that P ⊇ Q has no trivial corner. In fact, by using this new assertion instead of Lemma 4.7 (3) one gets the same equation (4.3) and the rest of the proof there works well. Let P ⊇ Q be an inclusion of von Neumann algebras without trivial corner. Assume that Q is commutative, P has a faithful normal semifinite trace TrP and there is a faithful normal conditional expectation EQ : P → Q satisfying TrP ◦ EQ = TrP . As in the proof of Lemma 4.7 (1) we choose the qm's and χ. Then we apply [8, Lemma 5.4 (3)] (note that it holds without assuming the separability of preduals, see Lemma 4.14 below) with q = p := qm and get a m) = 0 as long as k 6= 0. Letting u :=Pm∈N um we unitary um ∈ qmP qm satisfying that EQ(uk have u ∈ Pχ◦EQ and EQ(uk) = 0 as long as k 6= 0. Hence we are done. (cid:3) 22 Y. UEDA As remarked in [8, Lemma 5.3] the next lemma immediately follows from Rohlin's general theorem on Lebesgue spaces under the separability of preduals. Thus only the advantage of the proof below is no use of disintegration; hence the separability of preduals is not necessary in [8, Lemma 5.4]. Although it is a rather minor point, we do give it for the sake of completeness. Lemma 4.14. Let B ⊇ A be (unital) inclusion of commutative σ-finite von Neumann algebras with a faithful normal conditional expectation EA : B → A. If Bf 6= Af for any nonzero projection f ∈ B, then there is a unitary u ∈ B such that EA(uk) = 0 as long as k 6= 0. Proof. Choose non-zero f ∈ Bp. Since Bf 6= Af , there is x ∈ B such that x 6∈ Af and 0 ≤ x ≤ f . Since EA(x) ≤ EA(f ), one can choose a positive contraction c ∈ A so that cEA(f ) = EA(x) (since A is commutative). Letting y := x − cf ∈ Bf we have y = y∗ 6= 0 (due to x 6∈ Af ) and EA(y) = 0. Therefore, an idea given in the proof of [3, Lemma 2.1] enables us to construct projections e(ε1,...,εn) ∈ B, n ∈ N, εk ∈ {1, 2}, in such a way that e(ε1,...,εn) = e(ε1,...,εn,1) + e(ε1,...,εn,2) and EA(e(ε1,...,εn)) = 1 2n 1. The proof is done by induction. Assume that we have chosen up to n-th stage. Set Λe := {x = x∗ ∈ Bekxk∞ ≤ 1, EA(x) = 0} with e := e(ε1,...,εn). It is a σ-weakly compact convex subset, and thus has sufficiently many extremal points due to the Krein -- Milman theorem. Let a ∈ Λe be an extremal point. Then it suffices to prove a = 2e0 − e for some e0 ∈ Bp with e0 ≤ e, since it clearly implies that EA(e0) = 1 2 EA(e). On contrary, suppose that it is not the case. By the spectral decomposition of a one can find δ > 0 and non-zero f ∈ Bp in such a way that f ≤ e and −(1 − δ)f ≤ af ≤ (1 − δ)f . By what we have shown above, there is a non-zero y = y∗ ∈ Bf such that −δf ≤ y ≤ δf and EA(y) = 0, and hence a + y, a − y ∈ Λe and a = 1 2 (a − y), a contradiction. Thus e(ε1,...,εn,1) := e0 and e(ε1,...,εn,2) := e − e0 become desired ones in (n + 1)-th stage. Hence we have proved the claim. Let (C, ω) be the von Neumann algebraic infinite tensor product of C ⊕ C with equal weights {1/2, 1/2}. Once passing GNS representations one can construct an injective normal ∗-homomorphism from C ¯⊗A into B which intertwines ω ¯⊗idA and EA. Hence the desired assertion follows, since (C, ω) ∼= (L(Z), τZ) thanks to [19, Theorem III.1.22]. (cid:3) 2 (a + y) + 1 The entire non-triviality of an inclusion P ⊇ Q of von Neumann algebras is nothing but just the non-triviality of P when Q = C1, and hence Theorem 4.13 is no longer true under assuming only that M1 ⊇ N is entirely non-trivial instead. In fact, the plain free product of two 2-dimensional algebras with suitable states provides a counter example, see [27] for suitable references therein. Finally we conjecture that Corollary 4.4, especially a strong kind of irreducibility ((M1)ϕ)′ ∩ M ⊆ M1 for some faithful normal state ϕ, should also hold under the same set of assumptions of Theorem 4.13. This is rather technical, but such a property may have some potential in further analysis. We will consider it in future work beyond the case where Z(M ) = Z(M1) ∩ Z(M2) ∩ Z(N ) need not hold. Acknowledgment I thank Professors Cyril Houdayer and Stefaan Vaes for several fruitful conversations in Dec. 2011 and in Jan. 2012 and also for sending us a draft of [8] prior to putting it on the ArXiv. References [1] J. Asher, A Kurosh-type theorem for type III factors, Proc. Amer. Math. Soc., 137 (2009), 4109 -- 4116. [2] N.P. Brown and N. Ozawa, C∗-algebras and finite-dimensional approximations. Graduate Studies in Math- ematics, 88, American Mathematical Society, Providence, RI, 2008. [3] J. Cameron, J. Fang and K. Mukherjee, Mixing subalgebras of finite von Neumann algebras, Preprint, arXiv:1001.0169. NON-TRACIAL AMALGAMATED FREE PRODUCTS 23 [4] I. Chifan and C. Houdayer, Bass -- Serre rigidity results in von Neumann algebras, Duke Math. J., 153 (2010), 23 -- 54. [5] A. Connes, Almost periodic states and factors of type III1, J. Funct. Anal., 16 (1974), 415 -- 445. [6] K. Dykema, Free products of hyperfinite von Neumann algebras and free dimension, Duke Math. J., 69 (1993), 97 -- 119. [7] C. Houdayer, Construction of type II1 factors with prescribed countable fundamental group, J. reine angew. Math. 634 (2009), 169 -- 207. [8] C. Houdayer and S. Vaes, Type III factors with unique Cartan decomposition, Preprint, 2012. [9] A. Ioana, J. Peterson and S. Popa, Amalgamated free products of weakly rigid factors and calculation of their symmetry groups, Acta Math., 200 (2008), 85 -- 153. [10] M. Izumi, R. Longo and S. Popa, A Galois correspondence for compact groups of automorphisms of von Neumann algebras with a generalization to Kac algebras, J. Funct. Anal., 155 (1998), 25 -- 63. [11] R.V. Kadison, Diagonalizing matrices, Amer. J. Math., 106 (1984), 1451 -- 1468. [12] R.V. Kadison and J. Ringrose, Fundamentals of the theory of operator algebras, Vol. II, Advanced theory, Graduate Studies in Mathematics, 16, American Mathematical Society, Providence, RI, 1997. [13] H. Kosaki, Extension of Jones' theory on index to arbitrary factors, J. Funct. Anal., 66 (1986), 123 -- 140. [14] R. Longo, Index of subfactors and statistics of quantum fields, I, Commun. Math. Phys., 126 (1989), 217 -- 247. [15] S. Popa, Markov traces on universal Jones algebras and subfactors of finite index, Invent. math., 111 (1993), 375 -- 405. [16] S. Popa, On a class of type II1 factors with Betti numbers invariants, Ann. of Math. (2), 163 (2006), 809 -- 899. [17] S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I, Invent. math., 165 (2006), 369 -- 408. [18] S. Popa, On Ozawa's property for free group factors, Int. Math. Res. Not., IMRN 2007, Art. ID rnm036. [19] M. Takesaki, Theory of Operator Algebras, I, Encyclopedia of Mathematical Sciences, 124, Operator Alge- bras and Non-commutative Geometry, 5, Springer, Berlin, 2002. [20] M. Takesaki, Theory of Operator Algebras, II, Encyclopedia of Mathematical Sciences, 125, Operator Algebras and Non-commutative Geometry, 6, Springer, Berlin, 2003. [21] Y. Ueda, Amalgamated free product over Cartan subalgebra, Pacific J. Math., 191 (1999), 359 -- 392. [22] Y. Ueda, Remarks on free products with respect to non-tracial states, Math. Scand., 88 (2001), 111 -- 125. [23] Y. Ueda, Fullness, Connes' χ-groups, and ultra-products of amalgamated free products over Cartan subal- gebras, Trans. Amer. Math. Soc., 355 (2003), 349 -- 371. [24] Y. Ueda, Amalgamated free product over Cartan subalgebra, II: Supplementary Results & Examples, Advanced Studies in Pure Mathematics, 38 (2004), 239-265. [25] Y. Ueda, HNN extensions of von Neumann algebras, J. Funct. Anal., 225 (2005), 383 -- 426. [26] Y. Ueda, Remarks on HNN extensions in operator algebras, Illinois J. Math., 52 (2008), 705 -- 725. [27] Y. Ueda, Factoriality, type classification and fullness for free product von Neumann algebras, Adv. Math., 228 (2011), 2647-2671. [28] Y. Ueda, On type III1 factors arising as free products, Math. Res. Lett., 18 (2011), 909 -- 920. [29] S. Vaes, Rigidity results for Bernoulli actions and their von Neumann algebras (after Sorin Popa), S´eminaire Bourbaki, Vol. 2005/2006, Ast´erisque, 311 (2007), Exp. No. 961, viii, 237 -- 294. [30] B. J. Voeden, Normalcy in von Neumann algebras, Proc. London Math. Soc. (3), 27 (1973), 88 -- 100. [31] D. Voiculescu, Symmetries of some reduced free product C ∗-algebras, in Operator algberas and their con- nections with topology and ergodic theory, Lect. Notes in Math., 1132 (1985), 566 -- 588. Graduate School of Mathematics, Kyushu University, Fukuoka, 810-8560, Japan E-mail address: [email protected]
1701.01338	1	1701	2017-01-05T15:01:59	Functors induced by Cauchy extension of C*-algebras	[ "math.OA", "math.FA" ]	In this paper we give three functors $\mathfrak{P}$, $[\cdot]_K$ and $\mathfrak{F}$ on the category of C$^\ast$-algebras. The functor $\mathfrak{P}$ assigns to each C$^\ast$-algebra $\mathcal{A}$ a pre-C$^\ast$-algebra $\mathfrak{P}(\mathcal{A})$ with completion $[\mathcal{A}]_K$. The functor $[\cdot]_K$ assigns to each C$^\ast$-algebra $\mathcal{A}$ the Cauchy extension $[\mathcal{A}]_K$ of $\mathcal{A}$ by a non-unital C$^\ast$-algebra $\mathfrak{F}(\mathcal{A})$. Some properties of these functors are also given. In particular, we show that the functors $[\cdot]_K$ and $\mathfrak{F}$ are exact and the functor $\mathfrak{P}$ is normal exact.	math.OA	math	Functors induced by Cauchy extension of C∗-algebras Kourosh Nourouzi1 ∗, Ali Reza2 1,2 Faculty of Mathematics, K. N. Toosi University of Technology, P.O. Box 16315-1618, Tehran, Iran. October 14, 2018 Abstract In this paper we give three functors P, [·]K and F on the category of C∗- algebras. The functor P assigns to each C∗-algebra A a pre-C∗- algebra P(A) with completion [A]K . The functor [·]K assigns to each C∗-algebra A the Cauchy extension [A]K of A by a non-unital C∗-algebra F(A). Some properties of these functors are also given. In particular, we show that the functors [·]K and F are exact and the functor P is normal exact. Keywords: Pre-C∗- algebras; Extensions of C∗- algebras; Exact functors; Cauchy extension. 1 Introduction Given a complex C∗-algebra A, the algebra A[[Z]] consists of all sequences (an)∞ n=0 in A with pointwise linear operations and Cauchy product ((an)∞ n=0)((bn)∞ n=0) = (cn)∞ n=0, where each cn = Pn the formal power series in one variable of the form P∞ k=0 akbn−k. It is natural to think of elements of A[[Z]] as n=0 anZ n with product ( ∞Xn=0 ∞Xn=0 anZ n)( bnZ n) = ∞Xn=0 cnZ n, ∗Corresponding author, e-mail: [email protected] 2010 Mathematics Subject Classification. 46L05, 46M15 1 where cn's are as above. One may consider the complex subalgebra A[Z] = { ∞Xn=0 anZ n : ∞Xn=0 kank < ∞}, of A[[Z]]. It is of interest to find a C∗-algebra via A[Z] to be an extension of A. Recall that an extension B of C by A is a short exact sequence 0 −→ A f −→ B g −→ C −→ 0. (1) of C∗-algebras (see, e.g., [1, 5, 6, 10]). For any subset Kof [−1, 1] such that 0 is a limit point of K, we will define a pre-C∗-norm on A[Z]. The completion of A[Z], denoted by [A]K, is an extension of A (Proposition 7 (iii)) which will be called the Cauchy extension of A. The outline of this work is as follows. In Section 2 we introduce pre-C∗- In Proposition 5, it is shown that A[Z] is not a C∗-algebra. algebra A[Z]. Proposition 7 shows that the completion [A]K of pre-C∗-algebra A[Z] is an extension of A. We also introduce the functors P, [·]K and F on the category of C∗-algebras. The functor P assigns to each C∗-algebra A a pre-C∗-algebra P(A) = A[Z]. The functor [·]K assigns to each C∗-algebra A an extension [A]K of A by a non-unital C∗-algebra F(A), where the C∗-algebra F(A) is the completion of the ideal A1 = { ∞Xn=0 anZ n ∈ A[Z] : a0 = 0} of A[Z]. Some properties of functors P, [·]K and F are listed in Proposition 8. In Section 3 we show that the functors [·]K and F are exact. In Section 4, using the notion of normal exact sequence of the normed spaces introduced by Yang [16], we prove that the functor P is normal exact. More precisely, for any short exact sequence of C∗-algebra (1) the corresponding short exact sequence 0 −→ A[Z] f −→ B[Z] g −→ C[Z] −→ 0 is a normal exact sequence of pre-C∗-algebras. That is, B(Z)/ ker g −→ C[Z] is an isometry. Among other results we also show that for any closed ideal I of a C∗-algebra A, the pre-C∗-algebra I[Z] is a closed ideal of A[Z] (Proposition 8 (iii)) and the quotient A[Z]/I[Z] is a pre-C∗-algebra (Theorem 3) which is isometric ∗-isomorphic to (A/I)[Z] (Theorem 4). Finally in Section 5, we show that the Cauchy extension [A]K of a C∗-algebra A can be considered as a C∗- subalgebra of Cb(K, A), the C∗-algebra of all bounded continuous functions from K to A (Theorem 5 (i)). In particular, if K is compact, then [A]K is ∗-isomorphic to C(K, A). We also give some other results in Theorem 5. A minimax type result is given in Corollary 5. 2 2 Cauchy extension of C∗- algebras Let A be a complex Banach algebra and A[[Z]] be the complex algebra consisting of all formal power series in A. If A has a unit, then an element F = F (Z) = n=0 anZ n ∈ A[[Z]] is invertible if and only if a0 is an invertible element in A. In particular, 1 + Z 2 is invertible in A[[Z]] and we have P∞ (1 + Z 2)( ∞Xn=0 The subalgebra (−1)nZ 2n) = ( (−1)nZ 2n)(1 + Z 2) = 1. ∞Xn=0 A[Z] = { ∞Xn=0 anZ n ∈ A[[Z]] : ∞Xn=0 kank < ∞} can be equipped with a norm as kF k = ∞Xn=0 kank, (2) (3) for all F (Z) =P∞ n=0 anZ n ∈ A[Z]. Proposition 1. Let A be a Banach algebra. Then A[Z] with the norm given in (3) is a Banach algebra. n=0 aknZ n) be a Proof. To show that A[Z] is a Banach algebra, let (Fk) = (P∞ sequence in A[Z] such that P∞ ∞Xn=0 k=0 kFkk < ∞. Then kaknk < ∞. kaknk = ∞Xk=0 ∞Xn=0 k=0 akn and F =P∞ kPN ∞Xk=0 Let cn =P∞ There exists a positive integer N such thatP∞ n=0(P∞ n=0 kP∞ n=0P∞ k=N +1P∞ k=0 Fk − F k = kP∞ = P∞ ≤ P∞ = P∞ k=N +1 akn)Z nk k=N +1 aknk k=N +1 kaknk n=0 kaknk < ε. n=0 cnZ n. Then F ∈ A[Z]. Let ε > 0 be given. n=0 kaknk < ε. We have k=N +1P∞ This completes the proof. Proposition 2. Let A be a Banach algebra. If F (Z) = P∞ then PN n=0 anZ n → F (Z) as N → ∞, n=0 anZ n ∈ A[Z], 3 Proof. Since kF (Z) − NXn=0 anZ nk = k ∞Xn=N +1 anZ nk = ∞Xn=N +1 kank, we get the desired limit. gent series in A[Z]. Now one can consider any element F (Z) =P∞ P∞ lution and the norm given in (3) is a ∗-Banach algebra. If A is a C∗-algebra, we can define an involution ∗ in A[Z] by F ∗(Z) = n=0 a∗ nZ n for any F (Z) ∈ A[Z]. In this case, A[Z] equipped with this invo- n=0 anZ n ∈ A[Z] as a conver- Proposition 3. Let A be a C∗-algebra. There is no norm on involutive algebra (A[Z], ∗) which makes it a C∗-algebra. In particular, (A[Z], ∗) equipped with the norm given in (3) is not a C∗-algebra. Proof. We suppose on the contrary that there exists a norm k · k such that (A[Z], ∗, k · k) is a C∗-algebra. Suppose that A is unital. By (2) the element 1 + Z 2 is not invertible in A[Z]. This implies that −1 ∈ σ(Z 2) which is a contradiction. Now let A be non-unital and a ∈ A be self-adjoint with kak > 1. Applying (2) for aZ we get that 1 + a2Z 2 is not invertible in (A ⊕ C)(Z). That is, −1 ∈ σ(a2Z 2) which is again a contradiction. then For a C∗-algebra (A, k · k) if F (Z) = P∞ ∞Xn=0 kantnk ≤ ∞Xn=0 kank < ∞. n=0 anZ n ∈ A[Z] and −1 ≤ t ≤ 1 Hence F (t) =P∞ n=0 antn is norm-convergent in A. For any F (Z), G(Z) ∈ A[Z] and λ ∈ C, t ∈ [−1, 1] we have (λF (Z))(t) = λF (t), (F (Z) + G(Z))(t) = F (t) + G(t), (F (Z)G(Z))(t) = F (t)G(t). (4) (5) (6) Note that the equalities (4) and (5) are clear and the proof of (6) is similar to that of complex case (see [15, p. 74]). The following proposition has a straightforward proof which is omitted here. Proposition 4. Suppose that K is a subset of [−1, 1] such that 0 is a limit point of K and (an)∞ n=0 is a sequence in C∗-algebra A. If (i) F (Z) =P∞ n=0 anZ n such that P∞ (ii) F (t) = 0 for any t ∈ K, n=0 kank < ∞; then an = 0 for all n. 4 Hereafter, throughout the paper K will denote a subset of [−1, 1] such that 0 is a limit point of it. Proposition 5. The following statements hold: (i) The functional k · kK defined by kF kK = sup t∈K k ∞Xn=0 antnk, for all F = F (Z) =P∞ n=0 anZ n ∈ A[Z], is a norm; (ii) (A[Z], ∗, k · kK) is a pre-C∗-algebra but not a C∗-algebra; (iii) kF kK ≤ kF k for all F ∈ A[Z]; (iv) If F (Z) =P∞ n=0 anZ n, then PN n=0 anZ n → F (Z) as N → ∞ in k · kK. Proof. (i) From (4), (5), (6) and Proposition 4 it is easily seen that k · kK is a norm. (ii) By the definition of k · kK we have the identity kF ∗F kK = kF k2 K. Therefore (A[Z], ∗, k · kK) is a pre-C∗-algebra which by Proposition 3 is not a C∗-algebra. (iii) By the definition of k · kK is clear. (iv) The proof follows from Proposition 2 and Part (ii). We will call the completion [A]K of pre-C∗-algebra (A, ∗, k · kK) the K - It is clear that [A]K is a C∗- Cauchy or simply the Cauchy extension of A. algebra. Proposition 6. Let A be a C∗-algebra. The following hold: (i) If I is an ideal of A[Z], then the completion I of (I, k · kK) is a closed ideal of [A]K; (ii) If I is a closed ideal of A, then [I]K is a closed ideal of [A]K. Proof. (i) Let I be an ideal of A[Z]. Then the completion I of (I, k · kK) is a closed ideal of [A]K. Choose any element F ∈ I and G ∈ [A]K. Let (Fn) and (Gk) be two sequences in I and A[Z] respectively converging to F ∈ I and G ∈ [A]K . For any k, n ≥ 1 we have FnGk, GkFn ∈ I. This implies that F GK, GkF ∈ I, for all k ≥ 1 and so F G, GF ∈ I. That is I is a closed ideal of [A]K. (ii) Consider F ∈ I[Z] and G ∈ A[Z]. It is clear that F G, GF ∈ I[Z], i.e., I[Z] is an ideal of A[Z]. Now Part (i) implies that \(I[Z]) = [I]K is a closed ideal of [A]K. 5 For a C∗-algebra A define A0 = {F (Z) = ∞Xn=0 anZ n ∈ A[Z] : an = 0 f or n > 0}, A1 = {F (Z) = ∞Xn=0 anZ n ∈ A[Z] : a0 = 0}. Denote the completion of A1 by A1. It is clear that A1 is an ideal of A[Z] and by Proposition 6, A1 is a closed ideal of [A]K . Hence if A 6= 0, then [A]K has a proper closed ideal A1. Consequently no simple C∗-algebra is a Cauchy extension of some C∗-algebra. It is worth mentioning that there is no ideal I of A such that I[Z] = A1. Since A0 is naturally ∗-isomorphic to A we always use A instead of A0 as a subalgebra of A[Z]. Suppose that A, B, E are C∗-algebras such that B is an ideal of E. It is said to be E an extension of A by B if there is a short exact sequence 0 −→ B i−→ E p −→ A −→ 0, where i(B) = ker p and i, p are injective and surjective ∗-homomorphisms re- spectively (see, e.g.,[1]). Proposition 7. Let A be a C∗-algebra. The following statements hold: (i) Every element F of [A]K has a unique representation F = a + G, where a ∈ A and G ∈ A1; (ii) kakK = kak ≤ ka + GkK, for all a ∈ A and G ∈ A1; (iii) [A]K is an extension of A by A1; (iv) A1 is not unital as a C∗-subalgebra of [A]K . Proof. (i) Let (Fk) be a Cauchy sequence in (A[Z], k · kK), where Fk = ∞Xn=0 aknZ n ∈ A[Z]. Let ε > 0 be given. Then kFk − Fk′ kK < ε for sufficiently large k, k′ . Suppose that (tm) is a sequence in K such that tm −→ 0 as m −→ ∞. By the definition of k · kK we have kak0 − ak′0k = lim m→∞ k ∞Xn=0 (akn − ak′n)tn mk ≤ sup t∈K k ∞Xn=0 (akn − ak′n)tnk = kFk − Fk′ kK < ε, 6 for sufficiently large k, k′. Furthermore k sup t∈K ∞Xn=1 (akn − ak′n)tnk < 2ε. n=1 aknZ n) are Cauchy in A and A1, n=0 aknZ n ∈ n=1 aknZ n −→ G ∈ A1 as k → ∞. Since A1 ∩ A = 0, then this representation is unique. Hence [A]K is the internal direct sum of subspaces A and A1, i.e., [A]K = A ⊕ A1. Therefore the sequences (ak0) and (P∞ respectively. For F ∈ [A]K , let F = limk→∞ Fk, where Fk = P∞ A[Z]. Then F = a + G, where ak0 −→ a ∈ A and P∞ (ii) Note that if a ∈ A and G = limk→∞P∞ n=1 aknZ n ∈ A1, then ka + ∞Xn=1 aknZ nkK = sup t∈K ka + akntnk, ∞Xn=1 ka +P∞ for all k ≥ 1. A similar method to that used in Part (i) implies that kak ≤ n=1 aknZ nk, for all k ≥ 1. Therefore kak ≤ ka + Gk, for all a ∈ A and G ∈ A1. (iii) Define pA : [A]K −→ A by pA(a + G) = a, for all a ∈ A and G ∈ A1. It is easily seen that pA is a surjective ∗-homomorphism and ker pA = A1. Therefore we have the short exact sequence 0 → A1 i ֒→ [A]K pA→ A → 0. (7) This shows that [A]K is an extension of A by A1. (iv) Suppose on the contrary that A1 is unital with unit U (Z). Since aZU (Z) = aZ for all a ∈ A, we have taU (t) = ta for any t ∈ K and a ∈ A. This implies that aU (t) = a for all t 6= 0 and therefore limt→0 U (t) 6= 0, which is a contradiction. Remark 1. Each ∗-homomorphism f : A −→ B of C∗-algebras induces a ∗- homomorphism f : A[Z] −→ B[Z] between pre-C∗-algebras A[Z] and B[Z] by f ( ∞Xn=0 anZ n) = ∞Xn=0 f (an)Z n, (8) where P∞ n=0 anZ n ∈ A[Z]. Remark 2. If we define P(A) = A[Z] for any C∗-algebra A and P(f ) = f , for any ∗-homomorphism f : A −→ B of C∗-algebras, then P is a functor from the category of C∗-algebras to the category of pre-C∗-algebras. Each ∗- homomorphism f : A[Z] −→ B[Z] defined by (8) induces a ∗-homomorphism f : [A]K −→ [B]K. It is easy to see that [·]K is a functor from the category of C∗-algebras to itself as [f ]K = f . Now defining F(A) = A1 and F(A −→ B) = f A1 : A1 −→ B1, for C∗-algebras A, B and ∗-homomorphism f , we get a functor on the category of C∗-algebras which assigns, by Proposition 7 (iv), to any C∗-algebra a non-unital C∗-algebra. f 7 By A ∼= B we mean that the C∗-algebras A and B are ∗-isomorphic. Proposition 8. Let f : A −→ B be a ∗-homomorphism of C∗-algebras. Then (i) f is a contraction; (ii) f and f are isometries provided that f is an isometry; (iii) f is surjective provided that f is surjective; (iv) If f is a ∗-isomorphism, then both f and f are ∗-isomorphisms; (v) ker f = (ker f )[Z]; (vi) Im f = (Im f )[Z]; (vii) If I is a closed ideal of A, then I[Z] is a closed ideal of (A[Z], k · kK). In particular, and 0 −→ I[Z] ֒→ A[Z] p′ −→ A[Z]/I[Z] −→ 0 0 −→ I[Z] ֒→ A[Z] p −→ (A/I)[Z] −→ 0 are short exact sequences; (viii) [A ⊕ B]K ∼= [A]K ⊕ [B]K; (ix) \(A ⊕ B)1 ∼= A1 ⊕ B1. Proof. The proof of (iv) follows from (ii) and (iii). The proofs of (v) and (vi) are straightforward and the proof of (ix) is similar to Part (viii). We prove the others. (i) suppose that F (Z) =P∞ n=0 anZ n ∈ A[Z]. Then k f (F )kK = kP∞ n=0 f (an)Z nkK = supt∈K kP∞ = supt∈K kf (P∞ ≤ supt∈K kP∞ = kF kK. n=0 f (an)tnk n=0 antn)k n=0 antnk (ii) If f is an isometry, then the proof of (i) shows that k f (F )kK = kF kK, for all F ∈ A[Z]. That is f and consequently f is an isometry. 8 (iii) Let f be surjective and G = P∞ n=0 bnZ n ∈ B[Z]. For any integer n ≥ 0, there exists an ∈ A such that bn = f (an). For any integer n ≥ 0 there exists a′ n ∈ ker f such that kan + a′ nk ≤ kan + ker f k + 2−n. Since A/ ker f ∼= B, we have kan + ker f k = kf (an)k = kbnk. (9) (10) Define a′′ F (Z) =P∞ f (F ) = G. n = an + a′ n=0 a′′ n, for all n ≥ 0. Now we see from (9) and (10) that n) = bn, for each n ≥ 0, and therefore nZ n ∈ A[Z] and f (a′′ (vii) Exactness of first diagram is clear. Part (iii) shows that A[Z] p → (A/I)[Z] −→ A/I is surjective. By (v) ker p = I[Z] is p induces by the projection A a closed ideal of A[Z]. This completes the proof. (viii) It is easily seen that T : A[Z] ⊕ B[Z] −→ (A ⊕ B)[Z] defined by T ( anZ n, ∞Xn=0 ∞Xn=0 n=0 anZ n ∈ A[Z] and P∞ for all P∞ bnZ n) = ∞Xn=0 (an, bn)Z n, n=0 bnZ n ∈ B[Z], is a ∗-isomorphism. 3 Exactness of the functor [·]K In this section we show that [·]K is an exact functor. We first recall some definitions of the category theory [11]. f g Recall that a map X −→ Z and Y −→ Y in a category C is called an epimorphism if for h−→ Z in C with g ◦ f = h ◦ f , we have g = h. In all maps Y the category of C∗-algebras, a ∗-homomorphism f : A → B is an epimorphism if and only if it is surjective [13]. f Suppose that X j −→ X is a kernel of f if f ◦ j = 0 and for any map Z ′ −→ Y is a map in a category C with zero object. A map Z −→ X in C such that f ◦ g = 0, there exists a unique map Z ′ h−→ Z such that j ◦ h = g. For example, f −→ B is a ∗-homomorphism of C∗-algebras, then the inclusion ker f ֒→ A if A is a kernel of f . g Theorem 1. The functor [·]K is exact. 9 Proof. Suppose that 0 −→ A f −→ B g −→ C −→ 0 is a short exact sequence of C∗-algebras. We must show that 0 −→ [A]K f −→ [B]K g −→ [C]K −→ 0 (11) is a short exact sequence of C∗-algebras. We first show that if f : A → B is a surjective ∗-homomorphism of C∗-algebras, then f : [A]K → [B]K is also a h→ C surjective ∗-homomorphism of C∗-algebras. To do this suppose that [B]K g → C are ∗-homomorphism of C∗-algebras such that h ◦ f = g ◦ f . and [B]K From Proposition 8 (iii), we have f (A[Z]) = B[Z]. So for any G(Z) ∈ B[Z] there exists an element F (Z) ∈ A[Z] such that h(G(Z)) = h( f (F (Z))) = (g ◦ f )(F (Z)) = g(G(Z)). This implies that hB[Z] = gB[Z] and therefore g = h. Hence f is an epimorphism and consequently is surjective by [13]. Now we show that if A g ker f = [ker f ]K. To prove this, suppose that C of C∗-algebras such that f ◦ g = 0. If gC = pA ◦ g and C then gC ◦ iC = g. Since ker f ∗-homomorphism C h−→ ker f such that the diagram f −→ B is a ∗-homomorphism of C∗-algebras, then −→ [A]K is a ∗-homomorphism iC֒→ [C]K is the injection, j ֒→ A is a kernel of f , there exists a unique ker f h f B A gC C is commutative. Since [·]K is a functor we get the commutative diagram [ker f ]K j h f [B]K [A]K gC [C]K 10 Putting h′ = h ◦ iC we get j ◦ h′ = g, since j ◦ h = gC. Now we show that k−→ [ker f ]K such h′ is unique. Suppose that there is a ∗-homomorphism C that j ◦ k = g = j ◦ h′. Since j is an injection, then k = h′, which proves the uniqueness of h′. It is clear that ker f = [ker f ]K. Now the Parts (ii), (v) and (vi) of Proposition 8 imply that (11) is a short exact sequence of C∗-algebras, or equivalently [·]K is an exact functor. The diagram [ker f ]K j [A]K f [B]K h gC k [C]K g iC C shows the detials above. Corollary 1. If I is a closed ideal of a C∗-algebra A, then [A/I]K ∼= [A]K/[I]K. Proof. By Theorem 1, the short exact sequence 0 −→ I ֒→ A −→ A/I −→ 0 induces the short exact sequence 0 −→ [I]K ֒→ [A]K −→ [A/I]K −→ 0 which implies that [A]K/[I]K ∼= [A/I]K. In the following corollary we use 3 × 3 lemma in homological algebra for the C∗-algebras as complex vector spaces (see, e.g., [14]). Corollary 2. If 0 −→ A f −→ B g −→ C −→ 0 is a short exact sequence of C∗-algebras, then 0 −→ A1 f A1−→ B1 g B1−→ C1 −→ 0 is also a short exact sequence of C∗-algebras, i.e., F is an exact functor (see Remark 2). Furthermore, if I is a closed ideal of A, then (dA/I)1 ∼= A1/ I1 Proof. In the commutative diagram 11 0 0 0 f A1 f f 0 A1 [A]K pA A 0 0 B1 [B]K pB B 0 g B1 g g 0 C1 [C]K pC C 0 0 0 0 the middle row is exact by Theorem 1 and all columns are exact by (7). Now 3 × 3 Lemma [14] shows that the top row is also exact. By a similar argument as in Corollary, 1 we get (dA/I)1 ∼= A1/ I1. Recall that an ideal I of a C∗-algebra A is called modular if there is an element u ∈ A such that ua − a, au − a ∈ A, for all element a ∈ A. Note that I is modular if and only if A/I is unital [12]. Corollary 3. Let I be a closed ideal of a C∗-algebra A. Then I is a modular ideal of A if and only if [I]K is a modular ideal of [A]K. Proof. We first show that a C∗-algebra B is unital if and only if [B]K is unital. It can be easily seen that if B is unital, then [B]K is also unital. Now, by Proposition 7 (i) suppose that [B]K is unital with unit a + G for some a ∈ B and G ∈ B1. Consider an arbitrary element b + F ∈ [B]K with b ∈ B and F ∈ B1. Then (b + F )(a + G) = b + F or equivalently ba + F G + F a + bG = b + F . It follows that ba − b = H, for some H ∈ B1. Since B ∩ B1 = 0, then ba = b. Similarly ab = b. This shows that a is the unit of B. Now let I be a closed ideal of A. Then by Corollary 1, I is modular if and only if [A/I]K ∼= [A]K/[I]K is unital. Hence I is modular if and only if [I]K is modular. 12 4 Normal exactness of the functor P Suppose that A is a C∗-algebra and I is a closed ideal of A. It follows from Proposition 8 (vii) that A[Z]/I[Z] is a normed algebra with the usual quotient norm. In this section, we show that A[Z]/I[Z] is a pre-C∗-algebra. Also using Five Lemma and Theorem 2 below, we will show that A[Z]/I[Z] is isometric ∗-isomorphic to (A/I)[Z]. This implies that the functor P is, in fact, normal exact. We remind that the Five Lemma in homological algebra (see, e.g., [14]) says that in the commutative diagram A1 t1 B1 A2 t2 B2 A3 t3 B3 A4 t4 B4 A5 t5 B5 of commutative R-modules with exact rows if t1, t2, t4 and t5 are isomorphisms, so is t3. Definition 1. [16] The exact sequence · · · −→ An fn−→ An+1 fn+1−→ An+2 −→ · · · of normed spaces with contraction fn (kfnk ≤ 1 for any n) is called normal exact if the induced map An/ ker fn −→ fn(An) defined by x + ker fn 7−→ fn(x), is an isometry. Note that any short exact sequence of C∗-algebras is normal exact. The following theorem is the main one in [16]. Theorem 2. Suppose that 0 −→ Y i−→ X p −→ Z −→ 0 is a normal exact sequence of normed spaces. Then 0 −→ Y i −→ X p −→ Z −→ 0 is a normal exact sequence of corresponding completion Banach spaces. Theorem 3. Let I be a closed ideal of a C∗-algebra A. Then A[Z]/I[Z] is a pre-C∗-algebra. Proof. We first show that (i) If (uλ)λ∈Λ is an approximate unit for A, then (uλ)λ∈Λ is also an approximate unit for A[Z]; 13 (ii) If (uλ)λ∈Λ is an approximate unit for I, then for any F (Z) ∈ A[Z] we have kF (Z) + I[Z]k = limλ kF (Z) − uλF (Z)kK = limλ kF (Z) − F (Z)uλkK. n=0 anZ n ∈ A[Z] and ε > 0 be given. Since n=N +1 2kank < ε. Now for any λ ∈ Λ we have To prove (i), let F (Z) = P∞ P∞ n=0 kank < ∞, there is a positive integer N such that P∞ kF (Z) − uλF (Z)kK = kP∞ ≤ P∞ = PN < PN n=0 kan − uλank +P∞ n=0(an − uλan)Z nkK n=0 kan − uλank n=0 kan − uλank + ε. n=N +1 kan − uλank Therefore lim λ sup kF (Z) − uλF (Z)kK ≤ ε. Since ε > 0 was arbitrary, we have lim λ lim λ kF (Z) − uλF (Z)kK = 0. kF (Z) − F (Z)uλkK = 0. Similarly we get To prove (ii) let α = kF (Z) + I[Z]k = inf{kF (Z) + H(Z)kK : H(Z) ∈ I[Z]}. Let ε > 0 be given. There exists an element G(Z) ∈ I[Z] such that kF (Z) − G(Z)kK < α + ε. We have α ≤ kF (Z) − F (Z)uλkK ≤ k(F (Z) − G(Z)) − (F (Z) − G(Z))uλkK + kG(Z) − G(Z)uλkK = k(F (Z) − G(Z))(1 − uλ)kK + kG(Z) − G(Z)uλkK ≤ kF (Z) − G(Z)kK + kG(Z) − G(Z)uλkK < α + ε + kG(Z) − G(Z)uλkK. Now by Part (i) we have α ≤ limλ inf kF (Z) − F (Z)uλkK ≤ α + ε, α ≤ limλ sup kF (Z) − F (Z)uλkK ≤ α + ε. 14 Since ε > 0 was arbitrary, we have α = limλ kF (Z) − F (Z)uλkK. Similarly, α = limλ kF (Z) − uλF (Z)kK. To prove the theorem let (uλ)λ∈Λ be an approximate unit for I. If F (Z) ∈ A[Z] and G(Z) ∈ I[Z], by Parts (i), (ii) and Proposition 5 (i) we have kF (Z) + I[Z]k2 = limλ kF (Z) − F (Z)uλk2 K = limλ k(1 − uλ)F ∗(Z)F (Z)(1 − uλ)kK ≤ limλ k(1 − uλ)(F ∗(Z)F (Z) + G(Z))(1 − uλ)kK + limλ k(1 − uλ)G(Z)(1 − uλ)kK ≤ kF ∗(Z)F (Z) + G(Z)kK. Therefore kF (Z) + I[Z]k2 ≤ kF ∗(Z)F (Z) + I[Z]k and consequently we get the equality kF (Z) + I[Z]k2 = kF ∗(Z)F (Z) + I[Z]k, which completes the proof. Now we are ready to show that the functor P is normal exact. Theorem 4. The functor P is normal exact. Proof. Let I be a closed ideal of a C∗-algebra A. First we show that there exists an isometric ∗-isomorphism between A[Z]/I[Z] and (A/I)[Z]. Define T : A[Z]/I[Z] −→ (A/I)[Z] by T ( ∞Xn=0 anZ n + I[Z]) = ∞Xn=0 (an + I)Z n, n=0 anZ n ∈ A[Z]. It is clear that T is well defined, linear and preserves the involution. We are going to show that (a) T is injective, (b) T is surjective, (c) T is a contraction, and (d) T is an isometry. We proceed as follows: n=0(an + I)Z n = I, i.e., an ∈ I for n = 0, 1, 2, · · · . Therefore F (Z) ∈ I[Z] and so T is injective. n=0 anZ n ∈ A[Z] with T (F ) = I, thenP∞ for allP∞ (a) If F (Z) =P∞ (b) Let G =P∞ n=0(an + I)Z n ∈ (A/I)[Z]. For each n = 0, 1, 2, · · · there is an n=0 cnZ n, n=0 kan + Ik < ∞ element bn ∈ I such that kan + bnk < kan + Ik + 2−n. Let P∞ where cn = an + bn for each n = 0, 1, 2, · · · . Since P∞ we have F (Z) ∈ A[Z]. Therefore T (F (Z) + I[Z]) = P∞ = P∞ = G, n=0(cn + I)Z n n=0(an + I)Z n 15 that is T is surjective. (c) Let F (Z) =P∞ n=0 anZ n ∈ A[Z]. Then kT (F (Z) + I[Z])k = kP∞ n=0(an + I)Z nkK n=0(an + I)tnk n=0 antn + Ik = supt∈K kP∞ = supt∈K kP∞ = supt∈K inf b∈I kP∞ ≤ inf b∈I supt∈K kP∞ = inf G(Z)∈I[Z] supt∈K kP∞ = inf G kF (Z) + G(Z)kK n=0 antn + bk n=0 antn + bk n=0 antn + G(t)k = kF (Z) + I[Z]k, that is T is a contraction. (Note that sup inf f ≤ inf sup f for every real valued function f in two variables) (d) Suppose that \(A[Z]/I[Z]) is the completion of A[Z]/I[Z] with respect to the quotient norm and T : ( \A[Z]/I[Z]) −→ [A/I]K, is the extension of T . By Theorem 3, T is a ∗-homomorphism of C∗- algebras. Now we show that T is a ∗-isomorphism. The diagram n=0 anZ n P∞ n=0 anZ n P∞ p′ p n=0 anZ n + I[Z] P∞ T n=0(an + I)Z n P∞ shows that the diagram 0 0 I[Z] A[Z] I[Z] A[Z] p′ p A[Z]/I[Z] T (A/I)[Z] 0 0 of pre-C∗-algebras is commutative, where p′ is the quotient map and p is the map induced by the projection A −→ A/I (see Definition 1). The ex- actness of two rows follow from Proposition 8 (vii). Now the commutative diagram p 16 0 0 [I]K [A]K [I]K [A]K p′ p ( \A[Z]/I[Z]) T [A/I]K 0 0 of C∗-algebras have exact rows. In fact, the exactness of first row is a consequence of Theorem 2 and the second one follows from Theorem 1. Applying Five Lemma for commutative diagram [I]K [A]K t1 t2 [I]K [A]K p′ p ( \A[Z]/I[Z]) t3 = T [A/I]K t4 0 0 t5 0 0 with exact rows shows that T is a ∗-isomorphism. This implies, partic- ularly, that T is an isometry. Now consider the short exact sequence of C∗-algebras 0 −→ I i ֒→ A g −→ B −→ 0. Applying functor P we get a short exact sequence of pre-C∗-algebras 0 −→ I[Z] i ֒→ A[Z] g −→ B[Z] −→ 0. (12) Note that we have the ∗-isomorphism g1 : A/I −→ B, induced by g. By Part (d) we have the composition of isometric ∗-isomorphism of pre-C∗- algebras A[Z]/I[Z] T−→ (A/I)(Z) g1−→ B[Z] such that ∞Xn=0 anZ n + I[Z] 7→ ∞Xn=0 (an + I)Z n 7→ ∞Xn=0 g(an)Z n. That, is the induced map A[Z]/I[Z] −→ B[Z] by g is an isometry. There- fore (12) is a normal exact sequence of pre-C∗-algebras. From (c) and (d) of Theorem 4 we have the following. Corollary 4. Suppose that I is a closed ideal of a C∗-algebra A and a0, a1, a2, · · · , is a sequence in A such that P∞ n=0 kank < ∞. Then inf b∈I sup t∈K k ∞Xn=0 antn + bk = sup t∈K k inf b∈I 17 ∞Xn=0 antn + bk. 5 Cauchy extension [A]K as C∗-subalgebra of Cb(K, A) In this section, we characterize the Cauchy extensions of C∗-algebras as C∗- valued function spaces. Using the obtained characterization, we give some re- sults on the Cauchy extensions of C∗-algebras. Recall that for a C∗-algebra A and a topological space X, Cb(X, A) is the set of all bounded continuous functions from X to A. The addition, scalar multipli- cation and the product on Cb(X, A) are defined pointwise. The involution can be defined as α∗(x) = (α(x))∗, for all α ∈ Cb(X, A) and x ∈ X. Furthermore, defining kαk∞ = supx∈X kα(x)k for all α ∈ Cb(X, A), the algebra Cb(X, A) be- comes a C∗-algebra. If X is a locally compact Hausdroff space, then C0(X, A) consisting of all continuous functions f ∈ Cb(X, A) vanishing at infinity is a C∗-subalgebra of Cb(X, A) (see [12, p.37] ). If X is a compact Hausdorff space, then Cb(X, A) = C0(X, A) = C(X, A). It is easy to see that for C∗-algebras A1, A2, · · · , An, we have Cb(X, A1 ⊕ · · · ⊕ An) ∼= Cb(X, A1) ⊕ · · · ⊕ Cb(X, An). (13) In particular, if A = C, we use C(X), Cb(X) and C0(X) for C(X, C), Cb(X, C) and C0(X, C), respectively. Recall that a C∗-algebra A is called nuclear if for each C∗-algebra B, there is a unique C∗-norm on tensor product A ⊗ B. An ideal I of a C∗-algebra A is called essential if aI = 0 implies that a = 0. Theorem 5. Suppose that A and B are two C∗-algebras and K ⊆ J = [−1, 1] such that 0 is a limit point of K. Then (i) [A]K is ∗-isomorphic to a C∗-subalgebra of Cb(K, A); (ii) If K is a compact interval, then [A]K ∼= C(K, A); (iii) [A]K ∼= {f K : f ∈ C([−1, 1], A)}; (iv) If K is compact, then [A]K ∼= C(K, A). Furthermore, [A ⊗ B]K ∼= [A]K ⊗ B ∼= A ⊗ [B]K; (v) There is a closed ideal IK of [A]J such that [A]J /IK ∼= [A]K; (vi) A is nuclear if and only if [A]K is nuclear; (vii) I is an essential ideal of A if and only if [I]K is an essential ideal of [A]K ; (viii) If 0 /∈ K and K is a locally compact subspace of J such that K ′ = K ∪ {0} is compact, then [A]K ∼= C(K ′, A). If A is finite dimensional, then M ( A1) ∼= Cb(K, A), where M ( A1) is the multiplier algebra of A1; (ix) A ∼= B if and only if [A]K ∼= [B]K for any compact set K. 18 Proof. n=0 antn) = P∞ (i) It is clear that for any sequence (an) in A with P∞ tion f (t) = P∞ T : A(K) −→ A[Z] defined by T (P∞ n=0 kank < ∞ the summa- n=0 antn, where t ∈ K defines a function from K to A. Denote the set of all such functions by A(K). It is clear that f is a bounded contin- uous function on K and A(K) is a ∗-subalgebra of Cb(K, A). Now the map n=0 anZ n is an isometric ∗-isomorphism. That is [A]K is ∗-isomorphic to a C∗-subalgebra of Cb(K, A). (ii) For the case that A = C, since C(K) is a self-adjoint subalgebra of C(K) which separate points of K and contains the constant functions one can see, by Stone-Weierstrass Theorem (see [15, p.165]), that [C]K ∼= C(K). Now for any C∗-algebra A and any compact interval K one can use approximate Berstein Theorem (see [2, p.182] ), as follows: We may assume that K = [0, 1]. Let f ∈ C(K, A). Because f is uniformly continuous (see [8, p.60] ), define the Bernstein Polynomials βn(t) = f (m/n)(cid:18) n m(cid:19)tm(1 − t)n−m, nXm=0 for any t ∈ K and integer n > 0. Note that βn ∈ A(K) for any n = 1, 2, 3, · · · . By a similar argument as in the proof of the Berstein Theorem, we see that βn is convergent uniformly to f . This shows that [A]K ∼= C(K, A). (iii) Define T : A(J) −→ A(K) by T (f ) = f K, for each f ∈ A(J). It is clear that T is a bijective bounded linear operator. We claim that the extension T : [A]J −→ [A]K is surjective. Note that Parts (i) and (ii) imply that T is of the form T (g) = gK for all g ∈ [A]J . Suppose that H, G : [A]K −→ B are two ∗-homomorphisms such that G ◦ T = H ◦ T . This implies that H ◦ T A(J) = G ◦ T A(J) or HA(K) = GA(K). Since A(K) ∼= A[Z] is dense in [A]K , then H = G. Hence T is surjective (see [13]). By (ii) we have [A]J ∼= C(J, A) and therefore [A]K ∼= {f K : f ∈ C(J, A)}. (iv) By Tietze's Theorem ([9, Theorem 4.1]), any continuous function f : K −→ A has a continuous extension f1 : J −→ A. This fact together with Part (iii) show that [A]K ∼= C(K, A). From ([3, II.6.4.4] ) we have C(K, A) ∼= C(K) ⊗ A and therefore [A ⊗ B]K ∼= C(K) ⊗ (A ⊗ B) ∼= [A]K ⊗ B ∼= A ⊗ [B]K. (v) Let T : [A]J −→ [A]K be the given surjective ∗-homomorphism in Part (iii). If IK = ker T , then [A]J /IK ∼= [A]K . In fact, [A]J is an extension of any Cauchy extension [A]K . (vi) Let A be nuclear. By Part (ii) we have [A]J ∼= C(J, A). Since C(J) is nuclear (see [12, Theorem 6.4.15]) and C(J, A) ∼= C(J) ⊗ A (see [3, II.6.4.4]) we imply that [A]J is nuclear (see [3, IV.3.1.1]). Since every closed ideal of a nuclear C∗-algebra is nuclear (see [3, II.9.6.3]), then A1 is nuclear. In particular, the closed ideal IK (given in part (v)) is nuclear. Since [A]J /IK is nuclear (see 19 [3, IV 3.1.13]), Part (v) implies that [A]K is also nuclear. Conversely, if [A]K is nuclear, then the ideal A1 is nuclear. By (7) we have A ∼= [A]K / A1 which shows that A is nuclear too. (vii) Let I be an essential ideal of A. By Part (i) we can consider [A]K as a C∗- subalgebra of Cb(K, A). Choose G : K −→ A in [A]K such that f G = Gf = 0 for any f : K −→ I ∈ [I]K . For any t ∈ K we have f (t)G(t) = G(t)f (t) = 0. Let b be an arbitrary element in I and let fb : K −→ I be a constant function with value fb(t) = b. Now for any t ∈ K we have or fb(t)G(t) = G(t)fb(t) = 0 bG(t) = G(t)b = 0. This implies that G(t) = 0 for all t ∈ K. Therefore [I]K is an essential ideal of [A]K. The converse statement can be proved similarly. (viii) Suppose that C1(K) = {f ∈ C(K) : f (0) = 0}, where C(K) is as given in Part (i). For f ∈ C1(K) and ε > 0, suppose that X = {t ∈ K : f (t) ≥ ε} and x is a limit point of X. Then x 6= 0 and x is a limit point of K ′, and therefore x ∈ K. This implies that X is compact. That is f vanishes at infinity, so C1(K) ⊆ C0(K). Now suppose that 0 6= a ∈ C and g(x) = xa for all x ∈ K. Then g ∈ C1(K) and for any t ∈ K we have g(t) 6= 0. In addition, for any t1 6= t2 in K, g(t1) 6= g(t2), that is, C1(K) strongly separates points of K. It is clear that C1(K) is self-adjoint. By the Stone-Weierstrass Theorem (see [7, p.151]) we have C1 ∼= C0(K) and therefore [C]K ∼= C ⊕ C0(K) ∼= C(K ′) (see [3, p.53]). Parts (iii) and (iv) and the fact that kf kK = kf kK ′ for any f ∈ C(J, A) imply that the map f K 7→ f K ′ is a ∗-isomorphism between [A]K and C(K ′, A). Now suppose that A is a finite dimensional C∗-algebra. By ([12, p.194]) we have A ∼= Mn1(C) ⊕ Mn2(C) ⊕ · · · ⊕ Mnm(C). (14) We first show that for any positive integer n, \(Mn(C))1 ∼= Mn( C1). To see this, note that the completion of C1(K) is ∗-isomorphic to C0(K). Now the map G : (Mn(C))1(K) −→ Mn(C1(K)) defined by G(F ) = (Fij ), where F (t) = ∞Xm=1 Bmtm = (Fij (t)) and Fij ∈ C1(K), for any i, j = 1, 2, · · · , n is an isometric ∗-isomorphism with norm k(Fij)k = supt∈K kFij (t)k = kF k. Suppose that F = (Fij ) ∈ Mn(C0(K)), then Fij ∈ C0(K) for i, j = 1, 2, · · · , n. There exist sequences (Fmij ) in C1(K) for i, j = 1, 2, · · · , n such that Fmij −→ Fij as m → ∞ in norm k · kK. If F : K −→ Mn(C) is a continuous function such that for any t ∈ K, F (t) = (Fij (t)), then k(Fmij ) − (Fij )k = supt∈K k(Fmij(t)) − Fij (t)k ≤ supt∈KPi,j kFmij (t) − Fij (t)k ≤ Pi,j supt∈K kFmij (t) − Fij (t)k. 20 This implies that (Fmij ) −→ (Fij ) as m −→ ∞. Now by completion we see that \(Mn(C))1 ∼= Mn(C0(K)) ∼= Mn( C1). Also we have clearly the ∗-isomorphism Mn(Cb(K)) ∼= Cb(K, Mn(C)). (15) (16) Suppose that B, A1, A2, · · · , An are C∗-algebras. We have the following for the multipliers algebras (see [4, p.84]) M (Mn(B)) ∼= Mn(M (B)) (17) M (A1 ⊕ A2 ⊕ · · · ⊕ An) ∼= M (A1) ⊕ M (A2) ⊕ · · · ⊕ M (An), (18) (see [3, p.155]). We also have M (C0(K)) ∼= Cb(K) (see [12, p.83]). Now from (13) − (18), and Proposition 8 (ix), we have A1 ∼= Mn1(C0(K)) ⊕ Mn2(C0(K)) ⊕ · · · ⊕ Mnm(C0(K)). M ( A1) ∼= M (Mn1(C0(K))) ⊕ · · · ⊕ M (Mnm(C0(K))) ∼= Cb(K, Mn1(C)) ⊕ · · · ⊕ Cb(K, Mnm(C)) ∼= Cb(K, Mn1(C) ⊕ · · · ⊕ Mnm(C)) ∼= Cb(K, A). (ix) If A ∼= B, then [A]K ∼= [B]K by Proposition 8 (iv). Let ϕn : [A]Kn −→ [B]Kn be a ∗-isomorphism between [A]Kn and [B]Kn, where Kn = [−1/n, 1/n] for n=1 Kn = {0}. Now n = 1, 2, 3, · · · . ([A]Kn , pn)∞ It is clear that (Kn) is nested with T∞ n=1 is a direct sequence of C∗-algebras, where each map pn : [A]Kn −→ [A]Kn+1 defined by f Kn 7→ f Kn+1, for all f ∈ [A]K is a ∗-homomorphism. Part (iv) and [3, II.6.4.4] show that [A]Kn ∼= C(Kn, A) ∼= C(Kn) ⊗ A, for all n. Furthermore by [3, II.9.6.5] we have the direct limit [A]Kn lim −→ ∼= lim −→ (C(Kn) ⊗ A) ∼= (lim −→ C(Kn)) ⊗ A ∼= C({0}) ⊗ A ∼= C ⊗ A ∼= A. From the commutative diagram [A]Kn pn ϕn [B]Kn qn [A]Kn+1 ϕn+1 [B]Kn+1 21 where ([B]Kn, qn)∞ for any f ∈ [A]Kn , we conclude that n=1 is the direct sequence defined by qn(ϕn(f )) = ϕn+1(f Kn+1), A ∼= lim −→ [A]Kn ∼= lim −→ [B]Kn ∼= B, as desired. Any C∗-algebra of the form B = Mn1(C[a1, b1]) ⊕ · · · ⊕ Mnm(C[an, bn]) where ai < bi for i = 1, 2, · · · , n are real numbers, is a Cauchy extension of some C∗-algebra. In fact B ∼= Mn1(C[−1, 1]) ⊕ Mn2 (C[−1, 1]) ⊕ · · · ⊕ Mnm(C[−1, 1]). Therefore B ∼= [A]J , where A is the C∗-algebra defined in (14). Corollary 5. Suppose that A is a C∗-algebra and I is a closed ideal of A. If K = [0, 1] and F ∈ C(K, A), then inf b∈I sup t∈K kF (t) + bk = sup t∈K inf b∈I kF (t) + bk. Proof. Let ε > 0 be given. By Theorem 5 (ii) there exists an element Fn ∈ A(K) such that supt∈K kF (t) − Fn(t)k < ε. For any t ∈ K we have kF (t) + bk ≤ kF (t) − Fn(t)k + kFn(t) + bk < ε + kFn(t) + bk. On the other hand kFn(t) + bk ≤ kFn(t) − F (t)k + kF (t) + bk < ε + kF (t) + bk, for any t ∈ K. By Corollary 4 we have inf b∈I supt∈K kF (t) + bk ≤ ε + inf b∈I supt∈K kFn(t) + bk = ε + supt∈K inf b∈I kFn(t) + bk ≤ 2ε + supt∈K inf b∈I kF (t) + bk. Since ε > 0 was arbitrary, then inf b∈I sup t∈K kF (t) + bk ≤ sup t∈K inf b∈I kF (t) + bk. This completes the proof. 22 References [1] Arveson, W. Notes on extensions of C ∗ -algebras. Duke Math. J. 44 (1977), no. 2, 329 -- 355. [2] Bartle, R.G., The elements of real analysis. Second edition. John Wiley & Sons, New York-London-Sydney, 1976. [3] Blackadar, B., Operator algebras. Theory of C∗ -algebras and Von Neu- mann algebras. Encyclopaedia of Mathematical Sciences, 122. Operator Al- gebras and Non-commutative Geometry, III. Springer-Verlag, Berlin, 2006. [4] Blecher, D.P.; Le Merdy, C. Operator algebras and their modules an op- erator space approach. London Mathematical Society Monographs. New Series, 30. Oxford Science Publications. The Clarendon Press, Oxford Uni- versity Press, Oxford, 2004. [5] Brown, L.G.; Douglas, R. G.; Fillmore, P. A. Extensions of C∗-algebras and K-homology. Ann. of Math. (2) 105 (1977), no. 2, 265 -- 324. [6] Busby, R.C., Double centralizers and extensions of C∗ -algebras. Trans. Amer. Math. Soc. 132 (1968) 79 -- 99. [7] Conway, J.B., A course in functional analysis. Graduate Texts in Mathe- matics, 96. Springer-Verlag, New York, 1985. [8] Dieudonne, J., Foundations of modern analysis. Enlarged and corrected printing. Pure and Applied Mathematics, Vol. 10-I. Academic Press, New York-London, 1969. [9] Dugundji, J., An extension of Tietze's theorem. Pacific J. Math. 1, (1951). 353 -- 367. [10] Kasparov, G.G., The operator K -functor and extensions of C∗ -algebras. (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), no. 3, 571 -- 636, 719. [11] MacLane, S., Categories for the working mathematician. Graduate Texts in Mathematics, Vol. 5. Springer-Verlag, New York-Berlin, 1971. [12] Murphy, G.J., C∗ -algebras and operator theory. Academic Press, Inc., Boston, MA, 1990. [13] Reid, G. A., Epimorphisms and surjectivity. Invent. Math. 9 (1969/1970), 295 -- 307. [14] Rotman, J.J., An introduction to homological algebra. Second edition. Uni- versitext. Springer, New York, 2009. [15] Rudin, W., Principles of mathematical analysis. Third edition. Interna- tional Series in Pure and Applied Mathematics. McGraw-Hill Book Co., New York-Auckland-DCsseldorf, 1976. 23 [16] Yang, K.W., Completion of normed linear spaces. Proc. Amer. Math. Soc. 19 (1968), 801 -- 806. 24
1201.3879	1	1201	2012-01-18T19:01:49	The generator problem for Z-stable C*-algebras	[ "math.OA" ]	The generator problem was posed by Kadison in 1967, and it remains open until today. We provide a solution for the class of C-algebras absorbing the Jiang-Su algebra Z tensorially. More precisely, we show that every unital, separable, Z-stable C-algebras A is singly generated, which means that there exists an element x in A that is not contained in any proper sub-C-algebra of A. To give applications of our result, we observe that Z can be embedded into the reduced group C-algebra of a discrete group that contains a non-cyclic, free subgroup. It follows that certain tensor products with reduced group C-algebras are singly generated. In particular, the tensor product of two reduced free group C-algebras is singly generated.	math.OA	math	THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS HANNES THIEL AND WILHELM WINTER ABSTRACT. The generator problem was posed by Kadison in 1967, and it remains open until today. We provide a solution for the class of C ∗-algebras absorbing the Jiang-Su algebra Z tensorially. More precisely, we show that every unital, separa- ble, Z-stable C ∗-algebra A is singly generated, which means that there exists an element x ∈ A that is not contained in any proper sub-C ∗-algebra of A. To give applications of our result, we observe that Z can be embedded into the reduced group C ∗-algebra of a discrete group that contains a non-cyclic, free subgroup. It follows that certain tensor products with reduced group C ∗-algebras are singly generated. In particular, C ∗ r (F∞) is singly generated. r (F∞) ⊗ C ∗ . A O h t a m [ 1 v 9 7 8 3 . 1 0 2 1 : v i X r a 1. INTRODUCTION By an operator algebra we mean a ∗-subalgebra of B(H) that is either closed in the norm topology (a concrete C ∗-algebra) or the weak operator topology (a von Neumann algebra). One way of realizing an operator algebra is to take a subset of B(H) and consider the smallest operator algebra containing it. In a trivial way, every operator algebra can be obtained this way. The situation becomes interesting if one imposes restrictions on the generating set, and one nat- ural possibility is to require that it consists of only one element, i.e., to consider operator algebras that are generated by a single operator. It is an old problem to determine which operator algebras arise this way. More generally, one tries to compute the minimal number of elements that gen- erate a given operator algebra, see 2.1. It is often convenient to consider self- adjoint generators. Note that two self-adjoint elements a, b generate the same op- erator algebra as the element a + ib. Thus, if we ask whether an operator algebra is singly generated, it is equivalent to ask whether it is generated by two self-adjoint elements. In the case of von Neumann algebras, the generator problem was included in Kadison's famous 'Problems on von Neumann algebras', [Kad67]. This problem list has turned out to be very influential, yet its original form remains unpublished. It is indirectly available in an article by Ge, [Ge03], where a brief summary of the developments around Kadison's famous problems is given. Date: November 12, 2018. 2000 Mathematics Subject Classification. Primary 46L05, 46L85; Secondary 46L35. Key words and phrases. C ∗-algebras, generator problem, single generation, Z-stability. This research was partially supported by the Centre de Recerca Matem`atica, Barcelona. The first named author was partially supported by the Danish National Research Foundation through the Cen- tre for Symmetry and Deformation, Copenhagen. The second named author was partially supported by EPSRC Grants EP/G014019/1 and EP/I019227/1. 1 2 HANNES THIEL AND WILHELM WINTER Question 1.1 (Kadison, [Kad67, Problem 14], see also [Ge03]). Is every separably- acting1 von Neumann algebra singly generated? We just mention an incomplete list of results. As noted in [She09], there exist singly generated von Neumann algebras that are not separably-acting. However, the separably-acting von Neumann algebras are the natural class for which one might expect single generation. The answer to Question 1.1 is still open in general, but many authors have contributed to show that large classes of separably-acting von Neumann algebras are singly generated. It starts with von Neumann, [vN31], who showed that the abelian operator algebras named after him are gener- ated by a single self-adjoint element, thus implicitly raising the generator problem. Some thirty years later, this was extended by Pearcy, [Pea62], who showed that all von Neumann algebras of type I are singly generated. Then Wogen, [Wog69, Theorem 2], proved that all properly infinite von Neumann algebras are singly generated, thus reducing the generator problem to the type II1 case. Later, this was further reduced to the case of a II1-factor by Willig, [Wil74], and then to the case of a finitely-generated II1-factor by Sherman, [She09, Theorem 3.8]. This means that Question 1.1 has a positive answer if every separably-acting, finitely generated II1-factor is singly generated. There are many properties known to imply that a II1-factor is singly generated. We just mention that Ge and Popa, [GP98, Theorem 6.2], show that every tensori- ally non-prime2 II1-factor is singly generated. Our main result Theorem 3.5 can be considered as a partial C ∗-algebraic analog of this result. Let us also mention that the free group factors W ∗(Fk) are the outstanding examples of separably-acting von Neumann algebra for which it is not known whether they are singly generated. In the case of C ∗-algebras, the generator problem is more subtle. There is al- ready no obvious class of C ∗-algebras for which one conjectures that they are singly generated. Every singly generated C ∗-algebra is separable3. However, the converse is false, and counterexamples can be found among the commutative C ∗- algebras. In fact, the C ∗-algebra C0(X) is generated by n self-adjoint elements if and only if X can be embedded into Rn. Thus, C0(X) is singly generated if and only if X is planar, i.e., can be embedded into the plane R2. It is easy to see that a C ∗-algebra A is generated by n self-adjoint elements if and only if its minimal unitization eA is generated by n self-adjoint elements. Therefore, we will mostly consider the generator problem for separable, unital C ∗-algebra. In that case, taking the tensor product with a matrix algebra has the effect of reduc- ing the necessary number of generators. If A is generated by n2 + 1 self-adjoint elements, then A ⊗ Mn is singly generated, see e.g. [Nag04, Theorem 3]. One derives the principle that a C ∗-algebra needs less generators if it is 'more non-commutative'. Consequently, one might expect a (separable) C ∗-algebra to be singly generated if it is 'maximally non-commutative'. As a non-unital instance 1A von Neumann algebra is called 'separably-acting', or just 'separable', if it is a subalgebra of B(l2N), or equivalently if it has a separable predual. 2A II1-factor M is called tensorially non-prime if it is isomorphic to a tensor product, M1 ¯⊗M2, of two II1-factors M1, M2. 3A C ∗-algebra is called 'separable' if it contains a countable, norm-dense subset THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 3 of this principle, we note that the stabilization, A ⊗ K, of a separable unital C ∗- algebra A is singly generated, [OZ76, Theorem 8]. In the unital case, there are at least three natural cases when one considers a C ∗-algebra A to be 'maximally non-commutative', which are the following: (1) A contains a simple, unital, nonelementary sub-C ∗-algebra, (2) A contains a sequence of pairwise orthogonal, full elements, (3) A has no finite-dimensional irreducible representations. In general, the implications (1) ⇒ (2) ⇒ (3) hold; it is not known if the con- verses are true. Conditions (2) and (3) can also be considered for possibly non-unital C ∗-alge- bras, and we let (2∗) be the weaker statement that A contains two orthogonal, full elements. The implication '(3) ⇒ (2)' holds exactly if the implication '(3) ⇒ (2∗)' holds. The Global Glimm halving problem asks the following: Given a (possibly non- unital) C ∗-algebra A that satisfies condition (3), does there exist a full map from the cone over M2 to A? It is not known whether the Global Glimm halving prob- lem has a positive answer, but if it does then it shows that implication '(3) ⇒ (2)' holds, since the cone over M2 contains two orthogonal, full elements. Let us remark that the analogs of conditions (1)−(3) for von Neumann algebras are all equivalent. In fact, if a von Neumann algebra M has no finite-dimensional representations, then the hyperfinite II1-factor R unitally embeds into M . Historically, the generator problem for C ∗-algebras is mostly asked for C ∗-al- gebras that are simple ore more generally have no finite-dimensional representa- tions: Question 1.2. Is every simple, separable, unital C ∗-algebra singly generated? Question 1.3. Is a separable, unital C ∗-algebra singly generated provided it has no finite-dimensional irreducible representations? The answers to both questions are open. A positive answer to Question 1.3 implies a positive answer to Question 1.2, of course. The converse is not clear. Let us mention some results that solve the generator problem for particular classes of separable C ∗-algebras. It was shown by Topping, [Top68], that ev- ery UHF-algebra is singly generated. This was generalized by Olsen and Zame, [OZ76, Theorem 9], who showed that the tensor product, A ⊗ B, of any separable, unital C ∗-algebra A with a UHF-algebra B is singly generated. Later, it was shown by Li and Shen, [LS10, Theorem 3.1], that every unital, ap- proximately divisible4 C ∗-algebra is singly generated. This generalizes the result of Olsen and Zame, since the tensor product with a UHF-algebra is always ap- proximately divisible. In this article we prove that every separable, unital, Z-stable C ∗-algebra is singly generated, see Theorem 3.7. This generalizes the result of Li and Shen, since every approximately divisible C ∗-algebra is Z-stable, see [TW08, Theorem 2.3]. The notion of Z-stability has proven to be very important in the classifica- tion program of nuclear C ∗-algebras, see e.g. [Win07] or [ET08], and it is has been 4A unital C ∗-algebra A is 'approximately divisible' if for every ε > 0 and finite subset F ⊂ A there exists a finite-dimensional, unital sub-C ∗-algebra B ⊂ A such that B has no characters and kxb − bxk ≤ εkbk for all x ∈ F, b ∈ B. 4 HANNES THIEL AND WILHELM WINTER shown that many nuclear, simple C ∗-algebras are Z-stable, see e.g. [Win10]. Z- stability is also relevant in the non-nuclear context; for example, unital Z-stable C ∗-algebras satisfy Kadison's similarity property, see [JW11]. This paper proceeds as follows: In Section 2 we set up our notation and give some basic facts about the genera- tor rank, see 2.1, and C0(X)-algebras, see 2.4. Section 3 contains the proof of our main result, which states that the tensor product A ⊗max B of two separable, unital C ∗-algebras is singly generated, if A satisfies condition (2) from above (e.g. A is simple and non-elementary) and B admits a unital embedding of the Jiang-Su algebra Z, see Theorem 3.5. We derive that every separable, unital, Z-stable C ∗-algebra is singly generated, see Theorem 3.7. Our main result can be considered as a (partial) C ∗-algebraic analog of a theorem of Ge and Popa, [GP98, Theorem 6.2], which shows that a tensor product, M ¯⊗N , of two II1-factors M, N is singly generated. In fact, we can reprove their theorem with our methods, see Corollary 3.11. In Section 4 we give further applications of our main theorem to tensor prod- ucts with reduced group C ∗-algebras. We first observe that Z embeds unitally into r (F∞), the reduced group C ∗-algebra of the free group on infinitely many gen- C ∗ erators, see Lemma 4.1. Consequently, if a discrete group Γ contains a non-cyclic free subgroup, then Z embeds unitally into C ∗ r (Γ), see Proposition 4.2. We deduce that tensor products of the form A ⊗max C ∗ r (Γ) are singly generated if A is a separable, unital C ∗-algebra satisfying condition (2) from above, and Γ is a group containing a non-cyclic free subgroup, see Corollary 4.4. For example, r (F∞) is singly generated, although this C ∗-algebra is not Z-stable, C ∗ see Example 4.5. r (F∞) ⊗ C ∗ 2. PRELIMINARIES By a morphism between C ∗-algebras we mean a ∗-homomorphism, and by an ideal of a C ∗-algebra we understand a closed, two-sided ideal. If A is a C ∗-al- gebra, then we denote by eA its minimal unitization. Often, we write Mk for the C ∗-algebra of k-by-k matrices Mk(C). 2.1. Let A be a C ∗-algebra, and Asa ⊂ A the subset of self-adjoint elements. We say that a set S ⊂ Asa generates A, denoted A = C ∗(S), if the smallest sub-C ∗- algebra of A containing S is A itself. We denote by gen(A) the smallest number n ∈ {1, 2, 3, . . . , ∞} such that A contains a generating subset S ⊂ Asa of cardinality n, and we call gen(A) the generating rank of A. We stress that for the definition of gen(A), the generators are assumed to be self-adjoint. Two self-adjoint elements a, b generate the same C ∗-algebra as the (non-self-adjoint) element a + ib. Therefore, a C ∗-algebra A is said to be singly generated if gen(A) ≤ 2. For more details on the generating rank we refer the reader to Nagisa, [Nag04], where also the following simple facts are noted for C ∗-algebras A and B: (1) gen(eA) = gen(A), THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 5 (2) gen(C ∗(A, B)) ≤ gen(A) + gen(B), if A, B are sub-C ∗-algebras of a common C ∗-algebra, and where C ∗(A, B) denotes the sub-C ∗-algebra they generate together, (3) gen(A ⊕ B) = max{gen(A), gen(B)} if at least one of the algebras is unital. Let I ✁ A be an ideal in a C ∗-algebra A. It is easy to see that the generat- ing rank of the quotient A/I is not bigger than the generating rank of A, i.e., gen(A/I) ≤ gen(A), and the generating rank of A can be estimated as gen(A) ≤ gen(I) + gen(A/I). The following result gives an estimate for gen(I), and it is probably well-known to experts; since we could not locate it in the literature, we include a short proof. Proposition 2.2. Let A be a C ∗-algebra, and I ✁ A an ideal. Then gen(I) ≤ gen(A) + 1. Proof. We may assume gen(A) is finite. So let a1, . . . , ak be a set of self-adjoint generators for A. Then A and I are separable, and so I contains a strictly positive element h. It follows that C ∗(h) contains a quasi-central approximate unit, see [AP77, Corollary 3.3] and [Arv77]. It is straightforward to show that I is generated by the k + 1 elements h, ha1h, . . . , hakh. (cid:3) The following result is attributed to Kirchberg in [Nag04]. Theorem 2.3 (Kirchberg). Every separable, unital, properly infinite C ∗-algebra is singly generated. Proof. We sketch a proof based on the proof of [OZ76, Theorem 9]. Let A be a sep- arable, unital, properly infinite C ∗-algebra. Then there exist isometries s1, s2, . . . ∈ A with pairwise orthogonal ranges (i.e., A contains a unital copy of the Cuntz algebra O∞). Let a1, a2, . . . ∈ A be a sequence of (positive) generators for A such that their spectra satisfy σ(ak) ⊂ [1/2 · 1/4k, 1/4k]. A generator for A is given by: x :=Xk≥1 (skaks∗ k + 1/2ksk). As in in the proof of [OZ76, Theorem 9], one can show that σ(x) ⊂ {0} ∪ Sk≥1[1/2 · 1/4k, 1/4k]. Let B := C ∗(x) ⊂ A. Proceeding inductively, one shows that ak, sk ∈ B. We only sketch this for k = 1. Set p := s1s∗ 1. Let fn be a sequence of polynomials converging uniformly to 1 on [1/8, 1/4] and to 0 on [0, 1/16]. Then fn(x) converges to an element y ∈ B of the form y = p + pb(1 − p) for some b ∈ A. We compute yy∗ = p(1A + b(1 − p)b∗)p. Then for a continuous function f : R → R with f (0) = 0 and f (t) = 1 for t ≥ 1, we get f (yy∗) = p ∈ B. Then s1a1s∗ (cid:3) 1 = pxp ∈ B and s1 = 2 · px(1 − p) ∈ B, and then also a1 ∈ B. 2.4. Let X be a locally compact σ-compact Hausdorff space. A C0(X)-algebra is a C ∗-algebra A together with a morphism η : C0(X) → Z(M (A)), from the commu- tative C ∗-algebra C0(X) to the center of the multiplier algebra of A, such that for any approximate unit (uλ)Λ of C0(X), η(uλ)a → a for any a ∈ A, or equivalently, the closure of η(C0(X))A is all of A. Thus, if X is compact, then η is necessarily unital. We will usually suppress reference to the structure map, and simply write f a or f · a instead of η(f )a for the product of a function f ∈ C0(X) and an element a ∈ A. 6 HANNES THIEL AND WILHELM WINTER Let Y ⊂ X be a closed subset, and U := X \ Y its complement (an open subset). Then C0(U ) · A is an ideal of A, denoted by A(U ). The quotient A/A(U ) is denoted by A(Y ). Given a point x ∈ X, we write A(x) for A({x}), and we call this C ∗-algebra the fiber of A at x. For an element a ∈ A, we denote by a(x) the image of a in the fiber A(x). For each a ∈ A, we may consider the map a : x 7→ ka(x)k. This is a real-valued, upper-semicontinuous function on X, vanishing at infinity. The C0(X)-algebra A is called continuous if a is a continuous function for each a ∈ A. For more information on C0(X)-algebras we refer the reader to [Kas88, §1] or the more recent [Dad09, §2]. 2.5. The Jiang-Su algebra Z was constructed in [JS99]; it may be regarded as a C ∗- algebraic analog of the hyperfinite II1-factor. It can be obtained as an inductive limit of prime dimension drop algebras Zp,q := {f : [0, 1] → Mp ⊗ Mq f (0) ∈ 1p ⊗ Mq, f (1) ∈ Mp ⊗ 1q}. For more details, we refer the reader to [Win11], where Z is characterized in an entirely abstract manner, and to [Rør04] and [RW10], where it is shown that the generalized dimension drop algebra Z2∞,3∞ := {f : [0, 1] → M2∞ ⊗ M3∞ f (0) ∈ 1 ⊗ M3∞, f (1) ∈ M2∞ ⊗ 1} embeds unitally into Z; in fact, Z can be written as a stationary inductive limit of Z2∞,3∞ . 3. RESULTS Lemma 3.1. Let A be a separable, unital C ∗-algebra. Then gen(A ⊗ Z2∞,3∞) ≤ 5. Proof. Consider the ideal I := C0(0, 1) ⊗ M6∞ in B := A ⊗ Z2∞,3∞ . The quotient B/I is isomorphic to (A⊗M2∞ )⊕(A⊗M3∞). Thus, we have a short exact sequence: A ⊗ C0(0, 1) ⊗ M6∞ / A ⊗ Z2∞,3∞ / (A ⊗ M2∞) ⊕ (A ⊗ M3∞ ) It follows from [OZ76] that the tensor product of a unital, separable C ∗-alge- bra with a UHF-algebra is singly generated. In particular, gen(A ⊗ M2∞), gen(A ⊗ M3∞) ≤ 2. Thus, the quotient satisfies gen(B/I) = max{gen(A ⊗ M2∞), gen(A ⊗ M3∞)} ≤ 2, see 2.1. Note that I is an ideal in the C ∗-algebra C := A ⊗ C(S1) ⊗ M2∞ . We have gen(C) ≤ 2, and then gen(I) ≤ gen(C) + 1 ≤ 3, by Proposition 2.2. Then, the extension is generated by at most 2 + 3 = 5 self-adjoint elements. (cid:3) The following is a Stone-Weierstrass type result. We prove it using the factorial Stone-Weierstrass conjecture, which states that a sub-C ∗-algebra B ⊂ A exhausts A if it separates the factorial states of A. The factorial Stone-Weierstrass conjec- ture was proved for separable C ∗-algebras independently by Longo, [Lon84], and Popa, [Pop84]. See 2.4 for a short introduction to C0(X)-algebras. Lemma 3.2. Let A be a separable, continuous C0(X)-algebra, and B ⊂ A a sub-C ∗-al- gebra such that the following two conditions are satisfied: (i) For each x ∈ X, B exhausts the fiber A(x), (ii) B separates the points of X by full elements, i.e., for each distinct pair of points x0, x1 ∈ X there exists some b ∈ B such that b(x1) is full in B(x1) = A(x1) and b(x0) = 0. / / THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 7 Then A = B. Condition (ii) is for instance satisfied if B contains the image of the structure map η : C0(X) → Z(M (A)). Proof. Set Y := Prim(Z(M (A))), and identify Z(M (A)) with C(Y ). Let π : A → B(H) be a non-degenerate factor representation. Then π extends to a represen- It is straightforward to show π(A)′′ = π(M (A))′′, tation π : M (A) → B(H). so that π is a factor representation of M (A). For any c ∈ Z(M (A)), we have c ∈ π(A)′ ∩ π(M (A))′′ = C · 1H. Thus, there exists a point y ∈ Y such that π(c) = c(y) · 1H for all c ∈ Z(M (A)). Since η(C0(X)) contains an approximate unit for A, we have that π ◦ η is non-zero. Thus, there exists a point x ∈ X such that π ◦ η(f ) = f (x) · 1H for all f ∈ C0(X). This means that π ◦ η vanishes on the ideal A(X \ {x}), so that π factors through the fiber A(x). Let us show that B ⊂ A separates the factors states of A. So let ϕ1, ϕ2 be two different, non-degenerate factors states of A. We have shown above that there are two points x1, x2 ∈ X such that ϕi factors through A(xi), and we denote by ¯ϕi : A(xi) → C the induced factor state on A(xi), for i = 1, 2. We distinguish two cases: Case 1: x1 = x2. In this case, since ϕ1 6= ϕ2, there exists an element a ∈ A such that ϕ1(a) 6= ϕ2(a). By condition (i), there exists some element b ∈ B such that b(x1) = a(x1). Note that ϕi(b) = ¯ϕi(b(x1)) = ¯ϕi(a(x1)) = ϕi(a), for i = 1, 2. Thus, b separates the two states. Case 2: x1 6= x2. In this case, by condition (ii), there exists an element b ∈ B such that b(x2) is full in A(x2) and b(x1) = 0. Since ϕ2 6= 0, there exists an element a ∈ A such that ϕ2(a) = ¯ϕ2(a(x2)) ≥ 1. Since b(x2) is full, there exist finitely many elements gi, hi ∈ A(x2) such that ka(x2) −Pi cib(x2)dik < 1. By condition (i), there exist elements gi, hi ∈ B such that gi(x2) = gi and hi(x2) = hi. Set b′ := Pi cib di. Then ϕ2(b′) = ¯ϕ2(b′(x2)) > 0, while b′(x1) = 0. This shows that b′ separates the two states. It follows that B separates the factor states of A, and therefore B = A by the factorial Stone-Weierstrass conjecture, proved independently by Longo, [Lon84], and Popa, [Pop84]. (cid:3) Lemma 3.3. Let A be a unital C ∗-algebra with gen(A) ≤ 3. Then there exist a positive element x ∈ A ⊗ Z2,3 and two positive, full elements y′, z′ ∈ Z2,3 such that A ⊗ Z2,3 is generated by x and 1 ⊗ y′, and further y′ and z′ are orthogonal. Proof. We consider Z2,3 as the C ∗-algebra of continuous functions from [0, 1] to M6 with the boundary conditions Y f (0) =  Y   f (1) =(cid:18)Z Y QZQ∗(cid:19) , where Y ∈ M2 and Z ∈ M3 are arbitrary matrices, and Q ∈ M3 is the following fixed permutation matrix: Q =  1 1   . 1 8 HANNES THIEL AND WILHELM WINTER This means that f (0), f (1) ∈ M6 have the following form: f (0) = f (1) =   µ11 µ12 µ21 µ22 µ11 µ12 µ21 µ22 µ11 µ12 µ21 µ22     λ11 λ12 λ13 λ21 λ22 λ23 λ31 λ22 λ33   , λ33 λ31 λ32 λ13 λ11 λ12 λ23 λ21 λ22 for numbers µi,j, λi,j ∈ C. Note that Z2,3 is naturally a continuous C([0, 1])-algebra, with fibers Z2,3(0) ∼= M2, Z2,3(1) ∼= M3, and Z2,3(t) ∼= M6 for points t ∈ (0, 1) ⊂ [0, 1]. Let a, b, c ∈ A be a set of invertible, positive generators for A. Denote by ei,j the matrix units in M6. To shorten notation, for indices i, j set fi,j := ei,j + ej,i. For t ∈ [0, 1] we define the following element of A ⊗ M6: xt :=a ⊗ (e1,1 + (1 − t) · e3,3 + e5,5) +b ⊗ (f1,2 + (1 − t) · f3,4 + f5,6) +c ⊗ (e2,2 + (1 − t) · e4.4 + e6,6) +1A ⊗ (t · f2,3 + t · f4,5 + δ(t) · f1,3) where δ : [0, 1] → [0, 1] is a continuous function on [0, 1] that takes the value 0 at the endpoints 0 and 1, and is strictly positive at each point t ∈ (0, 1), e.g., δ could be given by δ(t) = 1/4 − (t − 1/2)2. We also define for t ∈ [0, 1] two elements of M6: y′ t :=e1,1 + (1 − t) · e3,3 + e5,5 z′ t :=e2,2 + (1 − t) · e4,4 + e6,6 It is easy to check that the assignment x : t 7→ xt defines an element x ∈ A⊗Z2.3. t. In matrix Similarly, we get two elements y′, z′ ∈ Z2.3 defined via t 7→ y′ form, these elements look as follows: t and t 7→ z′ xt := y′ t :=       a b δ(t) b c t δ(t) t (1 − t)a (1 − t)b (1 − t)c (1 − t)b 1 (1 − t) 1 t   t a b b c   z′ t := 1 (1 − t)   1 Set y := 1 ⊗ y′, and let D := C ∗(x + 1, y) be the sub-C ∗-algebra of E := A ⊗ Z2,3 generated by the two self-adjoint elements x + 1 and y. Since x ≥ 0, we get that both 1 and x lie in C ∗(x + 1). It follows that D = C ∗(1, x, y), and we will show that D = E. Note that E has a natural continuous C([0, 1])-algebra structure (induced THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 9 by the one of Z2,3), with fibers E(0) ∼= A ⊗ M2, E(1) ∼= A ⊗ M3, and E(t) ∼= A ⊗ M6 for points t ∈ (0, 1) ⊂ [0, 1]. Let J := E((0, 1)) ✁ E be the natural ideal corresponding to the open set (0, 1) ⊂ [0, 1]. Note that J ∼= A⊗ C0((0, 1))⊗ M6, and J is naturally a continuous C0((0, 1))- algebra. We will show in two steps that D exhausts the ideal J (i.e., D ∩ J = J) and the quotient E/J (i.e., D/(D ∩ J) = E/J). Step 1: We want to apply Lemma 3.2 to the C((0, 1))-algebra J with sub-C ∗- algebra D ∩ J. To verify condition (ii), note that the C ∗-algebra generated by y′ contains C0((0, 1)) ⊗ e3,3. Therefore, D ∩ J contains 1A ⊗ C0((0, 1)) ⊗ e3,3, which separates the points of (0, 1). Since 1A ⊗ e3,3 ∈ E(t) ∼= A ⊗ M6 is full, condition (ii) of Lemma 3.2 holds and it remains to verify condition (i). We need to show that D ∩ J exhausts all fibers of J. Fix some t ∈ (0, 1), and set Dt := C ∗(1, xt, yt) ⊂ A ⊗ M6. To simplify notation, we write ¯ei,j for the matrix units 1A ⊗ ei,j ∈ A ⊗ M6. We need to show that Dt is all of A ⊗ M6. This will follow if Dt contains all ¯ei,j, and for this it is enough to show that the off-diagonal matrix units ¯ei,i+1 are in Dt, for i = 1, . . . , 5. The spectrum of yt is {0, 1 − t, 1}. Applying functional calculus to yt we obtain that the following three elements lie in Dt: u := ¯e1,1 + ¯e5,5 v := ¯e3,3 w := 1 − v − u = ¯e2,2 + ¯e4,4 + ¯e6,6 Then, we proceed as follows: 1. ¯e1,3 = δ(t)−1uxtv ∈ Dt and so ¯e1,1, ¯e5,5 ∈ Dt. 2. g := b ⊗ e1,2 = ¯e1,1xtw ∈ Dt. It follows b ⊗ e1,1 = (gg∗)1/2 ∈ Dt, cf. [OZ76]. Then b−1 ⊗ e1,1 ∈ C ∗(b ⊗ e1,1) ⊂ Dt and so ¯e1,2 = (b−1 ⊗ e1,1) · g ∈ Dt and ¯e2,2 ∈ Dt. 3. b ⊗ e3,4 = (1 − t)−1¯e3,3xt(w − ¯e2,2) ∈ Dt. Arguing as above, it follows that ¯e3,4 ∈ Dt, and then ¯e4,4, ¯e6,6 ∈ Dt. 4. ¯e2,3 = t−1¯e2,2xt¯e3,3 ∈ Dt. 5. ¯e4,5 = t−1¯e4,4xt¯e5,5 ∈ Dt. 6. b ⊗ e5,6 = ¯e5,5xt¯e6,6 ∈ Dt and so ¯e5,6 ∈ Dt. This shows that D ∩ J exhausts the fibers of J. We may apply Lemma 3.2 and deduce D ∩ J = J, which finishes step 1. Let us denote the matrix units in M2 by e(0) M3 by e(1) 1A ⊗ e(k) Step 2: We want to show that D/J exhausts E/J = E({0, 1}) ∼= A ⊗ (M2 ⊕ M3). i,j , i = 1, 2, and the matrix units in i,j for the matrix units i,j ∈ A ⊗ (M2 ⊕ M3). Let us denote the image of x and y in D/J by v and w: 2,3 + ¯e(1) i,j , i = 1, 2, 3. To simplify notation, we write ¯e(k) 2,1) + c ⊗ (e(0) v = a ⊗ (e(0) 2,2) + ¯e(1) 2,2 + e(1) 1,2 + e(0) 2,1 + e(1) 1,2 + e(1) 3,2 0   . 0 =(cid:18)a b b w = ¯e(0) 1,1 + ¯e(1) 1 1,1 + e(1) c(cid:19) ⊕  1,1 =(cid:18)1 0 1,1) + b ⊗ (e(0) a b c b 1   0 0(cid:19) ⊕  1 10 HANNES THIEL AND WILHELM WINTER As in step 1, it is enough to show that D/J contains the off-diagonal matrix units ¯e(0) 1,2, ¯e(1) 1. g := wv(1 − w) = b ⊗ (e(0) 2,3. We argue as follows: 1,2 and ¯e(1) 1,2) ∈ D/J. As in step 1, it follows that b ⊗ (e(0) 1,2 + ¯e(1) 1,1 + 1,2 = 1,2 + e(1) 1,1) = (gg∗)1/2 ∈ D/J. Then b−1 ⊗ (e(0) e(1) 1,1 + e(1) (b−1 ⊗ (e(0) 3,3 = 1 − w − (¯e(0) 2,3 = v¯e(1) 1,1 + e(1) 1,1)) · g ∈ D/J. It follows that ¯e(0) 2,2 + ¯e(1) 3,3 ∈ D/J, and so ¯e(1) 2,2) ∈ D/J. 2,2 ∈ D/J. 1,1) ∈ D/J, and so ¯e(0) 2,2 + ¯e(1) 2,2 ∈ D/J. 2,2 ∈ D/J. Again, this implies ¯e(1) 1,2 ∈ D/J and so ¯e(1) 1,1 ∈ D/J. 1,2 = wv¯e(1) 2. ¯e(1) 3. ¯e(1) 4. b ⊗ e(1) 5. ¯e(0) 6. ¯e(0) 7. b ⊗ e(0) This finishes step 2. 1,1 = w − ¯e(1) 2,2 = 1 − w − ¯e(1) 1,1v¯e(0) 1,2 = ¯e(0) 1,1 ∈ D/J. 3,3 ∈ D/J. 2,2 − ¯e(1) 2,2 ∈ D/J. Again, this implies ¯e(0) 1,2 ∈ D/J. We have seen that A ⊗ Z2,3 is generated by x + 1 and y. Moreover, z′ is full, (cid:3) positive and orthogonal to y′. Lemma 3.4. Let A be a separable, unital C ∗-algebra. Then there exist a positive element x ∈ A ⊗ Z2∞,3∞ and two positive, full elements y′, z′ ∈ Z2∞,3∞ such that A ⊗ Z2∞,3∞ is generated by x and y := 1 ⊗ y′, and further y′ and z′ are orthogonal. Proof. Let B := A ⊗ Z2∞,3∞. Note that Z2∞,3∞ ⊗ Z2,3 is naturally a C([0, 1] × [0, 1])- algebra. Then, the quotient corresponding to the diagonal {(t, t) t ∈ [0, 1]} ⊂ [0, 1] × [0, 1] is isomorphic to Z2∞,3∞ , and we denote the resulting surjective mor- phism by π : Z2∞,3∞ ⊗ Z2,3 → Z2∞,3∞. We proceed in two steps. Step 1: We show that gen(B) ≤ k + 1 implies gen(B) ≤ k for k ≥ 2. So assume B is generated by the self-adjoint, invertible elements a1, . . . , ak+1. The sub-C ∗-al- gebra C := C ∗(ak−1, ak, ak+1) ⊂ B is unital and satisfies gen(C) ≤ 3. Consider the C ∗-algebra B ⊗ Z2,3. By Lemma 3.3, the sub-C ∗-algebra C ⊗ Z2,3 is generated by two self-adjoint elements, say b, c. One readily checks that B ⊗ Z2,3 is generated by the k self-adjoint elements a1 ⊗ 1, . . . , ak−2 ⊗ 1, b, c. Since B = A ⊗ Z2∞,3∞ is isomorphic to a quotient of B ⊗ Z2,3 = A ⊗ Z2∞,3∞ ⊗ Z2,3, we obtain gen(B) ≤ gen(B ⊗ Z2,3) ≤ k. Step 2: By Lemma 3.1, we have gen(B) ≤ 5. Applying Step 1 several times, we obtain gen(B) ≤ 3. It follows from Lemma 3.3 that there exists a positive element x ∈ B ⊗ Z2,3 and two positive, full elements y′, z′ ∈ Z2,3 such that B ⊗ Z2,3 is generated by x and 1 ⊗ y′, and further y′ and z′ are orthogonal. Consider the surjective morphism id ⊗π : A ⊗ Z2∞,3∞ ⊗ Z2,3 → A ⊗ Z2∞,3∞ . One checks that the elements x := (id ⊗π)(x) ∈ A ⊗ Z2∞,3∞ , and y′ := π(y′), z′ := π(z′) ∈ Z2∞,3∞ have the desired properties. (cid:3) Theorem 3.5. Let A, B be two separable, unital C ∗-algebras. Assume the following: (1) A contains a sequence a1, a2, . . . of full, positive elements that are pairwise or- thogonal, (2) B admits a unital embedding of the Jiang-Su algebra Z. Then A ⊗max B is singly generated. Every other tensor product A ⊗λ B is a quotient of A ⊗max B, and therefore is also singly generated. THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 11 Proof. There exists a unital embedding of Z2∞,3∞ in Z, so we may assume that there is a unital embedding of Z2∞,3∞ in B. We may assume that the elements a1, a2, . . . ∈ A are contractive. Choose a sequence b1, b2, . . . ∈ B of contractive, positive elements that is dense in the set of all contractive, positive elements of B. Consider the sub-C ∗-algebra A⊗ Z2∞,3∞ ⊂ A⊗max B. By Lemma 3.4, there exist a positive element x ∈ A ⊗ Z2∞,3∞ and two full, positive elements y′, z′ ∈ Z2∞,3∞ such that A ⊗ Z2∞,3∞ is generated by x and y := 1 ⊗ y′, and further y′ and z′ are orthogonal. Define the following two elements of A ⊗max B: v := x, w := 1 ⊗ y′ −Xk≥1 1/2k · ak ⊗ (z′bkz′). Let D := C ∗(v, w) be the sub-C ∗-algebra of A ⊗max B generated by v and w. We claim that D = A ⊗ B. Step 1: We show A ⊗ Z2∞,3∞ ⊂ D. Note that the two elements 1 ⊗ y′ and positive part of w, and therefore 1⊗y′ ∈ D. Therefore, C ∗(v, 1⊗y′) = A⊗Z2∞,3∞ ⊂ D. Pk≥1 1/2k · ak ⊗ (z′bkz′) are positive and orthogonal. It follows that 1 ⊗ y′ is the Step 2: We show 1 ⊗ B ⊂ D. We have g := Pk≥1 1/2k · ak ⊗ (z′bkz′) ∈ D. It k is full, there exist finitely many elements ci, di ∈ A such that 1A =Pi cia2 Step 1, we have ci ⊗ 1, di ⊗ 1 ∈ D. Then 1 ⊗ (z′bkz′) =Pi(ci ⊗ 1)(a2 k ⊗ (z′bkz′) = 2k · (ak ⊗ 1)g ∈ D. Since kdi. By k ⊗ (z′bkz′))(di ⊗ follows from Step 1 that ak ⊗ 1 ∈ D, and so a2 a2 1) ∈ D, for each k. Let b ∈ B be a contractive, positive element. Then b = limj bk(j) for certain in- dices k(j). Then 1 ⊗ (z′bz′) = limj 1 ⊗ (z′bk(j)z′) ∈ D. It follows that the hereditary sub-C ∗-algebra 1 ⊗ z′Bz′ is contained in D. Since z′ is full in Z2∞,3∞ , there exist finitely many elements ci, di ∈ Z2∞,3∞ such that 1B = Pi ciz′di. We have seen that 1 ⊗ z′bz′ ∈ D for any b ∈ B. Then 1 ⊗ bz′ = Pi(1 ⊗ ci)(1 ⊗ z′dibz′) ∈ D for any b ∈ B. Similarly 1 ⊗ b =Pi(1 ⊗ bciz′)(1 ⊗ di) ∈ D for any b ∈ B, as desired. It follows from Steps 1 and 2 that for each a ∈ A and b ∈ B the simple tensor a ⊗ b is contained in D. The conclusion follows since A ⊗max B is the closure of the linear span of simple tensors. (cid:3) Corollary 3.6. Let A, B be two separable, unital C ∗-algebras that both admit a unital embedding of the Jiang-Su algebra Z. Then A ⊗max B is singly generated. Proof. It is easy to verify that condition (i) of Theorem 3.5 is fulfilled if A admits a unital embedding of Z. (cid:3) Theorem 3.7. Let A be a unital, separable C ∗-algebra. Then A ⊗ Z is singly generated. Proof. Note that A ⊗ Z ∼= (A ⊗ Z) ⊗ Z. It is clear that both A ⊗ Z and Z admit unital embeddings of Z. Then apply the above Corollary 3.6. (cid:3) Corollary 3.8. Let A be a separable C ∗-algebra. Then gen(A ⊗ Z) ≤ 3. Proof. Let eA be the minimal unitization of A. gen(eA ⊗ Z) ≤ 2. Since A ⊗ Z is an ideal in eA ⊗ Z, we get gen(A ⊗ Z) ≤ gen(eA ⊗ Z) + 1 ≤ 3 from Proposition 2.2, as desired. It follows from Theorem 3.7 that (cid:3) 12 HANNES THIEL AND WILHELM WINTER Our results allow us to give new proofs for results about single generation of cer- tain von Neumann algebras. Proposition 3.9. Assume M, N are separably-acting von Neumann algebras that both admit a unital embedding of the hyperfinite II1-factor. Then M ¯⊗N is singly generated. Proof. Consider the GNS-representation π : Z → B(H) of the Jiang-Su algebra with respect to its tracial state. The weak closure, π(Z)′′, is isomorphic to the hyperfinite II1-factor R. Thus, there exists a weakly dense, unital copy of Z inside R. Choose weakly dense, separable, unital C ∗-algebras A0 ⊂ M , and similarly B0 ⊂ N . Consider Z ⊂ R ⊂ M and set A := C ∗(A0, Z) ⊂ M . Similarly set B := C ∗(B0, Z) ⊂ N . Then A and B are separable, unital C ∗-algebras that both contain unital copies of the Jiang-Su algebra. By Corollary 3.6, A ⊗max B is singly generated. Consider the sub-C ∗-algebra C := C ∗(A ¯⊗1, 1 ¯⊗B) ⊂ M ¯⊗N . Then C is a quo- tient of A ⊗max B, and therefore singly generated. Since C is weakly dense in M ¯⊗N , we obtain that M ¯⊗N is singly generated, as desired. (cid:3) Remark 3.10. We note that a von Neumann algebra M admits a unital embedding of R if and only if M has no (non-zero) finite-dimensional representations. The analogous statement for C ∗-algebras would be that a C ∗-algebra A admits a unital embedding of Z if and only if A has no (non-zero) finite-dimensional representations. It was shown by Elliott and Rørdam, [ER06], that this is true for C ∗-algebras of real rank zero. However, in [DHTW09] a simple, separable, unital, non-elementary AH-algebra is constructed into which Z does not embed. As a particular case of Proposition 3.9 we obtain the following result of Ge and Popa. Corollary 3.11 (Ge, Popa, [GP98, Theorem 6.2]). Assume M, N are separably-acting II1-factors. Then M ¯⊗N is singly generated. 4. APPLICATIONS In this section we show that the Jiang-Su algebra Z embeds unitally into the re- duced group C ∗-algebras, C ∗ r (Γ), of groups Γ that contain a non-cyclic free sub- group, see Proposition 4.2. We only consider discrete groups, and we let Fk denote the free group with k generators (k ∈ {2, 3 . . . , ∞}). We can apply Theorem 3.5 to show that certain tensor products of the form r (F∞) ⊗ r (Γ) are singly generated, see Corollary 4.4. In particular, C ∗ A ⊗max C ∗ C ∗ r (F∞) is singly generated, although it is not Z-stable, see Example 4.5. r (F∞). A key observation is that C ∗ 4.1. It was shown by Robert, [Rob10], that the Jiang-Su algebra Z embeds unitally into C ∗ r (F∞) has strict comparison of positive elements. This follows from the work of Dykema and Rørdam on reduced free product C ∗-algebras, see [DR98] and [DR00]. Dykema and Rørdam study the comparison of projections, but this can be gen- eralized to obtain results about the comparison of positive elements, as noted by Robert, [Rob10]. In particular, [DR98, Lemma 5.3] and [DR00, Theorem 2.1] can be generalized, and it follows that C ∗ r (F∞) has strict comparison of positive elements. THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 13 Proposition 4.2. If Γ is a discrete group that contains F∞ as a subgroup, then Z embeds unitally into C ∗ r (Γ). Proof. In general, for any subgroup Γ1 of a discrete group Γ, we have a unital embedding C ∗ r (Γ) contains a unital copy of C ∗ (cid:3) r (Γ). Hence, if F∞ is a subgroup of Γ, then C ∗ r (F∞), which in turn contains a unital copy of Z. r (Γ1) ⊂ C ∗ Remark 4.3. Every non-cyclic free group Fk (k ≥ 2) contains F∞ as a subgroup. In general, by the Nielsen-Schreier theorem, every subgroup of a free group is again free. Thus, if a, b are free elements, then the the elements akbk generate a subgroup Γ = hakbk, k ≥ 1i that is free, and since none of the elements akbk is contained in the subgroup generated by the other elements, we have Γ ∼= F∞. Thus, when we ask which discrete groups contain F∞ as a subgroup, we are equivalently asking which groups Γ contain a non-cyclic free subgroup. It is a necessary condition that Γ is non-amenable. The converse implication is known as the von Neumann conjecture, but this was disproved in 1980 by Ol'shanskij. A counterexample are the so-called Tarski monster groups, in which every non- trivial proper subgroup is cyclic of some fixed prime order. Clearly, such a group cannot contain F∞ as a subgroup, and it is Ol'shanskij's contribution to show that Tarski monster groups exist and are non-amenable. On the other hand, every group with the weak Powers property, as defined in [BN88], has a non-cyclic free subgroup. A proof can be found in [dlH07], which also lists classes of groups that have the (weak) Powers property. We just mention that all free products Γ1 ∗ Γ2 with Γ1 ≥ 2, Γ2 ≥ 3 have the Powers property, and therefore Proposition 4.2 applies. We may derive the following from Theorem 3.5 and Proposition 4.2: Corollary 4.4. Let A be a separable, unital C ∗-algebra that contains a countable sequence of pairwise orthogonal, full elements (e.g., A is simple and nonelementary), and let Γ be a group that contains a non-cyclic free subgroup. Then A ⊗max C ∗ r (Γ) is singly generated. r (Γ1 × Γ2) ∼= C ∗ Example 4.5. Let Γ1, Γ2 be two groups that contain non-cyclic free subgroups. Then C ∗ r (Γ2) is singly generated. For example, for any k, l ∈ {2, 3, . . . , ∞}, the C ∗-algebra C ∗ r (Fl) is singly generated. In particular, C ∗ r (F∞) is singly generated. r (F∞) ⊗max C ∗ r (Γ1) ⊗max C ∗ r (Fk) ⊗max C ∗ It was pointed out to the authors by S. Wassermann that C ∗ r (Fl) is not r (Fl) ∼= A ⊗ B ⊗ C, then Z-stable, for any k, l ∈ {2, 3, . . . , ∞}. In fact, if C ∗ one of the three algebras A, B or C is isomorphic to C. This is a generalization of the fact that C ∗ r (Fk) is tensorially prime, and it can be proved similarly. r (Fk) ⊗ C ∗ r (Fk) ⊗ C ∗ We note that it is a difficult open problem whether C ∗ r (Fk) is singly generated itself. Question 4.6. Given a non-amenable (discrete) group Γ. Does C ∗ tal embedding of Z? r (Γ) admit a uni- For each group Γ, the trivial group-morphism Γ → {1} induces a surjective mor- phism C ∗(Γ) → C. Thus, the Jiang-Su algebra can never unitally embed into a full r (Γ) ∼= C ∗(Γ), and consequently there group C ∗-algebra. If Γ is amenable, then C ∗ is no unital embedding of Z into the reduced group C ∗-algebra of an amenable group. 14 HANNES THIEL AND WILHELM WINTER On the other hand, if Γ contains a non-cyclic free subgroup, then Proposition 4.2 gives a positive answer to Question 4.6. As noted in Remark 4.3, not every non- amenable group contains a non-cyclic free subgroup. However, it is known that the reduced group C ∗-algebra of a non-amenable group has no finite-dimensional representations, which is a necessary condition for the Jiang-Su algebra to embed. ACKNOWLEDGMENTS The first named author thanks Mikael Rørdam for valuable comments, especially on the applications in Section 4. REFERENCES [AP77] C.A. Akemann and G.K. Pedersen, Ideal perturbations of elements in C ∗-algebras, Math. Scand. 41 (1977), 117 -- 139. [Arv77] W. Arveson, Notes on extensions of C ∗-algebras, Duke Math. J. 44 (1977), 329 -- 355. [BN88] F. Boca and V. Nit¸ica, Combinatorial properties of groups and simple C ∗-algebras with a unique trace, J. Oper. Theory 20 (1988), no. 1, 183 -- 196. [Dad09] M. Dadarlat, Continuous fields of C ∗-algebras over finite dimensional spaces, Adv. Math. 222 (2009), no. 5, 1850 -- 1881. [DHTW09] M. Dadarlat, I. Hirshberg, A.S. Toms, and W. Winter, The Jiang-Su algebra does not always [dlH07] [DR98] [DR00] [ER06] [ET08] [Ge03] [GP98] [JS99] [JW11] [Kad67] [Kas88] [Lon84] [LS10] embed, Math. Res. Lett. 16 (2009), no. 1, 23 -- 26. P. de la Harpe, On simplicity of reduced C ∗-algebras of groups, Bull. Lond. Math. Soc. 39 (2007), no. 1, 1 -- 26. K.J. Dykema and M. Rørdam, Projections in free product C ∗-algebras, Geom. Funct. Anal. 8 (1998), no. 1, 1 -- 16. , Projections in free product C ∗-algebras. II, Math. Z. 234 (2000), no. 1, 103 -- 113. G.A. Elliott and M. Rørdam, Perturbation of Hausdorff moment sequences, and an application to the theory of C ∗-algebras of real rank zero, Bratteli, Ola (ed.) et al., Operator algebras. The Abel symposium 2004. Proceedings of the first Abel symposium, Oslo, Norway, September 3 -- 5, 2004. Berlin: Springer. Abel Symposia 1, 97-115, 2006. G.A. Elliott and A.S. Toms, Regularity properties in the classification program for separable amenable C ∗-algebras, Bull. Am. Math. Soc., New Ser. 45 (2008), no. 2, 229 -- 245. L.M. Ge, On 'Problems on von Neumann algebras by R. Kadison, 1967', Acta Math. Sin., Engl. Ser. 19 (2003), no. 3, 619 -- 624. L. Ge and S. Popa, On some decomposition properties for factors of type II1, Duke Math. J. 94 (1998), no. 1, 79 -- 101. X. Jiang and H. Su, On a simple unital projectionless C ∗-algebra, Am. J. Math. 121 (1999), no. 2, 359 -- 413. M. Johanesova and W. Winter, The similarity problem for Z-stable C ∗-algebras, preprint, arXiv:1104.2067, 2011. R. Kadison, Problems on von Neumann algebras, unpublished manuscript, presented at Con- ference on Operator Algebras and Their Applications, Louisiana State Univ., Baton Rouge, La., 1967. G.G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), no. 1, 147 -- 201. R. Longo, Solution of the factorial Stone-Weierstrass conjecture. An application of the theory of standard split W ∗-inclusions, Invent. Math. 76 (1984), 145 -- 155. W. Li and J. Shen, A note on approximately divisible C ∗-algebras, preprint, arXiv:0804.0465, 2010. [Nag04] M. Nagisa, Single generation and rank of C ∗-algebras, Kosaki, Hideki (ed.), Operator alge- bras and applications. Proceedings of the US-Japan seminar held at Kyushu University, Fukuoka, Japan, June 7 -- 11, 1999. Tokyo: Mathematical Society of Japan. Advanced Stud- ies in Pure Mathematics 38, 135-143, 2004. C.L. Olsen and W.R. Zame, Some C ∗-algebras with a single generator, Trans. Am. Math. Soc. 215 (1976), 205 -- 217. [OZ76] THE GENERATOR PROBLEM FOR Z-STABLE C ∗-ALGEBRAS 15 [Pea62] [Pop84] [Rob10] [Rør04] [RW10] [She09] [Top68] [TW08] [vN31] [Wil74] C. Pearcy, W ∗-algebras with a single generator, Proc. Am. Math. Soc. 13 (1962), 831 -- 832. S. Popa, Semiregular maximal Abelian ∗-subalgebras and the solution to the factor state Stone- Weierstrass problem, Invent. Math. 76 (1984), 157 -- 161. L. Robert, Classification of inductive limits of 1-dimensional NCCW complexes, preprint, arXiv:1007.1964, 2010. M. Rørdam, The stable and the real rank of Z-absorbing C ∗-algebras, Int. J. Math. 15 (2004), no. 10, 1065 -- 1084. M. Rørdam and W. Winter, The Jiang-Su algebra revisited, J. Reine Angew. Math. 642 (2010), 129 -- 155. D. Sherman, On cardinal invariants and generators for von Neumann algebras, preprint, arXiv:0908.4565, 2009. D.M. Topping, UHF algebras are singly generated, Math. Scand. 22 (1968), 224 -- 226. A.S. Toms and W. Winter, Z-stable ASH algebras, Can. J. Math. 60 (2008), no. 3, 703 -- 720. J. von Neumann, no. 2, 191 -- 226. P. Willig, Generators and direct integral decompositions of W∗-algebras, Tohoku Math. J., II. Ser. 26 (1974), 35 -- 37. Uber Funktionen von Funktionaloperatoren, Ann. of Math. (2) 32 (1931), [Win07] W. Winter, Localizing the Elliott conjecture at strongly self-absorbing C ∗-algebras, preprint, arXiv:0708.0283, 2007. [Win10] [Win11] , Nuclear dimension and Z-stability of pure C ∗-algebras, preprint, arXiv:1006.2731, to appear in Inventiones, 2010. , Strongly self-absorbing C ∗-algebras are Z-stable, J. Noncommut. Geom. 5 (2011), no. 2, 253 -- 264. [Wog69] W.R. Wogen, On generators for von Neumann algebras, Bull. Am. Math. Soc. 75 (1969), 95 -- 99. DEPARTMENT OF MATHEMATICAL SCIENCES, UNIVERSITY OF COPENHAGEN, UNIVERSITETSPARKEN 5, DK-2100, COPENHAGEN Ø, DENMARK E-mail address: [email protected] MATHEMATISCHES INSTITUT DER UNIVERSIT AT M UNSTER, EINSTEINSTR. 62, 48149 M UNSTER, GERMANY E-mail address: [email protected]
1301.7347	1	1301	2013-01-30T20:12:10	C*-algebras associated with topological group quivers II: K-groups	[ "math.OA" ]	Topological quivers generalize the notion of directed graphs in which the sets of vertices and edges are locally compact (second countable) Hausdorff spaces. Associated to a topological quiver $Q$ is a $C^$-correspondence, and in turn, a Cuntz-Pimsner algebra $C^(Q).$ Given $\Gamma$ a locally compact group and $\alpha$ and $\beta$ endomorphisms on $\Gamma,$ one may construct a topological quiver $Q_{\alpha,\beta}(\Gamma)$ with vertex set $\Gamma,$ and edge set $\Omega_{\alpha,\beta}(\Gamma)= \{(x,y)\in\Gamma\times\Gamma\st \alpha(y)=\beta(x)\}.$ In \cite{Mc1}, the author examined the Cuntz-Pimsner algebra $\cO_{\alpha,\beta}(\Gamma):=C^*(Q_{\alpha,\beta}(\Gamma))$ and found generators (and their relations) of $\cO_{\alpha,\beta}(\Gamma).$ In this paper, the author uses this information to create a six term exact sequence in order to calculate the $K$-groups of $\cO_{\alpha,\beta}(\Gamma).$	math.OA	math	C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II: K-GROUPS SHAWN J. McCANN Abstract. Topological quivers generalize the notion of directed graphs in which the sets of vertices and edges are locally compact (second countable) Hausdorff spaces. Associated to a topological quiver Q is a C ∗-correspondence, and in turn, a Cuntz-Pimsner algebra C ∗(Q). Given Γ a locally compact group and α and β endomorphisms on Γ, one may construct a topological quiver Qα,β(Γ) with vertex set Γ, and edge set Ωα,β(Γ) = {(x, y) ∈ Γ × Γ(cid:12)(cid:12) α(y) = β(x)}. In [52], the author examined the Cuntz-Pimsner algebra Oα,β(Γ) := C ∗(Qα,β(Γ)) and found generators (and their relations) of Oα,β(Γ). In this paper, the author uses this information to create a six term exact sequence in order to calculate the K-groups of Oα,β(Γ). 1. Introduction and Notation 1.1. Background. Given a quintuple Q = (X, E, r, s, λ), where X and E are locally compact (second countable) Hausdorff spaces, r and s are continuous maps from X to E with r open, and λ = {λx}x∈E is a system of Radon measures, one can create a corresponding Cuntz-Pimsner C ∗-algebra C ∗(Q). In [24], Exel, an Huef and Raeburn define C ∗-algebras associated with a system (B, α, L) where α is an endomorphism of a unital C ∗-algebra B and L is a positive linear map L : B → B such that L(α(a)b) = aL(b) for all a, b ∈ B called a transfer operator. In fact, the C ∗-algebra they generate is a Cuntz-Pimsner algebra and under certain restrictions, a C ∗-algebra associated with a topological quiver; in particular, when B = C(Td) the continuous function on the d-torus, F ∈ Md(Z) and α is the endomorphism α(f )(e2πit) = f (e2πiF t) for f ∈ C(Td) and t ∈ Rd. Exel, an Huef and Raeburn then determine a six term exact sequence in which to use to calculate the K-groups of these C ∗-algebras. In [52], the author considers a certain class of topological quivers (which extend the notions of Exel, an Huef and Raeburn) Q = (Γ, Ωα,β(Γ), r, s, λ) where Γ is a locally compact group, α and β are endormorphism of Γ, Ωα,β(Γ) = {(x, y) ∈ Γ × Γ(cid:12)(cid:12) α(y) = β(x)} and λ is an appropriate family of Radon measures. The resulting Cuntz-Pimsner C ∗- algebra, denoted Oα,β(Γ), was then examined and certain generators and relations where found. We now proceed to generalize the six term exact sequence considered in [24] to C ∗-algebras of the form Oα,β(Γ) where Γ is a compact group. 1 2 SHAWN J. McCANN 1.2. Notation. The sets of natural numbers, integers, rationals numbers, real num- bers and complex numbers will be denoted by N, Z, Q, R, and C, respectively. Convention: N does not contain zero. Z+ 0 will denote the set N ∪ {0}, R+ denotes 0 = R+ ∪ {0}. Finally, Zp denotes the abelian group the set {r ∈ R(cid:12)(cid:12) r > 0} and R+ Z/pZ = {0, 1, ..., p − 1 mod p} and T denotes the torus {z ∈ C(cid:12)(cid:12) z = 1}. Whenever For a topological space Y , the closure of Y is denoted Y . Given a locally compact convenient, view Zp ⊂ T by Zp ∼= {z ∈ T(cid:12)(cid:12) zp = 1}. Hausdorff space X, let (1) C(X) be the continuous complex functions on X; (2) Cb(X) be the continuous and bounded complex functions on X; (3) C0(X) be the continuous complex functions on X vanishing at infinity; (4) Cc(X) be the continuous complex functions on X with compact support. The supremum norm is denoted · ∞ and defined by f ∞ = sup x∈X {f (x)} for each continuous map f : X → C. For a continuous function f ∈ Cc(X), denote supp f = osuppf . the open support of f by osupp f = {x ∈ X(cid:12)(cid:12) f (x) 6= 0} and the support of f by For C ∗-algebras A and B, A is isomorphic to B will be written A ∼= B; for example, we use C(Td) ⊗ MN (C) ∼= MN (C(Td)). Moreover, A⊕n denotes the n-fold direct sum A ⊕ · · · ⊕ A. Given a group Γ and a ring R, a normal subgroup, N, of Γ is denoted N ✁ Γ and an ideal, I, of R is denoted I ✁ R. Note if R is a C ∗-algebra then the term ideal denotes a closed two-sided ideal. Furthermore, End(Γ) (End(R)) and Aut(Γ) (Aut(R)) denotes the set of endomorphisms of Γ (R) and automorphisms of Γ (R), respectively. For a map γ : Γ → Aut(A), the fixed point set is denoted Aγ and defined by Let α ∈ C(X) then α# ∈ End(C(X)) denotes the endomorphism of C(X) defined Aγ = {a ∈ A(cid:12)(cid:12) γ(g)(a) = a for each g ∈ Γ}. by Let S be a set and define the Kronecker delta function δ : S × S → {0, 1} by α#(f ) = f ◦ α for each f ∈ C(X). δr s := δ(s, r) =(0 if s 6= r 1 if s = r . The set of n by n matrices with coefficients in a set R will be denoted Mn(R) and for any F ∈ Mn(R), the transpose of F is denoted F T . Given a function σ : R → S, we may create an augmented function σn : Mn(R) → Mn(S) via σn((ri,j)n i,j=1) = (σ(ri,j))n i,j=1 for each (ri,j)n i,j=1 ∈ Mn(R). Given vectors v = (v1, ..., vn) of length n and w = (w1, ..., wm) of length m, denote (v, w) to be the vector (v, w) = (v1, ..., vn, w1, ..., wm) of length n + m. C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 3 2. Preliminairies 2.1. Hilbert C ∗-modules. We begin by defining Hilbert C ∗-modules. Further de- tails and references can be found in [48, 63]. Definition 2.1. [48] If A is a C ∗-algebra, then a (right) Hilbert A-module is a Banach space EA together with a right action of A on EA and an A-valued inner product h·, ·iA satisfying (1) hξ, ηaiA = hξ, ηiAa (2) hξ, ηiA = hη, ξi∗ A (3) hξ, ξi ≥ 0 and ξ = hξ, ξi1/2 A A for all ξ, η ∈ EA and a ∈ A (if the context is clear, we denote EA simply by E). For Hilbert A-modules E and F , call a function T : E → F adjointable if there is a function T ∗ : F → E such that hT (ξ), ηiA = hξ, T ∗(η)iA for all ξ ∈ E and η ∈ F . Let L(E, F ) denote the set of adjointable (A-linear) operators from E to F . If E = F , then L(E) := L(E, E) is a C ∗-algebra (see [48].) Let K(E, F ) denote the closed two-sided ideal of compact operators given by where θE,F ξ,η (ζ) = ξhη, ζiA K(E, F ) := span{θE,F ξ,η (cid:12)(cid:12) ξ ∈ E, η ∈ F } for each ζ ∈ E. Similarly, K(E) := K(E, E) and θE ξ,η (or θξ,η if understood) denotes θE,E ξ,η . For Hilbert A-module E, the linear span of {hξ, ηi(cid:12)(cid:12) ξ, η ∈ E}, denoted hE, Ei, once closed is a two-sided ideal of A. Note that EhE, Ei is dense in E ([48]). The Hilbert module E is called full if hE, Ei is dense in A. The Hilbert module AA refers to the Hilbert module A over itself, where ha, bi = a∗b for all a, b ∈ A. An algebraic generating set for E is a subset {ui}i∈I ⊂ E for some indexing set I such that E equals the linear span of {ui · a(cid:12)(cid:12) i ∈ I, a ∈ A}. Definition 2.2. [37] A subset {ui}i∈I ⊂ E is called a basis provided the following reconstruction formula holds for all ξ ∈ E : ui · hui, ξi (in E, · .) ξ =Xi∈I If hui, uji = δj i as well, call {ui}i∈I an orthonormal basis of E. Remark 2.3. The preceding definition is in accordance with the finite version in [37], but many other versions exist such as in [24] where {ui}n i=1 is called a finite Parseval frame, or in [68] where this is taken as the definition for finitely generated. There has been substantial work done on similar frames (see [32]). The following notions of C ∗-correspondence and morphism may be found in [56, 9, 10, 11, 24, 25, 26, 39] Definition 2.4. [10, 11] If A and B are C ∗-algebras, then an A−B C ∗-correspondence E is a right Hilbert B-module EB together with a left action of A on E given by a ∗-homomorphism φA : A → L(E), a · ξ = φA(a)ξ for a ∈ A and ξ ∈ E. We may 4 SHAWN J. McCANN occasionally write, AEB to denote an A − B C ∗-correspondence and φ instead of φA. Furthermore, if A1EB1 and A2FB2 are C ∗-correspondences, then a morphism (π1, T, π2) : E → F consists of ∗-homomorphisms πi : Ai → Bi and a linear map T : E → F satisfying (i) π2(hξ, ηiA2) = hT (ξ), T (η)iB2 (ii) T (φA1(a1)ξ) = φB1(π1(a1))T (ξ) (iii) T (ξ)π2(a2) = T (ξa2) for all ξ, η ∈ E and ai ∈ Ai. Notation 2.5. When A = B, we refer to AEA as a C ∗-correspondence over A. For E a C ∗-correspondence over A and F a C ∗-correspondence over B, a morphism (π, T, π) : E → F will be denoted by (T, π). Definition 2.6. [56] If F is the Hilbert module CCC where C is a C ∗-algebra with the inner product hx, yiB = x∗y then call a morphism (T, π) : AEB → C of Hilbert modules a representation of AEB into C. Remark 2.7. Note that a representation of AEB need only satisfying (i) and (ii) of definition 2.4 as it was unnecessary to require (iii). For a proof, see [52, Remark 2.7]. A morphism of Hilbert modules (T, π) : E → F yields a ∗-homomorphism ΨT : K(E) → K(F ) by ΨT (θE ξ,η) = θF T (ξ),T (η) for ξ, η ∈ E and if (S, σ) : D → E, and (T, π) : E → F are morphisms of Hilbert modules then ΨT ◦ ΨS = ΨT ◦S. In the case where F = B a C ∗-algebra, we may first identify K(B) as B, and a representation (T, π) of E in a C ∗-algebra B yields a ∗-homomorphism ΨT : K(E) → B given by ΨT (θξ,η) = T (ξ)T (η)∗. Definition 2.8. [56] For a C ∗-correspondence E over A, denote the ideal φ−1(K(E)) of A by J(E), and let JE = J(E) ∩ (ker φ)⊥ where (ker φ)⊥ is the ideal {a ∈ A(cid:12)(cid:12) ab = If AEA and BFB are C ∗-correspondences over A and B re- 0 for all b ∈ ker φ} . spectively and K ✁ J(E), a morphism (T, π) : E → F is called coisometric on K if for all a ∈ K, or just coisometric, if K = J(E). ΨT (φA(a)) = φB(π(a)) Notation 2.9. We denote C ∗(T, π) to be the C ∗-algebra generated by T (E) and π(A) where (T, π) : E → B is a representation of AEA in a C ∗-algebra B. Further- more, if ρ : B → C is a ∗-homomorphism of C ∗-algebras, then ρ ◦ (T, π) denotes the representation (ρ ◦ T, ρ ◦ π) of E. Definition 2.10. [56] A morphism (TE , πE) coisometric on an ideal K is said to be universal if whenever (T, π) : E → B is a representation coisometric on K, there exists a ∗-homomorphism ρ : C ∗(TE , πE) → B with (T, π) = ρ ◦ (TE, πE). The universal C ∗-algebra C ∗(TE, πE ) is called the relative Cuntz-Pimsner algebra of E C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 5 determined by the ideal K and denoted by O(K, E). If K = 0, then O(K, E) is denoted by T (E) and called the universal Toeplitz C ∗-algebra for E. We denote O(JE, E) by OE . 2.2. Topological Quivers. Definition 2.11. [56] A topological quiver (or topological directed graph) Q = (X, E, Y, r, s, λ) is a diagram s .. ...................................................................................... ......... ... ..r .......... ..................................................................................... ... Y E X where X, E, and Y are second countable locally compact Hausdorff spaces, r and s are continuous maps with r open, along with a family λ = {λyy ∈ Y } of Radon measures on E satisfying (1) supp λy = r−1(y) for all y ∈ Y , and (2) y 7→ λy(f ) =RE f (α)dλy(α) ∈ Cc(Y ) for f ∈ Cc(E). Remark 2.12. If X = Y then write Q = (X, E, r, s, λ) in lieu of (X, E, X, r, s, λ). Remark 2.13. The author provides a broad history and a series of examples of topological quivers in [51, 52]. Given a topological quiver Q = (X, E, Y, r, s, λ), one may associate a correspon- dence EQ of the C ∗-algebra C0(X) to the C ∗-algebra C0(Y ). Define left and right actions (a · ξ · b)(e) = a(s(e))ξ(e)b(r(e)) by C0(X) and C0(Y ) respectively on Cc(E). Furthermore, define the Cc(Y )-valued inner product hξ, ηi(y) =Zr−1(y) ξ(α)η(α)dλy(α) for ξ, η ∈ Cc(E), y ∈ Y, and let EQ be the completion of Cc(E) with respect to the norm ξ = hξ, ξi1/2∞ = λy(ξ2)1/2 ∞ . Definition 2.14. Given topological quiver Q over a space X, define the C ∗-algebra, C ∗(Q) associated with Q to be the Cuntz-Pimnser C ∗-algebra OEQ of the correspon- dence EQ over A = C0(X). 2.3. Topological Group Quivers. Definition 2.15. Let Γ be a (second countable) locally compact group and let α, β ∈ End(Γ) be continuous. Define the closed subgroup, Ωα,β(Γ), of Γ × Γ, and let Qα,β(Γ) = (Γ, Ωα,β(Γ), r, s, λ) where r and s are the group homomorphisms defined by Ωα,β(Γ) = {(x, y) ∈ Γ × Γ(cid:12)(cid:12) α(y) = β(x)} for each (x, y) ∈ Ωα,β(Γ) and λx for x ∈ Γ is the measure on r(x, y) = x and s(x, y) = y r−1(x) = {x} × α−1(β(x)) 6 SHAWN J. McCANN defined by λx(B) = µ(y−1s(B ∩ r−1(x)) ∩ ker α) (for any y ∈ α−1(β(x))) for each measurable B ⊆ Ωα,β(Γ) where µ is a left Haar measure (normalized if possible) on r−1(1Γ) = {1} × ker α (a closed normal subgroup of Γ × Γ; hence, a locally compact group). Note if r−1(x) = ∅ then α−1(β(x)) = ∅ and so λx = 0. This measure is well-defined, supp λx = {x} × y ker α = {x} × α−1(β(x)) = r−1(x) and y 7→ λy(f ) is a continuous compactly supported function (cf. 3.1]. [52, Definition Call Qα,β(Γ) a topological group relation. Define Eα,β(Γ) to be the C0(Γ)-correspondence EQα,β(Γ) and form the Cuntz-Pimsner algebra Oα,β(Γ) := C ∗(Qα,β(Γ)) = O(JEα,β (Γ), Eα,β(Γ)) and the Toeplitz-Pimsner algebra Tα,β(Γ) := T (Qα,β(Γ)). Remark 2.16. It will be implicitly assumed that Γ is second countable. Further- more, since Γ is locally compact Hausdorff, r−1(x) is closed and locally compact. Moreover, whenever r is a local homeomorphism, r−1(x) is discrete and hence, λx is counting measure (normalized when ker α < ∞.) Example 2.17 ([52]). For the compact abelian group Td, note End(Td) ∼= Md(Z) ([67]); that is, an element σ ∈ End(Td) is of the form σF for some F ∈ Md(Z) where σF (e2πit) = e2πiF t for each t ∈ Zd. To simplify notation, use F and G in place of σF and σG whenever convenient. For instance, and the C ∗-correspondence QF,G(Td) := QσF ,σG(Td) EF,G(Td) := EσF ,σG(Td) where F, G ∈ Md(Z). We will consider the cases when these maps are surjective; that is, det F and det G are non-zero. Let F, G ∈ Md(Z) where det F, det G 6= 0. Then ker σF = det F and so, the C(Td)-valued inner product becomes hξ, ηi(x) = 1 det F XσF (y)=σG(x) ξ(x, y)η(x, y) for ξ, η ∈ EF,G(Td) and x ∈ Td. This is a finite sum since the number of solutions, y, to σF (y) = σG(x) given any x ∈ Td is det F < ∞. C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 7 Remark 2.18. The left action, φ, is defined by φ(a)ξ(x, y) = a(y)ξ(x, y) for a ∈ C(Td), ξ ∈ C(ΩF,G(Td)) and (x, y) ∈ ΩF,G(Td). Note: φ is injective. To see this claim, let a ∈ C(Td) and assume φ(a)ξ = 0 for each ξ ∈ C(ΩF,G(Td)). Then a(y)ξ(x, y) = 0 for each (x, y) ∈ ΩF,G(Td) and ξ ∈ C(ΩF,G(Td)). Since s(ΩF,G(Td)) = {y ∈ Td(cid:12)(cid:12) (x, y) ∈ ΩF,G(Td)} = Td by the surjectivity of σF , a = 0. Remark 2.19. It was shown in [52] that one may assume the matrix F is positive diagonal. Let F = Diag(a1, ..., ad) ∈ Md(Z), G = (bjk)d j = 1, ..., d, det G 6= 0 and let Gj denote the j-th row of G, (bjk)d j,k=1 ∈ Md(Z) where aj > 0 for each k=1. Further, let N = det F =Qd j=1 aj > 0 and let I(F ) = {ν = (νj)d The C(Td)-valued inner product becomes j=1 ∈ Zd(cid:12)(cid:12) 0 ≤ νj ≤ aj − 1}. N XσF (y)=σG(x) ξ(x, y)η(x, y) 1 hξ, ηi(x) = for all ξ, η ∈ C(ΩF,G(Td)) and x ∈ Td. Given ν ∈ I(F ), define uν ∈ C(ΩF,G(Td)) by uν(x, y) = yν = yνj d Yj=1 for (x, y) ∈ ΩF,G(Td). It was shown in [52] that {uν}ν∈I(F ) is a basis for EF,G(Td) and also the following: Theorem 2.20. [52, Theorem 3.23] Let F = Diag(a1, ..., ad), G ∈ Md(Z) where det F, det G 6= 0 and let Gj be the j-th row vector of G. Further, let I(F ) denote the set {ν = (νj)d algebra generated by isometries {Sν}ν∈I(F ) and (full spectrum) commuting unitaries {Uj}d j=1 ∈ Zd(cid:12)(cid:12) 0 ≤ νj ≤ aj − 1}. Then OF,G(Td) is the universal C ∗- j=1 that satisfy the relations (1) S∗ (2) U νS = Sν for all ν ∈ I(F ), (3) U aj ν Sν ′ = huν, uν ′i = δν ′ ν , j S = SU Gj , for all j = 1, ..., d and (4) 1 =Pν∈I(F ) SνS∗ where U ν denotesQd j=1 U νj ated by isometries {Sν}ν∈I(F ) and commuting unitaries {Uj}d (1)-(3) j . Furthermore, Tα,β(Γ) is the universal C ∗-algebra gener- j=1 that satisfy relations ν =Pν∈I(F ) U νSS∗U −ν 8 SHAWN J. McCANN 3. Six Term Exact Sequence for Oα,β(Γ) In this section, we follow and extend the approach of [24] to create a six term exact sequence. Let Γ be a compact group with α, β ∈ End(Γ). Suppose the left action for the correspondence, φ, is injective where φ is defined by φ(a)ξ(x, y) = a(y)ξ(x, y) for a ∈ C(Γ), ξ ∈ C(Ωα,β(Γ)) and (x, y) ∈ Ωα,β(Γ). Furthermore, we shall assume the existence of an orthonormal basis (see Defintion 2.2) {ui}N −1 i=0 for Eα,β(Γ). In order to construct our exact sequence for K∗(Oα,β(Γ)), note the short exact sequence 0 .......... ........................................................................... ... .. ker q ..ι .......... ........................................................................ ... Tα,β(Γ) ..q .......... ............................................................................. ... Oα,β(Γ) .......... ................................................................................ ... .. 0, where q : Tα,β(Γ) → Oα,β(Γ) is the canonical quotient map and ι : ker q → Tα,β(Γ) is the inclusion homomorphism, induces the six-term exact sequence of K-groups (see [65]) (3.1) K0(ker q) δ1 .. . . . . . . . . . . .. . .. . . . . . . . ... ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ι∗ ......... ............................................................................................................... ... .. K1(Oα,β(Γ)) ............................................................................................. .......... .. .. q∗ K0(Tα,β(Γ)) K1(Tα,β(Γ)) q∗ ... ......... ............................................................................................. .. K0(Oα,β(Γ)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . .. .. . . . . . . . . . . . . . . .. .δ0 K1(ker q) ............................................................................................................... .......... .. .. ι∗ Let (T, π) denote the universal Toeplitz representation on Eα,β(Γ); that is, π = q◦π is the morphism C(Γ) → Oα,β(Γ). As shown in [60, Theorem 4.4], the homomor- phism π : C(Γ) → Tα,β(Γ) induces an isomorphism of Ki(C(Γ)) onto Ki(Tα,β(Γ)). Thus we may replace Ki(Tα,β(Γ)) with Ki(C(Γ)) provided we can identify the re- sulting maps. We intend to show that (3.1) induces the six-term exact sequence (3.2) K0(C(Γ)) .. . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 − Ω∗ ... ......... .............................................................................................................................. .. K1(Oα,β(Γ)) ............................................................................................................ .......... .. .. π∗ K0(C(Γ)) K1(C(Γ)) π∗ ... ......... ............................................................................................................ .. K0(Oα,β(Γ)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .. .. . . . . . . . . . . . . . . .. .. .. K1(C(Γ)) .. .............................................................................................................................. .......... .. 1 − Ω∗ for π = q ◦ π : C(Γ) → Oα,β(Γ) and an appropriately chosen homomorphism Ω : C(Γ) → MN (C(Γ)). Lemma 3.1. Define Ω : C(Γ) → MN (C(Γ)) by Ω(a) = (hui, a · uji)N −1 i,j=0. Then Ω is a unital homomorphism and Ω(α#(a)) is the diagonal matrix β#(a)1N for all a ∈ C(Γ). C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 9 Proof. Let a, b ∈ C(Γ). Then the (i, j)-entry of Ω(a)Ω(b) is N −1 (Ω(a)Ω(b))i,j = hui, a · ukihuk, b · uji Xk=0 = hui, a · (Xk = hui, a · (b · uj)i = Ω(ab)i,j. uk · huk, b · uji)i Furthermore, for a∗ denoting the map a∗(x) = a(x) for x ∈ Γ, Ω(a∗) = (hui, a∗ · uji)i,j = (ha · ui, uji)i,j = (huj, a · uii∗)i,j = Ω(a)∗ and Finally, let x ∈ Γ. Then Ω(1) = (hui, uji)i,j = (δj i )i,j = 1N . Ω(α#(a))i,j(x) = hui, α#(a) · uji(x) ui(e)a(α(s(e)))uj(e) dλx(e) ui(e)a(β(x))uj(e) dλx(e) ui(e)uj(e) dλx(e) =Zr−1(x) =Zr−1(x) = a(β(x))Zr−1(x) = a(β(x))hui, uji(x) = δj i β#(a)(x); hence, Ω(α#(a)) = β#(a)1N . (cid:3) In order to describe ker q, use the notation E ⊗k := Eα,β(Γ)⊗k for the k-fold internal tensor product of C ∗-correspondences ([48]) Eα,β(Γ) ⊗ · · · ⊗ Eα,β(Γ), which is itself a C ∗-correspondence over A = C(Γ). For the universal covariant representation (T, π) : Eα,β(Γ) → Tα,β(Γ) (that is, q(Tj) is the isometry Sj with T (uj) = Tj,) there is, in i=1 T (ξi) for all elementary tensors ξ = ξ1 ⊗ · · · ⊗ ξk where ξi ∈ Eα,β(Γ) (see [27, Proposition 1.8] where the term "Hilbert bimodule" is used instead of C ∗-correspondence.) Note Eα,β(Γ)⊗0 := C(Γ) and T ⊗0 := π. By [27, Lemma 2.4], fact, a Toeplitz representation (T ⊗k, π) of Tα,β(Γ) such that T ⊗k(ξ) = Qk Tα,β(Γ) = span{T ⊗k(ξ)T ⊗k′ j . The proceeding lemmas and propositions are essentially those found in [24, Lemma 3.2, Lemma 3.3 & Proposition 3.4] with some changes. j=0 TjT ∗ Next, let p = PN −1 (η)∗(cid:12)(cid:12) k, k′ ≥ 0, ξ ∈ E ⊗k, η ∈ E ⊗k′ }. Lemma 3.2. With the preceding notation: (1) p is a projection which commutes with π(a) for all a ∈ C(Γ) 10 SHAWN J. McCANN (2) 1 − p is a full projection in ker q (3) (1 − p)T ⊗k(ξ) = 0 for all ξ ∈ E ⊗k and k ≥ 1 (4) ker q = span{T ⊗k(ξ)(1 − p)T ⊗k′(η)∗(cid:12)(cid:12) k, k′ ≥ 0, ξ ∈ E ⊗k, η ∈ E ⊗k′} i Tj = π(hui, uji) = δj Proof. (1) Recall that T ∗ more, i . Thus, p2 = p and p∗ = p. Further- pπ(a)p = = = and so, N −1 N −1 Xj,k=1 Xj,k=0 Xk=0 N −1 TjT ∗ j π(a)TkT ∗ k = N −1 Xj,k=0 T (ujhuj, a · uki)T ∗ k = Tj π(huj, a · uki)T ∗ k N −1 Xk=0 T ( N −1 Xj=0 ujhuj, a · uki)T ∗ k T (a · uk)T ∗ k = π(a)p pπ(a) = (π(a)∗p)∗ = (pπ(a)∗p)∗ = pπ(a)p = π(a)p. (2) Recall φ(a) =PN −1 j=0 θa·uj ,uj , so ΨT (φ(a)) = T (a · uj)T (uj)∗ = π(a)p N −1 Xj=0 and q(1 − p) = q(π(1) − π(1)p) = q(π(1) − ΨT (φ(1))) = 0. Hence, 1 − p = π(1) − π(1)p ∈ ker q and since ker q is the ideal in Tα,β(Γ) generated by {π(a) − ΨT (φ(a))(cid:12)(cid:12) a ∈ C(Γ)} and 1 − p ∈ ker q, ker q is the ideal generated by {π(a)(1 − p)(cid:12)(cid:12) a ∈ C(Γ)}. Hence 1 − p is full. (3) Let ξ ∈ Eα,β(Γ) then N −1 N −1 pT (ξ) = Xj=0 T (uj)T (uj)∗T (ξ) = Xj=0 T (ujhuj, ξi) = T (ξ). Thus, (1 − p)T (ξ) = 0. Now for k > 1, let ξ = ξ1 ⊗ ... ⊗ ξk. Then (1 − p)T ⊗k(ξ) = (1 − p) T (ξj) = 0 k Yj=0 and hence, by linearity and continuity, (3) has been proven. (4) Since ker q = Tα,β(Γ)(1−p)Tα,β(Γ), the description of Tα,β(Γ) preceding Lemma 3.2 paired with (3) gives the desired result. (cid:3) Lemma 3.3. There exists a homomorphism ρ : C(Γ) → ker q ⊂ Tα,β(Γ) such that ρ(a) = π(a)(1 − p) and ρ is an isomorphism of C(Γ) onto the full corner C ∗-algebra C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 11 (1 − p)ker q(1 − p). Proof. By the previous lemma, (1 − p)π(a)(1 − p) = π(a)(1 − p) ∈ ker q. Thus, ρ(a) = π(a)(1 − p) defines a homomorphism ρ : C(Γ) → (1 − p) ker q(1 − p) ⊂ ker q. Using the previous lemma, (1 − p) ker q(1 − p) = span{(1 − p)T ⊗k(ξ)(1 − p)T ⊗k′ (η)∗(1 − p)(cid:12)(cid:12) k, k′ ≥ 0, ξ ∈ E ⊗k, η ∈ E ⊗k′ } = span{(1 − p)π(a)(1 − p)π(b)∗(1 − p)(cid:12)(cid:12) a, b ∈ C(Γ)} = span{π(a)(1 − p)(cid:12)(cid:12) a ∈ C(Γ)} = ran ρ. Hence, ρ is surjective. In order to show the injectivity of ρ, choose a faithful representation π0 : C(Γ) → B(H) and consider the Fock representation (TF , πF ) of Eα,β(Γ) induced from π0 as described in [27, Example 1.4]. The underlying space of this Fock representation is F (Eα,β(Γ)) ⊗A H := ⊕k≥0(E ⊗k ⊗A H) where A = C(Γ) acts diagonally on the left and Eα,β(Γ) acts by creation operators. Then TF (ξ)∗ is an annihilation operator vanishing on the subspace A ⊗A H of F (Eα,β(Γ)) ⊗A H. Now, for a ∈ A, 0 = (TF × πF )(ρ(a)) = (TF × πF )(π(a)(1 − p) = πF (a)(1 − TF (uj)TF (uj)∗). N −1 Xj=0 Since TF (uj)∗ vanishes on A ⊗A H, we have that ρ(a) = 0 implies πF (a)(1 − TF (uj)TF (uj)∗)(1 ⊗A h) = 0 N −1 Xj=0 for all h ∈ H and so, πF (a)(1 ⊗A h) = 0 for all h ∈ H. Thus, a ⊗A h = 0 for all h ∈ H and hence, π0(a)h = 0 for all h ∈ H which implies a = 0 since π0 is faithful. Hence, ρ is injective. Lemma 3.4. [24, Lemma 3.5] Suppose that A is a C ∗-algebra, r ≥ 1 and N ≥ 2 are integers, and (cid:3) is a subset of A. For m, n satisfying 0 ≤ m, n < rN − 1, define {bj,s;k,t(cid:12)(cid:12) 0 ≤ j, k < N and 0 ≤ s, t < r} cm,n = bj,s;k,t where m = sN + j and n = lN + k, and dm,n = bj,s;k,t where m = jr + s and n = kr + t. Then there is a scalar unitary permutation matrix U such that the matrices C := (cm,n)m,n and Dm,n := (dm,n)m,n are related by C = UDU ∗. The following is standard (and also appears in [24]): Lemma 3.5. Suppose that S is an isometry in a unital C ∗-algebra A. Then is a unitary element of M2(A) and its class in K1(A) is the identity. S∗ (cid:19) U :=(cid:18)S 1 − SS∗ 0 12 SHAWN J. McCANN Proposition 3.6. Let (T, π) denote the universal Toeplitz representation on Eα,β(Γ) and let {uj}N −1 j=0 TjT ∗ j where Tj = T (uj). Then, with the maps Ω : C(Γ) → MN (C(Γ)) and ρ : C(Γ) → ker q ⊂ Tα,β(Γ) defined by j=0 be an orthonormal basis of Eα,β(Γ). Further, let p = PN −1 and Ω(a) = (hui, a · uji)N −1 i,j=0 ρ(a) = π(a)(1 − p) as in Lemmas 3.1 and 3.3, the following two diagrams (i = 0, 1) commute: (3.3) Ki(C(Γ)) ρ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . .. .. . . . . . . . . . . . . . . . . . .. Ki(ker q) 1 − Ω∗ ........... ............................................................................................................................... . . ι∗ ........ .................................................................................................................... . ... .. Ki(C(Γ)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . .. .. . . . . . . . . . . . . . . . . . .. π∗ Ki(Tα,β(Γ)) Proof. First, let i = 0. Let z = (zs,t) ∈ Mr(C(Γ)) be a projection and let πr denote the augmentation map, π ⊗ idr, of π on Mr(C(Γ)). Then ρ∗([z]) = [(ρ(zs,t))s,t] = [(π(zs,t)(1 − p))s,t] = [πr(z)] − [πr(z)(p1r)] and Hence, it suffices to show that π∗ ◦ (1 − Ω∗)([z]) = [πr(z)] − π∗ ◦ Ω∗([z]). π∗ ◦ Ω∗([z]) = [πr(z)(p1r)]. Note that so Ω∗([z]) = [(Ω(zs,t))s,t] = [((huj, zs,t · uki)j,k)s,t], π∗ ◦ Ω∗([z]) = [((π(huj, zs,t · uki))j,k)s,t] = [πrN ◦ Ωr(z)]. Set bj,s;k,t = π(huj, zs,t · uki) and C = (cm,n)m,n = πrN (Ωr(z)) as in Lemma 3.4. Let T01r T11r 0 ... 0 0 ... 0 T =  ... ... ... TN −11r ... 0 ... 0   ∈ MN (Mr(Oα,β(Γ))). Then T T ∗ = p1r ⊕ 0r(N1) and since πr(z) is a projection which commutes with p1r, (πr(z) ⊕ 0r(N −1))T is a partial isometry which implements a Murray-von Neumann equivalence between T ∗(πr(z) ⊕ 0r(N −1))T and thus, (πr(z) ⊕ 0r(N −1))T T ∗(πr(z) ⊕ 0r(N −1)) = πr(z)(p1r) ⊕ 0r(N −1); [πr(z)(p1r)] = [T ∗(πr(z) ⊕ 0r(N −1))T ]. C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 13 Furthermore, T ∗(πr(z) ⊕ 0r(N −1))T = T ∗  πr(z)T0 ... πr(z)TN −1 ... 0 ... ... 0 ...   = (T ∗ j πr(z)Tk)j,k so the (j, k) entry is (π(huj, zs,t · uki))s,t. Recall bj,s;k,t = π(huj, zs,t · uki) and so T ∗(πr(z) ⊕ 0r(N −1))T = D = (dm,n)m,n as in Lemma 3.4. Thus, by Lemma 3.4, there exists a unitary U such that C = U ∗DU which gives us [πr(z)(p1r)] = [D] = [C] = [πrN ◦ ΩR(z)] as desired. For the case i = 1, let u ∈ Mr(C(Γ)) be a unitary. Note ρ∗ : K1(C(Γ)) → K1(ker q) is the composition of a unital isomorphism of C(Γ) onto (1 −p) ker q(1 −p) with the inclusion of (1−p) ker q(1−p) as a full corner in the non-unital algebra ker q; that is, [u] 7→ [ρr(u)] = [πr(u)((1−p)1r)] 7→ [πr(u)((1−p)1r)+p1r] ∈ K1((ker q)+) = K1(ker q). Furthermore, π∗ ◦ Ω∗([u]) = [πr(u)] − [πrN ◦ Ωr(u)] and hence, we need only show [(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1)] = [πr(u) ⊕ 1r(N −1)] − [πrN ◦ Ωr(u)] in K1(Tα,β(G)). We take a brief moment to make an aside: If C ∈ M2rN (Tα,β(Γ)) is invertible with K1-class the identity 1 then the K1-class is unchanged by pre- and post- multiplication by C. In particular, when C is equal to: (1) (Lemma 3.5) a unitary of the form S∗ (cid:19) (cid:18)S 1 − SS∗ 0 where S is an isometry (2) an upper- or lower-triangular matrix of the form (cid:18)1 A 0 1(cid:19) or (which are connected to 1 via t 7→(cid:18)1 tA 0 (3) any constant invertible matrix in GL2rN (C) (because GL2rN (C) is path con- nected); this implies that row and column operations may be used without changing the K1-class. (cid:18) 1 0 A 1(cid:19) 1 (cid:19) and likewise for the transpose) Recall: T =  T01r T11r 0 ... 0 0 ... 0 ... ... ... TN −11r ... 0 ... 0 .   14 SHAWN J. McCANN With this in mind, calculate [(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1)] 0rN =h(cid:18)(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1) 0rN =h(cid:18)(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1) 0rN =h(cid:18)((πr(u)(1 − p)1r) + p1r) ⊕ 1r(N −1))T π(u)((1 − p)1r) ⊕ 1r(N −1) 1rN(cid:19)ih(cid:18) T 1rN(cid:19)ih(cid:18) T 1rN − T T ∗ 0rN 0rN 0rN 0rN T ∗ T ∗ T ∗ (cid:19)i (cid:19)i (1 − p)1r ⊕ 1r(N −1) and recall (1 − p)Ti = 0 by Lemma 3.2(3), hence (1 − p)1rT = 0 and (cid:19)i [(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1)] 0rN =h(cid:18) T =h(cid:18) T =h(cid:18) T 0rN 0rN πr(u)((1 − p)1r) ⊕ 1r(N −1) T ∗ πr(u)((1 − p)1r) ⊕ 1r(N −1) T ∗ (cid:19)i (cid:19)ih(cid:18)1rN T ∗(πr(u) ⊕ 1r(N −1)) 1rN 0rN (cid:19)i πr(u) ⊕ 1r(N −1) T ∗ (cid:19)i since T T ∗ = p1r ⊕0r(N −1) and (p1r)πr(u) = πr(u)(p1r). Using elementary operations, compute [(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1)] T T 0rN(cid:19)i 0rN(cid:19)ih(cid:18)1rN −(πr(u∗) ⊕ 1r(N −1))T −T ∗(πr(u∗) ⊕ 1r(N −1))T(cid:19)i 1rN 0rN 0rN (cid:19)i T ∗ T ∗ =h(cid:18)πr(u) ⊕ 1r(N −1) =h(cid:18)πr(u) ⊕ 1r(N −1) =h(cid:18)πr(u) ⊕ 1r(N −1) =h(cid:18) =h(cid:18)πr(u) ⊕ 1r(N −1) 0rN 1rN T ∗ 0rN −T ∗(πr(u∗) ⊕ 1r(N −1)) 1rN(cid:19)ih(cid:18)πr(u) ⊕ 1r(N −1) −T ∗(πr(u∗) ⊕ 1r(N −1))T(cid:19)ih(cid:18)1rN −T ∗(πr(u∗) ⊕ 1r(N −1))T(cid:19)i 0rN −1rN(cid:19)i 0rN 0rN 0rN T ∗ = [πr(u) ⊕ 1r(N −1)] + [T ∗(πr(u∗) ⊕ 1r(N −1))T ]. Furthermore, [T ∗(πr(u∗) ⊕ 1r(N −1))T ] = [πrN (Ωr(u−1))] = −[πrN (Ωr(u))]. Hence, [(πr(u)((1 − p)1r) + p1r) ⊕ 1r(N −1)] = [πr(u) ⊕ 1r(N −1)] − [πrN ◦ Ωr(u)] as desired. (cid:3) C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 15 Theorem 3.7. Let (S, π) = q ◦ (T, π) be the universal Cuntz-Pimsner covariant representation of Eα,β(Γ) in Oα,β(Γ). Then the following diagram is exact: (3.4) .. . . . . . . . . . . . . . . . . . . . . . . ... ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ρ−1 ∗ ◦ δ0 K1(Oα,β(Γ)) K0(C(Γ)) 1 − Ω∗ . .. ......... .............................................................................................................................. .. K0(C(Γ)) π∗ . .. ......... ............................................................................................................ .. K0(Oα,β(Γ)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . .. .. . . . . . . . . . . . . . . .. .. ..ρ−1 ∗ ◦ δ1 ... ........................................................................................................... .......... .. π∗ K1(C(Γ)) ... ............................................................................................................................. .......... .. 1 − Ω∗ K1(C(Γ)) Proof. Note ρ : C(Γ) → ker q is an isomorphism onto a full corner, implying ρ∗ is an isomorphism. Further note π∗ : Ki(C(Γ)) → Ki(Tα,β(Γ)) is an isomorphism (see comments prior to Lemma 3.1). Then (3.1) and the previous proposition give the stated result. (cid:3) 4. K-groups of OF,G(Γ) In this section, the approach of [24] is made easier and extended. For this section, let α, β ∈ End(C(Td)) defined by α = σ# G where F = Diag(a1, ..., ad) ∈ Md(Z)+ and G ∈ Md(Z) such that det F > 0 and det G 6= 0. We know there exists F and β = σ# an orthonormal basis for EF,G(Td), {uj(cid:12)(cid:12) 0 ≤ j ≤ N − 1}; this is the basis {uν}ν∈I(F ), described in Section 3.3, reindexed by 0, 1, ..., N − 1. Let Uj be the unitary defined by Uj(x) = xj for x = (xi)d i=1 ∈ Td for j ∈ {1, ..., d}. Further, let and J ′ = {1, ..., d} \ J in increasing order. Define {[1]0} Ik = {J ⊂ {1, ..., d}(cid:12)(cid:12) J = k, J = {j1 < ... < jk}} {[UJ ]0 = [Uj1]0 ∧ ... ∧ [Ujk]0(cid:12)(cid:12) J ∈ Ik} if k > 0 is even {[UJ ]1 = [Uj1]1 ∧ ... ∧ [Ujk]1(cid:12)(cid:12) J ∈ Ik} if k > 0 is odd if k = 0 Ek =  If it is understood, the notation [·] will be used in lieu of [·]i. It is well known (see [34] and [33, Example 3.11 and 3.15]) that with basis {Ek}k even and K0(C(Td)) ∼= ^evens K1(C(Td)) ∼= ^odds Zd = Z2d−1 Zd = Z2d−1 with basis {Ek}k odd. For subsets J and I of the same size, define FJ,I to be the square submatrix of F whose entries belong to the rows in J and the columns in I. j=1 → j=1 → span{[Uj]}d With these identifications, the (K1-group) induced map α∗V1 Zd : span{[Uj]}d j=1 is multiplication by F T = F, and β∗V1 Zd : span{[Uj]}d j=1 span{[Uj]}d is multiplication by GT , the transpose of G. We have U bjk k β∗([Uj]) = [β(Uj)] = [U Gj ] = [Yk bjk[Uk]. ] =Xk 16 SHAWN J. McCANN Do likewise to prove α∗V1 Zd is multiplication by F. One can also check that α∗ and α∗[1] = [α(1)] = 1 = [β(1)] = β∗[1] since α and β are group homomorphisms. β∗ act on V0 Zd by Lemma 4.1. For 1 ≤ k ≤ d, the matrix Ak representating α∗ : Vk Zd → Vk Zd with respect to the basis Ek is the diagonal matrix Diag(aI)I∈Ik (aI =Qi∈I ai) and matrix Bk representating β∗ :Vk Zd →Vk Zd is (det GJI)I,J∈Ik. Proof. Begin by noting, Ak = Vk F T and Bk = Vk GT . Let [UI] ∈ Ek with I = {i1 < ... < ik}. Then k k β∗([UI]) = ( = GT [Ui1] ∧ ... ∧ GT [Uik ] ^ GT )([Ui1] ∧ ... ∧ [Uik ]) d = = ^ GT )[UI] = ( Xm1,...,mk=1 X[UJ ]∈Ek,J={m1,...,mk} = X[UJ ]∈Ek Xσ∈Sk = X[UJ ]∈Ek Xσ∈Sk = X[UJ ]∈Ek det GI,J[UJ ] bi1,m1...bik,mk ([Um1] ∧ ... ∧ [Umk ]) bi1,m1...bik,mk[UJ ] bi1,σ(j1)...bik,σ(jk)([Uσ(j1)] ∧ ... ∧ [Uσ(jk)]) (−1)deg σbi1,σ(j1)...bik,σ(jk)([Uj1] ∧ ... ∧ [Ujk]) where Sk denotes the symmetric group on k elements. The result for Ak follows by specializing to the diagonal case. (cid:3) Let Ak and Bk (for k = 0, ..., d) denote the matrices described in Lemma 4.1; that is, A0 = B0 = 1, Ak = Diag(aI)I∈Ik and Bk = (det GJI)I,J∈Ik for k ∈ {1, ..., d}. From Lemma 3.1, the map Ω(σ# F ) : C(Td) → MN (C(Td)) is the diagonal matrix Ω(σ# G and d = (ds,t) ∈ Mr(C(Td)), G (a)1N and so, for α = σ# F (a)) = σ# F , β = σ# Ω∗ ◦ α∗([d]) = [(Ω ◦ α)r(d)] = [(Ω(α(ds,t))s,t] = [(β(ds,t)1N )s,t] = N[βr(d)] = Nβ∗[d] where Ωr, αr, βr denote the appropriate augmented maps on Mr(C(Td)). Using the previous lemma and the above equation, the matrix Ck representing Ω∗ on Vk Zd C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 17 satisfies CkAk = NBk where it is expected that Ck is a matrix with integer entries for each k = 0, ..., d. Recall Ak = Diag(aI)I∈Ik. Let I ′ = {1, ..., d} \ I ordered so that I = {i1 < ... < ik}, I ′ = {ik+1 < ... < id}. Then set C0 = N = det F ∈ Z and Ck = BkDiag(aI ′)I∈Ik = (aJ ′ det GJI)I,J∈Ik for k = 1, ..., d. Note that A0 = B0 = 1 by the calculations before Lemma 4.1. Hence, C0A0 = NB0 and CkAk = BkDiag(aI ′)I∈IkDiag(aI)I∈Ik = BkDiag(aI∪I ′)I∈Ik = BkDiag(a1,...,d)I∈Ik = BkDiag(N)I∈Ik = NBk for k = 1, ..., d. In order to calculate the K-theory of OF,G(Td), one needs only to calculate ker(1− Ck) and coker(1 − Ck) for each k = 0, ..., d as the next theorem will demonstrate. Theorem 4.2. Let F = Diag(a1, ..., ad), G ∈ Md(Z) such that det G 6= 0 and aj ∈ N for each j = 1, ..., d. With Ck defined as above and OF,G(Td) defined as in Example 2.17, (1) K0(OF,G(Td)) =(cid:0) M0≤k≤d, even (2) K1(OF,G(Td)) =(cid:0) M0≤k≤d, odd coker(1 − Ck)(cid:1) ⊕(cid:0) M0≤k≤d, odd coker(1 − Ck)(cid:1) ⊕(cid:0) M0≤k≤d, even Proof. By 3.4 in Theorem 3.7, ker(1 − Ck)(cid:1), and ker(1 − Ck)(cid:1). 1 − Ck M0≤k≤d, even . .. ......... ................................................................................................................................................................................................................................................. .. K0(C(Td) ... .. . . . . . . . . . . . . . . . . . . . .. .. . . ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K0(C(Td)) . .. ......... ........................................................................................................................................................................................................................ .. K0(OF,G(Td)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . .. .. .. .. .. K1(OF,G(Td)) ... ....................................................................................................................................................................................................................... .......... .. K1(C(Td)) is exact, so there are two exact sequences K1(C(Td)) ... ........................................................................................................................................................................................................................................... .......... .. 1 − Ck M0≤k≤d, odd 0 .......... ................................................................................ .. .. . and 0 .......... .................................................................................. ... .. M0≤k≤d, even M0≤k≤d, odd coker(1 − Ck) .......... ............................................................ .. .. . K0(OF,G(Td)) .......... ............................................................................. .. .. . coker(1 − Ck) .......... .............................................................. ... .. K1(OF,G(Td)) .......... ........................................................................... ... .. M0≤k≤d, odd M0≤k≤d, even ker(1 − Ck) .......... ............................................... .. .. . ker(1 − Ck) .. .......... ............................................. ... 0 0 18 SHAWN J. McCANN which split since Vk Zd and hence, ker(1 − Ck) is free for each k. Thus, K0(OF,G(Td)) =(cid:0) M0≤k≤d, even K1(OF,G(Td)) =(cid:0) M0≤k≤d, odd coker(1 − Ck)(cid:1) ⊕(cid:0) M0≤k≤d, odd coker(1 − Ck)(cid:1) ⊕(cid:0) M0≤k≤d, even ker(1 − Ck)(cid:1) ker(1 − Ck)(cid:1). Definition 4.3. A matrix Z ∈ Md(Z) is called an integer dilation matrix provided each eigenvalue λ of Z satisfies λ > 1. Remark 4.4. The case where F is an integer dilation and G = 1d was computed in [24, Theorem 4.9] where it was found that and and K0(OF,1(Td)) =(cid:0) M0≤k≤d, even K1(OF,1(Td)) =(cid:0) M0≤k≤d, odd coker(1 − Qk)(cid:1) ⊕(cid:0) M0≤k≤d, odd coker(1 − Qk)(cid:1) ⊕(cid:0) M0≤k≤d, even ker(1 − Qk)(cid:1) ker(1 − Qk)(cid:1) for the matrix Qk satisfying the relation k). Qk(det FJI)I,J∈Ik = N1(d Recall the Smith normal form of F, F = UDV where U, V ∈ Md(Z) are unimodular matrices and D = Diag(aj)d j=1 ∈ Md(Z) is a positive diagonal matrix. Then by properties of matrix minors, (det FJI)I,J∈Ik = UkDkVk where Uk = (det UJI)I,J∈Ik, Vk = (det VJI)I,J∈Ik and Dk = Diag(aI)I∈Ik. Also note for (U −1)k = (det(U −1)JI)I,J∈Ik, Uk(U −1)k = (UU −1)k = 1(d k); that is, Uk is unimodular. Hence, for G = U −1V −1, QkUkDkVk = N1(d k Diag(aI ′)I∈Ik = BkDiag(aI ′)I∈Ik = Ck. U −1 k QkUk = U −1 k V −1 k) implies Therefore, and likewise, ker(1 − Ck) = ker(U −1 k (1 − Qk)Uk) ∼= ker(1 − Qk) coker(1 − Ck) ∼= coker(1 − Qk); that is, Theorem 4.2 extends Theorem 4.9 of [24]. Here we first consider the case where F = n1d and G ∈ Md(Z). To this end, note that Ck has a simple form. First of all, note Ak is nk1(d k) and hence, Ck = nd−kBk. (see the calculations preceding Theorem 4.2.) C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 19 Theorem 4.5. For d ∈ N, let G ∈ Md(Z) (det G 6= 0) and n ∈ N. Then with Ck = nd−kBk where Bk = (det GJ,I)I,J∈Ik ∈ M(d k)(Z) as above, (1) If d > 1 and n > 1, then coker(1 − Ck) coker(1 − Ck)(cid:1) coker(1 − Ck) if det G 6= 1 if det G = 1 if det G 6= 1 if det G = 1 (a) K0(On,G(Td)) = M0≤k≤d, even Z ⊕(cid:0) M0≤k≤d−1, even  (b) K1(On,G(Td)) = M0≤k≤d, odd Z ⊕(cid:0) M0≤k≤d−1, odd  (a) K0(On,G(Td)) = Z ⊕(cid:0) M1≤k≤d, even (b) K1(On,G(Td)) = Z ⊕(cid:0) M1≤k≤d, odd coker(1 − Ck)(cid:1) coker(1 − Ck)(cid:1) coker(1 − Ck)(cid:1) (2) If n = 1 and G is an integer dilation matrix, then (3) If d = 1, then (a) K0(On,m(T)) = Zn−1 and K1(On,m(T)) = Zm−1 for n > 1 and m 6= 0, 1 (b) K0(O1,m(T)) = Z and K1(O1,m(T)) = Z ⊕ Zm−1 for m 6= 0, 1 (c) K0(On,1(T)) = Z ⊕ Zn−1 and K1(On,1(T)) = Z for n > 1 Proof. For (1), let d, n > 1 and det G 6= 1. Then C0 = nd 6= 1 and Cd = det G 6= 1; that is, 1 − C0 and 1 − Cd are injective. Furthermore, for 1 ≤ k ≤ d − 1, Ck = nd−k(det GJ,I)J,I∈Ik. Since the characteristic polynomial of a matrix is monic, it follows from Gauss' Lemma that any rational eigenvalue of a matrix in Md(Z) must actually be an 1 nd−k /∈ σ((det GJ,I)I,J) if and only if 1 /∈ σ(Ck). Thus, 1 − Ck is integer. That is, injective for each k = 1, ..., d − 1 and so, ker(1 − Ck) = 0 for each k = 0, ..., d. By Theorem 4.2, and K0(On,G(Td)) = M0≤k≤d, even K1(On,G(Td)) = M0≤k≤d, odd coker(1 − Ck), coker(1 − Ck). Now assume det G = 1, then, as above, ker(1 − Ck) = 0 for k = 0, ..., d − 1. But ker(1 − Cd) = Z and coker(1 − Cd) = Z whether d is even or odd. Theorem 4.2 gives and K0(On,G(Td)) = Z ⊕(cid:0) M0≤k≤d−1, even K1(On,G(Td)) = Z ⊕(cid:0) M0≤k≤d−1, odd coker(1 − Ck)(cid:1) coker(1 − Ck)(cid:1). 20 SHAWN J. McCANN For (2), let n = 1 and λ > 1 for all eigenvalues λ of G. Then C0 = 1 and Cd = det G 6= 1. We now wish to show that det(1 − Ck) 6= 0 for k = 1, ..., d. To this end, choose a basis of Cd such that G becomes upper triangular (not necessarily with integer entries); that is, a11 a12 0 a22 ... . . . 0 0 G =  ... a1d ... a2d . . . ... add .   Then Ck is lower triangular with diagonal entries det GII = Qi∈I aii 6= 1 (since λ > 1 for all λ ∈ σ(G)) and so det(1 − Ck) 6= 0. Hence, 1 − Ck is injective for k = 1, ..., d and ker(1 − C0) = coker(1 − C0) = Z. Theorem 4.2 now gives and K0(On,G(Td)) = Z ⊕(cid:0) M1≤k≤d, even K1(On,G(Td)) = Z ⊕(cid:0) M1≤k≤d, odd coker(1 − Ck)(cid:1), coker(1 − Ck)(cid:1). Finally, let us prove (3). If n > 1 and m 6= 0, 1 then 1 − n and 1 − m are injective and coker(1 − n) = Zn−1 := Z/(n − 1)Z; likewise, coker(1 − m) = Zm−1. Thus, K0(On,m(T)) = Zn−1 and K1(On,m(T)) = Zm−1. If n = 1 and m 6= 0, 1, then ker(1 − n) = coker(1 − n) = Z, coker(1 − m) = Zm−1 and ker(1 − m) = 0 thus, K0(O1,m(T)) = Z and K1(O1,m(T)) = Z ⊕ Zm−1. If n > 1 and m = 1, then similarly, K0(On,1(T)) = Z ⊕ Zn−1 and K1(On,1(T)) = Z. Corollary 4.6. Let d = 2, n = 1, and 1 /∈ σ(G), the spectrum of G. Then (cid:3) (1) if det G = 1, then (a) K0(O1,G(T2)) = Z2 (b) K1(O1,G(T2)) = Z2 ⊕ coker(1 − GT ) (2) if det G 6= 1, then (a) K0(O1,G(T2)) = Z ⊕ Z1−det G (b) K1(O1,G(T2)) = Z ⊕ coker(1 − GT ) Proof. Begin with calculating C0 = n2 = 1, C1 = nGT = GT and C2 = det G. With 1 not an eigenvalue of G, it is guaranteed that det(1 − GT ) 6= 0 and hence, 1 − C1 is injective. This leaves us to calculate (1) ker(1 − C0) = Z (2) coker(1 − C0) = Z (3) ker(1 − C1) = 0 (4) coker(1 − C1) = coker(1 − GT ) (5) ker(1 − C2) =(Z if det G = 1 if det G 6= 1 0 , and C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 21 (6) coker(1 − C2) =(Z if det G = 1 Z1−det G if det G 6= 1 We now calculate the K-theory when F, G ∈ Md(Z) are both diagonal matrices. Let F = Diag(a1, ..., ad) and G = Diag(b1, ..., bd) be diagonal integral matrices of non-zero determinant and such that 1 ≤ a1 ≤ ... ≤ ad. Let f denote the number of (cid:3) 1's in F ; that is, 1 = a1 = ... = af < af +1 ≤ ... ≤ ad. Let aI = Qi∈I ai for I ∈ Ik. Then Ak = Diag(aI)I∈Ik, Bk = Diag(bI)I∈Ik and Ck = (det F )BkA−1 k = Diag(bI aI ′)I∈Ik. So ker(1 − Ck) = Zdk where dk is the number of 1's in Ck; that is, the number of I making bIaI ′ = 1. Furthermote, bI aI ′ = 1 implies bI = aI ′ = 1 and so {f + 1, ..., d} ⊂ I. So let v be the number of negative ones in {bf +1, ..., bd} and let p be the number of ones in the same set. Then dk =Xr∈I (cid:18) v 2r(cid:19)(cid:18) p (k − d + f ) − 2r(cid:19) where I = {r ∈ N(cid:12)(cid:12) 0 ≤ 2r ≤ v, p − (k − d + f ) ≤ 2r ≤ k − d + f }. Lemma 4.7. With the context of the preceding paragraph, let p > 0 be the number of ones while v is the number of negative ones. Let pk(p, v) be the number of combinations of choosing k digits of 1′s and −1′s that multiply to 1 and let vk(p, v) be similar, but multiplying to −1; that is, and pk(p, v) = X0,k−p≤2r≤v,k(cid:18) v vk(p, v) = X0,k−p≤2r+1≤v,k(cid:18) v 2r(cid:19)(cid:18) p 2r + 1(cid:19)(cid:18) k − 2r(cid:19) k − 2r − 1(cid:19) p pk(p, v) =(2p 2p+v−1 if v = 0 and p > 0 if v 6= 0 . X0≤k≤p for 0 < k ≤ p, p0(p, v) = 1 if p > 0, and v0(p, v) = vk(p, v) = pk(p, v) = p0(0, v) = 0 for k > p. Then Proof. First realize that pk(p, v) + vk(p, v) =(cid:0)p+v nations of choosing k digits to multiply to either 1 or −1. Thus, k (cid:1) since it is the number of combi- Xk pk(p, v) +Xk So if v = 0, then the proof is done. vk(p, v) =Xk (cid:18)p + v k (cid:19) = 2p+v. Assume v 6= 0 and let v be odd. Then claim vk(p, v) = pp−k(p, v). This can be easily seen by realizing that, for each choice of k digits to multiply to −1, the 22 SHAWN J. McCANN remaining digits multiply to 1. Thus, Pk vk(p, v) = Pk pp−k(p, v) = Pk pk(p, v) and so, Pk pk(p, v) = 1 Now let v be even. Then, for even k 6= 0 22p+v = 2p+v−1. p p k − 2r − 1(cid:19) 2r + 1(cid:19)(cid:18) vk(p, v) − vk(p, v − 1) = X0,k−p≤2r+1≤v,k(cid:18) v − 1 k − 2r − 1(cid:19) 2r + 1(cid:19)(cid:18) − X0,k−p+1≤2r+1≤v,k(cid:18) v − 1 2r + 1(cid:19)](cid:18) 2r + 1(cid:19) −(cid:18) v − 1 [(cid:18) v = X0,k−p≤2r+1≤v,k k − 2r − 1(cid:19) 2r (cid:19)(cid:18) = X0,k−p≤2r+1≤v,k(cid:18)v − 1 (k − 1) − 2r(cid:19) 2r(cid:19)(cid:18) X0,(k−1)−p≤2r≤v−1,k−1(cid:18) v p − 1 = p = pk−1(p, v − 1). p k − 2r − 1(cid:19) Since v − 1 is odd and v0(p, v) = 0, we get Xk vk(p, v) =Xk>0 and consequently, vk(p, v − 1) +Xk>0 pk−1(p, v − 1) = 2p+v−2 + 2p+v−2 = 2p+v−1 pk(p, v) = 2p+v−1. Xk (cid:3) Theorem 4.8. Let F = Diag(a1, ..., ad) and G = Diag(b1, ..., bd) be diagonal integral matrices of non-zero determinant and such that 1 ≤ a1 ≤ ... ≤ ad. Let f > 0 denote the number of 1's in F , let v be the number of negative ones in {bf +1, ..., bd} and let p be the number of ones in that same set. Then (1) if p = 0 and v = 0, then (2) if p = 0 and v > 0, then Z1−bI aI ′ ] Z1−bI aI ′ ] [MI∈Ik [MI∈Ik (a) K0(OF,G(Td)) = M0≤k≤d, even (b) K1(OF,G(Td)) = M0≤k≤d, odd (a) K0(OF,G(Td)) = Z2v−1−1 ⊕(cid:0) M0≤k≤d, odd (b) K1(OF,G(Td)) = Z2v−1−1 ⊕(cid:0) M0≤k≤d, odd (a) K0(OF,G(Td)) = Z2p ⊕(cid:0) M0≤k≤d, even (3) if v = 0, p > 0, we have [ MI∈Ik,bI aI ′ 6=1 [ MI∈Ik,bI aI ′ 6=1 [ MI∈Ik,bI aI ′ 6=1 Z1−bI aI ′ ](cid:3) Z1−bI aI ′ ](cid:1) Z1−bI aI ′ ](cid:1) C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 23 (4) if v 6= 0, p > 0, then (b) K1(OF,G(Td)) = Z2p ⊕(cid:0) M0≤k≤d, odd (a) K0(OF,G(Td)) = Z2p+v−1 ⊕(cid:0) M0≤k≤d, even (b) K1(OF,G(Td)) = Z2p+v−1 ⊕(cid:0) M0≤k≤d, odd Proof. Begin by calculating [ MI∈Ik,bI aI ′ 6=1 Z1−bI aI ′ ](cid:1) [ MI∈Ik,bI aI ′ 6=1 [ MI∈Ik,bI aI ′ 6=1 Z1−bI aI ′ ](cid:1) Z1−bI aI ′ ](cid:1) Zdk ker(1 − Ck) = Mk, even(odd) coker(1 − Ck) = Mk, even(odd)(cid:0) MbI aI ′ 6=1,I∈Ik (1) Mk, even(odd) (2) Mk, even(odd) K0(OF,G(Td)) =(cid:0) M0≤k≤d, even coker(1 − Ck)(cid:1) ⊕(cid:0) M0≤k≤d, odd Thus, Zdk(cid:1). Z1−bI aI ′(cid:1) ⊕(cid:0) Mk, even(odd) ker(1 − Ck)(cid:1) Z1−bI aI ′(cid:1) ⊕(cid:0) Mk, even Zdk(cid:1) Z1−bI aI ′(cid:1) Z1−bI aI ′(cid:1). Zdk(cid:1) ⊕(cid:0) Mk, even(cid:0) MbI aI ′ 6=1,I∈Ik =(cid:0) Mk, odd =(cid:0)Mk Zdk(cid:1) ⊕(cid:0) Mk, even(cid:0) MbI aI ′ 6=1,I∈Ik = ZPk dk ⊕(cid:0) Mk, even(cid:0) MbI aI ′ 6=1,I∈Ik Similarly for K1; K1(OF,G(Td) = ZPk dk ⊕(cid:0) Mk, odd(cid:0) MbI aI ′ 6=1,I∈Ik Z1−bI aI ′(cid:1). If p = 0 and v = 0, then there is no way to multiply to get 1. Hence, 1 − Ck is injective for all k and the result follows readily. If p = 0 and v > 0, then dk = 0 for all odd k or k > v and dk =(cid:0)v dk = X2≤2k≤v(cid:18) v k(cid:1) for all even k ≤ v. Thus, 2k(cid:19) = ( X0≤2k≤v(cid:18) v Xk 2k(cid:19)) − 1 = 2v−1 − 1 and the rest follows. If p > 0, then dk =Xr∈I (cid:18) v 2r(cid:19)(cid:18) p (k − d + f ) − 2r(cid:19) where I = {r ∈ N(cid:12)(cid:12) 0 ≤ 2r ≤ v, p − (k − d + f ) ≤ 2r ≤ k − d + f }. Let k′ = k − d + f and then dk = pk′(p, v) 24 SHAWN J. McCANN where pk′(p, v) was defined in the above lemma and so, Xk dk =Xk and once again the rest follows. pk(p, v) =(2p 2p+v−1 if v = 0 if v 6= 0 Remark 4.9. If f, the number of ones in F, is 0 then Ck is injective for k = 0, ..., d − 1. It is then very simple to calculate the K-groups whether det G = 1 or det G 6= 1. Corollary 4.10. If F = n1d, G = m1d ∈ Md(Z) where n ∈ N and m ∈ Z are non-zero, then (1) if either n > 1 or m 6= 1, then (cid:3) (a) K0(On,m(Td)) = M0≤k≤d, even (b) K1(On,m(Td)) = M0≤k≤d, odd k) (d 1−nd−kmk Z k) (d 1−nd−kmk Z (2) if n = m = 1, then K0(O1,1(Td)) = K1(O1,1(Td)) = Z2d (3) if n = 1, m = −1, then K0(O1,−1(Td)) = K1(O1,−1(Td)) = Z2d−1 ⊕ Z2d−1 2 5. Acknowledgements This paper is the second product of my doctoral dissertation at the University of Calgary. It would not have been possible without Dr. Berndt Brenken for his insight on corrections and enhancements of the material discussed here. I am also indebted to NSERC, the Department of Mathematics and Statistics at the University of Calgary and the Department of Mathematics and Statistics at the University of Regina in providing funding to finance my mathematical studies. References [1] an Huef, A., and Raeburn, I., The ideal structure of Cuntz-Krieger algebras, Ergodic Theory Dyn. Sys. 17 (1997) 611-624. [2] Bates, T., Hong, J. H., Raeburn, I., and Szymanski, W., The ideal structure of C∗-algebras of infinite graphs, Illinois J. Math 46 (2002), 1159-1176. [3] Bates, T., Pask, D., Raeburn, I., and Szymanski, W., The C ∗-algebras of row-finite graphs, New York J. Math. 6 (2000), 307-324. [4] Blackadar, B., K-theory for Operator Algebras, M. S. R. I. Monographs, vol. 5, Springer- Verlag, Berlin and New York, 1986. [5] Blackadar, B., Operator Algebras: Theory of C-Algebras and von Neumann Algebras, En- cyclopaedia of Mathematical Sciences, vol. 122, Springer-Verlag, Berlin, 2006. [6] Brenken, B., The local product structure of expansive automorphisms of solenoids and their associated C ∗-algebras, Can. J. Math., 48, (1996), 692-709. [7] Brenken, B., C ∗-algebras associated with topological relations, J. Ramanujan Math. Soc. 19, No. 1 (2004), 1-21. [8] Brenken, B., Endomorphisms of type I von Neumann algebras with discrete center, J. Operator Theory 51 (2004), no. 1, 19-34. C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 25 [9] Brenken, B., The isolated ideal of a correspondence associated with a topological quiver. New York J. Math., 12, (2006), 1-16. [10] Brenken, B., A Dynamical Core for Topological Directed Graphs, Munster J. of Math. 3 (2010), 111-144. [11] Brenken, B., Topological Quivers as Multiplicity Free Relations, Math. Scand., 106, (2010), 217-242. [12] Brown, N.P., and Ozawa, N., C ∗-Algebras and Finite-Dimensional Approximations, Graduate Studies in Mathematics, 88. Amer. Math. Soc. Providence, Rhode Island, 2008. [13] Conway, J.B., A Course in Functional Analysis, Second Edition, Graduate Texts in Mathe- matics, 96. Springer, New York, 1990. [14] Cuntz, J., Simple C ∗-algebras generated by isometries. Comm. Math. Phys. 57 (1977), no. 2, 173-185. [15] Cuntz, J., and Krieger, W., A class of C ∗-algebras and topological Markov chains, Inventiones Math., 56 (1980), 251-268. [16] Davidson, K., C ∗-Algebras by Example, Fields Institute Monograph, Amer. Math. Soc. Prov- idence, Rhode Island, 1996. [17] Deaconu, V., Groupoids associated with endomorphisms, Trans. Amer. Math. Soc. 347 (1995), 1779-1786. [18] Deaconu, V., A path model for circle algebras, J. Operator Theory 34 (1995), 57-89. [19] Deaconu, V., Generalized Cuntz-Krieger algebras, Proc. Amer. Math. Soc. 124 (1996), 3427- 3435. [20] Deaconu, V., Continuous graphs and C ∗-algebras, in Operator theoretical methods (Timisoara, 1998), 137-149, Theta Found., Bucharest, 2000. [21] Deaconu, V., and Muhly, P., C ∗-algebras associated with branched coverings, Proc. Amer. Math. Soc. 129 (2001), 1077-1086. [22] Dummit, D.S., and Foote, R.M., Abstract Algebra, 3rd edition, Wiley and Sons, 2004. [23] Evans, D.E., The C-algebras of topological Markov chains, Lecture Notes, Tokyo Metropolitan University (1983). [24] Exel, R., an Huef, A., and Raeburn, I., Purely Infinite Simple C ∗-algebras associated to Integer Dilation Matrices, Indiana Univ. Math. J. 60 (2011), no. 3, 1033-1058. [25] Fowler, N.J., Laca, M., and Raeburn, I., The C ∗-algebras of infinite graphs, Proc. Amer. Math. Soc. 8 (2000), 2319-2327. [26] Fowler, N.J., Muhly, P.S., and Raeburn, I., Representations of Cuntz-Pimsner Algerbas, Indi- ana Univ. Math. J., 52(3) (2003), 569-605. [27] Fowler, N.J., Raeburn, I., The Toeplitz algebra of a Hilbert bimodule, Indiana Univ. Math. J. 48 (1999), 155-181. [28] Gabriel, P., Unzerlegbare Darstellungen I (Oberwolfach 1970), Manuscr. Math. 6 (1972), 71- 103. [29] Hajac, P. M., Matthes, R., and Szyma´nski, W., Graph C ∗-algebras and Z2-quotients of quan- tum spheres. Proceedings of the XXXIV Symposium on Mathematical Physics (Toru, 2002). Rep. Math. Phys. 51 (2003), no. 2-3, 215-224. [30] Hajac, P.M., Matthes, R., and Szyma´nski, W., Quantum real projective space, disc and spheres, Algebr. Represent. Theory 6 (2003), 169-192. [31] Hong, J.H., and Szyma nski, W., Quantum spheres and projective spaces as graph algebras, Comm. Math. Phys. 232 (2002), 157-188. [32] Han, D., Jing, W., Larson, D., and Mohapatra, R., Riesz bases and their dual modular frames in Hilbert C ∗-modules. J. Math. Anal. Appl. 343 (2008), no. 1, 246-256. [33] Hatcher, A., Algebraic Topology, Cambridge University Press, New York, 2001. [34] Ji, R., On Crossed Product C ∗-Algebras Associated with Furstenberg Transformations on Tori, PhD Thesis, State University of New York, Stony Brook, 1986. 26 SHAWN J. McCANN [35] Kadison, R.V., and Ringrose, J.R., Fundamentals of the theory of operator algebras. Vol. I. Elementary theory, Graduate Studies in Mathematics, 15. Amer. Math. Soc., Providence, RI, 1997. [36] Kadison, R.V., and Ringrose, J.R., Fundamentals of the theory of operator algebras. Vol. II. Advanced theory, Graduate Studies in Mathematics, 16. Amer. Math. Soc., Providence, RI, 1997. [37] Kajiwara, T., and Watatani, Y., Hilbert C ∗-bimodules and continuous Cuntz-Krieger algebras. J. Math. Soc. Japan 54 (2002), no. 1, 35-59. [38] Katsura, T., On C ∗-algebras associated with C ∗-correspondences. J. of Functional Analysis 217 (2004), 366-401. [39] Katsura, T., A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗- algebras. I. Fundamental results. Trans. Amer. Math. Soc. 356 (2004), no. 11, 4287-4322. [40] Katsura, T., A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗- algebras. II. Examples. Internat. J. Math. 17 (2006), no. 7, 791-833. [41] Katsura, T., A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗- algebras. III. Ideal structures. Ergodic Theory Dynam. Systems 26 (2006), no. 6, 1805-1854. [42] Katsura, T., A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗- algebras. IV. Pure infiniteness. J. Funct. Anal. 254 (2008), no. 5, 1161-1187. [43] Kumjian, A.,Notes on C-algebras of graphs, Contemporary Math. 228 (1998) 189-200. [44] Kumjian, A., and Pask, D., Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1-20. [45] Kumjian, A., and Pask, D., Actions of Zk associated to higher rank graphs, Ergodic Theory & Dynamical Systems, 23 (2003), 1153-1172. [46] Kumjian, A., Pask, D., and Raeburn, I., Cuntz-Krieger algebras of directed graphs, Pacific J. Math. 184 (1998), 161-174. [47] Kumjian, A., Pask, D., Raeburn, I., and Renault, J., Graphs, groupoids, and Cuntz-Krieger algebras, J. Funct. Anal. 144 (1997), 505-541. [48] Lance, E.C., Hilbert C ∗-modules: A toolkit for operator algebraists, London Mathematical Society Lecture Note Series, vol. 210, Cambridge University Press, 1995. [49] Mann, M.H., Raeburn, I., and Sutherland, C.E., Representations of finite groups and Cuntz- Krieger algebras, Bull. Austral. Math. Soc. 46 (1992), 225243. [50] Mann, M.H., Raeburn, I., and Sutherland, C.E., Representations of compact groups, Cuntz- Krieger algebras, and groupoid C ∗-algebras in Miniconference on probability and analysis (Sydney, 1991), 135144, Proc. Centre Math. Appl. Austral. Nat. Univ., 29, Austral. Nat. Univ., Canberra, 1992. [51] McCann, S.J., C ∗-algebras associated with topological group quivers, PhD Thesis, University of Calgary, Calgary, Alberta, Canada, 2012. [52] McCann, S.J., C ∗-algebras associated with topological group quivers I: generators, relations and spatial structure, preprint [53] Muhly, P., and Solel, B., Tensor algebras over C ∗-correspondences (representations, dilations, and C ∗-envelopes), J. Funct. Anal. 158 (1998), 389-457. [54] Muhly, P., and Solel, B., On the Morita Equivalence of Tensor algebras, Proc. London Math. Soc. 81 (2000), 113-168. [55] Muhly, P., and Tomforde, M., Adding tails to C ∗-correspondences, Doc. Math. 9 (2004), 79- 106. [56] Muhly, P., and Tomforde, M., Topological quivers. Internat. J. Math. 16 (2005), no. 7, 693-755. [57] Munkres, J. R., Topology, 2nd edition, Prentice-Hall, 2000. [58] Pask, D., and Sutherland, C.E., Filtered inclusions of path algebras; a combinatorial approach to Doplicher-Roberts duality, J. Operator Theory 31 (1994), 99-121. [59] Paulsen, V.I., Completely bounded maps and operator algebras, Cambridge Studies in Ad- vanced Math., 78, Cambridge University Press, Cambridge, 2002. C ∗-ALGEBRAS ASSOCIATED WITH TOPOLOGICAL GROUP QUIVERS II 27 [60] Pimsner, M., A class of C ∗-algebras generating both Cuntz-Krieger algebras and crossed prod- ucts by Z, in Free Probability Theory, fields inst. Commun., vol. 12, Amer. Math. Soc., Providence, 1997, pages 189-212. [61] Schweizer, J., Crossed Product by C ∗-correspondences and Cuntz-Pimsner Algebras, 'in C ∗- algebras, Proceedings of the SFB-Workshop on C ∗-algebras, Muenster, 1999,' (Eds.) J. Cuntz, S. Echterhoff, Springer Verlag, Berlin, 2000. [62] Stacey, P.J., Crossed products of C ∗-algebras by endomorphisms, J. Austral. Math. Soc. (Series A) 54 (1993), 204-212. [63] Raeburn, I., Williams, D.P., Morita Equivalence and Continuous-Trace C ∗-Algebras, Math. Surveys & Monographs, vol. 60, Amer. Math. Soc., Providence, 1998. [64] Rieffel, M., C ∗-algebras associated with irrational rotations, Pacific J. Math. 93(2) (1981), 415-429. [65] Rφrdam, M., Larsen, F., Lausten, N.J., An Introduction to K-theory for C-algebras. 256 pp. London Mathematical Society, Student Text 49, Cambridge University Press, Cambridge, 2000. [66] Rφrdam, M., Stφrmer, E., Operator Algebras and Non-Commutative Geometry, Vol VII: Classification of Nuclear C*-Algebras. Entropy in Operator Algebras. Encyclopaedia of Math- ematical Sciences 126. Springer Verlag, Heidelberg, 2001. [67] Walters, P., An introduction to ergodic theory, Springer, New York, 1982. [68] Yamashita, S., Circle Correspondence C ∗-algebras, Houston J. Math. 37 (2011), no. 4, 1181- 1202.
1906.09723	1	1906	2019-06-24T04:54:34	Absolutely compatible pair of elements in a von Neumann algebra-II	[ "math.OA" ]	Let $A$ be a unital C$^*$-algebra with unity $1_A$. A pair of elements $0 \le a, b \le 1_A$ in $A$ is said to be \emph{absolutely compatible} if, $\vert a - b \vert + \vert 1_A - a - b \vert = 1_A.$ In this paper we provide a complete description of absolutely compatible pair of strict elements in a von Neumann algebra. The end form of such a pair has a striking resemblance with that of a `generic pair' of projections on a complex Hilbert space introduced by Halmos.	math.OA	math	ABSOLUTELY COMPATIBLE PAIR OF ELEMENTS IN A VON NEUMANN ALGEBRA-II ANIL KUMAR KARN Abstract. Let A be a unital C∗-algebra with unity 1A. A pair of elements 0 ≤ a, b ≤ 1A in A is said to be absolutely compatible if, a − b + 1A − a − b = 1A. In this paper we provide a complete description of absolutely compatible pair of strict elements in a von Neumann algebra. The end form of such a pair has a striking resemblance with that of a 'generic pair' of projections on a complex Hilbert space introduced by Halmos. 1. Introduction Let A be a C∗-algebra. A pair of elements a, b ∈ A is said to be orthogonal, if ab = 0 = ba = a∗b = ab∗. Orthogonal pair of positive elements play an important role in the theory of C∗- algebras. For example, it follows from the functional calculus that every self-adjoint element a ∈ Asa has a unique decomposition: a = a+ − a− in A+, where a+ is algebraically orthogonal to a−. Recently, the author proved an order theoretic characterization of algebraic orthogonality among positive elements of a C∗-algebra [7]. (Also see [5, 6].) Orthogonal pairs of positive elements of norm ≤ 1 exhibit an interesting property. Let A be a unital C∗-algebra with unity 1A. For a pair of elements 0 ≤ a, b ≤ 1A in A, we have ab = 0 (a is algebraically orthogonal to b) if and only if a + b ≤ 1A and a − b + 1A − a − b = 1A. We isolate the later part and propose the following definition: A pair of elements 0 ≤ a, b ≤ 1A in A is said to be absolutely compatible, ([7]), if a − b + 1A − a − b = 1A. It was proved in [7] that a projection p in a C∗-algebra A is absolutely compatible with a positive element a of A with kak ≤ 1 if, and only if, ap = pa. However, the two notions are distinct in general. The notion of absolute compatibility was introduced as an instrument to prove a spectral decomposition theorem in the context of "absolute order unit spaces" [7]. As an absolute order unit space is an order theoretic generalization of unital C∗-algebras [7], absolute compatibility appears to be a significant property. Keeping this point of view, the author, along with Jana and Peralta, initiated a study of absolute compatibility in operator algebras [3, 4]. Let M be a von Neumann algebra and let 0 ≤ a ≤ 1M . We write s(a) := sup{p ∈ P(M ) : p ≤ a} 2010 Mathematics Subject Classification. Primary 46L10; Secondary 46B40. Key words and phrases. Absolutely ordered space, absolute oder unit space, order isometry, absolute value preserving maps, absolute matrix order unit space. 1 2 and ANIL KUMAR KARN n(a) := sup{p ∈ P(M ) : pa = 0}. For 0 ≤ a ≤ 1M , we say that a is strict in M , if s(a) = 0 and n(a) = 0. In [3], it was proved that an absolutely compatible pair of positive elements in a von Neumann algebra has a (matricial) decomposition as a direct sum of commuting and 'strict' elements. Let 0 ≤ a, b ≤ 1M such that a is absolutely compatible with b. Then there exist mutually orthogonal projections p1, p2, s, n1, n2 ∈ P(M ) with p1 + p2 + s + n1 + n2 = 1M such that and a = p1 ⊕ a1 ⊕ a2 ⊕ 0 ⊕ a3 b = b1 ⊕ p2 ⊕ b2 ⊕ b3 ⊕ 0 with respect to {p1, p2, s, n1, n2} with a2 and b2 are strict and absolutely compatible in sM s [3, Theorem 2.10]. Thus the study of absolutely compatible pair of elements reduces to such strict pairs. It was further proved that Theorem 1.1. [3, Theorem 2.12] Let M be a von Neumann algebra, and let a, b be elements in [0, 1]M . Then a is absolutely compatible with b if and only if there exists a projection p1 in M so that a and b have matrix representations, say a = (cid:18) a11 a12 12 a22 (cid:19) , and b = (cid:18) b11 b∗ 12 a∗ b12 b22 (cid:19) with respect to the set {p1, 1 − p1 = p2} (i.e., aij = piapj and bij = pibpj) satisfying: (i) a12 + b12 = 0; (ii) a12a∗ (iii) a∗ (iv) a12 = a11a12 + a12a22 = b11a12 + a12b22. 12 = (p1 − a11)(p1 − b11); 12a12 = a22b22 = b22a22; Using these decompositions, a complete description of absolutely compatible pair of strict elements was given for the finite dimensional algebras Mn in the same paper [3, Theorem 3.9]. In the present paper we continue the investigation and provide a complete de- scription of absolutely compatible pair of strict elements in a von Neumann algebra. However, as expected, the finite dimensional (matricial) techniques used in [3] fail to work. At this juncture, the author would like to thank Professor Kalyan B. Sinha for his suggestion to use polar decomposition. This idea works just perfectly. The main results of the paper are Theorems 1.2 and 1.3. Theorem 1.2. Let M be a von Neumann algebra with the underlying Hilbert space H. Assume that a, b ∈ [0, 1]M be strict and absolutely compatible pair. Put p = 1 − r(a ◦ b). (1) Then pH is isometrically isomorphic to (1−p)H. In particular, H ≡ K⊕K, where K = pH. (2) There exist strict elements a1, b1 ∈ [0, p] ∩ M with a1b1 = b1a1, a1 + b1 ≤ p together with p − (a1 + b1) strict in pM p; and a unitary U : H → K ⊕ K such that a = U ∗(cid:20) a1 (a1b1) 1 2 1 2 (a1b1) p − a1 (cid:21) U and b = U ∗(cid:20) b1 −(a1b1) 1 2 1 2 −(a1b1) p − b1 (cid:21) U. ABSOLUTE COMPATIBILITY 3 Theorem 1.3. Let M be a von Neumann algebra with the underlying Hilbert space H. Assume that a, b ∈ [0, 1]M be strict and commuting pair such that a2 + b2 ≤ 1 with 1 − (a2 + b2) strict. Put a1 = (cid:20)a2 ab ab 1 − a2(cid:21) and b1 = (cid:20) b2 −ab −ab 1 − b2(cid:21). Then a1, b1 ∈ [0, 1]M2(M) and a1 is absolutely compatible with b1. It is interesting and surprising to note that the end form of an absolutely com- patible pair of strict elements bears a striking resemblance with that of a 'generic pair' of projections on a complex Hilbert space (studied by Halmos [2]). A pair of closed subspaces M and N of a Hilbert space H is said to be in generic position, if each of the following four subspaces of H: M ∩ N , M ∩ N ⊥, M ⊥ ∩ N and M ⊥ ∩ N ⊥ are trivial. Let P and Q be the projections of H on M and N respectively. In [2], Halmos proved that if M and N are in generic position, then there exists a Hilbert space K, commuting, positive and invertible contractions C and S in B(K) and a unitary operator U : H → K ⊕ K such that P = U ∗(cid:20) C2 CS CS S2(cid:21) U and Q = U ∗(cid:20) C2 −CS S2 (cid:21) U. −CS This observation further signifies the importance of a pair of strict absolutely compatible elements in an operator algebra. This paper is in the sequel of [3]. We obtain some basic results to prove Theorems 1.2 and 1.3. 2. The main results Lemma 2.1. Let M be a von Neumann algebra with the underlying Hilbert space H and let a ∈ [0, 1]M be strict. If p ∈ P(M ), then, for the matrix representation a = (cid:20)a11 a12 12 a22(cid:21) with respect to {p, 1 − p}, we have a11 ∈ [0, p] and a22 ∈ [0, 1 − p] a∗ are also strict in the von Neumann algebras pM p and (1 − p)M (1 − p) respectively. Proof. Without any loss of generality, we may assume that p 6= 0, p 6= 1. First, we show that r(a11) = p. Put p − r(a11) = p0. Then p0a11 = 0 so that 0 ≤ (cid:20)p0 0 0 1 − p(cid:21)(cid:20)a11 a12 12 a22(cid:21)(cid:20)p0 a∗ 0 Thus p0a12 = 0 and consequently, 0 1 − p(cid:21) = (cid:20) 0 a∗ 12p0 p0a12 a22 (cid:21) . (cid:20)p − p0 0 0 1 − p(cid:21) a(cid:20)p − p0 0 0 1 − p(cid:21) = a. Now, it follows that 1 ≤ r(a) ≤ 1 − p0 so that p0 = 0. Thus r(a11) = p. Similarly, r(a22) = 1 − p. Since n(x) = 1 − r(x) for any x ∈ [0, 1]M , we have n(a11) = 0 and n(a22) = 0. Next, as a is strict, so is 1 − a = (cid:20)p − a11 −a∗ 12 −a12 1 − p − a22(cid:21). Thus, as above, r(p − a11) = p and r(1 − p − a22) = 1 − p. Therefore, s(a11) = 0 and s(a22) = 0. Hence a11 ∈ [0, p] and a22 ∈ [0, 1 − p] are strict in the von Neumann algebras pM p and (1 − p)M (1 − p) respectively. (cid:3) Lemma 2.2. Let M be a von Neumann algebra with the underlying Hilbert space H and let a ∈ [0, 1]M be strict. If ax = 0 for some x ∈ M , then x = 0. In particular, if aξ = 0 for some ξ ∈ H, then ξ = 0. 4 ANIL KUMAR KARN Proof. Let ax = 0. Then ax∗2 = axx∗ = 0 so that ax∗ = 0 and consequently, ar(x∗) = 0. Thus r(x∗) ≤ n(a) = 0 so that r(x∗) = 0. Now, it follows that x = 0. Next, if aξ = 0, then ap = 0 where p is the projection of H on the span of ξ. (cid:3) Now, by the first step, p = 0 so that ξ = 0. Lemma 2.3. Let M be a von Neumann algebra with the underlying Hilbert space H and let a, b ∈ [0, 1]M be strict. 1 (1) Then a (2) If ab = ba, then ab is also strict in M . 2 is also strict in M . Proof. (1) Let p = s(a 1 2 ). Then p = pa 1 2 = a 1 2 p. Squaring, we get p = pa = ap 2 ) = 0. Also, 1 so that p ≤ s(a) = 0. Thus p = 0 and consequently, s(a n(a 2 ) = n(a) = 0. Therefore, a 2 is strict. 1 1 (2) Let ab = ba. Then ab ∈ [0, 1]M with ab ≤ a. Thus s(ab) ≤ ab ≤ a so that s(ab) ≤ s(a) = 0. Therefore, s(ab) = 0. Next, we have abn(ab) = 0. As a and b are strict, a repeated use of Lemma 2.2 yields that n(ab) = 0. Thus ab is also strict. (cid:3) Remark 2.4. Let a ∈ [0, 1]M . Then a is strict if and only if a2 is strict, The following results are a compilation of operator algebra folklore which can easily be found in the literature. (See, for example, [1, Section I.5.2].) Folklore 2.5. Let M be a von Neumann algebra with the underlying Hilbert space H and let x ∈ M . If x = ux is the polar decomposition of x, then (1) u : (1 − r(x))H → r(x)H is an unitary. In particular, u is a partial isometry in M . (2) x = u∗x. (3) r(x) = u∗u. (4) x∗k = uxku∗ for all k ∈ N. (5) xk = u∗x∗ku for all k ∈ N. Now we prove the main results of the paper. Proof of Theorem 1.2. (1) Put r(a ◦ b) = p1 so that p = 1 − p1. Now, by Theorem 1.1, a and b have matrix representations a = (cid:20)a11 a12 b22 (cid:21) 12 a22(cid:21) and b = (cid:20) b11 −a12 −a∗ 12 a∗ 12a12 = a22b22; and 12 = (p1 − a11)(p1 − b11); with respect to {p1, p} such that (a) a12a∗ (b) a∗ (c) a12 = a11a12 + a12a22 = b11a12 + a12b22. Since a and b are strict, we have p1 6= 0, p1 6= 1. Also, by Lemma 2.1, a11 and b11 are strict in p1M p1 and a22 and b22 are strict in pM p. Consider the polar decomposition a12 = ua12 ABSOLUTE COMPATIBILITY 5 so that U : u∗uH → uu∗H is a unitary. By (a) and Folklore 2.5(4), we have (p1 − a11)(p1 − b11) = a∗ 122 = ua122 = u(a22b22)u∗ ≤ uu∗ ≤ p1. Now, as a11 and b11 are strict elements in [0, p1] with a11b11 = b11a11, by Lemma 2.3(2), we may conclude that (p1 − a11)(p1 − b11) is also a strict element in [0, p1]. Thus r((p1 −a11)(p1 −b11)) = p1. Thus uu∗ = p1 = 1−p. In a similar way, by using (b) and Folklore 2.5(5), we may prove that u∗u = p. Thus u : pH → (1 − p)H is a unitary. (2) By (b), we have a22b22 = a122 so that a22, b22 and a12 commute with each other. Thus by (c), we get ua12 = a12 = a11a12 + a12a22 = a11ua12 + ua12a22 = a11ua12 + ua22a12. In other words, a12(a11u + ua22 − u)∗(a11u + ua22 − u) = 0. By Lemma 2.3, a12 is strict in pM p so that, by Lemma 2.2, a11u+ua22 = u. Therefore, p1 = uu∗ = a11uu∗ + ua22u∗ = a11 + ua22u∗ whence ua22u∗ = p1 − a11. In the same way we can show that ub22u∗ = p1 − b11. Now, put a22 = a1, b22 = b1 and U = (cid:20) 0 u∗ p 0(cid:21). Then a1 and b1 are strict elements in [0, p] and U : H → K ⊕ K is a unitary. Also, the matrix multiplications yield that U ∗(cid:20) a1 (a1b1) 1 2 1 2 (a1b1) p − a1 (cid:21) U = a and U ∗(cid:20) b1 −(a1b1) 1 2 1 2 −(a1b1) p − b1 (cid:21) U = b. Finally, we show that p − a1 − b1 is a strict element of [0, p]. Note that a ◦ b = (cid:20)a11 + b11 − p1 0 0 0(cid:21). Thus r(a11 + b11 − p1) = r(a ◦ b) = p1. Next, a11 + b11 − p1 = p1 − ua1u∗ + p1 − ub1u∗ − p1 = u(p − a1 − b1)u∗ for upu∗ = uu∗uu∗ = p2 1 = p1. Also, then u∗(a11 + b11 − p1)ku = (p − a1 − b1)k for all k ∈ N. Since for any x ∈ M +, r(x) is limit of the increasing sequence {( 1 n + x)−1x} in the strong operator topology, we may conclude that r(p − a1 − b1) = u∗r(a1 + b11 − p1)u = u∗p1u = p. Thus n(p − a1 − b1) = 0. Also s(p − a1 − b1) ≤ p − a1 − b1 ≤ p − a1 so that s(p − a1 − b1) ≤ s(p − a1) = 0 for p − a1 is strict. Therefore, p − a1 − b1 is a strict element in [0, p]. (cid:3) 6 ANIL KUMAR KARN Proof of Theorem 1.3. Let ξ, η ∈ H. Then habη, ξi2 = hbη, aξi2 ≤ hb2η, ηiha2ξ, ξi ≤ h(1 − a2)η, ηiha2ξ, ξi. Thus we obtain that (cid:28)(cid:20)a2 ab ab 1 − a2(cid:21)(cid:20)ξ η(cid:21)(cid:29) = ha2ξ, ξi + habη, ξi + habξ, ηi + h(1 − a2)η, ηi Therefore, a1 ≥ 0. Similarly, we can also show that b1 ≥ 0. Again, ≥ (cid:16)ha2ξ, ξi 1 2 − h(1 − a2)η, ηi ≥ 0. 1 2(cid:17)2 1M2(M) − a1 = (cid:20)1 − a2 −ab a2 (cid:21) and 1M2(M) − b1 = (cid:20)1 − b2 ab b2(cid:21) −ab ab so that a1, b1 ∈ [0, 1]M2(M). Now, we show that a1 is strict. Let p be a projection in M2(M ) such that p ≤ a1. Then p = a1p. For (cid:20)ξ η(cid:21) ∈ p(H ⊕ H), we have a1(cid:20)ξ η(cid:21) = (cid:20)ξ η(cid:21) so that ξ = a2ξ + abη and η = abξ + (1 − a2)η. By the second condition, we get a2η = abξ so that a(aη − bξ) = 0. Since a is strict, by Lemma 2.2, we may conclude that aη = bξ. Using this in the first relation, we get ξ = a2ξ + b2ξ so that (1 − a2 − b2)ξ = 0. Since 1 − a2 − b2 is also strict, invoking Lemma 2.2 again we conclude that ξ = 0. But then, aη = 0 whence η = 0 as a is also strict. Now, it follows that p = 0 so that s(a1) = 0. Next, let q be a projection in M2(M ) such that a1q = 0. If (cid:20)ξ have a1(cid:20)ξ η(cid:21) = 0. Thus η(cid:21) ∈ q(H ⊕ H), we a2ξ + abη = 0 and abξ + (1 − a2)η = 0 and, as above we again conclude that q = 0. Therefore, n(a1) = 0 too whence a1 is strict. In a similar manner, we can conclude that b1 is also strict. Finally, we show that a1 is absolutely compatible with b1. We have = (cid:20)(a2 + b2)2 0 0 (a2 + b2)2(cid:21) so that a1 − b1 = (cid:20)a2 + b2 0 2ab 2ab (a1 − b1)2 = (cid:20)a2 − b2 b2 − a2(cid:21)2 a2 + b2(cid:21). Also 1M2(M) − a1 − b1 = (cid:12)(cid:12)(cid:12)(cid:12) 0 (cid:20)a − a2 − b2 = (cid:20)1 − a2 − b2 0 0 0 a2 + b2 − 1(cid:21)(cid:12)(cid:12)(cid:12)(cid:12) 1 − a2 − b2(cid:21) . 0 It follows that a1 − b1 + 1M2(M) − a1 − b1 = 1M2(M) so that a1 is absolutely compatible with b1. (cid:3) Acknowledgements: The author is thankful to Antonio M. Peralta for intro- ducing to him the notion of 'strict' elements. ABSOLUTE COMPATIBILITY 7 References [1] B. Blackadar, Operator algebras, Springer-Verlag Heidelberg Berling, 2006. [2] P. R. Halmos, Two subspaces, Trans. Amer. math. Soc., 144(1969), 181-189. [3] N. K. Jana, A. K. Karn and A. M. Peralta, Absolutely compatible pairs in a von Neumann algebra, (https://arxiv.org/abs/1801.01216), (Communicated for publication). [4] N. K. Jana, A. K. Karn and A. M. Peralta, Contractive linear preservers of absolutely compatible pairs between C∗-algebras, Revista de la Real Academia de Ciencias Exactas, Fsicas y Naturales. Serie A. Matemticas (RCSM) 113(2019) no. 3, 2731-2741. [5] A. K. Karn, Orthogonality in lp-spaces and its bearing on ordered Banach spaces, Positivity, 18(02) (2014), 223-234. [6] A. K. Karn, Orthogonality in C∗-algebras, Positivity, 20(03) (2016), 607-620. [7] A. K. Karn, Algebraic orthogonality and commuting projections in operator algebras, Acta Sci. Math. (Szeged), 84(2018), 323-353. School of Mathematical Sciences, National Institute of Science Education and Re- search, HBNI, Bhubaneswar, P.O. - Jatni, District - Khurda, Odisha - 752050, India. E-mail address: [email protected]
1904.06431	5	1904	2019-07-13T15:03:16	Invariant subspaces of generalized Hardy algebras associated with compact abelian group actions on W*-algebras	[ "math.OA" ]	We consider an action of a compact group whose dual is archimedean linearly ordered or a direct product (or sum) of such groups on a von Neumann algebra, M. We define the generalized Hardy subspace of the Hilbert space of a standard representation the algebra, and the Hardy subalgebra of analytic elements of M with respect to the action. We find conditions in order that the Hardy algebra is a hereditarily reflexive algebra of operators. In particular if every non zero spectral subspace, contains a unitary operator, the condition is satisfied and therefore the Hardy algebra is hereditarily reflexive. This is the case if the action is the dual action on a crossed product, or an ergodic action, or, if, in some situations, the fixed point algebra is a factor.	math.OA	math	Invariant subspaces of generalized Hardy alge- bras associated with compact abelian group ac- tions on W-algebras Costel Peligrad Department of Mathematical Sciences, University of Cincinnati, PO Box 210025, Cincinnati, OH 45221-0025, USA. E-mail address: [email protected] Key words and phrases. W-dynamical system, invariant subspaces, analytic elements, generalized Hardy algebra, reflexive algebra. 2013 Mathematics Subject Classification. Primary 46L10, 46L40, 47L75; Secondary 30H10, 47B35. ABSTRACT. We consider an action of a compact abelian group whose dual is any subgroup of the additive group of real numbers (so, an archimedean linearly ordered group) or a direct product (or sum) of such groups on a W- algebra, M . We define the generalized Hardy subspace of the Hilbert space of a standard representation the algebra, and the Hardy subalgebra of analytic elements of M with respect to the action. We find conditions in order that the Hardy algebra is a hereditarily reflexive algebra of operators. In particular if every non zero spectral subspace, contains a unitary operator, the condition is satisfied and therefore the Hardy algebra is hereditarily reflexive. This is the case if the action is the dual action on a crossed product, or an ergodic action, or, if, in some situations, the fixed point algebra is a factor. 1 Introduction This paper is concerned with the study of invariant subspaces and reflexivity of operator algebras associated with compact group actions on W-algebras. Recall first the definition of a reflexive operator algebra. Let A ⊂ B(X) be a weakly closed algebra of operators on a Banach space X. Denote by Lat(A) the lattice of closed subspaces of X that are invariant for all operators a ∈ A. Let algLat(A) = {b ∈ B(X) : bK ⊂ K for all K ∈ Lat(A)} . The algebra A is called reflexive if A = algLat(A). Hence, a reflexive operator algebra is completely determined by the lattice of its invariant subspaces. An algebra A ⊂ B(X) is called hereditarily reflexive if every unital weakly closed subalgebra of A is reflexive. Sarason [19], proved two results: (1) every com- mutative von Neumann algebra is hereditarily reflexive and (2) the algebra of 1 analytic Toeplitz operators on the Hardy space H 2(T ) where T is the the unit circle T = {z ∈ C : z = 1} , is hereditarily reflexive. In [14] we extended this result in two directions: (1) to the case of H p(T ), 1 < p < ∞ and (2) to the not necessarily commutative case of non selfadjoint crossed products of finite von Neumann algebras by the semigroup Z+. Later, in [9], Kakariadis has con- sidered the more general case of reduced w-semicrossed products and, among other results, he has extended the particular case of our reflexivity result in [14, Proposition 4.5,] for p = 2 to the semicrossed product setting [9, 2.10.H.]. This result was considered later by Helmer [5] in the context of W-correspondences [12]. Further, in [15], we studied a related problem in a more general setting than the crossed product or the reduced w∗semicrossed product considered in [14] and [9, 2.9] for the case of von Neumann algebras. We considered a W- dynamical system (M, T,α) where T = {z ∈ C : z = 1} is the circle group and M is a σ−finite W-algebra. We constructed a standard covariant representa- tion of the system on a certain Hilbert space, H, a generalized Hardy space, H+ and the corresponding Hardy algebra M+ ⊂ B(H+). We have shown that if the spectral subspace corresponding to the smallest positive element of the spectrum contains a unitary element, then, the algebra M+ is reflexive. Actu- ally, [15, Theorem 3.5.] shows that if M ⊂ B(H) is a σ-finite von Neumann algebra in its standard representation such that each spectral subspace contains a unitary operator (as is, in particular, the algebra of analytic Toeplitz opera- tors considered by Sarason), then, M+ ⊂ B(H+) is a reflexive operator algebra. Recently Bickerton and Kakariadis [2] have obtained results about reflexivity of algebras associated with actions of Zd + (the direct product of d copies of Z+). In this paper we make two significant steps towards solving the reflexiv- ity problem of Hardy algebras associated to one-parameter dynamical systems (M, R,α): 1. We consider a W-dynamical system (M, G, α) where M is a von Neumann algebra in standard form, and 2. G is a compact abelian group whose dual is an arbitrary subgroup of R (possibly R itself with the discrete topology), so, a discrete group with a linear archimedean order. We also consider actions of compact abelian groups, G, whose duals, Γ, are direct products or direct sums of discrete groups with linearly archimedean order and consider the lattice order on Γ (see for instance [3]). In Section 2.1. we define a standard covariant repre- sentation of the system that will be the framework for the rest of the paper. In the Corollary to Proposition 2.4. we show that every von Neumann algebra in standard form (in particular every maximal abelian von Neumann algebra) is hereditarily reflexive, thus extending the first result of Sarason mentioned above to every von Neumann algebra in its standard representation. In Section 3 we consider the case when the dual Γ of G has an archimedean linear order, or is a direct product of such groups, we define a generalized Hardy space H+ ⊂ H and a corresponding Hardy algebra M+ ⊂ B(H+), where H is.the Hilbert space of the standard covariant representation of the system (M, G, α) and we prove that, in some conditions, including the conditions in [15], M+ ⊂ B(H+) is hereditarily reflexive (in [15] we proved only reflexivity for the particular case when Γ = Z).We do not assume as in [15] that M is σ-finite. Also, if Γ is an arbitrary archimedean linearly ordered discrete group, it can be any subgroup 2 of R with the discrete topology, not only Z as in [15], [9], [14]. Examples include the Hardy algebra of analytic Toeplitz operators, H ∞(T), the results in [15], w∗-crossed products by abelian archimedean ordered discrete groups or a direct product of such groups, some reduced w∗-semicrossed products considered in [9], [2] and other situations as stated in the Corollaries 3.14., 3.15., 3.16. and 3.17. 2 Preliminary results and notations 2.1 Standard representations of W-algebras In this section we review some concepts and results related to the standard representation of a von Neumann algebra. Some of these results are certainly known, but we did not find an exact reference for them. We provide proofs of these results for the convenience of the reader. In Proposition 2.4. and its Corollary we prove that every von Neumann algebra in its standard representa- tion is hereditarily reflexive. In particular, every abelian von Neumann algebra is hereditarily reflexive ([19, Theorem 1]). Let M be a W-algebra and let ρ be a weight on the positive part, M +, of M, that is a mapping ρ : M + → [0, ∞) ∪ {∞} such that ρ(m + n) = ρ(m) + ρ(n), m, n ∈ M + and ρ(λm) = λρ(m), m ∈ M +, λ ∈ R, λ > 0 with the convention 0 · ∞ = 0. As it is customary ([8], [20]), denote Nρ = {m ∈ M : ρ(m∗m) < ∞} . Nρ = {m ∈ M : ρ(m∗m) = 0} . Fρ =(cid:8)m ∈ M + : ρ(m) < ∞(cid:9) . Mρ = linear span of Fρ. It is immediate that Nρ is a left ideal of M . The weight ρ is called faithful if Nρ = {0} , normal if it is the sum of a family {ϕι} of positive normal linear functionals and semifinite if Mρ or, equivalently [20, 2.1.], Nρ is w- dense in M. Now let M be a W-algebra, M0 ⊂ M a W-subalgebra, and P0 : M → M0 a w-continuous projection of norm 1 of M onto M0 which is, in addition, faithful on the set of positive elements of M. Let ρ0 be a faithful normal semifinite weight on M . It is known that such a weight exists. Indeed, consider a family {ϕι} of positive normal linear functionals of M0 such that their supports {pι} 3 form a maximal family of mutually orthogonal projections of M0, in particular P pι = I. Then ρ0 =P ϕι is a faithful normal semifinite weight of M0. The following fact is stated in [19, Corollary 10.5] as a consequence of a theorem of Takesaki [20, Theorem 10.1.]. We present a short proof of this fact in our setting for the convenience of the reader. Lemma 2.1. ρ = ρ0 ◦ P0 is a faithful normal semifinite weight on M. Proof. Since ρ0 and P0 are faithful, it follows that ρ is faithful. Since ρ0 is normal and P0 is w-continuous, it follows that ρ is normal. To prove that ρ is semifinite, notice that from the definition of Nρ we have that M Nρ0 ⊂ Nρ, where M Nρ0 denotes the linear span of {mn : m ∈ M, n ∈ Nρ0 } . Since Nρ0 is w-dense in M0 it follows that M Nρ0 and therefore Nρ is dense in M so ρ is semifinite. By [8, Theorem 7.5.3.], there exists a faithful normal representation πρ of M on the completion Hρ of Nρ ⊂ M with respect to the inner product hm, ni = ρ(n∗m). This representation is uniquely determined up to unitary equivalence and is, in that sense, independent of the choice of the weight ρ0. We will use the version of Tomita-Takesaki Theorem from [8, Theorem 9.2.37.]. If S is the conjugate ρ by S(n) = n∗, then S is a preclosed densely linear operator defined on Nρ ∩ N ∗ 2 , in defined operator on Hρ and its closure has the polar decomposition J∆ which ∆ is an invertible positive operator and J is a conjugate linear isometry acting on Hρ such that J 2 = I and J πρ(M )J = πρ(M ) is the commutant of πρ(M ) in B(Hρ). We have Nρ0 ⊂ Nρ, where, as above where πρ(M ) 1 ′ ′ and Nρ0 = {m ∈ M0 : ρ0(m∗m) < ∞} . Nρ = {m ∈ M : ρ(m∗m) < ∞} . In the rest of the paper if ρ is a faithful, normal semifinite weight on M + we will identify πρ(M ) with M and will write m instead of πρ(m), m ∈ M. We will call this representation the standard representation of M and we will refer to the inclusion M ⊂ B(Hρ) as the standard form of M. Also, we will denote Hρ by H and the closure of Nρ0 in H by H0. Lemma 2.2. The restriction of P0 to Nρ extends to the orthogonal projec- tion of H onto H0. 4 Proof. Clearly, P0(Nρ0 ) = Nρ0 . We will prove next that P0(Nρ) ⊂ Nρ0 . Indeed, let n ∈ Nρ. Then n = P0(n) + (I − P0)(n) = n0 + n1. Since n ∈ Nρ, we have ρ(n∗n) < ∞, so ρ(n∗ 0n0 + n∗ 0n1 + n∗ 1n0 + n∗ 1n1) = ρ0(P0(n∗ 0n0 + n∗ 0n1 + n∗ 1n0 + n∗ 1n1)) = ρ0(n∗ 0n0 + n∗ 0P0(n1) + P0(n∗ 1)n0 + P0(n∗ 1n1)) = ρ0(n∗ 0n0 + P0(n∗ 1n1)) < ∞. Therefore, ρ0(n∗ 0n0) < ∞ and we are done. On the other hand, hP0(n), mi = hn, P0(m)i . since both of the above terms equal ρ0(P0(m∗)P0(n)), so P0 is.self adjoint. Lemma 2.3. i) M H0 is dense in H and H0 is a separating set for M that is, if m ∈ M is such that mξ0 = 0 for all ξ0 ∈ H0, then m = 0. Here, M H0 denotes the linear span of {mξ : m ∈ M, ξ ∈ H0} . ii) M ′H0 is dense in H and H0 is a separating set for M ′ that is, if m′ ∈ M ′ is such that m′ξ0 = 0 for all ξ0 ∈ H0, then m′ = 0. Proof. i) We will prove that M H0 is dense in Nρ ⊂ H and, since Nρ is dense in H, the first part of i) will follow. Let x ∈ Nρ. Therefore, ρ(x∗x) = ρ0(P0(x∗x)) < ∞. Since ρ0 is a faithful normal semifinite weight on M0 we can assume that ρ0 is the sum of a family of normal positive linear functionals {ϕι} on M0 such that their suports {pι} form a maximal family of mutually ortog- onal projections in M0 and P pι = I. Since ρ0(P0(x∗x)) = P ϕι(P0(x∗x)) < P∞ i=1 ϕi(P0(x∗x)) < ∞. Let qn = Pi=n ∞, it follows that the summable family of positive numbers {ϕι(P0(x∗x))}ι is at most countable, say {ϕi(P0(x∗x))}i∈N with ρ(x∗x) = ρ0(P0(x∗x)) = i=1 pi, where, for each i ∈ N, pi is the supportof ϕi. Then, qn ∈ Nρ0 ⊂ H0 and xqn ∈ M Nρ0 ⊂ M H0 for every n ∈ N. Clearly, since pi is the support of ϕi we have ϕi(y) = ϕi(piy) = ϕi(ypi) for all i ∈ N, y ∈ M0 and ϕi(qny) = ϕi(y) for every i, n ∈ N with i 6 n, y ∈ M0 and ϕi(qnx) = 0 if i > n We will show that limn→∞ xqn = x in H.Let ǫ > 0. i=N +1 ϕi(P0(x∗x)) < ǫ2. Therefore, Then, there exists N = Nǫ > 0 such thatP∞ if n > N hqnx − x, qnx − xi = ρ((qnx − x)∗(qnx − x)) = ϕi(P0((qnx − x)∗(qnx − x)))+ nXi=1 + ∞Xi=n+1 ϕi(P0((qnx − x)∗(qnx − x)) Using the preceding observations, if i 6 n, we get ϕi(P0((qnx − x)∗(qnx − x))) = 5 ϕi(qnP0(x∗x)qn − qnP0(x∗x) − P0(x∗x)qn + P0(x∗x)) = 0 and, if i > n > N ϕi(P0((qnx − x)∗(qnx − x) = ϕi(P0(x∗x)) Hence hqnx − x, qnx − xi < ǫ2 and the claim is proven. To prove the second part, let m ∈ M such that mξ0 = 0 for every ξ0 ∈ H0, in particular, mn = 0 for every n ∈ Nρ0 ⊂ H0. Since Nρ0 is w-dense in M0 ⊂ M and I ∈ M0, it follows that m = 0 and part i) is proven. ii). Denote by K the closure of M ′H0. Then K is a closed subspace of H which is invariant for every m′ ∈ M ′, so, the orthogonal projection, p, of H on K comutes with M ′, and therefore p ∈ M. Since H0 ⊂ K, it follows that (1 − p)H0 = {0} . Since by i) H0 is a separating set for M, we have 1 − p = 0 and thus K = H.To prove the second part of ii), let m′ ∈ M ′ be such that m′H0 = {0} . It follows that M m′H0 = {0} , so m′M H0 = {0} . Since, by i) M H0 is dense in H, it follows that m′ = 0. Some of the statements in the next Proposition are probably known, but we did not find a reference for any of them. Proposition 2.4. i) Let M ⊂ B(H) be a von Neumann algebra in standard form. Then, every normal linear functional, ϕ, on M is a vector functional, that is, there exist ξ, η ∈ H such that ϕ(m) = hmξ, ηi , m ∈ M. ii) If N ⊂ B(H) is an abelian von Neumann algebra, not necessarily in standard form, then every normal linear functional on N is a vector functional. iii) If M0 ⊂ B(H0) is a maximal abelian von Neumann algebra, then it is spatially isomorphic with its standard form. Proof. i) Let p ∈ M be a countably decomposable projection. According to [8, 9.6.18.], the hypotheses of [8, 9.6.20.] are satisfied, so there exists ξ0 ∈ H such that J ξ0 = ξ0 and M ′ξ0 = pH. Therefore, every countably decomposable projection p ∈ M is a cyclic projection. Now, let ϕ be a normal linear functional on M. By the polar decomposition of normal linear functionals [8, Theorem 7.3.2.], it is enough to prove the statement for normal positive linear functionals. Let ψ be a normal positive functional on M and p its support (that is, p is the complement of the supremum of all projections q ∈ M for which ψ(q) = 0). Then, p is countably decomposable, so by the previous arguments, p is a cyclic projection. Applying [8, Proposition 7.2.7.] it follows that ψ is a vector normal positive functional. ii) Let ϕ be a normal linear functional on M. As argued in i), using the polar decomposition of normal linear functionals it is enough to prove the statement in ii) for normal positive functionals. Let ψ be a normal positive functional on M and p its support which is a countably decomposable projection. Without loss of generality we can assume that p = I. We will show that there exists a separating 6 vector, ξ0 ∈ H for N and therefore, cyclic for N ′. Let {ξi ∈ H : i ∈ A} be a maxi- mal family of orthogonal unit vectors such that, the projections {pi ∈ N : i ∈ A} onto(cid:8)Hi = N ′ξi ⊂ H : i ∈ A(cid:9) are mutually orthogonal, so H =P⊕ Hi. Since Suppose A ⊆ N. Let ξ0 =Pi∈A p = I is countably decomposable, it follows that the set A is at most countable. 2i ξi. and let m ∈ N + be such that mξ0 = 0. Hence hmξ0, ξii = 0, i ∈ A. Then, since N is abelian, so N ⊆ N ′, it follows that hmξi, ξii = 0 for every i. Therefore, since m ∈ N +, it follows that mξi = 0, so mHi = {0} for every i, and thus m = 0. Hence ξ0 is separating for N. The statement ii) follows from [8, 7.2.7.]. 1 iii) Let {pι} be a maximal family of mutually orthogonal countably decom- posable projections of M0 and {ϕι} a family of positive linear functionals such that the support of ϕι is pι. ClearlyP pι = I. Let ρ0 =P ϕι. Then ρ0 is a faith- ful normal semifinite weight on M + 0 . By ii) for every ι there exists ξι ∈ pιH0 such that ϕι(m) = hmξι, ξιi , m ∈ M0. Obviously, ξι is a cyclic and separating vector of M pιpιH0 for every ι.It is also clear that hm1ξι, m2ξιi = ϕι(m∗ 2m1), m1, m2 ∈ M0, for every ι. If Nρ0 is as above, Nρ0 = {m ∈ M0 : ρ0(m∗m) < ∞} , then the mapping mpι → mξι extends to a unitary operator from Hρ0 to H0 and we are done. Corollary i) Every von Neumann algebra in standard form is hereditarily reflexive. ii) [19, Theorem 2] Every abelian von Neumann algebra, not necessarily in standard form is hereditarily reflexive. Proof. i) Follows from Proposition 2.4. i) and [10, Theorem 3.5.]. ii) Follows from Proposition 2.4. ii) and [10 Theorem 3.5.]. 2.2 W-dynamical systems with compact abelian groups Let (M, G,α) be a W-dynamical system, where M is a W ∗−algebra, G is a compact abelian group with dual Γ, and α a faithful w∗−continuous action of G on M, .that is αg 6= id if g 6= 0, where id is the identity automorphism of M and the mapping g → ϕ(αg(m)) for every m ∈ M and every ϕ ∈ M∗, where M∗ denotes the predual of M. For each γ ∈ Γ, denote by Mγ =(cid:26)Z hg, γiαg(m)dg : m ∈ M(cid:27) . where the integral is taken in the w∗−topology. In particular, if γ = 0, M0 is the fixed point algebra of the system. It can immediately be checked that Mγ = {m ∈ M : αg(m) = hg, γi m} 7 It is clear that the mapping Pγ : M → Mγ defined by Pγ(m) =Z hg, γiαg(m)dg is a w-continuous projection of M onto the closed subspace Mγ ⊂ M. In particular, P0 is a w-continuous projection of M onto M0 which is clearly faithful (on M +). It is well known that M is the w∗-closed linear span of {Mγ : γ ∈ Γ} . The Arveson spectrum of the action α is, by definition ([1], [13]) sp(α) = {γ ∈ Γ : Mγ 6= {0}} . Lemma 2.5. i) M−γ = M ∗ ii) Mγ1Mγ2 ⊂ Mγ1+γ2 where Mγ1Mγ2 is the linear span of {xy : x ∈ Mγ1 , y ∈ Mγ2} . iii) If m ∈ Mγ has polar decomposition m = u x , then u ∈ Mγ and γ = {m∗ : m ∈ Mγ} M ∗ γ = {m∗ : m ∈ Mγ}. γ ,where M ∗ x ∈ M0. Proof. i) and ii) are obvious. iii) is a straightforward consequence of the uniqueness of the polar decomposition of m. Let (M, G, α) be as above, ρ0 a faithful normal semifinite weight on M0 and ρ = ρ0 ◦ P0. Consider the corresponding normal faithful representation πρ on Hρ and the Tomita-Takesaki operators S, J as in 2.1. above. As in 2.1. we will write H instead of Hρ and M instead of πρ(M ). In the case when Γ is a partially ordered group, this representation will allow us to construct a generalized Hardy space on which, in certain situations, the subalgebra of analytic elements of the system (M, G,α) is hereditarily reflexive. For every g ∈ G define the unitary operator Ug ∈ B(H) as the unique extension of Ug(n) = αg(n), n ∈ Nρ to H. Then, since clearly, Nρ is an α- invariant left ideal of M, it is straightforward to check that the group of unitary operators {Ug : g ∈ G} implements the action α. Also, from the definition of S it follows that SUg = UgS and S ∗Ug = UgS ∗ for all g ∈ G. Therefore, J Ug = UgJ, g ∈ G. It follows that the group {Ug : g ∈ G} implements an action α′ of G on M , namely ′ α′ g(J mJ) = UgJ mJ U ∗ g = J αg(m)J. Similarly with the projections Pγ of M onto Mγ, γ ∈ Γ one can define the projections P ′ γ of M ′ onto M ′ γ =(cid:8)x ∈ M ′ : α′ γ(x) =Z hg, γiα′ P ′ g(x) = hg, γi x(cid:9) g(x)dg. The proof of the following lemma is a straghtforward application of the definitions. Lemma 2.6. With the notations above, we have the following: i) α′ g is an action of G on M ′, where M ′ is the commutant g(m′) = Ugm′U ∗ of M in B(H). 8 ii) If g ∈ G, then Ug commutes with J, and J αg(m)J = α′ g(J mJ), m ∈ M, g ∈ G. iii) ( M ′)γ = JM−γJ, γ ∈ Γ. iv) Ug(m′ξ) = α′ v) sp(α) = sp(α′). g(m′)ξ, ξ ∈ H0, m′ ∈ M ′. We will need also the following Remark 2.7. If M is a finite W-algebra, then M ′ is a finite W- algebra.This fact is immediate from the definition of the standard representa- tions. Let Hγ =(cid:26)Z hg, γiUg(ξ)dg : ξ ∈ H(cid:27) = {ξ ∈ H : Ugξ = hg, γi ξ} . Then, the map P H γ from H to Hγ defined as follows P H γ (ξ) =Z hg, γiUg(ξ)dg : γ ∈ Γ, ξ ∈ H. is an orthogonal projection of H onto the closed supspace Hγ . Applying Lemma 2.2., we see that if γ = 0, the Hilbert subspace H0 ⊂ H coincides with the Hilbert subspace H0 considered in Section 2.1. Lemma 2.8. i) If γ1 6= γ2, then Hγ1 and Hγ2 are orthogonal. ii) For every γ ∈ sp(α),we have MγH0 = Hγ , where MγH0 = {mξ0 : m ∈ Mγ, ξ0 ∈ H0} . iii) The direct sum of Hilbert spaces XHγ equals H. iv) For every γ ∈ sp(α) we have (M ′ )γ H0 = Hγ , where ( M v) For all γ, γ ′ ∈ sp(α) we have MγHγ ′ ⊂ Hγ+γ ′ and M ′ )γ , ξ0 ∈ H0}. )γ H0 = {m ξ0 : m ′ ′ ′ ∈ (M ′ γHγ ′ ⊂ Hγ+γ ′. Proof. i) Let ξ ∈ Hγ1 , η ∈ Hγ2 . Then, by definition, Ug(ξ) = hg, γ1i ξ and Ug(η) = hg, γ2i η, for all g ∈ G. Since the operators Ug are unitary, we have hξ, ηi = hUgξ, Ugηi = hg, γ1 − γ2i hξ, ηi , g ∈ G. Hence, if γ1 6= γ2 it follows that hξ, ηi = 0. ii) Since, by Lemma 2.3. i), H0 is a cyclic set for M, the subspace M H0 = γ and Pγ are the above pro- {mξ0 : m ∈ M, ξ0 ∈ H0} is dense in H. Then, if P H jections, we have MγH0 = Pγ(M )H0 =(cid:26)Z hg, γiαg(m)ξ0dg : m ∈ M, ξ0 ∈ H0(cid:27) = =(cid:26)Z hg, γiUg(ξ)dg : ξ = mξ0, m ∈ M, ξ0 ∈ H0(cid:27) = . 9 =(cid:8)P H γ (ξ) : ξ = mξ0, m ∈ M, ξ0 ∈ H0(cid:9) . Since by Lemma 2.3. i) the subspace M H0 is dense in H and P H projection, the result stated in ii) follows. γ is an orthogonal iii) Let η ∈ H be such that η ⊥ Hγ for all γ ∈ Γ. Since M is the w∗-closed linear span of {Mγ : γ ∈ Γ} , it follows from ii) that η ⊥ M H0, so, since by Lemma 2.3. i) H0 is cyclic for M, it follows that η = 0. iv) The proof is similar with that of ii) taking into account that, according to Lemma 2.3. ii), H0 is cyclic for M ′ as well. v) Immediate from definitions. 3 Hereditary reflexivity of generalized Hardy al- gebras In this section we will construct the generalized Hardy space and the generalized Hardy algebra and prove the main results of this paper, Theorem 3.8. and Theorem 3.9. Let (M, G, α) be a W-dynamical system with G compact abelian. Through- out this section we will assyme that M ⊂ B(H) where H is the Hilbert space constructed in Section 2.1. Suppose, in addition, that Γ is an archimedean lin- early ordered discrete group, or Γ ⊆ Πι∈I Γι is the direct product or the direct sum of archimedean linearly ordered (discrete) groups Γι. If (Γι)+ is the semi- group of non negative elements of Γι, denote by Γ+ = Πι∈I (Γι)+ . Then, Γ+ is a sub semigroup of Γ such that and Γ+ ∩ (−Γ+) = {0} Γ+ − Γ+ = Γ so Γ+ defines a partial order on Γ, namely γ1 6 γ2 if γ2 − γ1 ∈ Γ+. Lemma 3.1. Let γ1, γ2 ∈ Γ+, γ1 6= γ2. Then, either i) γ1, γ2 are not comparable under the above order relation, or, ii) γ1 < γ2, or, iii) γ1 > γ2 and, in this case, there exists p ∈ N such that either iii a) γ1 > pγ2 and γ1 < (p + 1)γ2, or, iii b) γ1 > pγ2 and γ1 and (p + 1)γ2 are not comparable. Proof. Suppose that γ1, γ2 are comparable. Thus, either γ1 < γ2, or γ1 > γ2. Suppose that γ1 > γ2. Since γ1, γ2 ∈ Πι∈I (Γι)+ , we can write 1 > γι0 γ1 = (γι 2 for some ι0 ∈ I. Since for every ι ∈ I , Γι has an archimedean order, there exists a largest pι ∈ N such that γι 2. If L ⊂ N is the set of non repeating p′ ιs 2 ∈ (Γι)+ , so γι 2, ι ∈ I and γι0 1)ι∈I , γ2 = (γι 2)ι∈I with γι 1, γι 1 > γι 1 > pιγι 10 then, since N is well ordered, there exists p = min L. Therefore, γ1 > pγ2. By the definition of p ∈ N, γ1 (cid:11) (p + 1)γ2, so either iii a) or iii b) must hold. The following consequence of the above Lemma will be used Corollary 3.2. Let γ0 ∈ Γ+ r {0} and γ ∈ Γ+.Then, there exists p ∈ Z+ such that either pγ0 6 γ < (p + 1)γ0 or pγ0 6 γ and γ is not comparable with (p + 1)γ0. Proof. If γ < γ0 or γ and γ0 are not comparable, then p = 0 satisfies the conclusion. If γ > γ0, the statement follows from Lemma 3.1. iii). Lemma 3.3. If Γ ⊆ Πι∈IΓι is the direct product or the direct sum of linearly ordered discrete groups Γι, then Γ is lattice ordered (see [3]), i.e. if A = {γ1, γ2, ...γn} is a finite subset of Γ then there exists inf A and sup A in Γ. Proof. sup A = (νι)ι∈I . j(cid:1) , 1 6 j 6 n, let µι = min(cid:8)γι j : j = 1, 2, ...n(cid:9) and If γj = (cid:0)γι j : j = 1, 2, ...n(cid:9) for each ι ∈ I. Then clearly inf A = (µι)ι∈I and νι = max(cid:8)γι If (M, G, α), M ⊂ B(H), H0, bG = Γ and Γ+ are as above, define H+ = Xγ∈Γ+ Hγ Let p+ be the orthogonal projection of H onto H+. By Lemma 2.8. v), the (closed) subspace H+ ⊂ H is invariant for ∨γ>0Mγ, and for ∨γ>0(M ′)γ .We will denote by M+ the weak operator closure in B(H+) and similarly M+ = p+(∨γ∈Γ+Mγ)p+ wo (M ′)+ = p+(∨γ>0(M ′)γ)p+ wo where ∨γ>0Mγ , is the algebra generated by {Mγ : γ > 0}) and ∨γ>0(M ′)γ is the algebra generated by {(M ′)γ : γ > 0}). Then, we will call H+ the generalized Hardy space and M+ the generalized Hardy algebra of analytic elements of the dynamical system (M, G, α). To prove hereditary reflexivity, we also need the following Lemma 3.4. If ψ is a weakly continuous functional on M+ ⊂ B(H+), then ψ is a vector functional (that is, there exist ξ, η ∈ H+ such that ψ(m) = hmξ, ηi for all m ∈ M+). Proof. Since ψ is weakly continuous, there exist n ∈ N and ξi, ηi ∈ H+, 1 6 i 6 n such that ψ(m+) =Pi hm+ξi, ηii , m+ ∈ M+. Now let eψ be the functional 11 Proposition 2.4. to this restriction, it follows that there exist ξ, η ∈ H such on defined by eψ(b) =Pi hbξi, ηii , b ∈ B(H). Since ξi, ηi ∈ H+, it follows that eψ(b) = eψ(p+bp+), b ∈ B(H), where, as above, p+ is the projection of H onto H+. The restriction of eψ to M is a normal linear functional of M. Applying that eψ(m) = hmξ, ηi , m ∈ M. Since, as noticed before, eψ(m) = eψ(p+mp+), we can take ξ, η ∈ H+. Therefore, in particular, ψ(mγ) = eψ(mγ) = eψ(p+mp+) = ψ(mγ) = hmγ ξ, ηi for every mγ ∈ Mγ, γ ∈ Γ+. The definition of M+ implies that ψ(m+) = hm+ξ, ηi , m+ ∈ M+ and the proof is completed. Loginov and Sul'man [10, Theorem 2.3.] have shown, in particular, that if a reflexive algebra satisfies the hypothesis of Lemma 3.4. then it is hereditarily reflexive, that is, all unital weakly closed subalgebras are reflexive. Under the name of super reflexivity this fact has been also considered in [4, Proposition 2.5. (1)]. In Theorem 3.8. below we will assume that (M, G, α) satisfies the following condition: (C) For every γ ∈ sp(α) \ {0} there exists an element uγ ∈ Mγ such that γ = eγ where eγ is a central projection of M and Mγ = M0uγ. For u∗ γuγ = uγu∗ γ = 0 we will take u0 = I. Examples of dynamical systems (M, G, α) satisfying this condition include the following: a) If M is a finite W-algebra and the center, Z(M0), of M0 is contained in the center, Z(M ), of M [18]. This is the case, in particular, when M is a finite W-algebra and M0 is a factor (this case will be discussed in a more general context in part b)). The conditions M finite and Z(M0) ⊂ Z(M ) also hold if M = ⊕Mi and (Mi, G, α) is a finite W-algebra and the fixed point algebra is a factor. Corollary 3.14. below will refer to these example. b) If M is a semifinite injective von Neumann algebra such that M0 is a factor, except when M is type III and M0 is a type II1 factor [21]. Thomsen has proved that in these cases, the action α has full unitary spectrum, that is every nonzero spectral subspace contains unitary operators. This is the case, in particular, when M is a finite W-algebra and M0 is a factor. In particular, this latter situation occurs if α is a prime action of the compact abelian group G on the hyperfinite type II1 factor [6], [7], in particular if α is egodic. Recall that an action is called prime if the fixed point algebra is a factor. In particular if the action α is faithful, then the all the examples in this part b) satisfy sp(α) = Γ. Corollary 3.15. below will refer to these examples. Also, the Condition (C) is satisfied if M is the crossed product of a von Neumann algebra M0. by an abelian discrete group Γ. Corollaries 3.16. and 3.17. will consider this case. Lemma 3.5. Suppose that condition (C).is satisfied. Then 12 i) Mγ = uγM0 ii) There exists an element wγ ∈ M ′ = JM J such that wγ w∗ γ = w∗ γ wγ = eγ, and (M ′)γ = (M ′)0wγ = wγ(M ′)0. then x = meγuγ = eγmuγ = uγ(u∗ i) Clearly, if x = muγ for some Proof. γmuγ) ∈ uγM0 and conversely. ii) Obviously, wγ = J u∗ γJ satisfies the equality wγ w∗ γ = w∗ γ wγ = J eγJ.Since eγ is a central projection of M, we can apply [8, 9.6.18.] to get J eγJ = eγ. Lemma 3.6. Let (M, G, α), M ⊂ B(H) be a W-dynamical system with 6= 0, we have G compact abelian as above. Then, if γ, γ ′ ∈ sp(α) and eγeγ ′ γ ′ − γ ∈ sp(α) (therefore, γ − γ ′ ∈ sp(α)) and eγeγ ′ 6 eγ−γ ′. Proof. Since eγeγ ′ 6= 0. Hence, applying Lemma 2.5. ii), it follows that Mγ ′−γ 6= 0, so γ ′ − γ ∈ sp(α). To prove the last statement of the lemma, notice that by Lemma 2.5. and Lemma 3.5. 6= 0, we have uγu∗ 6= 0. so u∗ γuγ ′u∗ γ ′ γuγ ′ eγ eγ ′ = uγu∗ γuγ ′u∗ γ ′ = uγuγ ′−γmu∗ γ ′ = uγ uγ ′−γ meγ ′u∗ γ ′ = uγ uγ ′−γeγ′ mu∗ γ ′ for some m ∈ M0. Further, using repeatedly Lemma 3.5. we get eγeγ ′ = uγuγ ′−γeγ′ mu∗ γ ′ = uγuγ ′−γ u∗ γ ′uγ ′mu∗ γ ′ = = uγu∗ γ ′(uγ ′uγ ′−γ u∗ γ ′)uγ ′mu∗ γ ′ = uγ−γ ′m1uγ ′−γ m2uγ ′mu∗ γ ′ = uγ−γ ′uγ ′−γ m3m2eγ−γ ′m1uγ ′mu∗ γ ′ = eγ−γ ′m4 for some m1, m2, m3, m4 ∈ M0\ {0} . Therefore, eγeγ ′ = eγ−γ ′m4m∗ 4eγ−γ ′ 6 km4m∗ 4k eγ−γ ′ So eγeγ ′ 6 eγ−γ ′. Lemma 3.7. Suppose that Condition (C) is satisfied. Then i) M+ is the w-closed subalgebra of B ( H+) generated by M0 and {uγ : γ ∈ sp(α), γ > 0} . ii) (M ′)+ is the w-closed subalgebra of B ( H+) generated by (M ′)0 and {wγ : γ ∈ sp(α′) = sp(α), γ > 0} . Proof. Follows from Lemma 3.5. We will prove next our results about reflexivity. In Theorem 3.8. we assume that Γ is archimedean linearly ordered and that Condition (C) is satisfied. Theorem 3.8. Let (M, G, α) be such that bG = Γ is archimedean linearly ordered and Condition (C) is satisfied. Then M+ ⊂ B(H+) is hereditarily reflexive. 13 In Theorem 3.9. we assume that Γ is a direct product (or a direct sum) of archimedean linearly ordered discrete groups, but we assume a stronger condi- tion than Condition (C). Theorem 3.9. Let (M, G, α) be such that bG = Γ ⊆ Πι∈I Γι is the direct product, or the direct sum, of archimedean linearly ordered discrete groups, Γι. Suppose that sp(α) = Γ and that for every γ ∈ sp(α) = Γ, there exists a unitary operator uγ ∈ Mγ. Then M+ ⊂ B(H+) is hereditarily reflexive. The proofs of these theorems will be given after some auxiliary results. Lemma 3.10. Suppose that Γ is archimedean linearly ordered and that Condition (C) is satisfied. Then i) (M+)′ = (M ′)+, where ( M+)′ denotes the commutant of M+ in B(H+) and ii) ((M ′ ′ )+) = M+ Proof. i) Let x ∈ (M ′)γ1 and m ∈ Mγ2 , γ1, γ2 ∈ sp(α) ∩ Γ+ . Then, p+xp+mp+ = p+xmp+ = p+mxp+ = p+mp+xp+ so, (M ′)+ ⊂ (M+)′. To prove the converse inclusion, let x ∈ (M+)′ ⊂ B(H+). Consider the following dense subspace of H+ H ′ =Xγ∈F Hγ : F ⊂ sp(α) ∩ Γ+ a finite subset. Then, clearly, the subspace . H ′′ = lin {uγξ : ξ ∈ H ′, γ ∈ sp(α)} . Pn where uγ is the partial isometry in Condition (C), is dense in H. Let η = i=1 uγi ξi ∈ H". Without loss of generality, we will assume in the rest of this proof that γ1 6 γ2 6 ... 6 γn. If uγi u∗ uγi = eγi , 1 6 i 6 n are the central projections from Condition (C) above, then, a standard calculation in the commutative W-algebra Z(M ), shows that γi = u∗ γi eγ1 ∨ eγ2 ∨ ... ∨ eγn = eγ1 + nXi=2 (1 − eγ1)...(1 − eγi−1)eγi . (1) where eγ1 ∨ eγ2 ∨ ... ∨ eγn = sup {eγi : 1 6 i 6 n} . Clearly, the terms of the above sum are mutually orthogonal central projections in Z(M ). So if η = i=1 uγi ξi ∈ H" it follows that Pn η = eγ1 η + nXi=2 (1 − eγ1)...(1 − eγi−1)eγi η. (2) 14 or, η = nXi=1 piη (3) where p1 = eγ1 and pi = (1 − eγ1)...(1 − eγi−1 )eγi , 2 6 i 6 n. Define the operator bx on H ′′ as follows We prove first that bx is well defined. Indeed, suppose thatPn will show thatPn nXi=1 i=1 uγi xξi = 0.Now, if η =Pn nXi=1 bx( piη = 0, 1 6 i 6 n. uγi ξi) = uγi xξi. i=1 uγi ξi = 0. We i=1 uγi ξi = 0, it follows that We must show that pi uγj xξj = 0, 1 6 i 6 n. nXj=1 These equalities imply thatPn i=1 uγi xξi = 0, sobx is well defined. Since piη = 0, nXj=1 piuγj piξj = 0. nXj=i piuγj ξj = we have (4) Thus, factoring out uγi uγi nXj=1 piu∗ γi uγj piξj = 0 By multiplying the above equality by u∗ γi we get , and taking into account that pi 6 eγi where u∗ γi x ∈ (M+)′, it follows that uγj ∈ Mγj −γi . Since γi 6 γj, if i 6 j 6 n so, Mγj −γi ⊂ M+ and x(pieγi ξi + nXi=2 piu∗ γi uγj ξj) = pieγi xξi + nXj=i+1 piu∗ γiuγj xξj = 0. By multiplying the above equality by uγi , we get piuγixξi + nXj=i+1 piuγj xξj = pi uγj xξj = pi nXj=i uγj xξj = 0. nXj=i 15 so piu∗ γi uγj ξj = 0. nXj=i pieγi ξi + piu∗ γi uγj ξj = 0. nXj=i+1 Further, since γi 6 γj, and pjuγi = 0 when i 6 j and x ∈ (M+)′ we have and this proves (4). Therefore, bx is well defined. From the definition of bx it follows that bxm = mbx on H" for all m ∈ M andbx(η) = x(η) for every η ∈ H ′, so if, as we will prove,bx is bounded, it follows thatbx ∈ M ′. Next we prove that the operatorbx is bounded. Indeed, if, as above, η =P uγi eγi ξi, then, using the equality (3) and the fact thatbx commutes with pi, 1 6 i 6 n, we have nXi=1 kpibx (piη)k2. piuγj xξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXj=i γi uγj ξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Therefore, kbx(η)k 6 kxk kηk so bx is bounded. As noticed above, bx ∈ M ′ and, since obviously, p+bxp+ = x it follows that x ∈ (M ′)+ and we are done. =vuut kbx(η)k =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)bx nXi=1 piη!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) pibx (piη)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi=1 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) piuγj ξj)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kpibx (piη)k2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXj=i nXj=1 pibx( 6(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) uγj xξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) piuγj xξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXj=i nXj=i 6 kxk2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) γi uγj xξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXj=1 nXj=i γiη(cid:13)(cid:13)2 = kxk2(cid:13)(cid:13)piu∗ uγj xξj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ii) follows from i) by replacing M with M 6(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 6 kxk2 kpiηk2 . pix piu∗ pi piu∗ γi pi piu∗ 2 = 2 = 2 6 pi piuγi piu∗ γi 2 pi 2 2 2 ′ . The following version of Lemma 3.10. will be used in the proof of Theorem 3.9. Here and in Theorem 3.9 we assume that Γ is a direct product or a direct sum of archimedean linearly ordered discrete groups, but we also assume a stronger version of Condition (C), namely that for every γ ∈ sp(α), Mγ contains a unitary operator and that the action is faithful (i.e. αg = id implies g = 0). These conditions imply that sp(α) = Γ (see for instance [21 Lemma 2.2.]). Lemma 3.11. Let (M, G, α) be such that bG = Γ is a direct product, Γ = Πι∈IΓι, (or a direct sum) of archimedean linearly ordered discrete groups Γι. Suppose that α is faithful and for every γ ∈ sp(α), there exists a unitary operator uγ ∈ Mγ (as noticed above, these conditions imply that sp(α) = Γ). Then i) (M+)′ = (M ′)+, where ( M+)′ denotes the commutant of M+ in B(H+) and ii) ((M ′ ′ )+) = M+. Proof. it follows that (M ′)+ ⊂ (M+)′. To prove the opposite inclusion, let x ∈ (M+)′. Further, let us i) As in the proof of the previous Lemma 3.10. 16 denote and H ′ =Xγ∈F Hγ : F ⊂ Γ+ a finite subset. H ′′ = lin {uγξ : ξ ∈ H ′, γ ∈ Γ} . on H" as follows Clearly H ′ is dense in H+ and H" is dense in H. Define the linear operator bx uγi ξi) = uγi xξi. We will prove first that bx is well defined. Suppose that uγiξi = 0. (5) nXi=1 bx( nXi=1 nXi=1 Since, by Lemma 3.3., Γ is lattice ordered, let ν = inf {γi : 1 6 i 6 n} ∈ Γ. Since, by hypothesis sp(α) = Γ, we have ν ∈ sp(α). Let uν ∈ Mν be a unitary operator as in the hypothesis. By multiplying (5) by u∗ ν we get uγi−ν ξi = 0. nXi=1 Since γi − ν ∈ Γ+, i = 1, 2, ..., n and x ∈ (M+)′ it follows that x uν so u∗ uγi ξi) = 0 u∗ ν uγi ξi = u∗ ν uγi xξi = nXi=1 ν uγi xξi = 0. nXi=1 nXi=1 nXi=1 uγi xξi =bx( sobx is well defined. We will prove next that bx is continuous. Indeed ν uγi xξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) uγi xξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi=1 nXi=1 ν uγi ξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 6 kxk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ν uγiξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi=1 nXi=1 uγi ξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = kxk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) uγi ξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi=1 nXi=1 sobx is continuous. As in the proof of the previous Lemma 3.10. we can see that bx ∈ M ′ and p+bxp+ = x, so x ∈ (M ′)+. nXi=1 uγiξi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)bx( nXi=1 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ν uγi xξi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXi=1 kxk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ii) follows from i) by replacing M with M x u∗ u∗ ν uν u∗ u∗ u∗ = = . ′ . 17 It is worth mentioning that the previous two Lemmas imply, in particular, that (M+) " = M+ but we will not use this fact. In the next two lemmas we will assume that (M, G, α) is a W-dynamical system that satisfies Condition (C) and M is in standard form. We also assume that Γ is a direct product or sum of archimedean linearly ordered abelian discrete groups. If γ0 ∈ sp(α) ∩ Γ+ r {0}and γ ∈ sp(α) ∩ Γ+ according to Corollary 3.2. there exists p ∈ Z+ such that either pγ0 6 γ < (p + 1)γ0 or pγ0 6 γ and γ is not comparable with (p + 1)γ0. If, in addition, eγeγ0 6= 0 we will denote for every λ ∈ C, λ < 1 x(λ, γ0, γ, ξ) =Xn>0 λnun γ0 y(λ, γ0, γ, ξ) =Xn>0 λnwn and Kγ0,γ,λ = Lγ0,γ,λ = uγ−pγ0ξ : ξ ∈ eγ0 H0 γ0 wγ−pγ0 ξ : ξ ∈ eγ0 H0 ⊂ eγ0 H+. ⊂ eγ0 H Lemma 3.12. Let (M, G, α) be a W-dynamical system that satisfies Con- dition (C). Γ and γ0, γ ∈ sp(α) ∩ Γ+ be as above. Then, Kγ0 = lin {Kγ0,γ,λ : γ ∈ sp(α) ∩ Γ+, λ ∈ C, λ < 1} is dense in eγ0 H+. Proof. Notice first that if γ0, γ are as in the hypothesis of the lemma, then, by the definition of eγ0 in the Condition (C),we have that pγ0 ∈ sp(α) for every p ∈ Z+ and, by Lemma 3.6., γ − pγ0 ∈ sp(α), so uγ−pγ0 exists, and thus the definition of Kγ0,γ,λ in the hypothesis of the Lemma is consistent. Now, taking λ = 0 in Kγ0,γ,λ, it follows that eγ0 uγ−pγ0 H0 ⊂ Kγ0,γ ⊂ Kγ0 . In particular, for p = 0 (so, when either 0 6 γ < γ0 or 0 6 γ and γ is not comparable with γ0), we have eγ0 uγH0 = eγ0 Hγ ⊂ Kγ0. We will prove that eγ0 Hγ = uγeγ0 H0 ⊂ Kγ0 eγ0 Hγ ⊂ Kγ0 and eγ0 Hγ ⊂ K γ0 , so Kγ0 is dense in eγ0 H+ as claimed. uγ−pγ0 H0 ⊂ Kγ0,γ . The case p = 0 was proved for every γ ∈ sp(α) ∩ Γ+. This fact will imply that Pγ∈Γ+ therefore, eγ0 H+ =Pγ∈Γ+ We will prove first that eγ0 up γ0 above. Suppose that p > 0. We will prove by induction on k that eγ0 uk γ0 uγ−pγ0 H0 ⊂ Kγ0 for every k, in particular for k = p. If k = 0, the above inclusion follows, as noticed at the beginning of this proof from the definition of Kγ0,γ for λ = 0. Suppose by induction that ul γ0uγ−pγ0 ξ ∈ Kγ0 for l = 0, 1, ...k − 1.Then λnun γ0 uγ−pγ0 ξ ∈ Kγ0 Xn>k 18 for every ξ ∈ eγ0 H0, λ ∈ C, λ < 1. Thus λkuk γ0 uγ−pγ0ξ + Xn>k+1 λnun γ0 uγ−pγ0 ξ ∈ Kγ0 By dividing the above relation by λk, λ 6= 0 and then taking the limit as λ → 0, we get that uk γ0 uγ−pγ0 ξ ∈ Kγ0 for every ξ ∈ H0. Since, obviously, up ii), M−pγ0Mγ ⊂ Mγ−pγ0, it follows that γ0)p = eγ0 for p > 0 and, by Lemma 2.5. γ0 uγ−pγ0ξ ∈ Kγ0 , so, in particular, up γ0(u∗ eγ0 Hγ = eγ0 uγH0 = up γ0(u∗ γ0)puγH0 ⊂ up γ0 uγ−pγ0H0 ⊂ Kγ0 and we are done. The following lemma can be proven similarly with the previous Lemma 3.12. Lemma 3.13. Let (M, G, α) be a W-dynamical system that satisfies Con- dition (C), Γ and γ0, γ ∈ sp(α) ∩ Γ+ be as above. Then, Lγ0 = lin {Lγ0,γ,λ : pγ0 6 γ < (p + 1)γ0 for some p ∈ Z+, λ ∈ C, λ < 1} is dense in eγ0 H+. ′ ′ ′ )0p+ ⊂ B(H+). Therefore, for every projection f ∈ (M Proof. of Theorem 3.8. We will prove first that M+ ⊂ B(H+) is reflexive and then apply Lemma 3.4. and the subsequent discussion to infer that M+ is hereditarily reflexive. Let γ0 ∈ sp(α)∩Γ+. Since uγ0 ∈ M, it follows that uγ0 f = f uγ0 for every projection f ∈ (M ′), in particular for every projection f ∈ (M ′)0. By Lemma 2.8. iv), since (M ′)γ H0 = Hγ , γ ∈ sp(α), we have, in particular that )0, f H+belongs to (M Lat(M+). Let x ∈ algLat(M+) ⊂ B(H+). We will prove that x ∈ ((M and, then, applying Lemma 3.11. ii) it will follow that x ∈ M+, so M+ is a reflexive operator algebra. The way to prove this fact is to use Lemma 3.14. to show that x∗ ∈ ((M = M+.As noticed above, f H+ ∈ LatM + for every projection f ∈ (M )0,so xf = f x and therefore x∗f = f x∗ for every projection f ∈ (M )0. It follows that x commutes with every element of (M )0. We will prove next that xwγ = wγ x for every γ ∈ sp(α′) ∩ Γ+ = sp(α) ∩ Γ+ and then apply Lemma 3.10. ii) to infer that x ∈ = M+. To this end, let γ0 ∈ sp(α) ∩ Γ+. If γ0 = 0, then as convened, ((M uγ0 = wγ0 = I, so nothing to prove. Let γ0 ∈ sp(α) = sp(α′), γ0 > 0. Denote ) the operator uγ0 ∈ M+ (respectively wγ0 ∈ (M ′)+) by Tuγ0 defined on H+. Then the adjoints of Tuγ0 and then, clearly, it will follow that x ∈ ((M (respectively Twγ0 on H+ are , Twγ0 )+) ′ ′ )∗ +) )+) ′ ′ )+) ′ ′ ′ ′ ′ ′ T ∗ uγ0 ξγ = 0 if 0 6 γ < γ0 and T ∗ uγ0 ξγ = u∗ γ0 ξγ if γ > γ0. and similarly T ∗ wγ0 ξγ = 0 if 0 6 γ < γ0 and T ∗ wγ0 ξγ = w∗ γ0 ξγ if γ > γ0. 19 Since uγ0 , wγ0 (so Tuγ0 well. Notice that if, for λ ∈ C, λ < 1, we denote , Twγ0 ) commute, it follows that T ∗ uγ0 , T ∗ wγ0 commute as eLγ0,λ =nξ ∈ H+ : T ∗ wγ0 ξ = λξo . wγ0 then, since T ∗ x ∈ algLat(M+) it follows that.Lγ0,γ,λ ∈ Lat(x∗) for every λ ∈ C, λ < 1. Since for the eigenvalue λ, then, it follows that commutes with (M+)∗, we haveeLγ0,γ,λ ∈ Lat(M+)∗ and, since x∗ξ for every ξ ∈eLγ0,λ . On the other hand, it is clear that eLγ0,γ,λ consists of eigenvectors of T ∗ ξ = T ∗ x∗T ∗ wγ0 wγ0 wγ0 Lγ0,γ,λ ⊂eLγ0,λ wγ0 where Lγ0,γ,λ is as in Lemma 3.13., so x∗T ∗ By Lemma 3.13., ξ = T ∗ wγ0 x∗ξ for every ξ ∈ Lγ0,γ,λ . Lγ0 = lin {Lγ0,γ,λ : pγ0 6 γ < (p + 1)γ0 for some p ∈ Z+, λ ∈ C, λ < 1} x∗ξ for every ξ ∈ eγ0 H+, so is dense in eγ0 H+, and therefore x∗T ∗ + . Since x∗ x∗ commutes with T ∗ wγ0 commutes with (M ′)∗ + , it follows that x commutes with (M ′)+, so by Lemma 3.10. ii) x ∈ M+ so M+ is reflexive. Finally, by applying Lemma 3.4. and the discussion following it, we see that M+ is hereditarily reflexive and we are done. , γ0 ∈ sp(α) ∩ Γ+ and therefore with (M ′)∗ ξ = T ∗ wγ0 wγ0 Corollary 3.14. below refers to the Example a) to Condition (C). Corollary 3.14. Let (M, G, α) be a W -dynamical system with G compact abelian and M a finite W-algebra in standard form such that Z(M0) ⊂ Z(M ). Suppose that the dual Γ of G has an archimedean linear order. Then M+ ⊂ B(H+) is reflexive. Proof. According to [18, Theorem 2.3.], if M is finite and Z(M0) ⊂ Z(M ), then the Condition (C) is satisfied and therefore the result follows from Theorem 3.8. Proof. of Theorem 3.9. The proof is very similar with the proof of Theorem 3.8. The only modification is using Lemma 3.11 instead of Lemma 3.10. The next Corollary refers to Examples b) to Condition (C). Corollary 3.15. Let (M, G, α) be a W -dynamical system with M an injective von Neumann algebra in standard form and G a compact abelian group such that the dual Γ of G is a direct product (or a direct sum) of archimedean linearly ordered discrete groups. Suppose that α is prime and faithful and it is not the case that M is of type III and M0 is of type II 1. Then M+ is hereditarily reflexive. 20 Proof. Since α is faithful, we have sp(α) = Γ (see for instance [21, Lemma 2.2.]). By [21, Theorem 2.3.] each spectral subspace Mγ contains a unitary operator. The conclusion of the Corollary follows from Theorem 3.9. The concept of nonselfadjoint crossed product, or more generally that of w-semicrossed product were defined in [3], [11], [16]. Corollary 3.16. Let (M0, Γ, β) be a W-dynamical system such that M0 ⊂ B(H0) is in standard form and Γ is a discrete abelian group. Suppose that Γ ⊆ ΠΓι is a direct product (or a direct sum) of archimedean linearly ordered groups. Let M = M0 ×α Γ ⊂ B(l2(Γ, H0)) be the corresponding crossed product . Then, the non selfadjoint crossed product M+ = M0 ×β Γ+ (i.e. the algebra of elements of M with non negative spectrum) is a hereditarily reflexive operator algebra in B(H+) where H+ = l2(Γ+, H0). Proof. If G denotes the (compact) dual of Γ, and α = bβ is the dual action of β on M, then, consider the canonical conditional expectation P0 : M → M0. Since M0 ⊂ B(H0) is in standard form, from Lemma 2.1. and the subsequent discussion it follows that M ⊂ B(l2(Γ, H0)) is in standard form and by the definition of the crossed product, Mγ contains a unitary operator uγ for every γ ∈ Γ and thus the W-dynamical system (M, G,bβ) satisfies the conditions of Corollary 3.15. Corollary 3.17. Let (M0, Γ, β) be a W-dynamical system such that M0 ⊂ B(H0) is a maximal abelian von Neumann algebra and Γ is a discrete abelian group. Suppose that Γ ⊆ ΠΓι is a direct product (or a direct sum) of archimedean linearly ordered groups. Let M = M0 ×α Γ ⊂ B(l2(Γ, H0)) be the corresponding crossed product . Then, the non selfadjoint crossed product M+ = M0 ×β Γ+ is a hereditarily reflexive operator algebra in B(H+) where H+ = l2(Γ+, H0). Proof. Since M0 ⊂ B(H0) is a maximal abelian von Neumann algebra, by Proposition 2.4. it is spatially isomorphic with its standard form, so the result folows from the previous Corollary 3.16. In [2, Corollary 5.14.] it is stated that if M0 ⊂ B(H0) is a maximal abelian von Neumann algebra and Γ = Zd, d ∈ N, then M0 ×β Γ+ is reflexive, so the Corollary 3.18. above extends that result by showing also hereditary reflexivity in the special case Γ = Zd, d ∈ N. REFERENCES 1. W. B. Arveson, The harmonic analysis of automorphism groups, operator algebras and applications, Proc. Sympos. Pure Math., vol. 38, Amer. Math. Soc., Providence, R. I., 1982. 2. R. T. Bickerton, and E. T. A. Kakariadis, Free multivariate w-semicrossed products: reflexivity and the bicommutant property, Canad. J. Math. 70(2018), 1201 -- 1235. 21 3. K, R. Davidson, A. H. Fuller and E. T. A. Kakariadis, Semicrossed Products of Operator Algebras by Semigroups, Memoirs of the AMS, Vol. 247, 2017. 4. D. W. Hadwin and E.A. Nordgren, Subalgebras of reflexive algebras. J. Operator Theory 7 (1982), 3 -- 23. 5. L. Helmer, Reflexivity of non-commutative Hardy algebras, J. Funct. Anal. 272(2017), 2752 -- 2794. 6. V. F. R. Jones, Prime actions of compact abelian groups on the hyperfinite type II1 factor, J. Operator Theory, 9(1983), 181-186. 7. V. F. R. Jones and M. Takesaki, Actions of compact abelian groups on semifinite injective factors, Acta Math. 153 (1984), 213 -- 258. 8. R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of operator algebras, Vol. II Advanced Theory,Academic Press 1986. 9. E. T. A. Kakariadis, Semicrossed products and reflexivity, J. Operator Theory, 67(2012), 379-395. 10. A I Loginov and V S Sul'man, Hereditary and intermediate reflexivity of W-algebras, Mathematics of the USSR-Izvestiya 9(1975), 1189-1202. 11. M. McAsey, P. S. Muhly, and K.-S. Saito, Nonselfadjoint crossed prod- ucts (invariant subspaces and maximality), Trans. Amer. Math. Soc., 248(1979), 381 -- 409. 12. P. S. Muhly and B. Solel, Hardy algebras, W-correspondences and interpolation theory, Math. Ann. 330 (2004), 353 -- 415. 13. G. K. Pedersen, C-algebras and their automorphism groups, Academic Press 1979. 14. C. Peligrad, Reflexive operator algebras on noncommutative Hardy spaces, Math. Annalen, 253(1980), 165-175. 15. C. Peligrad, Invariant subspaces of algebras of analytic elements associ- ated with periodic flows on von neumann algebras, Houston J. Math., 42(2016), 1331-1445. 16. J. R. Peters, Semicrossed products of C-algebras. J. Funct. Anal., 59(1984), 498 -- 534. 17. H. Radjavi and P. Rosenthal, Invariant subspaces, 2nd edition, Dover Publications, Mineola, New, York, 2003. 18. K.-S. Saito, Nonselfadjoint subalgebras associated with compact abelian group actions on finite von Neumann algebras, Tohoku Math. Journ. 34(1982), 485-494. 19. D. Sarason, Invariant subspaces and unstarred operator algebras, Pacific J. Math., 17(1966), 511-517. 20. S. Stratila, Modular theory in operator algebras, Bucharest; Abacus Press, Tunbridge Wells, 1981 21. K. Thomsen, Compact abelian prime actions on von Neumann algebras, Trans. Amer. Math. Soc., 315(1989), 255-273. 22
1003.2676	1	1003	2010-03-13T05:22:46	Automorphisms of the bipartite graph planar algebra	[ "math.OA" ]	For any abstract subfactor planar algebra $P$, there exists a finite index extremal subfactor $M_0 \subset M_1$ with $P$ as its standard invariant. In this paper, we classify the automorphism group of a bipartite graph planar algebra, and obtain subfactor planar subalgebras by taking fixed points under groups of automorphisms. This construction provides both new examples of subfactors and new descriptions of the planar algebras of previously known examples.	math.OA	math	Automorphisms of the bipartite graph planar algebra R. D. Bursteina,∗ aDepartment of Mathematics 1326 Stevenson Center Vanderbilt University Nashville, TN 37240 1. Introduction 0 1 0 2 r a M 3 1 ] . A O h t a m [ 1 v 6 7 6 2 . 3 0 0 1 : v i X r a Planar algebras were introduced by Jones in [13]. They are a powerful tool for studying subfactors, providing a graphical calculus on the standard invariant of an finite index extremal subfactor of type II1 ([13],[22]). The planar operad is the set of planar tangles with zero or more internal disks and a checkerboard shading, with a distinguished region of the boundary and each internal disk, all taken up to isotopy. The operation is gluing: a tangle may be pasted into an internal disk of another tangle (matching up the distinguished boundary regions) if the number of strands and the shading are compatible. A planar algebra is a graded vector space P = (V ± n ), n ≥ 0, along with an associative action of the planar operad. If M0 ⊂ M1 is a finite index extremal II1 subfactor with Jones tower M0 ⊂ M1 ⊂ M2 ⊂ ..., then we may take V + 1 ∩ Mn+1. There is an operad action defined in [13] making this into a planar algebra. Conversely, Jones describes a certain list of additional properties which make P into a Popa system, implying that there exists a subfactor M0 ⊂ M1 of which P is the standard invariant [22]. A planar algebra with this list of properties is called a subfactor planar algebra (SPA). Other methods of constructing a subfactor from an SPA have since been obtained ([11],[15],[16]), which use more diagrammatic notation than Popa's original method. 0 ∩ Mn, V − n = M ′ n = M ′ Constructing an SPA abstractly therefore implies the existence of a corre- sponding subfactor. This method may be used to find new subfactors (as in [6] or [1]), or provide new proofs of the existence of subfactors with specified proper- ties. Abstract constructions of SPAs of previously known subfactors can provide new insight into the structure of their standard invariants (e.g. [3],[4],[19]). A planar algebra may be constructed from any finite bipartite graph [14], and some infinite graphs as well [11]. These bipartite graph planar algebras (BGPAs) are almost never of subfactor type, because their vector spaces are too large. However, they possess several of the necessary properties required for ∗Corresponding author. Tel. 1-615-322-6672 Fax 1-615-343-0215 Email address: [email protected] (R. D. Burstein) Preprint submitted to Elsevier October 30, 2018 SPAs, which are inherited by planar subalgebras. We may therefore try to find SPAs by looking at small planar subalgebras of BGPAs. It is generally difficult to show that a graded subspace of a BGPA is closed under the action of the planar operad, although progress has been made in the single generator case (e.g. [7],[8],[19],[1]). In this paper, we describe a method for doing so by taking fixed points of a BGPA under a group of automorphisms. In section 2 we describe BGPAs, with particular attention to infinite graphs. We introduce a slightly different notation from [11] and [14]; this simplifies the computations of section 3, where we compute the automorphism group of an arbitrary BGPA PΓ. In section 4 we find conditions for a planar subalgebra of a BGPA to be an SPA, and we conclude in section 5 by presenting some examples of SPAs obtained by this planar fixed point construction. 2. The bipartite graph planar algebra The bipartite graph planar algebra (BGPA) is described by Jones in [14]. The data required are a finite bipartite graph Γ and a spin vector: i.e. a function from the vertices of the graph to the positive real numbers. Given such data, there is a planar algebra PΓ = (V ± n has a basis labelled by loops of length 2n in the graph, starting at even or odd vertices depending on sign. n ), n ≥ 0, where each V ± We will consider BGPAs on infinite graphs as well. In this case, the vector spaces will be infinite dimensional, so there are some new topological consider- ations. To simplify later computations, we will describes the vector spaces of the planar algebra as operators on a certain Hilbert space. In our diagrams, we will use the convention of [15] that a thick line represents as many parallel strands as necessary. Also we will omit shading in a diagram when both shadings can occur. When not otherwise specified, the distinguished boundary region of a tangle is on the left. The discussion that follows is taken from [14], adjusted where necessary to allow for infinite graphs. Let Γ be a locally finite bipartite graph, i.e. each vertex of Γ is an endpoint of at most A edges. Let µ be a function from the vertices of Γ to the positive real numbers obeying the following local boundedness condition: there is some M > 0 such that for any two adjacent vertices v and w, we have µ(v)/µ(w) < M . Let l± n be the set of loops of length 2n on Γ, starting at an even (+) or odd (−) vertex. Then the vector space V ± n of the bipartite graph planar algebra PΓ is the set of bounded functions from l± n to C. The boundary type of a tangle is the ordered pair (n,±), where 2n strands intersect the tangle boundary transversely and the distinguished region of the tangle is shaded (+) or unshaded (−). The same definition is used for the boundary type of each internal disk of a tangle. Let T be a planar tangle with k internal disks. Let the boundary type of T be (n,±), and let the internal disks of T have boundary types respectively (n1,±), (n2,±), ..., (nk,±). Then to describe the planar operad, for each set of 2 inputs (x1, ..., xk) (with xk ∈ V ± n . This assignment must agree with gluing, be multilinear in the inputs, and be isotopy invariant. We define Z(T ) following [14]. nk ) we must assign the output Z(T ) ∈ V ± A state of the tangle is a function σ which maps the strands of T to edges of Γ, and the regions of T to vertices of Γ. Shaded regions are mapped to positive vertices, and unshaded regions to negative ones. A state must obey a compatibility condition: if a strand S is adjacent to a region R, then σ(R) must be one of the endpoints of σ(S). A state σ is compatible with a given loop if treading the output of σ coun- terclockwise around the boundary of T , starting from the distinguished region, produces that loop. A singularity of T is a local maximum or local minimum of a strand. Fix a state σ; let v be the vertex associated to the concave side of the singularity by σ, and w the vertex associated to the convex side. Then the value of the singularity is µ(v)/µ(w). A state σ associates a loop Li to each internal disk Di of T , obtained by reading the output of σ counterclockwise starting from the distinguished region of the disk. The value of the state on Di is then the value of the ith input xi at this loop Li. Now we can define the value of Z(T ) on a specified loop L ∈ l± n . This is Xσ compatible with L   σ(Di)  Y σ(s) YDi ∈internal disks s∈singularities of T where σ is evaluated on disks and internal singularities as above. By local finiteness of Γ, only finitely many states are compatible with L. Specifically, let m be the largest number of strands that must be crossed to get from the distinguished boundary region of T to any other region (boundary or internal). Then if v is the first vertex of L, every compatible state must assign all regions of T to vertices within a distance of m from v on the graph. A state is determined by its value on strands. Let each vertex of Γ contact at most emax edges. Then each strand must be chosen from at most em max possibilities, and if T contains a distinct strands, each loop is compatible with at most eam max states. The local boundedness condition on µ (µ(v)/µ(w) < M < ∞ for all adjacent vertices v and w) means that each singularity has value at most M , for any state. Let T have b singularities. Elements of V ± ni are bounded; let N be the largest bound of any input xi, 1 ≤ i ≤ k. on each loop L as a finite sum. Moreover, this sum is bounded by eam so the output is a bounded function on l± element of V ± n . Putting this together, we find that the tangle output Z(T ) may be evaluated maxM bN k, n and this evaluation rule produces an The proofs of [14] that this map is multilinear, isotopy invariant and respects gluing may be used without alteration, since local finiteness of Γ implies that all necessary sums are finite. Therefore the above definition of Z(T ) produces a planar algebra. 3 There is a natural antilinear involution, which we refer to as ∗, of each vector space V ± n , the reversed loop L′ consists of the n (see [14]). For each loop L ∈ l± same list of vertices and edges, taken in the opposite order. Then the involution is defined by A∗(L) = A(L′), for all A ∈ V ± Let ρT be the map corresponding to some tangle T with k internal disks, i.e. ρT (x1 ⊗ x2 ⊗ ... ⊗ xk) is equal to Z(T ) when the inputs are (x1, x2, ..., xk). k) = ρT ′ (x1 ⊗ x2 ⊗ ... ⊗ xk)∗, where Then as in [14] we have ρT (x∗ the tangle T ′ is the mirror image of T . From [13], this means that the BGPA PΓ is a planar ∗-algebra, with involution as above. of bounded linear operators on a certain Hilbert space. For ease of later computation, it will be convenient to describe V ± n and L ∈ l± n . 2 ⊗ ... ⊗ x∗ n as a set 1 ⊗ x∗ Let H ± n be the Hilbert space with basis {xp} labelled by the paths of length n on Γ whose initial vertex is even (+) or odd (−). These vector spaces are all infinite dimensional when Γ is infinite. We also define the total path Hilbert space H as the direct sum of the H ± n 's: H =Pn,± H ± n . n . A loop L ∈ l± Each element A of Vn± naturally defines a linear map on H ± n may be described as a pair of paths (π, ǫ) of length n whose endpoints are the same. Then π and ǫ correspond to basis vectors xπ, xǫ in H ± n , and we take hA(xǫ), xπi = A(L). is equal to (ǫ, π). hA∗(xǫ), xπi = hA(xπ), xǫi. In other words, the involution acts on V ± adjoint operation for B(H ± n is equal to the pair of paths (π, ǫ), then the reversed loop L′ It follows from the above definition of the involution that n as the If L ∈ l± n ). Every path p has a starting vertex s(p) and a terminal vertex t(p). For each vertex v of Γ we may define a projection sv ∈ B(H ± n ) which fixes the closed linear span of {xps(p) = v}, and likewise tv which fixes the closed span of {xpt(p) = v}. The sv's and tw's form an abelian algebra; in fact the projections {svtwv, w ∈ vertices of Γ} are a partition of unity. Since π and ǫ above always have the same endpoints, we have each A ∈ V ± n commuting with sv and tv for all vertices v. This means that A ∈ V ± n may be thought of as a (potentially infinite) sum of operators Avw each acting on the subspace svtw(H ± n ). Each svtw is of finite rank, and this rank is universally bounded by local finiteness of Γ, so these subspaces have bounded dimension. Since the components of A are bounded by definition, it follows that A is bounded in norm as a linear operator on H. We then have V ± n ⊂ B(Hn)±, and in fact V ± n ⊂ {sv, twv, w ∈ vertices of Γ}′. Each svtwB(H ± n ) is a finite dimensional matrix algebra, with matrix units given by partial isometries from xp to xq where p and q are paths from of length n from v to w. All such partial isometries are contained in V ± n ) ⊂ V ± n for all v, w. {sv, tw}′ consists of bounded formal sums of elements of svtwB(H ± n ), so from the definition of V ± n = {sv, tw}′ as operators on H ± n . This is a von Neumann algebra by the bicommutant theorem. n 's; these are just the scalars We will freely refer to scalar elements of the V ± n it then follows that V ± n , so svtwB(H ± as operators on the appropriate Hilbert space. To assist with computations using the above notation, we now define a con- 4 catenation operation c on the H ± n 's: m → H ± Definition 2.1. Let p and q be two paths on the graph Γ. Then c(xp, xq) is zero if t(p) 6= s(q), and otherwise is xr where r is the path obtained by first following p and then q. This operation extends linearly and continuously to maps H ± n ⊗ H ± From the definition of the bases of the V ± n 's, every concatenation map is surjective. Furthermore, these maps are associative: c(c(x, y), z) = c(x, c(y, z)). This means we may freely apply c to multiple inputs via the inductive definition c(x1, x2, ..., xn) = c(c(x1, x2, ..., xn−1), xn). k , where k = n + m and the signs are chosen appropriately. We also define an antilinear path reversal operator rev: Definition 2.2. Let p be a path on the graph Γ. Let q be the reverse path of p, i.e. the same edges and vertices taken in the opposite order. Then rev(xp) = xq. This operation extends antilinearly and continuously to maps from H ± n to H ± n or H ∓ n , depending on the value of n. Both c and rev are bounded, so they extend to the total Hilbert space H. rev is an antilinear involution. Later, we will be interested in demonstrating that certain maps on the V ± n 's commute with the action of the planar operad. To do this it suffices to show that such maps commute with particular tangles that generate the planar operad (see e.g. [4]). We now describe the action of one set of such generating tangles in terms of the above notation. All of these actions may be readily verified directly from the operad definition. The multiplication tangle: B A This tangle corresponds to operator multiplication. The output is AB. The left embedding tangle l(A): A For A ∈ V ± where x ∈ H ± n , we have l(A) ∈ V ∓ n and v ∈ H ∓ n+1 defined by l(A)(c(v, x)) = c(v, A(x)), 1 . This operation is bounded in norm by emax. 5 The right embedding tangle r(A) : A For A ∈ V ± 1 or H − where x ∈ H ± is norm bounded by emax as well. n , we have r(A) ∈ V ± n and v ∈ H + n+1 defined by r(A)(c(x, v)) = c(A(x), v), 1 depending on the value of n. This operation Both the left and right embedding operators are strongly continuous; this follows directly from the definition. Temperley-Lieb generators (T L+ and T L−): Using the above operad definition, the T L+ element acts on the path basis and as follows: Let a and b be paths of length 1 on Γ which start at positive vertices, c and d paths of length 1 which start at negative vertices, with xa, xb, xc, xd the corresponding basis vectors in V + 1 . Then we should have 1 hT L(c(xa, xc)), c(xb, xd)i = 0 unless a is the reverse of c and b is the reverse of d, with additionally s(a) = s(b). If these conditions hold, then the inner product should be µ(t(a))µ(t(c))/µ(s(a))2. In other words, for each vertex v ∈ P + and V − 0 , let yv = Xes(e)=v µ(t(e))c(xe, rev(xe)) where the sum is taken over all paths of length 1 on Γ. Then T L(yv) = Xe1,e2s(e1,2)=v = Pes(e)=v µ(t(e))2 µ(v)2 !  Xes(e)=v = Pes(e)=v µ(t(e))2 µ(v)2 yv µ(te1 )µ(te2 )2/µ(v)2c(xe2 , rev(xe2 )) µ(t(e))c(xe, rev(xe))  For each positive vertex v, yv is an eigenvector for T L with eigenvalue δv = Pes(e)=v µ(t(e))2/µ(v)2, and T L is zero off the closed linear span of these yv's. We also have a T L− element. This is the same diagram as above but with reversed shading, and is defined by reversing all signs in the above definition. yv is defined as above, but now for v being a negative vertex. Then T L− has zero off the closed linear span of these yv's. each such yv as an eigenvector with eigenvalue Pes(e)=v µ(t(e))2/µ(v)2, and is We can now compute left and right capping operators, respectively LC(A) : n → V ∓ V ± n−1 and and RC(A) : V ± n → V ± n−1. 6 A A These operators may be described in terms of embedding and the TL gen- erators: rn−1(T L)l(A)rn−1(T L) = rn−2(T L)l2(LC(A)) ln−1(T L)r(A)ln−1(T L) = ln−2(T L)r2(RC(A)) This uniquely defines the capping operation: from the definitio of left and right embedding rn−2(T L)l2(LC(A)) and ln−2(T L)l2(RC(A)) are each zero only if LC(A), RC(A) respectively are zero. Note that a graded bounded linear map on the V ± n 's which fixes the Temperley-Lieb algebra and commutes with multiplication and embedding also commutes with capping; for such a map ω we have ln−2(T L)r2(RC(ω(A))) = ω(ln−2(T l)r2(RC(A))) = ln−2(T L)r2(ω(RC(A))) by the above definition, and the same holds for LC. 0 -valued sesquilinear form on V ± We now describe the interaction of the involution on V ± n (defined above) with these tangles. Since conjugation by the involution corresponds to tangle reflection (see [13]), it follows that (AB)∗ = B∗A∗, T L± are self-adjoint, and the involution commutes with left and right embedding. This is consistant with the above description of this involution as the adjoint operation on each B(H ± n ). One important property of the BGPA in [14] was the existence of a positive definite V ± n , namely hx, yi = RCn(y∗x) in the above notation. We would like this form to be positive definite here as well. Let p, q, r, s be path basis elements in V ± n , and A, B rank one partial isometries from (respectively) p to q and r to s. Then it follows directly from the operad definition that hA, Bi is a positive scalar multiple of a rank one projection in V + 0 if p = q and r = s, and is zero otherwise. It follows that the form is positive definite on bounded formal sums of such elements, which constitute all of V ± n . In order for a BGPA to be useful in a subfactor context, we should have both shaded and unshaded circles being equal to some scalar δ. This condition on a planar algebra is called modulus δ [13]. By capping off the single vertical strand l(1), we see this is true when T L2 = δT L with both shadings. This occurs when is independent of v, or δv = Xes(e)=v µ(t(e))2/µ(v)2 µ(t(e))2 = δµ(v)2 Xes(e)=v 7 for all v. Another way of describing this situation is that the vector with com- ponents µ(v)2 is an eigenvector for the connection matrix of the graph Γ, with eigenvalue δ. Under these circumstances we will say that the spin vector itself (or the BGPA) has modulus δ. A modulus δ spin vector is necessarily locally bounded and locally finite, √δ and emax ≤ δ2. Note that on a finite graph, there is only one with M ≤ modulus δ spin vector (up to normalization). Its entries are the square roots of the Perron-Frobenius eigenvector entries for the inclusion matrix, and δ is the Perron-Frobenius eigenvalue. On an infinite locally finite graph, there may be many modulus δ spin vectors, with different values of δ. 3. Automorphisms of BGPAs An automorphism of a planar algebra is a graded linear map on the V ± n 's which commutes with the entire planar operad. If the planar algebra has an involution, we will require the map to commute with the involution as well. In this section we describe the automorphism group of an arbitrary BGPA. These automorphisms of planar algebras may be viewed as a generalization of the automorphisms of the standard invariant of a subfactor (see [18]; c.f. [23],[12] for specific examples). In the BGPA case, these automorphisms are similar to those of a Jones tower of finite-dimensional C∗-algebras, computed in [10]. Lemma 3.1. Let Γ be a bipartite graph, with path Hilbert space H as in section 2. Let U be a unitary operator on H which respects the grading and commutes with the concatenation operator c. Then the action of U on H ± 0 is described by a graph automorphism. Moreover, AdU leaves the BGPA PΓ invariant. Proof. Let x = xv be a standard basis element of H + 0 , corresponding to a vertex v. We have c(x, x) = x, so c(U (x), U (x)) = U (x) as well by the proper- ties of U . The only elements of H + 0 which have this property are of the form Pw∈S xw, where S is some subset of the even vertices of Γ. Unitarity of U implies that in fact U (xv) = xw for some w. So U acts by permutation on the even vertices of Γ, and likewise on the odd vertices by an identical argument. Let this permutation be σ. As in section 2, let sv be the projection onto the basis elements corresponding to paths starting at v, and tv the projection onto paths terminating at v. The dimension of tvsw(H + 1 ) is the number of edges between v and w, or zero if they are not adjacent. Let n(v, w) be this number of edges. For any path p from v to w (with corresponding basis element xp), we have c(xv, xp, xw) = xp, implying that c(xσ(v), U (xp), xσ(w)) = U (xp). In other words, U tvsw(H + 1 ). Moreover U ∗ also commutes with con- catenation and respects grading, so applying the above argument to U ∗ gives us equality of the above subspaces. This means that n(σ(v), σ(w)) = n(v, w), and σ is a graph automorphism. 1 ) ⊂ tσ(v)sσ(w)(H + 8 The above also implies that AdU leaves the algebra generated by the s's and t's invariant, since U svtwU ∗ = sσ(v)tσ(w). Therefore AdU leaves the commutant of this algebra invariant as well. This commutant is precisely V ± n , implying that AdU acts on the planar algebra as desired. We recall that the path reversal operator rev defined in section 2 is an involution sending H + 1 to H − 1 . From the definition, this operator obeys hrev(x), rev(y)i = hy, xi 1 or H − 1 . where h·,·i is the inner product on H + Theorem 3.1. Let Γ be a locally finite bipartite graph with locally bounded spin vector. Let H be the path Hilbert space on Γ as above. Let U be a unitary element of B(H) which agrees with the grading of H and commutes with the concatena- tion operator. Assume further that the restriction of U to H ± 1 commutes with the path reversal operator, and that the vertex permutation induced by U scales the measure on H0. Then AdU commutes with the entire planar operad (and adjoint) on V ± n . Proof. We know from lemma 3.1 that AdU acts on the planar algebra. n , v ∈ H ± Recall l is the left embedding operator. Let x be in V ± n . Then since U commutes with concatentation, 1 and w ∈ H ∓ U ∗l(x)U c(v, w) = U ∗l(x)c(U (v), U (w)) = U ∗c(U (v), xU (w)) = c(v, U ∗xU (w)) l(U ∗xU )c(v, w) = c(v, U ∗xU (w)) while which is the same. So AdU commutes with left embedding, and with right embedding as well by a similar argument. Since U is unitary, AdU commutes with operator multiplication and taking adjoint. It remains only to show that AdU commutes with the Temperley- Lieb diagrams, ie with the elements of the operad representing tangles with no internal disks. This algebra is generated by the generators T L+ and T L− along with multiplication and embedding, so we need to show that AdU fixes these generators, i.e. that U commutes with them. Let v and w be two adjacent vertices of Γ. Let H vw 1 be the subspace of H1 the subspace of H2 spanned by paths 1 . Take spannned by paths from v to w, and H vw from v to w and back to v. Let {ai} be any orthonormal basis for H vw yvw = Pi c(ai, rev(ai)), using the reversal operator defined above. Because of the interaction of rev with inner product described above, the inner product of yvw with any element of the form c(v, rev(v)) is equal to the squared norm of v regardless of which basis is chosen. Such elements span H vw 2 , so yvw is independent of the choice of specific basis. This is true if v is odd or even. 2 9 Let σ be the permutation action of U on vertices of Γ. As in lemma 3.1, we , and so U maps an orthonormal basis for the first have U (H vw vector space to one for the second. This means that U (yvw) = yσ(v)σ(w). 1 ) = H σ(v)σ(w) 1 We recall from section 2 that the T L generators in B(H ± 2 ) each leave the closed linear span of certain vectors {xv} invariant, where v is taken from the set of positive or negative vertices depending on the sign of the generator, and are zero off the span of these vectors. We defined yv = Xes(e)=v µ(t(e))c(xe, rev(xe)) for each vertex v but we can also write yv = Xwn(v,w)6=0 µ(w)yvw using the notation above. From lemma 3.1 σ is a graph automorphism of Γ. By assumption, σ scales the spin vector by a fixed constant λ. Therefore U (µ(w)yvw) = λ−1(µ(σ(w))yσ(v)σ(w) Since σ is a graph automorphism, it maps the set of vertices adjacent to v to the set of vertices adjacent to σ(v), and we have as well λ−1µ(w)yσ(v)w U  Xwn(w,v)6=0 µ(w)yvw  = Xwn(w,σ(v))6=0 So U maps one standard basis vector of the subspace acted on by either T L generator to λ−1 times another such vector. properties of σ, We have T L(yv) = δvyv, where δv = Pes(e)=v µ(t(e))2/µ(v)2. From the δσ(v) = Xes(e)=σ(v) = Xes(e)=v µ(t(e))2/µ(σ(v))2 = Xes(e)=v λ2µ(t(e))2/λ2µ(σ(v))2 = Xes(e)=v µ(σ(t(e)))2/µ(σ(v))2 µ(t(e))2/µ(v)2 = δv So U leaves each eigenspace of both T L generators invariant. This means that U commutes with the T L generators. So U commutes with a set of generating tangles for the planar operad, and hence with the entire operad. Now we describe two classes of linear maps on a BGPA which meet all the above conditions. 10 Definition 3.1. Let Γ be a bipartite graph with spin vector µ. Let κ be a permutation of the vertices of Γ, preserving connection numbers and parity and preserving or scaling spin. Label each n-fold multiple edge by {1, ..., n}. We may then extend κ to the edges of Γ by asserting that it preserves this numbering. Then κ gives rise to a permutation of the paths on Γ, and therefore a map U on the path Hilbert space H. Then AdU is the graph automorphism operator associated with κ. Lemma 3.2. Graph automorphism operators are automorphisms of the BGPA. Proof. Let AdU be a graph automorphism operator as above. It follows di- rectly from the definition that AdU commutes with path reversal and concate- nation, and agrees with the grading. We have also assumed that the underlying graph automorphism is trace scaling. So all of the conditions of theorem 3.1 are satisfied, and AdU is a planar algebra automorphism. Definition 3.2. Let O be an element of V + acting on H + path basis element of H + 1 . Let O′ = rev ◦ O ◦ rev, acting on H − 1 (for i even) or H − 1 which is unitary as an operator 1 . For 1 ≤ i ≤ n, let pi be a 1 (for i odd). Then let U (c(p1, p2, ..., pn)) = c(O(p1), O′(p2), O(p3)...) and U extends to a unique bounded linear operator on H + on H − n as the extension of the map n . Define U similarly U (c(p1, p2, ..., pn)) = c(O′(p1), O(p2), O′(p3)...) where the pi's are again in H + 1 or H − 1 as appropriate. Then AdU is the multiplication operator associated with O. If O acts nontrivially on only one of the subspaces svtw(H + 1 ), while leaving all others fixed, then we will call it a basic multiplication operator associated to the vertex pair {v, w}. Every multiplication operator is a product of basic mul- tiplication operators, and basic multiplication operators associated to different vertex pairs commute with each other. A multiplication operator is scalar if the restriction to B(svtw(H + 1 )) for each v, w is a scalar multiple of the identity. The scalar multiplication operators are the center of the multiplication operator group. If the graph has no multiple edges then every multiplication operator is scalar. Lemma 3.3. Multiplication operators are automorphisms of the BGPA. Proof. Let AdU be a multiplication operator as above. From the definition, it commutes with concatenation and path reversal and respects grading. Since O comes from an element of V + 1 , it commutes with sv and tv for vertices v. Therefore the associated graph automorphism σ is trivial, and preserves the trace. So all the conditions of theorem 3.1 are satisfied, and AdU is a planar algebra isomorphism. 11 Direct computation shows that conjugating a basic multiplication operator by a graph automorphism operator produces a basic multiplication operator associated to different vertex pair. So the group of automorphisms generated by these two types of operators has a crossed product structure: the subgroup of multiplication operators is normal. Now we will show that the two types of operators described above in fact generate the entire automorphism group of a BGPA. Lemma 3.4. Any automorphism of a BGPA is strongly continuous. Proof. Let α be an automorphism of a BGPA. n ⊂ B(H ± n ), where V ± From the definition in section 3.2, we have V ± n is a (potentially infinite) direct sum of type I factors. Every automorphism of such a von Neumann algebra can be written AdU , where U is a unitary element of B(H ± commutes with multiplication and involution, this restric- tion of α is a von Neumann algebra isomorphism, and may be written AdU ± n for unitary U ∈ B(H ± n s provides a unitary operator on the total Hilbert space H whose adjoint action on the graded vector space V ± n agrees with α. n ). Summing all the U ± n ). Since αV ± n Multiplication is strongly continuous. This means that two automorphisms of a BGPA are equal if and only if they agree on loops, since the loops span a strongly dense set in each V ± n . Lemma 3.5. Let α be an automorphism of a BGPA. Then there is a graph automorphism operator β such that α and β agree on V ± 0 . 0 , V − Proof. Let pv, qw be the atomic projections in V + 0 associated with even and odd vertices v and w. Since α is a BGPA automorphism, it must send pv to some other atomic projection pσ(v) in V + 0 , and likewise for qw. Therefore α induces a permutation σ on the vertices of Γ. To see that this permutation is a graph automorphism, note that l(pv)r(qw) is a minimal central projection of V + 1 (or zero), and the dimension of l(pv)r(qw)V + 1 is the square of the number of edges between v and w. This dimension is pre- served by α, i.e. 1 . There- fore (v, w) and (σ(v), σ(w)) have the same number of edges between them, i.e. n(v, w) = n(σ(v), σ(w)). it agrees with the dimension of l(pσ(v))r(qσ(w))V + Next we must show that σ preserves or scales the trace. For this we note that pv with a circle around it evaluates to qw µ(v) µ(w) Xwn(v,w)6=0 Since α commutes with the tangle, we must have µ(v)/µ(w) = µ(σ(v))µ(σ(w)) for all adjacent v, w, implying the desired result. 12 From the above description of graph automorphism operators, there is a such an operator β whose induced permutation action on the vertices of Γ is the same as α's. These two operators then agree on V ± 0 . Lemma 3.6. Let α be an automorphism of a BGPA which acts trivially on V ± 0 . Then there is a multiplication operator β such that β and α agree on V ± 1 . Proof. Since α acts trivially on V ± 0 , it fixes all elements of the form r(pv)l(qw) in V + 1 . These elements are the center of V + taken as a von Neumann algebra, 1 so α acts as an inner automorphism on V + 1 . There is a multiplication operator β whose action on V + is any desired inner automorphism. Then α and β agree 1 on V + 1 . Both of these automorphisms commute with the half rotation A which is a bijective map from V + 1 to V − 1 , so they agree on V − 1 as well. Lemma 3.7. Let α be an automorphism of a BGPA which acts trivially on V ± 1 . Then α is a scalar multiplication operator. Proof. From the proof of lemma 3.4, we can write the action of α on each V ± n as AdU ± n is a unitary in B(H ± n , where U ± n ). First note that for any path p on the graph Γ, with corresponding basis n . This n , there is a rank one projection onto xp contained in V ± vector xp ∈ H ± projection may be written as the product rn(pe1 )rn−1l(pe2 )...rln−1(pen−1)ln(pen ) 1 or V − where ei is the ith edge of p and pei is the rank one projection in V + 1 onto the vector xei corresponding to the path ei. Since α acts trivially on V ± 1 , it fixes this projection, and U ± n must therefore map each xp to a scalar multiple of itself. Now let bl ∈ V ± n be a rank one partial isometry corresponding to a loop n . We have bl = xpblxq for certain rank one projections xp, xq as above, l ∈ l± implying that α(bl) = α(xpblxq) = xpα(bl)xq The only way this can be true is if α sends bl to a scalar multiple of itself as well. In other words, every loop is an eigenvector of α, and α induces a map ρ from the set of loops to the complex scalars of modulus 1. Since α commutes with the half rotation, ρ(l) is independent of the basepoint 1 , ρ(l) = 1 for any loop l of length 2. Finally, since α of l. Because α fixes V ± commutes with this diagram 13 bl1 bl2 if l3 is the concatenation of the loops l1 and l2 then we have ρ(l3) = ρ(l1)ρ(l2). Putting this together we find that ρ is necessarily a 1-dimensional repre- sentation of the fundamental group of Γ. But any such representation may be obtained from a scalar multiplication operator. We may always find a set of free generators {l1, l2, ...} for the fundamental group with the property that each generator li contains some edge ei which does not appear in any other loop. Then a basic scalar multiplication operator with value λ associated to the endpoints of some ek corresponds to the representation of π1(Γ) sending lk to λ (or possible λ, depending on the direction of lk) and all other generators to 1. All other representations of π1 may be obtained similarly from basic scalar multiplication operators associated to various ei's. This implies that there is a scalar multiplication operator which agrees with α on loops. Since the loops span a strongly dense subset of each V ± n , and automorphisms of a BGPA are strongly continuous by lemma 3.4, it follows that α itself is a scalar multiplication operator. Since scalar multiplication operators themselves act trivially on V ± 1 , this lemma in fact shows that the scalar multiplication operators of a BGPA are isomorphic to the 1-dimensional representations of the fundamental group of the graph. Theorem 3.2. Let PΓ be a BGPA, with multiplication operators E and graph automorphism operators A. Let α be an automorphism of PΓ. Then α = ae for some a ∈ A, e ∈ E. Proof. This follows from the proceeding lemmas. There is β1 ∈ A such that β−1 1 α acts trivially on V ± 1 α acts trivially on V ± 1 . So β−1 1 α is a scalar multiplication operator β3, and α = β1β2β3 with β1 ∈ A and β2β3 ∈ E. 0 . There is β2 ∈ E such that β−1 2 β−1 2 β−1 Since conjugation by graph automorphisms leaves the multiplication oper- ator group invariant, we may in fact write AutPΓ = E ⋊ A with notation as above. 4. Planar fixed point subfactors A subfactor planar algebra is a planar algebra with the following additional properties [13]: • dimV ± • dimV ± 0 = 1 n < ∞ ∀n 14 • Spherical: Since V + the scalars. A planar algebra is spherical if for any A ∈ V + right caps LC(A) and RC(A) agree. 0 = C, we may equate these vector spaces with 1 , the left and 0 = V − • Involution: There is an antilinear isometry on each V ± with tangles as reflection. n which interacts • Positive definiteness: Involution gives us a scalar sesquilinear form, namely hx, yi = RCn(y∗x) where the multiplication and right capping tangles are as in section 2. This form should be positive definite. described as a subfactor planar algebra ([13],[22]). We take V + V − n = M ′ The standard invariant of any finite index extremal II1 subfactor may be 0 ∩ Mn, Conversely, if P is a subfactor planar algebra, then there exists a finite index extremal II1 subfactor such that the standard invariant of this subfactor is P ([22], c.f. [13],[11], [16],[15]). 1 ∩ Mn+1, and the operad definition is given in [13]. n = M ′ For finite graphs with the correct spin vector, any sufficiently small planar subalgebra is of subfactor type. Lemma 4.1. Let PΓ be a bipartite graph planar algebra with spin vector µ. Let x be in V + 0 obtained by capping off to the left, and xr ∈ V + from capping off to the right. Suppose both xl and xr are scalars. Then there is some constant α, independent of x, such that xl = αxr. 1 ; let xl represent the element of V − 0 Proof. From [14], there is a partition function defined on V + 0 and V0− as the linear extension of xv → µ4(v), and this function has the same value on xl and xr. If xl and xr are scalars, this means xl = λ1Pv xv, xr = λ2Pv xv. The partition function on x1 is λ1Pv µ4(xv), and on xr is λ2Pv µ4(xv). Since these are the same we must have λ1 = λ2 α = Pv µ4(xv ) Pv µ4(xv ) and xl = αxr for all x such that xl and xr are scalars. Pv µ4(xv) Pv µ4(xv) . In other words 0 = C. Then X is a subfactor planar algebra. Theorem 4.1. Let PΓ be a finite bipartite graph planar algebra whose spin vector µ has modulus δ. Let X be a planar ∗-subalgebra of PΓ such that X∩V + 0 = X ∩ V − Proof. The BGPA has an involution, giving rise to a positive definite V + 0 - valued sesquilinear form (see section 2). The involution and form are inherited by X, and the restriction of the form to X is scalar valued since X ∩ V ± 0 = C. Since Γ is finite, each V ± n is finite dimensional, and so this is true of their intersections with X as well. To show that X is of subfactor type, it remains only to demonstrate spher- icality. Let x ∈ X be an element of V + 1 . Capping off to left or right produces scalars, since X ∩ V ± is scalar. Therefore the conditions of the lemma above are satisfied. Because µ has modulus δ, shaded and unshaded circles represent the same scalar. These diagrams are the left and right caps of a single vertical 0 15 strand, so the constant α in the lemma is equal to 1. It follows that xl and xr are equal as scalars, and X is spherical. A small subalgebra of an infinite BGPA still corresponds to a subfactor, but the subfactor need not be extremal. 0 = C. Then X ∩ V ± Theorem 4.2. Let PΓ be a locally finite bipartite graph planar algebra. Let X ⊂ PΓ be a spherical planar ∗-subalgebra with X ∩ V ± n is finite dimensional. Proof. Let pv ∈ V + 0 be a minimal projection corresponding to some even n , and let RCn be the diagram con- vertex v, and q = rn(pv). Take x ∈ X ∩ V + sisting of capping off all strands to the right (this is the form from section 2). Then RCn(x) is a scalar from the properties of X, and RCn(qx) = qRCn(x) by isotopy invariance. q commutes with x and x∗ with respect to the usual multipli- cation tangle. If qx = 0, then qx∗x = 0, giving RCn(qx∗x) = 0 = qRCn(x∗x). Since RCn(x∗x) is a scalar, this means that px = 0 implies RCn(x∗x) = 0. But RCn gives a positive definite form on V + n , so x itself is zero in this case. This means that the map x → qx is injective on X ∩ V + n has basis labelled by loops of length 2n which start and end at v. By local finiteness of Γ, this set is finite. Therefore pV + n is as well. n is finite dimensional, and X ∩ V + n . But qV + n is finite dimensional. The same argument shows that X ∩ V − Burns described a class of rigid planar C∗-algebras in [9], generalizing Jones' definition of a SPA. Every rigid planar C∗-algebra is the standard invariant of a finite index II1 subfactor, but this subfactor need not be extremal. From Burns' definition, a planar algebra having all the characteristics of a SPA except sphericality is a rigid planar C∗-algebra. Let PΓ be a modulus δ BGPA coming from an infinite finite graph, and X ⊂ PΓ a planar subalgebra with dim(X ∩ V ± 0 ) = 1. The above lemma tells us that X∩V ± n is finite dimensional. The BGPA has a positive definite sesquilinear form and an involution with the right properties, which are inherited by X. This means that X is a rigid planar C∗-algebra. If X is additionally spherical, then it is an SPA. This is automatic when X is irreducible, but is not true in general. This means that there is a finite index II1 subfactor whose standard invariant is X. The subfactor will be extremal if and only if X is spherical. 5. Examples 5.1. Introduction In order for the fixed points P G We can construct a wide range of subfactor planar algebras under this fixed Γ to be an SPA, we need it point technique. to have 1-dimensional intersection with V + 0 , which is equivalent to having the graph automorphism part of G act transitively on both positive and negative vertices. We should also check sphericality for non-irreducible infinite graph examples. Some SPAS thus are the standard invariants of previously known subfactors, while others seem to be previously unclassified. A few such examples are described below. 0 and V − 16 Definition 5.1. If PΓ is a BGPA, and G ⊂ AutPΓ with P G corresponding subfactor is a planar fixed point subfactor. Γ an SPA, then the 5.2. Group-subgroup subfactors A specific example of the planar fixed point construction is found in [12]. In this paper, Gupta starts with the BGPA on the graph with n odd vertices all connected to one even vertex. Then G is some group acting by permutation on the set of odd vertices, and H is the subgroup which fixes some specified vertex. Gupta shows that the fixed points of the BGPA by this action constitute a subfactor planar algebra, and that this is in fact the standard invariant of the group-subgroup subfactor (see e.g. [17]) corresponding to the inclusion H ⊂ G, namely M G ⊂ M H for some outer action of some finite group G on a II1 factor M . 5.3. Wassermann subfactors Here we let Γ be the graph with two vertices connected by an n-fold multiple edge. Let G be any compact subgroup of the unitaries Mn(C). Then G may be embedded in the multiplication operators on this graph. The fixed points of this G-action are of subfactor type. They are identical to the standard invariant of the Wasserman subfactor (1 ⊗ Mn ⊗ Mn ⊗ ... st )G ⊂ (Mn ⊗ Mn ⊗ Mn ⊗ ... st )G where G acts pointwise on the tensor products (see [24]). 5.4. Diagonal subfactors Let G be a finitely generated group of outer automorphisms of a II1 factor M . Let G have distinct generators {g1, ..., gn}, and take Γ to be the graph which has one odd and one even vertex for each element of G. Two vertices v+ x and v− y are connected by a single edge if y = xh, for h ∈ {1, g1, ..., gn}. The spin vector of this graph is 1 at every vertex; it has modulus n + 1. g ) = vxg. Associating each such graph automorphism with the corresponding graph automorphism operator on PΓ gives an action of G on PΓ. Then G acts on the graph by left translation: αx(v± sphericality may be directly verified: V + projection has left and right trace 1/(n + 1). Therefore P G This action is transitive on even and odd vertices. When the graph is infinite, 1 has dimension n+ 1, and each minimal Γ is of subfactor type. Since vertices are labelled by group elements, loops of length 2n in this graph may be written as a list of vertices x − xa1 − xa1a−1 2 xa1a−1 2 a3... − xa1a−1 2 ...a−1 2n = x where each ai comes from the set {1, g1, ..., gn}. So we may think of this loop as a starting point x along with a list of generators (a1, a2, ...) whose alternating product is the identity in G. 17 The group action moves the base point (via left translation) while keeping the generator list invariant. It follows that these generator lists label a basis for the intersection of P G Γ with V ± n . This basis is precisely that described in [3] for the planar algebra of the diagonal subfactor (see [21],[2]). It may be verified that the operations of left and right embedding, involution, and multiplication on P G Γ agree with the planar algebra of [3], and the Temperley-Lieb algebra embeds in the same way, so these planar algebras are isomorphic. It follows that the planar algebra constructed in this way is that of a diagonal subfactor without cocycle. 5.5. Bisch-Haagerup subfactors Let G be a group of outer automorphisms of a II1 factor M , generated by finite subgroups H and K. For simplicity we require H ∩ K = {1}. Let Γ be the graph which has one even vertex for each H right coset in G, and one odd vertex for each K right coset. Then the edges of Γ are labelled by group elements. The endpoints of each edge eg are the even vertex vgH and the odd vertex vgK . The spin vector has value H on each even vertex and value K We may write a loop of length 2n as a list of edges: on each odd vertex, and it has modulus pHK. ex1 − ex2... − ex2n − e1 Assume first that the loop begins at a positive vertex. For the path to be connected, ex1 and ex2 must share a vertex, so x1 and x2 are in the same K- coset and x2 = x1k1. Likewise x2 and x3 are in the same H-coset. So we can view this loop as a starting point x along with a list of alternating elements of k and h, with the restriction that k1h1k2h2...knhn is equal to the identity in G. Again G acts on the graph by left translation, and this gives a G-action on PΓ. This action shifts the basepoint of loops while leaving the list of ki's and hi's the same. So we may identify a basis for P G n , namely n-tuples of elements of K and H obeying k1h1k2h2...knhn = 1; the basis for V − n is obtained by reversing the roles of H and K. This labelling set is the same as that of the planar algebra of [4] (c.f. [5]); again it may be shown that the two planar algebras are isomorphic. It follows that this fixed point planar algebra is the standard invariant of the Bisch-Haagerup subfactor M H ⊂ M ⋊ K. 5.6. The cube graph Γ ∩ V + Let Γ be the graph of a cube, and PΓ the corresponding BGPA. We obtain several subfactor planar algebras by taking fixed points under various group actions. We note that the automorphism group of the bipartite graph is S4, since each such automorphism may be described uniquely as a certain permutation of the even vertices. The only modulus δ spin vector assigns weight 3 to every vertex; δ = 3, so every subfactor planar subalgebra produces an index 9 subfactor. We can write down a biprojection which is invariant under every automorphism, so there is always an index 3 intermediate subfactor. 18 generated by (12)(34) and (13)(24), then P G subfactor; the group generators are each order 2, and the group is Z 2 A4, then P G order-3 automorphisms of M together generate the group A4. We mention some of the possibilities described above. If G is the subgroup Γ is the planar algebra of the diagonal 2 . If G is Γ is the Bisch-Haagerup subfactor M Z3 ⊂ M ⋊ Z3, where the two Its Γ is some other subfactor planar algebra. If we take G = S4, then P G principal graph may be directly computed from the group action: ∗ The dual principal graph is the same. It does not appear to be of any previously categorized type, although we have tentatively identified it as a com- position of two group-subgroup subfactors. Finally, we may take G = AutPΓ. Since the multiplication operator group of PΓ is (S1)5, this group is infinite, and P G Γ is infinite depth. 5.7. The degree (3,2) tree graph Let Γ be the graph which branches twice at each even vertex and three times at each odd vertex. The operator norm of this graph is 4√2. There are no multiplication operators on PΓ, so all automorphisms will come from graph automorphisms. First assign weight 2 to every even vertex and 3 to every odd vertex. This spin vector has modulus √6. Every automorphism, of Γ here gives rise to an automorphism of PΓ. Taking G = Aut(PΓ), we obtain a subfactor planar algebra P G Γ . This subfactor is irreducible, hence automatically spherical, and is non-amenable; we conjecture it is obtained as a composition of two group- subgroup subfactors, where the groups involved are variations of the Grigorchuk lamplighter group. We may also obtain a transitive subgroup as follows: color the edges of the graph with three colors so that no vertex contacts two edges of the same color, and then consider all automorphisms which leave the coloring invariant or permute the colors. This group is Z3 ∗ Z2, and the resulting subfactor is the index 6 Bisch-Haagerup subfactor described in [5] corresponding to this group. Now we describe a non-irreducible example. We color the edges of the graph red and blue so that each even vertex contacts a red and blue edge, and each odd vertex contacts two reds and a blue. We now consider G to be the group of color preserving automorphisms. It may be seen that this group is transitive on odd and even vertices, but the fixed points have 2-dimensional intersection with V + space, corresponding to the 2 G-orbits of edges. With the above spin 1 vector, sphericality fails: the 'blue edge' element has left and right traces of 1/3 and 1/2. To find a subfactor planar subalgebra of this BGPA, we will need a different spin vector. We take µ(v0) = 1 for some arbitary base vertex v0, and then define 19 b b b b b c b c b c b c µ elsewhere so that for adjacent vertices x even and y odd, we have µ(y) = µ(x) if the x − y edge is red, but µ(y) = 21/4µ(x) if the edge is blue. This trace has modulus 1 + √2. It may be seen that any color preserving automorphism of Γ will multiply µ by a constant factor, hence providing an automorphism of PΓ. With this choice of spin vector and G the color preserving graph automorphisms, the left and right traces of the 'blue' element are both √2 − 1, and sphericality holds. So from section 5.4, P G √2)2 = 3 + 2√2, and it does not have any intermediate subfactors. Γ is a subfactor planar algebra. The resulting subfactor is infinite depth non-amenable. Its index is (1 + From [20], this is the minimum index for an extremal non-irreducible sub- factor. The above construction is a new way of getting such a subfactor. Any group which is transitive on even and odd vertices but the same two orbits of edges as above will also produce a subfactor with this list of properties, so we actually have many such examples. For example, we might partition the red edges into 'red' and 'white' so that each odd vertex contacts one vertex of each color; G′ might then be the group of graph automorphisms which either preserve color or swap the colors red and white. There are also spin vectors of any modulus greater than 3 + 2√2 on this graph which are preserved or scaled by every element of G. Taking fixed points then gives us a non-extremal subfactor, as in [9]. So we obtain a continuous family of non-extremal subfactors with the same principal graphs, but different indices. Other non-irreducible subfactors with index (1 + √3)2 = 4 + 2√3 may be constructed similarly from a graph which branches twice at each even vertex and four times at each odd vertex. This is the third on the list from [20]. Acknowledgement. I would like to thank Professor Dietmar Bisch for his useful suggestions and corrections. References [1] Stephen Bigelow, Scott Morrison, Emily Peters and Noah Snyder. Con- structing the extended Haagerup planar algebra. arxiv.org, arXiv 0909.4099 [math.OA]:45 pp., 2009. [2] Dietmar Bisch. Entropy of groups and subfactors J. Funct. Anal. 103 (1992), no. 1, 190 -- 208. [3] Dietmar Bisch, Paramita Das and Shamindra K. Ghosh. The planar algebra of diagonal subfactors. Clay Mathematics Proceedings, Vol. 10 (2008), 25 pp. [4] Dietmar Bisch, Paramita Das and Shamindra K. Ghosh. The planar algebra of group-type subfactors. J. Funct. Anal. 257 (2009), no. 1, 20 -- 46. [5] Dietmar Bisch and Uffe Haagerup. Composition of subfactors: new exam- ples of infinite depth subfactors. Ann. Sci. ´Ecole Norm. Sup. (4), 29(3):329 -- 383, 1996. 20 [6] Dietmar Bisch and Vaughan F. R Jones. Algebras associated to interme- diate subfactors Invent. Math. 128 (1997), no. 1, 89 -- 157. [7] Dietmar Bisch and Vaughan F. R. Jones. Singly generated planar algebras of small dimension. Duke Math. J. 101 (2000), no. 1, 41 -- 75. [8] Dietmar Bisch and Vaughan F. R Jones. Singly generated planar algebras of small dimension, II. Adv. Math. 175 (2003), no. 2, 297 -- 318. [9] Michael Burns. Subfactors, planar algebras and rotations. PhD dissertation, University of California, Berkeley, Department of Mathematics, 2003. [10] Richard D. Burstein. Group-type subfactors and Hadamard matrices. To appear, Trans. Amer. Math. Soc. [11] Alice Guionnet, Vaughan F. R. Jones and Dimitri Shlyakhtenko. Ran- dom matrices, free probability, planar algebras and subfactors. arXiv.org, arXiv:0712.2904:31 pp., 2008. [12] Ved P. Gupta. Planar algebra of the subgroup-subfactor. Proc. Indian Acad. Sci. Math. Sci. 118 (2008), no. 4, 583 -- 612. [13] Vaughan F. R. Jones. Planar algebras, I. arXiv.org, arXiv:9909027 [math.OA]:122 pp, 1999. To appear, New Zealand J. Math. [14] Vaughan F. R. Jones. The planar algebra of a bipartite graph. in Knots in Hellas '98 (Delphi), 94 -- 117, World Sci. Publ., River Edge, NJ. [15] Vaughan F. R. Jones, Dimitri Shlyakhtenko, Kevin Walker. An orthogonal approach to the subfactor of a planar algebra. arXiv.org, arXiv:0806.4146 [math.OA]: 12pp., 2008. [16] Vijay Kodiyalam and V. S. Sunder. From subfactor planar algebras to subfactors Internat. J. Math. 20 (2009), no. 10, 1207 -- 1231. [17] Hideki Kosaki and Shigeru Yamagami. Irreducible bimodules associated with crossed product algebras. Internat. J. Math. 3 (1992), no. 5, 661 -- 676. MR1189679 [18] Phan H. Loi. On automorphisms of subfactors. J. Funct. Anal. 141 (1996), no. 2, 275 -- 293. [19] Emily Peters. A planar algebra construction of the Haagerup subfactor. arXiv.org, arXiv:0808.0764 [math.OA]: 57pp., 2008. To appear, Int. J. of Math. [20] Mihai Pimsner and Sorin Popa. Entropy and index for subfactors. Ann. Sci. ´Ecole Norm. Sup. (4) 19 (1986), no. 1, 57 -- 106. [21] Sorin Popa. Sousfacteurs, actions des groupes et cohomologie. C. R. Acad. Sci. Paris S´er. I Math. 309 (1989), no. 12, 771 -- 776. 21 [22] Sorin Popa. An axiomatization of the lattice of higher relative commutants of a subfactor, Invent. Math. 120 (1995), no. 3, 427 -- 445. [23] Anna L. Svendsen. Automorphisms of subfactors from commuting squares. Trans. Amer. Math. Soc. 356 (2004), no. 6, 2515 -- 2543 (electronic). [24] Antony Wassermann. Coactions and Yang-Baxter equations for ergodic actions and subfactors. in Operator algebras and applications, Vol. 2, 203 -- 236, Cambridge Univ. Press, Cambridge. 22
0906.1401	2	0906	2010-02-22T01:04:17	Nonseparable UHF algebras I: Dixmier's problem	[ "math.OA", "math.LO" ]	There are three natural ways to define UHF (uniformly hyperfinite) C*-algebras, and all three definitions are equivalent for separable algebras. In 1967 Dixmier asked whether the three definitions remain equivalent for not necessarily separable algebras. We give a complete answer to this question. More precisely, we show that in small cardinality two definitions remain equivalent, and give counterexamples in other cases. Our results do not use any additional set-theoretic axioms beyond the usual axioms, namely ZFC.	math.OA	math	NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM ILIJAS FARAH AND TAKESHI KATSURA Abstract. There are three natural ways to define UHF (uniformly hy- perfinite) C-algebras, and all three definitions are equivalent for separa- ble algebras. In 1967 Dixmier asked whether the three definitions remain equivalent for not necessarily separable algebras. We give a complete answer to this question. More precisely, we show that in small cardi- nality two definitions remain equivalent, and give counterexamples in other cases. Our results do not use any additional set-theoretic axioms beyond the usual axioms, namely ZFC. 1. Introduction Let A be a C-algebra and let ε be a positive number. For an element x of A and a subset F of A, we write x ∈ε F if there exists y ∈ F such that kx − yk < ε. For two subsets F,G of A, we write F ⊆ε G if x ∈ε G for all x ∈ F. For each n ∈ N, we denote by Mn(C) the unital C-algebra of all n × n matrices with complex entries. A C-algebra which is isomorphic to Mn(C) for some n ∈ N is called a full matrix algebra. Definition 1.1. A C-algebra A is said to be of full matrix algebras. matrix subalgebras with dense union. • uniformly hyperfinite (or UHF ) if A is isomorphic to a tensor product • approximately matricial (or AM ) if it has a directed family of full • locally matricial (or LM ) if for any finite subset F of A and any ε > 0, there exists a full matrix subalgebra M of A with F ⊆ε M , For a definition of tensor products, see Definition 2.16. The property LM was called matroid in [6, Definition 1.1]. A UHF algebra is unital by definition, and it is easy to see that UHF implies AM and that AM implies LM. In [11, Theorem 1.13], Glimm shows that a unital separable LM algebra is UHF (see also [6, Remark 1.3 and Theorem 1.6]). Thus for separable C-algebras, the three conditions UHF, unital AM and unital LM coincide. Dixmier asked whether these three conditions coincide for general C-algebras in [6, Problem 8.1]. We show that this is not the case. To state our results precisely, we need the following notion. Date: November 21, 2018. 1991 Mathematics Subject Classification. 46L05, 03E75. 1 2 ILIJAS FARAH AND TAKESHI KATSURA Definition 1.2. The character density χ(A) of a C-algebra A is the small- est cardinality of a dense subset of A. Hence A is separable if and only if its character density χ(A) is the first infinite cardinal ℵ0. Note that χ(A) is equal to the smallest cardinality of an infinite generating subset of A. The following are our main results which completely answer [6, Prob- lem 8.1]. Note that ℵ1 is the smallest uncountable cardinal. Theorem 1.3. AM and LM are equivalent. (1) For a C-algebra with character density at most ℵ1, (2) For every cardinal κ > ℵ1, there exists a unital LM algebra with (3) For every cardinal κ ≥ ℵ1, there exists a unital AM algebra with character density κ which is not AM. character density κ which is not UHF. Proof. (1) Follows from Proposition 5.2 and Proposition 5.6. (2) Follows from Proposition 6.10 and Proposition 6.12. (3) Follows from Proposition 4.5. (cid:3) In (3), we can also control the representation density (defined in Defi- nition 7.1) of the example (Theorem 7.17). In particular, we distinguish between AM algebras and UHF algebras faithfully represented on a separa- ble Hilbert space. Results similar to (1) and (2) hold for approximately finite-dimensional (AF) algebras. Definition 1.4. A C-algebra A is said to be of finite-dimensional subalgebras with dense union. • approximately finite-dimensional (or AF ) if it has a directed family • locally finite-dimensional (or LF ) if for any finite subset F of A and any ε > 0, there exists a finite-dimensional subalgebra D of A with F ⊆ε D. It is easy to see that AF implies LF. In [3, Theorem 2.2] Bratteli proved that for a separable C-algebra, AF and LF are equivalent. We get the following. Theorem 1.5. AF and LF are equivalent. (1)' For a C-algebra with character density at most ℵ1, (2)' For every cardinal κ > ℵ1, there exists an LF algebra with character density κ which is not AF. Proof. (1)' Follows from Proposition 5.6. (2)' Follows from Proposition 6.10 and Proposition 6.12. (cid:3) A C-algebra is AM (resp. AF) if and only if it is obtained as a direct limit of full matrix algebras (resp. finite-dimensional algebras) over a general directed set (not necessarily a sequence). On the other hand, it is not hard NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 3 to see that a C-algebra is LM (resp. LF) if and only if it is obtained as a direct limit of (separable) AM (resp. AF) algebras (Lemma 2.13). Hence the two theorems above imply the following. Corollary 1.6. The classes of AM algebras and AF algebras are not closed under taking direct limits. Some of the results of the present paper were announced in [14]. By extending our methods the first author constructed an AM algebra that has faithful irreducible representations both on a separable Hilbert space and on a nonseparable Hilbert space ([8]). In the sequel to this paper [10] we show that the classification problems for UHF and AM algebras are significantly different. Organization of the paper. In §2 we set up the toolbox used in the paper. In §3 we use the Jiang -- Su algebra to distinguish LM algebras from UHF algebras. σ-complete directed systems are used in §4 to distinguish between AM and UHF algebras. The relation between AM and LM algebras as well as the one between AF and LF algebras are explained in §5 and §6. In §7 we introduce the representation density, and using it distinguish between AM algebras and UHF algebras faithfully represented on a given Hilbert space. 2. Preliminary In the present section we fix the terminology and prove some standard facts from set theory, σ-complete directed systems and tensor products (re- spectively). 2.1. Set theory. By X ∐ Y we denote the disjoint union of sets X and Y . If f : X → Y and Z ⊆ X then we write f [Z] = {f (z) : z ∈ Z} instead of the notation f (Z) commonly accepted outside of set theory. Let us denote the cardinality of a set X by X. The countable infinite cardinal and the smallest uncountable cardinal are denoted by ℵ0 and ℵ1, respectively. The smallest uncountable ordinal is denoted by ω1. Lemma 2.1. Let X be a set. For each x ∈ X, choose a countable subset Yx ⊆ X with x ∈ Yx. If X > ℵ1 then one can find two elements x, y ∈ X such that x /∈ Yy and y /∈ Yx. Proof. Take Z ⊆ X with Z = ℵ1. Choose x ∈ X \Sz∈Z Yz and y ∈ Z \ Yx. Then x and y are as required. (cid:3) Remark 2.2. The conclusion may be false if X ≤ ℵ1. To see this consider X = ω1 and Yx = {y ∈ ω1 : y ≤ x} for x ∈ ω1. Definition 2.3. A directed set Λ is said to be σ-complete if every countable directed Z ⊆ Λ has the supremum sup Z ∈ Λ. The ordered set ω1 is σ-complete. The following is another σ-complete directed set considered in this paper. 4 ILIJAS FARAH AND TAKESHI KATSURA Definition 2.4. For an infinite set X, we denote by [X]ℵ0 the set of all countable infinite subsets of X, considered as a directed set with respect to the inclusion. Definition 2.5. Let Λ be a σ-complete directed set. A subset Λ0 of Λ is said to be closed if for every countable directed Z ⊆ Λ0 we have sup Z ∈ Λ0, and cofinal if for every λ ∈ Λ there exists λ0 ∈ Λ0 such that λ (cid:22) λ0. A closed and cofinal subset is called a club. A club is an abbreviation of a closed and unbounded set. The condition 'unbounded' (meaning 'not having an upper bound') is equivalent to 'cofinal' for totally ordered sets such as ω1, but is strictly weaker than 'cofinal' for general directed sets. A widely accepted custom among set theorists is calling closed and cofinal subsets of [X]ℵ0 closed and unbounded sets (or clubs). Reluctantly, we continue this unfortunate abuse of terminology in our paper. This can be justified by the fact that ω1 and [X]ℵ0 are the only σ-complete directed sets that we will consider from the next section on. 0 be clubs of 0 be an order isomorphism. Then there exists a club Λ00 Lemma 2.6. Let Λ be a σ-complete directed set. Let Λ0 and Λ′ Λ and φ : Λ0 → Λ′ of Λ such that Λ00 ⊆ Λ0 ∩ Λ′ Proof. Set Λ00 := {λ ∈ Λ0 ∩ Λ′ 0 : φ(λ) = λ}. It is easy to see that Λ00 is closed. We will see that it is cofinal. Take λ ∈ Λ. Since Λ0 is cofinal, there exists λ1 ∈ Λ0 with λ (cid:22) λ1. Since Λ′ 0 with 1 and φ(λ1) (cid:22) λ′ λ1 (cid:22) λ′ 0 for n = 1, 2, . . . such that 0 and φ ↾Λ00= id. 0 is cofinal, there exists λ′ 1. Recursively, we can find λn ∈ Λ0 and λ′ 1 ∈ Λ′ n ∈ Λ′ λn (cid:22) λ′ n, φ(λn) (cid:22) λ′ n, λ′ n (cid:22) λn+1, φ−1(λ′ n) (cid:22) λn+1. Then λ00 := sup{λn}∞ n=1 = sup{λ′ n}∞ n=1 ∈ Λ0 ∩ Λ′ 0 satisfies φ(λ00) = λ00. Thus we have found λ00 ∈ Λ00 with λ (cid:22) λ00. Lemma 2.7. Let X and Y be infinite sets. For a club C in [X ∐ Y ]ℵ0, there exists a club C0 in [X]ℵ0 such that for every µ0 ∈ C0 there exists µ ∈ C with µ0 = µ ∩ X. Proof. This is a well-known and very useful fact. We provide a proof for the reader's convenience. (cid:3) Let [X]<ℵ0 denote the set of all finite subsets of X. Since C is cofinal, we can find an increasing map f : [X]<ℵ0 → C satisfying s ⊆ f (s) for all s ∈ [X]<ℵ0 by induction on s. We define g : [X]ℵ0 → [X ∐ Y ]ℵ0 by g(µ0) := Ss⊆µ0 f (s) for µ0 ∈ [X]ℵ0 . For every µ0 ∈ [X]ℵ0 , we have µ0 ⊆ g(µ0) and g(µ0) ∈ C because f is increasing and C is closed. We set C0 := {µ0 ∈ [X]ℵ0 : µ0 = g(µ0) ∩ X}. NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 5 Then C0 is closed because for a countable directed Z ⊆ [X]ℵ0 , we have [µ0∈Z g(µ0) = g(cid:16) [µ0∈Z µ0(cid:17). It remains to show that C0 is cofinal in [X]ℵ0 . Take λ0 ∈ [X]ℵ0 arbitrarily. We define λ1, λ2, . . . ∈ [X]ℵ0 by λn+1 := g(λn) ∩ X for n = 0, 1, . . .. Then {λn}∞ n=0 λn is in C0. Thus C0 is cofinal. Therefore we get a club C0 in [X]ℵ0 as required. (cid:3) n=0 is an increasing sequence in [X]ℵ0 and µ0 := S∞ We note that by a well-known result of Kueker for every club C in [X]ℵ0 there exists h : [X]<ℵ0 → X such that every µ ∈ [X]ℵ0 closed under h belongs to C. 2.2. σ-complete directed families of subalgebras. By a subalgebra of a C-algebra we mean a C-subalgebra, and by a unital subalgebra of a unital C-algebra we mean a C-subalgebra containing the unit of the original C- algebra. By a directed family {Aλ}λ∈Λ of subalgebras of a C-algebra A, we mean that Λ is a directed set, and λ (cid:22) µ if and only if Aλ ⊆ Aµ. Thus by definition Λ ∋ λ 7→ Aλ is injective. Definition 2.8. A directed family {Aλ}λ∈Λ of subalgebras of a C-algebra A is said to be σ-complete if Λ is σ-complete and for every countable directed Z ⊆ Λ, Asup Z is the closure of the union of {Aλ}λ∈Z . In other words, a directed family {Aλ}λ∈Λ is σ-complete if Sλ∈Z Aλ is in the family for every countable directed Z ⊆ Λ. Lemma 2.9. Let A be a C-algebra, and let {Aλ}λ∈Λ be a σ-complete di- rected family of subalgebras of A with dense union. Then for a club Λ0 ⊆ Λ, the restriction {Aλ}λ∈Λ0 is also a σ-complete directed family with dense union. Proof. The restriction {Aλ}λ∈Λ0 is σ-complete because Λ0 is closed, and its union is dense because Λ0 is cofinal. (cid:3) Lemma 2.10. Every C-algebra A has a σ-complete directed family of sep- arable subalgebras with dense union. Proof. We can take the family of all separable subalgebras of A ordered by the inclusion. (cid:3) Lemma 2.11. Let A be a C-algebra, and let {Aλ}λ∈Λ be a σ-complete directed family of subalgebras of A with dense union. For every separable subalgebra A0 of A there exists λ ∈ Λ such that A0 ⊆ Aλ. Proof. Let {a1, a2, . . .} be a dense sequence of A0. For each n ∈ N, one can inductively find λn ∈ Λ such that ai ∈1/n Aλn for i = 1, 2, . . . , n and λn−1 (cid:22) λn because the family {Aλ}λ∈Λ is directed and its union is dense in A. Then λ := sup{λn : n ∈ N} ∈ Λ satisfies A0 ⊆ Aλ. (cid:3) 6 ILIJAS FARAH AND TAKESHI KATSURA By the lemma above, we can see that the union of a σ-complete directed 0 ⊆ Λ′ and an order isomorphism φ : Λ0 → Λ′ family is automatically closed. Proposition 2.12. Let A and B be C-algebras, and {Aλ}λ∈Λ and {Bλ′}λ′∈Λ′ be σ-complete directed families of separable subalgebras of A and B with dense union. Let Φ : A → B be an isomorphism. Then there exist clubs Λ0 ⊆ Λ and Λ′ 0 such that Φ[Aλ] = Bφ(λ) for all λ ∈ Λ0. If Λ = Λ′, then one can take Λ0 = Λ′ 0 and φ = id. Proof. Let Λ0 be the set of all λ ∈ Λ such that there exists λ′ ∈ Λ′ with 0 ⊆ Λ′ as the set of all λ′ ∈ Λ′ such that Φ[Aλ] = Bλ′. Similarly we define Λ′ there is λ ∈ Λ with Φ−1[Bλ′] = Aλ. Then there exists an order isomorphism φ : Λ0 → Λ′ 0 such that Φ[Aλ] = Bφ(λ) for all λ ∈ Λ0. We are going to show that Λ0 ⊆ Λ is a club. It is clear that Λ0 is closed. Take λ ∈ Λ. Since Aλ is separable, there exists λ′ by Lemma 2.11. By the same reason, there exists λ1 ∈ Λ such that Φ−1[Bλ′ ] ⊆ Aλ1. Then we have Aλ ⊆ Aλ1. In this way, we can find sequences Aλ ⊆ Aλ1 ⊆ Aλ2 ⊆ Aλ3 ⊆ ··· 1 ∈ Λ′ such that Φ[Aλ] ⊆ Bλ′ 1 1 Bλ′ 1 ⊆ Bλ′ such that Bλ′ and λ′ 2 ⊆ Bλ′ n ⊆ Φ[Aλn] and Φ[Aλn] ⊆ Bλ′ 0 ∈ Λ′ be the supremums of {λn}∞ for n = 1, 2, . . .. Let λ0 ∈ Λ n=1. Then we have , we get λ0 ∈ Λ0. 0 ⊆ Λ′ is This shows that Λ0 is cofinal, and hence it is a club. Similarly Λ′ a club. This shows the former assertion. The latter assertion follows from Lemma 2.6. (cid:3) n}∞ n. Since Φ[Aλ0] = Bλ′ Aλ0 =S∞ n=1 and {λ′ =S∞ n=1 Aλn and Bλ′ 0 n=1 Bλ′ 3 ⊆ ··· n+1 0 Lemma 2.13. A C-algebra A is LF if and only if it has a σ-complete directed family of separable AF algebras with dense union. Proof. We only need to prove the direct implication. We see that A has a σ-complete directed family of separable subalgebras {Aλ}λ∈Λ with dense union by Lemma 2.10. Since by [3, Theorem 2.2] every separable LF algebra is AF, it suffices to show that the set Λ0 of all λ ∈ Λ such that Aλ is LF is a club. Clearly Λ0 is closed. To show that Λ0 is cofinal, it suffices to see that for any separable subalgebra A0 of A, there exists a separable subalgebra A′ 0 containing A0 such that for any finite subset F of A0 and any ε > 0, there exists a finite-dimensional subalgebra M of A′ 0 with F ⊆ε M . This is easy to see. (cid:3) In the same way, one can show that a C-algebra A is LM if and only if it has a σ-complete directed family of separable AM subalgebras with dense union. Remark 2.14. Lemma 2.13 is just a special case of the downward Lowen- heim -- Skolem theorem for logic of metric structures ([1], or [9] for a version NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 7 suitable for study of C-algebras and II1 factors). Similar arguments have been used by C-algebraists to reflect properties of nonseparable algebras to separable subalgebras (see [2, II.8.5]) such as for example simplicity or the existence of the unique trace. 2.3. Tensor products. In this subsection, we give a definition and some properties of tensor products of C-algebras. We try to avoid using results on nuclear C-algebras as much as possible. In fact, we use the nuclear- ity only in Proposition 2.24 (and Lemma 2.22) which is used in the proof of Proposition 4.5 (3). We are interested in tensor products of possibly uncountably many unital C-algebras, and for this purpose the maximal tensor products are easier to treat than the minimal ones. We remark that we mainly deal with nuclear C-algebras for which there is no distinction between the minimal tensor products and the maximal ones. Definition 2.15. A family {Ax}x∈X of subalgebras of a C-algebra A is said to mutually commute if for distinct x, y ∈ X, every element of Ax commutes with every element of Ay. Definition 2.16. For a family {Ax}x∈X of unital C-algebras, its (maximal) tensor product Nx∈X Ax is the C-algebra having (an isomorphic copy of) Ax as unital subalgebras for x ∈ X satisfying the following two properties: (1) the family {Ax}x∈X of subalgebras ofNx∈X Ax mutually commutes, and its union Sx∈X Ax generates Nx∈X Ax. (2) for a unital C-algebra B and a family {ϕx}x∈X of unital ∗-homomor- phisms ϕx : Ax → B such that {ϕx[Ax]}x∈X is a mutually commuting family of unital subalgebras of B, there exists a unital ∗-homomor- phism ϕ : Nx∈X Ax → B such that ϕ ↾Ax= ϕx for all x ∈ X. When Ax = A for all x ∈ X, we simply write NX A for Nx∈X Ax. It is not difficult to see that the tensor product exists and is unique. A nice exposition of tensor products of C-algebras can be found e.g., in [4]. The condition (2) is called the universal property of the tensor product. A nice exposition of universal C-algebras can be found e.g., [2, II.8.3]. Let A and B be unital C-algebras. Since we consider A and B as unital subalgebras of A⊗ B, each a ∈ A and each b ∈ B are considered as elements of A ⊗ B. Thus the product ab ∈ A ⊗ B makes sense whereas this element is usually denoted by a ⊗ b ∈ A ⊗ B. Similarly, for a family {Ax}x∈X of unital C-algebras and a finite family {ax}x∈Y of elements with ax ∈ Ax for x ∈ Y ⊆ X, we denote by Qx∈Y ax ∈ Nx∈X Ax the product of {ax}x∈Y . Note that this product does not depend on the order of multiplications because the family {ax}x∈Y in Nx∈X Ax mutually commutes. The referee pointed out that the version of the next lemma when A is as- sumed to be nuclear and simple instead of LM is true (cf. [4, Corollary 9.4.6]). Since one can prove that LM algebras are nuclear and simple, this gives a proof of this lemma. We give an elementary proof for the reader's conve- nience. 8 ILIJAS FARAH AND TAKESHI KATSURA Lemma 2.17. Let A and B be unital subalgebras of a unital C-algebra D commuting with each other. If A is LM, then the natural map from A⊗ B to the C-subalgebra C ∗(A∪B) of D generated by A∪B ⊆ D is an isomorphism. Proof. We first show the statement in the case that A is a full matrix algebra i,j=1 be a matrix unit of A ∼= Mn(C). Then every element Mn(C). Let {ei,j}n of A ⊗ B can be written as Pn i,j=1 ei,jbi,j for bi,j ∈ B. In C ∗(A ∪ B) ⊆ D, we have ek,i′(cid:16) nXi,j=1 nXk=1 ei,jbi,j(cid:17)ej′,k for i′, j′ = 1, 2, . . . , n. Hence if an element Pn i,j=1 ei,jbi,j ∈ A ⊗ B is sent to 0 ∈ D by the natural map A ⊗ B → D, then bi,j = 0 for all i, j which implies Pn i,j=1 ei,jbi,j = 0 in A ⊗ B. Thus when A is a full matrix algebra, the natural map A⊗ B → C ∗(A∪ B) is injective, and hence an isomorphism. Now suppose that A is LM. Let π : A⊗B → C ∗(A∪B) be the natural map. Take x ∈ A ⊗ B. Take ε > 0 arbitrarily. Then there exist a1, a2, . . . , an ∈ A and b1, b2, . . . , bn ∈ B such that bi′,j′ = (cid:13)(cid:13)(cid:13)x − nXi=1 aibi(cid:13)(cid:13)(cid:13) < ε. Since A is LM, we may assume (by perturbing ai's slightly if necessarily) that a1, a2, . . . , an ∈ M for some unital full matrix subalgebra M of A. Then by the first part of the proof, we have (cid:13)(cid:13)(cid:13) Hence we get aibi(cid:17)(cid:13)(cid:13)(cid:13). aibi(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)π(cid:16) nXi=1 nXi=1 (cid:12)(cid:12)kxk − kπ(x)k(cid:12)(cid:12) < 2ε. Since ε was arbitrary, we have kxk = kπ(x)k. This shows that the natural map π : A ⊗ B → C ∗(A ∪ B) is injective, and hence an isomorphism. (cid:3) We take advantage of Lemma 2.17 and use the notation A ⊗ B whenever it is justified by this lemma. Note that this lemma is false if we replace LM by LF. To see this, just consider A = B = D = C ⊕ C. For a family {Ax}x∈X of unital C-algebras, and unital subalgebras Dx ⊆ Ax, we some- times denote by Nx∈X Dx the subalgebra of Nx∈X Ax generated by the mutually commuting family {Dx}x∈X of unital subalgebras ofNx∈X Ax. In fact, this unital subalgebra is the image of the ∗-homomorphism from the tensor product Nx∈X Dx to Nx∈X Ax, but no confusion should arise. We use the following well-known fact without mentioning. We give its proof for the reader's convenience. Lemma 2.18. Let A and B be unital C-algebras, and A0 ⊆ A and B0 ⊆ B be unital subalgebras. Then we have (A0 ⊗ B0) ∩ B = B0 in A ⊗ B. NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 9 Proof. Take a state ϕ of A. Define a linear map E : A⊗ B → B by E(ab) = ϕ(a)b for a ∈ A and b ∈ B. Since E(b) = b for b ∈ B and E(A0 ⊗ B0) ⊆ B0, we get (A0 ⊗ B0) ∩ B ⊆ B0. The inverse inclusion is easy to see. (cid:3) For two families {Ax}x∈X1 and {Ax}x∈X2 of unital C-algebras, the tensor product Nx∈X1∐X2 Ax is naturally isomorphic to (cid:16) Ox∈X1 Ax(cid:17) ⊗(cid:16) Ox∈X2 Ax(cid:17). We identify these two tensor products. In particular, we can and will con- sider Nx∈Y Ax as a unital subalgebra of Nx∈X Ax for a subset Y of X. We use the convention that Nx∈Y Ax = C for Y = ∅. We remark that the subalgebra Nx∈Y Ax coincides with Nx∈X Ax for a subset Y of X if and only if Ax = C for all x ∈ X \ Y . The following is easy to see. Lemma 2.19. Let {Ax}x∈X be an infinite family of unital C-algebras, and set A =Nx∈X Ax. Then {Nx∈λ Ax}λ∈[X]ℵ0 is a σ-complete directed system of subalgebras of A with dense union. (cid:3) Lemma 2.20. If A = Nx∈X Ax, X is infinite, and each Ax is separable and not C, then the character density χ(A) of A is equal to X. Proof. Fix a countable dense Cx ⊆ Ax for each x. Their union has cardi- nality X and generates A. This shows χ(A) ≤ X. Take a subset Z ⊆ A with cardinality less than X. For each z ∈ Z, there exists λz ∈ [X]ℵ0 with Ax by Lemma 2.19. Since the set Sz∈Z λz ⊆ X has cardinality z ∈ Nx∈λz less than X, we can find x ∈ X outside of this set. Since Ax is not C, Z is not dense in A. Hence χ(A) = X. (cid:3) For a unitary u of a unital C-algebra A, an automorphism Ad u on A is defined by Ad u(a) = uau∗ for a ∈ A. Let {Ax}x∈X be a family of unital C-algebras. By the universality, a family {αx}x∈X of automorphisms αx on Ax determines the automorphism α on Nx∈X Ax with α ↾Ax= αx which we denote by Nx∈X αx. For a subset Y ⊆ X and a family {αx}x∈Y of automorphisms αx on Ax, we denote by Nx∈Y αx the automorphism Nx∈X αx ofNx∈X Ax where αx = idAx for x ∈ X\Y . For unitaries ux ∈ Ax for x ∈ Y , we get an automorphism Nx∈Y Ad ux on A =Nx∈X Ax. When Y is finite, we get Nx∈Y Ad ux = Ad u where u = Qx∈Y ux ∈ A, but in general, Nx∈Y Ad ux is not in the form Ad u for a unitary u of A. 2.4. Relative commutants. For a subset A of a C-algebra B, we denote by ZB(A) the relative commutant (or centralizer ) of A inside B; ZB(A) := {b ∈ B : ab = ba for all a ∈ A} which is a subalgebra of B if A is closed under the ∗-operation (for example if A is a subalgebra). We avoid the common notation A′ ∩ B for ZB(A) in order to increase the readability of certain formulas. For a subset A of a 10 ILIJAS FARAH AND TAKESHI KATSURA If A is LM, then C-algebra B, we denote by C ∗(A) the subalgebra generated by A. Note that ZB(C ∗(A)) = ZB(A) for a subset A closed under the ∗-operation. We also note that ZB(A1 ∪ A2) = ZB(A1) ∩ ZB(A2). Lemma 2.21. Let A and D be unital C-algebras. ZA⊗D(A) = D. Proof. It is clear from the definition of tensor products that ZA⊗D(A) ⊃ D. Take x0 ∈ ZA⊗D(A). For any ε > 0, there exist elements a1, . . . , an ∈ A and d1, . . . , dn ∈ D such that kx0 −Pn i=1 aidik < ε. Since A is LM, we may assume that a1, . . . , an are in a full matrix unital subalgebra M of A. Let E : A⊗D → A⊗D be a contractive linear map defined by E(x) =RU uxu∗du for x ∈ A⊗D where du is the normalized Haar measure on the unitary group U of M . Since x0 ∈ ZA⊗D(A), we have E(x0) = x0. For a ∈ M and d ∈ D, we have E(ad) = tr(a)d where tr : M → C is the normalized trace. Hence we have kx0 −Pn i=1 tr(ai)bik < ε. This means that x0 ∈ε D. Since ε was arbitrary, x0 ∈ D. Thus we get ZA⊗D(A) ⊆ D, and therefore ZA⊗D(A) = D. We are done. (cid:3) By letting D = C in the lemma above, we see that the center ZA(A) of an LM algebra A is C. Thus one can write the conclusion of Lemma 2.21 as ZA⊗D(A) = ZA(A)⊗D. The referee pointed out that ZA⊗D(A) = ZA(A)⊗D holds for minimal tensor products by [12, Theorem 1]. Since one can prove that LM algebras are nuclear and satisfy ZA(A) = C, this gives an indirect proof of Lemma 2.21. To prove Proposition 4.5 (3), we need some facts on nuclear C-algebras (Lemma 2.22 and Proposition 2.24). When we apply Proposition 2.24 in the proof of Proposition 4.5 (3), we use the fact that a UHF algebra is a tensor product of separable nuclear C-algebras because full matrix algebras are nuclear. A nice exposition of nuclearity of C-algebras can be found e.g., in [4]. Lemma 2.22. Let A and D be unital C-algebras, and A0 a unital subal- gebra of A. Suppose that D is nuclear. Then ZA⊗D(A0) = ZA(A0) ⊗ D. Proof. Clearly we have ZA(A0) ⊗ D ⊆ ZA⊗D(A0). Let F := {c ∈ A ⊗ D : (id ⊗ω)(c) ∈ ZA(A0) for all ω ∈ D∗}. For a ∈ A ⊆ A ⊗ D and c ∈ A ⊗ D, we have (id⊗ω)(ac) = a(id⊗ω)(c), (id⊗ω)(ca) = (id⊗ω)(c)a for all ω ∈ D∗. Hence we get ZA⊗D(A0) ⊆ F . We claim that F = ZA(A0)⊗ D. This equality is usually referred to as the slice map property of the triple (D, A, ZA(A0)) (see [4, Definition 12.4.3]). Here we remark that the (maximal) tensor product considered in this paper coincides with the minimal one because D is nuclear. By [4, Theorem 12.4.4 (2)] (see [4, Defi- nition 12.4.1] and note nuclear⇔CPAP⇒SOAP), the triple (D, A, ZA(A0)) NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 11 has the slice map property because D is nuclear. Thus we have ZA(A0)⊗D = ZA⊗D(A0). (cid:3) Definition 2.23. Let A be a unital C-algebra, and A0 a unital subalgebra of A. We say that A0 is complemented in A if C ∗(A0 ∪ ZA(A0)) = A. In a tensor product A =Nx∈X Ax of unital C-algebras Ax, a subalgebra AY =Nx∈Y Ax is complemented for every subset Y of X. Proposition 2.24. Let A be a unital C-algebra. Suppose that there exists a unital C-algebra D such that A⊗ D is a tensor product of separable nuclear C-algebras. Then for a σ-complete directed system {Aλ}λ∈Λ of separable subalgebras of A with dense union, there exists a club Λ0 ⊆ Λ such that for each λ ∈ Λ0, Aλ is complemented in A. Proof. Fix a dense X ⊆ A and a dense Y ⊆ D. Then {C ∗(µ)}µ∈[X∐Y ]ℵ0 is a σ-complete directed family of separable subalgebras of A ⊗ D with dense union. Since A⊗ D is a tensor product of separable C-algebras, A⊗ D has a σ-complete directed system of separable complemented subalgebras with dense union by Lemma 2.19. Hence by Proposition 2.12, there exists a club C ⊆ [X ∐ Y ]ℵ0 such that C ∗(µ) is complemented in A ⊗ D for all µ ∈ C. By Lemma 2.7 there exists a club C0 ⊆ [X]ℵ0 such that for every µ0 ∈ C0 there exists µ ∈ C with µ0 = µ ∩ X. By Lemma 2.9, {C ∗(µ0)}µ0∈C0 is a σ-complete directed family of separable subalgebras of A with dense union. Hence by Proposition 2.12 applied with id : A → A, we get a club Λ0 ⊆ Λ such that for each λ ∈ Λ0 there exists µ0 ∈ C0 with Aλ = C ∗(µ0). It remains to prove that Aλ is complemented in A for every λ ∈ Λ0. Take λ ∈ Λ0. Then by the arguments above, there exists µ ∈ C such that Aλ = C ∗(µ∩ X). Then C ∗(µ) is complemented in A⊗ D, and we have Aλ ⊆ C ∗(µ) ⊆ Aλ ⊗ D. Since A⊗ D is a tensor product of nuclear C-algebras, D is nuclear by [4, Proposition 10.1.7]. Hence we get ZA(Aλ)⊗ D = ZA⊗D(Aλ) by Lemma 2.22. Therefore we have A ⊗ D = C ∗(cid:0)C ∗(µ) ∪ ZA⊗D(C ∗(µ))(cid:1) ⊆ C ∗(cid:0)(Aλ ⊗ D) ∪ ZA⊗D(Aλ)(cid:1) = C ∗(cid:0)(Aλ ⊗ D) ∪ (ZA(Aλ) ⊗ D)(cid:1) = C ∗(Aλ ∪ ZA(Aλ)) ⊗ D. This shows that Aλ is complemented in A and finishes the proof. (cid:3) 3. LM but not UHF Proposition 3.2 gives examples of unital LM algebras that are not UHF, answering part of [6, Problem 8.1]. Recall that Z is the Jiang -- Su algebra. We shall need the following properties of Z proved in [13]: • Z is a unital, separable C-algebra which is not UHF. • Nℵ0 Z ∼= Z. • A ⊗ Z ∼= A for any infinite-dimensional separable UHF algebra A. 12 ILIJAS FARAH AND TAKESHI KATSURA M2(C) is called the CAR algebra. Definition 3.1. The UHF algebra Nℵ0 Proposition 3.2. For two sets X and Y , define AX,Y := NX M2(C) ⊗ NY Z. Suppose that X is infinite. Then we have the following. (1) AX,Y is a unital LM algebra with χ(AX,Y ) = X + Y . (2) AX,Y is UHF if and only if X ≥ Y . (3) AX,Y ⊗ D is UHF for any UHF algebra D with χ(D) ≥ Y . Proof. Since X is infinite, we can identify AX,Y with NX A ⊗NY Z where A is the CAR algebra. For each λ ∈ [X]ℵ0 and λ′ ∈ [Y ]ℵ0 we set Dλ,λ′ :=Oλ A ⊗Oλ′ Z ⊆ AX,Y Then Dλ,λ′ is the CAR algebra for all λ and λ′. Since {Dλ,λ′} is a σ-complete directed system with dense union, we see that AX,Y is LM. By Lemma 2.20, we have χ(AX,Y ) = X + Y . This shows (1). By rearranging the factors, we see that AX,Y is UHF if X ≥ Y and that AX,Y ⊗ D is UHF for a UHF algebra D with χ(D) ≥ Y . It remains to show that AX,Y is UHF only if X ≥ Y . For the sake of obtaining a contradiction, assume that X < Y and AX,Y is UHF. Let us denote by Ax = M2(C) for x ∈ X and Ay = Z for y ∈ Y the unital subalgebra of AX,Y so that AX,Y =Nx∈X Ax⊗Ny∈Y Ay. Let Φ : AX,Y →Nz∈Z Mz be an isomorphism where Z is a set and {Mz}z∈Z is a family of full matrix algebras. For each x ∈ X, there exists a finite Fx ⊆ Z such that Φ[Ax] ⊆Nz∈Fx Mz. If we set Z1 = Sx∈X Fx ⊆ Z, then we get Z1 = X and Φ(cid:2)Nx∈X Ax(cid:3) ⊆ Nz∈Z1 Mz. Similarly, for each z ∈ Z1 there exists a finite Gz ⊆ Y such that Mz ⊆ Φ(cid:2)Nx∈X Ax ⊗Ny∈Gz Gz ⊆ Y , then we get Y1 ≤ Z1 = X and Oz∈Z1 Ay(cid:3). If we set Y1 = Sz∈Z1 Ayi. Ax ⊗ Oy∈Y1 Mz ⊆ ΦhOx∈X If we set Z2 = Z1 ∪Sy∈Y1 Ax ⊗ Oy∈Y1 ΦhOx∈X Mz. Next for each y ∈ Y1 there exists a countable Cy ⊆ Z such that Φ[Ay] ⊆ Nz∈Cy Cy ⊆ Z, then we get Z1 ⊆ Z2, Z2 = X and Ayi ⊆ Oz∈Z2 Recursively, we can find increasing sequences {Yk}n k=1 and {Zk}n sets of Y and Z, respectively, such that X ∐ Yk = Zk = X and k=1 of sub- Mz. Oz∈Zk Mz ⊆ ΦhOx∈X for k = 1, 2, . . .. We set Y ′ :=S∞ we have X ∐ Y ′ = Z ′ = X and Ayi ⊆ Oz∈Zk+1 Ax ⊗ Oy∈Yk k=1 Yk ⊆ Y and Z ′ :=S∞ Mz ΦhOx∈X Ax ⊗ Oy∈Y ′ Ayi = Oz∈Z ′ Mz. k=1 Zk ⊆ Z. Then NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 13 Since Nz∈Z ′ Mz is UHF and hence LM, we have ZAX,Y(cid:16)Ox∈X Ax ⊗ Oy∈Y ′ ZNz∈Z Mz(cid:16)Oz∈Z ′ Ay(cid:17) = Oy∈Y \Y ′ Mz(cid:17) = Oz∈Z\Z ′ Ay, Mz. by Lemma 2.21. Thus we get Φ(cid:2)Ny∈Y \Y ′ Ay(cid:3) =Nz∈Z\Z ′ Mz. Since Y ′ ≤ X < Y , we see that Y \ Y ′ is infinite. Hence Z \ Z ′ is also infinite. By Proposition 2.12 and Lemma 2.19, there exist C ∈ [Y \ Y ′]ℵ0 and C ′ ∈ [Z \ Z ′]ℵ0 such that Φ(cid:2)Ny∈C Ay(cid:3) = Nz∈C′ Mz. This is a contradiction because Ny∈C Ay ∼= Z is not UHF. We thank the referee who pointed out an error of a proof of Proposi- tion 3.2 (3) in an earlier draft. By Proposition 3.2, the unital C-algebra AX,Y is LM but not UHF if X < Y . When X = ℵ0 and Y = ℵ1, we see that AX,Y is AM by Theorem 1.3 (1). In the other case, we do not know whether AX,Y is AM or not. Problem 3.3. Let X, Y be sets such that ℵ0 ≤ X < Y and ℵ1 < Y . Is AX,Y =NX M2(C) ⊗NY Z AM? (cid:3) 4. AM but not UHF In this section, for each infinite set X we define a unital AM-algebra BX with χ(BX ) = X, and show that BX, or even BX ⊗ D for a unital C-algebra D, is not UHF when X ≥ ℵ1. Lemma 4.1. A C-algebra generated by two self-adjoint unitaries v, w with vw = −wv is always isomorphic to M2(C). Proof. A C-algebra A generated by two self-adjoint unitaries v, w with vw = −wv is spanned (as a vector space) by 4 elements {1, v, w, vw}, and it is noncommutative. Hence it is isomorphic to M2(C) which is the unique noncommutative C-algebra with dimension ≤ 4. A concrete isomorphism from A to M2(C) is given by sending v and w to the unitaries 0 (cid:18)1 0 −1(cid:19) and (cid:18)0 1 1 0(cid:19) in M2(C). (cid:3) Let us take a set X. For each x ∈ X, let Ax be a C-algebra generated by two self-adjoint unitaries vx, wx with vxwx = −wxvx. By Lemma 4.1, Ax is isomorphic to M2(C). We define a UHF algebra AX by AX :=Nx∈X Ax ∼= NX M2(C). We define an automorphism α on AX by α := Nx∈X Ad vx. Note that α2 = id. Let {ei,j}2 and define an embedding i,j=1 be a system of matrix units of M2(C), ι : AX ∋ a 7→ ae1,1 + α(a)e2,2 ∈ AX ⊗ M2(C). 14 ILIJAS FARAH AND TAKESHI KATSURA Let u ∈ AX ⊗ M2(C) be a self-adjoint unitary defined by u := e1,2 + e2,1. Set BX := C ∗(ι(AX ) ∪ {u}). We consider AX as a unital subalgebra of BX and omit ι. Then we have ua = α(a)u for a ∈ AX and BX = {au + a′ : a, a′ ∈ AX}. Remark 4.2. The C-algebra BX is nothing but the crossed product AX ⋊α (Z/2Z). For Y ⊆ X, we denote by AY the subalgebra Nx∈Y Ax ⊆ AX ⊆ BX, and define BY := C ∗(AY ∪ {u}) ⊆ BX . It is easy to see that AY ⊆ AX is globally invariant under α, and hence BY = {au + a′ : a, a′ ∈ AY }. Lemma 4.3. If Y is infinite then ZBX (AY ) = AX\Y . Proof. Since ZAX (AY ) = AX\Y by Lemma 2.21, it suffices to show that ZBX (AY ) ⊆ AX. Take au + a′ ∈ ZBX (AY ) with a, a′ ∈ AX, and we will show a = 0. For any ε > 0 there is a finite F ⊆ X such that a ∈ε AF . Since Y is infinite, pick y ∈ Y \ F . The unitary wy ∈ AY satisfies uwy = −wyu. Hence wy(au + a′) = (au + a′)wy yields (wya + awy)u + (wya′ − a′wy) = 0. Since bu + b′ = 0 for b, b′ in AX implies b = b′ = 0, we have wya = −awy. Thus we get kak = kwyak = kwya + wyak/2 = kwya − awyk/2. Since a ∈ε AF and wy commutes with AF , we have kak = kwya−awyk/2 < ε. Since ε was arbitrary, a = 0. We are done. (cid:3) Lemma 4.4. If Y ( X and Y is infinite, then BY is not complemented in BX . Proof. Since BY = C ∗(AY ∪ {u}), we have ZBX (BY ) = ZBX (AY ) ∩ ZBX ({u}) = AX\Y ∩ ZBX ({u}) by Lemma 4.3. Hence C ∗(BY ∪ ZBX (BY )) =(cid:8)au + a′ : a, a′ ∈ AY ⊗ (AX\Y ∩ ZBX ({u}))(cid:9) which does not contain wy ∈ AX\Y for y ∈ X \ Y . Proposition 4.5. (1) If X is infinite, then BX is a unital AM algebra (cid:3) with χ(BX) = X. (2) If X is uncountable, then BX is not UHF. (3) If X is uncountable, then BX ⊗ D is not UHF for any unital C- algebra D. Proof. (1) Suppose X is infinite. Let us set and define Λ = {(F, y) : F ⊆ X finite, and y ∈ X \ F} D(F,y) = C ∗(cid:0)BF ∪ {wy}(cid:1) ⊆ BX . NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 15 for (F, y) ∈ Λ. Then it is clear that {D(F,y)}(F,y)∈Λ is a directed family with dense union. It remains to show that D(F,y) is a full matrix algebra for (F, y) ∈ Λ. Take (F, y) ∈ Λ with F = n ∈ N. Then we have AF ∼= M2n(C), and the restriction of α to AF coincides with Ad v where v = Yx∈F vx ∈ AF . Then the two self-adjoint unitaries uv and wy in D(F,y) satisfy wy(uv) = −(uv)wy and commute with AF . By Lemma 4.1, the subalgebra of D(F,y) generated by uv and wy is isomorphic to M2(C), and commute with AF . Since D(F,y) is generated by AF and this subalgebra, D(F,y) is isomorphic to M2n+1(C). We are done. (2) Suppose X is uncountable. Then {BY }Y ∈[X]ℵ0 is a σ-complete di- rected family of separable subalgebras of BX with dense union. By Lemma 4.4, neither one of these subalgebras is complemented. By Lemma 2.19, a UHF algebra has a σ-complete directed system of separable complemented subal- gebras with dense union. Hence BX cannot be UHF by Lemma 2.7. (3) As in (2), BX has a σ-complete directed system of separable subal- gebras with dense union neither one of which is complemented. By Propo- sition 2.24, BX ⊗ D cannot be UHF for any unital C-algebra D because every UHF algebra is a tensor product of separable nuclear C-algebras. (cid:3) Note that an example of a unital LM algebra A that is not UHF given in Proposition 3.2 has the property that A ⊗ D is UHF for some UHF algebra D, but the one given in Proposition 4.5 does not have this property. The following answers [6, Problem 8.3] negatively although it was cer- tainly known. Corollary 4.6. There is a proper subalgebra A of the CAR algebra B such that A is also CAR algebra and ZB(A) = C1. In particular, B 6= A⊗ZB(A). Proof. Use Proposition 4.5 with X = N. Then AX is the CAR algebra. The C-algebra BX is also the CAR algebra because it is a separable unital LM algebra obtained as a direct limit of algebras of the form M2n(C) by the proof of Proposition 4.5. By Lemma 4.3, we have ZBX (AX ) = C1. (cid:3) 5. AM = LM and AF = LF for character density ≤ ℵ1 We first show AM = LM + AF. We use the following well-known result repeatedly. Recall that a finite-dimensional C-algebra D is isomorphic to a direct sum of finitely many full matrix algebras (e.g., [5, Theorem III.1.1]), and the cardinality F of a system F of matrix units of D as defined after [5, Theorem III.1.1] coincides with the dimension of D. Lemma 5.1 ([5, Corollary III.3.3]). Given d ∈ N, there exists δ > 0 so that if D is a finite-dimensional subalgebra of a C-algebra A with a system F of matrix units such that F = d and B is a subalgebra of A such that F ⊆δ B, 16 ILIJAS FARAH AND TAKESHI KATSURA there exists a unitary u in the unitization of A satisfying uDu∗ ⊆ B and commuting with D ∩ B. Moreover, for a previously given ε > 0 in addition to d, there exists δ > 0 such that one can choose u as above so that ku − 1k < ε. (cid:3) Proposition 5.2. A C-algebra is AM if and only if it is LM and AF. Proof. We only need to prove that if a C-algebra A is LM and AF, then it is AM. Take a directed family {Dλ}λ∈Λ of finite-dimensional subalgebras of A with dense union. To show that A is AM, it suffices to show that for any λ ∈ Λ there exists a full matrix subalgebra M containing Dλ and contained in Dλ′ for some λ′ (cid:23) λ. Then the set of such full matrix subalgebras is directed and has dense union. Take λ ∈ Λ. Let F be a system of matrix units of Dλ. Let δ > 0 be as in Lemma 5.1 for d = F ∈ N. Since A is LM, it has a full matrix subalgebra M0 such that F ⊆δ M0. By Lemma 5.1, there exists a unitary u in the unitization of A satisfying uDλu∗ ⊆ M0. Let F ′ be a system of matrix units of u∗M0u. Let δ′ > 0 be as in Lemma 5.1 for d = F ′. Since {Dλ}λ∈Λ has dense union, there exists λ′ ∈ Λ such that λ′ (cid:23) λ and F ′ ⊆δ′ Dλ′. By Lemma 5.1, there exists a unitary u′ in the unitization of A satisfying u′(u∗M0u)u′∗ ⊆ Dλ′ and commuting with (u∗M0u) ∩ Dλ′. Since Dλ ⊆ (u∗M0u) ∩ Dλ′, the full matrix subalgebra M := u′(u∗M0u)u′∗ of A satisfies Dλ ⊆ M ⊆ Dλ′. This completes the proof. (cid:3) By Proposition 5.2, the statement AM = LM is reduced to AF = LF because LM implies LF. Thus we only show that AF = LF for character density at most ℵ1 although the same argument as below works for showing AM = LM directly by just changing "F" to "M" and "finite-dimensional" to "full matrix" in all statements and proofs. Lemma 5.3. Let A be a separable AF algebra. For an increasing sequence {Dn}n∈N of finite-dimensional subalgebras of A there exists an increasing sequence {D′ n}n∈N of finite-dimensional subalgebras with dense union such that Sn∈N Dn ⊆Sn∈N D′ Proof. This is well-known to specialists, and can be shown in a similar way to [5, Theorem III.3.5]. For the reader's convenience, we give a proof. n. Let {Bk}k∈N be an increasing sequence of finite-dimensional subalgebras of A with dense union. We construct inductively an increasing sequence {kn}n∈N in N and a sequence {un}n∈N of unitaries in the unitization of A with kun − 1k < 2−n such that for each n ∈ N the finite-dimensional algebra is contained in Bkn and commutes with un+1. We first construct k1 and u1. Choose k1 ∈ N such that F ⊆δ Bk1 where F is a system of matrix units of D1 and δ > 0 be as in the latter statement of Lemma 5.1 for d = F and ε = 2−1. By Lemma 5.1, there exists a unitary u1 in the unitization of A satisfying u1D1u∗ 1 ⊆ Bk1 and ku1− 1k < 1/2. Suppose that k1, . . . , kn−1 ∈ N un ··· u2u1Dnu∗ 1u∗ 2 ··· u∗ n NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 17 and unitaries u1, . . . , un−1 were chosen. Choose kn ∈ N such that kn > kn−1 and F ′ ⊆δ′ Bkn where F ′ is a system of matrix units of un−1 ··· u2u1Dnu∗ 1u∗ 2 ··· u∗ n−1 and δ′ > 0 be as in the latter statement of Lemma 5.1 for d = F ′ and ε = 2−n. Lemma 5.1 gives us a unitary un in the unitization of A satisfying unun−1 ··· u2u1Dnu∗ 1u∗ 2 ··· u∗ n−1u∗ n ⊆ Bkn, commuting with un−1 ··· u2u1Dn−1u∗ 1u∗ 2 ··· u∗ n−1 ⊆ un−1 ··· u2u1Dnu∗ 1u∗ 2 ··· u∗ n−1 ∩ Bkn and satisfying kun − 1k < 2−n. Thus we get the desired sequences {kn}n∈N and {un}n∈N. Since kun − 1k < 2−n for all n ∈ N andPn∈N 2−n = 1 < ∞, the sequence {un ··· u2u1}n∈N converges to a unitary u in the unitization of A. Since un ··· u2u1Dnu∗ 2 ··· u∗ 1u∗ n = un+1un ··· u2u1Dnu∗ 1u∗ ⊆ un+1un ··· u2u1Dn+1u∗ 2 ··· u∗ 1u∗ nu∗ 2 ··· u∗ n+1 nu∗ n+1, 1u∗ 1u∗ 1u∗ n−1u∗ 2 ··· u∗ 2 ··· u∗ λ for each λ ∈ Λ. 2 ··· u∗ n := u∗Bknu for n ∈ N. Then {D′ n. un ··· u2u1Dnu∗ n commutes with un+2. By repeating this argument, one can see that un ··· u2u1Dnu∗ n commutes with um for all m > n. Hence we get uDnu∗ = unun−1 ··· u2u1Dnu∗ n ⊆ Bkn. We set D′ n}n∈N is an increasing sequence of finite-dimensional subalgebras with dense union such that Sn∈N Dn ⊆ Sn∈N D′ (cid:3) In the next lemma, for two families Υ = {Dλ}λ∈Λ and Υ′ = {D′ λ}λ∈Λ′ of subalgebras, Υ ⊆ Υ′ means that Λ ⊆ Λ′ and Dλ = D′ Lemma 5.4. Let A be a separable AF algebra contained in a separable AF algebra A′. For a countable directed family Υ of finite-dimensional subalge- bras of A with dense union, there exists a countable directed family Υ′ of finite-dimensional subalgebras of A′ with dense union such that Υ ⊆ Υ′. Proof. Let us write Υ = {Dλ}λ∈Λ. Since Λ is countable, we can choose a subsequence {λn}n∈N of Λ such thatSλ∈Λ Dλ =Sn∈N Dλn. By Lemma 5.3, there exists an increasing sequence {D′ n}n∈N of finite-dimensional subalge- bras of A′ with dense union such thatSn∈N Dλn ⊆Sn∈N D′ n. For each λ ∈ Λ, there exists n ∈ N such that Dλ ⊆ D′ n because Dλ is finite-dimensional. Let Λ′ := Λ ∐ N, ordered by requiring that Λ and N have their natural order- n. Then the family Υ′ := {D′ ings and λ (cid:22) n if Dλ ⊆ D′ λ}λ∈Λ′ defined by D′ λ := Dλ for λ ∈ Λ satisfies the desired properties. (cid:3) Lemma 5.5. Each LF algebra of character density at most ℵ1 has a σ- complete directed family of separable AF subalgebras with dense union in- dexed by the ordinal ω1. 18 ILIJAS FARAH AND TAKESHI KATSURA Proof. Let A be an LF algebra with χ(A) ≤ ℵ1. Fix a dense subset {xγ : γ ∈ ω1} of A, and define Aλ := C ∗({xγ : γ < λ}) for each λ ∈ ω1. Then {Aλ}λ∈ω1 is a σ-complete directed family of separable subalgebras of A. By Lemma 2.13, A also has a σ-complete direct family of separable AF subalgebras with dense union. By Proposition 2.12 applied with id : A → A, there is a club Λ ⊆ ω1 such that Aλ is AF for λ ∈ Λ. As ordered sets, Λ is isomorphic to ω1, and {Aλ}λ∈Λ is the desired family. (cid:3) Proposition 5.6. Each LF algebra of character density at most ℵ1 is an AF algebra. Proof. Let A be an LF algebra with χ(A) ≤ ℵ1. Let {Aξ}ξ∈ω1 be a σ- complete directed family of separable AF subalgebras of A with dense union as in Lemma 5.5. Using transfinite recursion, we are going to construct an increasing family of countable directed families Υξ of finite-dimensional subalgebras whose union is dense in Aξ for each ξ ∈ ω1. For ξ = 0, choose an increasing sequence of finite-dimensional subalgebras of A0 with dense union, and set it Υ0. If Υξ has been defined, then Υξ+1 is defined using Lemma 5.4. If η is a limit ordinal and Υξ has been defined for all ξ < η, required. let Υη =Sξ<η Υξ. Since Aη is the closure of the union of {Aξ}ξ<η, Υη is as Finally let Υ = Sξ∈ω1 Υξ. Then this is a directed family of finite- dimensional subalgebras of A with dense union. Thus A is an AF alge- bra. (cid:3) The example of the following section easily shows that the version of Lemma 5.4 for nonseparable algebras is false. 6. AM 6= LM and AF 6= LF for character density > ℵ1 In this section, we construct an LM algebra which is not AF. This C- algebra shows the difference between the classes of AM and LM algebras as well as between the classes of AF and LF algebras. To show that a given C-algebra is not AF, we use the following criterion. The converse direction in the following lemma was proved by George Elliott, following a remark by Tamas Matrai, during the first author's talk at a set theory seminar in Toronto in April 2009. Lemma 6.1. A C-algebra A is AF if and only if there exists a map ρ : A → A such that ka − ρ(a)k < 1 for every a ∈ A and C ∗({ρ(a)}a∈F ) is finite- dimensional for every finite subset F of A. Proof. Assume A is AF and let {Aλ}λ∈Λ be a directed family of finite- dimensional subalgebras of A with dense union. For each a ∈ A there exists λa ∈ Λ such that there exists ρ(a) ∈ Aλa with ka − ρ(a)k < 1. For every finite subset F of A there exists λ ∈ Λ such that λ (cid:23) λa for all a ∈ F . Then C ∗({ρ(a)}a∈F ) ⊆ Aλ is finite-dimensional. NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 19 Now assume that ρ : A → A is as in the statement of the lemma. If Λ is the family of all finite subsets of A then Aλ = C ∗({ρ(a)}a∈λ) form a directed family of finite-dimensional subalgebras of A. Fix a ∈ A and ε > 0. Let λ = {a/ε}. Then ερ(a/ε) ∈ Aλ and ka − ερ(a/ε)k < ε. Since a and ε were arbitrary, we conclude A is AF. (cid:3) We also use the following lemma (for the case when A is the CAR algebra) in the proof of Proposition 6.12 Lemma 6.2. Let A be a unital LM subalgebra of a unital C-algebra B. Take a1, a2, . . . , an ∈ A and b1, b2, . . . , bn ∈ ZB(A). i=1 is linearly independent in A and Pn Proof. Since A is LM, the natural map from A ⊗ ZB(A) to B is injective by Lemma 2.17. It is well known that the inclusion map from the algebraic tensor product of A and ZB(A) to the (maximal) tensor product A⊗ ZB(A) is injective (see [2, II.9.1.3]). The conclusion follows from these lemmas. (cid:3) i=1 aibi = 0 in B, then we have bi = 0 for all i. If (ai)n Definition 6.3. We say that a pair (v1, v2) of self-adjoint unitaries v1, v2 in a unital C-algebra is generic if the family (cid:0)(v1v2)n, (v1v2)nv1(cid:1)n∈Z = (1, v1, v2, v1v2, v2v1, v1v2v1, v2v1v2, v1v2v1v2, v2v1v2v1, v1v2v1v2v1, . . .) is linearly independent. In other words, (v1, v2) is generic if and only if the map sending the natural generators of the group algebra C((Z/2Z) ∗ (Z/2Z)) to v1, v2 is injective. Lemma 6.4. Let v1, v2, w1, w2 be the four self-adjoint unitaries in the C- algebra C([0, 1], M2(C)) defined by 0 v1(t) =(cid:18)1 0 −1(cid:19) , 1 0(cid:19) , w1(t) =(cid:18)0 1 sin(πt) v2(t) =(cid:18)cos(πt) w2(t) =(cid:18)− sin(πt) sin(πt) − cos(πt)(cid:19) , sin(πt)(cid:19) cos(πt) cos(πt) for t ∈ [0, 1]. Then v1, v2, w1, w2 satisfy v1w1 = −w1v1, v2w2 = −w2v2 and the pair (v1, v2) is generic. Proof. It is routine to check the two equalities v1w1 = −w1v1 and v2w2 = −w2v2. That the pair (v1, v2) is generic comes from the fact that {cos(nπt)+ √−1 sin(nπt)}n∈Z is linearly independent in C([0, 1]). We leave the details to the readers. (cid:3) Let X be an infinite set, and [X]2 be the set of all subsets of X with cardinality 2. For ξ = {x, y} ∈ [X]2 let Aξ be the CAR algebra. We fix four self-adjoint unitaries vx,y, vy,x, wx,y, wy,x in Aξ such that vx,ywx,y = −wx,yvx,y, vy,xwy,x = −wy,xvy,x and the pair (vx,y, vy,x) is generic. Such unitaries exist by Lemma 6.4 because there exists a unital embedding from C([0, 1], M2(C)) to the CAR algebra. 20 ILIJAS FARAH AND TAKESHI KATSURA M2(C). We define a UHF algebra A[X]2 by A[X]2 =Nξ∈[X]2 Aξ ∼=N[X]2×ℵ0 For a subset Y of X, we set A[Y ]2 =Nξ∈[Y ]2 Aξ ⊆ A[X]2. Definition 6.5. For a set X, we denote by GX the abelian group consisting of all finite subsets of X where the operation is the symmetric difference ∆. We often identify an element x of X with a subset {x} of X. Thus the group GX is generated by the family {x}x∈X of mutually commuting involutions. Hence GX is isomorphic to the group LX(Z/2Z) of the direct sum of X copies of Z/2Z. For g ∈ GX we define an automorphism αg on A[X]2 by αg = Ox ∈ g and y /∈ g Ad vx,y. If x /∈ g define unitaries Vg;x and Vx;g in A[X]2 via Vg;x =Yy∈g vy,xvx,y and vx,yvy,x. Vx;g =Yy∈g Lemma 6.6. If x /∈ g then αg ◦ αx = Ad(Vg;x) ◦ αg∪{x} and αg∪{x} ◦ αx = Ad(Vx;g) ◦ αg. Proof. Note that vx,y and vz,t commute unless z = y and x = t. Using x /∈ g we have αg ◦ αx =(cid:16) Oy∈g and z /∈g Ad vy,z(cid:17) ◦(cid:16)Oz6=x Ad vy,x ◦ Oy∈g and z /∈g∪{x} Ad vx,z(cid:17) Ad vy,z(cid:17) ◦(cid:16)Oz∈g Ad vx,z ◦ Oz /∈g∪{x} Ad vx,z(cid:17) =(cid:16)Oy∈g = Ad(Vg;x) ◦ αg∪{x} This proves the first equality. Since αx is an involution and Vx;g = V ∗ first equality implies the second equality. g;x, the (cid:3) Let us choose a faithful representation A[X]2 ⊆ B(H) on some Hilbert space H (see Section 7 for one construction of such a representation). Let ℓ2(GX , H) be the Hilbert space consisting of functions ξ : GX → H with Pg∈GX kξ(g)k2 < ∞. We embed A[X]2 into B(ℓ2(GX , H)) by for a ∈ A[X]2, ξ ∈ ℓ2(GX , H) and g ∈ GX . For each x ∈ X, we define ux ∈ B(ℓ2(GX , H)) by (aξ)(g) = αg(a)ξ(g) ∈ H (uxξ)(g) = Vg;x ξ(g ∪ {x}) ∈ H (uxξ)(g ∪ {x}) = Vx;g ξ(g) ∈ H for ξ ∈ ℓ2(GX , H) and g ∈ GX with x /∈ g. Lemma 6.7. For each x ∈ X, ux is a self-adjoint unitary such that Ad ux and αx agree on A[X]2 ⊆ B(ℓ2(GX , H)). NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 21 Proof. For g ∈ GX such that x /∈ g the subspace ℓ2({g, g ∪ {x}}, H) ⊆ ℓ2(GX , H) is invariant for ux, and ux is represented on it as 0 (cid:19) . This shows that ux is a self-adjoint unitary. To show that Ad ux and αx agree on A[X]2 ⊆ B(ℓ2(GX , H)), it suffices to see ux =(cid:18) 0 Vg;x Vx;g Ad(Vx;g) ◦ αg = αg∪{x} ◦ αx Ad(Vg;x) ◦ αg∪{x} = αg ◦ αx which is Lemma 6.6. (cid:3) By Lemma 6.7 we see that for {x, y} ∈ [X]2 and z ∈ X we have Ad ux ↾A{x,y}= Ad vx,y, and Ad uz ↾A{x,y}= id if z /∈ {x, y}. In particular, uz commutes with vx,y unless y = z. Lemma 6.8. For {x, y} ∈ [X]2 the two self-adjoint unitaries uxvx,y and uyvy,x commute. Proof. Take {x, y} ∈ [X]2. First note that for h ∈ GX we have αh(vx,y) = vy,xvx,yvy,x if y ∈ h and x /∈ h, and αh(vx,y) = vx,y otherwise. Fix g ∈ GX such that x /∈ g and y /∈ g. The subspace Hg = ℓ2(cid:0){g, g ∪ {x}, g ∪ {y}, g ∪ {x, y}}, H(cid:1) ⊆ ℓ2(GX , H) is invariant for each of ux, uy, vx,y and vy,x. Using the observation in the beginning of this proof, we see that ux, uy, vx,y and vy,x are represented on Hg by ux =  uy =  0 Vx;g 0 0 0 0 Vy;g 0 0 0 0 Vg;x 0 0 0 Vx;g∪{y} Vg;y 0 0 Vg∪{y};x 0 0 0 Vg∪{x};y 0 Vy;g∪{x} 0 0 0 0 0 0  vx,y =  ,    , vy,x =  vx,y 0 0 0 vy,x 0 0 0 0 vx,y 0 0 0 vy,xvx,yvy,x 0 0 0 vx,yvy,xvx,y 0 0 0 0 vy,x 0 0 0 0 vx,y 0 0 0 vy,x   ,   . Using the computations such as Vx;g∪{y} = Vx;gvx,yvy,x, we see that uxvx,y and uyvy,x are represented on Hg by uxvx,y =  uyvy,x =  0 Vg;xvx,y Vx;gvx,y 0 0 0 0 Vy;gvy,x 0 0 0 0 0 0 0 Vy;gvx,y 0 0 0 0 0 Vg;xvy,x Vx;gvy,x Vg;yvy,x 0 0 0 0 0 Vg;yvx,y 0 0   ,   . 22 ILIJAS FARAH AND TAKESHI KATSURA The unitaries Vg;x, Vx;g, Vg;y, Vy;g, vx,y and vy,x occurring in entries of these two matrices commute with each others except that vx,y does not commute with vy,x. Using this fact, one can show that both (uxvx,y)(uyvy,x) and (uyvy,x)(uxvx,y) are equal to Therefore uxvx,y and uyvy,x commute. (cid:3) 0 0 0 0 0 0 0 0 0 Vx;gVy;g Vg;xVg;y Vg;xVy;gvy,xvx,y Vx;gVg;yvx,yvy,x     . B[X]2 := C ∗(A[X]2 ∪ {ux}x∈X ) ⊆ B(cid:0)ℓ2(GX , H)(cid:1). 0 0 0 B[Y ]2 := C ∗(A[Y ]2 ∪ {ux}x∈Y ) ⊆ B[X]2. Let For a subset Y ⊆ X, we define Remark 6.9. The C-algebra B[X]2 does not depend on the choices of em- beddings A[X]2 ⊆ B(H), and is isomorphic to a cocycle crossed product A[X]2 ⋊(α,c) GX for an appropriate cocycle action (α, c) (see [15] for defini- tions of cocycle actions and cocycle crossed products). In fact, the proof of Proposition 6.10 shows that any C-algebra generated by A[X]2 ∪ {ux}x∈X with the relations in Lemma 6.7 and Lemma 6.8 is isomorphic to B[X]2. Proposition 6.10. The C-algebra B[X]2 is a unital LM algebra with χ(B[X]2) = X. Proof. By Lemma 2.20, we have χ(A[X]2) = X. This implies χ(B[X]2) = X. We are going to show that B[X]2 is a direct limit of CAR algebras. This implies that B[X]2 is LM. For a finite subset F ⊆ X and an injective map ι : F → X \ F , define a subalgebra D(F,ι) ⊆ B[X]2 by D(F,ι) := C ∗(cid:0)B[F ]2 ∪ {wx,ι(x)}x∈F(cid:1) ⊆ B[X]2. The family {D(F,ι)}(F,ι) of subalgebras is directed because X is infinite, and its union is dense in B[X]2. Thus it suffices to show that D(F,ι) is the CAR algebra for every finite subset F ⊆ X and every injective map ι : F → X\F . Take a finite subset F ⊆ X and an injective map ι : F → X \ F . For x ∈ F , we define which is a self-adjoint unitary. Since Lemma 6.7 shows u′ x := ux Yy∈F \{x} vx,y ∈ D(F,ι). Ad ux ↾A[F ]2 = αx ↾A[F ]2 = Ad(cid:16) Yy∈F \{x} vx,y(cid:17) ↾A[F ]2 , x commutes with the subalgebra A[F ]2. The family {u′ u′ x}x∈F mutually com- mutes by Lemma 6.8. For each x ∈ F , the self-adjoint unitary wx,ι(x) ∈ NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 23 D(F,ι) commutes with A[F ]2 and {wy,ι(y), u′ x. Therefore C ∗(u′ −wx,ι(x)u′ Lemma 4.1, and the family {C ∗(u′ x, wx,ι(x))}x∈F ∪ {A[F ]2} xwx,ι(x) = x, wx,ι(x)) is isomorphic to M2(C) for x ∈ F by y}y∈F \{x}, and satisfies u′ mutually commutes. Since D(F,ι) is generated by these mutually commuting subalgebras, we get D(F,ι) = A[F ]2 ⊗Ox∈F C ∗(u′ x, wx,ι(x)) ∼= O[F ]2×ℵ0+F M2(C). We are done. (cid:3) Lemma 6.11. Let Y be a nonempty proper subset of X. Take x ∈ Y and y ∈ X \ Y . Then every element in B[Y ]2 ⊆ B[X]2 can be written as avx,y + a′ for a, a′ ∈ ZB[X]2 (A{x,y}). Proof. Since vx,y is a self-adjoint unitary in A{x,y}, the set of all elements in the form avx,y +a′ for a, a′ ∈ ZB[X]2 (A{x,y}) is a subalgebra of B[X]2. Hence it suffices to show that the generators A[Y ]2 ∪{uz}z∈Y of B[Y ]2 are in this form. We have A[Y ]2 ⊆ ZB[X]2 (A{x,y}) since y /∈ Y . We have uz ∈ ZB[X]2 (A{x,y}) for z ∈ Y \{x}. Finally, we get ux = (uxvx,y)vx,y and uxvx,y ∈ ZB[X]2 (A{x,y}). We are done. (cid:3) Proposition 6.12. If X > ℵ1 then B[X]2 is not AF. Proof. For the sake of obtaining a contradiction, assume that B[X]2 is AF. Then by Lemma 6.1 there exists a family {bx}x∈X in B[X]2 with kux−bxk < 1 for all x ∈ X such that C ∗({bx}x∈F ) ⊆ B[X]2 is finite-dimensional for all finite subsets F of X. For each x ∈ X, there exists a countable subset Yx of X with x ∈ Yx such that bx ∈ B[Yx]2. Since X > ℵ1, we can apply Lemma 2.1 to get {x, y} ∈ [X]2 such that x /∈ Yy and y /∈ Yx. By Lemma 6.11, there exists x and by = ayvy,x+a′ ax, a′ y. Since kux − bxk < 1, we have x, ay, a′ y ∈ ZB[X]2 (A{x,y}) such that bx = axvx,y +a′ (cid:13)(cid:13)(cid:0)(ux − bx) − wx,y(ux − bx)wx,y(cid:1)/2(cid:13)(cid:13) < 1. x) − (−axvx,y + a′ (cid:0)bx − wx,ybxwx,y(cid:1)/2 =(cid:0)(axvx,y + a′ x)(cid:1)/2 = axvx,y, We have and similarly (ux − wx,yuxwx,y)/2 = ux. Hence we get kux − axvx,yk < 1. Thus kuxvx,y−axk < 1. Since uxvx,y is a unitary, ax is an invertible element. Similarly, one can show that ay is also invertible. By the assumption, C ∗({bx, by}) is finite-dimensional. Therefore {(bxby)n}∞ is linearly dependent. Hence there exist N ∈ N and λ0, λ1, . . . , λN ∈ C with n=0 24 ILIJAS FARAH AND TAKESHI KATSURA λN 6= 0 such that PN λn(bxby)n = NXn=0 n=0 λn(bxby)n = 0. We can write NXn=0 λn(cid:0)(axvx,y + a′ x)(ayvy,x + a′ y)(cid:1)n = Xv∈V fvv x, ay, a′ where V := {1, vx,y, vy,x, vx,yvy,x, vy,xvx,y, vx,yvy,xvx,y, vy,xvx,yvy,x, . . . , (vx,yvy,x)N} and for each v ∈ V , fv ∈ ZB[X]2 (A{x,y}) is a sum of products of λ0, λ1, . . . , λN ∈ C and ax, a′ y ∈ ZB[X]2 (A{x,y}). Since V ⊆ A{x,y} is linearly in- dependent, we get fv = 0 for all v ∈ V by Lemma 6.2. In particular, f(vx,yvy,x)N = λN (axay)N ∈ ZB[X]2 (A{x,y}) is 0. This cannot happen because λN 6= 0 and both ax and ay are invertible. Thus we get a contradiction. We are done. (cid:3) Remark 6.13. When X = ℵ0, B[X]2 is a UHF algebra (in fact CAR algebra) by Glimm's theorem [11, Theorem 1.13]. When X = ℵ1, B[X]2 is a unital AM algebra by Proposition 6.10 and Theorem 1.3 (1). In this case one can show that B[X]2 is not UHF in a similar (but much more complicated) way to the proof of Proposition 4.5 (2) (see [10]). Remark 6.14. As we pointed out in Remark 4.2, the examples in Section 4 of unital AM algebras which are not UHF are obtained as crossed products of UHF algebras by the group Z/2Z. The examples in this section of unital LM algebras which are not AM are obtained as cocycle crossed products (see Remark 6.9). However we do not know the following. Problem 6.15. Find an example of a unital LM algebra which is not AM such that it is obtained as a crossed product of a unital AM (or UHF) algebra by a discrete group. Remark 6.16. We can solve the non-unital version of this problem using the examples in this section. In fact, by [15, Corollary 3.7] the tensor product B[X]2 ⊗ K is obtained as an (ordinary) crossed product of A[X]2 ⊗ K by the group GX where K := K(cid:0)ℓ2(GX )(cid:1) is the non-unital AM algebra of all compact operators on the Hilbert space ℓ2(GX ). Thus for every cardinal κ > ℵ1, there exists an example of a non-unital LM algebra with character density κ which is not AM such that it is obtained as a crossed product of a non-unital AM algebra by a discrete group. Note that B[X]2 ⊗ K is not AM if B[X]2 is not AM because every corner of an AM algebra is AM. The same comments can be applied to LF and AF instead of LM and AM. 7. Representation density and character density The purpose of this section is to give an answer to the half of the question raised by Masamichi Takesaki when the second author gave a talk on this paper. We could not answer the other half (Problem 7.19). The proof uses NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 25 the construction (Proposition 7.12) that was given by Bruce Blackadar when the first author gave a talk. Both authors would like to thank Masamichi Takesaki and Bruce Blackadar. For a Hilbert space H, we also denote by χ(H) the smallest cardinality of a dense subset of H. Note that for an infinite-dimensional Hilbert space H and an infinite set X, we get χ(H) = X if and only if H is isomorphic to ℓ2(X). Definition 7.1. The representation density χr(A) of a C-algebra A is the smallest cardinal χ(H) where H is a Hilbert space on which A can be faithfully represented. Note that both the representation density χr and the character density χ (Definition 1.2) are monotonic in the sense that if A is a subalgebra of B then the density of B is not smaller than the density of A. Since these cardinal invariants of C-algebras were apparently not consid- ered previously, the reader will hopefully excuse us for starting this section by listing a few trivial statements. Lemma 7.2. For every C-algebra A we have that χ(A) ≥ sup(cid:8)X : X is a family of commuting projections in A(cid:9) χr(A) ≥ sup(cid:8)X : X is a family of nonzero orthogonal projections in A(cid:9). Proof. For the first part note that if p and q are distinct commuting projec- tions then kp − qk = 1. The second part is obvious. (cid:3) Lemma 7.3. For every infinite-dimensional Hilbert space H we have χ(cid:0)B(H)(cid:1) = B(H) = 2χ(H). Proof. Let us choose an infinite set X with X = χ(H), and identify H with ℓ2(X). For a subset Y ⊆ X, let pY ∈ B(H) be the projection onto the subspace ℓ2(Y ) ⊆ H. Then {pY }Y ⊆X is a family of commuting projections of size 2X. Thus we have χ(B(H)) ≥ 2X by Lemma 7.2. For x, y ∈ X, p{x}B(H)p{y} is one dimensional, and the map B(H) ∋ T 7→(cid:0)p{x}T p{y}(cid:1)x,y∈X ∈ Yx,y∈X(cid:0)p{x}B(H)p{y}(cid:1) ∼= Yx,y∈X C is injective. Hence we get χ(B(H)) ≤ B(H) ≤ CX×X = 2X. We are done. If K = 22ℵ0 with the product topology then C(K) ∼= N2ℵ0 C2 is an abelian C-algebra with character density 2ℵ0 and representation density ℵ0. The first claim follows by Lemma 2.20. The second claim follows from the fact that K is, being a product of 2ℵ0 separable spaces, separable by the Hewitt -- Marczewski -- Pondiczery Theorem (see e.g., [7, Corollary 2.3.16]). See also Corollary 7.7, Theorem 7.17 and Problem 7.19. Lemma 7.4. For every C-algebra A we have χr(A) ≤ χ(A) ≤ 2χr(A). (cid:3) 26 ILIJAS FARAH AND TAKESHI KATSURA Proof. Choose a subset X ⊆ A with X = χ(A). For each x ∈ X, there exists a cyclic representation πx : A → B(Hx) with kπx(x)k = kxk (see [2, Corollary II.6.4.9]). Since Hx has a cyclic vector for πx, we have χ(Hx) ≤ χ(A). Then the representation π := Mx∈X Hx(cid:17) = Xx∈X is faithful, and χ(cid:16)Mx∈X πx : A → B(cid:16)Mx∈X Hx(cid:17) χ(Hx) ≤ X × χ(A) = χ(A) Hence χr(A) ≤ χ(A). The second inequality χ(A) ≤ 2χr(A) follows from Lemma 7.3. Lemma 7.5. Let X0 ∋ x 7→ ξx ∈ H be a map from a set X0 to a Hilbert space H such that X0 > χ(H). Then for every ε > 0, there exists X1 ⊆ X0 with X1 > χ(H) such that kξx − ξyk < ε for every x, y ∈ X1. Proof. Choose a dense subset Y ⊆ H with Y = χ(H). For each x ∈ X0 there exists η(x) ∈ Y such that kξx−η(x)k < ε/2. Since X0 > χ(H) = Y , there exists η ∈ Y such that the set X1 := {x ∈ X0 : η(x) = η} ⊆ X0 satisfies X1 > χ(H). Then for every x, y ∈ X1, we get (cid:3) kξx − ξyk ≤ kξx − ηk + kξy − ηk < ε. (cid:3) Proposition 7.6. For a family {Ax}x∈X of nonabelian unital C-algebras, the representation density of the tensor product A = Nx∈X Ax is at least X. Proof. Assume the contrary and fix a faithful representation π : A → B(H) for a Hilbert space H with X > χ(H). Note that this assumption implies that X is uncountable. For each x ∈ X, fix ax and bx in the unit ball of Ax such that axbx 6= bxax. Since π is faithful, we can choose a vector ξx ∈ H such that π(axbx − bxax)ξx 6= 0. Since X is uncountable, there exist δ > 0 and a subset X0 ⊆ X with X0 > χ(H) such that for all x ∈ X0 we have kπ(axbx − bxax)ξxk ≥ δ. In this proof, we write a ≈ε b if ka − bk < ε. Since Set ε = δ/4 > 0. X0 > χ(H), we can apply Lemma 7.5 to {ξx}x∈X0 and ε > 0 to get X1 ⊆ X0 with X1 > χ(H) such that ξx ≈ε ξy for every x, y ∈ X1. By applying Lemma 7.5 three more times to {π(ax)ξx}x∈X1 and so on, we get X4 ⊆ X1 with X4 > χ(H) such that π(ax)ξx ≈ε π(ay)ξy, π(bx)ξx ≈ε π(by)ξy, π(bxax)ξx ≈ε π(byay)ξy NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 27 for every x, y ∈ X4. Since X4 > χ(H) ≥ ℵ0, we can take two distinct x, y ∈ X4. Then we have π(axbx)ξx = π(ax)π(bx)ξx ≈ε π(ax)π(by)ξy = π(axby)ξy ≈ε π(axby)ξx = π(bxax)ξx ≈ε π(byay)ξy = π(by)π(ay)ξy ≈ε π(by)π(ax)ξx = π(byax)ξx because ax ∈ Ax ⊆ A and by ∈ Ay ⊆ A commute. Thus we get (cid:13)(cid:13)π(axbx − bxax)ξx(cid:13)(cid:13) < 4ε = δ, which is a contradiction. This completes the proof. (cid:3) Corollary 7.7. If A is a UHF algebra then χ(A) = χr(A). (cid:3) With the possible exception of the algebras AX,Y as defined in §3, each example of an AM, or even LM, algebra given so far has a UHF subalge- bra with the same character density. Since the algebras AX,Y are tensor products of separable algebras, Proposition 7.6 implies that for each AM or LM algebra A so far defined in this paper we have χ(A) = χr(A). We are going to show that χ(A) can be any cardinality between χr(A) and 2χr(A) for unital AM algebras A. Let X be an infinite set. As in Section 4, let Ax be a C-algebra generated by two self-adjoint unitaries vx, wx with vxwx = −wxvx for each x ∈ X, and let AX := Nx∈X Ax. By Lemma 4.1, Ax ∼= M2(C) for each x ∈ X and hence AX ∼=NX M2(C) is a UHF algebra. For each Y ⊆ X, we set AY :=Ox∈Y Ax ⊆ AX . We are going to use the GNS representation of AX associated with the unique tracial state of AX. For the reader's convenience we explain what it is. For each finite subset F ⊆ X, there exists a unique linear functional τF : AF → C satisfying the trace condition τF (ab) = τF (ba) for a, b ∈ AF and the normalized condition τF (1) = 1. If F = n, then we have τF = 2−nTr where Tr is the usual trace of AF ∼= M2n(C). It is easy to see that τF is positive and faithful, that is, τF (a∗a) > 0 for all a ∈ AF \ {0}. Let Afin X := SF ⊆X AF ⊆ AX where F runs all finite subsets of X. By the uniqueness of the tracial state τF , we get τF ′ ↾AF = τF for two finite subsets F ⊆ F ′ ⊆ X. Thus we get a linear map τ : Afin X → C such that τ ↾AF = τF for every finite subset F ⊆ X. Although we do not need it, we would like to remark that τ can be extended to the unique tracial state of AX (cf. [5, Lemma I.9.5]). We define an inner product on Afin X ∋ (a, b) 7→ τ (ab∗) ∈ C. Then the completion HX of Afin X with respect to the norm coming from the inner product defined as above becomes a Hilbert space. The embedding from Afin X to HX is denoted by Afin X ∋ a 7→ a ∈ HX. The image of this embedding is dense in HX. For each finite subset F ⊆ X and each a ∈ AF , it is easy to see that the map b 7→ bab extends to a bounded operator on HX. Thus we get a ∗- homomorphism πF : AF → B(HX) such that πF (a)(bb) = bab for a ∈ AF and X × Afin X by Afin 28 ILIJAS FARAH AND TAKESHI KATSURA X . We have πF ′ ↾AF = πF for two finite subsets F ⊆ F ′ ⊆ X. Since b ∈ Afin the family {π{x}[A{x}]}x∈X mutually commutes, we get a representation π : AX → B(HX) such that π ↾AF = πF for every finite subset F ⊆ X. This representation is called the GNS representation associated with τ . Since π(a)(ba∗) = daa∗ 6= 0 for all F ⊆ X and all a ∈ AF \ {0}, π is injective. In order to simplify the notation we identify AX with the subalgebra π[AX ] of B(HX). Lemma 7.8. We have χ(HX) = X. Proof. Since the union of finite-dimensional subspaces {a ∈ HX a ∈ AF} for finite subsets F ⊆ X is dense in HX, we have χ(HX ) ≤ X. For distinct x, y ∈ X, we have τ (uxuy) = 0 because τ (uxuy) = τ (wx(wxuxuy)) = τ ((wxuxuy)wx) = τ (wxux(wxuy)) = τ (wx(−wxux)uy) = −τ (uxuy). Hence we get kcux −cuyk2 = τ ((ux − uy)(ux − uy)) = τ (2 − 2uxuy) = 2 for all x, y ∈ X with x 6= y. This shows that χ(HX) ≥ X. Thus we get χ(HX) = X. (cid:3) We can consider the power-set P(X) of a set X as an abelian group with respect to the symmetric difference. This group is naturally isomorphic to the direct product of X copies of Z/2Z. For g ∈ P(X) consider an automorphism of AX defined by αg =Ox∈g Ad vx. X . Then α defines an action of P(X) on AX . For each g ∈ P(X), the automor- phism αg preserves the subalgebra AF ⊆ AX and satisfies τF ◦ αg = τF for every finite subset F ⊆ X. Hence we get an element ug ∈ B(HX) such that ug(bb) = αgb(b) for b ∈ Afin Lemma 7.9. The elements {ug}g∈P(X) ⊆ B(HX) are self-adjoint unitaries satisfying ugaug = αg(a) and uguh = ugh for a ∈ AX ⊆ B(HX) and g, h ∈ P(X). Proof. Take g ∈ P(X). Since αg preserves τ , the element u∗ isfies u∗ because α−1 αg(aαg(b)) = αg(a)b and αg(αh(b)) = αgh(b) for b ∈ AX . Definition 7.10. For an infinite set X and a subgroup Γ ⊆ P(X) we define g ∈ B(HX) sat- X . Hence ug is a unitary. This is self-adjoint g = αg. The latter two equalities follow from the equations (cid:3) g(bb) = α−1 g b(b) for b ∈ Afin BX,Γ := C ∗(AX ∪ {ug}g∈Γ) ⊆ B(HX). NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 29 Remark 7.11. One can show that BX,Γ is isomorphic to the crossed product AX ⋊α Γ. In particular BX in Section 4 is isomorphic to BX,Γ for Γ = {∅, X} ∼= Z/2Z. Proposition 7.12. The C-algebra BX,Γ satisfies χ(BX,Γ) = X + Γ and χr(BX,Γ) = X. Proof. We have χ(AX ) = X by Lemma 2.20. On the other hand, we have χ(C ∗({ug}g∈Γ)) ≥ Γ by Lemma 7.2 because {(ug + 1)/2}g∈Γ is a family of commuting projections. Since BX,Γ is generated by AX and {ug}g∈Γ, we get X + Γ = max{X,Γ} ≤ χ(BX,Γ) ≤ X + Γ This shows χ(BX,Γ) = X +Γ. Since BX,Γ ⊆ B(HX), we have χr(BX,Γ) ≤ χ(HX) = X by Lemma 7.8. We also have χr(BX,Γ) ≥ χr(AX ) = X by Corollary 7.7. Hence we get χr(BX,Γ) = X. (cid:3) Proposition 7.13. The unital C-algebra BX,Γ is AM if every finite subset of Γ is included in a subgroup generated by g1, g2, . . . , gn ∈ Γ which are infinite and mutually disjoint. Proof. Take mutually disjoint infinite elements g1, g2, . . . , gn ∈ Γ. Take a finite subset F of X and choose xi ∈ gi \ F for i = 1, 2, . . . , n. Let Λ be the set of all such data λ = ({gi}n Dλ := C ∗(cid:0){ugi}n i=1, F,{xi}n i=1 ∪ AF ∪ {wxi}n i=1(cid:1) ⊆ BX,Γ. By the assumption of Γ, the family {Dλ}λ∈Λ of subalgebras is directed and its union is dense in B[X]2. We are going to show Dλ ∼= M2m+n (C) for λ = ({gi}n i=1) as above where m = F. This implies that B[X]2 is AM, and hence completes the proof. For i ∈ {1, 2, . . . , n} define i=1, F,{xi}n i=1), and define u′ i = ugi Yx∈F ∩gi vx ∈ Dλ. Since Ad ugi ↾AF = Ad(cid:16) Yx∈F ∩gi vx(cid:17) ↾AF , u′ i is a self-adjoint unitary and commutes with the subalgebra AF . It is easy to see that the family {u′ i}n i=1 mutually commutes. Since xi ∈ gi \ F and gi is disjoint from gj for j 6= i, we have that wxi commutes with AF and {u′ i anti-commute because so do wxi and ugi. Therefore C ∗(u′ i, wxi) is isomorphic to M2(C) for i ∈ {1, 2, . . . , n} by Lemma 4.1, and the family j, wxj}j6=i. Finally wxi and u′ {C ∗(u′ i, wxi)}n i=1 ∪ {AF} mutually commutes. Since Dλ is generated by these mutually commuting subalgebras, we get Dλ =(cid:16) nOi=1 C ∗(u′ i, wxi)(cid:17) ⊗ AF ∼= On+F M2(C) ∼= M2n+m(C), 30 ILIJAS FARAH AND TAKESHI KATSURA (cid:3) as required. Remark 7.14. For finite g ∈ P(X), we have αg = Ad(cid:0)Qx∈g vx(cid:1). From this fact, one can show that BX,Γ is not AM if Γ contains a finite nonempty element g (one can also show that BX,Γ is always AF). Thus in order for BX,Γ to be AM it is necessary that every g ∈ Γ\{∅} is infinite. One can show that this is also sufficient although its proof becomes significantly complicated compared with Proposition 7.13. We shall not need such generality for proving Theorem 7.17. Remark 7.15. One can show that BX,Γ is not UHF when X ≥ ℵ1 and Γ 6= {∅} in a similar way to the proof of Proposition 4.5 (2). We omit the proof because we do not need this (see the proof of Theorem 7.17 for some special cases). One can also show that ZBX,Γ(AX ) = C1 holds when every g ∈ Γ\{∅} is infinite (even in the case χ(AX ) < χ(BX,Γ)). This shows that a generalization of question [6, Problem 8.3] for nonseparable AM algebras has a very strong negative answer (see Corollary 4.6). The authors would like to thank Bruce Blackadar for pointing out the phenomenon ZBX,Γ(AX ) = C1. This strong phenomenon does not occur for UHF algebras because we can show χ(ZB(A)) = χ(B) for a subalgebra A of a UHF algebra B with χ(A) < χ(B), and hence in this case ZB(A) is huge. Lemma 7.16. For every cardinal κ with X ≤ κ ≤ 2X, there exists a subgroup Γ ⊆ P(X) with Γ = κ such that every finite subset of Γ is included in a subgroup generated by g1, g2, . . . , gn ∈ Γ which are infinite and mutually disjoint. Proof. Take a subset Y ⊆ P(X) with Y = κ. Let Γ0 be the Boolean subal- gebra of P(X) generated by Y , that is the smallest subset of P(X) containing Y and closed under taking unions, intersections and complements. Then Γ0 is a subgroup of P(X) with Γ0 = κ. Choose a bijection ι : X × N → X and define an injective homomorphism ϕ : P(X) ∋ g 7→ ι[g × N] ∈ P(X). Let Γ := ϕ[Γ0] ⊆ P(X). Then every finite subset of Γ is included in a finite Boolean subalgebra of Γ. If g1, g2, . . . , gn ∈ Γ are the atoms of this subalgebra then they clearly satisfy the requirements. (cid:3) Theorem 7.17. For every pair of infinite cardinals κ and ν with κ ≥ ℵ1 and ν ≤ κ ≤ 2ν, there exists a unital AM algebra of character density κ and representation density ν which is not UHF. Proof. For κ = ν ≥ ℵ1, the example BX in Proposition 4.5 for X = κ is a unital AM algebra of character density κ and representation density ν which is not UHF. Suppose ν < κ ≤ 2ν . Take a set X with X = ν. By Lemma 7.16, there exists a subgroup Γ ⊆ P(X) with Γ = κ satisfying the assumption of Proposition 7.13. Then BX,Γ is a unital AM algebra of character density κ and representation density ν by Proposition 7.12 and Proposition 7.13. This is not UHF by Corollary 7.7. (cid:3) NONSEPARABLE UHF ALGEBRAS I: DIXMIER'S PROBLEM 31 From Theorem 7.17 we have the following. Corollary 7.18. There is a unital AM algebra faithfully represented on a separable Hilbert space that is not a UHF algebra. (cid:3) This corollary answers a half of the question raised by Masamichi Take- saki. The following is the other half which we could not answer. Problem 7.19. Is there an LM algebra faithfully represented on a separable Hilbert space which is not AM? Since χ(cid:0)B(ℓ2(N))(cid:1) = 2ℵ0, by Theorem 1.3 (1) there is no such a C- algebra if we assume the continuum hypothesis 2ℵ0 = ℵ1. We do not know what happens if we do not assume the continuum hypothesis. Acknowledgment. Many of the results presented in this paper were proved while both authors were visiting the Fields Institute in Fall 2007 and in early 2008. I.F. would like to thank colleagues from his department, in particular Juris Stepr¯ans and Man Wah Wong, for making his stay at the Fields Insti- tute possible. Both authors would like to thank George Elliott, Toshihiko Masuda, Narutaka Ozawa, N. Christopher Phillips, Juris Stepr¯ans, Reiji Tomatsu and Andrew Toms for illuminating conversations. We would also like to thank the anonymous referee for several useful remarks. References [1] I. Ben Yaacov, A. Berenstein, C.W. Henson, and A. Usvyatsov, Model theory for metric structures, Model Theory with Applications to Algebra and Analysis, Vol. II (Z. Chatzidakis et al., eds.), Lecture Notes series of the London Math. Society., no. 350, Cambridge University Press, 2008, pp. 315 -- 427. [2] B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122, Springer-Verlag, Berlin, 2006, Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non-commutative Geometry, III. [3] O. Bratteli, Inductive limits of finite dimensional C ∗-algebras, Trans. Amer. Math. Soc. 171 (1972), 195 -- 234. [4] N. Brown and N. Ozawa, C ∗-algebras and finite-dimensional approximations, Grad- uate Studies in Mathematics, vol. 88, Amer. Math. Soc., Providence, RI, 2008. [5] K.R. Davidson, C ∗-algebras by example, Fields Institute Monographs, vol. 6, Amer. Math. Soc., Providence, RI, 1996. [6] J. Dixmier, On some C ∗-algebras considered by Glimm, J. Functional Analysis 1 (1967), 182 -- 203. [7] R. Engelking, General topology, Heldermann, Berlin, 1989. [8] I. Farah, Graphs and CCR algebras, Indiana Univ. Math. Journal (to appear). [9] I. Farah, B. Hart, and D. Sherman, Model theory of operator algebras II: Model theory, preprint, 2010. [10] I. Farah and T. Katsura, Nonseparable UHF algebras II: Classification, in preparation, 2010. [11] J.G. Glimm, On a certain class of operator algebras, Trans. Amer. Math. Soc. 95 (1960), 318 -- 340. [12] R. Haydon and S. Wassermann, A commutation result for tensor products of C*- algebras, Bull. London Math. Soc. 5 (1973), 283 -- 28. 32 ILIJAS FARAH AND TAKESHI KATSURA [13] X. Jiang and H. Su, On a simple unital projectionless C ∗-algebra, Amer. J. Math. 121 (1999), no. 2, 359 -- 413. [14] T. Katsura, Non-separable UHF algebras, RIMS proceedings (2009). [15] J.A. Packer and I. Raeburn, Twisted crossed products of C ∗-algebras, Math. Proc. Cambridge Philos. Soc. 106 (1989), no. 2, 293 -- 311. Department of Mathematics and Statistics, York University, 4700 Keele Street, North York, Ontario, Canada, M3J 1P3, and Matematicki Institut, Kneza Mihaila 34, Belgrade, Serbia URL: http://www.math.yorku.ca/∼ifarah E-mail address: [email protected] Department of Mathematics, Faculty of Science and Technology, Keio Uni- versity, 3-14-1 Hiyoshi, Kouhoku-ku, Yokohama, JAPAN, 223-8522 E-mail address: [email protected]
1608.03515	1	1608	2016-08-11T16:06:35	Eta-diagonal distributions and infinite divisibility for R-diagonals	[ "math.OA", "math.CO", "math.PR" ]	The class of R-diagonal -distributions is fairly well understood in free probability. In this class, we consider the concept of infinite divisibility with respect to the operation $\boxplus$ of free additive convolution. We exploit the relation between free probability and the parallel (and simpler) world of Boolean probability. It is natural to introduce the concept of an eta-diagonal distribution that is the Boolean counterpart of an R-diagonal distribution. We establish a number of properties of eta-diagonal distributions, then we examine the canonical bijection relating eta-diagonal distributions to infinitely divisible R-diagonal ones. The overall result is a parametrization of an arbitrary $\boxplus$-infinitely divisible R-diagonal distribution that can arise in a C-probability space, by a pair of compactly supported Borel probability measures on $[ 0, \infty )$. Among the applications of this parametrization, we prove that the set of $\boxplus$-infinitely divisible R-diagonal distributions is closed under the operation $\boxtimes$ of free multiplicative convolution.	math.OA	math	ETA-DIAGONAL DISTRIBUTIONS AND INFINITE DIVISIBILITY FOR R-DIAGONALS HARI BERCOVICI, ALEXANDRU NICA, MICHAEL NOYES, AND KAMIL SZPOJANKOWSKI Abstract. The class of R-diagonal ∗-distributions is fairly well understood in free prob- ability. In this class, we consider the concept of infinite divisibility with respect to the operation ⊞ of free additive convolution. We exploit the relation between free proba- bility and the parallel (and simpler) world of Boolean probability. It is natural to in- troduce the concept of an η-diagonal distribution that is the Boolean counterpart of an R-diagonal distribution. We establish a number of properties of η-diagonal distributions, then we examine the canonical bijection relating η-diagonal distributions to infinitely di- visible R-diagonal ones. The overall result is a parametrization of an arbitrary ⊞-infinitely divisible R-diagonal distribution that can arise in a C ∗-probability space, by a pair of com- pactly supported Borel probability measures on [0, ∞). Among the applications of this parametrization, we prove that the set of ⊞-infinitely divisible R-diagonal distributions is closed under the operation ⊠ of free multiplicative convolution. 1. Introduction Free additive convolution ⊞ is a binary operation on the set P of Borel probability measures on R, reflecting the addition operation for free selfadjoint elements in a noncom- mutative probability space. The properties of this operation parallel in many respects the ones of the usual convolution on P, for instance in the treatment of infinite divisibility. One way to approach ⊞-infinite divisibility is to use a bijection constructed in [4] which re- lates free independence to another form of noncommutative independence, namely Boolean independence. In this paper we focus on probability measures with compact support, so we view this bijection as a map B : Pc → P (inf−div) , where Pc is the set of probability measures with compact support on R, while P (inf−div) consists of those measures µ ∈ Pc which are ⊞-infinitely divisible, that is, have the property that for every n ∈ N, there exists µn ∈ Pc satisfying c c (1.1) µn ⊞ · · · ⊞ µn = µ. n {z } The bijection B connects the fundamental transforms of free and Boolean probability, the R-transform and respectively the η-series. For µ ∈ Pc, both of these transforms Rµ(z) and ηµ(z) are convergent power series. The bijection B is described by the equation (1.2) RB(µ) = ηµ, µ ∈ Pc. More precisely, for every µ ∈ Pc there exists a uniquely determined measure ν ∈ P (inf−div) such that Rν = ηµ, and one defines B(µ) := ν. c At the level of compactly supported distributions, the bijection B is precisely the para- metrization of ⊞-infinitely divisible distributions provided in [14]. This was extended in [4] HB: supported in part by a grant from the National Science Foundation of the USA. AN: research supported by a Discovery Grant from NSERC, Canada. 1 2 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI to the space P of all Borel probability measures on R. In a different direction, the bijection B was extended in [2] to the space of joint distributions for k-tuples of selfadjoint elements in a C ∗-probability space. Our goal in this paper is to use a multivariate version of the bijection B in order to study ⊞-infinitely divisibile R-diagonal distributions, a significant class of ∗-distributions considered in free probability. To explain our results, we introduce some notation. We let Dc(1, ∗) stand for the collec- tion of all ∗-distributions of (generally, not selfadjoint) elements in a (generally not tracial) C ∗-probability space. There is a natural operation ⊞ on Dc(1, ∗) which corresponds to the addition a + b of two variables a, b in the same space such that {a, a∗} is free from {b, b∗}. Infinite divisibility in Dc(1, ∗) is defined as in (1.1), and we denote by Dc(1, ∗)(inf−div) the collection of ⊞-infinitely divisible elements of Dc(1, ∗). The notions of R-transform and η-series also have natural extensions to the context of ∗-distributions. The results of [2], specialized to two selfadjoint variables, can be applied to Dc(1, ∗) after a simple change of coordinates. There is again a bijection B(1,∗) : Dc(1, ∗) → Dc(1, ∗)(inf−div) defined by the requirement that (1.3) RB(1,∗)(µ) = ηµ, µ ∈ Dc(1, ∗). This is analogous to the condition (1.2) satisfied by the original bijection B, but proving the existence of B(1,∗) is more than a trivial extension of the proof for B, and requires a mixture of combinatorial and analytic methods. We turn now to R-diagonal ∗-distributions, which can be succinctly described as the distributions in Dc(1, ∗) that are invariant under multiplication by a free Haar unitary (see [11, Theorem 15.10, p. 244]). For our purposes, it is more useful to consider the original definition [10] of R-diagonal distributions which asks that the R-transform of the distribution be in some sense 'diagonal' [11, Definition 15.3, p. 241]. From this point of view, it is clear how to define the Boolean counterpart of R-diagonality: we simply say that a ∗-distribution is η-diagonal if its η-series is diagonal. The map B (1,∗) defined by (1.3) will then give a bijection between the set of all η-diagonal distributions in Dc(1, ∗) and the set of R-diagonal distributions in Dc(1, ∗) which are ⊞-infinitely divisible. The above discussion shows that there is some interest in studying η-diagonal distribu- tions. In this paper we point out a few general algebraic and combinatorial properties of such a distribution µ, which actually hold for µ in a larger, purely algebraic space Dalg(1, ∗). The property of a distribution µ ∈ Dalg(1, ∗) of being η-diagonal has an elegant description phrased directly in terms of the ∗-moments of µ. This result (Theorem 2.8) is reminiscent of (but simpler than) the description [9, Theorem 1.2.1] of R-diagonal variables in terms of their ∗-moments. The η-diagonal distributions also have other algebraic and combinatorial properties that are analogous to known properties of R-diagonal distributions. In particu- lar, if a is an η-diagonal element in a ∗-probability space (A, ϕ) (which means, by definition, that a has η-diagonal ∗-distribution with respect to ϕ) then it follows that aa∗ and a∗a are Boolean independent elements of A, and that the coefficients of the η-series of aa∗ and a∗a are read from the so-called determining sequences for the ∗-distribution of a. For details on the terms used above and for a discussion of why this is indeed analogous to known facts about R-diagonals, see Remark 3.4 below. In the case in which the η-diagonal distribution µ is in Dc(1, ∗), we point out a natural parametrization for µ, given by a pair of compactly supported Borel probability measures on [0, ∞). That is, we establish a canonical bijection (1.4) {µ ∈ Dc(1, ∗) : µ is η-diagonal} ∋ µ ↔ (σ1, σ2) ∈ P + c × P + c , ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 3 := {σ ∈ Pc : σ(cid:0) [0, ∞)(cid:1) = 1}. Without going into details, we mention that all where P + c the ∗-distributions appearing in this paper are defined as linear functionals on the algebra ChZ, Z ∗i of complex polynomials in the non-commuting indeterminates Z and Z ∗, and that the correspondence µ ↔ (σ1, σ2) from (1.4) amounts to the equalities (1.5) µ((ZZ ∗)n) =Z ∞ 0 tn dσ1(t) and µ((Z ∗Z)n) =Z ∞ 0 tn dσ2(t), n ∈ N. In other words, the probability measures σ1 and σ2 which parametrize µ in (1.4) are simply the distributions of ZZ ∗ and of Z ∗Z with respect to the functional µ. The relevant point here is that for any given σ1, σ2 ∈ P + c there exists a unique η-diagonal distribution µ ∈ Dc(1, ∗) such that (1.5) holds. When the bijection B(1,∗) is applied to {µ ∈ Dc(1, ∗) : µ is η-diagonal} in (1.4), we obtain a bijection (1.6) nν ∈ Dc(1, ∗) : ν is R-diagonal and ⊞-infinitely divisible o ∋ ν ↔ (σ1, σ2) ∈ P + c × P + c . Thus, we have a parametrization of a general ⊞-infinitely divisible R-diagonal distribution by a pair of probability measures from P + c . Analogously to (1.5), it is possible to write explicitly the relation connecting σ1, σ2 to the distributions of the elements ZZ ∗ and Z ∗Z in the noncommutative probability space (ChZ, Z ∗i, ν). Theorem 6.4 below realizes this parametrization by providing precise formulas for the R-transforms of ZZ ∗ and of Z ∗Z. An important subclass of R-diagonal distributions are those which satisfy the KMS con- dition for some parameter t ∈ (0, ∞). This is a generalization of the trace condition, where the latter corresponds to the special case t = 1 (see the review in Section 3 below). For an R-diagonal distribution ν which satisfies KMS with parameter t, one can process further the result of Theorem 6.4 in order to obtain explicit formulas for the distributions of ZZ ∗ and of Z ∗Z, in terms of the probability measures σ1 and σ2 which parametrize ν. These formulas invoke some commonly used elements of free harmonic analysis on P + c , and are given in Proposition 6.8. As an application of the parametrization from (1.6), we prove that the set of ⊞-infinitely divisible R-diagonal ∗-distributions is closed under the operation ⊠ of multiplicative con- volution. In addition to the present introduction, the paper contains 6 sections. Section 2 intro- duces η-diagonal ∗-distributions and discusses some of their algebraic properties. Section 3 is devoted to a review of R-diagonal ∗-distributions, with emphasis on facts that are needed in the present paper. In Section 4 we verify that the bijection B(1,∗) does indeed work on Dc(1, ∗) in the way described in (1.3). Section 5 presents the operator model for η-diagonals that was announced in (1.4). Section 6 contains our results concerning the parametriza- tion of infinitely divisible R-diagonal distributions, and a discussion of the KMS example. Finally, Section 7 discusses the application to free multiplicative convolution. 2. η-series and η-diagonal ∗-distributions Notation 2.1. (1) We denote by W + the setF∞ n=1{1, ∗}n consisting of all non-empty words over the two-letter alphabet {1, ∗}. This is a semigroup (without unit) under the natural operation of concatenation. We denote by w the number of letters in a word w ∈ W +. 4 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI (2) The algebra of complex polynomials in two non-commuting variables Z and Z ∗ is denoted, as usual, by ChZ, Z ∗i. For every word w = (ℓ1, . . . , ℓn) ∈ W + we write Z w = Z ℓ1 · · · Z ℓn ∈ ChZ, Z ∗i. The set {1} ∪ {Z w : w ∈ W +} is a basis of ChZ, Z ∗i as a complex vector space. (3) An algebraic ∗-distribution is a linear functional µ : ChZ, Z ∗i → C such that µ(1) = 1. (At this stage we do not require µ to have any additional properties.) The values of µ on monomials Z w (with w ∈ W +) will be referred to as ∗-moments of µ. (4) The collection of all algebraic ∗-distributions from (3) is denoted Dalg(1, ∗). Notation 2.2. (Series and their coefficients.) (1) The algebra of formal power series in two non-commuting indeterminates z and z∗ is denoted, as usual, by Chhz, z∗ii. The collection C0hhz, z∗ii ⊂ Chhz, z∗ii of power series with vanishing constant coefficient is a two-sided ideal in Chhz, z∗ii. An arbitrary element f ∈ C0hhz, z∗ii is of the form (2.1) f (z, z∗) = ∞Xn=1 Xℓ1,...,ℓn∈{1,∗} α(ℓ1,...,ℓn)zℓ1 · · · zℓn = Xw∈W + αwzw, where the coefficients αw are complex numbers, and for w = (ℓ1, . . . , ℓn) ∈ W + we use the notation zw = zℓ1 · · · zℓn. (2) Given w ∈ W +, we denote by Cfw : C0hhz, z∗ii → C the linear functional which extracts the coefficient of zw from a series f ∈ C0hhz, z∗ii. That is, if f is given by (2.1), we have Cfw(f ) = αw, w ∈ W +. (3) Given a positive integer n, a word w = (ℓ1, . . . , ℓn) ∈ W +, and a partition π of {1, . . . , n}, we define a functional (non-linear unless π consists of only one block) Cfw;π : C0hhz, z∗ii → C, as follows. For every block B = {b1, . . . , bm} of π, where 1 ≤ b1 < · · · < bm ≤ n, we set wB = (ℓ1, . . . , ℓn)B := (ℓb1, . . . , ℓbm) ∈ {1, ∗}m. Then we define (2.2) Cfw;π(f ) := YB∈π CfwB(f ), f ∈ C0hhz, z∗ii. [Suppose, for instance, that n = 5, π = {{1, 4, 5}, {2, 3}}, and w = (ℓ1, . . . , ℓ5). Then Cfw;π(f ) = Cf(ℓ1,ℓ4,ℓ5)(f ) · Cf(ℓ2,ℓ3)(f ), f ∈ C0hhz, z∗ii.] Definition and Remark 2.3. (Moment series, η-series.) Fix µ ∈ Dalg(1, ∗). (1) The moment series of µ is defined as Mµ :=Pw∈W + µ(Z w)zw ∈ C0hhz, z∗ii. (2) The η-series of µ is defined as (2.3) ηµ := Mµ(1 + Mµ)−1 = (1 + Mµ)−1Mµ ∈ C0hhz, z∗ii, where all the algebraic operations are performed in the algebra Chhz, z∗ii. (3) It is immediate from (2.3) that the series Mµ can be retrieved from ηµ by the formula (2.4) Mµ = ηµ(1 − ηµ)−1 = (1 − ηµ)−1ηµ. (4) The right-hand side of (2.4) can be written as a geometric series P∞ converges in the sense that the series P∞ µ (which µ) contains only finitely many non-zero n=1 Cfw(ηn n=1 ηn ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 5 terms for every w ∈ W +). This leads to an explicit formula for the coefficients of Mµ in terms of those of ηµ, namely Cfw(Mµ) = Xπ∈Int(n) Cfw;π(ηµ), w ∈ W + with w = n. (2.5) (2.6) Here Int(n) denotes the set of interval partitions of {1, . . . , n}, that is, partitions which have the property that every block B of π is of the form {a, a + 1, . . . , b} for some a ≤ b in {1, . . . , n}. An analogous argument converts (2.3) into the formula Cfw(ηµ) = Xπ∈Int(n) (−1)1+πCfw;π(Mµ), w ∈ W + with w = n, where π denotes the number of blocks of the partition π. Remark 2.4. It is clear that the map Dalg(1, ∗) ∋ µ 7→ Mµ ∈ C0hhz, z∗ii is bijective. Equations (2.3) and (2.4) show that the map Dalg(1, ∗) ∋ µ 7→ ηµ ∈ C0hhz, z∗ii is a bijection as well. In other words, we can define a distribution µ ∈ Dalg(1, ∗) by specifying its η-series. Definition 2.5. A word w ∈ W + is said to be alternating when it is of the form ) = (∗, 1)m, ) = (1, ∗)m or w = (∗, 1, ∗, 1, . . . , ∗, 1 w = (1, ∗, 1, ∗, . . . , 1, ∗ 2m {z } 2m {z } for some positive integer m. In the first case w is said to be of type (1, ∗), and in the second case w is said to be of type (∗, 1). In these formulas, powers are taken relative to concatenation. Note in particular that alternating words have positive, even length. Definition 2.6. (1) A distribution µ ∈ Dalg(1, ∗) is said to be η-diagonal if Cfw(ηµ) = 0 for every word w ∈ W + which is not alternating. (2) If µ is η-diagonal, its η-series is thus of the form ηµ(z, z∗) = αn(zz∗)n + βn(z∗z)n, ∞Xn=1 ∞Xn=1 with αn, βn ∈ C for n ∈ N. The sequences (αn)∞ ing sequences of the η-diagonal distribution µ. n=1 and (βn)∞ n=1 will be called the determin- The main goal of the present section is to reveal an equivalent characterization of η- diagonal distributions, which is phrased directly in terms of ∗-moments. For this purpose, we require one more concept related to words in W +. Definition and Remark 2.7. (1) A word w = (ℓ1, . . . , ℓn) ∈ W + is said to be mixed- alternating if n = 2m is even and if the letters of w are such that ℓ2k−1 6= ℓ2k, for k = 1, 2, . . . , m. Equivalently, w is mixed-alternating when it belongs to the concatenation subsemigroup of W + generated by the words (1, ∗) and (∗, 1). (2) By grouping factors, it is easily seen that every mixed-alternating word w can be written in a unique way as a concatenation of alternating words such that consecutive words are of different types. Indeed, if we write such a word as (2.7) w = w1w2 · · · wp where p ≥ 1, each wi is alternating, and wi is not of the same type as wi+1, i = 1, 2, . . . , p−1, then the boundaries between the words w1, . . . , wp can be retrieved at the places where w has two consecutive identical letters. [For example, w = (1, ∗, 1, ∗, 1, ∗, ∗, 1, ∗, 1, 1, ∗, ∗, 1, ∗, 1) is mixed-alternating and its canoni- cal factorization (2.7) is w1w2w3w4 with w1 = (1, ∗)3, w2 = (∗, 1)2, w3 = (1, ∗), w4 = (∗, 1)2.] 6 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Theorem 2.8. For every distribution µ ∈ Dalg(1, ∗), the statements (a) and (b) are equiv- alent. (a) µ is η-diagonal. (b) µ satisfies the following conditions1: (ηDM1) Whenever w ∈ W + is not mixed-alternating, it follows that µ(Z w) = 0. (ηDM2) Whenever w = w1 · · · wp ∈ W + is mixed-alternating and factored as in (2.7), it follows that µ(Z w) = µ(Z w1) · · · µ(Z wp). The proof of the implication (b) ⇒ (a) in the above theorem requires two auxilliary results. Lemma 2.9. Suppose that a distribution µ ∈ Dalg(1, ∗) satisfies the condition (ηDM1). Then Cfw(ηµ) = 0 for every word w ∈ W + that is not mixed-alternating. Proof. Suppose that w is not mixed alternating and w = n. We prove that Cf w(ηµ) = 0 by showing that each term in the right-hand side of (2.6) vanishes. Indeed, let π = {J1, . . . , Jm} ∈ Int(n) be a partition where the intervals J1, . . . , Jm are listed in increasing order. Observe that w = (wJ1) · · · (wJm) (concatenation product). Since w is not mixed- alternating, there must exist an index 1 ≤ k ≤ m such that wJk is not mixed-alternating. For this k, condition (ηDM1) yields CfwJk(Mµ) = 0. Therefore the term indexed by π in (2.6) vanishes as well because CfwJk(Mµ) is one of its factors. The lemma follows. (cid:3) Lemma 2.10. Let µ, ν ∈ Dalg(1, ∗) be such that (1) Both µ and ν satisfy conditions (ηDM1) and (ηDM2), and (2) Cfw(ηµ) = Cfw(ην ) for every alternating word w ∈ W +. Then µ = ν. In fact, it suffices to prove this equality when w is alternating. Proof. If w ∈ W + is not mixed-alternating then µ(Z w) = ν(Z w) = 0 because µ and ν satisfy (ηDM1). Thus it suffices to verify that µ(Z w) = ν(Z w) for mixed-alternating words w. Indeed, suppose for the moment that the equality has been proved for alternating words and let w ∈ W + be a mixed-alternating word. Consider the canonical factorization w = w1 · · · wp indicated in (2.7). We have µ(Z w) = µ(Z w1) · · · µ(Z wp) (by (ηDM2) for µ) = ν(Z w1) · · · ν(Z wp) (by assumption on alternating moments) = ν(Z w) (by (ηDM2) for ν). We conclude the proof by showing that µ(Z w) = ν(Z w) for every alternating word w. By symmetry, it suffices to verify that µ((Z ∗Z)m) = ν((Z ∗Z)m) for every m ∈ N Fix m and write µ((Z ∗Z)m) and ν((Z ∗Z)m) as sums indexed by Int(2m), in the way indicated in (2.5). We show that for every π ∈ Int(2m), the terms indexed by π in the two sums (for µ and for ν) are equal to each other. If π has a block B of odd cardinality, then the terms we are looking at are both equal to 0 because they include the factors CfwB(ηµ) and respectively CfwB(ην ), and these factors are zero by Lemma 2.9. If all the blocks of π are even, we write π = {J1, . . . , Jk} where the intervals J1, . . . , Jk are listed in increasing order, and where J1 = 2d1, . . . , Jk = 2dk for some d1, . . . , dk ∈ N. The terms indexed by π in 1The acronym ηDM is meant to suggest η-Diagonality-in-Moments. ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 7 the two sums we consider are then (2.8) Cf(∗,1)di (ηµ) and respectively Cf(∗,1)di (ην) kYi=1 kYi=1 where we used the fact that (∗, 1)mJi = (∗, 1)di , for 1 ≤ i ≤ k. The two products in (2.8) are indeed equal by assumption (2) in the statement. (cid:3) Proof of Theorem 2.8. Suppose first that µ is η-diagonal. We verify that it satisfies (ηDM1) and (ηDM2). The hypothesis on µ says in particular that Cfw(ηµ) = 0 for every w ∈ W + which is not mixed-alternating. To prove (ηDM1), we must show that Cfw(Mµ) = 0 for every such word. This argument is carried precisely as in the proof of Lemma 2.9, with the roles of Mµ and ηµ being reversed and with (2.5) in place of (2.6). The reader will have no difficulty verifying the details. In order to show that µ also satisfies (ηDM2), fix a mixed-alternating word w ∈ W +, with canonical factorization w = w1 · · · wp as in (2.7). Set n = w = w1 + · · · + wp, and let ρ0 be the partition of {1, . . . , n} into intervals J1, . . . , Jp (written in increasing order) with lengths J1 = w1, . . . , Jp = wp. Given a partition π of {1, . . . , n}, we write π ≤ ρ0 if every block B of π is contained in one of the blocks J1, . . . , Jp of ρ0. (This relation is usually called the reverse refinement order on partitions.) Next, we use (2.5) to express the coefficient Cfw(Mµ) = µ(Z w) as a sum indexed by Int(n). The special structure of the coefficients of ηµ implies that a partition π ∈ Int(n) has a zero contribution to that sum unless π ≤ ρ0. It is immediate that the partitions π ∈ Int(n) satisfying π ≤ ρ0 are in natural bijective correspondence to tuples of partitions (π1, . . . , πp) where π1 ∈ Int(J1), . . . , πp ∈ Int(Jp). This correspondence is such that for π ↔ (π1, . . . , πp) we have These observations lead to the formula Cfw;π(ηµ) = Cfw1;π1(ηµ) · · · Cfwp;πp(ηµ). µ(Z w) = Xπ1∈Int(J1),..., πp∈Int(Jp) Cfw1;π1(ηµ) · · · Cfwp;πp(ηµ) = pYi=1(cid:16) Xπi∈Int(Ji) Cfwi;πi(ηµ)(cid:17). In the latter product, one more application of (2.5) identifies Xπi∈Int(Ji) Cfwi;πi(ηµ) = µ(Z wi), 1 ≤ i ≤ p, i=1 µ(Z wi). thus implying the desired conclusion that µ(Z w) =Qp Conversely, assume now that µ satisfies conditions (ηDM1) and (ηDM2). We show that it is η-diagonal by an indirect argument: we construct an η-diagonal distribution ν ∈ Dalg(1, ∗) and then prove that µ = ν. The distribution ν is defined by specifying its η-series (see Remark 2.4), namely Cfw(ην) = Cfw(ηµ) if w is alternating and Cfw(ην) = 0 otherwise. To prove that µ = ν we show that µ and ν satisfy the hypothesis of Lemma 2.10. Indeed, both µ and ν satisfy (ηDM1) and (ηDM2): µ does so by hypothesis, while ν does so because it is η-diagonal and by virtue of the implication (a) ⇒ (b) proved above. On the other hand, if w ∈ W + is an alternating word, the equality Cfw(ηµ) = Cfw(ην) is true by the definition of ν. This concludes the proof of the theorem. (cid:3) 8 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Remark 2.11. Let (A, ϕ) be a noncommutative probability space (that is, A is a unital algebra over C, ϕ : A → C is a linear functional, and ϕ(1) = 1), and let a1, a2 ∈ A. Recall [13] that a1, a2 are said to be Boolean independent provided that, given positive integers n, p1, . . . , pn and indices i1, . . . , in ∈ {1, 2} such that ik 6= ik+1 for i = 1, . . . , n − 1, the following identity is satisfied: ϕ(ap1 i1 · · · apn in ) = ϕ(ap1 i1 ) · · · ϕ(apn in ). Now consider the noncommutative probability space (ChZ, Z ∗i, µ), where µ ∈ Dalg(1, ∗) is η-diagonal. Condition (ηDM2) of Theorem 2.8 can be restated as saying that ZZ ∗ and Z ∗Z are Boolean independent in (ChZ, Z ∗i, µ). Remark and Notation 2.12. Another relevant fact concerning an η-diagonal ∗-distribu- tion µ concerns the individual η-series of ZZ ∗ and Z ∗Z in the noncommutative probability space (ChZ, Z ∗i, µ). If a is an element in a noncommutative probability space (A, ϕ), then its moment series and η-series Ma, ηa ∈ C[[z]] are defined as in Definition 2.3 but using moments in place of ∗-moments: first we set Ma(z) =P∞ ηa(z) = Ma(z)/(1 + Ma(z)) ∈ C[[z]]. n=1 ϕ(an)zn, and then define The coefficients of Ma and ηa are related to each other via summations over interval par- titions which are analogous to those shown in (2.5), (2.6) (and are derived the same way, by starting from the algebraic relations satisfied by the series themselves). We explicitly record here the analogue of (2.6): (2.9) Cf n(ηa) = Xρ∈Int(n) (−1)1+ρYB∈ρ ϕ(aB), n ∈ N, where (by analogy with Notation 2.2(2)) we use the notation Cfn : C[[z]] → C for the linear map that extracts the nth coefficient of a series in C[[z]]. When applied to the elements ZZ ∗ and Z ∗Z from the framework of Theorem 2.8, these observations yield the following result. n=1, (βn)∞ Proposition 2.13. Suppose that µ ∈ Dalg(1, ∗) is an η-diagonal distribution, and let (αn)∞ n=1 be its determining sequences (as introduced in Definition 2.6(2)). Then in the noncommutative probability space (ChZ, Z ∗i, µ), the elements ZZ ∗ and Z ∗Z have η-series given by αnzn and ηZ ∗ Z (z) = βnzn. ∞Xn=1 ηZZ ∗ (z) = ∞Xn=1 Cfn(ηZZ ∗ ) = Xρ∈Int(n) Proof. By symmetry, it suffices to prove the first formula. Equation (2.9) yields (2.10) (−1)1+ρYB∈ρ µ((ZZ ∗)B), n ∈ N. For the remainder of the proof, we fix n ∈ N and verify that the right-hand side of (2.10) is equal to αn. For every partition ρ = {J1, . . . , Jk} ∈ Int(n), with intervals J1, . . . , Jk written in increas- ing order, we define a doubled partition bρ = {bJ1, . . . ,cJk} ∈ Int(2n). This is the interval partition uniquely determined by the requirement that bJ1, . . . ,cJk come in increasing order and satisfy bJi = 2Ji for 1 ≤ i ≤ k. With this notation, it is easily seen that the right-hand ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 9 side of (2.10) can be written as This, however, is the same as Xρ∈Int(n) Xπ∈Int(2n) (−1)1+bρCf(1,∗)n;bρ(Mµ). (−1)1+πCf(1,∗)n;π(Mµ). Indeed, due to the special structure of the ∗-moments of µ described in Theorem 2.8, all the terms in the latter sum, corresponding to partitions π ∈ Int(2n) which are not of the form bρ, are equal to 0. We conclude that Cfn(ηZZ ∗ ) = Xπ∈Int(2n) (−1)1+πCf(1,∗)n;π(Mµ) = Cf(1,∗)n (ηµ) = αn, where (2.6) is used in the second equality. (cid:3) 3. R-transforms and R-diagonal ∗-distributions The discussion in Section 2 is better put into perspective when one compares it to the parallel (more elaborate) free probability framework. In the free probability framework, instead of η-series one works with R-transforms, and one has the concept of what it means for a ∗-distribution µ ∈ Dalg(1, ∗) to be R-diagonal. The class of R-diagonal ∗-distributions is in fact rather well-studied in the free probability literature. Here we review some of their basic properties, mostly following [11, Lecture 15], and with emphasis on the facts we need in the present work. Remark 3.1. On a combinatorial level, switching to the world of free probability comes to using non-crossing partitions instead of interval partitions. We recall that a crossing of a partition π of {1, . . . n} consists of integers 1 ≤ a < b < c < d ≤ n such that the set {a, c} is contained in a block of π and {b, d} is contained in a different block of π. A partition is non-crossing if it has no crossings. We denote by N C(n) the collection of all non-crossing partitions of {1, . . . n}. Similarly to the lattice Int(n), the set N C(n) is partially ordered by reverse refinement. The minimal and maximal elements with respect to this partial order are denoted by 0n (the partition of {1, . . . , n} into n singleton blocks) and respectively 1n (the partition of {1, . . . , n} into one block). We record a notation and an elementary observation needed in the final part of this section. For every n ∈ N, we denote by N CE(2n) the collection of all the partitions π ∈ N C(2n) with the property that every block of π has even cardinality. Observe that if π ∈ N CE(2n) and if V = {k1 < k2 < · · · < k2m} is a block of π, then the numbers k1, k2, . . . , k2m have alternating parities. Indeed, for every i = 1, . . . , 2m − 1, the set {ki + 1, ki +2, . . . , ki+1 −1} is a union of blocks of π, and hence has even cardinality, which implies that ki+1 is of opposite parity from ki. For a discussion of other elementary facts concerning N C(n), we refer to [11, Lecture 9]. Remark 3.2. (Review of R-transforms.) (1) The R-transform of a ∗-distribution µ ∈ Dalg(1, ∗) is the series Rµ ∈ C0hhz, z∗ii, whose coefficients are uniquely determined by the requirement that they relate to the ∗-moments 10 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI of µ by the formula (3.1) Cfw(Mµ) = Xπ∈N C(n) Cfw;π(Rµ), w ∈ W + and n = w. Equation (3.1) is the free probabilistic counterpart of (2.5). It is often referred to as the moment-cumulant formula for free cumulants (see [11, Lecture 11] for an explanation of this terminology). One can also define the series Rµ by an equation involving the series Mµ and Rµ them- selves (rather than their coefficients, as in (3.1)). More precisely, Rµ is the unique series in C0hhz, z∗ii that satisfies the functional equation (3.2) Rµ(z(1 + Mµ(z, z∗)), z∗(1 + Mµ(z, z∗)) = Mµ(z, z∗) (see [11, Corollary 16.16]). This is the free probabilistic analogue of (2.3), but now we only have an implicit functional equation rather than an explicit formula describing the series Rµ. An easy inductive argument in the moment-cumulant formula (3.1) shows that one can recover Mµ from Rµ and that (as in Remark 2.4) we have a bijection Dalg(1, ∗) ∋ µ 7→ Rµ ∈ C0hhz, z∗ii. In other words, one can define a distribution µ ∈ Dalg(1, ∗) by specifying its R-transform. (2) Consider the framework of Remark 2.11, where we discussed the moment series and η-series Ma, ηa ∈ C[[z]] associated to an element a in a noncommutative probability space (A, ϕ). In that framework one also has an R-transform associated with the element a ∈ A. This is the series Ra ∈ C[[z]] that relates to Ma by (3.3) and (3.4) Cfn(Ma) = Xπ∈N C(n) Cfn;π(Ra), n ∈ N, Ra(z(1 + Ma(z))) = Ma(z). These formulas are analogous to (3.1) or (3.2), respectively. For a detailed discussion of the algebraic aspects of R-transforms (covering both the series Rµ in part (1) of this remark and the series Ra in part (2)), see [11, Lecture 16]. Definition 3.3. (1) A ∗-distribution µ ∈ Dalg(1, ∗) is said to be R-diagonal when Cfw(Rµ) = 0 for every word w ∈ W + that is not alternating. (2) If µ is R-diagonal, its R-transform is thus of the form (3.5) Rµ(z, z∗) = αn(zz∗)n + βn(z∗z)n, ∞Xn=1 ∞Xn=1 with αn, βn ∈ C for n ∈ N. The sequences (αn)∞ sequences of µ. n=1 and (βn)∞ n=1 are called the determining Remark 3.4. The concept of an η-diagonal ∗-distribution from Section 2 obviously parallels the one of an R-diagonal ∗-distribution, with the η-series in place of the R-transform. The basic properties of η-diagonal distributions proved in Section 2 are the counterparts of known facts concerning R-diagonal distributions, as noted below. (1) Remark 2.11 is the Boolean counterpart of [11, Corollary 15.11, p. 244]: if µ is R-diagonal, then Z ∗Z and ZZ ∗ are freely independent in the noncommutative probability space (ChZ, Z ∗i, µ). ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 11 (2) Theorem 2.8 is the counterpart of [9, Theorem 1.2.1] which describes R-diagonal distributions in terms of their ∗-moments. (3) Proposition 2.13 is analogous to [11, Proposition 15.6, p. 241] which gives a precise formula, first found in [7], for the coefficients of the one-variable R-transforms RZZ ∗ and RZ ∗Z in terms of the determining sequences (αn)∞ n=1 from (3.5). This formula is more elaborate than the relation found in Proposition 2.13 for η-diagonal elements. It states that n=1 and (βn)∞ (3.6) Cfn(RZZ ∗) = Xπ={V1,...,Vk}∈N C(n) αV1βV2 · · · βVk, where the blocks of π are arranged so 1 ∈ V1. The coefficients of RZ ∗Z are obtained by interchanging the roles of α and β in these formulas. For example, the first three coefficients of RZZ ∗ are α1, α2 + α1β1 and α3 + 2α2β1 + α1β2 + α1β2 1 . Equations (3.6) lead to the following observation: an easy induction on n (where one singles out the terms indexed by the partition 1n ∈ N C(n) on the right-hand sides of these equations) shows that the determining sequences (αn)∞ n=1 can be retrieved from the coefficients of RZZ ∗ and RZ ∗Z . Hence the R-diagonal ∗-distribution µ ∈ Dalg(1, ∗) is completely determined by these R-transforms. n=1 and (βn)∞ In Section 6 we require a reformulation of (3.6) in terms of operations with series rather than individual coefficients. This reformulation is given in the next proposition. The for- mulas (3.7) bear a striking resemblance to the functional equation (3.4) of the R-transform. In fact, (3.7) collapse to RZZ ∗(z) = RZ ∗Z(z) = Ma(z) in the special case αn = βn, n ∈ N, in which case we can take a = b. Proposition 3.5. Let µ ∈ Dalg(1, ∗) be an R-diagonal ∗-distribution with determining sequences (αn)∞ n=1. Suppose we are given some elements a and b in a noncom- mutative probability space (A, ϕ) such that n=1 and (βn)∞ Ra(z) = ∞Xn=1 αnzn and Rb(z) = βnzn. ∞Xn=1 Then, in the noncommutative probability space (ChZ, Z ∗i, µ), we have (3.7) RZZ ∗(z) = Ra(z(1 + Mb(z))) and RZ ∗Z (z) = Rb(z(1 + Ma(z))). Proof. The argument is analogous to the proof of (3.2) (see, for instance, the proof of [11, Theorem 16.15]). For the reader's convenience, we describe the basic idea. By symmetry, it suffices to prove the first equality in (3.7). We show that the coefficients of zn in the series RZZ ∗(z) and Ra(z(1 + Mb(z))) are equal to each other for every n ∈ N. The formal series expansion Ra(z(1 + Mb(z))) = ∞Xm=1 αm(cid:0)z(1 + Mb(z))(cid:1)m yields Cfn(cid:0)Ra(z(1 + Mb(z)))(cid:1) = nXm=1 αmCfn−m(cid:0)(1 + Mb)m(cid:1), n ∈ N. 12 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Recall that the coefficients of 1 + Mb are moments of b, and expand (1 + Mb)m, to obtain (3.8) Cfn[Ra(z(1 + Mb)(z))] = nXm=1 Xk1,...,km≥0 with k1+···+km=n−m αmϕ(bk1) · · · ϕ(bkm). On the other hand, (3.6) yields Cfn(RZZ ∗) = nXm=1 XS⊆{1,...,n} with Xπ={V1,...,Vp}∈N C(n) S=m and 1∈S with V1=S αmβV2 · · · βVp, n ∈ N. For a fixed set S = {s1, . . . , sm} ⊆ {1, . . . , n} with 1 = s1 < s2 < · · · < sm ≤ n, the collection of non-crossing partitions {π ∈ N C(n) : S is a block of π} is naturally identified with the Cartesian product N C(s2 − s1 − 1) × · · · × N C(sm − sm−1 − 1) × N C(n − sm), in a way that converts the sum βV2 · · · βVp with V1=S Xπ={V1,...,Vp}∈N C(n)  Xπℓ∈N C(sℓ+1−sℓ−1) YV ∈πℓ mYℓ=1 βV  , into the product (3.9) that QV ∈πℓ We conclude that where we set sm+1 = n + 1. An application of the moment-cumulant formula (3.3) shows βV = ϕ(bsℓ+1−sℓ−1). (See the proof of [11, Theorem 16.15] for more details.) (3.10) Cfn(RZZ ∗) = nXm=1 XS⊆{1,...,n} with S=m and 1∈S αm mYℓ=1 ϕ(bsℓ+1−sℓ−1), n ∈ N. Finally, observe that for every fixed m ∈ {1, . . . , n} there is a natural bijection between tuples (k1, . . . , km) ∈ (N ∪ {0})m with k1 + · · · + km = n − m (on the one hand) and subsets 1 ∈ S ⊆ {1, . . . , n} with S = m (on the other), given by the formula (k1, . . . , km) 7→ S = {1, k1 + 2, k1 + k2 + 3, . . . , k1 + · · · + km−1 + m}. The inner sums on the right-hand sides of (3.8) and (3.10) are identified term by term via this bijection, and this concludes the proof. (cid:3) In the remainder of this section we discuss the R-diagonal ∗-distributions that satisfy the KMS condition. This is a special case of the class of ∗-distribution studied in [12] (see, for instance [12, Remark 2.10]). The best known example of a KMS R-diagonal distribution is the one where, in the framework of the next definition, one sets α1 = λ, β1 = 1 and αn = βn = 0 for n ≥ 2; this is called the λ-circular distribution, and is studied in detail in [12, Section 4]. ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 13 Definition 3.6. Let µ ∈ Dalg(1, ∗) be an R-diagonal ∗-distribution with determining se- quences (αn)∞ n=1, and let t be a positive real number. We say that µ satisfies the KMS condition with parameter t if n=1 and (βn)∞ (3.11) αn = tβn, n ∈ N. The following result shows that the KMS condition is a generalization of the trace prop- erty, where the latter property occurs for the value t = 1 of the parameter. Proposition 3.7. Let µ ∈ Dalg(1, ∗) be an R-diagonal ∗-distribution, let t be a positive real number, and suppose that µ satisfies the KMS condition with parameter t. Denote by the unique unital algebra homomorphism such that Ut : ChZ, Z ∗i → ChZ, Z ∗i Then (3.12) Ut(Z) = tZ and Ut(Z ∗) = Z ∗. 1 t µ(P Q) = µ(QUt(P )), P, Q ∈ ChZ, Z ∗i. Proof. Both sides of (3.12) are bilinear in P and Q, so it suffices to check the equation when both P and Q are monomials. Using the notation Z ∅ = 1, we must show that (3.13) µ(Z vZ w) = µ(Z w Ut(Z v)), v, w ∈ W + ∪ {∅}. Equivalently, we must show that the set S = {v ∈ W + ∪ {∅} : µ(Z vZ w) = µ(Z wUt(Z v)), w ∈ W + ∪ {φ}} is equal to W + ∪ {∅}. The set S is clearly closed under concatenation and contains ∅. Therefore, it suffices to show that {1, ∗} ⊂ S. In other words it suffices to prove that (3.14) µ(ZZ w) = tµ(Z wZ) and µ(Z ∗Z w) = µ(Z wZ ∗), w ∈ W + ∪ {∅}. 1 t We only prove the first equality in (3.14); the verification of the second one is analogous. The case w = ∅ follows from the fact (incorporated in the definition of an R-diagonal distribution) that µ(Z) = 0. For the remainder part of the proof we fix a word w = (ℓ1, . . . , ℓn) ∈ W +, for which we prove that µ(ZZ w) = t µ(Z wZ). Observe that µ(ZZw) = Cfw1(Mµ), µ(ZwZ) = Cfw2(Mµ), where w1 := (1, ℓ1, . . . , ℓn) and w2 := (ℓ1, . . . , ℓn, 1). It is convenient to view w1 and w2 as functions from {1, . . . , n + 1} to {1, ∗} and to record the fact that (3.15) w2 = w1 ◦ γn+1, where γn+1 is the cyclic permutation 1 7→ 2 7→ · · · 7→ n + 1 7→ 1 of {1, . . . , n + 1}. For every partition π = {V1, . . . , Vk} ∈ N C(n+1) we denote by γ−1 in N C(n+1)) whose blocks are γ−1 tCfw2(Mµ) is obtained from n+1(V1), . . . , γ−1 n+1(π) the partition (still n+1(Vk). The desired conclusion Cfw1(Mµ) = (3.16) Cfw1;π(Rµ) = tCfw2;γ−1 n+1(π)(Rµ), π ∈ N C(n + 1), using the moment-cumulant formula. Indeed, sum both sides of (3.16) over π ∈ N C(n + 1) and invoke (3.1) applied to the words w1 and w2. The sums thus obtained are precisely Cfw1(Mµ) on the left side and tCfw2(Mµ) on the right. 14 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Thus, it remains to prove (3.16). Fix a partition π = {V1, . . . , Vk} ∈ N C(n + 1) such that n+1(Vj), 1 ≤ j ≤ k, and n + 1 ∈ W1. n+1(π) = {W1, . . . , Wk}, where Wj = γ−1 1 ∈ V1. Then γ−1 It follows from (3.15) that w1 Vj = w2 Wj ∈ W + if 2 ≤ j ≤ k, and thus (3.17) Cfw1Vj (Rµ) = Cfw2Wj (Rµ), 2 ≤ j ≤ k For the remaining block, we show that (3.18) Cfw1V1(Rµ) = t · Cfw2W1(Rµ). Indeed, suppose that V1 = {j1, . . . , jm} with 1 = j1 < j2 < · · · < jm, and therefore W1 = {j2 − 1, . . . , jm − 1, n + 1}. Both sides of (3.18) are 0 if m is odd or if m is even but w1V1 is not an alternating word. If m is even and w1V1 is alternating, then we find that w1V1 = {1, ∗}m/2 and w2V2 = {∗, 1}m/2, which implies that Cfw1V1(Rµ) = αm/2 and Cfw2W1(Rµ) = βm/2. In this case, (3.18) follows from the KMS hypothesis. Finally, for the partition π fixed in the preceding paragraph we write: Cfw1;π(Rµ) = Cfw1V1(Rµ) Cfw1Vj (Rµ) = tCfw2W1(Rµ) Cfw2Wj (Rµ) (by (3.17) and (3.18)) kYj=2 kYj=2 = tCfw2;γ−1 n+1(π)(Rµ), thus concluding the proof of (3.16). (cid:3) Remark 3.8. The converse of Proposition 3.7 is also true. More precisely, every R-diagonal distribution µ ∈ Dalg(1, ∗) that satisfies (3.12) for some t ∈ (0, +∞) must also satisfy the KMS condition for the same value of t. To see this, let (αn)∞ n=1 be the determining sequences of µ. Equation (3.12) yields, in particular, the identity n=1 and (βn)∞ (3.19) µ((ZZ ∗)n) = tµ((Z ∗Z)n), n ∈ N. This identity implies αn = tβn, n ∈ N, by induction on n. For the induction step one invokes the moment-cumulant formula in order to expand both sides of (3.19) as sums over N C(2n); then the action of the cyclic permutation γ−1 2n on N C(2n) can be used in the same way as it was done in the proof of Proposition 3.7. 4. The framework of Dc(1, ∗) and Dc(k), BBP bijections We now introduce the analytic framework which is of interest for the present paper. Definition 4.1. (1) Let (A, ϕ) be a C ∗-probability space (which means that A is a unital C ∗-algebra, ϕ : A → C is a positive linear functional, and ϕ(1) = 1), and let a ∈ A. The ∗-distribution of a is the functional µ ∈ Dalg(1, ∗) determined by the requirement that µ(Z w) = ϕ(aw), w ∈ W +. (2) We denote by Dc(1, ∗) the set of all elements of Dalg(1, ∗) that are equal to the ∗-distribution of some element in a C ∗-probability space. (3) Free additive (respectively, multiplicative) convolution is a binary operation on Dc(1, ∗) denoted by ⊞ (respectively, ⊠). This operation is uniquely determined by the following property: given elements a, a′ in a C ∗-probability space (A, ϕ) such that {a, a∗} is free from {a′, (a′)∗}, the ∗-distribution of a + a′ (respectively, aa′) is the free additive (respectively, ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 15 multiplicative) convolution of the ∗-distributions of a and a′. See [11, Lectures 5 and 7] for more details. (4) An element µ ∈ Dc(1, ∗) is said to be ⊞-infinitely divisible if for every n ∈ N there exists µn ∈ Dc(1, ∗) such that The set of all ⊞-infinitely divisible distributions in Dc(1, ∗) is denoted by D(inf−div) c (1, ∗). µ = µn ⊞ · · · ⊞ µn . n times {z } The main result of this section is the following theorem. The series Rν and ηµ appearing in the statement of the theorem are as defined in Sections 2 and 3. Theorem 4.2. (BBP bijection on Dc(1, ∗).) There exists a bijection B (1,∗) : Dc(1, ∗) → D(inf−div) c (1, ∗), determined by the requirement that (4.1) RB(1,∗)(µ) = ηµ, µ ∈ Dc(1, ∗). More precisely, for every µ ∈ Dc(1, ∗) there exists a unique ∗-distribution ν ∈ D(inf−div) such that Rν = ηµ, and we define B (1,∗)(µ) := ν. c (1, ∗) Definition and Remark 4.3. (Framework of Dc(k).) We reduce Theorem 4.2 to an analogous theorem proved in [2] for the space, denoted by Dc(2), of joint distributions of pairs of selfadjoint elements in a C ∗-probability space. The passage from Dc(1, ∗) to Dc(2) is natural, and essentially amounts to the change of variables (a, a∗) 7→(cid:18) a + a∗ 2 , a − a∗ 2i (cid:19) , for a in a C ∗-probability space (A, ϕ). In order to clarify this idea, we review briefly the framework of Dc(k). Fix k ∈ N. (1) We denote by ChX1, . . . , Xki the algebra of polynomials in the non-commuting inde- terminates X1, . . . , Xk. (2) Let (A, ϕ) be a C ∗-probability space and let b1, . . . , bk ∈ A be selfadjoint. The joint distribution of b1, . . . , bk is the linear functional λ : ChX1, . . . , Xki → C which is determined by the requirement that λ(1) = 1 and λ(Xi1 · · · Xin) = ϕ(bi1 · · · bin), n ∈ N, i1, . . . , in ∈ {1, . . . , k}. (3) We denote by Dc(k) the set of all linear functionals λ : ChX1, . . . , Xki → C that can arise as joint distributions of k-tuples of selfadjoint elements in some C ∗-probability space. (4) Free additive convolution is a binary operation on Dc(k) denoted 2 by ⊞. This operation is uniquely determined by the following property: given selfadjoint elements a1, . . . , ak and b1, . . . , bk in a C ∗-probability space (A, ϕ) such that {a1, . . . , ak} is free from {b1, . . . , bk}, the joint distribution of a1 + b1, . . . , ak + bk is the free additive convolution of the joint distributions of a1, . . . , ak and b1, . . . , bk. The concept of ⊞-infinite divisibility in Dc(k) is introduced as in Definition 4.1(4). The set of ⊞-infinitely divisible distributions in Dc(k) is denoted by D(inf−div) (k). c (5) We denote by C0hhx1, . . . , xkii the space of those formal power series with complex coefficients in k non-commuting indeterminates x1, . . . , xk whose constant term is equal to 2It is customary to always denote free additive convolution by "⊞". The setting in which the symbol ⊞ is used should be clear, in each case, from the context. 16 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI 0. We denote by Cf(i1,...,in)(f ) the coefficient of xi1 · · · xin in a series f ∈ C0hhx1, . . . , xkii. Every joint distribution λ ∈ Dc(k) has a moment series Mλ, an R-transform Rλ and an η-series ηλ. These are elements of C0hhx1, . . . , xkii, and their definitions are analogous to Definitions 2.3 and 3.2. A detailed description of these power series and of the relations between their coefficients can be found in [2, pp. 14-17]. The proof of Theorem 4.2 will be reduced to the following result from [2] (see also [4] for the case k = 1). Theorem 4.4. (BBP bijection on Dc(k).) Let k be a positive integer. There exists a bijection Bk : Dc(k) → D(inf−div) (k), determined by the requirement that c (4.2) RBk(λ) = ηλ, λ ∈ Dc(k). More precisely, for every λ ∈ Dc(k) there exists a unique ν ∈ D(inf−div) ηλ, and we define Bk(λ) := ν. c (k) such that Rν = (cid:3) Remark 4.5. We only need Theorem 4.4 for k = 1 and k = 2. When k = 1, the space Dc(1) is naturally identified with the space Pc of compactly supported Borel probability measures on R. Indeed, given b = b∗ in a C ∗-probability space (A, ϕ), Definition 4.3(2) produces a linear functional λ : Chx1i → C which becomes, via the Riesz representation theorem, a Borel probability measure supported on the spectrum of b. The original BBP bijection from [4] was defined on Pc (and on the larger set P of all Borel probability measures on R). In Section 6, we will simply talk about B(σ) for σ ∈ Pc. In other words if λ denotes the functional in Dc(1) corresponding to σ, then B(σ) ∈ Pc denotes the probability measure corresponding to B1(λ). The following result creates bijections C and D that we use in conjunction with the case k = 2 of Theorem 4.4. (The letters C and D are meant to suggest complexification and decomplexification.) The proof is immediate, and therefore omitted. Proposition 4.6. There exists a bijection D : Dc(1, ∗) → Dc(2) defined as follows. Given µ ∈ Dc(1, ∗) that is the ∗-distribution of an element a in a C ∗-probability space (A, ϕ), D(µ) is the joint distribution of the pair (cid:18) a + a∗ 2 , a − a∗ 2i (cid:19) . The inverse of D is the bijection C : Dc(2) → Dc(1, ∗) defined as follows. Given λ ∈ Dc(2) that is the joint distribution of a pair b1, b2 of selfadjoint elements in a C ∗-probability space (A, ϕ), C(λ) is the ∗-distribution of b1 + ib2. (cid:3) Definition and Remark 4.7. In addition to the transformations C and D, the proof of Theorem 4.2 requires the corresponding change of variables for power series. Denote by (4.3) t1,1 = t1,∗ = 1 2 and t2,1 = 1 2i , t2,∗ = − 1 2i , the coefficients of the linear transformation b1 = (a + a∗)/2, b2 = (a − a∗)/2i. This trans- formation can now be written more compactly as (4.4) bi = Xℓ∈{1,∗} ti,ℓaℓ, for i = 1, 2. ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 17 Using the coefficients ti,ℓ we define a map eD : C0hhz, z∗ii → C0hhx1, x2ii as follows: given a series f ∈ C0hhz, z∗ii, the coefficients of the series g = eD(f ) ∈ C0hhx1, x2ii are given by (4.5) Cf(i1,...,in)(g) := Xℓ1,...,ℓn∈{1,∗} The map eD is clearly linear and bijective. Its inverse eC : C0hhx1, x2ii → C0hhz, z∗ii is defined by a formula analogous to (4.5), but with [ti,ℓ] replaced by the inverse matrix ti1,ℓ1 · · · tin,ℓnCf(ℓ1,...,ℓn)(f ), n ∈ N, i1, . . . , in ∈ {1, 2}. t′ 1,1 = 1, t′ 1,2 = i and t′ ∗,1 = 1, t′ ∗,2 = −i. Lemma 4.8. For every ∗-distribution µ ∈ Dc(1, ∗), we have (4.6) MD(µ) = eD(Mµ), RD(µ) = eD(Rµ), and ηD(µ) = eD(ηµ). Proof. Suppose that µ ∈ Dc(1, ∗) is the ∗-distribution of an element a in a C ∗-probability space (A, ϕ), and set bi =Pℓ∈{1,∗} ti,ℓaℓ for i = 1, 2. By definition, the joint distribution of b1, b2 is D(µ) ∈ Dc(2). To verify the first identity in (4.6), fix n ∈ N and i1, . . . , in ∈ {1, 2}, and calculate directly Cf(i1,...,in)(MD(µ)) = ϕ(bi1 · · · bin) ti1,ℓ1 · · · tin,ℓn ϕ(aℓ1 · · · aℓn) = ϕ(cid:16)(ti1,1a1 + ti1,∗a∗) · · · (tin,1a1 + tin,∗a∗)(cid:17) = Xℓ1,...,ℓn∈{1,∗} = Xℓ1,...,ℓn∈{1,∗} = Cf(i1,...,in)eD(Mµ). ti1,ℓ1 · · · tin,ℓnCf(ℓ1,...,ℓn)(Mµ) The second equality in (4.6) follows from a similar multilinearity argument, using the fact that the C ∗-probability space (A, ϕ) carries a family of multilinear functionals (κn : An → C)∞ n=1, called free cumulant functionals, such that Cf(i1,...,in)(RD(µ)) = κn(bi1 , . . . , bin), n ∈ N, i1, . . . , in ∈ {1, 2}, and Cf(ℓ1,...,ℓn)(Rµ) = κn(aℓ1, . . . , aℓn), n ∈ N, ℓ1, . . . , ℓn ∈ {1, ∗}. (This multilinearity argument is precisely the one used to describe the behavior of the R-transform under linear transformations [11, Proposition 16.12].) The third equality (4.6) follows from a similar multilinearity argument, using the Boolean n=1 (for a discussion of Boolean cumulants see, for (cid:3) cumulant functionals (βn : An → C)∞ instance, [8, Section 4.6]). Lemma 4.9. Let D : Dc(1, ∗) → Dc(2) be the bijection defined in Proposition 4.6. Then: (1) D(µ ⊞ µ′) = D(µ) ⊞ D(µ′) for every µ, µ′ ∈ Dc(1, ∗). (2) D(D(inf−div) (1, ∗)) = D(inf−div) (2). c c 18 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Proof. (1) Since every λ ∈ Dc(2) is uniquely determined by its R-transform, it suffices to verify that D(µ ⊞ µ′) and D(µ) ⊞ D(µ′) have the same R-transform. Indeed, RD(µ⊞µ′) = eD(Rµ⊞µ′ ) (by Lemma 4.8) = eD(Rµ) + eD(Rµ′) (since Rµ⊞µ′ = Rµ + Rµ′ and eD is linear) = RD(µ) + RD(µ′) (by Lemma 4.8) = RD(µ)⊞D(µ′). Part (2) follows immediately from (1) and from the definition of ⊞-infinite divisibility. (cid:3) Proof of Theorem 4.2. We define the required bijection B (1,∗) so that the diagram Dc(1, ∗) D↓ c B(1,∗)−→ D(inf−div) ↓D (1, ∗) Dc(2) B2−→ D(inf−div) c (2) is commutative, where D is defined in Proposition 4.6 and B2 is provided by Theorem 4.4 for k = 2. More precisely, let D0 : D(inf−div) (2) be the restriction of D; this is a bijection by Lemma 4.9(2)). Then define (1, ∗) → D(inf−div) c c B(1,∗) := D−1 0 ◦ B2 ◦ D. Pick an arbitrary µ ∈ Dc(1, ∗), denote B injective, it suffices to verify that eD(Rν) = eD(ηµ). Indeed, the definition of B B2(D(µ)) = D(ν), and the definition of B2 yields RD(ν) = ηD(µ). Thus (1,∗)(µ) = ν. We prove that Rν = ηµ. Since eD is (1,∗) implies eD(Rν ) = RD(ν) = ηD(µ) = eD(ηµ), as required. (cid:3) 5. Parametrization of η-diagonal distributions in Dc(1, ∗) We show that an η-diagonal ∗-distribution µ ∈ Dc(1, ∗) is naturally parametrized by the pair of compactly supported probability measures on [0, ∞) that arise as the distributions of ZZ ∗ and of Z ∗Z in the noncommutative probability space (ChZ, Z ∗i, µ). Definition and Remark 5.1. Suppose that µ ∈ Dc(1, ∗) is the ∗-distribution of an element a in a C ∗-probability space (A, ϕ). Basic considerations on positive elements in a C ∗- probability space (see, for instance, [11, Propositions 3.13 and 3.6]) show the existence of compactly supported Borel probability measures σ1, σ2 on [0, ∞) such that 0 tn dσ1(t) and ϕ((a∗a)n) =Z ∞ ϕ((aa∗)n) =Z ∞ Z ∞ Z ∞ 0 0 0 tn dσ2(t) = µ((Z ∗Z)n) n ∈ N. tn dσ1(t) = µ((ZZ ∗)n), n ∈ N Thus σ1 and σ2 satisfy (5.1) and (5.2) tn dσ2(t), n ∈ N. ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 19 Moreover, σ1 and σ2 are uniquely determined by (5.1) and (5.2) since a compactly supported probability measure on R is determined by its moments. We refer to σ1 and σ2 as the distributions of ZZ ∗ and Z ∗Z, respectively, in the ∗-probability space (ChZ, Z ∗i, µ). The following theorem provides the parametrization announced in the title of the section. c denote the set of all compactly supported Borel probability measures Theorem 5.2. Let P + on [0, ∞). There is bijective map c × P + Φ : P + c → {µ ∈ Dc(1, ∗) : µ is η-diagonal} described as follows: given σ1, σ2 ∈ P + c , Φ(σ1, σ2) is the unique η-diagonal ∗-distribution µ ∈ Dc(1, ∗) such that the distributions of ZZ ∗ and Z ∗Z in (ChZ, Z ∗i, µ) are equal to σ1 and σ2, respectively. c × P + The point of Theorem 5.2 is that the map Φ is defined on all of P + c . In other words, for every σ1, σ2 ∈ P + c there exists an η-diagonal ∗-distribution µ ∈ Dc(1, ∗) such that (5.1) and (5.2) hold. We prove this by producing an operator model for µ: starting from σ1 and σ2 we construct explicitly an operator A on a Hilbert space K such that the ∗-distribution of A with respect to a suitably chosen functional on B(K) is the required η-diagonal distribution. The bulk of this section is devoted to the description of the operator model. At the end, we complete the proof of Theorem 5.2. The construction of the operator model is described in the next remark. Remark and Notation 5.3. (Description of the operator model.) Fix σ1, σ2 ∈ P + c which we take as the input for our construction of an η-diagonal operator. In the description of the construction, it is convenient to use the symmetric square roots of σ1 and σ2. These moments given by the formula are the symmetric compactly supported Borel probability measures eσ1 and eσ2 on R with Z ∞ −∞ tn deσj(t) =(0, R ∞ 0 n odd tn/2 dσj(t), n even, for j = 1, 2. Our construction of an η-diagonal operator proceeds in three steps. Step 1. We construct a Hilbert space H, an operator X ∈ B(H), and vectors ξ1, ξ2 ∈ H with the following properties: (1a) ξ1 = ξ2 = 1, (1b) hξ1, ξ2i = 0, (1c) hX kξ1, ξ2i = hX kξ2, ξ1i = 0 for k ∈ N, 0 tk dσj(t) for k ∈ N and j = 1, 2. algebra B(H). For the actual construction of X consider, for j = 1, 2, Hilbert spaces Mj, operators Tj ∈ B(Mj), and unit vectors ηj ∈ Mj, such that the distribution of Tj with (1d) hX 2k−1ξj, ξji = 0 and hX 2kξj, ξji =R ∞ In other words, property (1d) says that X has distribution eσ1 with respect to the vector state defined by ξ1, and distributioneσ2 with respect to the one defined by ξ2, on the operator respect to the vector state ηj is eσj. Then set H = M1 ⊕ M2, X = T1 ⊕ T2, ξ1 = η1 ⊕ 0, and ξ2 = 0 ⊕ η2. Properties (1a) -- (1d) are then easily verified. (In subsequent steps we only use the properties (1a) -- (1d). The precise description of H, X, ξ1, and ξ2 is not necessary.) Step 2. Define a rank-one partial isometry Y ∈ B(H) by setting Y (ζ) = hζ, ξ1iξ2 for ζ ∈ H. We have Y ξ1 = ξ2, Y ∗ζ = hζ, ξ2iξ1 for ζ ∈ H, and (5.3) Thus Y Y ∗ and Y ∗Y are the orthogonal projections onto the 1-dimensional spaces generated by ξ2 and ξ1, respectively. Y Y ∗ζ = hζ, ξ2iξ2, Y ∗Y ζ = hζ, ξ1iξ1. 20 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI Step 3. Consider the Hilbert space K = H ⊗ H and the unit vector ξ = ξ1 ⊗ ξ2 ∈ K, then consider the C ∗-probability space (B(K), ϕξ), where ϕξ(T ) = hT ξ, ξi for T ∈ B(K). Let V ∈ B(H⊗H) be the flip operator determined by the requirement that by V (ζ ⊗ζ ′) = ζ ′ ⊗ζ, ζ, ζ ′ ∈ H. Note that V is a symmetry (that is, it is self-adjoint and V 2 = I). Finally, define A = V (Y ⊗ X). This concludes the construction of the variable A in (B(K), ϕξ). We now take on the proof that the operator A constructed above has the desired η- diagonal distribution with respect to the functional ϕξ. We start by recording some easily verified identities satisfied by A, the proof of which is left to the reader. Lemma 5.4. Consider the framework of Remark 5.3. We have AA∗ = X 2 ⊗ Y Y ∗, A∗A = Y ∗Y ⊗ X 2, (5.4) and (5.5) A2 = XY ⊗ Y X, (A∗)2 = Y ∗X ⊗ XY ∗. The following lemma establishes the distributions of AA∗ and A∗A along with a few non-alternating ∗-moments of A. (cid:3) Lemma 5.5. Let A be as above, then for any integer k ≥ 0 we have (2) ϕξ(A(AA∗)k) = 0, (3) ϕξ(A∗(AA∗)k) = 0, (1) ϕξ((AA∗)k) =R ∞ (4) ϕξ((A∗A)k) =R ∞ (5) ϕξ(A(A∗A)k) = 0, (6) ϕξ(A∗(A∗A)k) = 0. 0 0 tk dσ1(t), tk dσ2(t), Proof. We verify only the first three equations. The proof of (4) -- (6) is similar. We have so (AA∗)k = X 2k ⊗ (Y Y ∗)k = X 2k ⊗ (Y Y ∗), ϕξ((AA∗)k) = h(X 2k ⊗ (Y Y ∗))ξ, ξi = hX 2kξ1, ξ1i, and (1) follows from property (1d) in Step 1 of the construction of A. To prove (2), we calculate A(AA∗)k = V (Y ⊗ X)(X 2k ⊗ Y Y ∗), thus ϕξ(A(AA∗)k) = h(Y ⊗ X)(X 2k ⊗ Y Y ∗)ξ1 ⊗ ξ2, V ξ1 ⊗ ξ2i = h(Y ⊗ X)(X 2kξ1 ⊗ ξ2), ξ2 ⊗ ξ1i = hY X 2kξ1, ξ2ihXξ2, ξ1i = 0, because hXξ2, ξ1i = 0, thereby concluding the proof of (2). Similarly, A∗(AA∗)k = (X ⊗ Y ∗)V (X 2k ⊗ Y Y ∗), so ϕξ(A∗(AA∗)k) =h((X ⊗ Y ∗)V (X 2k ⊗ Y Y ∗)ξ, ξi = hX ⊗ Y ∗(ξ2 ⊗ X 2kξ1), ξ1 ⊗ ξ2i =hXξ2, ξ1ihY ∗X 2kξ1, ξ2i = 0 by property (1c) in Step 1 of the construction of A. (cid:3) ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 21 The next lemma gives some properties of the operator A that are useful in verifying its η -- diagonality. Lemma 5.6. Let A be as above, then for any integer k ≥ 0 we have (1) A2(AA∗)kξ = 0, (2) (A∗)2(AA∗)kξ = 0, (4) A2(A∗A)kξ = 0, (5) (A∗)2(A∗A)kξ=0, (3) (A∗A)(AA∗)kξ =R ∞ (6) (AA∗)(A∗A)kξ =R ∞ 0 0 tk dσ1(t) · A∗Aξ, tk dσ2(t) · AA∗ξ. Proof. As in the previous proof we only verify (1) -- (3). We have (A)2(AA∗)k = (XY ⊗ Y X)(X 2k ⊗ Y Y ∗) = XY X 2k ⊗ Y XY Y ∗, and using the fact that Y Y ∗ξ2 = ξ2 we see that Y XY Y ∗ξ2 = hXξ2, ξ1iξ2 = 0 by property (1c). Similarly, (A∗)2(AA∗)k = (Y ∗X ⊗ XY ∗)(X 2k ⊗ Y Y ∗) = Y ∗X 2k+1 ⊗ XY ∗, and (2) follows because Y ∗X 2k+1ξ1 = hX 2k+1ξ1, ξ2iξ1 = 0 by (1c). Finally, (5.4) yields (A∗A)(AA∗)k = (Y ∗Y ⊗ X 2)(X 2k ⊗ Y Y ∗) = Y ∗Y X 2k ⊗ X 2Y Y ∗. Observe that Y ∗Y X 2kξ1 = hX 2kξ1, ξ1iξ1 = R ∞ Therefore 0 (A∗A)(AA∗)k =Z ∞ 0 tk dσ1(t) · (ξ1 ⊗ X 2ξ2) =Z ∞ 0 tk dσ1(t) · ξ1 by (1d), while Y Y ∗ξ2 = ξ2. tk dσ1(t) · A∗Aξ, thus proving (3). (cid:3) Corollary 5.7. Let W = W1W2 · · · Wd be a mixed-alternating word in A and A∗, factored as in (2.7) with d ≥ 2. Then W ξ = ϕξ(W2)ϕξ(W3) · · · ϕξ(Wd)W1ξ. Proof. Parts (3) and (6) of the preceding lemma yield the conclusion when d = 2. The general case follows easily by induction on d. (cid:3) Proposition 5.8. Let σ1 and σ2 be probability measures in P + c , and let the operator A in (B(K), ϕξ) be constructed as in Remark 5.3. Then the ∗-distribution of A is η-diagonal. Moreover, the distributions of AA∗ and A∗A are σ1 and σ2, respectively. Proof. The second assertion follows from parts (1) and (4) of Lemma 5.5. It remains to prove that the distribution of A is η-diagonal, and to do this we verify the conditions in Theorem 2.8. Let W = W1W2 · · · Wd be a mixed-alternating word in A and A∗, factored as in (2.7). Corollary 5.7 yields ϕξ(W ) = ϕξ(W2)ϕξ(W3) · · · ϕξ(Wd)hW1ξ, ξi = ϕξ(Wj), dYj=1 thus verifying condition (ηDM2). Finally we verify condition (ηDM1). Suppose that V is a word in A and A∗ that is not mixed-alternating, and choose a mixed-alternating word W of 22 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI maximum length with the property that V can be written as V = U W for some non-empty word U . Also, write W = W1W2 · · · Wd as in (2.7). We have ϕξ(V ) = hV ξ, ξi = ϕξ(Wj)hU W1ξ, ξi dYj=2 by Corollary 5.7. If U = 1, the equality ϕξ(V ) = 0 follows from Lemma 5.5. If U ≥ 2, then U is of the form U ′AA or U ′A∗A∗ for some (possibly empty) word U ′. In this case, ϕξ(V ) = 0 by Lemma 5.6. (cid:3) We conclude the discussion of the parametrization announced at the beginning of the section. Proof of Theorem 5.2. We first note that the map Φ is well-defined. Indeed, let σ1, σ2 ∈ P + c be given. The existence of an η-diagonal ∗-distribution µ ∈ Dc(1, ∗) which fulfils the conditions (5.2) is ensured by Proposition 5.8. The uniqueness of µ follows from the fact that an η-diagonal ∗-distribution is completely determined by its alternating ∗-moments, as we saw in Theorem 2.8. The surjectivity of Φ is immediate from its definition: every η-diagonal ∗-distribution c are the distributions of ZZ ∗ µ ∈ Dc(1, ∗) can be written as Φ(σ1, σ2), where σ1, σ2 ∈ P + and respectively Z ∗Z in (ChZ, Z ∗i, µ), in the sense discussed in Definition 5.1. Finally, the injectivity of Φ is immediate as well. Indeed, if Φ(σ1, σ2) = µ, then the mo- ments of σ1 and σ2 can be retrieved as alternating moments of µ, and compactly supported probability measures on R are determined by their moments. (cid:3) 6. Parametrization of infinitely divisble R-diagonal distributions In this section we use the BBP method to characterize ⊞-infinitely divisible R-diagonal distributions. The parametrization mentioned in the title of the section arises naturally, in the way indicated in the following remark. Remark and Notation 6.1. Let R(inf−div) denote the set of all the R-diagonal distribu- tions in Dc(1, ∗) that are ⊞-infinitely divisible. It is immediate that the bijection B(1,∗) from Theorem 4.2 induces a bijection (still denoted B c (1,∗)) (6.1) B (1,∗) : {µ ∈ Dc(1, ∗) : µ is η-diagonal} → R(inf−div) c . On the other hand, Theorem 5.2 provides a natural bijection (6.2) The map (6.3) Φ : P + c × P + c → {µ ∈ Dc(1, ∗) : µ is η-diagonal}. Ψ := B(1,∗) ◦ Φ : P + c × P + c → R(inf−div) c is therefore a bijection as well. We refer to Ψ as the BBP parametrization of R(inf−div) c yields a distribution ν = Ψ(σ1, σ2) ∈ R(inf−div) Every choice of parameters σ1, σ2 ∈ P + and every ν ∈ R(inf−div) arises from a unique pair σ1, σ2. c c c . , We emphasize that the bijection Ψ works in a really straightforward way -- the coefficients of the η-series of σ1 and σ2 give the determining sequences of ν = Ψ(σ1, σ2). It is actually worth recording a direct consequence of this fact, as follows. Notation 6.2. We denote by E + that f = ησ for some σ ∈ P + c the collection of those series f ∈ C[[z]] with the property c (where σ is, a fortiori, uniquely determined). ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 23 Proposition 6.3. Let ν ∈ Dc(1, ∗) be an R-diagonal distribution, and let (αn)∞ (βn)∞ ν is ⊞-infinitely divisible if and only if both f and g belong to E + c . n=1 be its determining sequences. Set f (z) =P∞ n=1 αnzn and g(z) =P∞ n=1 and n=1 βnzn. Then Proof. If ν ∈ R(inf−div) g = ησ2, and so f, g ∈ E + for some σ1, σ2 ∈ P + c , then ν = Ψ(σ1, σ2) for some σ1, σ2 ∈ P + c . Conversely, suppose that f, g ∈ E + c , hence f = ησ1 and c , so f = ησ1 and g = ησ2 , and the c hence ν ∈ R(inf−div) definition of Ψ shows that eν has the same determining sequences as ν. This forces ν =eν, c . Then the distribution eν := Ψ(σ1, σ2) belongs to R(inf−div) The criterion provided by Proposition 6.3 is useful because one can (following the work in terms of the associated analytic functions. We will in [1]) characterize the series from E + c follow up on this in the application presented in Section 7. (cid:3) . c Since we are dealing with free probabilistic structures, it is natural to ask what is the description of the BBP parametrization Ψ in terms of R-transforms. Recall (Remark 3.4) that an R-diagonal ∗-distribution ν ∈ Dc(1, ∗) is uniquely determined by the R-transforms RZZ ∗, RZ ∗Z ∈ C[[z]]. The following result thus provides an alternative characterization of what is Ψ(σ1, σ2). Theorem 6.4. Let σ1, σ2 ∈ P + of Z ∗Z in the ∗-probability space (ChZ, Z ∗i, ν) are described as follows: c and set ν = Ψ(σ1, σ2). Then the R-transforms of ZZ ∗ and (6.4) RZZ ∗(z) = RB(σ1)(cid:0)z(1 + MB(σ2)(z))(cid:1) , RZ ∗Z (z) = RB(σ2)(cid:0)z(1 + MB(σ1)(z))(cid:1) , where B(σ1) and B(σ2) indicate the original BBP bijection (as discussed in Remark 4.5). Proof. We set µ := Φ(σ1, σ2), so ν is R-diagonal, µ is η-diagonal, and B (1,∗)(µ) = ν. Thus Rν(z, z∗) = ηµ(z, z∗) = αn(zz∗)n + βn(z∗z)n, ∞Xn=1 ∞Xn=1 n=1 and (βn)∞ where (αn)∞ n=1 are the (common) determining sequences for µ and for ν. By the definition of the bijection Φ in Theorem 5.2, σ1 has the same moments as the element ZZ ∗ in the noncommutative probability space (ChZ, Z ∗i, µ). This implies that ησ1 = ηZZ ∗ , n=1 αnzn. In a similar way we and then Proposition 2.13 gives us the formula ησ1(z) =P∞ find that ησ2(z) =P∞ Consider now the probability measures B(σ1), B(σ2) ∈ Pc. The definition of B implies n=1 βnzn. RB(σ1)(z) = ησ1 (z) = αnzn, RB(σ2)(z) = ησ2(z) = βnzn. ∞Xn=1 ∞Xn=1 But then Proposition 3.5 applies to the R-diagonal ∗-distribution ν and yields (6.4). (cid:3) As a consequence of Theorem 6.4, we obtain a natural connection between the notions of ⊞-infinite divisibility in Dc(1, ∗) and in Pc. This is stated in the next corollary. The converse of the corollary fails even in the tracial framework (see Remark 6.9 below). Corollary 6.5. Let ν ∈ Dc(1, ∗) be R-diagonal and let τ1, τ2 ∈ P + c be the distributions of ZZ ∗ and Z ∗Z in the ∗-probability space (ChZ, Z ∗i, ν) (as discussed in Definition 5.1). If ν is ⊞-infinitely divisible in Dc(1, ∗), then τ1 and τ2 are ⊞-infinitely divisible in Pc. Proof. By symmetry, it suffices to show that τ1 is ⊞-infinitely divisible. According to [14, Theorem 4.3], a compactly supported Borel probability measure on R is ⊞-infinitely divisible 24 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI if and only if its R-transform can be extended to an analytic self-map of the upper half- plane C+. Suppose that ν = Φ(σ1, σ2), where σ1, σ2 ∈ P + c . The R-transform Rτ1, which is the same as RZZ ∗, is given by the first Equation (6.4). By [5, Proposition 6.1], the moment series of the probability measure B(σ2) can be extended analytically to C+ and this extension satisfies z ∈ C+ ⇒ z(1 + MB(σ2)(z)) ∈ C+. Finally, since B(σ1) is ⊞-infinitely divisible, [14, Theorem 4.3] assures us that RB(σ1) extends analytically to a self-map of C+. We conclude that for every z ∈ C+, RB(σ1) is defined at z(1 + MB(σ2)(z)), and that z 7→ RB(σ1)(cid:0) z(1 + MB(σ2))(z)(cid:1) is an analytic self-map on C+, as required. (cid:3) In the remainder of this section, we discuss the KMS example. In this special case one can process further the formulas from Theorem 6.4 and arrive at explicit formulas (stated in Proposition 6.8) for the distributions of ZZ ∗ and of Z ∗Z in terms of the probability measures σ1, σ2 that parametrize ν. These formulas call on some commonly used operations from the free harmonic analysis of P + c , that are reviewed in the following remark. Remark 6.6. (Some elements of free harmonic analysis on P + c .) (1) Measures σ ∈ Pc have free additive convolution powers with real exponent t ∈ [1, ∞). More precisely, for every σ ∈ Pc and t ∈ [1, ∞), there exists a unique measure τ ∈ Pc such that Rτ = tRσ (see [11, pp. 228-231]). This measure τ is denoted σ⊞t. When t is an integer, σ⊞t is simply the t-fold convolution σ ⊞ · · · ⊞ σ. The argument in [11, pp. 228-231] also shows that σ⊞t ∈ P + for c all t ∈ [1, ∞) if σ ∈ P + c . The analogous result for Boolean convolution provides for every σ ∈ Pc and t ∈ (0, ∞) a Boolean convolution power σ⊎t ∈ Pc such that ησ⊎t = tησ (see [13, Theorem 3.6]). As in the free case, σ⊎t ∈ P + c (see, for instance, the operator model c constructed in [2, Proposition 4.8]). for every t ∈ (0, ∞) if σ ∈ P + (2) The original BBP bijection B : Pc → P (inf−div) c (Remark 4.5) can be expressed using convolution powers, by the formula B(σ) =(cid:0)σ⊞2(cid:1)⊎1/2 , σ ∈ Pc, which was proved in [3, Theorem 1.2]). The facts reviewed in (1) above imply that B(σ) ∈ P + c for every σ ∈ P + c . (3) Free multiplicative convolution ⊠ is another binary operation defined on the set P + c . This operation corresponds to the product of free random variables. Quite remarkably, BP + c was shown in [3, Remark 3.9] to be a homomorphism for ⊠, that is, B(σ ⊠ σ′) = B(σ) ⊠ B(σ′), σ, σ′ ∈ P + c . (4) The free counterpart of the standard Poisson distribution is the Marchenko-Pastur distribution Π1 (also known as the the free Poisson distribution). This distribution is supported on the interval [0, 4] and it is Lebesgue absolutely continuous with density Its R-transform is dΠ1(t)/dt = 1 2πp(4 − t)/t, 0 ≤ t ≤ 4. RΠ1(z) = z/(1 − z), ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 25 and a simple calculation using the definition of B shows that (6.5) Π1 = B(cid:18) 1 2 (δ0 + δ2)(cid:19) . A useful property of Π1 is that it converts moment series into R-transforms via the formula (6.6) Rσ⊠Π1 = Mσ, σ ∈ P + c . See, for instance, [11, Propositions 17.2 and 17.4]. Remark 6.7. Let σ ∈ P + (αn)∞ R(inf−div) n=1 and (βn)∞ satisfy c c and let t > 0 be a real number. The determining sequences n=1 of the infinitely divisible R-diagonal ∗-distribution ν := Ψ(σ⊎t, σ) ∈ βnzn = ησ(z), αnzn = ησ⊎t (z) = tησ(z). ∞Xn=1 ∞Xn=1 Thus ν satisfies the KMS condition with parameter t: αn = tβn, n ∈ N (Definition 3.6). Proposition 6.8. With the notation of the preceding remark, let τ1, τ2 ∈ P + c be the dis- tributions of ZZ ∗ and of Z ∗Z in the noncommutative probability space (ChZ, Z ∗i, ν) (as discussed in Definition 5.1). Then (6.7) τ1 = (B(σ) ⊠ Π1)⊞t and τ2 =(cid:0)B(σ)⊞t ⊠ Π1(cid:1)⊞1/t . Proof. The two formulas in (6.7) have similar proofs. We only verify the first one. Since τ1 is the distribution of ZZ ∗, we have Rτ1 = RZZ ∗, and Theorem 6.4 yields (6.8) where σ1 = σ⊎t and σ2 = σ. The relation σ1 = σ⊎t and (1.2) imply B(σ1) = B(σ)⊞t, and hence RB(σ1) = t RB(σ). The equality (6.8) can be continued as follows: Rτ1(z) = RB(σ1)(cid:0)z(1 + MB(σ2)(z))(cid:1) , Rτ1(z) = t · RB(σ)(cid:0)z(1 + MB(σ)(z))(cid:1) = t · MB(σ)(z) (by (3.4)) = t · RB(σ)⊠Π1 (z) (by (6.6)) = R(B(σ)⊠Π1)⊞t (z). Thus the probability measures τ1 and (B(σ) ⊠ Π1)⊞t are equal because they have the same R-transform. (cid:3) Remark 6.9. (Tracial case.) In the special case when t = 1, the preceding proposition reduces to (6.9) τ1 = τ2 = B(σ) ⊠ Π1. Using (6.5) and invoking the multiplicativity of B (Remark 6.6(3)), we can rewrite (6.9) as (6.10) τ1 = τ2 = B(cid:18)σ ⊠ 1 2 (δ0 + δ2)(cid:19) . This confirms the fact (Corollary 6.5) that τ1 and τ2 are ⊞-infinitely divisible in Pc. c that cannot be written as σ ⊠ 1 We conclude with an argument showing that the converse of Corollary 6.5 does not hold. 2 (δ0 + δ2) for any σ ∈ P + c . 3 (δ0 + δ1 + δ2) is such a distribution.) Let ν ∈ Dc(1, ∗) be the tracial R-diagonal ∗-distribution defined by the requirement that the common distribution of ZZ ∗ Choose a distribution eσ ∈ P + (For instance, eσ = 1 and Z ∗Z in (ChZ, Z ∗i, ν) is equal to B(eσ ) (see [11, Proposition 15.13] for an argument that 26 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI ν exists). The distributions of ZZ ∗ and Z ∗Z are ⊞-infinitely divisible in Pc, by construction. We show that ν is not ⊞-infinitely divisible in Dc(1, ∗). Suppose, to get a contradiction, that ν is ⊞-infinitely divisible. Then ν = Ψ(σ, σ) for some σ ∈ P + 2 (δ0 + δ2)(cid:1), and thuseσ = σ ⊠ 1 (6.10) yields B(eσ) = B(cid:0)σ ⊠ 1 contrary to the choice ofeσ. before Definition 3.6. Indeed, it is immediate that in this case the series P∞ P∞ Example 6.10. (λ-circular distribution.) Let λ > 0 be a parameter. If in the setting of Remark 6.7 and Proposition 6.8 we take σ = δ1 (Dirac mass at 1) and t = λ, then the resulting ∗-distribution ν ∈ R(inf−div) is the λ-circular distribution mentioned right n=1 βnzn and n=1 αnzn from Remark 6.7 are reduced to ηδ1(z) = z and respectively to tηδ1(z) = λz; hence we have α1 = λ, β1 = 1 and αn = βn = 0 for all n ≥ 2, as required in the definition of the λ-circular distribution. 2 (δ0 + δ2) because B is injective, c . Since τ1 = τ2 = B(eσ), c In this example, the formulas indicated in Proposition 6.8 for the distributions of ZZ ∗ and of Z ∗Z give free Poisson distributions. In order to make this precise, we need to review another bit of notation: for any two parameters p, q > 0 one has a free Poisson distribution of rate p and jump size q, which we will denote as Πp;q, and which appears in the free analogue of the Poisson limit theorem (see e.g. Proposition 12.11 in [11]). The Marchenko- Pastur distribution reviewed in Remark 6.6(4) corresponds to p = q = 1 (so "Π1" from there becomes "Π1;1"). For general p, q > 0, the formula given in Remark 6.6(4) for the R-transform of Π1 extends to RΠp;q (z) = pqz 1 − qz . Returning to the example of the λ-circular distribution, an immediate processing of the formulas (6.4) from Theorem 6.4 gives us that the R-transforms of ZZ ∗ and of Z ∗Z in the noncommutative probability space (ChZ, Z ∗i, ν) are RZZ ∗(z) = λz 1 − z , RZ ∗Z (z) = z 1 − λz . For our example, this shows that the distributions τ1 and τ2 appearing in (6.7) (Proposition 6.8) are free Poisson distributions: τ1 = Πλ;1 and τ2 = Π1/λ;λ. 7. Stability of R(inf−div) c under free multiplicative convolution Remark 7.1. In this section we consider the operation ⊠ on Dc(1, ∗), which follows the multiplication of ∗-free random variables (cf. Definition 4.1(3)). One has the remarkable fact that whenever µ, µ′ ∈ Dc(1, ∗) and at least one of µ, µ′ is R-diagonal, it follows that µ ⊠ µ′ is R-diagonal as well (see [11, Proposition 15.8]). If we make the additional assumption that both µ and µ′ are R-diagonal, then we have explicit formulas for the determining sequences of µ ⊠ µ′ in terms of the determining sequences of µ and of µ′. To be precise, denote the determining sequences of µ by (αn)∞ n=1. n=1, and those of µ′ by (α′ n=1, (βn)∞ n=1, (β′ n)∞ n)∞ ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 27 Tthen the determining sequences (bαn)∞ Xπ⊔ρ∈N C(2n) n=1, (bβn)∞ bαn = π={V1,...,Vp}∈N C(1,3,...,2n−1) with 1∈V1, and ρ={W1,...,Wr}∈N C(2,4,...,2n) (7.1) n=1 of µ ⊠ µ′ are given by: αV1βV2 · · · βVpα′ W1 · · · α′ Wr, bβn = Xπ⊔ρ∈N C(2n) π={V1,...,Vp}∈N C(1,3,...,2n−1) with 1∈V1, and ρ={W1,...,Wr}∈N C(2,4,...,2n) β′ V1α′ V2 · · · α′ VpβW1 · · · βWr.  The formulas (7.1) were proved in [7, Proposition 3.9]. They can also be rephrased in terms of equations for power series, as shown in the next proposition. The formulas (7.2) in the proposition have appeared before (but only as a conjecture, without proof), in [9, Section 5.3]. For the reader's convenience, we include the proof of how (7.2) is derived out of (7.1). Proposition 7.2. With the notation of Remark 7.1, suppose that we have elements a, b, a′, b′ in a noncommutative probability space (A, ϕ) such that {a, b} is free from {a′, b′} and such that Ra(z) = Ra′(z) = αnzn, Rb(z) = α′ nzn, Rb′(z) = ∞Xn=1 ∞Xn=1 βnzn, β′ nzn. ∞Xn=1 ∞Xn=1  n=1bαnzn =(cid:16)Ra ◦ Rh−1i P∞ n=1bβnzn =(cid:16)Rb′ ◦ Rh−1i P∞ a′ b ◦ Mba′(cid:17) (z), ◦ Ma′b(cid:17) (z). Assume moreover that β1 6= 0 6= α′ relative to composition. Then: 1, so the series Rb and Ra′ have inverses Rh−1i b and Rh−1i a′ (7.2) Proof. The second equation in (7.2) follows from the first one if we substitute b′, a′, b for a, b, a′, respectively. To prove the first equation, we fix an n ∈ N and we suitably structure the formula for bαn provided in (7.1). Let us also momentarily fix an m ≤ n and a set V1 = {2i1 − 1, . . . , 2im − 1}, where 1 = i1 < · · · < im ≤ n. Denote nk = ik+1 − ik, k = 1, . . . , m, where im+1 = n + 1. Note that n1 + · · · + nm = n and that V1 can recovered from n1, . . . , nm. Use the moment-cumulant formula as in the proof of Proposition 3.5 (using the cumulant functionals and the fact that the mixed cumulants of a′ and b vanish on account of freeness) to obtain Xπ⊔ρ∈N C(2n) π={V1,...,Vp}∈N C(1,3,...,2n−1) ρ={W1,...,Wr}∈N C(2,4,...,2n) αV1βV2 · · · βVpα′ W1 · · · α′ Wr = αm ϕ(a′(ba′)nk−1). mYk=1 Letting V1 vary, (7.1) yields, for the n ∈ N that we had fixed: nXm=1 αm Xn1+···+nm=n mYk=1 bαn = ϕ(a′(ba′)nk−1). 28 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI We now let n vary in N, and get that (7.3) where (7.4) ∞Xn=1 ∞Xn=1bαnzn = ∞Xn=1 g(z) = αm(g(z))m = Ra(g(z)), ϕ(a′(ba′)n−1)zn ∈ C[[z]]. It remains to show that the series g introduced in (7.4) is equal to Rh−1i ◦ Mba′ or, equivalently, that one has Rb ◦ g = Mba′. To see this, apply again the moment-cumulant formula (using, as in [11, Theorem 14.4] the fact that b is free from a′) to obtain b (7.5) ϕ((ba′)n) = βV1 · · · βVpα′ W1 · · · α′ Wr. Xπ⊔ρ∈N C(2n) π={V1,...,Vp}∈N C(1,3,...,2n−1) ρ={W1,...,Wr}∈N C(2,4,...,2n) In (7.5) we can list the blocks of π such that 1 ∈ V1. A similar argument to the one used above to structure the formula for bαn shows now that mYk=1 βm Xn1+···+nm=n nXm=1 ϕ((ba′)n) = and this implies the desired relation Mba′ = Rb ◦ g. ϕ(a′(ba′)nk−1), n ∈ N, (cid:3) Remark 7.3. With the notation of the preceding proposition, suppose that βn = 0 for every n ∈ N. Then the only non-zero term in the first equality in (7.1) corresponds to The next corollary presents a reformulation of (7.2) which has the advantage that it π = 1n and ρ = 0n, and therefore bαn = αn(α′ introduces in discussion two power series F and eF , related with the subordination results Corollary 7.4. In the framework of Proposition 7.2, we have of [6]. 1)n. (7.6) where F = M h−1i b (P∞ n=1bαnzn = Ra (F (z) (1 + Mb(F (z)))) n=1bβnzn = Rb′(cid:16)eF (z)(cid:16)1 + Ma′(eF (z))(cid:17)(cid:17) , P∞ ◦ Mba′ and eF = M h−1i ◦ Ma′b. a′ Proof. By symmetry, it suffices to prove the first of the two equations. Using (7.2), we see that we must verify the identity (Rh−1i (7.7) b ◦ Mba′)(z) = F (z)(cid:0)1 + Mb(F (z))). Recalling the assumption that ϕ(b) 6= 0, the functional equation Mb(z) = Rb(z(1 + Mb(z))) can be rewritten as (7.8) Rh−1i b (w) = (1 + w)M h−1i b (w) (see [11, Remark 16.18]). Substitute Mba′ for w in (7.8) to find that (cid:0)Rh−1i b ◦ Mba′(cid:1)(z) = (1 + Mba′(z)) ·(cid:0)M h−1i b ◦ Mba′(cid:1)(z) = (1 + Mba′ (z)) · F (z). Finally, the definition of F implies that Mba′ (z) = Mb(F (z)), and using this equality in the right hand side of the preceding equality yields (7.7). (cid:3) ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 29 The following lemma is an immediate consequence of the definition of Ψ (Remark 6.1). Lemma 7.5. Consider a distribution ν ∈ R(inf−div) n=1 be its determining sequences. There exist positive elements a, b in a C ∗-probability space (A, ϕ) such that n=1 and (βn)∞ , and let (αn)∞ c Ra(z) = αnzn, and Rb(z) = βnzn. ∞Xn=1 Moreover, the distributions of a and b are ⊞-infinitely divisible. Proof. Write ν = Ψ(σ1, σ2), with σ1, σ2 ∈ P + c . Then ∞Xn=1 ∞Xn=1 αnzn = ησ1(z) = RB(σ1)(z), where B(σ1) is a ⊞-infinitely divisible distribution in P + c (cf. Remark 6.6(2)). Thus taking a to be a positive element with distribution B(σ1) in some C ∗-probability space will fulfill the required conditions. The argument for b is similar. (cid:3) In reference to the set of power series E + c introduced in Notation 6.2, we record a result which follows easily from [1, Proposition 2.2]. Proposition 7.6. A series f ∈ C[[z]] belongs to the set E + c following three conditions: if and only if it satisfies the (i) f has real coefficients; (ii) f has positive convergence radius; (iii) f can be extended to an analytic map (still denoted f ) of C+ into C+ such that f (0) = 0 and Arg(z) ≤ Arg(f (z)) for z ∈ C+. (cid:3) is not identically zero. Corollary 7.7. Suppose that the series f (z) = P∞ Then α1 > 0. n=1 αnzn ∈ E + c Proof. Let n be the smallest integer such that αn 6= 0 and suppose, to get a contradiction, that either n > 1 or n = 1 and αn < 0. Choose γ ∈ C+ such that γ = 1 and ℑ(γnαn) < 0. We have limr↓0(f (rγ)/rn)) = γnαn, and therefore ℑf (rγ) < 0 for sufficiently small r, contrary to Proposition 7.6(iii). (cid:3) We are now ready for the main result of this section. c . n)∞ c n)∞ Theorem 7.8. For every ν, ν′ ∈ R(inf−div) Proof. Let (αn)∞ n=1) denote the determining sequences of ν (respectively, ν′). Two applications of Lemma 7.5, combined with a free product construction, allow us to construct a C ∗-probability space (A, ϕ) and positive elements a, b, a′, b′ ∈ A such that we have ν ⊠ ν′ ∈ R(inf−div) n=1 (respectively (α′ n=1 and (βn)∞ n=1 and (β′ n=1 βnzn, n=1 β′ (c) {a, b} is free from {a′, b′}. (a) Ra(z) =P∞ (b) Ra′(z) =P∞ We know from Remark 7.1 that ν ⊠ ν′ is an R-diagonal distribution in Dc(1, ∗). Let (bαn)∞ and (bβn)∞ n=1 αnzn and Rb(z) =P∞ nzn and Rb′(z) =P∞ ∞Xn=1bαnzn, n=1 denote the determining sequences of ν ⊠ ν′, and set ∞Xn=1bβnzn. bf (z) := bg(z) := nzn, and n=1 α′ n=1 30 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI because α′ yields α′ c . By symmetry, it suffices to show that n=1 αn(α′ n=1 αnzn belongs to E + We dispose first of the simple case in which β1 = 0. Corollary 7.7 yields βn = 0 for all 1z)n. The desired conclusion follows c . Similarly, if α′ 1 = 0, Corollary 7.7 By Proposition 6.3, we have to prove that bf ,bg ∈ E + c , and this is done by verifying that bf satisfies conditions (i) -- (iii) of Proposition 7.6. bf ∈ E + n ∈ N, and Remark 7.3 implies that bf (z) =P∞ 1 ≥ 0 and the seriesP∞ n = 0 for all n ∈ N, and then (7.1) implies that bf = 0. It remains to show that bf ∈ E + that bf = Ra (F (z) (1 + Mb(F (z)))), where F = M h−1i Therefore bf ∈ E + composition of the three power series Ra(z), z(1 + Mb(z)), and F (z). We know that Ra ∈ E + c by c . [5, Proposition 6.1]. Proposition 7.6 shows that the set E + is closed under composition. c (cid:3) c , thus concluding the proof. Corollary 7.9. Suppose that ν ∈ R(inf−div) C ∗-probability space. Then the ∗-distribution of an belongs to R(inf−div) In other words, bf is the c . The series z(1 + Mb(z)) also belongs to E + is the ∗-distribution of an element a in some 1. In this case, Corollary 7.4 shows It was proved in [6] that F ∈ E + c when β1 6= 0 6= α′ for every n ∈ N. c ◦ Mba′. b c Proof. This follows from Theorem 7.8 and the known fact [7, Proposition 3.11] that the distribution of an is equal to ν ⊠n. (cid:3) We conclude the section by looking again at the KMS example, and by describing explic- itly the BBP parametrization for the powers of a λ-circular element. Remark 7.10. (1) Suppose that ν, ν′ ∈ Dc(1, ∗) are R-diagonal and satisfy the KMS condition with parameters t, t′ ∈ (0, ∞), respectively. Consider the ∗-distribution ν ⊠ ν′, which is R-diagonal as well (see Remark 7.1). We claim that ν ⊠ ν′ also satisfies the KMS condition, with parameter tt′. Using the same notations for determining sequences as in Remark 7.1, this claim amounts to the fact that bαn = (tt′)bβn for every n ∈ N. In order to prove this, we replace αV1 and α′ (7.1) to obtain W1, respectively, in the first formula W1 by tβV1 and tβ′ β′ W1α′ W2 · · · α′ WlβV1βV2 · · · βVk. bαn = (tt′) Xπ⊔ρ∈N C(2n) π={V1,...,Vk}∈N C(1,3,...,2n−1) ρ={W1,...,Wl}∈N C(2,4,...,2n) 1∈V1,2∈W1 To see that the last sum equals bβn, we observe that pairs (π, ρ) as above are in a bijective correspondence with pairs (eπ,eρ) such that eπ ⊔ eρ ∈ N C(2n) and eπ = {fW1, . . . ,fWl} ∈ N C(1, 3, . . . , 2n − 1) andeρ = {eV1, . . . ,eVk} ∈ N C(2, 4, . . . , 2n). Indeed,eρ andeπ are obtained aseπ ⊔eρ = γ−1 2n (π ⊔ ρ), where we use the permutation γ2n from the proof of Proposition 3.7. Thus, the sum above is equal to β′ fW1 α′ fW2 · · · α′ fWl β eV1 · · · β eVk, Xπ′⊔ρ′∈N C(2n) eπ={fW1,...,fWl}∈N C(1,3,...,2n−1) eρ={ eV1,..., eVk}∈N C(2,4,...,2n) 1∈fW1,2∈ eV1 and this equals bβn by the second formula (7.1). (2) Now fix a real number λ > 0 and consider the λ-circular distribution ν = Ψ(δλ, δ1), as in Example 6.10. If a is an element in some ∗-probability space such that the ∗-distribution ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 31 of a is equal to ν, then we will say that a is a λ-circular element. Such elements do of course exist, for instance we can just take a = Z in the ∗-probability space (ChZ, Z ∗i, ν). If a is a λ-circular element, then Theorem 7.8 and Corollary 7.9 tell us that every power ak has ∗-distribution ν ⊠k ∈ R(inf−div) . Moreover, part (1) of the present remark assures us that ν ⊠k satisfies the KMS condition with parameter λk. Hence for every k ∈ N we have a BBP parametrization of the form c ν ⊠k = Ψ(σ⊎λk c . For k = 1, we know from Example 6.10 that σ1 is , σk), for some probability measure σk ∈ P + the Dirac mass δ1. The next proposition gives a way of describing σk for k ≥ 2. Proposition 7.11. Let λ and (σk)∞ probability measures with finite support (τk)∞ k=1 defined by k=1 be as above, and consider on the other hand the k Then one has (7.9) τk := λk 1 + λk δ0 + 1 1 + λk δ1+λk , k ∈ N. σk = τ1 ⊠ · · · ⊠ τk−1, k ≥ 2. Proof. As in Remark 7.10(2), we use the notation ν for the λ-circular distribution. We fix a k ∈ N and invoke Proposition 7.2 in the special case in which the ∗-distributions µ, µ′ considered there are ν ⊠k and ν, respectively. The power series (7.10) βnzn, β′ nzn, ∞Xn=1 ∞Xn=1 ∞Xn=1bβnzn σ1 and σk+1, respectively. (For instance the equalityP∞ σ1 = δ1, for the second power series in (7.10) we actually haveP∞ from Proposition 7.2 are equal in this case to the η-series of the probability measures σk, n=1 βnzn = ησk (z) follows from the , σk).) Note that, since The notation of Proposition 7.2 also include some non-commutating random variables comments at the end of Remark 6.1 and the fact that ν ⊠k = Ψ(σ⊎λk k n=1 β′ nzn = z. a, b, a′, b′, where b is such that (7.11) Rb(z) = βnzn = ησk (z) = RB(σk)(z). ∞Xn=1 From (7.11) we infer that the distribution of b is B(σk). Similar reasoning, based on the formulas Rb′(z) = z and Ra′(z) = λz, leads to the fact that a′ and b′ have distributions δλ and δ1, respectively. As a consequence, we may assume without loss of generality that a′ = λ and b′ = 1 in their noncommutative probability space. We are interested in the second relation (7.2) from Proposition 7.2. Due to the very simple form of Ra′ and Rb′, this equation simplifies to (7.12) ∞Xn=1bβnzn = 1 λ Mλb(z). The same argument as used in (7.11) shows that the left-hand side of (7.12) is equal to RB(σk+1)(z). On the right-hand side of (7.12) we perform the obvious transformation Mλb(z) = Mb(λz) = MB(σk )(λz), and this leads us to a direct connection between σk and σk+1: (7.13) RB(σk+1)(z) = 1 λ MB(σk)(λz). 32 H. BERCOVICI, A. NICA, M. NOYES, AND K. SZPOJANKOWSKI In order to make use of (7.13), it is convenient to resort to another well-known transform of free probability, the S-transform. For a probability measure σ ∈ Pc with non-vanishing mean, one defines the S-transform of σ as the power series Sσ(z) = 1 z Rh−1i σ (z) = z + 1 z M h−1i σ (z) (see, for instance, [11, Definition 18.15 and Remark 18.16 on p. 294]). Some straightforward processing of Equation (7.13) (multiply both sides by λ, take inverses under composition, and write the resulting series in terms of the suitable S-transforms) then leads to the formula (7.14) SB(σk+1)(z) = 1 1 + λz SB(σk)(λz). The formula (7.14) was obtained for a fixed (but arbitrary) k ∈ N. We now unfix k and use a straightforward induction argument, with base case SB(σ1)(z) = Sδ1(z) = 1, in order to infer that (7.15) SB(σk)(z) = 1 1 + λjz k−1Yj=1 , ∀ k ∈ N. It remains to make the connection to the τk indicated in the statement of the proposition. For every j ∈ N, an elementary calculation shows that B(τj) is the free Poisson distribution Π1/λj ;λj , where the notation "Πp;q" is as in Example 6.10. Another elementary calculation shows that the S-transform of Π1/λj ;λj is 1/(1 + λjz). Thus the right-hand side of (7.15) can be written as SB(τ1)(z) SB(τ2)(z) · · · SB(τk−1)(z). Now, the S-transform is multiplicative with respect to the operation ⊠ ([11, Corollary 18.17]). Since B is multiplicative as well (Remark 6.6(3)), the observations made in the preceding paragraph lead to the formula SB(σk) = SB(τ1 ⊠···⊠τk−1), k ≥ 2. The required Equation (7.9) follows from here, since B is injective and since a probability measure with non-vanishing mean is uniquely determined by its S-transform. (cid:3) References [1] S.T. Belinschi, H. Bercovici. Partially defined semigroups relative to free multiplicative convolution, International Mathematics Research Notices (2005), 65-101. [2] S.T. Belinschi, A. Nica. η-series and a Boolean Bercovici-Pata bijection for bounded k-tuples, Advances in Mathematics 217 (2008), 1-41. [3] S.T. Belinschi, A. Nica. On a remarkable semigroup of homomorphisms with respect to free multiplica- tive convolution, Indiana University Mathematics Journal 57 (2008), 1679-1713. [4] H. Bercovici, V. Pata. Stable laws and domains of attraction in free probability theory. With an appendix by P. Biane, The density of free stable distributions, Annals of Mathematics 149 (1999), 1023-1060. [5] H. Bercovici, D. Voiculescu. Free convolution of measures with unbounded support, Indiana University Mathematics Journal 42 (1993), 733-773. [6] P. Biane. Processes with free increments, Mathematische Zeitschrift 227 (1998), 143-174. [7] B. Krawczyk, R. Speicher. Combinatorics of free cumulants, Journal of Combinatorial Theory, Series A 90 (2000), 267-292. [8] F. Lehner. Cumulants in noncommutative probability I. Noncommutative exchangeable systems. Math- ematische Zeitschrift 248 (2004), 67-100. [9] A. Nica, D. Shlyakhtenko, R. Speicher. R-diagonal elements and freeness with amalgamation, Canadian Journal of Mathematics 53 (2001), 355-381. [10] A. Nica, R. Speicher. R-diagonal pairs -- a common approach to Haar unitaries and circular elements, Fields Institute Communications 12 (1997), 149-188. ETA-DIAGONALS AND INFINITE DIVISIBILITY FOR R-DIAGONALS 33 [11] A. Nica, R. Speicher. Lectures on the combinatorics of free probability, London Mathematical Society Lecture Note Series 335, Cambridge University Press, 2006. [12] D. Shlyakhtenko. Free quasi-free states, Pacific Journal of Mathematics, 177 (1997), 329-368. [13] R. Speicher, R. Woroudi. Boolean convolution, Fields Institute Communications 12 (1997), 267-279. [14] D. Voiculescu. Addition of certain noncommuting random variables, Journal of Functional Analysis 66 (1986), 323 -- 346. Hari Bercovici: Department of Mathematics, Indiana University, Bloomington, Indiana, USA. E-mail address: [email protected] Alexandru Nica: Department of Pure Mathematics, University of Waterloo, Ontario, Canada. E-mail address: [email protected] Michael Noyes: Department of Mathematics, Bard High School Early College, New York, New York, USA. E-mail address: [email protected] Kamil Szpojankowski: Department of Pure Mathematics, University of Waterloo, Ontario, Canada, and Faculty of Mathematics and Information Science, Warsaw University of Tech- nology, Poland. E-mail address: [email protected], [email protected]
1110.4476	2	1110	2013-01-10T11:49:52	The Weyl group of the Cuntz algebra	[ "math.OA", "math.DS", "math.FA", "math.GR" ]	The Weyl group of the Cuntz algebra O_n, with n finite, is investigated. This is (isomorphic to) the group of polynomial automorphisms of O_n, namely those induced by unitaries that can be written as finite sums of words in the canonical generating isometries and their adjoints. A necessary and sufficient algorithmic combinatorial condition is found for deciding when a polynomial endomorphism restricts to an automorphism of the canonical diagonal MASA. Some steps towards a general criterion for invertibility of such endomorphisms on the whole of O_n are also taken. A condition for verifying invertibility of a certain subclass of polynomial endomorphisms is given. First examples of polynomial automorphisms of O_n not inner related to permutative ones are exhibited, for every n. In particular, the image of the Weyl group in the outer automorphism group of O_n is strictly larger than the image of the reduced Weyl group analyzed in previous papers. Results about the action of the Weyl group on the spectrum of the diagonal are also included.	math.OA	math	The Weyl Group of the Cuntz Algebra Roberto Conti∗, Jeong Hee Hong†and Wojciech Szyma´nski*‡ Abstract The Weyl group of the Cuntz algebra On is investigated. This is (isomorphic to) the group of polynomial automorphisms λu of On, namely those induced by unitaries u that can be written as finite sums of words in the canonical generating isometries Si and their adjoints. A necessary and sufficient algorithmic combinato- rial condition is found for deciding when a polynomial endomorphism λu restricts to an automorphism of the canonical diagonal MASA. Some steps towards a gen- eral criterion for invertibility of λu on the whole of On are also taken. A condition for verifying invertibility of a certain subclass of polynomial endomorphisms is given. First examples of polynomial automorphisms of On not inner related to those induced by unitaries from the core UHF subalgebra are exhibited, for ev- ery n ≥ 2. In particular, the image of the Weyl group in the outer automorphism group of On is strictly larger than the image of the restricted Weyl group analyzed in previous papers. Results about the action of the Weyl group on the spectrum of the diagonal are also included. MSC 2010: Primary 46L40, 46L55, Secondary 37B10 Keywords: Cuntz algebra, MASA, automorphism, endomorphism, Cantor set ∗This research was supported through the programme "Research in Pairs" by the Mathematisches Forschungsinstitut Oberwolfach in 2011. †This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (Grant No. 2012R1A1A2039991). ‡Partially supported by the FNU Rammebevilling 'Operator algebras and applications', the Nord- Forsk Research Network 'Operator algebra and dynamics' (grant #11580), and the FNU Forskningspro- jekt 'Structure and Symmetry'. 1 1 Introduction Consider a finite alphabet {1, 2, . . . , n} with n ≥ 2 letters, and let W be the set of finite words on this alphabet. We say that two words are orthogonal if one is not the initial subword of the other. Let Σ be the collection of finite subsets of W consisting of mutually orthogonal words. We consider the set of n words {α1, α2, . . . , αn} (beginning with the same subword α and ending with all the distinct letters of the alphabet) equivalent to the single word α, and this extends to an equivalence relation on Σ. The set of equivalence {β1, . . . , βr}, with the property that both {α1, . . . , αr} and {β1, . . . , βr} are equivalent to the empty word. Such a U determines recursively a sequence of transformations classes is denoted eΣ. We fix U ∈ Σ, comprised of two ordered subsets: {α1, . . . , αr} and Tk : eΣ → eΣ such that: if γ = αjµ for some j then T1(γ) = βjµ, and if Tk−1(γ) = ναjµ for some j and a word ν of length k − 1 then Tk(γ) = νβjµ. Thus each transformation Tk is determined by a certain Turing machine, [10], and hence it is computable for any finite set of inputs. We are interested in the following stabilization problem of the recursive process Tk ◦ Tk−1 ◦ . . . ◦ T1 : eΣ → eΣ. For what U it holds that for each A ∈ eΣ there exists an m such that for all k ≥ m we have Tk ◦ . . . ◦ Tm ◦ . . . ◦ T1 = Tm ◦ . . . ◦ T1? We provide a surprisingly simple complete solution to this (suitably reformulated in more algebraic terms) stabilization problem in Theorem 3.7, below. We can reformulate the above described combinatorial setup in topological terms, as follows. Let Xn be the space of all (one-sided) infinite words. Then Xn is a Cantor set with the product topology and elements of eΣ are in bijective correspondence with its clopen subsets. Our stabilization problem is then equivalent to injectivity of a certain continuous map ψU : Xn → Xn determined naturally by U. Then, by the Gelfand duality, this problem is equivalent to surjectivity of a unital, injective ∗-homomorphism ψU : C(Xn) → C(Xn), dual to ψU . That is, we ask if ψU is a homeomorphism of Xn or, equivalently, if ψU is an automorphism of C(Xn). identity. In our setting, the element U gives rise to a unitary u = Pr Somewhat paradoxically, it is most natural to view this problem in the context of much larger and noncommutative Cuntz algebras On, [8]. These are C ∗-algebras generated by n isometries S1, . . . , Sn of a Hilbert space with ranges adding up to the βj in On, which in turn leads to a necessarily injective, unital ∗-endomorphism λu of On such that λu(Sj) = uSj for all j = 1, . . . , n. The C ∗-subalgebra Dn of On generated by ranges of all finite products of S1, . . . , Sn is maximal abelian in On and naturally isomorphic to C(Xn). The restriction of endomorphism λu to Dn coincides with ψU . Thus, our combinatorial stabilization problem is equivalent to the problem of surjectivity of λuDn. The question of surjectivity of λu itself is very interesting as well and closely related to the so called Weyl group of the Cuntz algebra. This last problem appears very difficult and an algorithm for deciding surjectivity of an arbitrary λu has not been found yet, although we make some headway towards its solution, below. j=1 Sαj S∗ 2 The present paper is a continuation of our investigations of the subgroup Aut(On, Dn) of automorphisms of On which globally preserve the canonical diagonal MASA Dn, and of related endomorphisms of On, [6, 4, 5, 11, 1, 2]. As shown in [9], the quotient of Aut(On, Dn) by its normal subgroup AutDn(On), consisting of those automorphisms which fix Dn point-wise, is discrete. Since AutDn(On) is a maximal abelian subgroup of Aut(On), [9], it is natural to call this quotient the Weyl group of On. The Weyl group contains a natural interesting subgroup corresponding to those automorphisms which also globally preserve the core UHF-subalgebra Fn of On, called the restricted Weyl group of On. It was shown in [2] that the image of the restricted Weyl group in the outer automorphism group of On can be embedded into the quotient of the automorphism group of the full two-sided n-shift by its center, and this embedding is surjective whenever n is prime. In the present article, we focus our attention on the (full) Weyl group. It was shown in [6] that the Weyl group is isomorphic with the group of those automorphisms λu ∈ Aut(On) whose corresponding unitaries u may be written as a sum of words in {Si, S∗ j }. (The collection of all such unitaries in On is denoted Sn.) The structure of the Weyl group is highly complicated. For example, it contains the Thompson F group in its intersection with Inn(On), [14]. Our main objective here is investigation of the structure of the Weyl group of On, its action on the diagonal MASA, and determining which unitaries u ∈ Sn give rise to automorphisms. The present paper is organized as follows. In section 2, we set up notation and review some basic facts on Cuntz algebras and their endomorphisms. In section 3, we study the restriction of an endomorphism λu, u ∈ Sn, to the diagonal Dn. We give an algorithmic criterion for λuDn to be an automorphism of Dn, Theorem 3.7. Its proof is combinatorial and involves equivalence of surjectivity of λuDn with the stabilization problem mentioned above. In section 4, we investigate the problem when λu is an automorphism of the entire On. In Proposition 4.3, we present a combinatorial procedure for deciding this question for a certain large class of unitaries u ∈ Sn. In section 5, we exhibit endomorphisms λu, u ∈ Sn, which are not inner related to the ones of the form λw with w a unitary in the core UHF-subalgebra Fn. In particular, we show with concrete examples that the image in Out(On) of the Weyl group is strictly larger then the image of the restricted Weyl group, Theorem 5.2 and Corollary 5.3. Finally, in section 6, we look at the action induced by λu on the space Xn, the spectrum of the diagonal Dn. We characterize homeomorphisms of Xn corresponding to automorphisms Ad(u), u ∈ Sn, and describe the fixed points in Xn for some exotic automorphisms λu. 2 Notation and preliminaries infinite C ∗-algebra generated by n isometries S1, . . . , Sn satisfying Pn If n is an integer greater than 1, then the Cuntz algebra On is a unital, simple, purely i = 1, [8]. We denote by W k n the set of k-tuples µ = (µ1, . . . , µk) with µm ∈ {1, . . . , n}, and by Wn the union ∪∞ n , where W 0 n = {0}. We call elements of Wn multi-indices. If µ ∈ W k k=0W k n then µ = k is the length of µ. For µ, ν ∈ Wn we write µ ≺ ν if µ is an initial subword n and µ ≺ ν, then we denote by ν − µ the word in W m−k of ν. If µ ∈ W k i=1 SiS∗ n , ν ∈ W m n 3 n obtained from ν by removing its initial segment µ. Also, if µ ∈ W k n then we denote by s(µ) its first letter, and by µ the word in W k−1 obtained from µ by removing s(µ). We denote by µ ∧ ν the collection of all non-empty words η such that both η ≺ µ and η ≺ ν. If µ = (µ1, . . . , µk) ∈ Wn then Sµ = Sµ1 . . . Sµk (S0 = 1 by convention) is an isometry with range projection Pµ = SµS∗ i i = 1, . . . , n} can be uniquely expressed as SµS∗ µ. Every word in {Si, S∗ ν , for µ, ν ∈ Wn [8, Lemma 1.3]. k=0F k We denote by F k n the C ∗-subalgebra of On spanned by all words of the form SµS∗ ν , µ, ν ∈ W k n , which is isomorphic to the matrix algebra Mnk (C). The norm closure Fn of ∪∞ n is the UHF-algebra of type n∞, called the core UHF-subalgebra of On, [8]. We denote by τ the unique normalized trace on Fn. The core UHF-subalgebra Fn is the fixed-point algebra for the gauge action γ : U(1) → Aut(On), such that γz(Sj) = zSj for z ∈ U(1) and j = 1, . . . , n. We denote by E the faithful conditional expectation from On onto Fn given by averaging with respect to the normalized Haar measure: E(x) =Zz∈U (1) γz(x)dz. For an integer m ∈ Z we denote O(m) subspace for γ. Then O(0) have O(m) n = FnSα and O(−m) = S∗ n αFn. n = Fn and for each positive integer m and each α ∈ W m := {x ∈ On : γz(x) = zmx, ∀z ∈ U(1)}, a spectral n we n The C ∗-subalgebra of On generated by projections Pµ, µ ∈ Wn, is a MASA (maximal abelian subalgebra) in On. We call it the diagonal and denote Dn. Every projection in Dn of the form Pα for some α ∈ Wn will be called standard. The spectrum of Dn is naturally identified with Xn -- the full one-sided n-shift space. For d ∈ Dn we denote by Md a map Md : Dn → Dn such that Md(x) = dx. As shown by Cuntz in [9], there exists the following bijective correspondence be- tween unitaries in On (whose collection is denoted U(On)) and unital ∗-endomorphisms of On (whose collection we denote End(On)). A unitary u ∈ U(On) determines an endomorphism λu by λu(Si) = uSi, i = 1, . . . , n. Conversely, if ρ : On → On is an endomorphism, then Pn i = u gives a unitary u ∈ On such that ρ = λu. Composition of endomorphisms corresponds to a 'convolution' multiplication of unitaries as follows: i=1 ρ(Si)S∗ λu ◦ λw = λλu(w)u. (1) If A is either a unital C ∗-subalgebra of On or a subset of U(On), then we denote λ(A) = {λu ∈ End(On) : u unitary in A} and λ(A)−1 = {λu ∈ Aut(On) : u unitary in A}. We denote by ϕ the canonical shift: ϕ(x) = nXi=1 SixS∗ i , x ∈ On. If we take u =Pn j then ϕ = λu. For all u ∈ U(On) we have Ad(u) = λuϕ(u∗). It is well-known that ϕ leaves Dn globally invariant. We denote by φ the standard left i,j=1 SiSjS∗ i S∗ 4 inverse of ϕ, defined as φ(x) = 1 n nXi=1 S∗ i xSi, x ∈ On. If u ∈ U(On) then for each positive integer k we denote uk = uϕ(u) · · · ϕk−1(u). (2) βu∗ β) = ukSαS∗ We often consider elements of On of the form w = P(α,β)∈J cα,βSαS∗ k stands for (uk)∗. If α and β are multi-indices of Here ϕ0 = id, and we agree that u∗ length k and m, respectively, then λu(SαS∗ m. This is established through a repeated application of the identity Six = ϕ(x)Si, valid for all i = 1, . . . , n and x ∈ On. β, where J is a finite collection of pairs (α, β) of words α, β ∈ Wn and cα,β ∈ C. We denote J1 = {α : ∃(α, β) ∈ J } and J2 = {β : ∃(α, β) ∈ J }. Of course, such a presentation (if it exists) is not unique, but once it is chosen then we associate with it two integers: ℓ = ℓ(J ) = max{α : (α, β) ∈ J } and ℓ′ = ℓ′(J ) = max{α, β : (α, β) ∈ J }. Note that if w ∈ Fn then w ∈ F ℓ ν , where ϕ(J ) = {((i, α), (β, i)) : i ∈ W 1 n , (α, β) ∈ J } and c(i,α),(β,i) = cα,β. Then ℓ(ϕ(J )) = ℓ(J ) + 1 and ℓ′(ϕ(J )) = ℓ′(J ) + 1. In particular, we consider the group Sn of those unitaries in On which can be written as finite sums of words, i.e. β. n. We have ϕ(w) = P(µ,ν)∈ϕ(J ) cµ,νSµS∗ in the form u = P(α,β)∈J SαS∗ Note that such a sum is a unitary if and only if Pα∈J1 Pα = 1 = Pβ∈J2 Pβ. We also n for the subgroups of Sn consisting of permutative n = Sn ∩ F k write Pn = Sn ∩ Fn and P k unitaries. For algebras A ⊆ B we denote by NB(A) = {u ∈ U(B) : uAu∗ = A} the normalizer of A in B and by A′ ∩ B = {b ∈ B : (∀a ∈ A) ab = ba} the relative commutant of A in B. We also denote by Aut(B, A) the collection of all those automorphisms α of B such that α(A) = A, and by AutA(B) those automorphisms of B which fix A point-wise. AutDn(On) is a normal subgroup of Aut(On, Dn), and the corresponding quotient is called the Weyl group of On. It was shown in [9] that the Weyl group is discrete, and more recently in [6] that it is isomorphic to λ(Sn)−1. The quotient of Aut(On, Dn) ∩ Aut(On, Fn) by AutDn(On) is called the restricted Weyl group of On. It is isomorphic to λ(Pn)−1, [6]. The image of λ(Sn)−1 in Out(On) is called the outer Weyl group of On and such image of λ(Pn)−1 is called the restricted outer Weyl group of On. As shown in [3, Theorem 3.7], the outer Weyl group is just the quotient of λ(Sn)−1 by {Ad(u) : u ∈ Sn}. Likewise, the restricted outer Weyl group is the quotient of λ(Pn)−1 by {Ad(w) : w ∈ Pn}. 3 The automorphisms of the diagonal In this section, we give an algorithmic crierion for deciding if the restriction to Dn of an endomorphisms λu, u ∈ Sn, gives rise to an automorphism of the diagonal Dn. Lemma 3.1 Let u ∈ Sn be such that u = P(α,β)∈J SαS∗ n for all k ∈ N. n) ⊆ Dkℓ λu(Dk β, and let ℓ = ℓ(J ). Then 5 Proof. We proceed by induction on k. For k = 1 and i ∈ W 1 n we have λu(Pi) = uPiu∗ = X(α,β), (α′,β ′)∈J SαS∗ βPiSβ ′S∗ α′ = X(α,β)∈J , s(β)=i SαS∗ α and thus λu(D1 n) ⊆ Dℓ n. For the inductive step, suppose that λu(Dk n) ⊆ Dkℓ n . Then λu(Dk+1 n ) = λu(D1 nϕ(Dk n)) = λu(D1 n)(Ad(u)ϕλu)(Dk n) ⊆ Dℓ n(Ad(u)ϕ)(Dkℓ n ) ⊆ Dℓ nD(k+1)ℓ n and thus λu(Dk+1 n ) ⊆ D(k+1)ℓ n . ✷ Proposition 3.2 Let u ∈ Sn. Then the following hold. 1. λuDn is an automorphism of Dn if and only if for each α ∈ Wn the sequence {u∗ kPαuk} eventually stabilizes. 2. λu is an automorphism of On if and only if: (a) λuDn is an automorphism of Dn, and (b) there exists a w ∈ Sn such that λwDn = (λuDn)−1. Proof. Ad 1. This is well-known, [9]. Indeed, the sequence {u∗ kPαuk} eventually stabi- lizes if and only if Pα belongs to the range of λu (and then λu(lim u∗ kPαuk) = Pα). Thus, condition 1. is equivalent to λu(Dn) = Dn, i.e. to λuDn being an automorphism of Dn. Ad 2. If λu is automorphism of On, then λu(Dn) ⊆ Dn since u ∈ NOn(Dn). Thus λu(Dn) = Dn, since Dn is a MASA in On. Also, there exists w ∈ Sn such that λ−1 u = λw, [15, 6, 12]. This gives one implication of part 2. For the reversed implication, suppose that (a) and (b) hold. Then λuλwDn = id. Thus λuλw is an automorphism of On by [1, Proposition 3.2]. Consequently, λu being surjective is automorphism of On. ✷ 1S∗ 2 +S2S∗ 1 +S1S2S∗ Example 3.3 (a) If u = S1S1S∗ Indeed, a straightforward calculation shows that λu(D2)P2 = CP2. (b) If u = S2S1S∗ projection P12 does not satisfy (Condition 1) of Proposition 3.2. (c) If u = u∗ = S1S∗ projection P11 does not satisfy (Condition 1) of Proposition 3.2. 2 + P21 + S2S2S∗ 1 + S2S2S∗ 2S∗ 1S∗ 2 + S1S∗ 2S∗ 2S∗ 1 ∈ S2 then λuD2 is not surjective. Indeed, Indeed, 2 ∈ S2 then λuD2 is not surjective. 2 ∈ S2 then λuD2 is not surjective. Our next result shows that in order to verify (Condition 1) in Proposition 3.2 it is enough to check it only for finitely many projections. Before that, we note the following. β. Then for each word µ ∈ Wn and for each Let u ∈ Sn be such that u = P(α,β)∈J SαS∗ (α, β) ∈ J we have In particular, Ad(Pβ) = Pα. Ad(u)(Pβµ) = Pαµ. (3) 6 Lemma 3.4 Let u ∈ Sn be such that u = P(α,β)∈J SαS∗ λuDn is an automorphism of Dn if and only if for each γ ∈ W ℓ′ eventually stabilizes. β, and let ℓ′ = ℓ′(J ). Then kPγuk} n the sequence {u∗ Proof. For short, say a projection Q ∈ Dn is "bad" (relative to u) if the sequence {u∗ kQuk} does not stabilize, and "good" otherwise. Also, let r be the non-negative integer uniquely defined by requiring that all projections in Dr n are good, but there is a bad projection in Dr+1 is bad as well. We claim that r + 1 ≤ ℓ′. . Then at least one of the minimal projections in Dr+1 n n n n Reasoning by way of contradiction, suppose that ℓ′ < r +1 and let p = Pγ, γ ∈ W r+1 , . Now, u∗pu can be computed using equation be such a bad minimal projection in Dr+1 (3), with u replaced by u∗, and hence it is still of the form Pγ1 for some γ1 ∈ Wn. In this process, by replacing the initial α-segment of γ with the corresponding β, the last r + 1 − ℓ′ digits will remain unaltered. Now, the assumption that p is bad easily implies that the projection Pδ := nφ(u∗pu), obtained from u∗pu by deleting the first digit of γ1, is still bad. By assumption, one must have δ ≥ r + 1, and hence u∗pu /∈ Dr+1 . In other words, when computing u∗pu we have replaced a word α in γ with a longer word β. This implies that when in the next step we consider Pγ2 := ϕ(u)∗u∗puϕ(u), the last r + 1 − ℓ′ digits of γ2 will coincide again with those of γ. Also, (nφ)2(Pγ2) must be bad, i.e. u∗ . Repeating this argument, one can indeed show that u∗ for all k = 1, 2, . . ., and moreover the last r + 1 − ℓ′ digits of γk kpuk = Pγk /∈ Dr+k coincide with those of γ for any k. All in all, this means that these last digits of γ indeed play no role in the whole process and defining γ′ simply to be the multi-index obtained from γ by deleting its last digit, the very same argument would readily show that Pγ ′ is still bad. But then Pγ ′ ∈ Dr n, contradicting our assumption. 2pu2 = Pγ2 /∈ Dr+2 n n n By the above, if there are bad projections at all, we can find at least one of them in n . As a sum of good projections is clearly good, it is also clear that in that case there ✷ Dℓ′ is always such a bad projection of the form Pγ, where γ = ℓ′. All in all, for u ∈ Sn one has λu(Dn) = Dn ⇔ Dℓ′ n ⊆ λu(Dn), (4) where ℓ′ is as in the statement of Lemma 3.4. In view of Lemma 3.4, the process of determination if an endomorphism λuDn, u ∈ Sn, is an automorphism of the diagonal can be reduced to verification if a certain finite collection of projections is contained in its range. This is a very significant reduction but still it is not clear a priori if this process can be carried out in finately many steps even for a single projection! This question has a positive answer in the case of a permutative unitary u ∈ Pn, as shown in [15, 6], but the present case is much more complicated. Now, we will describe a key construction of the present paper, producing a certain finite directed graph corresponding to a unitary u ∈ Sn. Non occurence of closed paths on the graph will turn out to be equivalent to λuDn being automorphism of Dn. Given u = P(α,β)∈J SαS∗ vertices Γ0 Γu, we proceed by induction. β in Sn, we define a finite directed graph Γu, whose u will be identified with certain subsets of J1. In order to construct the graph 7 The initial step. To begin with, we include in Γ0 the empty set ∅. Now, given (α, β) ∈ J , one of the following three cases takes place: u each singleton subset {α} of J1 and (i) β = (i) for some i ∈ W 1 n , (ii) β = (i, α′, µ) for some i ∈ W 1 n , α′ ∈ J1, and a word µ (possibly empty), (iii) β = (i, µ) for some i ∈ W 1 elements of J1, namely α′ 1, . . . , α′ r. n and a word µ which is an initial segment of at least two Depending on the case, we enlarge the graph Γu as follows. In case (i), we add an edge from vertex {α} to vertex ∅ with label i. In case (ii), we add an edge from vertex {α} to vertex {α′} with label i. In case (iii), we add a vertex A = {α′ r} and an edge from {α} to A with label i. The inductive step. Let A ⊆ J1 be a vertex added to Γ0 u in the preceding step, but A 6= ∅ and A not a singleton set. For each j ∈ W 1 n we proceed as follows. Let Bk, k = 1, . . . , m, be the collection of all those already constructed vertices of Γu that there k=1 Bk = J1 then we add k=1 Bk (if such a vertex does not exist already), and we add an edge from A to B with label j. exists an α ∈ A and an edge from {α} to Bk with label j. If Sm an edge from A to ∅ with label j. If Sm k=1 Bk 6= J1 then we add a vertex B =Sm 1, . . . , α′ Continuing inductively in the above described manner, we produce the desired graph Γu. This is a finite, directed, and labeled graph. Each vertex emits at most n edges, carrying distinct labels from the set W 1 n . Any finite path on the graph Γu may be uniquely identified with a pair (A, ν), where A ∈ Γ0 u is the initial vertex of the path and ν = (ν1, ν2, . . . , νk) is the word such that νj is the label of the jth edge entering this path. For such a path (A, ν), we denote its terminal vertex by ν(A). We will denote by Γ1 u the set of finite paths. Γ∗ u(A), respectively, are the sets of finite paths and paths of length k which begin at the vertex A. u the set of paths of length k, and by Γ∗ u the set of edges of the graph, by Γk u(A) and Γk Example 3.5 Let u = S1S∗ 1. Then the corresponding graph Γu has five vertices and five edges, and looks as follows. In particular, there is a closed (directed) path on the graph. 2 + S2S2S∗ 2 + S2S1S∗ 2 S∗ 1S∗ ..................................................... .... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. ... . 21 .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. ... ..................................................... (2) .................................................................................................................. .................. ............................................................. ...................................................... .. .. .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. . ... . 22 ... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. .. ... ..................................................... (1) .................................................................................................................. .................. ............................................................. ..................................................... .... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. ... . 1 .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. ... ..................................................... ...................................................... .. .. .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. . ... . ∅ ... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. .. ... ..................................................... (2) ............................................................................................................................................................................................................................................................................. .................. ................................. .. .. .... ................................. .................. (2) ............................................................................................................................................................................................................................................................................. ......................................................................................................................................................................................................................................................................................................................... (1) ................................. .................. . . .. . .. . .. .. .. .. ... ... ... .... ..... ... .................................... ............................. .. ........................... .. .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .. 21, 22 ........................ .................................... .. ... ........................... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. Example 3.6 Let u = S12S∗ the corresponding graph Γu looks as follows: 21 +S11S∗ 221 +S21S∗ 222 +S2222S∗ 11 +S2221S∗ 122 +S221S∗ 121. Then 8 ............................ ............................. .... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. ... ... .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .... 2222 ......................................................... ........................................................... ... .. .. .. . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. . .. ... 11 .. .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. ... ............................. ........................... .... ........................................................ .... ... ... .. .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. . .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. .. 221 .. .... ........................... ............................... .............................................................................................................................. ............................................................................................................................................................. (1) .................. .................................. ................ ............................ ............................ ... .... .. 11, 12 ... . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. ... ............................. ............................. ................ .. .. ... (2) ............................ .... .. ............................................................... ............................ ... .................. .................................. 11, 12, 21 .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. ... . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. .. ... ............................. ............................................................... .. .. ... ... ... ... ............................. .. ... ... ... ... . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2) .............................. .............................. ... .. . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .. .. .. ... 12 ... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .. ........................................................... .............................. .............................. ... .. . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .. .. .. ... 2221 ........................................................... ... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .. (1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. ... ... ... .. ... ... ... .. ... ... ... .. ... ... ... (2) .. ... ... ... ... .. ... ... ... .. ... ... ... .. ... ... . ... ... .. .. ... ... . . . . . . ... .. ... . ... ... . . .. . .. . ... ... .. . ... .. ..................... .................. . .. ... .. ... . ... ... . .. ... ... . . .. .. . . . . . . . . . ... ... ... ... ... ... .. .. ... ... ... ... .. ... ... ... .. ... ... ... .. ... (1) ... ... .. ... ... ... .. ... .............................. .............................. ... .. . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .. .. .. ... ∅ ... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .. ........................................................... (2) ......................................................................................................................................................................................... ..................................... . . . . . .. . .. . .. ... .. (1) .................. .................................. .............................................................................................................................. .............................................................................................................................. ........................................................ .... ... .. .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. . .. 21 ... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. .. .. .... ........................... ............................... (2) ............................................................................................................................................................. ............................... .. .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . ... .............................. .................. .................................. 221, 2221, 2222 ............................................................................................................................................................. ............................... ... ... .. ... ... .. .......................... .... ... .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. .. For α ∈ J1, we say that {α} is a splitting vertex if it emits an edge to a vertex A ⊆ J1 such that A contains at least two elements. This happens when for (α, β) ∈ J we have that β is an initial subword of more than one α ∈ J1. For example, α = 1 in Example 3.5 and α1 = 12, α2 = 21, α3 = 2221 and α4 = 2222 in Example 3.6 are all splitting vertices. The point of introducing graph Γu is that it conveniently captures the essential kPαuk, appearing in part 1 of Proposition 3.2. features of the process of calculating u∗ Indeed, for A ∈ Γ0 u denote PA :=Pα∈A Pα. Then we have u(A) Xα∈j(A) Clearly, if Ad(u∗)(Pµ) =Pk Pνk, then for each i ∈ W 1 Ad(ϕ(u∗))(Piµ) =Xk Ad(u∗)(PA) = X(A,j)∈Γ1 Piνk. Pjα. n we have (5) (6) Combining (5) with (6) and proceeding by induction on k, we see that for any A ∈ Γ0 u and a non-negative integer k we have Ad(u∗ k)(PA) = X(A,ν)∈Γk u(A) Xα∈ν(A) Pνα + k−1Xm=1 X(A,µ)∈Γm u (A) µ(A)=∅ Pµ. (7) Now, we are ready to prove a theorem which gives an algorithmic (finite) procedure for determining if an endomorphism λu, u ∈ Sn, restricts to an automorphism of the diagonal Dn. Theorem 3.7 Let u ∈ Sn and let Γu be the directed graph corresponding to u. Then λuDn is an automorphism of Dn if and only if graph Γu does not contain any closed (directed) paths. 9 Proof. Firstly, suppose that there is a closed path A1 (i1) −→ A2 (i2) −→ . . . (ir−1) −→ Ar (ir) −→ A1 in the graph Γu. We denote ν = (i1, i2, . . . , ir) and νk = νν · · · ν (k-fold composition). With help of formula (7) we see that Ad(u∗ kr)(PA1) = Pνk ϕkr(PA1) + Xµ∧νk=∅ Pµ. (8) Given any k < k′ there exists a non-zero projection q ∈ Dn such that q ≤ Pνk and qPνk′ = 0. But then formula (8) implies that q Ad(u∗ k′r)(PA1) = 0. Thus the sequence {Ad(u∗ m)(PA1)} never stabilizes and, consequently, projection PA1 does not belong to λu(Dn), Proposition 3.2. kr)(PA1) 6= 0 while q Ad(u∗ Conversely, suppose that graph Γu does not contain any closed paths. By virtue of k)(Pµ)} eventually stabilizes for Lemma 3.4, it suffices to show that the sequence {Ad(u∗ each µ ∈ W ℓ′ n with ℓ′ = ℓ′(J ). To this end, consider the following three cases. Firstly, we consider the case of Pα, α ∈ J1. Since Γu is a finite graph without closed paths, there are only finitely many paths and each of them terminates at a sink. By construction, graph Γu contains exactly one sink, namely vertex ∅. Thus formula (7) applied to A = {α} shows that for sufficiently large k we have Ad(u∗ k)(Pα) = X(A,µ)∈Γ∗ u({α}) µ({α})=∅ Pµ, and thus the sequence {Ad(u∗ k)(Pα)} eventually stabilizes. Secondly, we consider a word µ such that there exists an α ∈ J1 with µ ≺ α. Then In this case, the Pµ = P Pα′, where the sum is over all such α′ ∈ J1 that µ ≺ α′. k)(Pµ)} stabilizes by the preceding argument. sequence {Ad(u∗ Thirdly, we must consider the case with µ a word of length at most ℓ′ for which there exists an α ∈ J1 such that α ≺ µ. Then write µ = αν. Let ({α}, η) be the maximal path beginning at {α} and such that each vertex on the path is a singleton subset of J1. Let k be the length of this path. Using formula (7), we see that Ad(u∗ k)(Pµ) = Pηα′ν (9) for some α′ ∈ J1. Now, one of the following two cases happens: either {α′} emits an edge (with label j) to the sink ∅, or {α′} is a splitting vertex. In the former case, we have Ad(u∗ k+1)(Pµ) = Pηjν, and the question of stabilization of the sequence corresponding to the word µ reduces to the same question for the sequence corresponding to the word ν, which is strictly shorter then µ. In the latter case, let {α′} emit an edge (with label j=1 Piαj νj , for some αj ∈ J1 and words νj such that each νj is strictly shorter then ν. Taking into account formula (9), we obtain Ad(u∗ k)(Pµ)} stabilizes (with µ = αν) reduces to the same question for all µj = αjνj, where νj < ν. Consequently, the claim follows for all words µ = αν, α ∈ J1, by induction on ν. ✷ i) to a vertex A. Then we have Ad(u∗)(Pα′) = Pβ ′ =Pm j=1 Pηiαj νj . Thus, the question if the sequence {Ad(u∗ k+1)(Pµ) = Pm 10 Remark 3.8 We note that for certain special classes of unitaries u ∈ Sn, different criteria for λuDn ∈ Aut(Dn) were given earlier in [7]. 4 The invertibility In this section, we consider the problem when λu, u ∈ Sn, is an automorphism of On. Recall that E : On → Fn is the gauge invariant conditional expectation, and for a β ∈ W k obtained from β by removing its first letter. n the symbol β denotes the word in W k−1 n Lemma 4.1 If u ∈ Sn is arbitrary then there exists a v ∈ Sn such that E(w) 6= 0 for w = vuϕ(v∗). Proof. Let u = P(α,β)∈J SαS∗ P(α,β)∈J vSαS∗ β and suppose that E(u) = 0. If v ∈ Sn then vuϕ(v∗) = v∗S∗ β1. Thus, it suffices to find a v ∈ Sn such that for certain (α, β) ∈ J β v∗ ∈ O(1) we have vSαS∗ n . Since E(u) = 0, there exists (α, β) ∈ J with α > β. Now, β one of the following two cases takes place: either Pα is orthogonal to P β or β ≺ α and β 6= α. In the former case, put v = S2 β +(other terms). In the latter, we have α = βµ. Take any ν 6= µ with ν = µ and put v = S1S∗ + (other terms). βµ Then w = vuϕ(v∗) has the required form. α + S2S∗ 2 S∗ βν + Sµ 1S∗ ✷ Let u = P(α,β)∈J SαS∗ β and ℓ = ℓ(J ). Assume Dℓ n ⊆ λu(On). Then for each (α, β) ∈ J and j = 1, . . . , n the element SαS∗ βSj = Pαλu(Sj) belongs to λu(On). Denote by Zu the collection of all finite products of these elements SαS∗ and their adjoints. The β linear span of Zu is dense in λu(On). Also, we denote by hZui the collection of all sums of elements from Zu. Lemma 4.2 Let u = P(α,β)∈J SαS∗ β. Denote ℓ = ℓ(J ) and let k ≥ ℓ be any integer such that there exists a z ∈ Zu, a word of length 2k − 1, with z ∈ O(1) n . Assume that λu(Dn) = Dn and E(u) 6= 0. Then for λu to be an automorphism of On it suffices that F k n ⊆ λu(On). If this is the case then each SµS∗ n belongs to hZui. ν with µ, ν ∈ W k Proof. At first we note that a word z, as in the statement of this lemma, exists since E(u) 6= 0 by assumption. Then observe that ϕk(Si) belongs to λu(On) for all i = 1, . . . , n. Hence ϕk(Fn) ⊆ λu(On), and consequently λu(On) contains the entire Fn. Thus Fn and ϕk(Si) are contained in λu(On) and we conclude that λu(On) = On. We have hZui = λu(h{SµS∗ ν : µ, ν ∈ Wn}i) ⊆ h{SµS∗ invertible then there exists a unitary w ∈ Sn such that λ−1 ν : µ, ν ∈ Wn}i. Now, if λu is u = λw. Thus we have λ−1 u (h{SµS∗ ν : µ, ν ∈ Wn}i) ⊆ h{SµS∗ ν : µ, ν ∈ Wn}i and, consequently, hZui = h{SµS∗ ν : µ, ν ∈ Wn}i. ✷ 11 For a while, we restrict our attention to unitaries u = P(α,β)∈J SαS∗ β such that α − β ∈ {0, ±1} for all (α, β) ∈ J . (We note that endomorphisms corresponding to such unitaries were studied earlier in [5].) Given a unitary u as above, we may always find its presentation such that the lengths of all α coincide. Let k be this common length. Then the collection of all α entering the presentation of u is equal to W k n . Now, we define a new directed graph ∆u, as follows. The set of vertices of ∆u is just W k n . We put an edge from α1 to α2 whenever Pα2 ≤ P β1. In view of our assumptions on u, the difference α − β is 0, 1 or 2. We call this difference the degree of vertex α and of each edge emitted by α. If d is the degree of α then the vertex α emits exactly nd edges, which end at distinct vertices. With each edge of degree d > 0 from α1 to α2, we associate a label, which is the terminal subword of length d of α2. Edges of degree 0 carry empty labels. We extend so defined labels from edges to finite directed paths on ∆u by concatenation. Also, we define the degree of a path on ∆u as the sum of the degrees of its edges. We denote the label of a path x by L(x) and its degree by deg(x). Now, let ∆∗ u × ∆∗ u be the set of all finite directed paths. In what follows, we consider pairs (x, y) in ∆∗ u such that x and y end at the same vertex. Let x = x′e and y = y′f , where e from α1 to α and f from α2 to α are the last edges of x and y, respectively. Since e and f end at the same vertex, P β1P β2 6= 0 and thus either β1 ≺ β2 or β2 ≺ β1. Let µ be the word of length β1 − β2 such that β1 = β2µ or β2 = β1µ. We say that the pair (x, y) is balanced if the following condition holds: L(x′) = L(y′)µ if β1 = β2µ, and L(x′)µ = L(y′) if β2 = β1µ. Then we define the total label of (x, y) as L(x′) in the former case, and L(y′) in the latter. Now, we define a subset Ωu of the Cartesian product ∆∗ u, as follows. A pair (x, y) belongs to Ωu if and only if: u × ∆∗ (i) The paths x and y end at the same vertex, but they begin at distinct vertices. (ii) The paths x and y have identical degrees. (iii) The pair (x, y) is balanced. The importance of the set Ωu for our purposes comes from the following Proposition 4.3. Unfortunately, it is not clear to us at the present moment if its hypothesis may be algorithmically verified in all cases (i.e. for all applicable unitaries u ∈ Sn). However, in many concrete situations this can be done fairly easily, but preferably with the help of a computer. Thus, combined with Theorem 3.7, Lemma 4.1 and Lemma 4.2, Proposition 4.3 gives a criterion for deciding invertibility of endomorphism λu. Proposition 4.3 Let u =P(α,β)∈J SαS∗ the corresponding graph. We assume that λu(Dn) = Dn. Then the following hold. β be such that α − β ∈ {0, ±1} and let ∆u be 1. Let (x, y) ∈ Ωu have the total label γ and let the paths x, y begin at α and α′, respectively. Then Zu contains SαPγS∗ α′. 2. If α, α′ ∈ W k α′ belongs to hZui if and only if there exists a finite collection (x1, y1), . . . , (xm, ym) in ∆u with the total labels γ1, . . . , γm, respectively, and with n then SαS∗ 12 all xj beginning at α and all yj beginning at α′, such that 1 = mXj=1 Pγj . Proof. Ad 1. Let (α, α1, . . . , αm) be the consecutive vertices through which the path x passes, and likewise let (α′, α′ r) be such vertices for y. Then our definition of Ωu ensures that 1, . . . , α′ SαPγS∗ α′ = SαS∗ βSα1S∗ β1 · · · SαmS∗ βm (Sα′S∗ β ′Sα′ 1S∗ β ′ 1 · · · Sα′ r S∗ β ′ r )∗, and thus SαPγS∗ α′ ∈ Zu. α′ = Pm Ad 2. Suppose that SαS∗ α′ ∈ hZui, and let SαS∗ νj , with each Sµj S∗ νj in Zu. Since there are no cancellations among words, each Sµj S∗ νj must be of the form SαPγj S∗ α′ for some γj ∈ Wn. Now, it is not difficult to verify that an element of Zu has this form if and only if there exists a pair (xj, yj) in Ωu with the total label γj and such that xj and yj begin at α and α′, respectively. j=1 Sµj S∗ The reverse implication is an immediate consequence of part 1 of this proposition. ✷ We end this section with some examples of invertible endomorphisms λu, u ∈ Sn \Pn. Example 4.4 Let µ, ν be two words such that ν = j1 · · · jr µ with jk ∈ W 1 {µ1, ν1} for all k = 1, . . . , r. Let n and jk 6∈ u = SνS∗ µ + SµS∗ ν + 1 − Pν − Pµ. Suppose that λu(Dn) = Dn. We claim that then λu is automatically invertible. deed, it suffices to check that SνS∗ µSµS∗ Now, SνS∗ µ = Pνλu(Sµ1) and SνS∗ SνS∗ invertible, as claimed. In- ν SνS∗ µ. µ = ν ∈ λu(On). Consequently, λu is µ ∈ λu(On). But we have SνS∗ µ = λu(S∗ ν Sj1 · · · Sjr = Pνλu(Sj1) · · · λu(Sjr) and hence SµS∗ ν1)Pµ are both in λu(On). Also, SνS∗ µ = SνS∗ Example 4.5 Let α1, α2, α3 be such that {Pαj } are mutually orthogonal and each αj begins with the same letter i. Furthermore, suppose that αj = γjµ for some γj which do not contain the letter i. Let u = Sα1S∗ α2 + Sα2S∗ α3 + Sα3S∗ α1 + 1 − Pα1 − Pα2 − Pα3. If λu(Dn) = Dn then automatically λu ∈ Aut(On). Indeed, we have Sk = λu(Sk) for all k 6= i and thus Sα1S∗ i )Pα2 belongs to λu(On). Similarly, Sα2S∗ α3 and Sα3S∗ α1 are in λu(On) as well. Thus u ∈ λu(On) and λu is invertible. A concrete example in O2 is obtained by putting α2 = Pα1λ(Si)Sγ2S∗ γ3λ(S∗ w = S11S∗ 121 + S121S∗ 1221 + S1221S∗ 11 + P1222 + P2, and then indeed λw is an automorphism of O2. 13 5 The outer Weyl group In this section, we consider the question whether an endomorphism corresponding to a unitary in Sn may or may not be equivalent (via an inner automorphism) to one corresponding to a unitary in the core UHF-subalgebra Fn. Proposition 5.1 There exist unitaries u ∈ Sn such that λu 6∈ Aut(On)λ(Fn). Proof. At first we observe that if w ∈ U(Fn) and Q 6= 0 is a projection in On then the space λw(Dn)Q is infinite dimensional. Indeed, since E(Q) is a non-zero, positive element of Fn, there is a non-zero projection q ∈ Fn and a scalar t > 0 such that tq ≤ E(Q). There exists a sequence of indices jk ∈ W 1 n are defined recursively as α1 = j1, αk+1 = (αk, jk+1) then λw(Pαk)q 6= 0 for all k. The sequence {qλw(Pαk)q} never stabilizes. Indeed, if qλw(Pαk+m)q = qλw(Pαk)q for all m then n such that if αk ∈ W k 0 6= τ (qλw(Pαk )q) = τ (qλw(Pαk+m)q) = τ (λw(Pαk+m)qλw(Pαk+m)) ≤ τ (Pαk+m) −→ m→∞ 0, a contradiction. The inequality above holds since w being in U(Fn) the corresponding endomorphism λw is τ -preserving. Thus, there is a strictly decreasing, infinite sequence of projections f1 > f2 > . . . in λw(Dn) such that qfkq > qfk+1q for all k. Thus (fk − fk+1)q 6= 0 for all k, and hence 0 6= (fk − fk+1)tq(fk − fk+1) ≤ (fk − fk+1)E(Q)(fk − fk+1). Thus (fk−fk+1)Q 6= 0 and, consequently, {(fk−fk+1)Q} is an infinite sequence of linearly independent elements of λw(Dn)Q, since these are non-zero operators with mutually orthogonal ranges. Now, the same conclusion as above holds if λw is replaced by ψλw for some auto- morphism ψ ∈ Aut(On), since the dimension of (ψλw)(Dn)Q is the same as that of λw(Dn)ψ−1(Q). Thus, the conclusion of the proposition follows from Example 3.3 (a), where a unitary u ∈ S2 is exhibited such that λu(D2)P2 is one-dimensional. ✷ Of course, the method of Proposition 5.1 cannot give any information about auto- morphisms. We treat the automorphism case in Theorem 5.2, below. To the best of our knowledge, the automorphism entering its proof is the first known example of an automorphism of On not inner related to an automorphism induced by a unitary from the core UHF subalgebra Fn. Theorem 5.2 There exist automorphisms λu, u ∈ Sn, of On such that for all w ∈ U(On) and v ∈ U(Fn) we have λu 6= Ad(w)λv. Proof. At first we consider the following self-adjoint element of S2 (c.f. Example 4.4): u = S11S∗ 121 + S121S∗ 11 + P122 + P2. (10) One easily checks that λu(S1) = S1(S1S∗ 21 + S21S∗ 1 + P22) and λu(S2) = S2. 14 This yields λ2 u = id. Suppose by contradiction that there exist w ∈ U(O2) and v ∈ U(F2) such that λu = Ad(w)λv. Then we have D2 = λu(D2) = Ad(w)λv(D2) and λv(D2) ⊆ F2, thus w∗D2w ⊆ F2. Hence for all d ∈ D2 and z ∈ U(1) we have γz(w∗dw) = w∗dw. Therefore γz(w)w∗ ∈ D′ 2 ∩ O2 = D2. Thus for each z ∈ U(1) there exists a unitary dz ∈ U(D2) such that γz(w) = dzw. Now we calculate γz(u) = γz(wvϕ(w∗)) = dzwvϕ(w∗)ϕ(d∗ z) = dzuϕ(d∗ z). This means that if u =P SαS∗ β then zα−βSαS∗ βdz = dzSαS∗ β for all (α, β) and z ∈ U(1). Taking α = β = 2 this yields S2dz = dzS2. By [13], dz is a scalar, and this in turn implies that γz(u) = u for all z, a contradiction. Now, if n ≥ 2 is arbitrary, then we consider u = S11S∗ 121 + S121S∗ and the same argument as above applies. 11 + P122 + 1 − P1, ✷ As immediate consequences of Theorem 5.2, we obtain the following two corollaries. Corollary 5.3 The restricted outer Weyl group of On is a proper subgroup of the outer Weyl group of On. As shown in [2], the restricted outer Weyl group of On is residually finite and nona- menable. Thus the outer Weyl group is nonamenable as well, but we do not know if it is residually finite. Corollary 5.4 There exist unital subalgebras A of On isomorphic to the UHF algebra of type {n∞} such that Fn and A are conjugate inside On (by an automorphism of On) but not inner conjugate. 6 The action on the shift space Equality (3) easily implies that for all d ∈ Dn and all k > ℓ′(J ) we have Ad(u)(ϕk(d)) = X(α,β)∈J ϕk+α−β(d)Pα. (11) Consider a map f : Dn → Dn. We say that f eventually preseves standard projec- tions if there exists an integer m ∈ N such that for each α ∈ Wn, α ≥ m, the image If u ∈ Sn then Ad(u) eventually preserves standard f (Pα) is a standard projection. projections. Proposition 6.1 If f ∈ Aut(Dn) then there exists a unitary u ∈ Sn such that f = Ad(u)Dn if and only if; (i) f eventually preserves standard projections, and 15 (ii) there exist projections Pi, Qi, i = 1, . . . , r, in Dn and non-negative integers ki, mi, i = 1, . . . , r, such that Pr i=1 Pi = 1 =Pr i=1 Qi and f ◦ MPi ◦ ϕki = MQi ◦ ϕmi, i = 1, . . . , r. Proof. Let f ∈ Aut(Dn) satisfy conditions (i) and (ii) of the proposition. For a given i ∈ {1, . . . , r}, we note that for any subprojection p of Pi we have f ◦ Mp ◦ ϕki = Mf (p) ◦ ϕmi. Subdividing Pi into a sum of standard projections and using condition (i), we can assume in condition (ii) that all projections Pi, Qi are standard, say Pi = Pβi and Qi = Pαi. Define u =Pr (Ad(u∗) ◦ f ) ◦ MPβi i=1 SαiS∗ ◦ ϕki+αi+h = MPβi βi, a unitary element of Sn. Then we have ◦ ϕmi+β+h, i = 1, . . . , r, for all sufficiently large h ∈ N. We claim that ki + αi = mi + βi for each i. Indeed, fix an i and suppose that ki + αi ≥ mi + βi (otherwise consider (Ad(u∗) ◦ f )−1 instead). Then we have (Ad(u∗) ◦ f )(ϕki+αi−mi−βi(y)Pβi) = yPβi, ∀y ∈ ϕmi+βi+h(Dn), for all sufficiently large h ∈ N. Fix such an h and let r ≥ mi + βi + h be such that (Ad(u∗) ◦ f )(ϕki+αi−mi−βi(Dmi+βi+h n )Pβi) ⊆ Dr nPβi and Pβi ∈ Dr n. Then we have (Ad(u∗) ◦ f )(ϕki+αi−mi−βi(Dr n)Pβi) = (Ad(u∗) ◦ f )(ϕki+αi−mi−βi(Dmi+βi+h ⊆ Dr ⊆ Dr nPβiϕmi+βi+h(Dr−mi−βi−h nPβi. n )Pβi n )ϕmi+βi+h(Dr−mi−βi−h n )Pβi) Since Ad(u∗) ◦ f is injective and the dimension of ϕki+αi−mi−βi(Dr than the dimension of Dr can only happen when ki + αi − mi − βi = 0. Consequently, n)Pβi is not smaller nPβi, it follows that these two dimensions are identical, and this (Ad(u∗) ◦ f ) ◦ MPβi ◦ ϕh = MPβi ◦ ϕh, i = 1, . . . , r, for all sufficiently large h ∈ N. Summing over i we get (Ad(u∗) ◦ f ) ◦ ϕh = ϕh The opposite direction is clear. Indeed, let u ∈ Sn be such that u = P(α,β)∈J SαS∗ for all sufficiently large h ∈ N. Therefore Ad(u∗) ◦ f = Ad(w)Dn for some w ∈ Pn, by [2, Lemma 3.2]. Hence f = Ad(uw)Dn and uw ∈ Sn. This proves one direction. β. Then condition (i) holds, as noted just above this proposition. One easily checks that condition (ii) holds with projections Pβ and Pα instead of Pi and Qi, respectively, and with β and α instead of ki and mi, respectively. ✷ Given u ∈ Sn and considering the homeomorphism Ad(u)∗ of the spectrum Xn of Dn, we see that the set of fixed points has a very simple structure, as the following Proposition 6.2 shows. 16 Proposition 6.2 For u ∈ Sn, the set of fixed points in Xn for the homeomorphism Ad(u)∗ consists of the union of a clopen set and a finite set. Furthermore, each of the isolated fixed points is either a local attractor or a local repeller. Proof. Let u =P(α,β)∈J SαS∗ β. It is clear that Ad(u)∗ admits fixed points in Xn if and only if there exists (α, β) ∈ J such that either α ≺ β or β ≺ α. Thus we arrive at one of the following three cases. (1) If α = β then the clopen set {x ∈ Xn : β ≺ x} is fixed by Ad(u)∗. (2) If α = βµ, µ 6= ∅, then x = βµµ . . . is a fixed point and a local attractor. (3) If β = αµ, µ 6= ∅, then x = αµµ . . . is a fixed point and a local repeller. ✷ In contrast to Proposition 6.2 above, the set of fixed points in Xn corresponding to an outer automorphism λu, u ∈ Sn, may have a much more complicated structure, as the following example demonstrates. Example 6.3 Let u be the unitary in S2 defined by formula (10). It is not difficult to verify that the corresponding homeomorphism (λu)∗ of X2 fixes an x ∈ X2 if and only if x does not contain substrings (11) and (121). These fixed points form a compact, nowhere dense subset K of X2, in which there are no isolated points. Thus K itself is homeomorphic to the Cantor set and closed under the action of the one-sided shift ϕ∗. References [1] R. Conti, Automorphisms of the UHF algebra that do not extend to the Cuntz algebra, J. Austral. Math. Soc. 89 (2010), 309 -- 315. [2] R. Conti, J. H. Hong and W. Szyma´nski, The restricted Weyl group of the Cuntz algebra and shift endomorphisms, J. reine angew. Math. 667 (2012), 177 -- 191. [3] R. Conti, J. H. Hong and W. Szyma´nski, Endomorphisms of graph algebras, ac- cepted to J. Funct. Anal., arXiv:1101.4210. [4] R. Conti, J. Kimberley and W. Szyma´nski, More localized automorphisms of the Cuntz algebras, Proc. Edinburgh Math. Soc. 53 (2010), 619 -- 631. [5] R. Conti, M. Rørdam and W. Szyma´nski, Endomorphisms of On which preserve the canonical UHF-subalgebra, J. Funct. Anal. 259 (2010), 602 -- 617. [6] R. Conti and W. Szyma´nski, Labeled trees and localized automorphisms of the Cuntz algebras, Trans. Amer. Math. Soc. 363 (2011), 5847 -- 5870. [7] R. Conti and W. Szyma´nski, Automorphisms of the Cuntz algebras, accepted to the Proceedings of the EU-NCG 4th Annual Meeting in Bucharest (2011), arXiv:1108.0860. [8] J. Cuntz, Simple C ∗-algebras generated by isometries, Commun. Math. Phys. 57 (1977), 173 -- 185. 17 [9] J. Cuntz, Automorphisms of certain simple C ∗-algebras, in Quantum fields-algebras- processes, ed. L. Streit, 187 -- 196, Springer, 1980. [10] S. Eilenberg, Automata, languages and machines, Academic Press, New York, 1974. [11] J. H. Hong, A. Skalski and W. Szyma´nski, On invariant MASAs for endomorphisms of the Cuntz algebras, Indiana Univ. Math. J. 59 (2010), 1873 -- 1892. [12] K. Matsumoto, Orbit equivalence of topological Markov shifts and Cuntz-Krieger algebras, Pacific J. Math. 246 (2010), 199 -- 225. [13] K. Matsumoto and J. Tomiyama, Outer automorphisms on Cuntz algebras, Bull. London Math. Soc. 25 (1993), 64 -- 66. [14] V. Nekrashevych, Cuntz-Pimsner algebras of group actions, J. Operator Theory 52 (2004), 223 -- 249. [15] W. Szyma´nski, On localized automorphisms of the Cuntz algebras which preserve in 'New Development of Operator Algebras', R.I.M.S. the diagonal subalgebra, Kokyuroku 1587 (2008), 109 -- 115. Roberto Conti Dipartimenti di Scienze Universit`a di Chieti-Pescara 'G. D'Annunzio' Viale Pindaro 42, I -- 65127 Pescara, Italy present address: Dipartimento di Scienze di Base e Applicate per l'Ingegneria Sezione di Matematica Sapienza Universit`a di Roma Via A. Scarpa 16 00161 Roma, Italy E-mail: [email protected] Jeong Hee Hong Department of Data Information Korea Maritime University Busan 606 -- 791, South Korea E-mail: [email protected] Wojciech Szyma´nski Department of Mathematics and Computer Science The University of Southern Denmark Campusvej 55, DK-5230 Odense M, Denmark E-mail: [email protected] 18
1201.0609	2	1201	2012-01-04T15:18:36	Radial multipliers on reduced free products of operator algebras	[ "math.OA" ]	Let A_i be a family of unital C-algebras, respectively, of von Neumann algebras and phi: N_0 \to C. We show that if a Hankel matrix related to phi is trace-class, then there exists a unique completely bounded map M_phi on the reduced free product of the A_i, which acts as an radial multiplier. Hereby we generalize a result of Wysocza\'nski for Herz-Schur multipliers on reduced group C-algebras for free products of groups.	math.OA	math	RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS OF OPERATOR ALGEBRAS UFFE HAAGERUP AND SÖREN MÖLLER Abstract. Let Ai be a family of unital C-algebras, respectively, of von Neumann algebras and φ : N0 → C. We show that if a Hankel matrix related to φ is trace-class, then there exists a unique completely bounded map Mφ on the reduced free product of the Ai, which acts as an radial multiplier. Hereby we generalize a result of Wysoczański for Herz-Schur multipliers on reduced group C-algebras for free products of groups. 1. Introduction Let C denote the set of functions φ on the non-negative integers N0 for which the matrix h = (φ(i + j) − φ(i + j + 1))i,j≥0 is of trace class. Let G = ∗i∈IGi be the free product of discrete groups (Gi)i∈I. In [12] , J. Wysoczański proved that if φ ∈ C and φ : G → C is defined by φ(e) = φ(0) and φ(g1 . . . gn) = φ(n) for all n > 0 when gj ∈ Gij \ {e} and i1 6= i2 6= · · · 6= in, then φ is a Herz-Schur multiplier on G and k φkHS ≤ kφkC , where k · kC is the norm on C defined in (2.2) below. In particular, there is a unique completely bounded map Mφ : C ∗ r (G) → C ∗ r (G) such that Mφ(1) = φ(0)1 and Mφ(λ(g1 . . . gn)) = φ(n)λ(g1 . . . gn) when gj ∈ Gij \ {e} and i1 6= i2 6= · · · 6= in as above, and kMφkcb ≤ kφkC . Furthermore J. Wysoczański proved that kMφkcb = kφkC in the cases when I = ∞ and Gi = ∞ for all i ∈ I. In the special case of φs(n) = sn for n ≥ 0 and s < 1 it follows that 1 − s 1 − s kMφskcb ≤ kφskC = . In this paper we will show that every function φ from C gives rise to radial multipliers Mφ on reduced free products of C ∗-algebras and reduced free products of von Neumann algebras (cf. Theorem 2.2), satisfying kMφkcb ≤ kφkC . Radial multipliers of general reduced free products of C ∗-algebras were first considered by È. Ricard and Q. Xu Date: November 15, 2018. 2010 Mathematics Subject Classification. Primary 46L54; Secondary 46L07. 1 RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 2 in [10] and the weaker estimate kMφkcb ≤ φ(0) +P∞ be obtained from [10, Corollary 3.3]. n=1 4nφ(n) can The main result is proved in Section 5. In Section 6 we discuss a related set of functions C ′ (cf. Definition 6.3). It was used by T. Steenstrup, R. Szwarc and the first author in [6] to characterize radial multipliers on free groups Fn (2 ≤ n ≤ ∞). Moreover, C. Houdayer and È. Ricard used it in [7] to characterize multipliers on the free Araki-Woods factor Γ(HR, , Ut)′′ (cf. Section 6.3). In Section 7 we obtain an integral representation of functions in the class C which together with N. Ozawa's result in [9], shows that for every hyperbolic group Γ, and every φ ∈ C , the function φ(x) = φ(d(x, e)) is a completely bounded Fourier multiplier on Γ (cf. Remark 7.6). 2. The main results We start by defining the class C , crucial in what follows. Definition 2.1. Let C denote the set of functions φ : N0 → C for which the Hankel matrix h = (φ(i + j) − φ(i + j + 1))i,j≥0 is of trace- class. If φ ∈ C , then k = (φ(i + j + 1) − φ(i + j + 2))i,j≥0 is of trace-class, as well. Furthermore, we have ∞ φ(n) − φ(n + 1) ≤ khk1 + kkk1 < ∞, (2.1) Xn=0 where kxk1 = T r(x) is the trace-class norm for x ∈ B(l2(N0)). This implies that c = limn→∞ φ(n) exists. For φ ∈ C set kφkC = khk1 + kkk1 + c. (2.2) The main result of this paper is the following generalization of Wysocza- ński's result: Theorem 2.2. (1) Let A = ∗i∈I(Ai, ωi) be the reduced free product of unital C- algebras (Ai)i∈I with respect to states (ωi)i∈I for which the GNS- representation πωi is faithful, for all i ∈ I. If φ ∈ C , then there is a unique linear completely bounded map Mφ : A → A such that Mφ(1) = φ(0)1 and Mφ(a1a2 . . . an) = φ(n)a1a2 . . . an whenever aj ∈ Aij = ker(ωij ) and i1 6= i2 6= · · · 6= in. Moreover kMφkcb ≤ kφkC . (2) Let (M , ω) = ¯∗i∈I(Mi, ωi) be the w-reduced free product of von Neumann algebras (Mi)i∈I with respect to normal states (ωi)i∈I for which the GNS-representation πωi is faithful, for all i ∈ I. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 3 If φ ∈ C , then there is a unique linear completely bounded normal map Mφ : M → M such that Mφ(1) = φ(0)1 and Mφ(a1a2 . . . an) = φ(n)a1a2 . . . an whenever aj ∈ Mij = ker(ωij ) and i1 6= i2 6= · · · 6= in. Moreover kMφkcb ≤ kφkC . Remark 2.3. By J. Wysoczański's result in [12], the norm estimates in Theorem 2.2 are best possible, as equality is attained if I = ∞ and r (Gi), τi) for a family (Gi)i∈I of infinite discrete groups, (Ai, ωi) = (C ∗ where τi is the canonical trace on C ∗ r (Gi) coming from the left regular representation. It would be interesting to know for which (Ai, ωi)i∈I, the equality kMφkcb = kφkC holds for all φ ∈ C . We start by proving that the operator Mφ is unique, if it exists. Lemma 2.4 (Uniqueness in Theorem 2.2). The map Mφ is uniquely determined by the conditions in the hypothesis of Theorem 2.2. Proof. The algebra C1 +(cid:16)Pi∈I dense in A and, respectively, C1+(cid:16)Pi∈I Ai(cid:17) +(cid:16)Pi16=i2 Mi(cid:17)+(cid:16)Pi16=i2 is σ-weakly dense in M . As Mφ is bounded, it is then uniquely defined on all of A , respectively, on all of M . (cid:3) Ai1 Ai2(cid:17) + . . . is norm Mi2(cid:17)+. . . Mi1 Now to prove Theorem 2.2 we start by showing that it is enough to prove the result for the special case of the algebras M = B(Hi, Ωi) equipped with ωi, the vector state given by Ωi, as this will implie the result for general C*- and von Neumann-algebras. Proposition 2.5. If Theorem 2.2 part (2) holds for (Mi, ωi) = (B(Hi), ωΩi) for Hilbert spaces (Hi, Ωi) and associated vector states ωi then Theorem 2.2 holds in general. Proof. Assume Theorem 2.2 holds for (B(Hi), ωΩi) for arbitrary Hi and Ωi. Now let A = ∗i∈I(Ai, ωi), respectively, (M , ω) = ¯∗i∈I(Mi, ωi). Let (Hi, Ωi) = (Hωi, ξωi) be the Hilbert space and state coming from the GNS-representation of Ai, respectively, Mi, and let (H, Ω) = ∗i∈I(Hi, Ωi) be their Hilbert space free product. Now by [11, Definition 1.5.1], Ai, respectively, Mi can be realized as subalgebras of B(H) by the action defined as follows. If a ∈ Ai, γ1 ⊗ · · · ⊗ γn ∈ H with γj ∈ Hj := Ω⊥ j then a(γ1 ⊗ · · · ⊗ γn) = a(Ωi) ⊗ γ1 ⊗ · · · ⊗ γn if i 6= j, and otherwise a(γ1 ⊗ · · · ⊗ γn) = (a(γ1) − ha(γ1), ΩiiΩi) ⊗ γ2 ⊗ · · · ⊗ γn + ha(γ1), Ωiiγ2 ⊗ · · · ⊗ γn. Hence MφA and MφM can be obtained by restricting Mφ to the re- spective subalgebra of B(H) and we then have kMφA kcb ≤ kMφkcb ≤ RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 4 kφkC , respectively, kMφM kcb ≤ kMφkcb ≤ kφkC which gives the de- sired general result. (cid:3) We will prove the special case considered in Proposition 2.5 in the following sections. 3. Preliminaries We start by introducing some notation. Let (H, Ω) = ∗i∈I(Hi, Ωi). Also denote Hi = Ω⊥ i , for i ∈ I. Then by n=0 H(n), Hi1 ⊗ · · · ⊗ Hin, for n > 0, and H(0) = CΩ. We will denote the projection from H to H(n) by Pn ∈ B(H), and let the definition of the Hilbert space free product we have H =L∞ where H(n) := Li16=···6=in Qn :=P∞ Now choose orthonormal bases Γi for Hi, then Γi = Γi ∪ {Ωi} are bases for Hi. Put Λ(0) = {Ω} and Λ(n) = {γ1 ⊗ · · · ⊗ γn : γj ∈ Γij , i1 6= · · · 6= in} for all n ≥ 1. Then Λ(n) is an orthonormal basis for H(n), n=0 Λ(n) is an orthonormal basis for H. Note k=n Pk. for all n ≥ 0 and Λ = S∞ that Λ(1) =Si∈I Γi considered as a subset of H. Now we can define the basic operators in B(H). Let γ ∈ Λ(1). Let Lγ, Rγ ∈ B(H) be the operators for which LγΩ = RγΩ = γ, and for χ = χ1 ⊗ · · · ⊗ χn ∈ Λ(n) where χj ∈ Γij and γ ∈ Γi we have respectively, Lγ(χ) =(cid:26) γ ⊗ χ 0 Rγ(χ) =(cid:26) χ ⊗ γ 0 if i 6= i1 if i = i1 if i 6= in if i = in. Note that Lγ and Rγ are well-defined partial isometries in B(H). Moreover for all γ ∈ Λ(1) and n ≥ 0 we have LγH(n) ⊆ H(n + 1), respectively, RγH(n) ⊆ H(n + 1). For γ = γ1 ⊗ · · · ⊗ γn ∈ Λ(n) denote Lγ = Lγ1Lγ2 . . . Lγn, respectively, Rγ = RγnRγn−1 . . . Rγ1, where we set LΩ = RΩ = 1. Lemma 3.1. Let B(Hi)= {a ∈ B(Hi) : haΩi, Ωii = 0}. Then the set spann{Lγ : γ ∈ Γi} ∪ {L∗ in B(Hi)considered as a subset of B(H). γ : γ ∈ Γi} ∪ {LγL∗ δ : γ, δ ∈ Γi}o is σ-weakly dense Proof. Let (eγ,δ)γ,δ∈Γi be the matrix units of B(Hi) corresponding to the basis Γi. Then span{eγ,δ : (γ, δ) 6= (Ωi, Ωi)} is σ-weakly dense in B(Hi). Moreover, by the natural embedding of B(Hi) in B(H) one gets for γ, δ ∈ Γi that Lγ = eγ,Ωi, L∗ δ = eγ,δ, which proves the lemma. (cid:3) γ = eΩi,γ, and hence LγL∗ RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 5 Definition 3.2. Let a = (ai)i≥0 ∈ l∞(N0). Denote by Da the operator which is defined by Da(ξ) = anξ for ξ ∈ Λ(n), n ≥ 1, respectively, Da(Ω) = a0Ω and by linearity is extended to all of H. Note that Da = P∞ n=0 anPn and that Da is bounded with kDak = kak∞. Let S denote the standard shift on l∞(N0), i.e., for a = (ai)i≥0 ∈ l∞(N0), let S(a0, a1, a2, . . . ) = (0, a0, a1, a2, . . . ). To ease notation, con- sider seperately the following two cases, which together contain all pos- sible situations. Definition 3.3. For ξ ∈ Λ(k) and η ∈ Λ(l), k, l ≥ 0 we say that we are in • Case 1 if ξ = Ω or η = Ω or k, l ≥ 1 and ξ = ξ1 ⊗ · · · ⊗ ξk and η = η1 ⊗ · · · ⊗ ηl where ξk ∈ Γi, ηl ∈ Γj and i 6= j, i, j ∈ I, respectively, • Case 2 if k, l ≥ 1 and ξ = ξ1 ⊗ · · · ⊗ ξk and η = η1 ⊗ · · · ⊗ ηl where ξk, ηl ∈ Γi for some i ∈ I. 4. Technical lemmas Definition 4.1. For x, y ∈ l2(N0) and a ∈ B(H) set Φ(1) x,y(a) := D(S ∗)nxaD∗ (S ∗)ny + ∞ Xn=0 DSnxρn(a)D∗ Sny, ∞ Xn=1 respectively, Φ(2) x,y(a) := Xn=0 ∞ ∞ D(S ∗)nxaD∗ (S ∗)ny + DSnxρn−1(ǫ(a))D∗ Sny Xn=1 where ρ(a) := Pγ∈Λ(1) RγaR∗ projection onto span{ξ ∈ Λ(n) : n ≥ 1, ξ = γ1 ⊗ · · · ⊗ γn, γn ∈ Γi} for i ∈ I. γ and ǫ(a) := Pi∈I qiaqi and qi is the Lemma 4.2. Let k, l ≥ 0. Then for every ξ ∈ Λ(k) and η ∈ Λ(l) we have for all n ≥ 0, ρn(LξL∗ η) in Case 1, and, respectively, ǫ(LξL∗ ηQl+n and ǫ(LξL∗ η) = ρ(LξL∗ η) = LξL∗ η in Case 2. η) = LξL∗ Proof. For the first statement observe that ρn(LξL∗ RζLξL∗ ηR∗ ζ η) = Xζ∈Λ(n) = Lξ  Xζ∈Λ(n) = LξQnL∗ η = LξL∗ ηQl+n. RζR∗ ζ  L∗ η RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 6 For the second statement, let χ ∈ Λ(m). If m > l then ǫ(LξL∗ qiLξL∗ ηqi(χ) η)(χ) =Xi∈I = LξXi∈I qiqiL∗ η(χ) η(χ) = LξQ(1)L∗ = LξL∗ = LξL∗ ηQ(l + 1)(χ) η(χ) While if m = l, χ = η and ηl ∈ Γj for some j ∈ I we have ǫ(LξL∗ ηqi(η) = qjLξL∗ ηqj(η) = qjLξL∗ qiLξL∗ η(η) = qj(ξ), η)(η) =Xi∈I and this is equal to 0 in Case 1 (i.e., ξk /∈ Γj), respectively, equal to ξ in Case 2 (i.e., ξk ∈ Γj). Note that both sides vanish if m = l and χ 6= η, or m < l. To calculate the completely bounded norm of Φ(·) x,y we use the follow- (cid:3) ing result from [3]. Theorem 4.3. [3, Theorem 1.3] If kPi uiu∗ some ui, vi ∈ B(H), then Φ(a) =Pi uiavi defines a normal completely bounded operator on B(H) and kΦkcb ≤ kPi∈I uiu∗ i vik. Using this theorem we ontain the following cb-norm estimates. i kkPi∈I v∗ i k, kPi v∗ i vik < ∞ for Lemma 4.4. For x, y ∈ l2(N0) we have kΦ(1) tively, kΦ(2) x,ykcb ≤ kxk2kyk2. Proof. Let χ ∈ Λ(m). Then x,ykcb ≤ kxk2kyk2, respec- ∞ Xn=0 respectively, ∞ D(S ∗)nxD∗ (S ∗)nx(χ) = x(m + n)D(S ∗)nx(χ) ∞ ∞ = Xn=0 Xn=0 = ∞ Xn=m x(m + n)2(χ) x(n)2! (χ), (DSnxRζ)(DSnxRζ)∗(χ) = Xn=1 Xζ∈Λ(n) = ∞ m Xn=1 Xζ∈Λ(n) Xn=1 DSnxRζR∗ ζ D∗ Snx(χ) DSnxD∗ Snx(χ) = x(n)2(χ). m−1 Xn=0 RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 7 Here the second equality holds since RζR∗ or if n > m. ζ (χ) = 0 if ζ 6= χm−n+1 . . . χm On the other hand, ∞ (DSncRζqi)(DSncRζqi)∗(χ) ∞ = Xn=1 Xζ∈Λ(n−1)Xi∈I Xn=1 Xζ∈Λ(n−1)Xi∈I Xn=1 Xζ∈Λ(n−1) Xn=1 DSncD∗ = = m m DSncRζqiR∗ ζD∗ Snc(χ) DSncRζR∗ ζ D∗ Snc(χ) Snc(χ) = c(n)2(χ). m−1 Xn=0 The second equality holds since χm−n+1 ∈ Γi (the rightmost element ζ (χ)) for a unique i ∈ I if n − 1 < m, and there is no such of R∗ i ∈ I if n − 1 ≥ m. The third equality holds as RζR∗ ζ(χ) = 0 for ζ 6= χm−n+2 . . . χm. Hence D(S ∗)nxD∗ (S ∗)nx + ∞  Xn=0  = kxk2 2 (χ), ∞ Xn=1 Xζ∈Λ(n) (DSnxRζ)(DSnxRζ)∗  (χ) respectively, D(S ∗)nxD∗ (S ∗)nx + ∞  Xn=1  = kxk2 2 (χ). ∞ Xn=1 Xζ∈Λ(n−1)Xi∈I (DSnxRζqi)(DSnxRζqi)∗  (χ) Using these calculations for x, y ∈∈ l2(N0) and applying Theorem 4.3 to get the desired result. (cid:3) Lemma 4.5. Let k, l ≥ 0. If ξ ∈ Λ(k) and η ∈ Λ(l) then Φ(1) x,y(LξL∗ η) = ∞ Xt=0 x(k + t)y(l + t)! LξL∗ η and, respectively, Φ(2) x,y(LξL∗ η) =(cid:26) P∞ P∞ t=0 x(k + t)y(l + t)LξL∗ t=0 x(k + t − 1)y(l + t − 1)LξL∗ η η in Case 1 in Case 2. Proof. We prove this by showing that both sides act similarly on all simple tensors in H. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 8 Indeed, let m ≥ 0 and χ ∈ Λ(m) and let n ≥ 0. If χ = η ⊗ ζ, where ζ ∈ Λ(m − l) for some l ≥ 0 we have for the common type of terms in Φ(1) x,y and Φ(2) x,y that D(S ∗)nxLξL∗ ηD∗ (S ∗)ny(χ) = y(m + n)D(S ∗)nxLξL∗ η(χ) (4.1) = y(m + n)D(S ∗)nx(ξ ⊗ ζ) = x(k + m − l + n)y(m + n)LξL∗ η(χ). Otherwise, if there is no ζ such that χ = η ⊗ ζ, then both sides vanish, wherein we have used the convention that x(p) = 0 for p < 0. For the other type of terms in Φ(1) x,y, we get DSnxρn(LξL∗ η)D∗ Sny(χ) = y(m − n)DSnxρn(LξL∗ η)(χ) (4.2) = y(m − n)DSnxLξL∗ ηQl+n(χ) = y(m − n)DSnxQl+n(ξ ⊗ ζ) = x(k + m − l − n)y(m − n)LξL∗ ηQl+n(χ) where in the second equality we use Lemma 4.2 and the fact that both sides vanish if n > m − l. We now estimate the other type of terms in Φ(2) x,y. In Case 1 we similarly get DSnxρn−1(ǫ(LξL∗ η))D∗ Sny(χ) = DSnxρn(LξL∗ Sny(χ) = x(k + m − l − n)y(m − n)LξL∗ η)D∗ ηQl+n(χ) with both sides vanishing for n > m − l. In Case 2 we get by Lemma 4.2 DSnxρn−1(ǫ(LξL∗ η))D∗ Sny(χ) = DSnxρn−1(LξL∗ Sny(χ) = x(k + m − l − n)y(m − n)LξL∗ η)D∗ ηQl+n−1(χ) (4.3) (4.4) with both sides vanishing for n > m − l + 1. Combining (4.1) and (4.2) we get Φ(1) x,y(LξL∗ η)(χ) = = ∞ ∞ Xn=0 Xn=0 D(S ∗)nxLξL∗ ηD∗ (S ∗)ny + ∞ Xn=1 DSnxρn(LξL∗ η)D∗ Sny x(k + m − l + n)y(m + n)LξL∗ η(χ) m−l + Xn=1 = ∞ Xn=l−m x(k + m − l − n)y(m − n)LξL∗ η(χ) x(k + m − l + n)y(m + n)! LξL∗ η(χ) RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 9 = ∞ Xt=0 x(k + t)y(l + t))! LξL∗ η(χ) as desired. Similarly in Case 1, combining (4.1) and (4.3) we get Φ(2) x,y(LξL∗ η)(χ) = ∞ Xt=0 x(k + t)y(l + t)! LξL∗ η(χ). While in Case 2, combining (4.1) and (4.4) we get Φ(2) x,y(LξL∗ η)(χ) = ∞ Xn=0 x(k + m − l + n)y(m + n)LξL∗ η(χ) m−l+1 + = ∞ Xt=0 This completes the proof. x(k + m − l − n)y(m − n)LξL∗ Xn=1 x(k + t − 1)y(l + t − 1)! LξL∗ η(χ). η(χ) (cid:3) We now establish some technical results conserning maps φ ∈ C . Put ψ1(n) = P∞ Lemma 4.6. Let φ ∈ C and let h, k and c be as in Definition 2.1. i=0(φ(n + 2i) − φ(n + 2i + 1)) and ψ2(n) = ψ1(n + 1), for n ≥ 0. Then φ(n) = ψ1(n) + ψ2(n) + c for n ≥ 0 and the entries hi,j and ki,j of h and k are given by hi,j = ψ1(i + j) − ψ1(i + j + 2), respectively, ki,j = ψ2(i + j) − ψ2(i + j + 2), for i, j ≥ 0. Proof. By (2.1) we have ∞ lim n→∞ ψ1(n) ≤ lim n→∞ φ(n + 2i) − φ(n + 2i + 1) = 0. Xi=0 A similar statement holds for ψ2 and therefore limn→∞ ψ1(n) = 0 and limn→∞ ψ2(n) = 0. Next, let n ≥ 0 be fixed. Then simple computations give ψ1(n) + ψ2(n) = φ(n) − c, and ψ1(n) − ψ1(n + 2) = φ(n) − φ(n + 1), respectively, ψ2(n) − ψ2(n + 2) = φ(n + 1) − φ(n + 2). Using these equations, we get the desired formulas for hi,j, respectively ki,j. (cid:3) kkk1. Here we use the notation (u ⊙ v)(t) = ht, viu, for u, v, t ∈ l2(N0). Remark 4.7. Since h, k are trace-class, it is well-known (cf. [6, p. 13]) i=1 xi ⊙ yi and that there exist xi, yi, zi, wi ∈ l2(N0) such that h = P∞ P kxik2kyik2 = khk1, respectively, k =P∞ in Remark 4.7 we have ψ1(k + l) = P∞ ψ2(k + l) =P∞ i=1 zi⊙wi andP kzik2kwik2 = i=1P∞ Lemma 4.8. For ψ1 and ψ2 as in Lemma 4.6, and xi, yi, zi, and wi as t=0 xi(k + t)yi(l + t) and t=0 zi(k + t)wi(l + t). i=1P∞ RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 10 Proof. Let k, l ≥ 0, then ∞ ψ1(k + l + 2t) − ψ1(k + l + 2t + 2) hk+t,l+t ∞ xi(k + t)yi(l + t) xi(k + t)yi(l + t) ψ1(k + l) = = = = ∞ ∞ Xt=0 Xt=0 Xt=0 Xi=1 ∞ ∞ Xi=1 Xt=0 where the sums are absolutly convergent, and we use Lemma 4.6 for the first two equalities. A similar reasoning applies to ψ2. (cid:3) 5. Proof of the main result As shown in Section 2, it is enough to prove the following lemma in order to obtain the result of the main theorem. Proposition 5.1. Let (H, Ω) = ∗i∈I(Hi, Ωi) be the reduced free product of Hilbert spaces (Hi)i∈I with unit vector Ωi and let ωi(a) = haΩi, Ωii for a ∈ B(Hi) where we realize B(Hi) as subalgebras of B(H) via the standard embedding from [11, Definition 1.5.1]. Then for every φ ∈ C , there exists a linear completely bounded normal map Mφ on B(H) such that Mφ(1) = φ(0)1 and Mφ(a1a2 . . . an) = φ(n)a1a2 . . . an whenever n ≥ 1, i1, . . . in ∈ I with i1 6= i2 6= · · · 6= in and aj ∈ B(Hij )= ker(ωij ). Moreover, kMφkcb ≤ kφkC . The proof of Proposition 5.1 will be divided into a series of lemmas. Lemma 5.2. Let T : B(H) → B(H) be a bounded linear normal map, and let φ : N0 → C. The following statements are equivalent. (a) For all n ≥ 1, i1, . . . in ∈ I with i1 6= i2 6= · · · 6= in and aj ∈ B(Hij ) = ker(ωij ), we have T (1) = φ(0)1 and T (a1a2 . . . an) = φ(n)a1a2 . . . an. (b) For all k, l ≥ 0 and ξ ∈ Λ(k), η ∈ Λ(l) we have T (LξL∗ η) =(cid:26) φ(k + l)LξL∗ φ(k + l − 1)LξL∗ η η in Case 1 in Case 2. Proof. Assume (a) and let k, l ≥ 1, ξ ∈ Λ(k), η ∈ Λ(l). Now by the definition of Lξ we have LξL∗ η = Lξ1 . . . Lξk L∗ ηl . . . L∗ η1. If we are in Case 1, there exist i, j ∈ I, i 6= j such that ξk ∈ Γi and ηl ∈ Γj. Hence all adjacent terms above are from different B(Hi), hence LξL∗ η is of the form a1 . . . an in (a) with n = k + l. On the other hand, if we are in Case 2, there exists i ∈ I such that η is of the form ηl ∈ B(Hi), hence LξL∗ In this case LξkL∗ ξk, ηl ∈ Γi. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 11 a1 . . . an in (a) with n = k + l − 1. Applying (a) we get the conclusion of (b) for k, l ≥ 1. If k = 0 or l = 0, e.g., ξ = Ω or η = Ω the result follows similarly by using LΩ = 1. Assume (b). Using Kaplansky's density theorem [8, Theorem 5.3.5] and the fact that the product is jointly σ-strong continuous on bounded sets, by Lemma 3.1 it is enough to check that T (1) = φ(0)1 and T (a1 . . . an) = φ(n)a1 . . . an whenever n ≥ 1 and aj ∈ {Lγγ ∈ Γij } ∪ {L∗ δγ, δ ∈ Γij } where i1 6= i2 6= · · · 6= in. In γLδ = 0 when γ ∈ Γij and δ ∈ Γij+1, since ij 6= ij+1. δ1 for some particular, L∗ Hence a1 . . . an = 0, unless a1 . . . an = Lγ1 . . . Lγk L∗ δl γj ∈ Γij , δs ∈ Γrs, ij 6= ij+1, rs 6= rs+1 and i1, . . . , ik, r1, . . . , rl ∈ I. γLδ = 0 when γ, δ ∈ Λ(1), γ 6= δ. . . . L∗ γγ ∈ Γij } ∪ {LγL∗ It is easy to check that L∗ If we are in Case 1, we have ik 6= rl. Hence neighboring elements on the right hand side are from different B(Hi)and thus n = k + l. ∈ B(Hik ), thus If we are in Case 2, we have ik = rl. Hence Lγk L∗ δl n = k + l − 1. Now (b) gives the result for k ≥ 1 or l ≥ 1. Moreover the k = l = 0 case of (b) gives T (1) = φ(0)1. (cid:3) Next, we explicitly construct such a map T . Lemma 5.3. Let φ ∈ C . Define maps T1 = ∞ Xi=1 Φ(1) xi,yi and T2 = Φ(2) zi,wi ∞ Xi=1 where Φ(·) x,y are as in Definition 4.1, ψ1, ψ2 as in Lemma 4.6, and xi, yi, zi, wi as in Remark 4.7. Then for all k, l ≥ 0, ξ ∈ Λ(k), η ∈ Λ(l) we have T1(LξL∗ η) = ψ1(k + l)LξL∗ η, respectively, T2(LξL∗ η) =(cid:26) ψ2(k + l)LξL∗ ψ2(k + l − 2)LξL∗ η η in Case 1 in Case 2. Proof. Let k, l ≥ 0 and ξ ∈ Λ(k), η ∈ Λ(l). Now by Lemma 4.5 and Lemma 4.8 we have ∞ T1(LξL∗ η) = Φ(1) xi,yi(LξL∗ η) Xi=1 = ∞ Xi=1 ∞ Xt=0 xi(k + t)yi(l + t)! (LξL∗ η) = ψ1(k + l)LξL∗ η. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 12 Furthermore, in Case 1 we have by Lemma 4.5 and Lemma 4.8 ∞ T2(LξL∗ η) = Φ(2) zi,wi(LξL∗ η) Xi=1 = ∞ Xi=1 ∞ Xt=0 = ψ2(k + l)LξL∗ η, zi(k + t)wi(l + t)! (LξL∗ η) respectively, in Case 2 ∞ T2(LξL∗ η) = Φ(2) zi,wi(LξL∗ η) Xi=1 = ∞ Xi=1 ∞ Xt=0 = ψ2(k + l − 2)LξL∗ η. zi((k − 1) + t)wi((l − 1) + t)! (LξL∗ η) This completes the proof. (cid:3) Lemma 5.4. Define T = T1 + T2 + c Id where Id denotes the identity operator on B(H). Then for k, l ≥ 0 and ξ ∈ Λ(k), η ∈ Λ(l) we have T (LξL∗ η) =(cid:26) φ(k + l)LξL∗ φ(k + l − 1)LξL∗ η η in Case 1 in Case 2. Note that by Lemma 5.2 this implies that T (1) = φ(1)1 and that for n ≥ 1, Tφ(a1a2 . . . an) = φ(n)a1a2 . . . an. Proof. Assume we are in Case 1, then T (LξL∗ η) = T1(LξL∗ η) + T2(LξL∗ η) + cLξL∗ η = (ψ1(k + l) + ψ2(k + l) + c) LξL∗ η = φ(k + l)LξL∗ η. Here we use the definition of T , then Lemma 5.3, and lastly Lemma 4.6. If we are in Case 2, we similarly get η) + T2(LξL∗ η) = T1(LξL∗ η) + cLξL∗ T (LξL∗ η = (ψ1(k + l) + ψ2(k + l − 2) + c) LξL∗ η = (ψ2(k + l − 1) + ψ1(k + l − 1) + c) LξL∗ η = φ(k + l − 1)LξL∗ η. Here we furthermore use ψ2(n) = ψ1(n + 1), for n ≥ 0. (cid:3) By this result we have proven the existence of Mφ in Proposition 5.1, and it remains to calculate the cb-norm. Lemma 5.5. We have kT kcb ≤ kφkC . RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 13 Proof. Let xi, yi ∈ l2(N0) then we have by Lemma 4.4 that kΦ(1) xi,yikcb ≤ xi,yi we have by Remark i=1 Φ(1) 4.7 kxik2kyik2. Furthermore, since T1 = P∞ Xi=1 xi,yikcb ≤ kT1kcb ≤ ∞ Xi=1 kΦ(1) respectively, kxik2kyik2 = khk1, ∞ Xi=1 ∞ ∞ kT2kcb ≤ kΦ(2) zi,wikcb ≤ kzik2kwik2 = kkk1. Xi=1 Hence kT kcb ≤ kT1kcb + kT2kcb + kcIdkcb ≤ khk1 + kkk1 + c = kφkC as desired. (cid:3) Combining Lemmas 5.4 and 5.5 we obtain Porposition 5.1, and there- fore an application of Proposition 2.5 yields the conclusion of Theorem 2.2. 6. Examples 6.1. The case φs(n) = sn. As a first example we will look at a simple φ where kφkC can be calculated explicitly. Corollary 6.1. Let D = {s ∈ Cs < 1} and s ∈ D. Denote by φs the function φs(n) := sn. Then φs defines a radial multiplier Mφs on A = ∗i∈I(Ai, ωi), respectively, (M , ω) = ¯∗i∈I(Mi, ωi) as in Theorem 2.2. Moreover, kMφskcb ≤ 1 − s/(1 − s). Proof. The conclusion follows from Theorem 2.2, once we show that φ belongs to C and that kφkC = khk1 + kkk1 + c ≤ 1 − s/(1 − s). Observe first that c = limn→∞ φ(n) = limn→∞ sn = 0 as c < 1. Furthermore, φ(i + j + 1) − φ(i + j + 2) = s(φ(i + j) − φ(i + j + 1)) so k = s · h, hence kφkC = (1 + s)khk1. Moreover φ(i + j) − φ(i + j + 1) = (1 − s)si+j so h = (1 − s)m, where m is the matrix mi,j = si+j. This gives kφkC = (1+s)khk1 = (1+s)1−skmk1. Now m = a⊙¯a, where 2 = 1/(1 − s2). Combining a = (sk)k≥0 ∈ l2(N0), hence kmk1 = kak2 these calculations we get kφkC = 1 − s/(1 − s), which proves the corollary. (cid:3) 6.2. Wysoczański's theorem. As a second example, we will show that Wysoczański's result, apart from determining when equality holds, is a special case of Theorem 2.2. Theorem 6.2 ([12, Theorem 6.1]). Let G = ∗i∈IGi be the free product of a family of discrete groups, and let g ∈ G be g = g1g2 . . . gn where gj ∈ Gij \{e}, j1, . . . , jn ∈ I and j1 6= j2 6= · · · 6= jn. If φ ∈ C then φ(g) = φ(n) is a Herz-Schur multiplier on G. Moreover k φkHS ≤ kφkC . RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 14 Proof. Let φ ∈ C and g = g1 . . . gn ∈ G as above. Now by [1, p. 301] and [4] we have k φkHS = k φkM0A(G) = k M φkcb where M φ is the operator M φ(λ(g)) = φ(g)λ(g) for g ∈ G and λ the left regular rep- resentation. By the definition of φ this is M φ(λ(g)) = φ(n)λ(g) and by the definition of L(G) we have λ(g) = λ(g1)λ(g2) . . . λ(gn). Hence M φ(λ(g)) = Mφ(λ(g)), where Mφ is as defined in Theorem 2.2. Ap- plying the theorem one obtains that φ is a Herz-Schur-multiplier, and k φkHS ≤ kφkC . (cid:3) 6.3. Relation to Houdayer and Ricard's results. Recently, C. Houdayer and É. Ricard in [7] proved results concerning radial mul- tipliers on free Araki-Woods factors related to functions from a class C ′ quite similar to C . Although their results apply to different objects than those considered in this paper, we will discuss in this section the issue of how their methods could be applied to prove Theorem 2.2. We start by defining the class of functions C ′, mentioned above. Definition 6.3. Let C ′ denote the set of functions φ : N0 → C for which the Hankel matrix h = (φ(i + j) − φ(i + j + 2))i,j≥0 is of trace- class. Observe that this implies the existence of c1, c2 ∈ C and a unique ψ : N0 → C such that φ(n) = c1+(−1)nc2+ψ(n) and limn→∞ ψ(n) = 0. For φ ∈ C ′ put kφkC ′ = c1 + c2 + khk1. In [7] the following two results for functions in the class C ′ are proved, note the resemblance with Theorem 2.2. In what follows T denotes the Toeplitz algebra, and furthermore, Γ(H, Ut)′′ denotes the free Araki- Woods factor associated to a real Hilbert space H and a one parameter group of orthogonal transformations (Ut). (See [7, Sections 2.5 and 3.1] for more precise definitions). Theorem 6.4 ([7, Proposition 3.3]). A function φ belongs to C ′ if and only if the operator γ defined by γ(Si(S∗)j) = φ(i + j) extends to a bounded map on T . Moreover, kγkT ∗ = kφkC ′, and we say that γ is the radial functional associated with φ. Theorem 6.5 ([7, Theorem 3.5]). Let φ : N0 → C. Then φ de- fines a completely bounded radial multiplier on Γ(H, Ut)′′ if and only if the radial functional γ on T associated to φ is bounded. Moreover, kMφkcb = kγkT ∗. A similar argument as in the proof of these theorems could be used to prove Theorem 2.2, if one could prove the existence of a ∗-isomorphism π : C ∗(Lγγ ∈ Λ(1)) → C ∗(Lγγ ∈ Λ(1)) ⊗ C ∗(S2, SS∗) such that π(LξL∗ η) =(cid:26) LξL∗ LξL∗ η ⊗ S2k(S∗)2l η ⊗ S2k−1(S∗)2l−1 in Case 1 in Case 2 (6.1) RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 15 for all k, l ≥ 0 and ξ ∈ Λ(k), η ∈ Λ(l). Indeed, if this were the case we could choose w ∈ C ∗(S2, SS∗)∗ as w(Sk(S∗)l) =(cid:26) φ(cid:0) k+l 2 (cid:1) 0 if k + l even otherwise. This functional would be bounded if h were trace-class. Moreover kId ⊗ wkcb = kwk = kφkC . Letting T be defined by T = (Id ⊗ w) ◦ π we would get T (LξL∗ η) =(cid:26) φ(k + l)LξL∗ φ(k + l − 1)LξL∗ η η in Case 1 in Case 2 with kT kcb ≤ kwkkπkcb ≤ kwk = kφkC . Hence, by Lemma 5.2 T would be Mφ as defined in Theorem 2.2 and thus be completely bounded. It is however not possible to construct such an isomorphism. Let for instance I = 1 and dim(H) = 2 and let eij denote the matrix units with respect to the basis (Ω, γ) of H. Then we have e01e10 = e00, but Φ(e01)Φ(e10) = e00 ⊗ S∗S 6= 1 ⊗ 1 − e11 ⊗ SS∗ = Φ(e00). However, note that it would be sufficient if there existed a unital completely positive π satisfying (6.1). To find such an operator we can regard l2(N0) = l2(N0)even ⊕ l2(N0)odd. In this case S2 on l2(N0) can be realized as S ⊕ S and SS∗ on l2(N0) can be realized as SS∗ ⊕ 1. Then it would be enough to find unital completely positive operators π1, π2 such that π1(LξL∗ η) = LξL∗ η ⊗ Sk(S∗)l, respectively, π2(LξL∗ η) =(cid:26) LξL∗ LξL∗ η ⊗ Sk(S∗)l η ⊗ Sk−1(S∗)l−1 (6.2) (6.3) in Case 1 in Case 2. Since π(1) = π(2) = 1 we have kπ1kcb, kπ2kcb ≤ 1 and π = π1 ⊕ π2 is unital completely positive too. Hence T = (Id ⊗ w) ◦ π would be as desired. i=0 Pi+n⊗ei0, where eij are the matrix units in B(l2(N0)), and use the convention Pm = 0 if m < 0. Now it can be shown that Set Un =P∞ π1(x) = 0 Xn=−∞ 0 Un(x ⊗ 1)U ∗ n + ∞ Xn=1 ∞ Un(ρn(x) ⊗ 1)U ∗ n, Un(x ⊗ 1)U ∗ n + Un(ρn−1(ǫ(x)) ⊗ 1)U ∗ n respectively, π2(x) = Xn=−∞ Xn=1 are unital completely positive and fulfill (6.2) and (6.3). The proof of this fact can be given by an argument quite similar to that given in Sections 4 and 5. We leave the details to the reader. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 16 7. Integral representation of functions from C In [6] the following integral representation was proved for φ ∈ C ′, where C ′ is the set of functions φ : N0 → C from Definition 6.3. The set C ′ is not defined in [6], but the result follows form [6, Theorem 2.12 and Theorem 4.2]. Theorem 7.1. Let ψ : N0 → C be a function. Then the following are equivalent: (1) ψ ∈ C ′ (2) There exists a complex Borel measure µ on D and constants c+, c− ∈ C, such that ψ(n) = c+ + (−1)nc− +ZD and sndµ(s) < ∞ (7.1) 1 − s2 1 − s2 dµ(s) < ∞. ZD Moreover, for φ ∈ C ′, the measure µ in (7.1) can be chosen such that c+ + c− +ZD 1 − s2 1 − s2 dµ(s) ≤ 8 π kψkC ′. We will prove next a similar characterization of functions in C : Theorem 7.2. Let φ : N0 → C be a function. Then the following are equivalent: (1) φ ∈ C (2) There exists a constant c ∈ C and a complex Borel measure ν sndν(s) (7.2) Moreover, for φ ∈ C , the measure ν in (7.2) can be chosen such that c +ZD 1 − s 1 − s dν(s) ≤ 8 π kφkC . Proof. (1) implies (2). Let φ ∈ C and put φ(n) =(cid:26) φ( n 2 ) 0 if n is even if n is odd. Then by Definition 2.1 and Definition 6.3, φ ∈ C ′ and k φkC ′ = kφkC . From Theorem 7.1 there exists a complex measure µ on D and constants on D such that and φ(n) = c +ZD ZD 1 − s 1 − s dν(s) < ∞. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 17 c+, c− ∈ C such that and sndµ(s) < ∞ ψ(n) = φ(2n) = c+ + c− +ZD c+ + c− +ZD 1 − s2 1 − s2 dµ(s) ≤ Let ν be the range measure of µ by the map s → s2 of D onto D, and put c = c+ + c−. Then ν is less or equal to the range measure of µ by the map s → s2. Hence 8 π kψkC . sndν(s) s2ndµ(s) = c +ZD φ(n) = c +ZD dν(s) ≤ c+ + c− +ZD and c +ZD 1 − s 1 − s 1 − s2 1 − s2 dµ(s) ≤ 8 π kφkC . This proves (1) implies (2) and the last statement in Theorem 7.2. Conversely if (2) holds, the Hankel matrices h, k from Definition 2.1 have the entries and hij =ZD kij =ZD si+j(1 − s)dν(s) si+js(1 − s)dν(s). By the proof of Corollary 6.1, k(si+j)i,j≥0k1 = 1 1 − s2 , s ∈ D. Hence khk1 + kkk1 ≤ZD 1 − s + s(1 − s) 1 − s2 dν(s) =ZD 1 − s 1 − s dν(s) < ∞ which shows that φ ∈ C . (cid:3) In [9] N. Ozawa proved that if Γ is a discrete hyperbolic group (in the sense of M. Gromov [5]), then Γ is weakly amenable. The proof was obtained by showing that the metric d : Γ × Γ → N0 (w.r.t. the Cayley graph of Γ) satisfies three properties (1), (2) and (3) listed in [9, Theorem 1]. As an application of Theorem 7.2, we will show below, that the first condition (1) from [9] is sufficient to prove that Γ is weakly amenable. For the definition of weak amenability and of the constant Λ(Γ) for a weakly amenable group Γ, we refer to [2, Section 12.3]. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 18 Recall that a metric on a discrete metric space (X, d) is called proper if the ball B(x, r) = {y ∈ X : d(x, y) < r} is finite for all x ∈ X and all r > 0. Theorem 7.3. Let Γ be a discrete countable group and let d : Γ × Γ → N0 be a proper left invariant metric. Put φs(x) = sd(x,e), s ∈ D, x ∈ Γ. Assume that there exists a constant C ≥ 1, such that ψs ∈ M0A(Γ) for all s ∈ D and kφskM0A(Γ) ≤ C 1 − s 1 − s , s ∈ D. Then Γ is weakly amenable with constant Λ(Γ) ≤ C. Remark 7.4. As in [6] we have used the notation M0A(Γ) for the set of completely bounded Fourier multipliers on Γ. Note that in [2, Section 12.3] the space M0A(Γ) is denoted B2(Γ). We first prove Lemma 7.5. (1) Put χn(k) = δkn for n, k ≥ 0. Then χn ∈ C and kχnkC ≤ max{1, 4n}, n ≥ 0. (2) For r ∈ (0, 1) and l ≥ 0, put φr(k) = rk φr,n(k) =(cid:26) rk 0 ≤ k ≤ n k > n. 0 then φr, φr,n ∈ C , kφrkC = 1 and for fixed r ∈ (0, 1) lim n→∞ kφr − φr,nkC = 0. Proof. From Definition 2.1 we have χn ∈ C , and kχnkC = kHnk1 + kKnk1 where Hn(i, j) = χn(i + j) − χn(i + j + 1) Kn(i, j) = χn(i + j + 1) − χn(i + j + 2). If H = (hij)∞ i,j=0 is a matrix of complex numbers for whichPi,j hij < ∞, then H is of trace class and kHk1 ≤ P∞ i,j=0 hij. Hence kχnkC ≤ (2n+ 1) + (2n−1) for n ≥ 1 and kχ0kC ≤ 1 which proves (1). It follows from Corollary 6.1, that kφrkC = 1, 0 < r < 1. By (1), kφr − φr,nkC = k which proves (2). rkχkkC ≤ ∞ Xk=n+1 4krk ∞ Xk=n+1 (cid:3) RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 19 Proof of Theorem 7.3. Let φ : N0 → C be a function from C , and put φ(x) = φ(d(x, e)), x ∈ Γ. Then by (7.1) and the integral representation of φ from Theorem 7.2 it follows that φ is a completely bounded Fourier multiplier on Γ and that k φkM0A(Γ) ≤ 8 π kφkC . (7.3) Let φr and φr,n be as in Lemma 7.5. Then by (7.1) k φrkM0A(Γ) ≤ C. Moreover by (7.3) and Lemma 7.5 k φr − φr,nkM0A(Γ) = 0 lim n→∞ for fixed r ∈ (0, 1). Put rk = 1 − 1/k, k ≥ 1 and chose for each k ≥ 2 an nk ≥ k, such that k φrk − φrk,nkkM0A(Γ) ≤ 1 k . Then ψk = φrk,nk form a sequence of finitely supported functions on Γ, such that kψkkM0A(Γ) < C + 1/k, and limk→∞ ψk(x) = 1 for all x ∈ Γ. Hence Γ is weakly amenable and Λ(Γ) ≤ C. (cid:3) Remark 7.6. By [9, Theorem 1] and the proof of Theorem 7.3 it follows that for every hyperbolic group Γ and every φ ∈ C ′, the function φ(x) = φ(d(x, e)), x ∈ Γ is a completely bounded Fourier multiplier on Γ and k φkM0A(Γ) ≤ π kφkC ′, where C is the constant in [9, Theorem 1 (1)]. 8C References [1] M. Bożejko and G. Fendler. Herz-Schur multipliers and completely bounded multipliers of the Fourier algebra of a locally compact group. Boll. Un. Mat. Ital. A (6), 3(2):297 -- 302, 1984. [2] N. P. Brown and N. Ozawa. C ∗-algebras and finite-dimensional approxima- tions, volume 88 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2008. [3] E. Christensen and A. M. Sinclair. A survey of completely bounded operators. Bull. London Math. Soc., 21(5):417 -- 448, 1989. [4] J. De Cannière and U. Haagerup. Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups. Amer. J. Math., 107(2):455 -- 500, 1985. [5] M. Gromov. Hyperbolic groups. In Essays in group theory, volume 8 of Math. Sci. Res. Inst. Publ., pages 75 -- 263. Springer, New York, 1987. [6] U. Haagerup, T. Steenstrup, and R. Szwarc. Schur multipliers and spherical functions on homogeneous trees. Internat. J. Math., 21(10):1337 -- 1382, 2010, 0908.4424. [7] C. Houdayer and É. Ricard. Approximation properties and absence of Cartan subalgebra for free Araki-Woods factors. Adv. Math., 228(2):764 -- 802, 2011, 1006.3689. RADIAL MULTIPLIERS ON REDUCED FREE PRODUCTS 20 [8] R. V. Kadison and J. R. Ringrose. Fundamentals of the theory of operator algebras. Vol. I, volume 15 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1997. Elementary theory, Reprint of the 1983 original. [9] N. Ozawa. Weak amenability of hyperbolic groups. Groups Geom. Dyn., 2(2):271 -- 280, 2008, 0704.1635. [10] É. Ricard and Q. Xu. Khintchine type inequalities reduced free products and applications. J. Reine Angew. Math., 599:27 -- 59, 2006, arXiv:math/0505302. for [11] D. V. Voiculescu, K. J. Dykema, and A. Nica. Free random variables, volume 1 of CRM Monograph Series. American Mathematical Society, Providence, RI, 1992. A noncommutative probability approach to free products with applica- tions to random matrices, operator algebras and harmonic analysis on free groups. [12] J. Wysoczański. A characterization of radial Herz-Schur multipliers on free products of discrete groups. J. Funct. Anal., 129(2):268 -- 292, 1995. Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark E-mail address: [email protected] Department of Mathematics and Computer Science, University of Southern Denmark, Campusvej 55, 5230 Odense M, Denmark E-mail address: [email protected]
1703.07073	3	1703	2018-03-18T04:32:24	Heisenberg Modules over Quantum 2-tori are metrized quantum vector bundles	[ "math.OA" ]	The modular Gromov-Hausdorff propinquity is a distance on classes of modules endowed with quantum metric information, in the form of a metric form of a connection and a left Hilbert module structure. This paper proves that the family of Heisenberg modules over quantum two tori, when endowed with their canonical connections, form a family of metrized quantum vector bundles, as a first step in proving that Heisenberg modules form a continuous family for the modular Gromov-Hausdorff propinquity.	math.OA	math	HEISENBERG MODULES OVER QUANTUM 2-TORI ARE METRIZED QUANTUM VECTOR BUNDLES FRÉDÉRIC LATRÉMOLIÈRE Abstract. The modular Gromov-Hausdorff propinquity is a distance on classes of modules endowed with quantum metric information, in the form of a metric form of a connection and a left Hilbert module structure. This paper proves that the family of Heisenberg modules over quantum two tori, when endowed with their canonical connections, form a family of metrized quantum vector bundles, as a first step in proving that Heisenberg modules form a continuous family for the modular Gromov-Hausdorff propinquity. . A O h t a m [ 3 v 3 7 0 7 0 . 3 0 7 1 : v i X r a 1. Introduction The primary purpose of our research is to bring forth an analytic framework, con- structed around Gromov-Hausdorff-like hypertopologies on quantum metric spaces, to bear on problems from mathematical physics and noncommutative geometry [17, 13, 19, 18, 14, 12, 3, 15]. We constructed an hypertopology on classes of Hilbert modules over quantum metric spaces in [16] as a far-reaching generaliza- tion of the Gromov-Hausdorff propinquity. We constructed a distance, up to full quantum isometry, called the modular Gromov-Hausdorff propinquity, on a class of objects which generalize Hermitian vector bundles over Riemannian manifolds. These metrized quantum vector bundles are natural objects for noncommutative geometry and mathematical physics, as they carry a metric structure and a form of generalized connection, and we are now able to discuss such questions as continuity and approximations, not only of quantum compact metric spaces, but also of their associated modules. As modules are fundamental objects in C-algebra theory and their geometry, this new development allows us to further our goal of a geometric theory of the class of C-algebras. This paper brings into our noncommutative metric geometry framework some very important examples of modules, namely Heisenberg modules over quantum 2-tori. These modules come naturally equipped with a connection induced by the action of the Heisenberg Lie group. This noncommutative construct played the central role in the beginning of Connes' noncommutative geometry [5], where the Heisenberg modules over quantum 2-tori and their connections were first built. Rieffel [27] then proved that these Heisenberg modules, the finite rank free mod- ules, and their direct sums, describe all the finitely generated projective modules Date: September 17, 2018. 2000 Mathematics Subject Classification. Primary: 46L89, 46L30, 58B34. Key words and phrases. Noncommutative metric geometry, Gromov-Hausdorff convergence, Monge-Kantorovich distance, Quantum Metric Spaces, Lip-norms, D-norms, Hilbert modules, noncommutative connections, noncommutative Riemannian geometry, unstable K-theory. This work is part of the project supported by the grant H2020-MSCA-RISE-2015-691246- QUANTUM DYNAMICS. 1 2 FRÉDÉRIC LATRÉMOLIÈRE over quantum tori. Connes and Rieffel [7] proved that the natural connections on Heisenberg modules solve the noncommutative Yang-Mills problem. We will now prove that Heisenberg modules are fundamental examples of metrized quantum vector bundles. Doing so then allows us to discuss in [20] the continuity, for the modular propinquity, of family of Heisenberg modules as the quantum 2-tori vary continuously for the propinquity. This will be our first, significant application of the modular propinquity. Informally, the continuity result in [20] can be understood as a form of continuity of K-theory. Thus, this paper and [20] are two parts of the study of the metric geometry of Heisenberg modules. As a matter of convention throughout this paper, we will use the following no- tations. Notation 1.1. By default, the norm of a normed vector space E is denoted by k·kE. When A is a C-algebra, the space of self-adjoint elements of A is denoted by sa (A). The state space of A is denoted by S (A). In this work, all C-algebras A will always be unital with unit 1A. Convention 1.2. If P is some seminorm on a vector subspace D of a vector space E, then for all x ∈ E \ D we set P (x) = ∞. With this in mind, the domain D of P is the set {x ∈ E : P (x) < ∞}, with the usual convention that 0∞ = 0 while all other operations involving ∞ give ∞. Noncommutative metric geometry [6, 28, 30] studies noncommutative general- izations of Lipschitz algebras, defined as follows. Definition 1.3. An ordered pair (A, L) is a Leibniz quantum compact metric space when A is a unital C-algebra, and L is a seminorm defined on a dense Jordan-Lie subalgebra dom (L) of the space of self-adjoint elements sa (A) of A such that: (1) {a ∈ dom (L) : L(a) = 0} = R1A, (2) the Monge-Kantorovich metric mkL defined on the state space S (A) of A by setting, for any two ϕ, ψ ∈ S (A): mkL(ϕ, ψ) = sup{ϕ(a) − ψ(a) : a ∈ dom (L), L(a) 6 1} metrizes the weak topology restricted to S (A), (3) L is lower semi-continuous, (4) max(cid:8)L(cid:0) ab+ba 2 (cid:1) , L(cid:0) ab−ba 2i (cid:1)(cid:9) 6 kakAL(b) + kbkAL(a). Leibniz quantum compact metric spaces, and more generally quasi-Leibniz quan- tum compact metric spaces (a generalization we will not need in this paper), form a category with the appropriate notion of Lipschitz morphisms [21], con- taining such important examples as quantum tori [28], Connes-Landi spheres [22], group C-algebras for Hyperbolic groups and nilpotent groups [29, 23], AF alge- bras [12], Podlès spheres [2], certain C-crossed-products [1], among others. Any compact metric space (X, d) give rise to the Leibniz quantum compact metric space (C(X), Lip) where C(X) is the C-algebra of C-valued continuous functions over X, and Lip is the Lipschitz seminorm induced by d. Rieffel characterized the main property of Leibniz quantum compact metric spaces as follows: Theorem 1.4 ([28, Theorem 1.9]). Let (A, L) be a pair with a unital C-algebra A and a seminorm L defined on a dense subspace dom (L) of sa (A). The following assertions are equivalent: D-NORMS ON HEISENBERG MODULES 3 (1) the Monge-Kantorovich metric mkL defined for any two ϕ, ψ ∈ S (A) by mkL(ϕ, ψ) = sup{ϕ(a) − ψ(a) : L(a) 6 1}, metrizes the weak* topology on S (A), (2) the diameter diam (S (A), mkL) is finite and: {a ∈ sa (A) : L(a) 6 1 and kakA 6 1} is norm precompact. In [16], we extend this idea to noncommutative analogues vector bundles. Our classical prototype of a metrized quantum vector bundle is given by the module ΓV of continuous sections of a vector bundle V over a compact Riemannian man- ifold M with metric g, endowed with a hermitian metric h and some associated metric connection ∇. For any two ω, η ∈ ΓV , we then set hω, ηiV : x ∈ M 7→ RX hx(ωx, ηx) dVol(x) where Vol is the volume form over M for g, which turns ΓV into a C(M )-left Hilbert module. We also define, for all ω ∈ M , the norm D(ω) as the operator norm for the operator ∇ω : X ∈ Γ(T M ) 7→ ∇X ω ∈ ΓV -- noting that the space of vector fields ΓT M of M has a norm induced by the metric g. Our general definition for a metrized quantum vector bundle abstracts this picture. For the present paper, we shall only deal with so-called Leibniz metrized quantum vector bundles, even though our framework in [16] is more general. This is the main definition for this paper. Definition 1.5 ([16, Definition 3.8]). A 5-tuple (M ,h·,·iM , D, A, L) is a metrized quantum vector bundle when: (1) a Leibniz quantum compact metric space (A, L) called the base space, (2) a A-left Hilbert module (M ,h·,·iM ), (3) a norm D defined on a dense subspace of M such that D(ω) > phω, ωiM for all ω ∈ M , and such that the set: is compact in M , {ω ∈ M : D(ω) 6 1} (4) for all a ∈ sa (A) and for all ω ∈ M , we have: DM (aω) 6 (kakA + LA(a))DM (ω), which we call the inner Leibniz inequality for DM , (5) for all ω, η ∈ M , we have: max{LA (ℜhω, ηiM ) , LA (ℑhω, ηiM )} 6 2DM (ω)DM (η), which we call the modular Leibniz inequality for DM . We refer to [16] for a discussion of these objects, where in particular [16, Ex- ample 3.10] shows that the prototype of a hermitian vector bundle over a compact Riemannian manifold, as sketched above, is indeed an example of a metrized quan- tum vector bundle. We note that Definition (1.5) includes a compactness condition which mirrors the compactness condition in Theorem (1.4). Heisenberg modules, equipped with the analogue of a connection as in [5], over quantum 2-tori, have a similar signature to a metrized quantum vector bundle. The key difficulty is to prove that the connection can be used to define a D-norm, as in Definition (1.5), whose unit ball is actually compact in the Hilbert modules norm of Heisenberg modules. The main result of this paper is to prove that indeed, this is the case. 4 FRÉDÉRIC LATRÉMOLIÈRE We begin our work with a presentation of Heisenberg modules, which allow us to fix our notations for the rest of the paper and [20]. We then prove a series of lemmas about convergence in the Hilbert modules norm for the Heisenberg modules -- as these norms are complicated, these lemmas will prove very helpful both in this paper and in [20]. We prove in the process of this second section that Heisenberg modules form a continuous field of Banach spaces -- a result which will prove helpful in [20] and is of independent interest. This result uses the same tools as the proof that the action of the Heisenberg group on Heisenberg modules is strongly continuous, which is part of the next section of this paper, where properties of the Heisenberg group actions which we will need in our work are established. Now, with all these basic tools in hand, we show how to use Lie group actions to define D-norm candidates, which have all the desired properties of D-norms except maybe for the key compactness property of their unit ball. This compactness property for the Heisenberg modules D-norms is the subject of the last section of this paper, which conclude our main result. Importantly, our methods in this paper are designed not only in support of the main theorem here, but also as key tools for the study of the continuity of the Heisenberg modules in [20]. For the problem of continuity, we will need not just to be able to pick finite subsets of the compact unit ball of some D-norm which are ε-dense for some ε > 0, but also to pick such a finite set which is uniformly ε-dense across several Heisenberg modules as the D-norms vary. To do so, we will use the approximation operators introduced in the last section of this paper. 2. Background on Quantum 2-tori and Heisenberg modules Quantum 2-tori are the twisted convolution C-algebras of Z2. The projective finitely generated modules over quantum tori have been extensively studied, and next to the free modules, the most important class of projective, finitely generated modules over a quantum torus are the Heisenberg modules. This subsection intro- duces these modules, as well as the notations we will use throughout this section regarding quantum tori. Twisted group C-algebras are defined by twisting the convolution product over a locally compact group by a representative of a continuous 2-cocycle of the group. Notation 2.1. For any θ ∈ R, we define the skew bicharacter of R2: (2.1) eθ : ((x1, y1), (x2, y2)) ∈ R2 × R2 7−→ exp (iπθ(x2y1 − x1y2)) . By [10], any 2-cocycle of Z2 is cohomologous to the restriction of a skew bichar- acter eθ to Z2 × Z2 for some θ ∈ R. We shall use the same notation for eθ and its restriction to Z2. Moreover, for any θ, ϑ ∈ R, the skew bicharacters eθ and eϑ of Z2 are cohomol- ogous if and only if θ ≡ ϑ mod 1. We note that, as skew bicharacters of R2, they are cohomologous if and only if θ = ϑ. We define the twisted convolution products on ℓ1(Z2), where we use the following notation. Notation 2.2. For any (nonempty) set E and any p ∈ [1,∞), the set ℓp(E) is the set of all absolutely p-summable complex valued functions over E, endowed with D-NORMS ON HEISENBERG MODULES 5 the norm: for all ξ ∈ ℓp(E). element of ℓp(E) for all p. kξkℓp(E) = Xx∈E 1 p ξ(x)p! We write δn the function which is 1 at n and 0 otherwise; this function is an Moreover, if p = 2 then (ℓ2(E),k · kℓ2(E)) is a Hilbert space, where the inner product hξ, ηiℓ2(E) =Px∈E ξ(x)η(x) for all ξ, η ∈ ℓ2(E). Definition 2.3. Let θ ∈ R and eθ be defined by Expression (2.1). The twisted convolution product ∗θ is defined for all f, g ∈ ℓ1(Z2) and for all n ∈ Z2: We now define: f∗θg(n) = Xm∈Z2 f (m)g(n − m)eθ(m, n). The adjoint of any f ∈ ℓ1(Z2) is defined for all n ∈ Z2 by: f ∗(n) = f (−n). veloping C-algebra. To do so, we shall choose a natural faithful -representation One checks easily that (cid:0)ℓ1(Z2),∗θ,·∗(cid:1) is a -algebra. In particular, the adjoint operation is an isometry of (cid:0)ℓ1(Z2),k · kℓ1(Z2)(cid:1). We now wish to construct its en- of (cid:0)ℓ1(Z2),∗θ,·∗(cid:1) on ℓ2(Z2). This representation was a key ingredient in the con- struction of bridges between quantum tori in our work in [11] on convergence of quantum tori for the quantum propinquity and will play a role in the convergence of Heisenberg modules. Notation 2.4. If T : E → F is a continuous linear map between two normed spaces, we write its norm as TE Theorem 2.5 ([32]). Let θ ∈ R. We define, for any n ∈ Z2 and ξ ∈ ℓ2(Z2), the function: F . When E = F , we simply write TF . is a unitary eθ-projective representation of Z2, i.e. θ θ = eθ(n, m)U n+m θ ξ : m ∈ Z2 7→ eθ(m, n)ξ(m + n). U n The map n ∈ Z2 7→ U n for all n, m ∈ Z2. θ U m U n If, for all f ∈(cid:0)ℓ1(Z2),∗θ,·∗(cid:1), we define: πθ(f ) = Xn∈Z2 f (n)U n θ θ which is a bounded operator on ℓ2(Z2) with: πθ(f )ℓ2(Z2) 6 kfkℓ1(Z2), then πθ is a faithful -representation of (ℓ1(Z2),∗θ,∗). Thus, we may define a C-norm on ℓ1(Z2) by setting: kfkAθ = πθ(f )ℓ2(Z2) for all f ∈ ℓ1(Z2). We thus can define quantum 2-tori. Definition 2.6. The quantum 2-torus Aθ is the completion of (ℓ1(Z2),∗θ,·∗) for the norm πθ(·)ℓ2(Z2). 6 FRÉDÉRIC LATRÉMOLIÈRE As per our general convention, the norm on Aθ is denoted by k·kAθ for all θ ∈ R. Remark 2.7. Let θ ∈ R. By construction, ℓ1(Z2) is identified with a dense - subalgebra of Aθ, and we shall employ this identification all throughout this paper. With this identification, we also note that for all f ∈ ℓ1(Z2) we have kfkAθ 6 kfkℓ1(Z2), a fact which we will use repeatedly in the next section. We take one derogation to the convention of using the same symbol for an element of ℓ1(Z2) and its counter part in a given quantum torus, because the following notation is at once common and convenient. Notation 2.8. Let θ ∈ R. The element δ1,0 is denoted by uθ and the element δ0,1 is denoted by vθ when regarded as elements of Aθ. The geometry, and in particular the metric geometry [28], of the quantum tori is obtained by transport of structure using the dual action of the torus given as follows: Theorem-Definition 2.9. [32] For all z = (z1, z2) ∈ T2 there exists a unique -automorphism βz θ of Aθ such that, for any f ∈ ℓ1(Z2) and (n, m) ∈ Z2, we have: βz θ f (n, m) = zn 1 zm 2 f (n, m). The map z ∈ T2 7→ βz action. Moreover, β is ergodic, in the sense that: θ is a strongly continuous action of T2 on Aθ called the dual (cid:8)a ∈ Aθ : ∀z ∈ T2 βz(a) = a(cid:9) = C1Aθ . We now turn to the class of modules to which we shall apply our new modular propinquity. We construct these modules following [5] using the universal property of quantum 2-tori, which we now recall. Proposition 2.10 ([32]). Let θ ∈ R. If U , V are two unitary operators on some Hilbert space H such that U V = exp(2iπθ)V U for some θ ∈ [0, 1), then there exists a -morphism : Aθ → B(H ) such that (uθ) = U and θ(vθ) = V . The range of is C∗(U, V ). Another way to state Proposition (2.10) is that, for any θ ∈ R, if ς is some projective representation of Z2 on some Hilbert space H for some multiplier of Z2 cohomologous to eθ, then H is a module over Aθ. Indeed, Proposition (2.10) gives us a -morphism from Aθ to the C-algebra B(H ) of all bounded linear operators on H , with (uθ) = ς 1,0 and (vθ) = ς 0,1. Thus H is a Aθ module. With this observation in mind, we now turn to the construction of some particular projective representations of Z2. The idea, found in [5] and explicit in [25], is to take the tensor product of a projective representation of R2, restricted to Z2, and a finite dimensional projective representation of Zq for some q ∈ N\{0}. By adjusting the choice of the multipliers associated with each projective representation, we get the desired module structure. Projective representations of R2 are naturally related to the representations of the Heisenberg group, and we will make important use of this fact in our work. We thus begin with setting our notations for the Heisenberg group. Convention 2.11. The vector space Cd is endowed by default with its standard j=1 zjyj, whose associated norm is inner product h(z1, . . . , zd), (y1, . . . , yd)iCd = Pd denoted by k · kCd. D-NORMS ON HEISENBERG MODULES 7 Notation 2.12. The Heisenberg group is the Lie group given by: H3 =    1 x u 0 y 0 1 0  : x, y, u ∈ R 1  . We shall identify H3 with R3 via the natural map (x, y, u) ∈ R3 7→   1 0 which is a Lie group isomorphism once we equip R3 with the multiplication: 1 x u 0 y 0 1 , (x1, y1, u1)(x2, y2, u2) = (x1 + x2, y1 + y2, u1 + u2 + x1y2) for all (x1, y1, u1), (x2, y2, u2) ∈ R3. gleaned by looking at its Lie algebra, which is given by: The importance of the Heisenberg group for quantum mechanics [8] may be which is a 2-nilpotent Lie algebra. We easily compute that for all x, y, u ∈ R3: (2.2) h =    0 x u 0 y 0 0 0 exp  0 0 0 x u 0 y 0  : x, y, u ∈ R 0  0  = 1   .  1 x u + 1 2 xy 0 y 0 1 0 This expression for the exponential will be important for our construction. Note that the exponential map is both injective and surjective. We now set: 0 0 0 P =  1 0 0 0 0 0  , Q =  0 0 0 0 0 0 1 0 0  and T =  0 0 0 0 0 0 1 0 0  . We easily check that [P, Q] = T = −[Q, P ] while other other commutators between P , Q and T are null, and spanC{P, Q, T} = h. We note that in particular, T is central, and thus the relations defining h from the basis {P, Q, T} are the structural equations of quantum mechanics -- the canonical commutation relation, as proposed by Heisenberg, in order to express the uncer- tainty principle between two conjugate observables. We refer to [8] for a detailed analysis of the Heisenberg group and its connections to the Moyal product, pseudo- differential calculus, and more fascinating topics. Thus the study of the irreducible representations of H3 provide the irreducible representations of the canonical commutation relations. We first note that: H3.{(0, 0, u) : u ∈ R} = R2 is Abelian, and thus we get a collection of trivial, one-dimensional representations of H3 by simply lifting the irreducible representations of R2. If we set, for any ð ∈ R \ {0} and ξ ∈ L2(R): (2.3) αx,y,u ð,1 ξ : s ∈ R 7→ exp(2iπ(ðu + sx))ξ(s + ðy) then we define a unitary representation of H3, and any nontrivial irreducible unitary representations of the Heisenberg group is unitarily equivalent to αð,1 for some 8 FRÉDÉRIC LATRÉMOLIÈRE ð 6= 0 [8]. We note that they all are infinite dimensional (the other, trivial, unitary representations of H3 are one-dimensional). Let ð ∈ R \ {0}. For all (x, y) ∈ R2 and for all ξ ∈ L2(R), set: σx,y ð,1 ξ = α (xP +yQ) ξ expH3 ð,1 x,y, xy 2 ð,1 = α ξ : s ∈ R 7→ exp(iπðxy + 2iπsx)ξ(s + ðy). The map σx,y ð,1 is a unitary on L2(R) for all (x, y) ∈ R2. Moreover, for all (x1, y1), (x2, y2) ∈ R2, we note that: ð,1 σx2,y2 σx1,y1 ð,1 = eð((x1, y1), (x2, y2))σx1+x2,y1+y2 ð,1 , i.e. σð,1 is a projective representation of R2 on L2(R) for the bicharacter eð, namely the Schrödinger representation of "Plank constant" ð. Moreover, every nontrivial irreducible unitary projective representation of R2 is unitarily equivalent to one of σ1,ð for some ð 6= 0 (by nontrivial, we mean associated with a nontrivial cocycle). We introduce one more notation which will prove very useful in defining our D- norm on Heisenberg modules. If d ∈ N with d > 0, we define the following unitarry operators on L2(R) ⊗ Cd: αx,y,u ð,d = αx,y,u ð,d = σx,y ð,1 ⊗ id and σx,y ð,1 ⊗ id for all x, y, u ∈ R, where id is the identity map on Cd. We trivially check that αð,d is a unitary representation of H3 on L2(R) ⊗ Cd, while σð,d is a eð-projective representation of R2 on L2(R) ⊗ Cd. Moreover, we also check immediately that αx,y,0 ð,d = σx,y ð,d for all x, y ∈ R. We now turn to the projective representations of Z2 note that, for any p ∈ Z, the skew bicharacter e p Z2 q -- which we keep denoting by e p q q q, where q ∈ N\{0}. We first of Z2 induces a skew bicharacter of q is cohomologous . By [10], any multiplier of Z2 to e p q for some p ∈ N. For our purpose, we will thus get, up to unitary equivalence, every possible q for arbitrary finite dimensional unitary projective representations of the groups Z2 q ∈ N \ {0} by considering the following family. Notation 2.13. Let p ∈ Z and q ∈ N \ {0}. Let n ∈ Z 7→ [n] ∈ Zq be the canonical surjection. Let: z z2 . . . zq−1 0 1 . . . and vp,q =  . . . 0 . . . . . . 1 1 0 ,     up,q = 1   p,q = vq p,q = 1, the map: q (cid:17). Since uq q 7→ ρz,w with z = exp(cid:16) 2iπp ρp,q,1 : (z, w) ∈ Z2 is well-defined. An easy computation shows that ρp,q,1 is a projective representation of Z2 q. p,q,1 = exp(cid:18) iπpnm p,q where [n] = z and [m] = w q (cid:19) un p,qvm D-NORMS ON HEISENBERG MODULES 9 For all d ∈ qN, d > 0, we now set: ρn,m p,q,d = ρn,m p,q,1 ⊗ id d q where id d q is the identity map on C d q . We remark that ρp,q,d acts on Cd, i.e. we parametrized ρ by the dimension of the space on which it acts rather than the multiplicity of ρp,q,1, as it will make our notations much simpler. If p and q are relatively prime, the representation ρp,q,1 is irreducible, with range the entire algebra of q × q matrices -- it is in fact, the only irreducible e p - projective representation of Z2 q up to unitary equivalence. Thus in general, any finite dimensional unitary representation of Z2 q is unitarily equivalent to some ρl,r,d for some l ∈ Z, r ∈ N \ {0}, d ∈ rN \ {0}, with l = 0 and r = 1 or l, r relatively prime. q In order to construct the inner product on the Heisenberg modules, we shall need to first work on a space of well-behaved functions inside the Hilbert space ℓ2(Z2) on which quantum tori will act. This space will consist of the Schwarz functions. Definition 2.14. Let E be a finite dimensional vector space. A function f : R → E is a E-valued Schwarz function over R when it is infinitely differentiable on R and, for all j ∈ N and all polynomial p ∈ R[X], we have: lim t→±∞(cid:13)(cid:13)(cid:13) p(t)f (j)(t)(cid:13)(cid:13)(cid:13)E = 0. 1+sj for all s ∈ R. The space of all E-valued Schwarz functions over R is denoted by S(E). We note that if f ∈ S(E) for some finite dimensional space E, then in particular, f ∈ Lp(R) for all p ∈ [1,∞], since for any j ∈ N, there exists M > 0 such that kf (s)kE 6 M We now implement the scheme which we described a few paragraphs above to construct modules over quantum tori. We refer to the mentioned works of Connes and Rieffel for the details and justification behind the following construction. Theorem-Definition 2.15 ( [5], [24], [7] ). Let θ ∈ R and q ∈ N\{0}. Let p ∈ Z, q ∈ N \ {0} , and let d ∈ qN \ {0}. The Heisenberg module H p,q,d is the module over Aθ defined as follows. of d the same cocycle. Up to unitary conjugation, we assume that ρp,q,d acts on Cd. Let ρp,q,d be the projective action of Z2 , consisting of the sum q copies of the unique, up to unitary equivalence, irreducible representation with q with cocycle e p θ q Let: ð = θ − p q . Let αð,1 be the action of the Heisenberg group H3 on L2(R) given by Expression (2.3). and [m], we set: For (n, m) ∈ Z2, denoting the class of n and m in Z(cid:14)qZ , respectively, by [n] For all n, m ∈ Z, the map n,m p,q,ð,d is a unitary of L2(R) ⊗ Cd, and moreover ð,1 ⊗ ρ[n],[m] p,q,d . p,q,ð,d = σn,m n,m p,q,ð,d is an eθ-projective representation of Z2. 10 FRÉDÉRIC LATRÉMOLIÈRE By universality, the Hilbert space L2(R) ⊗ Cd is a module over Aθ, with, in particular, for all f ∈ ℓ1(Z2) and ξ ∈ L2(R, Cd) = L2(R) ⊗ Cd: f ξ = Xn,m∈Z f (n, m)n,m p,q,ð,dξ. Let S p,q,d θ = S(Cd) ⊆ L2(R) ⊗ Cd. For all ξ, ω ∈ S p,q,d θ , define hξ, ωiH p,q,d θ as the function in ℓ1(Z2) given by: hξ, ωiH p,q,d θ : (n, m) ∈ Z2 7−→Dn,m p,q,ð,dξ, ωEL2(R)⊗E is the completion of S p,q,d . θ for the norm associ- The Heisenberg module H p,q,d θ ated with the Aθ-inner product h·,·iH p,q,d θ . We note that S p,q,d θ the action of the subalgebra: is not closed under the action of Aθ but it is closed under {f ∈ ℓ1(Z2) : ∀p ∈ R[X, Y ] lim n,m→±∞ p(n, m)f (n, m) = 0} of (ℓ1(Z2),∗θ,·∗), often referred to as the smooth quantum torus. We will not use this observation later on, though it is notable that the completion of S p,q,d is indeed a Aθ-module. θ 3. A continuous fields of C∗-Hilbert norms θ , as θ varies in R, form a continuous family. All Heisenberg modules are completions of S(Cd) for some d ∈ N, d > 0. For a fixed d, it thus becomes possible to ask whether the various C∗-Hilbert norms k · kH p,q,d To this end, we establish a succession of lemmas whose primary goal is to provide us with estimates on the Heisenberg modules' C∗-Hilbert norms in terms of the norm of ℓ1(Z2). While the Heisenberg modules' C∗-Hilbert norms are in general delicate to work with as they involve the no-less abstract quantum tori norms, the ℓ1(Z2) norm, which dominates all of the quantum tori norms, is much more amenable to computations. For our purpose, we will take full advantage of the regularity of Schwarz functions, which will enable us to apply various analytic tools to derive our desired result. The first step is a lemma which provides a first upper bound to the ℓ1(Z2) norm of the difference between certain Heisenberg module inner products. Lemma 3.1. If θ, ϑ ∈ R and p ∈ Z, q ∈ N \ {0}, d ∈ qN \ {0}, and if ω, η and ξ are C2 functions from R to Cd such that for all f ∈ {ω, η, ξ}: (1) all of f , f ′ and f ′′ are integrable on R, (2) limt→±∞ f (t) = limt→±∞ f ′(t) = limt→±∞ f ′′(t) = 0, then, writing ðθ = θ − p q and ðϑ = ϑ − p q , we have: − hξ, ηiH p,q,d 1 ϑ (cid:13)(cid:13)(cid:13)ℓ1(Z2) θ (cid:13)(cid:13)(cid:13)hω, ηiH p,q,d 6 Xn∈Z 4π2n2 ZR Xm∈Zkω′′(t + ðθm) − ξ′′(t + ðϑm)kCd kη(t)kCd dt+ + 2ZR Xm∈Zkω′(t + ðθm) − ξ′(t + ðϑm)kCd kη′(t)kCd dt D-NORMS ON HEISENBERG MODULES 11 + ZR Xm∈Z kω(t + ðθm) − ξ(t + ðϑm)kCd kη′′(t)kCd dt! . Proof. We begin with the observation that for all (n, m) ∈ Z2 we have: (n, m) p,q,d ω(t + ðθm), η(t)ECd p,q,d ξ(t + ðϑm), η(t)ECd exp(2iπnt) dt exp(2iπnt) dt (ω(t + ðθm) − ξ(t + ðϑm)), η(t)ECd exp(2iπnt) dt. θ ϑ hω, ηiH p,q,d (n, m) − hξ, ηiH p,q,d =ZRDρ[n],[m] −ZRDρ[n],[m] =ZRDρ[n],[m] For all n, m ∈ Z, the function: fn,m : t 7→Dρ[n],[m] p,q,d p,q,d ω(t + ðθm) − ξ(t + ðϑm), η(t)ECd has a first and continuous second derivative which are integrable, and: lim t→±∞ fn,m(t) = lim t→±∞ f ′ n,m(t) = lim t→±∞ f ′′ n,m(t) = 0. We consequently may apply integration by part and obtain, for all m, n ∈ Z: ZRDρ[n],[m] p,q,d ω(t + ðθm) − ξ(t + ðϑm), η(t)ECd exp(2iπnt) dt =ZR =ZR fn,m(t) exp(2iπnt) dt = −ZR f ′′ n,m(t) exp(2iπnt) 4π2n2 dt. f ′ n,m(t) exp(2iπnt) dt 2iπn We compute trivially that for all t ∈ R and m, n ∈ Z: (ω′′(t + ðθm) − ξ′′(t + ðϑm)), η(t)ECd n,m(t) =Dρ[n],[m] p,q,d f ′′ + 2Dρ[n],[m] p,q,d (ω′(t + ðθm) − ξ′(t + ðϑm)), η′(t)ECd +Dρ[n],[m] p,q,d (ω(t + ðθm) − ξ(t + ðϑm)), η′′(t)ECd . Thus using Cauchy-Schwarz and since ρ[n],[m] p,q,d is a unitary, we conclude: (cid:13)(cid:13)(cid:13)hω, ηiH p,q,d θ − hξ, ηiH p,q,d ϑ (cid:13)(cid:13)(cid:13)ℓ1(Z2) 12 FRÉDÉRIC LATRÉMOLIÈRE dt 1 exp(2iπnt) dt(cid:12)(cid:12)(cid:12)(cid:12) p,q,d ZRDρ[n],[m] n,m(t)(cid:12)(cid:12) 4π2n2 (ω(t + ðθm) − ξ(t + ðϑm)), η(t)ECd 4π2n2 (cid:18)ZR kω′′(t + ðθm) − ξ′′(t + ðϑm)kCdkη(t)kCd dt = Xn,m∈Z(cid:12)(cid:12)(cid:12)(cid:12) 6 Xm,n∈ZZR (cid:12)(cid:12)f ′′ 6 Xm,n∈Z + 2ZR kω′(t + ðθm) − ξ′(t + ðϑm)kCdkη′(t)kCd dt +ZR kω(t + ðθm) − ξ(t + ðϑm)kCdkη′′(t)kCd dt(cid:19) = Xn∈N + 2ZR Xm∈Nkω′(t + ðθm) − ξ′(t + ðϑm)kCd! kη′(t)kCd dt +ZR Xm∈Nkω(t + ðθm) − ξ(t + ðϑm)kCd! kη′′(t)kCd dt# by Tonelli's theorem. kω′′(t + ðθm) − ξ′′(t + ðϑm)kCd!kη(t)kCd dt 1 4π2n2 "ZR Xm∈N This concludes our lemma. (cid:3) Our next lemma focuses on the type of estimates given in Lemma (3.1), and gives a sufficient condition for these upper bounds to converge to 0 when various parameters are allowed to converge to appropriate values. Lemma 3.2. Let d ∈ N, d > 0. Let N = N∪{∞} be the one point compactification of N. If (ωk)k∈N and (ηk)k∈N are two families of C2-functions from R to Cd and (ðk)k∈N is a sequence of nonzero real numbers converging to some ð∞ 6= 0 such that: (1) (t, k) ∈ R × N 7→ ωk(t) and (t, k) ∈ R × N 7→ ηk(t) are jointly continuous, (2) there exists M > 0 such that for all k ∈ N and t ∈ R: M 1 + t2 , max {kωk(t)kCd ,kηk(t)kCd} 6 then: (3.1) lim k→∞Xn∈N 1 4π2n2 ZR Xm∈Z kωk(t + ðkm) − ω∞(t + ð∞m)kCdkηk(t)kCd dt = 0. Proof. First, we observe that Expression (3.1) is left unchanged if we replace ðk with −ðk for all k ∈ N, thanks to the summation over m ∈ Z. Consequently, we may assume without loss of generality that ð∞ > 0 and assume that ðk > 0 for all k ∈ N (since (ðk)k∈N converges to ð∞ 6= 0, we must have that ðk and ð∞ have the same sign for k larger than some K ∈ N; we thus can truncate our sequence to start at K and flip all the signs if necessary to work with positive values). With this in mind, since (ðk)k∈N is positive and converges to ð∞ > 0, there exists 0 < ð− < ð+ such that for all k ∈ N, we have ðk ∈ [ð−, ð+]. D-NORMS ON HEISENBERG MODULES 13 We shall employ the Lebesgue dominated convergence theorem. To this end, we introduce the following function to serve as our upper bound. For all t, m ∈ R we set: (3.2) 1+(t+mð−)2 if m > 0 and t > −ð−m, or if m < 0 and t 6 −ð−m, M otherwise. M b(t, m) =( For a fixed t ∈ R, we note that: b(t, m) ∼m→±∞ M −m2 , ð2 soPm∈Z b(t, m) < ∞. Moreover, by construction, for all t, m ∈ R and ð ∈ [ð−, ð+], we have: M 1 + (t + ðm)2 6 b(t, m). Therefore, using our hypothesis, for all t ∈ R, m ∈ Z, k ∈ N and ð ∈ [ð−, ð+]: kωk(t + mð) − ω∞(t + mð∞)kCd 6 1 + (t + mð)2 + 1 + (t + mð∞)2 M M Thus for a fixed t ∈ R, we may apply Lebesgue dominated convergence theorem to conclude: 6 2b(t, m). (3.3) lim k→∞ Xm∈Zkωk(t + mðk) − ω∞(t + mð∞)kCd = 0, since (t, k) ∈ R × N 7→ ωk(t) is jointly continuous. We now make another observation. For any fixed ð > 0 and k ∈ N, The function: t ∈ R 7→ Xm∈Zkωk(t + ðm)kCd is ð-periodic. If t ∈ [0, ð+], k ∈ N and ð ∈ [ð−, ð+], then since: kωk(t + ðm)kCd 6 sup x∈[0,ð+] b(x, m) while, as can easily be checked: sup x∈[0,ð+] b(x, m) ∼m→±∞ M −m2 , ð2 we conclude that the series: (cid:16)(t, k, ð) ∈ R × N × [ð−, ð+] 7→Xkωk(t + ðm)kCd(cid:17)m∈Z converges uniformly to its limit on [0, ð+] × N × [ð−, ð+]. In particular: (t, k, ð) ∈ [0, ð+] × N × [ð−, ð+] 7→ Xm∈Zkωk(t + ðm)kCd is continuous on a compact domain and so it is bounded. Let C > 0 such that for all (t, k, ð) ∈ [0, ð+] × N × [ð−, ð+], we have: Xm∈Zkωk(t + ðm)kCd 6 C. 14 FRÉDÉRIC LATRÉMOLIÈRE (3.5) lim 2CM 1 + t2 . 1+t2 is integrable over R. Once again, we apply Lebesgue dominated We conclude that t 7→ Pm∈Z kωk(t − ðkm)kCd is bounded by C on R, since it is an ðk-periodic function with ðk 6 ð+, for all k ∈ N. We thus have that for all t ∈ R and k ∈ N: (3.4) Xm∈Zkωk(t + mðk) − ω∞(t + mð∞)kCdkηk(t)kCd 6 2Ckηk(t)kCd 6 Now t ∈ R 7→ 2CM convergence theorem, and we conclude from Expression (3.3) that: k→∞ZR Xm∈Z Last, using Inequality (3.4) again, we note that for all k ∈ N: kωk(t + mðk) − ω∞(t + mð∞)kCdkηk(t)kCd dt 6ZR kωk(t + mðk) − ω∞(t + mð∞)kCdkηk(t)kCd dt = 0. ZR Xm∈Z and thus for all n ∈ Z and k ∈ N: 4π2n2 ZR Xm∈Zkω(t + mðk) − ω(t + mð∞)kCdkη(t)kCd dt 6 with Pn∈Z k→∞Xn∈Z 2πn2 < ∞; hence we may apply Lebesgue dominated convergence the- 4π2n2 ZR Xm∈Zkωk(t + mðk) − ω∞(t + mð∞)kCdkηk(t)kCd dt = 0. orem once more to conclude from Expression (3.5): 2CM π 4π2n2 = CM 2πn2 , 2CM 1 + t2 CM 1 1 lim dt = 2CM π This concludes our lemma. (cid:3) Remark 3.3. One may check that Lemma (3.1) and Lemma (3.2) together prove qo, that if p, q ∈ N, ξ, ω ∈ S(Cd), for any d ∈ qN with d > 0, and if θ ∈ R \n p ∈ ℓ1(Z2). It is a well-known fact (indeed a basic fact for the very then hξ, ωiH p,q,d construction of Heisenberg modules) though maybe not apparent from Theorem- Definition (2.15) without consulting such sources as [24]. θ We now bring together Lemma (3.1) and Lemma (3.2) to obtain a first result of continuity on the Heisenberg module inner products, albeit using the ℓ1(Z2) norm. This is the core result of this section, and it is phrased at a somewhat higher level of generality that what is needed for the proof of continuity of the family of Heisenberg C∗-Hilbert norms. Indeed, this level of generality will prove useful twice later in this paper: when proving that the Heisenberg group representations αð,d define strongly continuous actions on Heisenberg modules, and when establishing that our prospective D-norms on Heisenberg modules will also form a continuous family of norms in [20]. Lemma 3.4. Let p, q ∈ N with q > 0 and d ∈ qN with d > 0. If (ξk)k∈N is a family of Cd-valued C2-functions over R such that: (1) there exists M > 0 such that for all k ∈ N and t ∈ R: M 1 + t2 , max{kξk(t)kCd,kξ′ k(t)kCd ,kξ′′ k (t)kCd} 6 (2) (t, k) ∈ R × N 7→ ξk(t) is continuous, (cid:13)(cid:13)(cid:13)(cid:13) 6(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13) (3.7) observe that: θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) +(cid:13)(cid:13)(cid:13)hξk, ξ∞iH p,q,d θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) =(cid:13)(cid:13)(cid:13)(cid:13) θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) hξk, ξkiH p,q,d k (t + ðkm) − ξ′′ θ∞ (cid:13)(cid:13)(cid:13)ℓ1(Z2) θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) . D-NORMS ON HEISENBERG MODULES 15 and if (θk)k∈N is a sequence converging to θ∞ such that θk − p then we have: q 6= 0 for all k ∈ N, lim k→∞(cid:13)(cid:13)(cid:13)(cid:13) hξk, ξkiH p,q,d θk − hξ∞, ξ∞iH p,q,d = 0. θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) Proof. To fix notations, for all k ∈ N, we set ðk = θk − p sequence of nonzero real numbers converging to ð∞ 6= 0. We shall prove our result from the following inequality: q . Note that (ðk)k∈N is a (3.6) hξk, ξkiH p,q,d θk − hξ∞, ξ∞iH p,q,d θk − hξk, ξ∞iH p,q,d hξk, ξkiH p,q,d We begin with the first term of the right hand side of Inequality (3.6). We θ∞ − hξ∞, ξ∞iH p,q,d . θk − hξk, ξ∞iH p,q,d hξk, ξkiH p,q,d By Lemma (3.1), we then have for all k ∈ N: θk − hξ∞, ξkiH p,q,d θk − hξ∞, ξkiH p,q,d (cid:13)(cid:13)(cid:13)(cid:13) hξk, ξkiH p,q,d 6 Xn∈Z 1 4π2n2 ZR Xm∈Zkξ′′ + 2ZR Xm∈Z ∞(t + ð∞m)kCd kξk(t)kCd dt+ ∞(t + ð∞m)kCd kξ′ kξ′ k(t + ðkm) − ξ′ k(t)kCd dt + ZR Xm∈Zkξk(t + ðkm) − ξ∞(t + ð∞m)kCd kξ′′ k (t)kCd dt! . Our assumptions allow us to apply Lemma (3.2) to each term in the right hand side of Inequality (3.7) to conclude that: lim k→∞(cid:13)(cid:13)(cid:13)(cid:13) hξk, ξkiH p,q,d θk − hξk, ξ∞iH p,q,d = 0. θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) We handle the second term of Inequality (3.6) in a similar manner. From Inequality (3.6), our lemma is proven. (cid:3) We now conclude this section with the proof that indeed, Heisenberg C∗-Hilbert norms form continuous families of norms for a fixed projective representation of some Z2 q. Proposition 3.5. Let p, q ∈ N and d ∈ qN with d > 0. Let (ξ)k∈N be a family in S(Cd) such that (k, t) ∈ N × R 7→ ξk(t) is (jointly) continuous and there exists M > 0 such that kξ(s) 1+t2 for all k ∈ N, t ∈ R and s ∈ {0, 1, 2}. k (t)kCd 6 M If (θk)k∈N is a sequence in R converging to θ∞ and such that θk − p q = 0 for all k ∈ N, then: lim k→∞ kξkkH p,q,d θk = kξ∞kH p,q,d θ∞ . θk θk hξk, ξkiH p,q,d hξk, ξkiH p,q,d hξk, ξkiH p,q,d θk − hξ∞, ξ∞iH p,q,d =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθ∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)Aθk (cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d 6(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)Aθk (cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d +(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθ∞(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d θ∞ (cid:13)(cid:13)(cid:13)Aθk −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d 6(cid:13)(cid:13)(cid:13)(cid:13) θ∞ (cid:13)(cid:13)(cid:13)(cid:13)Aθk +(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d θ∞ (cid:13)(cid:13)(cid:13)Aθ∞(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθk −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d 6(cid:13)(cid:13)(cid:13)(cid:13) θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) +(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d θ∞ (cid:13)(cid:13)(cid:13)Aθ∞(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθk −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d θ∞ (cid:13)(cid:13)(cid:13)(cid:13)ℓ1(Z2) θ∞ (cid:13)(cid:13)(cid:13)Aθ∞(cid:12)(cid:12)(cid:12)(cid:12) θ∞ (cid:13)(cid:13)(cid:13)Aθk −(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d θk − hξ∞, ξ∞iH p,q,d hξk, ξkiH p,q,d θk − hξ∞, ξ∞iH p,q,d = 0. = 0. . We now apply Lemma (3.4) to conclude that: lim hξk, ξkiH p,q,d k→∞(cid:13)(cid:13)(cid:13)(cid:13) k→∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13)hξ∞, ξ∞iH p,q,d lim 16 FRÉDÉRIC LATRÉMOLIÈRE Proof. For each k ∈ N ∪ {∞}, we set ðk = θk − p q 6= 0. We first compute: (3.8) (cid:12)(cid:12)(cid:12)(cid:12) kξkk2 H p,q,d θk − kξ∞k2 H p,q,d θ∞ (cid:12)(cid:12)(cid:12)(cid:12) Now, for any f ∈ ℓ1(Z2), the function θ ∈ R 7→ kfkAθ is continuous by [26, Corollary 2.7]. Hence, using Remark (3.3): Thus, we conclude from Inequality (3.8) that: k→∞ kξkk2 lim H p,q,d θk = kξ∞k2 H p,q,d θ∞ which, by continuity of the square root, proves our lemma. Corollary 3.6. Let p, q ∈ N and d ∈ qN with d > 0. Let ξ ∈ S(Cd). If (θk)k∈N is a sequence in R converging to θ∞ and such that θk − p q = 0 for all k ∈ N, then: . = kξkH p,q,d k→∞ kξkH p,q,d lim (cid:3) θ∞ θk Proof. We apply Proposition (3.5) to the family k ∈ N 7→ ξ. We note that since ξ is a Schwarz function, our assumptions are met. (cid:3) 4. The action of the Heisenberg group on Heisenberg modules Our goal in this paper is to prove that Heisenberg modules may be endowed with a metrized quantum vector bundle structure over quantum 2-tori using a D- norm built from a Lie group action and inspired by the construction of [28], albeit involving a projective action of a locally compact group, which will not act via D-NORMS ON HEISENBERG MODULES 17 isometries of the D-norm. These changes will introduce new difficulties which we will handle in the next few sections. As a first step, we study the actions of the Heisenberg group on Heisenberg modules. One motivation for the results in this section is to establish the properties which will meet the hypothesis of the main results in our next section, from which our D-norm will emerge. We also note that the actions αð,d, for all ð ∈ R \ {0} and d ∈ N\{0}, is a strongly continuous action by isometries of L2(R)⊗Cd, but we need these results to be proven for the Heisenberg C∗-Hilbert norms, which dominate the norm of L2(R) ⊗ Cd. in this section, and thus we group them in the following. We shall use the same hypotheses for a series of lemmas and our main definition Hypothesis 4.1. Let p ∈ Z, q ∈ N \ {0}, and let d ∈ qN with d > 0. Let θ ∈ R \n p qo. We write ð = θ − p We shall employ the notations of Theorem-Definition (2.15). q . We begin with two lemmas which will prove that H3 acts via isometries of the norm of the Heisenberg modules on the subspace of Schwarz functions -- where we have an explicit formula for our inner product -- and thus can indeed be extended to the entire module. Lemma 4.2. We assume Hypothesis (4.1). For all (x, y, u) ∈ H3, if z1 = exp (2iπðy) and z2 = exp (−2iπðx), and if ξ, ω ∈ S p,q,d ð,d (ω)EH p,q,d (cid:16)hξ, ωiH p,q,d Dαx,y,u ð,d (ξ), αx,y,u = βz1,z2 (cid:17) . , then: θ θ θ θ Proof. Let n, m ∈ Z. We compute: Dαx,y,u ð,d (ω)EH p,q,d ð,d (ξ), αx,y,u θ (n, m) ð,1 αx,y,u p,q,ð,dαx,y,u ð,d ξ, αx,y,u ð,1 ⊗ ρ[n],[m] ð,d ωEL2(R)⊗Cd p,q,d (cid:17) ξ, αx,y,u ð,1 ⊗ ρ[n],[m] αx,y,u ð,1 ⊗ ρ[n],[m] =Dn,m =D(cid:16)σn,m =D(cid:16)α(x,y,u)−1 =Dexp(2iπð(yn − xm))(cid:16)σn,m p,q,ð,dξ, ωEL2(R)⊗Cd n,m, nm 2 ð,1 = zn 1 zm ð,1 α . ð,d ωEL2(R)⊗Cd p,q,d (cid:17) ξ, ωEL2(R)⊗Cd p,q,d (cid:17) ξ, ωEL2(R)⊗Cd = βz1,z2 θ (cid:16)hξ, ωiH p,q,d θ (cid:17) (cid:3) Therefore, by definition of the dual action β: 2 Dn,m ð,d (ω)EH p,q,d θ ð,d (ξ), αx,y,u Dαx,y,u as desired. To ease our notations in this section, we set: Notation 4.3. For all (x, y) ∈ R2 and ð > 0, we define: υð(x, y) = (exp(2iπðy), exp(−2iπðx)) ∈ T2. 18 FRÉDÉRIC LATRÉMOLIÈRE We now show that the Heisenberg group acts by isometries for the C∗-Hilbert θ θ ð,d ,k · kH p,q,d norm. Lemma 4.4. We assume Hypothesis (4.1). For all (x, y, u) ∈ H3, the map αx,y,u (cid:17). is an isometry of (cid:16)H p,q,d Proof. Let (x, y, u) ∈ H3 and ξ ∈ S p,q,d =(cid:13)(cid:13)(cid:13)(cid:13)Dαx,y,u =(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)hξ, ξiH p,q,d ð,d ξEH p,q,d (cid:13)(cid:13)(cid:13)Aθ hξ, ξiH p,q,d (cid:13)(cid:13)(cid:13)Aθ ð,d ξ, αx,y,u (cid:13)(cid:13)(cid:13)(cid:13)Aθ by Lemma (4.2), . We compute: kαx,y,u βυr(x,y) θ ξk2 = kξk2 H p,q,d H p,q,d . θ ð θ θ θ θ θ This completes our proof. (cid:3) θ θ θ ð,d ,k · kH p,q,d may thus be extended to H p,q,d Notation 4.5. We use the notations of Hypothesis (4.1). The action αð,d of H3 on S p,q,d for all (x, y, u) ∈ H3; we shall keep the notation of this extension as αð,d. We note that it also acts via isometry on (cid:16)H p,q,d We also use the same notation for σð,d extended to (cid:16)H p,q,d by extending by continuity αx,y,u The actions of the Heisenberg group on Heisenberg modules is by morphism modules, in the sense of [16, Definition 3.5]. This result will play a role in the proof that our D-norm satisfies the modular version of the Leibniz inequality. Lemma 4.6. We assume Hypothesis (4.1). For all a ∈ Aθ, ξ ∈ H p,q,d (x, y, u) ∈ H3, then: ,k · kH p,q,d (cid:17). (cid:17). and θ θ θ θ αx,y,u ð,d (aξ) = βυð(x,y) θ (a)αx,y,u ð,d (ξ). Proof. Let n, m ∈ Z and ξ ∈ S p,q,d θ and fm,m ∈ ℓ1(Z2) be defined by: fn,m : (z, w) ∈ Z2 7−→(1 if n = z and m = w, 0 otherwise. We compute: αx,y,u ð,d (fn,mξ) = αx,y,u ð,d n,m p,q,ð,dξ n,m, nm 2 ð,d ð,d α p,q,d (cid:17) ξ =(cid:16)αx,y,u ⊗ ρ[n],[m] = exp(2iπð(yn − xm))(cid:16)α = exp(2iπð(yn − xm))n,m = βυð(x,y) (fn,m)αx,y,u ð,d ξ. θ n,m, nm 2 ð,d p,q,ð,d(cid:17) ξ ð,d ⊗ ρ[n],[m] αx,y,u p,q,ð,dαx,y,u ð,d ξ Since βθ is an action by -morphisms, we conclude that for all a ∈ Aθ: (4.1) ð,d (aξ) = βυð(x,y) αx,y,u θ (a)αx,y,u ð,d (ξ) D-NORMS ON HEISENBERG MODULES 19 as desired. The lemma is concluded by extending Equality (4.1) to H p,q,d continuity. θ by (cid:3) An important corollary of Lemma (4.6) is as follows: Corollary 4.7. We assume Hypothesis (4.1). For all a ∈ Aθ, ξ ∈ H p,q,d (x, y, u) ∈ H3, we observe that: αx,y,u θ 6 kakAθkξkH p,q,d θ . Proof. Let a ∈ Aθ, ξ ∈ H p,q,d θ θ (cid:13)(cid:13)(cid:13) ð,d (aξ)(cid:13)(cid:13)(cid:13)H p,q,d ð,d (aξ)(cid:13)(cid:13)(cid:13)H p,q,d =(cid:13)(cid:13)(cid:13) 6 kβυð(x,y) θ θ αx,y,u (cid:13)(cid:13)(cid:13) and (x, y, u) ∈ H3. We compute: βυð(x,y) θ (a)αx,y,u ð,d ξ(cid:13)(cid:13)(cid:13)H p,q,d θ akAθkξkH p,q,d θ by Lemma (4.6), by Lemma (4.4). This completes our proof. and (cid:3) We have checked that the actions of the Heisenberg group on Heisenberg modules, which the latter were constructed from, act by isometric module morphisms on the entire module. Note that we already observed that Heisenberg modules can be regarded as dense subspaces of L2(R) ⊗ Cd spaces on which the same action of the Heisenberg group is defined, strongly continuous and isometric; however we needed to ensure that these actions are well-behaved with respect to the inner product and norm of the Heisenberg modules. In order to define our D-norms, we shall require one more important analytic property: we want our actions to be strongly continuous for the Heisenberg C∗- Hilbert norms. This is the subject of the next proposition. We actually include in the next proposition a somewhat more general hypothesis and estimate than needed for the strong continuity of our actions, as this stronger statement will play an important role in our study of the continuity properties of our D-norms later on in [20]. Proposition 4.8. Let p ∈ Z, q ∈ N \ {0} and d ∈ qN with d > 0. Let C > 0 and M > 0 some constant. Let 0 < ð− < ð+. There exists K > 0 such that for all ξ ∈ S (Cd) satisfying: (4.2) max{kξ(s)kCd,ksξ(s)kCd,kξ′(s)kCd ,ksξ′(s)kCd , kξ′′(s)kCd,ksξ′′(s)kCd} 6 M 1 + s2 , the following holds for all s ∈ R, ð ∈ [ð−, ð+] and (x, y, u) ∈ R3 with x+y+u 6 C: : n ∈ {0, 1, 2}o 6 K(x + y + u) 1 + s2 . (4.3) maxn(cid:13)(cid:13)(cid:13) αx,y,u ð,d ξ(n)(s) − ξ(n)(s)(cid:13)(cid:13)(cid:13)Cd In particular, for all ð 6= 0 and θ = ð + p q : αx,y,u lim (x,y,u)→0(cid:13)(cid:13)(cid:13) ð,d ξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d θ = 0. 20 FRÉDÉRIC LATRÉMOLIÈRE Proof. Let ξ ∈ S(Cd) and (x, y, u) ∈ R3. We note that for all s ∈ R, using the continuity of ξ, we of course have: αx,y,u ð,d ξ(s) − ξ(s) = exp(2iπ(u + xs))ξ(s + ðy) − ξ(s) (x,y,u)→0 −−−−−−→ 0. However, we wish to apply Lemma (3.4) to obtain convergence in norm, so we seek a more precise estimate. To this end, let: fs(t) = αtx,ty,tu ð,d ξ(s) = exp(2iπ(ðtu + txs))ξ(s + ðty) for all t, s ∈ R. We compute for all t, s ∈ R: f ′ s(t) = exp(2iπ(ðtu + txs)) (2iπ(ðu + xs)ξ(s + ðty) + ðyξ′(s + ðty)) . Let k(x, y, u)k1 = x +y +u for all (x, y, u) ∈ R2, i.e. k·k1 is the usual 1-norm on R3. Let us now assume k(x, y, u)k1 6 C -- in particular, y < C. We observe that for all s ∈ R, using the function b introduced in Expression (3.2) in the proof of Lemma (3.2): αx,y,u (cid:13)(cid:13)(cid:13) 0 f ′ s(t) dt(cid:13)(cid:13)(cid:13)(cid:13)Cd ð,d ξ(s) − ξ(s)(cid:13)(cid:13)(cid:13)Cd Z 1 = kfs(1) − fs(0)kCd =(cid:13)(cid:13)(cid:13)(cid:13) 6Z 1 0 kexp(2iπ(ðtu + txs)) (2iπ(ðu + xs)ξ(s + ðty) + ðyξ′(s + ðty))kCd dt =Z 1 0 k2iπ(ðu + xs)ξ(s + ðty) + ðyξ′(s + ðty)kCd dt  0 k(u, x, y)k1 max 6Z 1   6 2π max{1, ð+}k(x, y, u)k1Z 1 b(s, y)! . 6 2π max{1, ð+}k(x, y, u)k1 sup k2iπðξ(s + ðty)kCd , k2iπsξ(s + ðty)kCd , kðξ′(s + ðty)kCd b(s, ty) dt y∈[−C,C] dt 0 Since: (4.4) lim s→±∞ (1 + s2) sup b(s, y) = M , y∈[−C,C] we conclude that there exists R > 0 such that for all s ∈ R \ [−R, R], we have: kαx,y,u ð,d ξ(s) − ξ(s)kCd 6 M1k(x, y, u)k1 1 + s2 for M1 = 4M π max{1, ð+}. We note that M1 depends only on M , ð+ and C through Expression (4.4), and not on ξ. 1+s2 is continuous and strictly positive, we may adjust M1 to a Since s ∈ R 7→ 1 larger value if necessary such that: min s∈[−R,R] M1 1 + s2 > 2πM max{1, ð+}. D-NORMS ON HEISENBERG MODULES 21 Therefore, we have, for all s ∈ R and (x, y, u) ∈ R3 with k(x, y, u)k1 6 C: kαx,y,u ð,d ξ(s) − ξ(s)kCd 6 M1k(x, y, u)k1 1 + s2 6 M1C 1 + s2 . Now, all the above computations may be applied equally well to ξ′ and ξ′′. We conclude that indeed, Expression (4.3) holds as stated. Let now ξ ∈ S ⊗ Cd be chosen. Since ξ is a Schwarz function, there exists M > 0 such that for all s ∈ R, we have: max {kξ(s)kCd ,ksξ(s)kCd,kξ′(s)kCd ,ksξ′(s)kCd,kξ′′(s)kCd ,ksξ′′(s)kCd} 6 M 1 + s2 . Thus we can apply our previous work to conclude that Expression (4.3) holds for some K > 0, having chosen C = 1 for this last part of our proof. Furthermore, we can apply now Lemma (3.4). For this part, we pick ð > 0; we need not to worry about the uniformity in ð (we may as well assume ð− = ð+ = ð here). Thus, if (xn, yn, un)n∈N converges to 0, Lemma (3.4) implies that:   ð,d ξ − ξkH p,q,d θ 6rkDαxn,yn,un ð,d 0 6 kαxn,yn,un limn→∞rkDαxn,yn,un To prove our result for a general ð 6= 0, we simply observe that for all (x, y, u) ∈ and thus our proposition is completely proven. (cid:3) ξ − ξ, αxn,yn,un kℓ1(Z2) = 0 which concludes the proof of our proposition for ð > 0. ξ − ξEH p,q,d ξ − ξ, αxn,yn,un R3 we have αx,y,u ð,d = αx,−y,−u −ð,d ξ − ξEH p,q,d θ ð,d ð,d ð,d θ kℓ1(Z2) We wish to use the actions of H3 on Heisenberg modules to define our D-norms. The next section presents a general source of possible D-norms from actions of Lie groups satisfying the properties we have established in this section. 5. Seminorms from Lie group actions Connes introduced a quantized differential calculus on quantum tori in [5] using the dual action of the tori, using the Lie group structure of the tori. Moreover, he introduced a noncommutative connection on Heisenberg modules, and these connections proved to be solutions of the Yang-Mills problem for quantum 2-tori [7]. These connections were also useful in Rieffel's work on the classification of modules over quantum tori [24]. Moreover, ergodic actions of metric compact groups on C-algebras were the first example of L-seminorms constructed by Rieffel in [28]. In this section, we begin investigating how to build D-norms from Lie group actions. We will employ as assumptions the properties which we derived for the action of the Heisenberg group on Heisenberg modules. Our construction, as we shall see, lies at the intersection of the purely metric picture of Rieffel and the differential picture of Connes, and is a noncommutative version of [16, Example 3.10]. Our D-norm will be constructed using the following theorem. Definition 5.1. Let α be a strongly continuous action of a Lie group G on a Banach space E . Let w be a nonzero subspace of the Lie algebra of G. An element 22 FRÉDÉRIC LATRÉMOLIÈRE ξ ∈ E is α-differentiable with respect to w when for all X ∈ w, the limit: X(ξ) = lim t→0 αexp(tX)ξ − ξ t exists. In any vector space E, and for any function f : E → R, we denote as usual: lim sup x→0 f (x) = inf δ>0 sup{f (x) : 0 < kxk 6 δ} . Theorem 5.2. Let α be a strongly continuous action by linear isometries of a Lie group G on a Banach space E . Let g be the Lie algebra of G and let h ⊆ g be a nonzero subspace of g. Let S ⊆ E be the subspace of E consisting of α-differentiable elements of E with Let k · k be a norm on h. For all ξ ∈ S , the norm of the linear map: respect to h. We note that S is dense in E . ∇ξ : X ∈ h 7→ ∇X ξ = X(ξ) is denoted by ∇ξ. If ξ ∈ S , then, for any δ > 0: ∇ξ = sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E = sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E X→0 (cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E = lim sup kXk kXk kXk : X ∈ h \ {0}) : X ∈ h \ {0},kXk 6 δ) . Proof. A smoothing argument [4] proves that the set: (cid:26)ξ ∈ E : t > 0 7→ αexp(tX)ξ − ξ t has a limit at 0 for all X ∈ g(cid:27) is dense in E . Therefore, since S contains this set, S is dense in E as well. Fix ξ ∈ S . Let X ∈ h. We define: F : t ∈ R 7→ αexp(tX)ξ. The function F is continuously differentiable, and in particular, F (0) = ξ and F (1) = αexp(X)ξ. Moreover, using the fact that t ∈ R 7→ exp(tX) is a continuous group homomor- phism: F ′(t) = lim s→0 αexp((t+s)X)ξ − αexp(tX)ξ h = lim s→0 αexp(tX)(cid:0)αexp(hX)ξ − ξ(cid:1) h = αexp(tX)∇X ξ. Thus: αexp(X)ξ − ξ =Z 1 0 F ′(t) dt =Z 1 0 αexp(tX) (∇X ξ) dt D-NORMS ON HEISENBERG MODULES 23 so that: (cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E kXk This proves that: αexp(X) (∇X ξ)(cid:13)(cid:13)(cid:13)E dt 0 ∇ξkXk dt = ∇ξ. 1 1 6 = = (cid:13)(cid:13)(cid:13)R 1 0 F ′(t) dt(cid:13)(cid:13)(cid:13)E kXk kXkZ 1 0 (cid:13)(cid:13)(cid:13) kXkZ 1 kXkZ 1 sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E kXk 6 1 : X ∈ h \ {0}) 6 ∇ξ. 0 k∇X ξkE dt since αexp(tX) is an isometry by hypothesis, On the other hand, let us now fix some δ > 0. let us now assume that kXk = 1. We first note that: ∇X ξ = F ′(0) = lim t→0 where lim is used for the topology of (E ,k · kE ), F (t) − F (0) αexp(tX)ξ − ξ t tkXk = lim t→0 = lim t→0 αexp(tX)ξ − ξ . ktXk Thus for all X ∈ h with kXk = 1, since ktXk 6 δ for all t ∈ R with t < δ: and thus: We have thus concluded our argument, as the function: : Y ∈ h \ {0},kY k 6 δ) : X ∈ h \ {0},kXk 6 δ) : X ∈ h \ {0}) . kY k k∇X ξk 6 sup((cid:13)(cid:13)αexp(Y )ξ − ξ(cid:13)(cid:13)E ∇ξ 6 sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E 6 sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E δ ∈ (0,∞) 7→ sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E kXk kXk kXk : X ∈ h \ {0},kXk 6 δ) (cid:3) has been shown to be constant. We note that the seminorms constructed in Theorem (5.2) include Rieffel's L- seminorms in [28] from actions of compact Lie groups. Corollary 5.3. Let α be a strongly continuous action by linear isometries of a compact connected Lie group G on a Banach space E . As a compact Lie group, G admits an Ad-invariant inner product h·,·ig on g. Let k · k be the norm associated with h·,·ig. For any g ∈ G, since G is connected and compact, we may define ℓ(g) as the distance from 1G to g for the Riemannian metric induced by h·,·ig. 24 FRÉDÉRIC LATRÉMOLIÈRE If ξ ∈ S then: sup(cid:26)kαgξ − ξkE ℓ(g) : g ∈ G \ {1G}(cid:27) = ∇ξ. Proof. As G is a compact group, it admits a right Haar probability measure µ. Let h·,·i be any inner product on g. If we set, for all X, Y ∈ g: hX, Y iG =ZG hAdgX, AdgY i dµ(g) then one easily verifies that h·,·iG is an Ad-invariant inner product on g. Now, we endow G with the Riemannian metric induced by left translation of the inner product h·,·iG. As this metric is induced by an Ad-invariant inner product, it is in fact right invariant as well. In particular, G, as a connected compact Riemannian manifold, is geodesically complete by Hopf-Rinow theorem. As a first application, we let ℓ(g) be the dis- tance from 1G to g in G for this Riemannian metric, for all g ∈ G. As a second application, we note that the Riemannian exponential map of G for our metric is indeed surjective. It is now possible to check that the exponential map for the Lie group G and the exponential map for the Riemannian metric coincide. This is done by checking that the Riemannian exponential map defines a 1-parameter subgroup of G. With this in mind, we conclude that for all X ∈ g, we have: ℓ(exp(X)) = inf {kY k : exp(X) = exp(Y )} . We note that the Lie exponential map is certainly not injective, at least as long as G is of dimension at least one, though this does not affect our conclusion. Moreover, since G is a compact connected Lie group, exp is surjective since the Riemannian exponential is surjective. Thus, our corollary is proven using Theorem (5.2). (cid:3) Now, Rieffel proved in [28] that the obvious necessary condition for a seminorm of the type given in Corollary (5.3) to be a L-seminorm is, remarkably, sufficient as well. This fact is highly non-trivial as well, and we record it here as it will be the source of quantum metrics we put on quantum tori. Theorem 5.4 ([28, Theorem 1.9]). Let β be a strongly continuous group action by -automorphisms of a compact group G on a unital C-algebra A. Let ℓ be a continuous length function on G. For all a ∈ A, we define: L(a) = sup(cid:26)kβg(a) − akA ℓ(g) : g ∈ G \ {e}(cid:27) , allowing for this quantity to be infinite. Then the following are equivalent: (1) (A, L) is a quantum compact metric space (which is necessarily Leibniz), (2) {a ∈ A : ∀g ∈ G βg(a) = a} = C1A. We note that the proof of Theorem (5.4) involves explicitly the fact that the spectral subspaces of the action β are finite dimensional under the condition of ergodicity [9]. This result is not trivial, and worse yet for our purpose, does not carry to locally compact group. In fact, besides the trivial representation, no irreducible representation of the Heisenberg group is finite dimensional -- so we are as far as we can to apply the idea in [28]. In this paper, we shall focus on the Heisenberg D-NORMS ON HEISENBERG MODULES 25 modules, and we will prove in this case that the seminorms constructed in Theorem (5.2) have compact unit balls using quite different techniques from Rieffel. The rest of this section introduces the general scheme to construct D-norms from Lie group actions which we will employ in this paper, and prove that this construction meets all our requirements except, maybe, for the compactness of the unit ball which, in the case of Heisenberg modules, will be the subject of our next section. Proposition 5.5. Let β be the action of a compact connected Lie group G on a unital C-algebra A via -automorphisms. Let α be the action by isometric C-linear isomorphisms of a Lie group H on a Hilbert module (M ,h·,·iM ) over A. We write g and h the respective Lie algebras of G and H, and expG : g → G and expH : h → H be the respective Lie exponential maps of G and H. Let w be a nonzero subspace of h. Let k · k♭ be a norm on g and k · k♯ be a norm We set for all a ∈ A: on w ⊆ h. and for all ξ ∈ E : : X ∈ g \ {0}) , kXk♭ L(a) = sup((cid:13)(cid:13)βexp(X)a − a(cid:13)(cid:13)A D(ξ) = sup(kξkM ,(cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)M : X ∈ w \ {0}) . If there exist two linear maps j : w → g and q : g → w such that: (1) for all ξ, ω ∈ M and X ∈ w: βexpG(X)hξ, ωiM =DαexpH (j(X))ξ, αexpH (j(X))ωEE kXk♯ (5.1) and: (5.2) αexpH (X)(aξ) = βexpG(q(X))(a)αexpH (X)ξ, (2) j is an isometry from (g,k · k♭) to (w,k · k♯), (3) q is a surjection of norm at most 1, i.e. kq(X)k♭ 6 kXk♯ for all X ∈ w, then: (1) L is a seminorm on a dense subspace of (A,k · kA), and moreover: L(a) = 0 ⇐⇒ ∀g ∈ G βg(a) = a, (2) D is a norm on a dense subspace of (M ,h·,·iM ) and D(·) > k · kM , (3) L and D are lower semicontinuous, (4) for all a ∈ A and ξ ∈ M : D(aξ) 6 kakAD(ξ) + L(a)kξkM , (5) for all ξ, ω ∈ M : L (hξ, ωiM ) 6 kξkM D(ω) + D(ξ)kωkM . Proof. Let Sg(A) be the subspace of A consisting of all the β-differentiable elements with respect to g, and Sh(M ) be the subspace of M consisting of all the α- differentiable elements of M with respect to w. 26 FRÉDÉRIC LATRÉMOLIÈRE For any a ∈ Sg(A), we define the linear map ∂a : X ∈ g 7→ X(a) whose norm is denoted by ∂ag A, where g is endowed with k · k♭. Since g is finite dimensional, ∂a is continuous and thus has finite norm for all a ∈ Sg(A). For any ξ ∈ Sw(M ), we also define ∇ξ : X ∈ w 7→ X(ξ) whose norm is ∇ξw where w is endowed by k · k♯ -- since w is finite dimensional, the norm of ∇ξ is finite as well. M By Theorem (5.2), for all a ∈ Sg(A) and for all ξ ∈ Sw(M ), then: L(a) = ∂ag A < ∞ and D(ξ) = ∇ξw M < ∞. Since Sg(A) and Sw(E ) are dense, we conclude that the domains of L and D are indeed dense. Since D(·) > k · kM by construction, D is in particular a norm on its domain. Moreover if L(a) = 0 for some a ∈ A, we immediately conclude that βga = a for all g ∈ G since the exponential map of G is surjective. The function ξ ∈ M 7→ αexp(X)ξ−ξ is continuous for all X ∈ w \ {0} and thus D is lower semi-continuous as the pointwise supremum of continuous functions. The same reasoning and conclusion applies to L. kXk♯ We are left to prove the two forms of the Leibniz inequalities, which can be easily checked by direct computation. Let ξ, ω ∈ M . We compute: L (hξ, ωiE ) = sup((cid:13)(cid:13)βexp(X)hξ, ωiE − hξ, ωiE(cid:13)(cid:13)A kXk♭ : X ∈ g \ {0}) : X ∈ w \ {0}) : X ∈ w \ {0}) kXk♯ kXk♯ kj(X)k♯ : X ∈ g \ {0}) = sup((cid:13)(cid:13)(cid:10)αexp(j(X))ξ, αexp(j(X))ω(cid:11)E − hξ, ωiE(cid:13)(cid:13)A 6 sup((cid:13)(cid:13)(cid:10)αexp(X)ξ, αexp(X)ω(cid:11)E − hξ, ωiE(cid:13)(cid:13)A 6 sup((cid:13)(cid:13)(cid:10)αexp(X)ξ, αexp(X)ω(cid:11)E −(cid:10)αexp(X)ξ, ω(cid:11)E(cid:13)(cid:13)A + sup((cid:13)(cid:13)(cid:10)αexp(X)ξ, ω(cid:11)E − hξ, ωiE(cid:13)(cid:13)A : X ∈ w \ {0}) : X ∈ w \ {0}) 6 sup((cid:13)(cid:13)αexp(X)ξ(cid:13)(cid:13)M (cid:13)(cid:13)αexp(X)ω − ω(cid:13)(cid:13)E + sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E 6 kξkM sup((cid:13)(cid:13)αexp(X)ω − ω(cid:13)(cid:13)E + sup((cid:13)(cid:13)αexp(X)ξ − ξ(cid:13)(cid:13)E : X ∈ w \ {0})kωkM : X ∈ w \ {0}) : X ∈ w \ {0})kωkM kXk♯ kXk♯ kXk♯ kXk♯ kXk♯ = kξkM D(ω) + D(ξ)kωkM . D-NORMS ON HEISENBERG MODULES 27 Now, let a ∈ A and ξ ∈ M . We compute: : X ∈ w \ ker q) kXk♯ kXk♯ : X ∈ w \ {0}) : X ∈ w \ {0}) sup ((cid:13)(cid:13)αexp(X) (aξ) − aξ(cid:13)(cid:13)M = sup((cid:13)(cid:13)βexp(q(X))(a)αexp(X) (ξ) − aξ(cid:13)(cid:13)M 6 sup((cid:13)(cid:13)βexp(q(X))(a)αexp(X) (ξ) − aαexp(X)ξ(cid:13)(cid:13)M + sup((cid:13)(cid:13)aαexp(X) (ξ) − aξ(cid:13)(cid:13)M : X ∈ w \ {0}) : X ∈ g \ {0})kξkM + kakAD(ξ) 6 sup((cid:13)(cid:13)βexp(q(X))(a) − a(cid:13)(cid:13)M kq(X)k♭ kXk♯ kXk♭ = L(a)kξkM + kakAD(ξ), as desired. (cid:3) Thus, Proposition (5.5) shows that if we follow the scheme suggested by Theorem (5.2), then we obtain potential D-norms on modules. The missing property is the compactness of the closed unit ball for the D-norm candidate. We conclude our section by connecting our metric framework with the noncom- mutative differential framework of connections on modules. Let us use the notations of Proposition (5.5). A direct computation shows that for all X ∈ w, the following holds: (5.3) ∇X (aξ) = q(X)a · ξ + a∇X ξ while for all X ∈ g, we also have: (5.4) X(hξ, ωiM ) = hj(X)ξ, ωiM + hξ, j(X)ωiM . We also denote A ⊗ g∗ by Ω1 and the space of β-differentiable elements of A by A1. We define ∂ : A1 → Ω1 by setting, for all a ∈ A1: ∂a : X ∈ g 7→ X(a). We observe trivially that Ω1 is an A-A-bimodule and that ∂ is a derivation, i.e. ∂(ab) = a∂(b) + ∂(a)b for all a, b ∈ A1. We first note that to get an interesting connection, we want q to be injective, i.e. g and w to be isomorphic. It is always possible to increase the dimension of g (the Lie algebra structure is actually not involved in the computations to follow, so this is always possible), but this would amount to define ∂X = 0 for all vector X not in g, and this is rather awkward and artificial. Since, for the differential picture, the norms k · k♭ and k · k♯ do not play a role in the construction of the connection, we will for now identify g and w and j and q with the identity map. With this assumption, Expressions (5.3) translates to the operator ∇ : M → M ⊗ g∗, defined by: ∇(ξ) : X ∈ g 7→ ∇X ξ 28 FRÉDÉRIC LATRÉMOLIÈRE for all α-differentiable ξ ∈ M with respect to g, to be a noncommutative connection. We indeed easily check that for all a ∈ A and ξ ∈ M : ∇(aξ) = a∇(ξ) + ∂(a)ξ. Expression (5.4) means that the connection ∇ is hermitian, i.e. it is compatible with the noncommutative equivalent of a metric on the quantum vector bundle M . It is tempting to call ∇ a Levi-Civita connection, although we do not address here the computation of the torsion of ∇. Nonetheless, we see that our structure provides a noncommutative Riemannian geometry. This is the structure which inspired our definition of metrized quantum vector bundle, and we now can see how it is implemented through our main example. In summary, we have constructed a natural D-norm candidate on modules carry- ing certain Lie group actions. The key difficulty, of course, regards the compactness of the unit ball of such a D-norm. 6. A D-norm from a connection on Heisenberg modules We now define our D-norms on Heisenberg modules. Our method employs the idea of Theorem (5.2) and Proposition (5.5), where the actions of the Heisenberg group on Heisenberg modules defines a norm which restricts to the operator norm of a connection constructed via the associated action of the Heisenberg Lie algebra. As noted at the end of the previous section, we want to only work with a subspace of the Heisenberg Lie algebra to build our D-norm and its associated connection, since the central element of the Heisenberg Lie algebra does not act, so to speak, as a derivation -- it simply acts by multiplication by a scalar. We follow a pattern which is common in the literature on the Heisenberg group: we only consider the action of the subspace span{P, Q} in the Lie algebra H. We thus endow span{P, Q} with a norm. If we were to construct a metric on the Heisenberg group using this data -- by defining the length of a curve whose tangent vector at (almost) every point lies in span{P, Q} in the usual manner by integrating the norm of the tangent vector along the curve, and then defining the distance between two points as the infimum of the length of all so-called horizontal curves -- we would actually obtain a sub-Finslerian metric (if our choice of norm comes from a Hilbert space structure, we would have a sub-Riemannian structure and our construction would give rise to a Carnot-Carathédory distance on the Heisenberg group). However, as discussed, we do not transport the Carnot-Carathédory metric from the Heisenberg group via its action in this paper. We prefer to carry the norm of the subspace span{P, Q} of the Heisenberg Lie algebra to our modules. This approach means that we work with a connection, and seems more natural. In essence, the Carnot-Caratheodory is the metric obtained on the group while our D-norms are the quantum metrics obtained on our modules; as the acting group is not compact, we have no reason to expect them to agree. With this in mind, we now introduce: D-NORMS ON HEISENBERG MODULES 29 Definition 6.1. Let p ∈ Z, q ∈ N \ {0} and d ∈ qN with d > 0. Let θ ∈ R \n p qo. Let k·k be a norm on R2. We endow the Heisenberg module H p,q,d : (x, y) ∈ R2 \ {0}  (ξ) = sup  ξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d 2πðk(x, y)k kξkH p,q,d with the norm: ,(cid:13)(cid:13)(cid:13) expH3 ð,d Dp,q,d (xP +yQ) α θ θ θ θ where ð = θ − p q . We now lighten our notation for the rest of our paper. θ simply by Dp,q,d Convention 6.2. We endow R2 with a fixed norm k·k for the rest of this paper. We shall denote Dk·k,p,q,d , as the norm on R2 will not be understood. We emphasize that k · k is independent of any of the parameters p, q, d and θ. The norm k · k on R2 provides us with a continuous length function on Aθ for all θ ∈ R. This length function arises from the invariant Finslerian metric induced by k · k. A direct computation simply shows that: θ ℓ(exp(ix), exp(iy)) = inf{k(x + 2nπ, y + 2mπ)k : n, m ∈ Z2}. For all θ ∈ R, we denote by Lθ the L-seminorm on Aθ associated with the action βθ on Aθ and the length function ℓ via [28, Theorem 1.9]. We note that since T2 is compact and Abelian, Corollary (5.3) implies that for all a ∈ Aθ: : (x, y) ∈ R2 \ {0}) Lθ(a) = sup(kβ ξ − ξkAθ expT2 (x,y) θ k(x, y)k and Lθ agrees with the operator norm of derivative for the natural differential calculus defined by βθ on βθ-differentiable elements. We refer to the previous section for a discussion of these matters. We begin by listing various equivalent expressions for our D-norm candidates, as we shall use whichever may prove useful in this paper. Remark 6.3. We recall from Notation (2.12) that: expH3 (xP + yQ) =(cid:18)x, y, 1 2 xy(cid:19) for all x, y ∈ R. identities hold: For all p, q ∈ N, d ∈ qN with d > 0, θ ∈ R\{pq−1} and ξ ∈ H p,q,d θ , the following θ θ α 2 xy x,y, 1 ð,d Dp,q,d kξkH p,q,d 2πðk(x, y)k : (x, y) ∈ R2 \ {0}  : (x, y) ∈ R2 \ {0}  Proposition 6.4. Let p, q ∈ N and d ∈ qN with d > 0. Let θ ∈ R \n p qo. ξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d ð,d ξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d (ξ) = sup  = sup  ,(cid:13)(cid:13)(cid:13) ,(cid:13)(cid:13)(cid:13) 2πðk(x, y)k kξkH p,q,d σx,y . θ θ θ 30 FRÉDÉRIC LATRÉMOLIÈRE We endow span{P, Q} with the norm 2πðk·k. We also define, for all (x, y) ∈ R2 and ξ ∈ S p,q,d θ : ∇ð x,yξ = lim t→0 = lim t→0 (t(xP +yQ)) expH3 ð,d α tx,ty, 1 ð,d α t 2 t2xy ξ − ξ t ξ − ξ . To ease notation, let ·2πð denote the operator norm for linear maps from ,k · kH p,q,d θ ). (R2, 2πðk · k) to (H p,q,d θ We record: (1) Dp,q,d (2) For all ξ ∈ S p,q,d θ θ is a norm on a dense subspace of H p,q,d , θ and for all δ > 0, the following expressions hold: Dp,q,d θ θ ,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)∇ðξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2πðo (ξ) = maxnkξkH p,q,d = sup(kσx,y (cid:13)(cid:13)(cid:13) = lim sup (x,y)→0 ð,dξ − ξkH p,q,d 2πðk(x, y)k σx,y ð,d ξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d 2πðk(x, y)k . θ θ : (x, y) ∈ R2, 0 < k(x, y)k < δ) (3) If a ∈ Aθ and ξ ∈ H p,q,d θ then: Dp,q,d θ (aξ) 6 kakAθ then: Dp,q,d θ (ξ) + Lθ(a)kξkH p,q,d θ . θ (4) If ξ, ω ∈ H p,q,d Lθ(cid:16)hξ, ωiH p,q,d (ξ)kωkH p,q,d Proof. The Lie algebra of T2 is R2 with the exponential map given as: (cid:17) 6 kξkH p,q,d (ω) + Dp,q,d Dp,q,d θ θ θ θ θ . expT2 : (x, y) ∈ R2 7→ (exp(ix), exp(iy)). Now, the map υð : (x, y) ∈ R2 7→ (2iπðy,−2iπðx) satisfies, according to Lemma (4.2), the relation: β expT2 (υð(x,y)) θ hξ, ωiH p,q,d θ expH3 ð,d =Dσ (x,y) ξ, σ expH3 ð,d (x,y,0) ωEH p,q,d θ . and, according to Lemma (4.6), the relation: (x,y) expH3 ð,d σ (aξ) = β expT2 (υð(x,y)) θ (a)σ expH3 ð,d (x,y) (ξ). In order to apply Proposition (5.5), since υð is indeed a linear isomorphism, we endow span{P, Q} with the norm: kxP + yQk∗ = 2πðk(x, y)k. We now are in the setting of Proposition (5.5), which allows us to conclude all but Assertion (2) in our proposition. Assertion (2), in turn, follows from Theorem (5.2), with our choice of norm. (cid:3) D-NORMS ON HEISENBERG MODULES 31 We now turn to the remaining, main issue of the compactness of the closed unit balls for our D-norm candidates. The strategy we employ relies on a particular source of finite rank operators naturally associated with the Schödinger represen- tations of R2 via the Weyl calculus. Our first step is to introduce the convolution-like operators at the core of our analysis. Lemma 6.5. Assume Hypothesis (4.1). If f ∈ L1(R2) and: σf ð,d =ZZR2 f (x, y)α x,y, xy 2 ð,d dxdy then σf ð,d is a well-defined operator on H p,q,d θ Proof. Let ξ ∈ H p,q,d of H p,q,d θ θ for all (x, y, u) ∈ H3, we simply compute: and (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) σf ð,d(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)H p,q,d θ 6 kfkL1(R2). . Using Lemma (4.4), i.e. the fact that αx,y,u is an isometry ð,d f (x, y)α x,y, xy 2 ð,d ZZR2(cid:13)(cid:13)(cid:13) (ξ)(cid:13)(cid:13)(cid:13)H p,q,d θ dxdy =ZZR2 f (x, y)(cid:13)(cid:13)(cid:13) x,y, xy 2 ð,d α (ξ)(cid:13)(cid:13)(cid:13)H p,q,d dxdy θ =ZZR2 f (x, y)kξkH p,q,d = kfkL1(R2)kξkH p,q,d . θ θ dxdy Thus σf ð,d is well-defined, and moreover: (cid:13)(cid:13)(cid:13) σf ð,d(ξ)(cid:13)(cid:13)(cid:13)H p,q,d θ =(cid:13)(cid:13)(cid:13)(cid:13) ZZR2 f (x, y)σx,y ð,d(ξ) dxdy(cid:13)(cid:13)(cid:13)(cid:13)H p,q,d θ 6 kfkL1(R2)kξkH p,q,d θ . (cid:3) This completes our proof. We now prove the first of two core lemmas of this section, which provides us with a mean to approximate elements in Heisenberg modules using our convolution-type operators, in a manner which is uniform in our prospective D-norms. This lemma is an adjustment of [30] to our context. Lemma 6.6. Assume Hypothesis (4.1). Let ε > 0. If f : R2 → [0,∞) is measurable and satisfies: then for all ξ ∈ H p,q,d 2πð , (1) RR2 f = 1, (2) RRR2 f (x, y)k(x, y)k dxdy 6 ε ð,dξ(cid:13)(cid:13)(cid:13)H p,q,d ξ − σf (cid:13)(cid:13)(cid:13) : θ θ 6 εDp,q,d θ (ξ). 32 FRÉDÉRIC LATRÉMOLIÈRE Proof. If ξ ∈ H p,q,d θ , then: (cid:13)(cid:13)(cid:13) ξ − σf ð,dξ(cid:13)(cid:13)(cid:13)H p,q,d θ f (x, y)α x,y, xy 2 ð,d ξ dxdy(cid:13)(cid:13)(cid:13)(cid:13)H p,q,d θ dxdy θ ξkH p,q,d kξ − α x,y, xy 2 ð,d ξkH p,q,d θ 2πðk(x, y)k dxdy θ(ξ) dxdy =(cid:13)(cid:13)(cid:13)(cid:13) ZZR2 6ZZR2 6ZZR2 6ZZR2 = Dp,q,d θ = εDρ θ(ξ), f (x, y)ξ dxdy −ZZR2 f (x, y)kξ − α x,y, xy 2 ð,d f (x, y)2πðk(x, y)k f (x, y)2πðk(x, y)kDρ (ξ)(cid:18)2πð 2πð(cid:19) ε as desired. (cid:3) We now ensure that we indeed have an ample source of functions which meet the hypothesis of Lemma (6.6). Notation 6.7. If (E, d) is a metric space then the closed ball {x ∈ E : d(x0, x) 6 r} of center x0 ∈ E and radius r > 0 is denoted by E[x0, r]. The following lemma is valid for any norm on R2; we shall work within our context with the fixed norm k · k. Lemma 6.8. For all n ∈ N, let ψn : R2 → [0,∞) be an integrable function supported on R2h0, If f : R2 → [0,∞) is integrable on some ball centered at 0 in (R2,k · k), and f 1 continuous at 0, then: n+1i and with RR2 ψn = 1. n→∞ZZR2 lim ψn(x, y)f (x, y) dxdy = f (0). Proof. Let δ > 0 such that f is integrable on R2[0, δ]. Let ε > 0. Since f is continuous at 0, there exists δc > 0 such that f (x)−f (0) 6 ε for all x ∈ R2[0, δc]. Let N ∈ N be chosen so that 6 min{δ, δc}. For all n > N , we first note that since ψn is supported on a subset of R2[0, δ], the function ψnf is integrable on R2. Moreover for all n > N : N +1 1 ZZR2 (cid:12)(cid:12)(cid:12)(cid:12) ψn(x, y)f (x, y) dxdy − f (0)(cid:12)(cid:12)(cid:12)(cid:12) 6ZR2 ψn(x, y)(f (x, y) − f (0)) dxdy =ZZR2[0,n−1] ψn(x, y)f (x, y) − f (0) dxdy 6ZZR2[0,n−1] ψn(x, y)ε dxdy 6 ε. Thus we have shown that limn→∞RR2 ψn(x, y)f (x, y) dxdy = f (0). (cid:3) D-NORMS ON HEISENBERG MODULES 33 We are now ready to prove the second core lemma of this section. We begin with an explanation of the ideas and reasons behind this lemma. operator which maps bounded subsets of E to totally bounded subsets of E. By a compact operator on a Banach space (E,k · kCd), we mean as usual an The map f ∈ L1(R2) 7→ σf ð,d is a -representation of the twisted convolution algebra L1(R2) for the convolution product defined for all f, g ∈ L1(R2) and x ∈ R2 by: and the involution: f (y)g(x − y)eð(y, x − y) dy f∗ðg(x) =ZR2 f ∈ L1(R2) 7→ f ∗ = x ∈ R2 7→ f (−x), σf operators. The fact that σf as can be directly checked, or is established in [8]. It is an important, well-known fact [8, Theorem 1.30] that this representation is valued in the algebra of compact operators on L2(R) ⊗ Cd, and is faithful; the completion of (L1(R2),∗ð,∗) for the is the entire algebra of compact ð,d is compact as an operator of L2(R)⊗ Cd does not immediately (cid:17) since in general, ð,1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)L2(R) norm f ∈ ℓ1(Z2) 7→ kfkC ∗(R2,eð) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) imply that it is compact for the Banach space(cid:16)H p,q,d . We thus must prove compactness of these we only know that k·kL2(R) 6 k·kH p,q,d operators for our C∗-Hilbert norm. However, we can extract the essential tools for our work from the expansive work on Laguerre expansion of functions and the study of the Moyal plane. We will prove that, at least when f is a radial function, then we can approximate σf ð,d by finite rank operators, in norm. To this end, we need a supply of finite rank operators, which provide a mean to approximate any σf ð,d for f radial. The theory of the quantum harmonic oscillator provides us with a well- suited family of finite rank projections, obtained as σψ ð,d for ψ a properly scaled Laguerre function [8, Ch. 1, sec. 9]. ,k · kH p,q,d θ θ θ To obtain the desired approximation result, however, we need to approximate our radial functions in the norm of L1(R2) using functions obtained from Laguerre functions. As Laguerre functions form an orthonormal basis for some L2 space, we certainly do have a Laguerre expansion which converges in some L2 norm, but convergence in L1(R2) is highly not trivial. The work of Sundaram Thangaveru in [31] comes to our rescue, however, by proving that we may obtain the desired convergence if we replace the Laguerre expansion series by the sequence of its Césaro averages. We now formalize our discussion in the next key lemma. Lemma 6.9. If f : R+ → R is a function such that r ∈ R 7→ rf (r) is Lebesgue integrable, and if we set: f ◦ : (x, y) ∈ R2 7→ f(cid:16)px2 + y2(cid:17) , then the operator σf ◦ ð,d is a compact operator for the Banach space(cid:16)H p,q,d ,k · kH p,q,d Proof. Our goal is to write σf ◦ ð,d as a limit, in the operator norm, of finite rank operators. To this end, let us first assume that ð > 0 and for all n ∈ N, we let ψn θ ð θ (cid:17). Note that these functions are given in [31, (6.1.17)] for ð = 1 which will be important for us in later proofs is that ψn ð = ðψn obtain all the Laguerre functions we are considering via a simple rescaling. 1 (√ð·), i.e. we can π . An observation By slight abuse of notation, we denote by Lp(R+, rdr) the p-Lebesgue space for the measure defined, for all measurable f : [0,∞] → [0,∞), by R ∞ 0 f (r) rdr. In particular, note that the inner product of L2(R+, rdr) is given for any two f, g ∈ L2(R, rdr), by: hf, giL2(R+,rdr) =Z ∞ 0 f (r)g(r) rdr. 34 FRÉDÉRIC LATRÉMOLIÈRE ψn be the nth Laguerre function defined for all r ∈ [0,∞) by: 2 (cid:19) Ln(cid:0)πðr2(cid:1) , (−1)j j! (cid:18) n n − j(cid:19)xj. where Ln is the nth Laguerre polynomials, given for all x ∈ R by: ð (r) = ð exp(cid:18)− Ln(x) = πðr2 n Xj=1 With all these notations set, we define, for each n ∈ N\{0}, the nth Césaro sum of the series given by the Laguerre expansion of f : Cn ð(f ) = n Xj=0 n + 1 − j n + 1 Df ψj ð, ψj ðEL2(R+,rdr) ψj ð. Then by the work of S. Thangavelu in [31, Theorem 6.2.1] -- where our ψj ð is a j in [31, Chapter 6] and we use the rescaled version of the function denoted by ψ0 Césaro sums for "δ = 1" in his notations -- we conclude: n→∞kCn lim ðf − fkL1(R+,rdr) = 0. Now, a quick computation shows that for all n ∈ N \ {0}: and therefore: (cid:13)(cid:13)(cid:13) (Cj ð(f ))◦ − f ◦(cid:13)(cid:13)(cid:13)L1(R2) =(cid:13)(cid:13)(cid:13) Cj ð(f ) − f(cid:13)(cid:13)(cid:13)L1(R+,rdr) , n→∞k(Cn lim ð (f ))◦ − f ◦kL1(R2) = 0 where of course, L1(R2) stands for the 1-Lebesgue space with respect to the usual Lebesgue measure on R2. By Lemma (6.5), writing κn = (Cn ð(f ))◦ for all n ∈ N, we then conclude: lim n→∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ð,d − σf ◦ σκn ð,d(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)H p,q,d θ = 0. By construction, σκn ð,d is finite rank. Indeed, the operator σκn bination of the operators σ turn, projections on CHj ð)◦ (ψj ð,d with j ∈ {0, . . . , n}. The operators σ ð ⊗ Cd ⊆ L2(R) ⊗ Cd, where Hn ð,d is a linear com- are, in ð is the Hermite function: (ψj ð,d ð)◦ Hj ð : t ∈ R 7→ exp − t2√2πð 2 ! Hj(cid:16)t√2πð(cid:17) 1 4 (2ð) pj!2j D-NORMS ON HEISENBERG MODULES 35 where Hj is the jth Hermite polynomial, given for instance by: Hj : t ∈ R 7→ (−1)j exp(t2) dj dtj exp(−t2). ð)◦ (ψj ð,1 Thus the image of the unit ball H p,q,d are projections on CHj ⊆ L2(R) for all Indeed, by [8, p. 65], the operators σ j ∈ N. We note that reassuringly, we will not need the explicit form of the Hermite polynomials or the Laguerre polynomials in our work. (cid:17) by σκn [0, 1] of (cid:16)H p,q,d (cid:17) for all n ∈ N, as a bounded subset of totally bounded in (cid:16)H p,q,d a finite dimensional space (as all norms are equivalent in finite dimension, this observation does not depend on k · kH p,q,d ,k · kH p,q,d ,k · kH p,q,d ð,d is compact as the norm limit of compact operators. Thus σf ◦ We are left to treat the case when ð < 0. We note that for all (x, y, u) ∈ H3, we ð,d is ). θ θ θ θ θ θ have: αx,y,u ð,d = αx,−y,−u −ð,d . We thus proceed as above with −ð in place of ð, and note that σκn κn is a radial function. The rest of the proof is left unchanged. ð,d = −σκn −ð,d since (cid:3) With Lemma (6.9) and Lemma (6.6), we are now able to prove the desired property for our D-norms: Lemma 6.10. We assume Hypothesis (4.1). The set: θ D1(cid:16)Dp,q,d ,k·kH p,q,d (cid:17) =nξ ∈ H p,q,d (cid:17). θ θ : Dp,q,d θ (ξ) 6 1o θ is compact in (cid:16)H p,q,d Proof. Let (ψn)n∈N be a sequence of smooth functions from [0,∞) to [0,∞) such that for all n ∈ N, the function ψn is supported on h− 1 n+1i and: n+1 , 1 Thus, using the notations of Lemma (6.9), we note that: ψn(r) rdr = 1 2π . 0 Z ∞ 2 Z ∞ − π π 2 0 ZR2 ψ◦ n =Z ψn(r) rdrdθ = 2π 2π = 1. Let ε > 0 be given. By Lemma (6.8), we have: Thus, there exists N ∈ N such that for all n > N , the following inequality holds: We may thus apply Lemma (6.6) to conclude that for all ξ ∈ D1(cid:16)Dp,q,d θ (cid:17) and n > N : lim n→∞ZZR2 ZZR2 ψ◦ n(x, y)k(x, y)k dxdy = 0. ψ◦ n(x, y)k(x, y)k dxdy < ε 4πð (cid:13)(cid:13)(cid:13) ξ − σ ψ◦ n ð,dξ(cid:13)(cid:13)(cid:13)H p,q,d θ 6 ε 2 . 36 FRÉDÉRIC LATRÉMOLIÈRE (cid:17) by Lemma (6.9), and D1(cid:16)Dp,q,d (cid:17) (cid:17) by σ (cid:17) for all n ∈ N. In particular, there exists ψ◦ n ð,d θ θ Now, σ ψ◦ n ð,d is compact in (cid:16)H p,q,d θ ,k · kH p,q,d θ θ θ θ θ a ε ψ◦ N ð,d 2 -dense subset Bε in σ by construction. Thus the image of D1(cid:16)Dp,q,d is bounded for k · kH p,q,d is totally bounded in (cid:16)H p,q,d ,k · kH p,q,d (cid:17). D1(cid:16)Dp,q,d (cid:17), then there exists η ∈ Bε such that: Consequently, if ξ ∈ D1(cid:16)Dp,q,d (cid:13)(cid:13)(cid:13) ð,d ξ(cid:13)(cid:13)(cid:13)H p,q,d η − σ (cid:17) is totally bounded. Thus kξ − ηkH p,q,d We thus conclude that D1(cid:16)Dp,q,d Moreover, for all (x, y) ∈ R2, the map ξ 7→ θ 2πðk(x,y)k 6 ε. H p,q,d ξ−ξk ε 2 ψ◦ N xy 2 kα 6 x,y, ð,d . θ θ θ θ is lower semi-continuous with respect to k·kH p,q,d ((−∞, 1]) is closed. Since H p,q,d θ θ is complete and D1(cid:16)Dp,q,d θ . Hence D1(cid:16)Dp,q,d θ (cid:17) = (cid:17) is closed (cid:3) and totally bounded, it is in fact compact, as desired. thus Dp,q,d θ θ (cid:16)Dp,q,d (cid:17)−1 is continuous, and We summarize the results of this section with the following theorem announcing that indeed, we have defined D-norms on Heisenberg modules, turning them into metrized quantum vector bundles over quantum 2-tori. Theorem 6.11. Let H p,q,d p ∈ Z, q ∈ N \ {0} and d ∈ qN \ {0}. Let ð = θ − p a norm on R2. If we set, for all ξ ∈ H p,q,d be the Heisenberg module over Aθ for some θ ∈ R, q and assume ð 6= 0. Let k · k be : θ θ kξkH p,q,d θ ,(cid:13)(cid:13)(cid:13) σx,y ð,dξ − ξ(cid:13)(cid:13)(cid:13)H p,q,d 2πðk(x, y)k θ : (x, y) ∈ R2 \ {0}  , (cid:13)(cid:13)(cid:13) βexp(ix),exp(iy) θ a − a(cid:13)(cid:13)(cid:13)Aθ : (x, y) ∈ R2 \ {0}  k(x, y)k ,Aθ, Lθ(cid:17) is a Leibniz metrized quantum vector bun- , Dp,d,q θ and for all a ∈ Aθ: Dp,d,q θ (ξ) = sup  Lθ(a) = sup  ,h·,·iH p,q,d θ then (cid:16)H p,q,d θ dle. Proof. Proposition (6.4) proves that Dp,q,d is a norm on a dense subspace of H p,q,d which satisfies the inner and modular quasi-Leibniz inequalities and, by construc- tion, Dp,q,d . θ θ θ > k · kH p,q,d θ Lemma (6.10) moreover gives us that D1(cid:16)Dp,q,d θ (cid:17) is compact for k · kH p,q,d θ . (cid:3) References 1. S. White J. Zacharias A. Hawkins, A. Skalski, On spectral triples on crossed products arising from equicontinuous actions, Math. Scand. 113 (2013), 262 -- 291, arXiv:1103.6199. D-NORMS ON HEISENBERG MODULES 37 2. K. Aguilar and J. Kaad, The Podlès sphere as a spectral metric space, Sub (2018), 23 pages, ArXiv: 1803.03027. 3. K. Aguilar and F. Latrémolière, Quantum ultrametrics on af algebras and the Gromov -- Hausdorff propinquity, Studia Mathematica 231 (2015), no. 2, 149 -- 194, ArXiv: 1511.07114. 4. O. Bratteli and D. Robinson, Operator algebras and quantum statistical mechanics i, Springer- Verlag, 1979. 5. A. Connes, C -- algèbres et géométrie differentielle, C. R. de l'academie des Sciences de Paris (1980), no. series A-B, 290. 6. A. Connes, Compact metric spaces, Fredholm modules and hyperfiniteness, Ergodic Theory and Dynamical Systems 9 (1989), no. 2, 207 -- 220. 7. A. Connes and M. A. Rieffel, Yang-mills for noncommutative two-tori, Contemporary Math 62 (1987), no. Operator algebras and mathematical physics (Iowa City, Iowa, 1985), 237 -- 266. 8. G. Folland, Harmonic analysis in phase space, Princteon University Press, 1989. 9. R. Hoegh-Krohn, M. B. Landstad, and E. Stormer, Compact ergodic groups of automorphisms, Annals of Mathematics 114 (1981), 75 -- 86. 10. A. Kleppner, Multipliers on Abelian groups, Mathematishen Annalen 158 (1965), 11 -- 34. 11. F. Latrémolière, Convergence of fuzzy tori and quantum tori for the quantum Gromov -- Hausdorff Propinquity: an explicit approach., Münster Journal of Mathematics 8 (2015), no. 1, 57 -- 98, ArXiv: math/1312.0069. 12. 13. 14. 15. , Curved noncommutative tori as Leibniz compact quantum metric spaces, Journal of Math. Phys. 56 (2015), no. 12, 123503, 16 pages, ArXiv: 1507.08771. , The dual Gromov -- Hausdorff Propinquity, Journal de Mathématiques Pures et Ap- pliquées 103 (2015), no. 2, 303 -- 351, ArXiv: 1311.0104. , Quantum metric spaces and the Gromov-Hausdorff propinquity, Noncommutative geometry and optimal transport, Contemp. Math., no. 676, Amer. Math. Soc., 2015, ArXiv: 150604341, pp. 47 -- 133. , Equivalence of quantum metrics with a common domain, Journal of Mathematical Analysis and Applications 443 (2016), 1179 -- 1195, ArXiv: 1604.00755. 16. F. Latrémolière, The modular Gromov -- Hausdorff propinquity, Submitted (2016), 67 pages, ArXiv: 1608.04881. 17. F. Latrémolière, The Quantum Gromov-Hausdorff Propinquity, Trans. Amer. Math. 2015, electronically published on May 22, no. Soc. 368 (2016), http://dx.doi.org/10.1090/tran/6334, ArXiv: 1302.4058. 365 -- 411, 1, 18. 19. , A compactness theorem for the dual Gromov-Hausdorff propinquity, Indiana Univer- sity Journal of Mathematics 66 (2017), no. 5, 1707 -- 1753, ArXiv: 1501.06121. , The triangle inequality and the dual Gromov-Hausdorff propinquity, Indiana Univer- sity Journal of Mathematics 66 (2017), no. 1, 297 -- 313, ArXiv: 1404.6633. 20. F. Latrémolière, Convergence of heisenberg modules for the modular Gromov-Hausdorff propinquity, Submitted (2018), 34 pages. 21. F. Latrémolière and J. Packer, Noncommutative solenoids and the gromov-hausdorff propin- quity, Proc. Amer 145 (2017), no. 5, 1179 -- 1195, ArXiv: 1601.02707. 22. H. Li, C ∗-algebraic quantum Gromov-Hausdorff distance, (2003), ArXiv: math.OA/0312003. 23. N. Ozawa and M. A. Rieffel, Hyperbolic group C ∗-algebras and free product C ∗-algebras as compact quantum metric spaces, Canad. J. Math. 57 (2005), 1056 -- 1079, ArXiv: math/0302310. 24. M. A. Rieffel, The cancellation theorem for the projective modules over irrational rotation C ∗-algebras, Proc. London Math. Soc. 47 (1983), 285 -- 302. 25. 26. , Projective modules over higher-dimensional non-commutative tori, Can. J. Math. XL (1988), no. 2, 257 -- 338. , Continuous fields of C* -- algebras coming from group cocycles and actions, Math. Ann. 283 (1989), 631 -- 643. 27. M. A. Rieffel, Deformation-quantization for actions of Rd, Memoirs of the AMS, vol. 106, American Mathematical Society, 1993. 28. M. A. Rieffel, Metrics on states from actions of compact groups, Documenta Mathematica 3 (1998), 215 -- 229, math.OA/9807084. 29. , Group C ∗-algebras as compact quantum metric spaces, Documenta Mathematica 7 (2002), 605 -- 651, ArXiv: math/0205195. 38 30. FRÉDÉRIC LATRÉMOLIÈRE , Gromov-Hausdorff distance for quantum metric spaces, Mem. Amer. Math. Soc. 168 (March 2004), no. 796, math.OA/0011063. 31. Sundaram Thangavelu, Lectures on hermite and laguerre expansions, Mathematical notes, vol. 42, Princteon University Press, 1993. 32. G. Zeller-Meier, Produits croisés d'une C*-algèbre par un groupe d' Automorphismes, J. Math. pures et appl. 47 (1968), no. 2, 101 -- 239. E-mail address: [email protected] URL: http://www.math.du.edu/~frederic Department of Mathematics, University of Denver, Denver CO 80208
1907.05832	1	1907	2019-07-12T16:43:17	An Introduction to Abstract Classification Theory in the Operator Algebraic Setting	[ "math.OA", "math.LO" ]	In the setting of modern mathematical logic and model theory, classification theory has been one of the landmark achievements of the field. Likewise, the classification of UHF-algebras and AF-algebras were substantial contributions to the field of operator algebra theory. These seemingly disparate topics of study in mathematics, model theory and operator algebras, have in recent years become closely related in many respects. I here attempt to bridge the gap between these two topics by discussing how operator algebraic classifications may be understood in terms of model-theoretic classification theory. This introductory article assumes basic familiarity with model theory and linear operator, but higher-level concepts are introduced when necessary. The focus of this introduction is conceptual and informal, and as such, many results are stated without proof, but relevant sources are cited for completeness. The reader should take this not as a detailed review, but rather as an overview of a general narrative thread connecting these two branches of modern mathematics.	math.OA	math	An Introduction to Abstract Classification Theory in the Operator Algebraic Setting Patrick Fraser [email protected] University of Toronto Abstract In the setting of modern mathematical logic and model theory, classifica- tion theory has been one of the landmark achievements of the field. Likewise, the classification of UHF-algebras and AF-algebras were substantial contri- butions to the field of operator algebra theory. These seemingly disparate topics of study in mathematics, model theory and operator algebras, have in recent years become closely related in many respects. I here attempt to bridge the gap between these two topics by discussing how operator algebraic classifications may be understood in terms of model-theoretic classification theory. This introductory article assumes basic familiarity with model theory and linear operator, but higher-level concepts are introduced when necessary. The focus of this introduction is conceptual and informal, and as such, many results are stated without proof, but relevant sources are cited for complete- ness. The reader should take this not as a detailed review, but rather as an overview of a general narrative thread connecting these two branches of modern mathematics. 1. Introduction Freeman Dyson famously wrote about a distinction common to the sci- ences between the diversifiers and the unifiers [1]. In order for a science to mature, it requires these two groups of people; the diversifiers who explore the vast jungle of the world, uncovering new, beautiful objects to study, pro- liferating novel ideas, and the unifiers, who fly above this jungle and piece together the distinct components, drawing them together into a coherent picture. Both are essential to understand the world fully. In mathematics, Based on an essay submitted for MAT437: K-theory and C ∗-algebras July 15, 2019 too, we see a vast array of diversifiers and unifiers. This essay focuses on the general goals and methods of the latter. One of the common ways of unifying ideas in mathematics is through so-called classification. Initially, the mathematician may notice that there is a particularly interesting class of objects which have certain properties that may be used to understand how things work under a new light. Such was the case with the early formal introduction of what are now standard algebraic structures, such as groups, rings, and fields. Once such basic objects have been noticed, one then seeks to explore their behaviour. One might notice, for instance, that not all rings are commutative, but when they are, certain features appear which may be exploited to further our understanding. The initial piece of discovery opens a path, and mathematicians eagerly map their way along this new path, until they have a fairly detailed understanding of its cartography. Once the objects under consideration become sufficiently well understood, it becomes reasonable to try to ask the strongest, most general questions. Specifically, once mathematicians realize that a certain class of mathematical objects has interesting behaviour, they might want an explicit list of all of these objects of a particular form (up to some sort of equivalence). Often, obtaining such a complete list is very difficult; in many cases, a classification program first identifies a collection of quantities which are invariant under isomorphisms, and then claim that the objects in question are classified if these invariants are enough to tell apart any two non-isomorphic objects. The invariants themselves may be quite difficult to compute, and so a complete list is often unattainable (such is the case with von Neumann algebras, whose classification program is considered complete, even though it is often quite difficult to say whether or not two von Neumann algebras are isomorphic) but in principle, these invariants would allow one to compare any two objects and determine if they are isomorphic or not. This is the game of classification. An ideal classification program would consist of a complete list, up to isomorphism, of all of the objects which behave in a certain way. In an elementary group theory course, one might, for instance see the historically significant classification of finite abelian groups. This classification tells us that a group G is a finite abelian group if and only if G ∼= Z/n1Z×...×Z/nk Z for sufficient ni's and k; it gives us a complete list, up to isomorphism, of all possible finite abelian groups. In more sophisticated settings, classification theorems provide a powerful correspondence between the abstract and the concrete, and are usually landmark theorems in their disciplines. 2 Classification theory is a field of study in its own right, generally sub- sumed as a topic within model theory and mathematical logic. Here, I present an elementary introduction to classification theory in its model-theoretic set- ting, and then discuss how it may be concretely applied to the theory of operator algebras, UHF and AF-algebras in particular. 2. Classification Theory I begin with the broad notions of classification theory from the model- theoretic standpoint, based on [2]. In lieu of a basic introduction to model theory, I point the interested reader towards the very accessible introduc- tion [3] and the slightly more sophisticated [4]. Assuming familiarity with the concept of a first-order predicate language,1 I take the following defini- tions: Definition: Fix a language L. An L-structure A consists of a set of elements (the universe), a set of constants of the same cardinality as the constants of L, and for every relation and function symbol in L, a relation and function between elements of the universe. Languages have no meaning, but one may map the symbols of a language to a structure in which those symbols take on a meaning using an assign- ment function. Generally, one is concerned with models of a particular set of formulas; that is, some structure in which all of a certain class of abstract formulas are true. Definition: Let L be a language. Given a set T of L-formulas, an L- structure M is a model of T if M (cid:15) φ for all φ ∈ T . Structures are objects such as the natural numbers equipped with a binary addition function and a binary ordering relation, which may be taken as 1Essentially, such a language is a meaningless assemblage of symbols which represent variables, constants, relations, and functions, with standard logical connectives like ∧, →, and ∨, as well as a negation symbol ¬ and universal and existential quantification symbols ∀ and ∃ respectively, which have basic rules for constructing grammatical terms, formulas, and sentences. The language itself is meaningless, but concrete meaning is prescribed to symbols and formulas when they are interpreted in a structure of that language. 3 models for certain abstract formulas in the language. For example, in a language with one constant and an order relation L = {0, ≤}, the abstract formula φ := (∃x)[(∀y)(x ≤ y)] is merely a string of symbols, but when interpreted in the structure of the natural numbers, this formula is satisfied because it means that the natural numbers have a smallest element, which is true. However, this same formula is not satisfied by the L-structure that is the rational numbers (equipped with the usual ordering), for instance, because they do not have a lowest element under their usual ordering. As such, N models φ, but Q does not (and we write N (cid:15) φ and Q 6(cid:15) φ). When we seek to "classify" a collection of objects, we require first that they are in fact objects of the same kind; if we are concerned with groups, we are concerned with groups. If we are concerned with approximately finite C ∗-algebras, we are concerned with approximately finite C ∗-algebras. But we are never comparing apples with oranges. Classification takes place al- ways within the confines of kind. Formally, this amounts to saying that we are restricting ourselves to L-structures which are all grounded in the same language L, and further, which all satisfy the same axioms. That is, they must be models of the same L-theory. With this in mind, I give the fol- lowing definitions which will help us, in broad strokes, to study the general behaviour of classes of models for a given theory: Definition: Let K = (K, ≤K) where K = Mod(T ) the set of all models of some first-order L-theory T , and ≤K is partial ordering on K. Then K is an abstract elementary class if: (i) K is closed under isomorphism and ≤K pre- serves isomorphisms, (ii) if M ≤K N , then M is a substructure of N , (iii) (downward Lowenheim-Skolem theorem) there exists a cardinal α ≥ ℵ0 + L such that for every M ∈ K and A ⊂ M, there is a N ∈ K with N ≤K M, A ⊂ N , and N ≤ A + α, and (iv) (Tarski-Vaught Chains) K is closed under ≤K-increasing chains of arbitrary regular cardinality. [2, 5] Essentially, abstract elementary classes are classes of models for a par- ticular theory which have enough particular constraints to ensure that the elements in each class all 'look alike' in a precise sense. The models of a theory, particularly the infinite models, are in general not easy to pin down. Indeed, the following famous theorem ensures that, if there are any infinite models of some some first-order theory, there are always many distinct infi- nite models. 4 Theorem (Lowenheim-Skolem): Suppose T is a countable theory in a first-order language L which admits an infinite model M. Then for all car- dinals κ > L, there exists a first-order theory of size κ which is elementary equivalent to M (see, for instance, [3, 4]). One may heuristically understand this as saying that an infinite model of some theory can always be made more or less infinite, for instance, by extending it to a non-standard model. To understand the behaviour of such infinite models, which may, in general, be very badly behaved, it is valuable to understand categoricity: Definition: An abstract elementary class K is categorical in λ if it has one and only one model of cardinality λ up to isomorphism). The class of cardinals for which K is categorical is called the categoricity spectrum of K and is denoted cat(K). Categoricity allows us to compare the relative size of elements of some abstract elementary class, and indeed, there are several nice facts about categoricity which lend themselves nicely to the demarcation of distinct iso- morphism classes. Notably, the following theorem is substantial [6]: Theorem (Morley): Fix some countable first-order L-theory T . Let λ be a cardinal, and I(λ, K) be the number of models in K of cardinality λ (up to isomorphism). Then if for some λ > T + ℵ0, I(λ, T ) = 1, then I(µ, T ) = 1 for all cardinals µ > T + ℵ0. That is, if the categoricity spectrum of some countable first-order theory includes a single model of some infinite cardinality larger than that of the language used, then indeed the categoricity spectrum of that theory extends to all such cardinals. Categoricity is the gateway to stable model theory which attempts to understand how well-constrained the infinite models of some theory are. If one wants to consider more specific properties unique to a models of a particular theory, the route to formalizing this is definability, which tells you when the a particular theory admits the specification of par- ticular subsets based on a precise set of properties which can be formulated in the language in question. Definitions: Let A be an L − structure with universe A and let B ⊂ A. 5 A set X ⊂ An is B-definable if there exists some φ(v1, ..., vn, w1, ..., wn) such that X = {a ∈ Anφ(a, b), b ∈ Bm}. We can now say that classification theory deals with the problems of un- derstanding isomorphism classes within abstract elementary classes of mod- els, and frequently, the understanding of the isomorphism classes relies on what features are definable within a particular first-order theory and how the categoricity of the models behaves. The high-level goal of classification theory is to look at a collection of such classes {K}I and present clear mecha- nisms for demarcating the "nice" or "well-behaved" classes from those which are "messy" or "complicated". In this way, one might obtain a strong way to pin down the well-behaved classes, and prove substantive claims about all of these classes through more straight-forward case work. The most important result along these lines has been due to Shelah who determined the precise conditions for which an abstract elementary class admits a classification (or in more technical language, has a structure theory) of its isomorphism classes (the so-called dichotomy theorem) [2, 7, 8]. This theorem states that: Theorem (Shelah):An abstract elementary class K has a structure the- ory if there is a class of subsets {A}I of the universe of each model in K (called cardinal invariants), which are definable in each model, such that, for any two M, N ∈ K, M ∼= N if and only if AN for all invariants Ai. i = AM i The low-level goal of classification theory is, given a particular class K which is known to admit a classification, to determine concretely all of the elements of K up to isomorphism, or at least to determine a cardinality bound for how many such distinct elements there may be. The later part of this project was, again, solved by Shelah [2, 7, 8]. In the operator algebra setting, once one has established that a particular kind of operator algebra with a first-order set of axioms in fact does admit a classification (a structure theory), then the important work is to determine every element of the class of models of those axioms up to isomorphism. The earliest results along these lines were the classification of UHF-algebras due to Glimm [9, 10] and the classification of AF-algebras due to Elliott [9, 11]. Before proceeding to the operator algebra setting, to see a nice case of classification theory in action in a comfortable setting, consider vector spaces. Example: Vector Spaces. The class of all vector spaces V is the class 6 of all models of the theory of vector spaces TV S. One can show readily that TV S is a first-order theory [3]. Therefore, the class of vector spaces forms an abstract elementary class (ordered under inclusion, where all of the other more complicated requirements are satisfied more or less trivially). Given a vector space V ∈ V, there is a definable subset of V which is the set of all bases of V (the relation under which this set is definable in the language of Marker [4], is span). It is an elementary fact of linear algebra (and thus a theorem provable in the theory TV S, and so true of all vector spaces) that the cardinality of every basis of a particular vector space is the same. Thus, the cardinality of every element of the set of bases of V is constant. We call this cardinality the dimension of V , and it is provable using elementary linear algebra that two finite-dimensional vector spaces are isomorphic if and only if their dimensions are the same. Therefore, the cardinal invariant for the structure theory of vector spaces is the dimension, which is the cardinality of any element of the definable subset of the universe of V which is the set of bases for V . 3. C ∗-algebras and their K-Theory In the classification of C ∗-algabras, K-theory is a valuable tool, as it al- lows one to construct a related structure to the algebra in question, namely the ordered K0-group (as well as higher K-groups), which encodes lots of the information about the algebra while allowing questions to be posed in the context of abelian groups rather than operator algebras. Because of this, K- theory is helpful for the classification of C ∗-algebras, as definable invariants for the algebra may be functorially pushed forward to the K-theory setting where they are more natural and easier to work with. In the special case of AF-algebras, the classification process is made very clear. After presenting the basic ideas pertinent to studying C ∗-algebras and their K-theory, I walk through the examples of UHF-algebras and AF-algebras to understand their classification in model theoretic terms. I then discuss more general ideas presently being pursued for more general classifications. As such, I base the following on [9]. Definition: A C ∗-algebra is a Banach algebra equipped with a conjugate- linear homomorphic isometric involution x 7→ x∗ such that x∗x = xx∗ = x2. 7 Broadly speaking, a C ∗-algebra is an abstract representation of some al- gebra of bounded, self-adjoint, norm-closed operators on a Hilbert space. In- deed, the Gelfand-Naimark-Segal (GNS) construction tells us precisely that every abstract C ∗-algebra admits such a Hilbert space representation. This notion of concrete representation is made stronger by the existence of the Gelfand transform, which acts as an isometric ∗-isomorphism between ele- ments of any given commutative, unital C ∗-algebra and the set of bounded continuous functions on some particular compact Hausdorff space2. The study of specific operator algebras over particular Hilbert spaces together with their associated spectra is a significant area of study in its own right, but the abstract treatment of general C ∗-algebras is likewise very rich, and particularly amenable for understanding the behaviour of such systems in the most unified, general setting. Even though 'zooming out' to the higher-level theory of abstract operator algebras allows one to forget about many of the precise details of the systems they are considering, C ∗-algebras are, in their own right, highly complicated systems. As such, it is valuable to have tools to simplify problems pertaining to C ∗-algebras further. One such method for gaining further understanding comes from K-theory. When considering any operator algebra, the class of idempotent elements is of particular importance (one need only look towards von Neumann's for- malism of quantum mechanics, or more modern POVM-based formalisms to see how practically powerful projections may be in the wild [12]). Specifi- cally, projections in an algebra 'pin down' much of the dimension structure of the algebra3 quite concretely in a precise sense using unitary equivalence and Murray-von Neumann equivalence: Definition:Let A be a C ∗-algebra. Let p, q ∈ A. Then p and q are Murray-von Neumann equivalent (written p ∼ q) if there exists some v ∈ A such that p = v∗v and q = vv∗. 2In this sense, it may indeed be reasonable to think of the study of C ∗-algebras as an abstract study of noncommutative topology. 3For a particularly easy example, given the matrix algebras Mn(C) and Mk(C), one has that Mn(C) ∼= Mk(C) if and only if n = T r([In]0) = T r([Ik]0) = k where [In]0 and [Ik] are the K-theoretic identities associated with each algebra (here, just the matrix identity up to Murray-von Neumann equivalence). 8 It can readily be shown that if p, q ∈ A are projections which are homo- topic or unitarily equivalent, then p ∼ q. There is a canonical matrix algebra semigroup of projections associated with any C ∗-algebra A given by: Definition: Let Mn(A) denote the algebra of m×m matrices with entries in the C ∗-algebra A. Let Pn(A) denote the class of projections in the ma- trix algebra Mn(A). Then the projection semigroup is P∞(A) := ∪∞ n=1Pn(A). Addition in this semigroup is given by the usual direct sum ⊕ on matri- ces. There is an associated notion of Murray-von Neumann equivalence over P∞(A) given by: Definition: Let p ∈ Pn(A) and q ∈ Pm(A). Then p ∼0 q if there is some v ∈ Mm,n(A) such that p = v∗v and q = vv∗. The triumph of K-theory is that, in light of this equivalence on the pro- jection semigroup associated with a C ∗-algebra A, one can readily construct a functor from the category of C ∗-algebras to the category of ordered abelian groups, which preserves most of the relevant information about the structure of the algebra through its idemptotent elements. This functor is called the K0 functor, and it has many beautiful features. The K0 functor (which is a map which takes a C ∗-algebra A to an ordered abelian group K0(A)) is defined as follows: Definition: Let S be a semigroup. The Grothendieck group G(S) is given by the quotient G(S) = S × S/ ∼G where (x1, y1) ∼G (x2, y2) if and only if there is some z with x1 + y2 + z = x2 + y1 + z. The K0 group associated with a C ∗-algebra A is given by the Grothendieck group of the projection semigroup of A. That is, K0(A) = G(P∞(A)/ ∼0). To better understand what K0(A) 'looks like', the following results are quite valuable: Theorem: Let [p]0 denote the ∼0 equivalence class of some element p ∈ P∞(A). Then K0(A) = {[p]0 − [q]0 : p, q ∈ Pn(A), n ∈ N}. Moreover, [p]0 + [q]0 = [p ⊕ q]0. Finally, if p and q are homotopic, then [p]0 = [q]0. Furthermore, the functoriality of K0 tells us that it distributes in the 9 natural sense over the composition of ∗-homomorphisms and preserves the identity map. Additionally, if one has the categorial concept of inductive limits in place (described below), there is a natural sense in which K0 is continuous. With the notion of K0 in place, we may begin to talk about specific classes of C ∗-algebras in terms of their K-theory. The most elementary class of C ∗-algebras is that of finite C ∗-algebras (each of which is isomorphic to the matrix algebra Mn(C) for some n). These algebras are not terribly ex- citing, but one can form many highly non-trivial classes of C ∗-algebras out of them using inductive limits. In their general construction, inductive lim- its act like limits of sequences of objects in some category which are linked to each other by connecting arrows; they generalize limits to more abstract categories. Specifically: Definition: Let C be some category. Let {Oi, φij} be a sequence of : Oi → Oj (i < j). Then the inductive objects Oi in C and arrows φij limit lim→{Oi, φij} is the object O∞, which has a canonical set of maps φi : Oi → O∞ such that O∞ = ∪∞ i=1φi(Oi) and where φj = φi ◦ φij (i.e. the below diagram commutes). φ1 φ23 φ12 O1 O2 ... φ2 O∞ It is important to note that not all categories admit inductive limits (for instance, the category of sets). When we take the category in question to be the category of C ∗-algebras, we can form inductive limits (taking the connecting maps to be ∗-homomorphisms), and it turns out that even the inductive limits of finite C ∗-algebras can be quite fascinating objects to study. For our interests here, this setting also provides us with a straightforward and historically significant instance of classification theory found in the wild. Specifically, the study of UHF-algebras and AF-algebras (which are inductive limits of finite C ∗-algebras) provide an interesting setting to connect Shelah's abstract classification theory to concrete and relevant classification programs. 10 4. The Classification of C ∗-algebras To proceed, we begin with the following definition of approximately finite (AF) algebras: Definition: An AF-algebra is a C ∗-algebra which is isomorphic to the inductive limit of a sequence of finite C ∗-algebras. The first example of classification theory as it pertains to C ∗-algebras which makes clear how these ideas are all inter-connected in a concrete way is the classification of uniformly hyper-finite (UHF) algebras due to Glimm [10]. An algebra A is a UHF-algebra if it is an AF-algebra where the connecting ∗-homomorphisms are unit preserving. UHF-algebras may be classified using supernatural numbers which are defined as [9]: Definition: Let {ni}i∈N be a (finite or countable) set of cardinals where i where pi ni ≤ ℵ0, and let n formally be expressed as the product n = Πi∈Npni is the ith prime number. Any such n is a supernatural number. It will be shown that {ni}i∈N is the set of cardinal invariants of models of 4, the first-order theory of UHF algebras. Without getting our hands TU HF too dirty, stated without proof, we have the following theorems [9]: Theorem: Let G ⊂ (Q, +) be a subgroup containing 1. Then there is a supernatural number n associated with G such that G = Q(n), the subgroup composed of elements of Q where the denominator is a factor of n. Theorem: Let A be a UHF algebra. Then (K0(A), [1A]0) ∼= (Q(n), 1) for some supernatural number n. From this, we see that every UHF-algebra has an associated supernat- 4It should be noted that, in general, C ∗-algebras are not universally first-order ax- iomatizable, and even when they are, they generally require an uncountable language, making certain results, such as Lowenheim-Skolem results, more subtle. However, it can be shown that abelian algebras, and unital finite C ∗-algebras (as well as many other ele- mentary kinds of algebras) are first-order axiomatizable, and so we do not get into trouble here [13]. 11 ural number n, and so to every UHF-algebra, a set of cardinal invariants {ni}i∈N may be assigned. This is a step towards classification, however, we do not yet know that this associated supernatural number determines a given UHF-algebra uniquely up to isomorphism; a priori, there may exist two non-isomorphic UHF-algebras (models of TU HF ) with the same associ- ated supernatural number. However, the major result of Glimm says that the associated supernatural number of a UHF-algebra is in fact enough to determine that UHF-algebra up to isomorphism. Specifically, we have: Theorem: (Glimm): If A and A′ are two UHF algebras with associated supernatural numbers n and n′ respectively, then A ∼= A′ if and only if n = n′. In the model-theoretic setting, if n = n′, since prime factorizations are unique, we see that ni = n′ i}i∈N, and so we see that A ∼= A′ if and only if ni = n′ i for all i ∈ N and so {ni}i∈N is a (countable) set of cardinal invariants which constitute the structure theory for TU HF . Therefore, we see a nice reflection of Shelah's classification theory in Glimm's theory for UHF-algebras. i for all elements of {ni}i∈N and {n′ In light of the fact that (K0(A), [1A]0) ∼= (Q(n), 1) from above, we see that in fact we may restate this classification theorem to be in the language of K- theory. Specifically, A ∼= A′ if and only if (K0(A), [1A]0) ∼= (K0(A′), [1A′]0), because the K groups are determined by the associated supernatural num- bers. Further connecting the classification of UHF-algebras to classification the- ory in general, if we recall the low-level goal of determining the cardinality of the set of isomorphism classes of models for a particular theory (TU HF in this case), we see that there is a converse theorem which comes in handy. It is shown in [9] that for every supernatural number n, there is a UHF-algebra with associated supernatural number n. A supernatural number is exactly specified by the set {ni}i∈N (noting that we have here ordered the ni's using the same usual ordering of the corresponding prime numbers, and that prime factorization is unique). Since there are ℵ0 elements of {ni}i∈N, and since each ni may take on ℵ0 possible values, there are ℵℵ0 0 many supernatural numbers. One may show that ℵℵ0 0 = 2ℵ0 = ℵ1 = R, and so there is a (small) uncountable set of distinct isomorphism classes of UHF algebras. I now look at unital AF-algebras in general. Let TAF be the first-order theory of unital AF-algebras (everything to fol- low may be generalized to the non-unital case [9, 11], but the details involved 12 do not lead to a clearer understanding of the general classification method). The question which must be answered to classify AF-algebras in the sense of Shelah is: What set of cardinals {λi}i∈I is definable in any model A (cid:15) TAF i}i∈I if and only if A ∼= A′? The theorem of Elliott [11] such that {λi}i∈I = {λ′ of course indicates how this may be answered: Theorem (Elliott): Two unital AF-algebras A and B are isomorphic if and only if (K0(A), K0(A)+, [1A]0) ∼= (K0(B), K0(B)+, [1B]0). Thus, the ordered K0 group with its unit is an algebraic invariant which is sufficient to classify AF-algebras. But how do we interpret this in terms of sets of cardinals? The following result, found in [9], brings clarity to the matter: Theorem: Let A be an AF algebra. Then (K0(A), K0(A)+) is a dimen- sion group. Conversely, for every dimension group (G, G+), there exists an AF-algebra A such that (G, G+) ∼= (K0(A), K0(A)+). The notion of dimension is valuable, because it allows us to tease out the cardinals from the discussion. As is explained clearly by Effros et al. [14], 'dimension' may take on many different meanings. In the context of matrix algebras, dimension refers to the rank of the projections in the algebra. In the von Neumman algebra setting, dimension becomes a non-negative real value. In the C ∗-algebra setting however, dimension refers to a value assigned to Murray-von Neumman equivalence classes of projections in the K0-group of the algebra; dimension then refers to values in some ordered abelian group. In this way, we see why the term dimension group is fitting in the AF-algebra setting; the dimension group is precisely the K0 group in which elements of the algebra may have their 'dimension' defined. If we wish to understand how Elliott's classification theorem for AF- algebras may be expressed in terms of cardinal invariants for the abstract elementary class of models of the first-order theory for AF-algebras, it is valuable to follow Bratteli [15] in his diagrammatic representation of AF- algebras using infinite directed graphs, and read off the cardinal invariants from this graph in terms of counting the vertices and edges in a particular way. Bratteli diagram for an AF-algebra A ∼= lim→{Ai, φij} are constructed thusly: For each Ai, there will be a row of the Bratteli diagram consisting 13 ∼= Ls of a collection of points. Each Ai is isomorphic to some finite factor space of finite-dimensional complex matrix algebras. That is, Ai k=1 Mki(C). Thus, in the ith row of the Bratteli diagram, draw one point for each factor Mki(C) of Ai, and label it by the integer ki. Given the ith row of a Bratteli diagram, the i + 1th row is obtained by likewise drawing one point for each factor space of Ai+1 (in its finite-dimensional complex matrix algebra repre- sentation, for instance). Then the ith row is connected to the i + 1th row by drawing edges connecting the points in the ith row to points in the i+1th row precisely when a given point in the ith row is mapped to the corresponding point in the i + 1th row under the connecting map φii+1. The number of lines connecting two points denotes the multiplicity of the map connecting the two points; if φij : Ai → Aj is a ∗-homomorphism, then the multiplicity of φij is defined to be T r(φij(p))/T r(p) for any non-zero projection p (it is easy to show that this quantity is independent of the choice of p). It turns out this graph representation completely characterizes the inductive limit and thus completely characterizes up to isomorphism the underlying AF-algebra. Two examples of Bratteli diagrams (taken from [9]) may be found in figures 1 and 2. ·1 ·2 ·3 ... ·1 ·1 ·2 ... Figure 1: The Bratteli diagram for the inductive sequence of finite-dimensional C ∗-algebras constructed by taking An = Mfn (C)⊕Mfn−1(C) where f0, f1, f2, . . . are the usual Fibonacci numbers, with connecting maps given by: (x, y) 7→ (cid:0)(cid:0) x 0 0 y (cid:1), x(cid:1). The ordered abelian K0 group (or the dimension group, as one chooses) associated with an AF-algebra, as Elliott's theorem showed, comes from the sequence of the K0-groups of the terms in the inductive sequence of finite C ∗- algebras which induce the AF-algebra in question. As such, to determine the 14 ·1 ·2 ·4 ·8 ... Figure 2: The Bratteli diagram for the inductive sequence of finite-dimensional C ∗-algebras given by An ∼= M2n(C), with the connecting maps given by φii+1(x) = diag(x, x). We see that the inductive limit of this sequence is isomorphic to M2∞ (C). It may be shown that this is also the (unique) C ∗-completion of the canonical anti-commutation relation (CAR) algebra over a separable infinite-dimensional Hilbert space, an algebra which is very important in mathematical physics, the study of quantum field theories and quan- tum statistical mechanics in particular (anything which makes use of antisymmetric Fock space). algebraically invariant dimension group of an AF-algebra, one may look at the dimension groups associated with each algebra in the sequence. But this is effectively the content of a Bratteli diagram! Bratteli diagrams show us, at each level, the dimension of each finite matrix sub-algebra (using the un- usual notion of dimension, which may readily be cast as a cardinal) together with the way in which these finite matrix algebras are 'stitched' together via their connecting maps, as well as the multiplicity of these connecting maps (which therefore determines how the rank of projections changes when pushed forward to the next row of the Bratteli diagram). The 'bottom row' of a Bratteli diagram thus determines exactly the dimension group of the in- ductive limit of the sequence, which is therefore isomorphic to the K0 group of the AF-algebra in question. The ith row of a Bratteli diagram may be completely specified by the number associated to each point, together with the edges connecting it to the previous row. Thus, the ith row may be specified completely by the set of cardinals {ni, {di,j}ni } where ni is the number of j=1, {m(i−1,s)→(i,r)}s=ni−1,r=ni s,r=1 15 s,r=1 elements of the ith row, di,j is the dimension associated with the point in the jth position (from left to right) of the ith row, and {m(i−1,s)→(i,r)}s=ni−1,r=ni denotes the multiplicities of each edge connecting the sth point in the i − 1th row to the rth point in the ith row (where each m(0,s)→(1,r) = 0 for all s and r since there are no connecting maps ending in the first row). These values are all finite integers for any given row. The 'bottom line' (i.e. the ith row in the limit when i → ℵ0, the countable infinity) may likewise be specified by this set of cardinals.5 The Bratteli diagram encodes all of the information about the K0 group of the AF-algebra in question, and so it is an equivalent description of the invariant necessary to classify AF-algebras, although it is more conducive to being read as a sequence of cardinals, formally putting us in a position to claim, in Shelah's language, that models of the theory TAF are classified up to isomorphism by a set of cardinal invariants. UHF and AF-algebras are not the only classes of C ∗-algebras to be classi- fied, nor are they the only classes of C ∗-algebras defined in terms of inductive limits to be classified. However, noting that, in the UHF-algebra case, the in- variant required could be equivalently expressed as the double (K0(A), [1A]0), and in the more general unital AF-algebra case, the invariant needed was only the slightly more sophisticated triple (K0(A), K0(A)+, [1A]0), one might be inclined to state a more general thesis, namely that all C ∗-algebras of some form may be classified by a more general K-theoretic invariant. This is the present program of classification, and one of the conjectured invariants is the so-called Elliott invariant [16]: Ell(A) = (K0(A), K0(A)+, [1A]0, K1(A), T (A), rA) Here, K1(A) is the standard K1 group (which is isomorphic to the K0 group of SA, the topological suspension of A), T (A) is the tracial space of A (i.e. the set of all tracial functionals in the dual space of A), and RA is the so-called coupling map which associates states with traces. The goal of the classification program was to classify all separable, unital, nuclear simple C ∗-algebras via this invariant. That is, ideally, one would end up with a theorem which states that, if A and B are two such algebras, then A ∼= B if and only if Ell(A) ∼= Ell(B). Unfortunately, there have been counter- 5For the set theorist, this is an terribly dull set of cardinals, as all of them are finite or countably infinite. However, boring or not, cardinals they are, and so they are satisfactory for us here. 16 examples preventing such a complete functorial classification, but it has held for many important cases, such as the aforementioned algebras, as well as irrational rotation algebras, and AT-algebras in general. There are presently new candidate invariants being proposed. However, these classification pro- grams have become extremely complicated and sophisticated, and a complete functorial classification may be far away; one of the major problems in this field currently is to try and understand precisely how complex the isomor- phism maps between C ∗-algebras may in principle be, thereby giving insight into the complexity of their classification [16]. 5. Conclusions I here presented a general overview of Shelah's classification theory, and the theory of cardinal invariants. After a cursory glance at the basics of K- theory for C ∗-algebras, I then recalled several known results, due primarily to Glimm and Elliott, about the classification of various classes of C ∗-algebras using particular K-theoretic invariants. I demonstrated how these invariants may be understood as cardinal invariants in terms of sequences defining su- pernatural numbers and sequences defining states of Bratteli diagrams. With this understanding in place, I described the goals of future classification. The topic of C ∗-algebra classification has, in recent years, become deeply rooted in model theory and mathematical logic, and so framing the discussion in terms of Shelah's classification theory reorients the setting of discussion to be more amenable to such discourse. In this way, we see that the marriage of model theory and operator algebra theory is, in a sense, preordained by the sophistication of the field, and will likely serve as an important tool for future investigation. References [1] Freeman Dyson. Infinite in All Directions: Gifford Lectures Given at Aberdeen, Scotland. Harper Perennial, 1985. [2] Saharon Shelah. Introduction to: classification theory for abstract ele- mentary class. arXiv:0903.3428, 2009. [3] Christopher C. Leary and Lars Kristiansen. A Friendly Introduction to Mathematical Logic. Milne Library, 2015. 17 [4] David Marker. Model theory: an introduction. Springer, 2002. [5] Rami Grossberg. Classification theory for abstract elementary classes. In Yi Zhang, editor, Logic and Algebra. AMS, 2002. [6] Michael Morley. Categoricity in power. Transactions of the American Mathematical Society, 114(2), 1965. [7] Wilfred Hodges. What is a structure theory? Bull. London Math. Soc., 19:209 -- 237, 1987. [8] Saharon Shelah. Classification theory. volume 92 of Studies in logic and the foundations of mathematics. North-Holland, 2 edition, 1990. [9] Flemming Larsen Mikael Rørdam and Niels J. Laustsen. An Introduction to K-theory for C-algebras. Cambridge University Press, 2000. [10] James G. Glimm. On a certain class of operator algebras. Transactions of the American Mathematical Society, 95(2):318 -- 340, 1960. [11] George A. Elliott. On the classification of inductive limits of sequences of semisimple finite-dimensional algebras. Journal of Algebra, 38:29 -- 44, 1976. [12] Isaac Chuang Michael Nielsen. Quantum Computation and Quantum Information: 10th Anniversary Edition. Cambridge University Press, 2010. [13] I. Farah, B. Hart, M. Lupini, L. Robert, A. Tikuisis, A. Vignati, and W. Winter. Model theory of c-algebras. arXiv:1602.08072, 2018. [14] David E. Handelman Edward G. Effros and Chao-Liang Shen. Dimen- sion groups and their affine representations. American Journal of Math- ematics, 102(2):385 -- 407, 1980. [15] Ola Bratteli. Inductive limits of finite dimensional c-algebras. Trans- actions of the American Mathematical Society, 171:195 -- 234, 1972. [16] Ilijas Farah. Logic for c-algebras. Draft of a book, April 2019. 18
1608.04795	1	1608	2016-08-16T22:24:29	Interpolation and Commutant Lifting with Weights	[ "math.OA" ]	Our two principle goals are generalizations of the commutant lifting theorem and the Nevanlinna-Pick interpolation theorem to the context of Hardy algebras built from $W^*$-correspondences endowed with a sequence of weights. These theorems generalize theorems of Muhly and Solel from 1998 and 2004, respectively, which were proved in settings without weights. Of special interest is the fact that commutant lifting in our setting is a consequence of Parrott's Lemma; it is inspired by work of Arias.	math.OA	math	INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS JENNIFER R GOOD Abstract. Our two principle goals are generalizations of the com- mutant lifting theorem and the Nevanlinna-Pick interpolation the- orem to the context of Hardy algebras built from W ∗-correspon- dences endowed with a sequence of weights. These theorems gener- alize theorems of Muhly and Solel from 1998 and 2004, respectively, which were proved in settings without weights. Of special interest is the fact that commutant lifting in our setting is a consequence of Parrott's Lemma; it is inspired by work of Arias. 1. Introduction This paper concerns weighted interpolation problems in the theory of operator algebras built from W ∗-correspondences. The classic origins of this study are summarized in the phrase weighted interpolation. The original interpolation problem, solved separately by Nevanlinna and Pick in the early twentieth century, sought conditions under which, given k ∈ N and two collections, {ωi}k i=1 ⊆ C, there exists a function φ ∈ H ∞(D) of norm at most 1 that interpolates this data; that is, i=1 ⊆ D and {λi}k (1.1) φ(ωi) = λi, 1 ≤ i ≤ k. Nevanlinna and Pick showed that such a φ exists if and only if is a positive semidefinite k × k matrix where K is the Szego kernel, K(w, z) = 1 1−w¯z for w, z ∈ D, reproducing kernel to the Hardy space H 2(D) [10] [13]. More generally, a weighted Hardy space, as described in [17], is a reproducing kernel Hilbert space with kernel K : Dr ×Dr → for a sequence of strictly positive numbers β = {βi}∞ i=0. Our primary goal is a weighted Nevanlinna- Pick type interpolation theorem that applies to the so-called weighted C, defined by K(w, z) =P∞ (wz)i i=0 β2 i Key words and phrases. Nevanlinna-Pick interpolation, Commutant Lifting, Double Commutant, W ∗-correspondence, noncommutative Hardy algebra, Se- quence of Weights. 1 (1.2) (cid:2)K(ωi, ωj)(1 − λiλj)(cid:3)k i,j=1 2 JENNIFER R GOOD Hardy algebra, a non-self-adjoint operator algebra recently studied in [9]. While we will expand on the details in a later section, in brief the setting of [9], inspired by Popescu's work in [14], begins with a W ∗- algebra M, a W ∗-correspondence over M denoted E, and a sequence of operator-valued weights Z that has been constructed from a so- called admissible sequence X. Our algebra of focus is the weighted Hardy algebra H ∞(E, Z), the algebra of adjointable operators on the Fock space F(E) that is generated by the weighted creation operators associated with Z and the left action maps of M on F(E). To gain insight into H ∞(E, Z) we may view it as an algebra of functions on its space of representations, as follows. Having fixed σ, a representation of M on a Hilbert space H, the associated representations of H ∞(E, Z) are in a sense determined by D(X, σ), a certain collection of operators in B(E ⊗σ H, H). Every point z ∈ D(X, σ) gives rise to a representation (σ × z) of H ∞(E, Z) on H, and we may consider an operator Y ∈ H ∞(E, Z) as a B(H)-valued function bY on D(X, σ) defined at z ∈ D(X, σ) by bY (z) = (σ × z)(Y ). In this setting, let us preview the main theorem. We form a weighted W ∗-version of the original Szego kernel, a map K from D(X, σ) × D(X, σ) to a collection of completely bounded maps on σ(M)′. For any choice of k ∈ N, {zi}k i=1 ⊆ D(X, σ), and two collections of opera- tors {Bi}k i=1 in B (H), there is an associated map A from the k × k matrices with entries in σ(M)′ to the k × k matrices with entries in B (H) that is defined at [Aij]k i=1 and {Fi}k i,j=1 to be j − Fi · K(zi, zj)(Aij) · F ∗ . i,j=1 j(cid:3)k The conclusion of our weighted Nevanlinna-Pick theorem is that A is completely positive if and only if there exists Y ∈ H ∞(E, Z) such that kY k ≤ 1 and (1.3) (1.4) (cid:2)Bi · K(zi, zj)(Aij) · B∗ Bi(cid:16)bY (zi)(cid:17) = Fi, 1 ≤ i ≤ k. A comparison of equations (1.1) and (1.4), as well as matrices (1.2) and (1.3), reveals the similarities with the original interpolation problem. Indeed, when M = E = C, σ is the one-dimensional representation of M, Z is the constant sequence at 1, and Bi = 1 for every i, matrix (1.3) reduces to matrix (1.2) with minor technical adjustments, and we obtain the original result of Nevanlinna and Pick. Even in the scalar case, our theorem is more general than that of Nevanlinna and Pick, for it applies to certain weighted Hardy spaces, precisely those that have the so-called complete Pick property described in [1]. For example, our INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 3 results apply to the Hardy and Dirichlet spaces, but not the Bergman space. Commutant lifting, pioneered by Sarason [16] and Sz.-Nagy and Foia¸s [18], serves as the principle tool for the proof of Muhly and Solel's unweighted Nevanlinna-Pick interpolation result, Theorem 5.3 in [8]. Wishing to follow suit, we must first establish that commutant lifting can be done in the weighted case. That is, an operator on a co-invariant subspace for an induced representation of H ∞(E, Z) that commutes with the compression of the image of the representation may be lifted to a commuting operator on the full induced space without increasing the norm. As it turns out, our proof of the weighted com- mutant lifting theorem is where we differ most substantially from the unweighted case. Muhly and Solel's commutant lifting result, Theo- rem 4.4 of [7], is proven using so-called isometric dilations, but in our setting, the weights create obstacles that make this method difficult or even impossible. Instead, we adapt the proof for a commutant lifting theorem given by Arias in [2], an argument that ultimately makes use of Parrott's lemma [11]. Along with the results themselves, this new W ∗-approach to commutant lifting utilizing technology that works in our more general setting is one of the paper's primary attractions. The W ∗-setting of our results provides for a generality that encompasses, for instance, the unweighted commutant lifting and Nevanlinna-Pick theorems of Muhly and Solel in [7] and [8] and the weighted lifting and interpolation theorems of Popescu in [14]. The paper is organized as follows. In Section 2 we establish defini- tions and notation; for convenience, we have organized this material into three subsections. Section 3 is a technical section devoted to the construction of a family of orthonormal bases in preparation for the weighted commutant lifting theorem, Theorem 4.1, the principal result of Section 4. In Section 5 we give a weighted double commutant theo- rem, Theorem 5.3, that extends its unweighted analogue in [8]. Section 6 contains our main result, the weighted Nevanlinna-Pick interpolation theorem, Theorem 6.5. 2. Preliminaries 2.1. W ∗-correspondences and the Unweighted Hardy Algebra. We let N0 = N∪{0}. Hilbert spaces have inner products that are linear in the second variable and conjugate linear in the first. Throughout the paper, M will denote a W ∗-algebra, that is a C ∗-algebra that is 4 JENNIFER R GOOD also a dual space, thought of abstractly, without reference to a partic- ular representation on Hilbert space. Likewise, E will denote a W ∗- correspondence over M in the sense of Section 2 in [8]. That is, E is a self-dual Hilbert C ∗-module over M in the sense of [5] and [12] with a second-linear inner product that is also a left M-module with respect to a faithful normal ∗-homomorphism ϕ : M → L(E) where L(E) de- notes the W ∗-algebra of adjointable operators on E. For simplicity, we assume that ϕ is unital. At certain points, we will add the assumption that the right action of E is full in the sense that the ultraweakly closed linear span of {hξ, ηi ξ, η ∈ E} is all of M. By Proposition 3.8 of [12], E is a dual space, and we refer to the weak-∗ topology on E as its ultraweak topology. If E is nonzero, then E has an orthonormal basis, that is, a family A ⊆ E that is maximal with respect to the following two properties: for every α ∈ A, hα, αi is a nonzero projection in N, and if α, β ∈ A and α 6= β, then hα, βi = 0 ([12], proof of Theorem 3.12). If N is a W ∗-algebra and F is an (M, N) W ∗-correspondence with left action map σ : M → L(F ), then there is an (M, N) W ∗-correspon- dence, denoted E ⊗σ F , formed by taking the self-dual completion of a quotient of the algebraic tensor product of E and F , balanced over N; the quotient is determined by the semi inner product satisfying hξ1 ⊗ η1, ξ2 ⊗ η2i = hη1, σhξ1, ξ2i(η2)i for ξ1, ξ2 ∈ E and η1, η2 ∈ F . To give the left action of M on E⊗σF , first form the induced representation of L(E), σE : L(E) → L(E ⊗σ F ), defined at S ∈ L(E) by σE(S) = S ⊗ IF , as in [15]. Then the left action of M on E ⊗σ F is σE ◦ ϕ, called the induced representation of M, mapping a ∈ M to ϕ(a)⊗IF in L(E ⊗σ F ). We observe that an (M, C) W ∗-correspondence is simply a Hilbert space H together with a normal unital ∗-homomorphism σ : M → B(H); thus, taking F to be H we obtain the Hilbert space E ⊗σ H, called the induced representation space, and the representation σE ◦ ϕ : M → B(E ⊗σ H). If, instead, we inductively take F to be E in the tensor product construction, we form the tensor powers of E, k=0, a family of W ∗-correspondences over M. We write ϕk for {E⊗k}∞ the left action of M on E⊗k for each k ∈ N0. To be precise, E⊗0 := M, E⊗1 := E, and E⊗k := E ⊗ϕk−1 E⊗k−1 for k ≥ 2 where ϕ0 is left multiplication and ϕ1 = ϕ. powers of E; that is F(E) := P∞ The Fock space of E, F(E), is the ultraweak direct sum of the tensor k=0 ⊕E⊗k. The Fock space is a W ∗- correspondence over M with respect to the left action map ϕ∞ : M → L(F(E)), defined at a ∈ M by ϕ∞(a) = diag[ϕ0(a), ϕ1(a), ϕ2(a), . . .]. Along with the class of left action maps, a second important class of Tξ = 0 T (0) ξ 0 0 0 0 T (1) ξ 0 0 0 0 T (2) ξ . . . . . . . . . ,   ξ INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 5 operators in L(F(E)) is the class of creation operators. The creation operator Tξ determined by ξ ∈ E is defined at η ∈ F(E) by Tξ (η) = ξ ⊗ η; matricially, Tξ has a subdiagonal matrix, where for each j ∈ N0, T (j) : E⊗j → E⊗(j+1) is defined at η ∈ E⊗j by T (j) ξ (η) = ξ ⊗ η. For arbitrary k ∈ N0 and ξ ∈ E⊗k, the creation operator Tξ is defined in an analogous fashion. The algebraic tensor algebra, T 0 + (E), is the subalgebra of L(F(E)) generated by the left action and creation operators. The principal object of study in the unweighted setting of [8] is the Hardy algebra, H ∞(E), the closure of + (E) in the ultraweak topology of the W ∗-algebra L(F(E)). T 0 If F is an (M, N) W ∗-correspondence with left action σ and ξ ∈ E, the (left) insertion operator associated with ξ is the map LF ξ ∈ L(F, E ⊗σ F ), defined at η ∈ F by LF ξ (η) = ξ ⊗ η. We omit the superscript F if it is clear from context. We note that a creation operator is a specific example of an insertion operator. The following facts are easily verified for ξ, η ∈ E. The operator Lξ is bounded and kLξk ≤ kξk. For a ∈ M and c ∈ N, La·ξ·c = (ϕ(a) ⊗ IF )Lξσ(c). If S ∈ L(E), T ∈ L(F ), and T is a left M-module homomorphism, LSξT = (S ⊗ T )Lξ. For any ζ ∈ F , L∗ ξ(η ⊗ ζ) = hξ, ηi · ζ; hence, L∗ η = θξ,η ⊗ IF where θξ,η ∈ L(E) denotes the rank one operator defined at ζ ∈ E by θξ,η(ζ) = ξ · hη, ζi. ξLη = σhξ, ηi. Finally, LξL∗ 2.2. Duality. Let σ : M → B(H) be a normal, unital ∗-homomor- phism for a Hilbert space H. If ψ : M → B(K) is another such map for Hilbert space K, I(σ, ψ) denotes the space of intertwiners, i.e. the collection of operators T ∈ B(H, K) such that T ◦ σ(a) = ψ(a) ◦ T for every a ∈ M. With this notation, we define Eσ := I(σ, σE ◦ ϕ), called the σ-dual of E, a subspace of B(H, E ⊗σ H). By Proposition 3.2 of [8], Eσ is a W ∗-correspondence over the W ∗-algebra σ(M)′ with a · ξ · b := (IE ⊗ a)ξb and hξ, ηi := ξ∗η for a, b ∈ σ(M)′ and ξ, η ∈ Eσ. We write ϕ′ : σ(M)′ → L(Eσ) for the left action map. Let us recall several maps that arise in [8], simultaneously providing notation for future use. In this subsection, we will suppose that E is full. For k ∈ N, there is a well-defined σ(M)′ W ∗-correspondence 6 JENNIFER R GOOD isomorphism Λσ k : (Eσ)⊗k → (E⊗k)σ with k ⊗ i=1 ξi ∈ (Eσ)⊗k Λσ k(cid:16) k ⊗ i=1 ξi(cid:17) = (Ik−1 ⊗ ξ1) · · · (I2 ⊗ ξk−2)(I1 ⊗ ξk−1)ξk, k LH 0 : (Eσ)⊗0 → (E⊗0)σ at A ∈ σ(M)′ and h ∈ H by (Λσ where Ij denotes the identity operator on E⊗j. To include k = 0, define Λσ 0 (A))(h) = 1M ⊗ Ah. The inclusion map ι : σ(M)′ → B(H) is a faithful, normal, unital k : (Eσ)⊗k ⊗ι H → ∗-homomorphism, and for every k ∈ N0, the map U σ E⊗k ⊗σ H defined for ξ ∈ (Eσ)⊗k and h ∈ H by U σ k (ξ ⊗ h) = Λσ k(ξ)(h) is a Hilbert space isomorphism identifying the induced representation k=0 ⊕U σ spaces. Equivalently, U σ k identifies the spaces F(Eσ) ⊗ι H and F(E) ⊗σ H ([8]; Lemma 3.8). Let Ad(U σ ∞) : B (F(Eσ) ⊗ι H) → B (F(E) ⊗σ H), denote the isomorphism that sends an operator T ∈ B (F(Eσ) ⊗ι H) to U σ ∞ . It follows that ∞) ◦ ιF(Eσ) is a ρσ : L(F(Eσ)) → B(F(E) ⊗σ H), defined by ρσ = Ad(U σ faithful, normal, unital ∗-homomorphism; ρσ is called "ρ" in Theorem ∞(Y ⊗ IH )U σ∗ 3.9 of [8]. Specifically, for Y ∈ L(F(Eσ)), ρσ(Y ) = U σ ∞ . The map πσ : L(F(E)) → B(F(Eσ) ⊗ι H), defined by πσ = Ad(U σ∗ ∞ ) ◦ σF(E) is also a faithful, normal, unital ∗-homomorphism; πσ is called "ρ" in Section 5 of [8]. For Y ∈ L(F(E)), πσ(Y ) = U σ∗ ∞ := P∞ k(ξ). The map U σ ξ = Λσ ∞T U σ∗ ∞ (Y ⊗ IH)U σ ∞. k, U σ k , U σ To summarize, using the W ∗-algebra M, the W ∗-correspondence over the algebra E, and the representation of the algebra σ, we've defined the maps Λσ ∞, ρσ, and πσ. We repeat the process with the W ∗-algebra σ(M)′, the correspondence Eσ, and the representation ι. The analogue of σ(M)′ is the commutant of the image of ι in B(H), which is precisely σ(M). The analogue of ι is the inclusion map that we denote by  : σ(M) → B(H). We write Eσι in place of (Eσ)ι, a W ∗- ∞ : σ(M) → L(F(Eσι)). correspondence over σ(M) with left action ϕ′′ We obtain the analogous collection of maps Λι ∞, ρι, and πι. Repeating the process once again would utilize the W ∗-algebra σ(M), the correspondence Eσι, and the representation , but if σ is faithful this is unnecessary since these constructs may be naturally identified with M, E, and σ as follows. The identifications of M with σ(M) and σ with  are immediate. After appropriate identifications, the map ω : E → Eσι defined at ξ ∈ E by ω(ξ) = U σ∗ is an isomorphism of W ∗-correspondences that is studied in and around Theorem 3.6 of [8]. Defining ωk := ω ⊗ · · · ⊗ ω identifies E⊗k with (Eσι)⊗k for each k, k=0 ωk identifies F(E) with F(Eσι). For every k ∈ N0, k(ωk ⊗ IH) = IE⊗k⊗σH , as demonstrated in and around Corollary for every ξ ∈ E⊗k. More U σ m ⊗ Λι m for k, m ∈ N0 and m denotes the identity operator on (Eσ)⊗m. It follows U σ 3.10 of [8]. Equivalently U σ generally, U σ k+m (I ′ ξ ∈ E⊗k, where I ′ and ω∞ := P∞ k(ωkξ)) = L(E⊗m⊗σH) k(ωkξ) = LH ξ k ◦ Λι 1 LH ξ k, U ι k, U ι k U ι ξ πσ(Tξ) = 0 ωξ 0 0 0 I ′ 1 ⊗ ωξ 0 0  0 0 0 I ′ 2 ⊗ ωξ . . . . . . . . . .  INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 7 ∞U ι that U σ how πσ acts on the generators of H ∞(E): I ′ ∞ ⊗ σ(a), and for ξ ∈ E, πσ(Tξ) has the subdiagonal matrix, ∞(ω∞ ⊗IH ) = IF(E)⊗σH . Equations (5.1) and (5.2) of [8] show for a ∈ M, πσ(ϕ∞(a)) = 2.3. The Weighted Hardy Algebra. We now turn to the weighted setting of [9]. Throughout, the sequence X = {Xk}∞ k=0 will denote an admissible sequence; this means that Xk ∈ ϕk(M)c for each k ∈ N0, Xk ≥ 0 for each k ∈ N0, X0 = 0, X1 is invertible, and finally that lim supk→∞ kXkk1/k < ∞, where ϕk(M)c denotes the commutant of ϕk(M) in L(E⊗k) ([9], Definition 4.1). Let R = {Rk}∞ k=0 denote the sequence IM , Rk = if k = 0 , if k > 0 , i=1 α(i) = ko if 1 ≤ j ≤ k j=1(cid:16)Pα∈F (k,j) (cid:16)Pk Xα(i)(cid:17)(cid:17)1/2 where F (k, j) := nα : {1, . . . , j} → N(cid:12)(cid:12)Pj j ⊗ i=1 ([9], Equation 4.4). Each Rk is a positive, invertible element in ϕk(M)c. A sequence of operators Z = {Zk}∞ k=0 is called a sequence of weights associated with X if Zk is invertible and belongs to ϕk(M)c for each k ∈ N0, Z0 = IM , and Z (k)∗Z (k) = R−2 for all k ∈ N0, where Z (k), thought of k as the product of the weights {Zk, Zk−1, . . . , Z1, Z0}, is defined Z (k) := Zk(I1 ⊗ Zk−1) · · · (Ik−1 ⊗ Z1)(Ik ⊗ Z0) ∈ ϕk(M)c ([9], Definition 4.6). We define Z (k,j) := Z (k)(Ik−j ⊗ Z (j))−1 when 0 ≤ j ≤ k, equivalently Z (k,j) = Zk(I1 ⊗ Zk−1) · · · (Ik−j−1 ⊗ Zj+1), which may be thought of as the product of the weights {Zk, Zk−1, . . . , Zj+1}. One example of a weight sequence is the canonical sequence Z = {Zk}∞ k=0, with Z0 := IM and Zk := R−1 k (I1 ⊗ Rk−1), k ≥ 1. We note that the unweighted setting studied, for instance, in [7] and [8] occurs when Z is the sequence of identity operators. To work in the setting of [14], we take the W ∗- correspondence E to be Cd and add the condition that X be a sequence of diagonal operators. If Z is a sequence of weights associated with X, then supj∈N0 kZjk < ∞ ([9], proof of Theorem 4.5). It follows that for every k ∈ N0 the operator DZ k = Dk := diag[0, . . . , 0, Z (k), Z (k+1,1), Z (k+2,2), . . .] belongs to L(F(E)). We omit the superscript 'Z' when the weight sequence 8 JENNIFER R GOOD is understood. For ξ ∈ E⊗k, we define the weighted creation operator W Z ξ = Wξ in L(F(E)) by Wξ := Dk · Tξ. The matrix for Wξ is k- subdiagonal; Wξ = Z (k)T (0) ξ 0 ... 0 0 ...  0 ... 0 Z (k+1,1)T (1) ξ 0 0 ... 0 Z (k+2,2)T (2) ξ . . . . . . . . . .  where T (j) ξ maps η ∈ E⊗j to ξ ⊗ η ∈ E⊗j+k for j ∈ N0. We observe that while our arrival at the definition of Wξ differs from that in [9], a comparison of the preceding matrix with equation (3.2) in [9] confirms that the two definitions agree. When k, l ∈ N0; a, b ∈ M; ξ ∈ E⊗k; and η ∈ E⊗l; Wa·ξ·b = ϕ∞(a) ◦ Wξ ◦ ϕ∞(b) and Wξ ◦ Wη = Wξ⊗η. For a ∈ M, Wa = ϕ(a). The Z-algebraic tensor algebra, T 0 + (E, Z), is the subalgebra of L(F(E)) generated by the left action and weighted creation operators. The Z-Hardy algebra, H ∞(E, Z), is the closure of T 0 + (E, Z) in the ultraweak topology of L(F(E)). Let σ : M → B(H) be a normal, unital ∗-homomorphism. We observe that I(σE ◦ ϕ, σ)∗ = I(σ, σE ◦ ϕ) = Eσ, which was defined above. For a point z ∈ I(σE ◦ ϕ, σ) and k ∈ N0, we define the kth (k) := z(I1⊗z) · · · (Ik−1⊗z). The intertwining property tensorial power, z of z guarantees that z(k) is a well-defined operator in B(E⊗k ⊗σ H, H); moreover, z(k) ∈ I(σE⊗k ◦ ϕk, σ). Computation shows that z(k+l) = z(k)(Ik ⊗ z(l)) for k, l ∈ N0. We define D(X, σ) :=(z ∈ I(σE ◦ ϕ, σ)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞Xk=1 (k)(Xk ⊗ IH )z z < 1) ; (k)∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ([9], Definition 4.3). Here and throughout unless stated otherwise, infi- nite sums indicate convergence with respect to the ultraweak topology for a W ∗-correspondence or W ∗-algebra and the norm topology for a Hilbert space. If z ∈ D(X, σ), there is an ultraweakly continuous, com- pletely contractive representation (σ × z) : H ∞(E, Z) → B(H) such that (σ × z)(ϕ∞(a)) = σ(a) for every a ∈ M and (σ × z)(Wξ) = z(k)Lξ for every ξ ∈ E⊗k with k ≥ 1 ([9], Corollary 5.9). For Y ∈ H ∞(E, Z), we define the noncommutative function bY : D(X, σ) → B(H) at the point z ∈ D(X, σ) by bY (z) = (σ × z)(Y ). INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 9 3. An Orthonormal Basis for Tensor Products The proof of our weighted commutant lifting theorem in Section 4 is modeled after the proof of a commutant lifting theorem by Alvaro Arias in [2]. Working in the setting of complex n-space, Arias makes use of the fact that if {ei}n i=1 is an orthonormal basis for Cn and k ∈ N, then one orthonormal basis for (Cn)⊗k consists of the simple tensors k eα(i) such that α is any function from {1, . . . , k} to {1, . . . , n}. In ⊗ i=1 search of a similarly constructed orthonormal basis for E⊗k or, more generally, for any correspondence formed by tensor product, let us con- sider an obstacle that must be overcome. Let F be a nonzero (M, N) W ∗-correspondence with orthonormal basis A, and let G be a nonzero (N, P ) W ∗-correspondence with left action σ and orthonormal basis B for W ∗-algebras M, N, and P . If M = N = P = C, then F and G are simply Hilbert spaces, the W ∗-tensor product of F and G is the same as their Hilbert space tensor product, and {α ⊗ β : α ∈ A, β ∈ B}, which is a Hilbert space orthonormal basis for the tensor product, is also an orthonormal basis in the sense of Hilbert W ∗-modules. For general M, N, and P , the issue is more complicated. If α ∈ A and β ∈ B, there is no reason to think that hα ⊗ β, α ⊗ βi is a nonzero projection in P , in which case α ⊗ β cannot belong to an orthonormal basis for F ⊗σ G. In fact, it is possible that F ⊗σ G is zero, in which case F ⊗σ G does not have an orthonormal basis. We begin this sec- tion by overcoming these obstacles to construct an orthonormal basis for a nonzero tensor product of two W ∗-correspondences that consists of simple tensors. Then we obtain, via an inductive process, a family of orthonormal bases, one basis for each nonzero E⊗k, which we use in Section 4 to prove the weighted commutant lifting theorem using Arias' technique. Along with M, let N and P be W ∗-algebras. Let F be an (M, N) W ∗-correspondence, and let G be an (N, P ) W ∗-correspondence with left action σ : N → L(G). We will write a · η in place of σ(a)(η) for a ∈ N and η ∈ G. If q is a projection in N, then q · G = {q · η η ∈ G} is the kernel of σ(1N − q). Thus q · G is an ultraweakly closed (right) P -submodule of G and is therefore a Hilbert W ∗-module over P ([3]; Consequence 1.8). As such, for every projection q ∈ N such that q · G 6= {0} we may fix an orthonormal basis for q · G. Define Q to be the, possibly empty, subset of F , Q := {ξ ∈ F hξ, ξi is a projection in N and hξ, ξi · G 6= {0}} . If Q is nonempty and ξ ∈ Q, define B(ξ) to be the orthonormal basis for hξ, ξi·G chosen above. Fix A, an orthonormal basis for F . If α ∈ A, 10 JENNIFER R GOOD then hα, αi is a projection in N. The subset of F ⊗σ G, C := {α ⊗ β α ∈ A ∩ Q and β ∈ B(α)}, is empty if and only if A ∩ Q is empty. For simplicity, when we say that "α ⊗ β ∈ C", we mean that α ∈ A ∩ Q and β ∈ B(α). The emphasis is needed since the expression of an element in F ⊗σ G in terms of simple tensors is not unique. If α ⊗ β ∈ C, then hα, αi · β = β. In our first theorem, we show that when F ⊗σ G is nonzero, C is an orthonormal basis. Theorem 3.1. The correspondence F ⊗σ G is nonzero if and only if C is nonempty. In this case, C is an orthonormal basis for F ⊗σ G, and if αi ⊗ βi ∈ C for i = 1, 2, then if α1 = α2 and β1 = β2 otherwise . if A ∩ Q 6= ∅ if A ∩ Q = ∅ . (3.1) (3.2) hα1 ⊗ β1, α2 ⊗ β2i =(hβ2, β2i ξ ⊗ η =(Pα∈A∩Q α ⊗ hα, ξi · η 0 0 Proof. First, we show that for any ξ ∈ F and η ∈ G, By Fourier expansion with respect to A, we have ξ =Pα∈A α · hα, ξi ([12], proof of Theorem 3.12). It follows that ξ ⊗ η =Pα∈A α ⊗ hα, ξi · η. For every α ∈ A, we have hα · hα, αi − α, α · hα, αi − αi = 0, so α · hα, αi = α. Thus hα, ξi · η belongs to hα, αi · G, so if α /∈ Q, then hα, ξi · η = 0. Equation (3.2) follows. When C is empty, A ∩ Q is also empty, so by equation (3.2) every simple tensor in F ⊗σ G is zero. Thus F ⊗σ G = {0} since F ⊗σ G is the ultraweak closure of the linear span of the simple tensors. Suppose that C is not empty. Using the properties of A and C, equation (3.1) follows from straightforward computations. It follows that for any α ⊗ β ∈ C, hα ⊗ β, α ⊗ βi is a nonzero projection in P , so α ⊗ β 6= 0 and F ⊗σ G 6= {0}. It only remains to show that C is an orthonormal basis for F ⊗σ G. By equation (3.1), C is an orthonormal set. Towards maximality, suppose that C ′ is an orthonormal subset of F ⊗σ G such that C ⊆ C ′. For any ξ ∈ F and η ∈ G, if α belongs to A ∩ Q, then as above, hα, ξi · η is an element in hα, αi · G. Using the Fourier expansion with respect to the orthonormal basis B(α), hα, ξi · η =Pβ∈B(α) β · hβ, hα, ξi · ηi =Pβ∈B(α) β · hα ⊗ β, ξ ⊗ ηi. Thus, by equation (3.2), (3.3) ξ ⊗ η = Xα∈A∩Q  Xβ∈B(α) (α ⊗ β) · hα ⊗ β, ξ ⊗ ηi . INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 11 Now if C is properly contained in C ′, then there exists ζ ∈ C ′ such that for every α ⊗ β ∈ C, hζ, α ⊗ βi = 0. Since F ⊗σ G is the ul- traweak closure of the linear span of the simple tensors, there is a net converging ultraweakly to ζ in F ⊗σ G. By equa- tion (3.3), nPNλ j=1 ξλ,j ⊗ ηλ,joλ NλXj=1  Xα∈A∩Q NλXj=1 hζ, ζi = lim λ = lim λ hζ, ξλ,j ⊗ ηλ,ji  Xβ∈B(α) hζ, α ⊗ βihα ⊗ β, ξλ,j ⊗ ηλ,ji = 0, contradicting the fact that hζ, ζi is a nonzero projection. Therefore, C is an orthonormal basis for F ⊗σ G. (cid:3) Let us inductively construct a family {Ck} of orthonormal bases, one basis for each nonzero E⊗k. First we establish some notation, mirroring the discussion prior to Theorem 3.1. Fix an orthonormal basis for the Hilbert W ∗-module q · E for every projection q ∈ M such that q · E 6= {0}. Define Q = {ξ ∈ E hξ, ξi is a projection in M and hξ, ξi · E 6= {0}} . If ξ ∈ Q, let B(ξ) denote the orthonormal basis for hξ, ξi · E chosen above. Since E is nonzero, let A be an orthonormal basis for E. Let C0 = {1M } where 1M is the identity element in M. Let C1 = A. When k ≥ 2, define Ck to be the possibly empty subset of E⊗k consisting of ξi = ξ1⊗· · ·⊗ξk such that the following properties the simple tensors hold: k ⊗ i=1 (1) ξ1 ∈ A ∩ Q, (2) ξi ∈ B(ξi−1) ∩ Q for every i such that 1 < i < k, and (3) ξk ∈ B(ξk−1). For simplicity, when we say that ξi ∈ Ck, we always mean that k ⊗ i=1 i=1 satisfies properties (1), (2), and (3). Notice that if {ξi}k ξi ∈ Ck, then for every i such that 1 < i ≤ k, we have hξi−1, ξi−1i · ξi = ξi. We arrive at the following theorem. k ⊗ i=1 Theorem 3.2. For k ∈ N0, the correspondence E⊗k is nonzero if and only if Ck is nonempty. In this case, Ck is an orthonormal basis for 12 JENNIFER R GOOD k ⊗ i=1 E⊗k, and for any ξ = (3.4) hξ, ηi =(hηk, ηki 0 ξi and η = k ⊗ i=1 ηi in Ck, if ξi = ηi for every i such that 1 ≤ i ≤ k otherwise . ξi(cid:17) ⊗ ξk+1(cid:12)(cid:12)(cid:12)(cid:12) Proof. The cases when k = 0 and k = 1 follow readily. Inductively, suppose the conclusions of the theorem are satisfied for some k ∈ N. Towards applying Theorem 3.1 to E⊗k ⊗ϕ E, define If ξ ∈ Qk, let B(ξ) denote the orthonormal basis for hξ, ξi · E. Define Qk :=(cid:8)ξ ∈ E⊗k hξ, ξi is a projection in M and hξ, ξi · E 6= {0}(cid:9) . ξi(cid:17)(cid:27) . k+1 :=(cid:26)(cid:16) k ξi ∈ Ck ∩ Qk and ξk+1 ∈ B(cid:16) k k ⊗ i=1 ⊗ i=1 ⊗ i=1 C ′ Using the inductive assumption we see that Ck+1 is the image of C ′ k+1 under the natural isomorphism between E⊗k ⊗ϕ E and E⊗k+1. It now follows from Theorem 3.1 that E⊗k+1 is nonzero if and only if Ck+1 is nonempty and that Ck+1 is an orthonormal basis in that case. If ξ = ηi belong to Ck+1, then (cid:16) k (cid:16) k ηi(cid:17) ⊗ ηk+1 belong to C ′ k+1 and(cid:16) k ξi(cid:17) and(cid:16) k Since hηk, ηki · ηk+1 = ηk+1 and hξ, ηi = Dξk+1,D k ξi(cid:17) ⊗ ξk+1 and ηi(cid:17) belong to Ck. ηiE · ηk+1E, equation (3.4) follows from Theorem 3.1 and the inductive hypothesis. (cid:3) ξi and η = k ⊗ i=1 ⊗ i=1 k+1 ⊗ i=1 k+1 ⊗ i=1 ⊗ i=1 ⊗ i=1 ⊗ i=1 ⊗ i=1 ξi, We devote the remainder of the section to establishing notation and giving technical lemmas that will be useful in proving our main theorems. Define S := (cid:8)k ∈ N0 E⊗k 6= {0}(cid:9), which is either N0 or {0, 1, . . . , n} for some n ∈ N. For k ∈ N0, define Ak to be Ck when k ∈ S and the zero set when k /∈ S. If k ∈ N0 then vk ∈ L(E⊗k, F(E)) denotes the isometry mapping ξ ∈ E⊗k to bξ := (δi=k ξ)∞ i=0 where δi=k is 1 when i = k and is otherwise zero. A handy fact is that a diagonal operator S = diag[S0, S1, S2, · · · ] ∈ L(F(E)) may be ex- pressed as S =P∞ j . The insertion operator Lbξ maps x ∈ H to bξ ⊗ x ∈ F(E) ⊗σ H. We define Qk in L(F(E)) to be the projection onto vk(E⊗k), an ultraweakly closed submodule of F(E), noting that Qk = vkv∗ k. For ξ ∈ E⊗k, θξ ∈ L(E⊗k) denotes the positive rank one operator that sends η ∈ E⊗k to ξ · hξ, ηi. j=0 vjSjv∗ Lemma 3.3. For k ∈ N0, (1) if S ∈ L(E⊗k, F ) for a Hilbert W ∗ M-module F , then SS∗ = θSξ in L(F ); Pξ∈Ak INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 13 (2) Qk ⊗ IH =Pξ∈Ak (3) Pξ∈Ak LbξL∗ bξ closed submodule {(ζi)∞ i=0 ζi = 0 when 0 ≤ i < k} of F(E). ξ is the projection in L(F(E)) onto the ultraweakly in B(F(E) ⊗σ H); and TξT ∗ Proof. If k /∈ S, then E⊗k = {0}, and all of the operators in parts (1), (2), and (3) are zero. Suppose k ∈ S. For η ∈ E⊗k and F a finite subset of Ak, θξ! η, η+ =Xξ∈F 0 ≤* Xξ∈F hη, ξihξ, ηi ≤ Xξ∈Ak fore,Pξ∈Ak algebra L(E⊗k) and for any η ∈ E⊗k, D(cid:16)Pξ∈Ak Thus by polarization, by a Parseval type identity ([12], proof of Theorem 3.12). There- θξ converges ultraweakly to a positive operator in the W ∗- θξ(cid:17) η, ηE = hη, ηi. hη, ξihξ, ηi = hη, ηi Ik = Xξ∈Ak θξ. (3.5) (3.6) (3.6), P∞ Xξ∈Ak The map Ad(S) : L(E⊗k) → L(F ) that sends Y ∈ L(E⊗k) to SY S∗ is linear and ultraweakly continuous, and Ad(S)(θξ) = θSξ for any ξ ∈ Ak. Therefore, part (1) follows from applying Ad(S) to equation (3.5). Taking S to be vk in part (1), we obtainPξ∈Ak Since σF(E) is linear and ultraweakly continuous, Qk ⊗IH =Pξ∈Ak IH =Pξ∈Ak θbξ = Qk in L(F(E)). θbξ ⊗ i−k is linear and ultraweakly continuous, so it follows from equation (3.5) that Let i ∈ N0 such that i ≥ k. The induced representation ϕE⊗k LbξL∗ , giving part (2). bξ Ii = Xξ∈Ak θξ ⊗ Ii−k . For any ξ ∈ Ak, we have TξT ∗ i=k vi(θξ ⊗ Ii−k)v∗ ξ = diag[0, . . . , 0, θξ, θξ ⊗ I1, θξ ⊗ I2, . . .] = i . Summing over all ξ ∈ Ak and using equation TξT ∗ ξ = Xξ∈Ak ∞Xi=k vi(θξ ⊗ Ii−k)v∗ i! vi Xξ∈Ak = ∞Xi=k θξ ⊗ Ii−k! v∗ i = Qi. ∞Xi=k 14 JENNIFER R GOOD noting that sums may be interchanged as all terms are non-negative. i=k Qi is the projection in L(F(E)) onto the (cid:3) It is readily shown thatP∞ ultraweakly closed submodule {(ζi)∞ i=0 ζi = 0 when 0 ≤ i < k}. The following technical lemma is similar in flavor to Lemma 5.2 of [9]. Lemma 3.4. If Z is a sequence of weights associated with X, WX 1/2 ∞Xj=1 ξ = IF(E) − Q0 Proof. First we note that if i ∈ N then Pi equation (4.7) of [9], so ξW ∗ X 1/2 j j in L(F(E)). j=1 Xj ⊗ R2 i−j = R2 i by (3.7) Z (i)(Xj ⊗ R2 i−j)Z (i)∗ = Z (i)R2 i Z (i)∗ = Ii. Also when j ∈ N,Pξ∈Aj (3.8) θX 1/2 j ξ = Xj by Lemma 3.3. Thus if i ≥ j ≥ 1, Z (i)(θX 1/2 ξ ⊗ R2 i−j)Z (i)∗ = Z (i)(Xj ⊗ R2 i−j)Z (i)∗. j For j ∈ N and ξ ∈ Aj, computation shows that WX 1/2 a diagonal matrix whose ith diagonal entry is Z (i)(θX 1/2 when i ≥ j and is otherwise zero. Thus by equation (3.8), j j has ξW ∗ ξ ⊗ R2 j ξ X 1/2 i−j)Z (i)∗ Xξ∈Aj iXj=1 Xξ∈Aj Xξ∈Aj WX 1/2 j ξW ∗ X 1/2 ξ j viZ (i)(θX 1/2 ξ ⊗ R2 j Z (i)(θX 1/2 j ξ ⊗ R2 i−j)Z (i)∗v∗ i! i−j)Z (i)∗ v∗ i viZ (i)(Xj ⊗ R2 i−j)Z (i)∗v∗ i . = = Xξ∈Aj ∞Xi=j viXξ∈Aj ∞Xi=j ∞Xi=j ξ = = ∞Xj=1 ∞Xi=j i−j)Z (i)∗! v∗ Z (i)(Xj ⊗ R2 ξW ∗ X 1/2 j viZ (i)(Xj ⊗ R2 i−j)Z (i)∗v∗ i = viv∗ i = IF(E) − Q0, ∞Xi=1 i! ∞Xj=1 j WX 1/2 Xξ∈Aj vi iXj=1 ∞Xi=1 = Summing over j and using equation (3.7), INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS which completes the proof. 15 (cid:3) To "factor" an element ξ ∈ Ak when k ∈ N we establish the following notation. If k ∈ S, ξ = ξj := j ⊗ i=1 ξi ∈ E⊗j k ⊗ i=1 ξi ∈ Ak, and 1 ≤ j ≤ k define and ξj ◦ :=( k ξi i=j+1 hξ, ξi ⊗ if j < k if j = k ∈ E⊗k−j. If k /∈ S, let ξj and ξj ◦ to be the zero elements in their respective spaces. Lemma 3.5. (1) For k ∈ N0 and ξ ∈ Ak, hξ, ξi is a projection in M that is zero if and only if k /∈ S. In any case, ξ · hξ, ξi = ξ. (2) For j ∈ N, l ∈ N0, and ξ ∈ Al+j, we have ξ = ξj ⊗ ξj ◦ and hξj, ξji · ξj ◦ = ξj ◦. If l + j ∈ S, then ξj ∈ Aj. (3) If j ∈ N, l ∈ N0, and l + j ∈ S, then as a disjoint union, (4) For j ∈ N and ξ ∈ Aj, Al+j = aξ∈Aj {η ∈ Al+j ηj = ξ}. ϕ∞hξ, ξi = ∞Xl=0  Xη∈Al+j ηj =ξ  , θcηj ◦ where any sum taken over an empty set is assumed to be zero. vl(ϕlhξ, ξi)v∗ Proof. The proofs of the first three assertions are straightforward. To show part (4), let l ∈ N0. If η ∈ Al+j, then T (l)∗ ◦ which, by considering separately the cases when l + j /∈ S and l + j ∈ S, we find to be equal to ηj ◦ when ξ = ηj and zero when ξ 6= ηj. Since ϕlhξ, ξi = T (l)∗ , it now follows from Lemma 3.3(1) that (η) = hξ, ηji · ηj ξ ξ T (l) ξ l = vl Xη∈Al+j θT (l)∗ ξ l = vl Xη∈Al+j η v∗ ηj =ξ θηj ◦ v∗ l = Xη∈Al+j ηj =ξ . θcηj ◦ Since ϕ∞hξ, ξi = diag[ϕ0hξ, ξi, ϕ1hξ, ξi, ϕ2hξ, ξi, . . .], we obtain part (4) by summing over l ∈ N0. (cid:3) For the final technical lemma of this section, we decompose certain induced representation spaces in terms of the family {Ak}∞ k=0. Suppose that H is a Hilbert space and σ : M → B(H) is a faithful, normal, unital ∗-homomorphism. If p is a projection in M, then pM is a Hilbert 16 JENNIFER R GOOD W ∗-module over M. Also, σ(p)(H) is a closed subspace of H, and it is readily shown that pM ⊗σ H and σ(p)(H) are isomorphic Hilbert spaces under the identification of a ⊗ x with σ(a)(x) for a ∈ pM and x ∈ H. If k ∈ N0 and ξ ∈ Ak, then hξ, ξi is a projection in M; we let Hξ ⊕Hξ. denote the Hilbert space σhξ, ξi(H), and we define Hk :=Pξ∈Ak Note that if k /∈ S, then Hk = {0}. Lemma 3.6. (1) For k ∈ N0, there is a Hilbert space isomorphism γk : E⊗k ⊗σ H → Hk such that for η ∈ E⊗k, h ∈ H, and h = (hξ)ξ∈Ak ∈ Hk, γk(η ⊗ h) =(cid:16) σhξ, ηi(h)(cid:17)ξ∈Ak and γ∗ ξ ⊗ hξ. k(cid:0)h(cid:1) = Xξ∈Ak (2) Matricially γk(Y ⊗ IH)γ∗ k = [σhξ, Y ηi]ξ,η∈Ak for Y ∈ L(E⊗k), k ∈ N0. P∞ (3) There is an isomorphism of Hilbert spaces Γ : F(E) ⊗σ H → k=0 in F(E), h ∈ H, and k=0 ⊕Hk such that for η = (ηk)∞ k=0 ∈P∞ k=0 h = ( (hk,ξ)ξ∈Ak )∞ k=0 ⊕Hk, Proof. When k /∈ S, we define γk to be the zero map. When k ∈ S, Ak Γ(η ⊗ h) =(cid:16) γk(ηk ⊗ h)(cid:17)∞ is an orthonormal basis for E⊗k, so E⊗k and(cid:16)Pξ∈Ak and Γ∗(cid:0)h(cid:1) = ∞Xk=0 Xξ∈Akbξ ⊗ hk,ξ! . ⊕ hξ, ξiM(cid:17) are isomorphic Hilbert W ∗-modules ([12], proof of Theorem 3.12). Using the discussion preceding the lemma, parts (1) and (3) readily follow. Part (2) follows from routine computation. (cid:3) 4. Weighted Commutant Lifting Commutant lifting is the principle tool used to obtain the Nevan- linna-Pick interpolation result, Theorem 5.3, in [8]. Towards obtaining our weighted interpolation theorem in Section 6, the purpose of the present section is to show that commutant lifting can be done in the weighted case. Since the weights present obstacles for the method producing the (unweighted) commutant lifting theorem, Theorem 4.4 in [7], our approach is instead inspired by Arias' proof of Theorem 3.1 in [2], a lifting result in complex n-space that ultimately makes use of Parrott's lemma, Theorem 1 of [11]. Theorem 4.1 (Weighted Commutant Lifting). Let σ : M → B(H) be a faithful, normal, unital, ∗-homomorphism for a Hilbert space H, and let Z be a sequence of weights associated with X. Suppose that INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 17 J is a closed linear subspace of F(E) ⊗σ H such that for every Y ∈ H ∞(E, Z), (Y ∗ ⊗ IH )(J) ⊆ J. With V the inclusion map of J into F(E) ⊗σ H, suppose there exists G ∈ B(J) such that for every Y ∈ H ∞(E, Z), G (V ∗(Y ⊗ IH)V ) = (V ∗(Y ⊗ IH)V ) G. Then there exists eG ∈ B(F(E) ⊗σ H) such that (1) eG∗(J) ⊆ J, (2) V ∗eGV = G, (3) eG(Y ⊗ IH) = (Y ⊗ IH )eG for all Y ∈ H ∞(E, Z), and (4) keGk = kGk. j=0 ⊕E⊗j(cid:17) ⊗σ H P . For n ∈ N0, let Kn be the isomorphic image of(cid:16)Pn Our construction of eG ∈ B(K) will utilize an ascending sequence of Proof. First, we consider the case when kGk = 1. Let K = F(E) ⊗σ H, and let P ∈ B(K) be the projection map onto J. Note that V ∗ is the range restriction of P to J, V ∗V is the identity map on J, and V V ∗ = subspaces {Ji} of K and operators {Gi} such that each Gi ∈ B(K, Ji). We construct these sequences inductively. To begin the process, let let J1 = J, and define G1 := GV ∗. To organize the n1 = −1, numerous properties we wish to maintain at each step of the induc- If m ∈ N, we say tive argument, we make the following definition. ({ni}m i=1) is an "m-triple" if for each i with 1 ≤ i ≤ m we have that ni ∈ Z, Ji is a closed subspace of K, Gi ∈ B(K, Ji), and the following conditions hold in all cases: with Vi the inclusion map of Ji into K, in K. For technical purposes, let K−1 be the zero subspace of K. i=1, {Gi}m i=1, {Ji}m (Y ∗ ⊗ IH)(Ji) ⊆ Ji, 1 ≤ i ≤ m, Y ∈ i (Y ⊗ IH)Vi)Gi = Gi(Y ⊗ IH ), 1 ≤ 1 ≤ i ≤ j ≤ m; 1 ≤ i ≤ m. n1 < n2 < · · · < nm, J1 ( J2 ( · · · ( Jm, (if m ≥ 2); (if m ≥ 2); 1 ≤ i ≤ m; Condition (1, m): Condition (2, m): Condition (3, m): Kni ⊆ Ji, Condition (4, m): H ∞(E, Z); Condition (5, m): i ≤ m, Y ∈ H ∞(E, Z); V ∗ Condition (6, m): i VjGj = Gi, kGik = 1, Condition (7, m): Our hypotheses guarantee that ({ni}1 (V ∗ i=1, {Ji}1 i=1) is a 1-triple. i=1) with m ∈ N such that Jm = K, then to satisfy the conclusion of Let us show that if there exists an m-triple ({ni}m i=1, {Gi}m i=1, {Ji}m our theorem, we may take eG to be Gm. Here, Gm ∈ B(K), Vm is the identity map on K, and V1 = V . By Condition (6, m), V ∗Gm = 1 VmGm = G1 = GV ∗. Thus P Gm = V V ∗Gm = V GV ∗ and P GmP = V ∗ i=1, {Gi}1 18 JENNIFER R GOOD i=1, {Ji}m i=1, {Gi}m i=1, {Ji}m i=1, {Gi}m (V GV ∗)P = V GV ∗ = P Gm. It follows that Gm satisfies property (1) in the conclusion of the theorem. Using again the fact that V ∗Gm = GV ∗, we have V ∗GmV = GV ∗V = G, so property (2) holds. Condition (5, m) gives us property (3), and property (4) follows from Condition (7, m). We are left to consider the case when there is no m ∈ N such i=1). Proving if m ∈ N and i=1) is an m-triple such that Jm ( K, then there that Jm = K for some m-triple ({ni}m the following fact will occupy us for quite some time: ({ni}m exist nm+1, Jm+1, and Gm+1 such that(cid:0){ni}m+1 ofS∞ We begin by defining nm+1 and Jm+1. Since K is the norm-closure j=0 Kj and Jm 6= K, the set {j ∈ N0 Kj * Jm} is nonempty, so we define nm+1 to be its least element. Define Jm+1 to be the closed subspace of K that is generated by Jm ∪Knm+1. It is readily shown that the first three conditions for an (m + 1)-triple are satisfied. To show Condition (4, m + 1), we first prove that for any j ∈ N and ξ ∈ E⊗j, i=1 (cid:1) is an (m + 1)-triple. i=1 , {Ji}m+1 i=1 , {Gi}m+1 (4.1) (4.2) Pm+1(Wξ ⊗ IH)Pm = Pm+1(Wξ ⊗ IH) = Pm+1(Wξ ⊗ IH)Pm+1, ξ ⊗IH )(Knm+1) ⊆ K(nm+1)−1 ⊆ Jm, so (W ∗ where Pm and Pm+1 are the projection maps in B(K) onto Jm and Jm+1, respectively. First, let j = 1 and ξ ∈ E. By Condition (4, m), (W ∗ ξ ⊗ IH)(Jm) ⊆ Jm. At the same time, from the definition of Wξ it follows that for any n ∈ N0, (W ∗ ξ ⊗IH )(Kn) ⊆ Kn−1. Therefore, by definition of nm+1, (W ∗ ξ ⊗IH )(Jm+1) ⊆ Jm. Thus, equation (4.1) holds when j = 1. Since PmPm+1 = Pm and E⊗k+1 is generated by elements of the form ξ = η ⊗ ζ where η ∈ E and ζ ∈ E⊗k, an inductive argument gives equation (4.1) in the general case. Equation (4.2) follows since PmPm+1 = Pm. For a ∈ M, it is readily seen that ϕ∞(a∗) ⊗ IH leaves Kn invariant for any n ∈ N and leaves Jm invariant by Condition (4, m). It follows that for Y = ϕ∞(a), (4.3) (Y ∗ ⊗ IH)(Jm+1) ⊆ Jm+1. By equation (4.2), containment (4.3) is also satisfied for Y = Wξ when ξ ∈ E⊗j for some j ∈ N. It follows that containment (4.3) holds for every Y ∈ H ∞(E, Z). This, together with Condition (4, m), implies Condition (4, m + 1). Before considering the remaining conditions we must define Gm+1, which will require some preliminary work. Let us begin by recalling and establishing notation. Let {Ak}∞ k=0 be the family defined in the paragraph before Lemma 3.3. For k ∈ N0 and ξ ∈ E⊗k, bξ denotes i=0 ∈ F(E). For ξ ∈ F(E), Lξ is the insertion operator vk(ξ) = (δi=k ξ)∞ INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 19 m is the range restriction of Pm, V ∗ ∈ B(H, Jm). Note mapping h ∈ H to ξ ⊗ h. Define g := GmLd1M that V ∗ mVm is the identity on Jm, and VmV ∗ m = Pm, with analogous properties for Vm+1. For notational convenience, define α : H ∞(E, Z) → B(Jm) and β : H ∞(E, Z) → B(Jm, Jm+1) at Y ∈ H ∞(E, Z) by, α(Y ) = V ∗ m(Y ⊗ IH)Vm and β(Y ) = V ∗ m+1(Y ⊗ IH)Vm. The following technical lemma collects facts about α and β that will be useful in future computations. Lemma 4.2. (1) α is contractive, linear, multiplicative, unital, ultraweakly con- tinuous, and for any Y ∈ H ∞(E, Z), α(Y )Gm = Gm(Y ⊗ IH). (2) β is contractive, linear, continuous with respect to the ultra- weak topology on H ∞(E, Z) and the weak operator topology on B(Jm, Jm+1), and for any k ∈ N, ξ ∈ E⊗k, and Y ∈ H ∞(E, Z), β(Wξ)α(Y ) = β(WξY ). (3) For any a ∈ M, k ∈ N0, and ξ ∈ E⊗k, β(Wξ)∗V ∗ m+1(ϕ∞(a) ⊗ IH)Vm+1 = β(Wa∗·ξ)∗. (4) For all k ∈ N0 and ξ ∈ E⊗k, (5) For all a ∈ M, k ∈ N0, and ξ ∈ E⊗k, α(Wξ·a)g = α(Wξ)gσ(a). (6) For all a ∈ M, k ∈ N0, and ξ ∈ E⊗k, g∗β(Wξ·a)∗ = α(W(Z (k))−1ξ)g = GmLbξ, σ(a∗)g∗β(Wξ)∗. (7) For any j ∈ N0, η ∈ E⊗j, and T ∈ L(E⊗j), α(WT η)g = Xξ∈Aj α(WT ξhξ,ηi)g, with convergence in the weak operator topology on B(H, Jm). (8) For any j ∈ N0, η ∈ E⊗j, and T ∈ L(E⊗j), g∗β(WT η)∗ = Xξ∈Aj g∗β(WT ξhξ,ηi)∗, with convergence in the weak operator topology on B(Jm+1, H). Proof of Lemma 4.2. (1): The induced representation σF(E) is a normal, unital ∗-homomor- phism. We form α by composing the restriction of σF(E) to H ∞(E, Z) with Ad(V ∗ mT Vm. It follows that α is linear, unital, contractive, and ultraweakly continuous. By Condition (4, m), for Y ∈ H ∞(E, Z), V ∗ m(Y ⊗ IH)Pm. m) : B(K) → B(Jm), which sends T to V ∗ m(Y ⊗ IH) = V ∗ 20 JENNIFER R GOOD It follows that α is multiplicative. By Condition (5, m), for any Y ∈ H ∞(E, Z), α(Y )Gm = Gm(Y ⊗ IH). (2): The linearity, contractivity, and continuity of β follow from arguments similar to those used above for α. By equation (4.1), if k ∈ N, ξ ∈ E⊗k, and Y ∈ H ∞(E, Z), then β(Wξ)α(Y ) = V ∗ m+1(Wξ ⊗ IH)Pm(Y ⊗ IH)Vm = V ∗ m+1(Wξ ⊗ IH )(Y ⊗ IH )Vm = β(WξY ). (3): Since Condition (4, m + 1) holds and ϕ∞(a∗)Wξ = Wa∗·ξ, m+1(ϕ∞(a∗)⊗IH )Vm+1β(Wξ) = V ∗ V ∗ m+1(ϕ∞(a∗)⊗IH )Pm+1(Wξ⊗IH )Vm = V ∗ m+1(ϕ∞(a∗) ⊗ IH )(Wξ ⊗ IH)Vm = β(Wa∗·ξ). The result follows from taking the adjoint. (4): By part (1), α(W(Z (k))−1ξ)g = Gm(W(Z (k))−1ξ ⊗ IH)Ld1M (5): Since ϕ(a) = Wa and Z (0) = IM , by part (4) α(ϕ∞(a))g = α(Wa)g = GmLba = GmLd1M (4.4) σ(a) = gσ(a). = GmLbξ. Since α is multiplicative, α(Wξ·a)g = α(Wξ)α(ϕ∞(a))g = α(Wξ)gσ(a). (6): By Condition (4, m) and equation (4.4), β(Wξ·a)g = V ∗ m+1(Wξ ⊗ IH)(ϕ∞(a) ⊗ IH )Vmg = V ∗ m+1(Wξ ⊗ IH)Pm(ϕ∞(a) ⊗ IH)Vmg = β(Wξ)α(ϕ∞(a))g = β(Wξ)gσ(a). The result follows by taking the adjoint. (7): We have η = Pξ∈Aj ξhξ, ηi since Aj is either an orthonormal basis or the zero set. As an element of L(E⊗j), T is an ultraweakly continuous, right M-module homomorphism; also the map that sends ζ ∈ E⊗j to Wζ ∈ L(F(E)) is linear and ultraweakly continuous, as is α. The result follows. (8): The result follows along similar lines as part (7). (cid:4) One fact we will require in defining Gm+1 is that (4.5) G∗ m(x) = x ∈ Jm. x! , ∞Xk=0 Xξ∈Akbξ ⊗ g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ (cid:16)(cid:0) g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x(cid:1)ξ∈Ak (cid:17)∞ k=0 To prove this fact, we will show that (4.6) k=0 ⊕Hk, using the notation of Lemma 3.6. Let k ∈ N0 and ξ ∈ Ak. By Lemma 3.5(1), ξ · hξ, ξi = ξ, so by Lemma is an element in P∞ INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 21 By Lemma 4.2(4) and Lemma 3.3(2), 4.2(5), we have g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x = σhξ, ξig∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x ∈ Hξ. (4.7) Xξ∈Ak(cid:13)(cid:13)g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x(cid:13)(cid:13)2 = Xξ∈AkDLbξL∗ bξG∗ = Xξ∈Ak(cid:13)(cid:13)(cid:13)L∗ mx(cid:13)(cid:13)(cid:13) bξG∗ mxE mx, G∗ 2 = h(Qk ⊗ IH)G∗ mx, G∗ mxi ≤ kG∗ mxk2. 2 = mx, G∗ mxi = kG∗ mxk2. h(Qk ⊗ IH)G∗ is otherwise zero. By Lemma 4.2 parts (5), (7), and (4), operator convergence in B(K), so by computation (4.7) we find, is a well- k=0(Qk ⊗ IH) = IK with strong k=0 ⊕Hk. By Lemma 3.6(3), its image under Γ∗ in K is the element in the right-hand side of equation (4.5). To show equality with G∗ m(x), let h ∈ H, j ∈ N0, and We conclude that for every k ∈ N0,(cid:0) g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x(cid:1)ξ∈Ak defined element of Hk. Moreover, P∞ x(cid:1)ξ∈Ak(cid:13)(cid:13)(cid:13) ∞Xk=0(cid:13)(cid:13)(cid:13)(cid:0) g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ ∞Xk=0 Thus, the element in expression (4.6) belongs toP∞ η ∈ E⊗j. For k ∈ N0 and ξ ∈ E⊗k,Dbξ,bηE equals hξ, ηi when k = j and * ∞Xk=0 Xξ∈Akbξ ⊗ g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x! ,bη ⊗ h+ = Xξ∈Aj(cid:10)x, α(cid:0)W(Z (j))−1ξ(cid:1) gσhξ, ηih(cid:11) = Xξ∈AjDx, α(cid:0)W(Z (j))−1ξhξ,ηi(cid:1) g(h)E mx,bη ⊗ hi. β(cid:0)W(Z (k))−1η(cid:1) gg∗β(cid:0)W(Z (k))−1η(cid:1)∗! ≤ IJm+1. =Dx, α(cid:0)W(Z (j))−1η(cid:1) g(h)E = hx, GmLbη(h)i = hG∗ Equation (4.5) follows from routine approximation arguments. Another fact we need before constructing Gm+1 is that ∞Xk=1 Xη∈Ak (4.8) 22 JENNIFER R GOOD For convenience, we break the proof into three main steps, as follows: (4.9) IJm+1 ≥ = (4.10) (4.11) = j j Xξ∈Aj β(cid:16)WX 1/2 ∞Xj=1 ∞Xk=1 kXj=1 Xη∈Ak ∞Xk=1 Xη∈Ak ξ(cid:17) GmG∗ β(cid:18)W(cid:16)X 1/2 ·g∗β(cid:18)W(cid:16)X 1/2 ξ(cid:17)∗ mβ(cid:16)WX 1/2 j ⊗(Z (k−j))−1(cid:17)η(cid:19) g j ⊗(Z (k−j))−1(cid:17)η(cid:19)∗(cid:19)(cid:19) β(cid:0)W(Z (k))−1η(cid:1) gg∗β(cid:0)W(Z (k))−1η(cid:1)∗! . We begin with inequality (4.9). ξW ∗ for any j ∈ N, Pξ∈Aj WX 1/2 j j X 1/2 a finite subset of Aj, then because σF(E) is a ∗-homomorphism and kVmGmk ≤ 1, It follows from Lemma 3.4 that If F is converges in L(F(E)). ξ j ξ(cid:17)∗ mβ(cid:16)WX 1/2 ξ ⊗ IH(cid:17) VmGmG∗ j j j V ∗ ≤ V ∗ Xξ∈F ξ(cid:17) GmG∗ β(cid:16)WX 1/2 m+1(cid:16)WX 1/2 =Xξ∈F m+1 Xξ∈F m+1Xξ∈Aj It follows that Pξ∈Aj ≤ V ∗ weakly in B(Jm+1) and V ∗ j j ξ X 1/2 X 1/2 mV ∗ ξW ∗ WX 1/2 WX 1/2 ⊗ IH(cid:19) Vm+1 m(cid:18)W ∗ ξ! ⊗ IH! Vm+1 ξ ⊗ IH Vm+1. ξ(cid:17)∗ mβ(cid:16)WX 1/2 ξ(cid:17) GmG∗ β(cid:16)WX 1/2 m+1(cid:18)(cid:18)Pξ∈Aj ξ(cid:19) ⊗ IH(cid:19) Vm+1 is converges ultra- WX 1/2 ξW ∗ ξW ∗ X 1/2 X 1/2 j j j j j j an upper bound. This, together with Lemma 3.4, implies that for INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 23 N ∈ N, NXj=1 Xξ∈Aj j j ξ(cid:17) GmG∗ mβ(cid:16)WX 1/2 β(cid:16)WX 1/2 m+1 Xξ∈Aj NXj=1 WX 1/2 j ξW ∗ ≤ V ∗ ξ(cid:17)∗ ξ ⊗ IH Vm+1 ≤ IJm+1. j X 1/2 Inequality (4.9) follows. Towards proving equation (4.10), fix j ∈ N and ξ ∈ Aj. Since ξ · hξ, ξi = ξ, it follows from Lemma 4.2 parts (1) and (2) that (4.12) β(cid:16)WX 1/2 j ξ(cid:17) Gm = β(cid:16)WX 1/2 j ξ(cid:17) Gm (ϕ∞hξ, ξi ⊗ IH) . Also, by Lemma 4.2(4), whenever l ∈ N0 and ζ ∈ E⊗l, Gm(cid:16)θbζ ⊗ IH(cid:17) G∗ m = GmLbζL∗ Therefore, by Lemma 3.5(4), (4.13) Gm (ϕ∞hξ, ξi ⊗ IH ) G∗ m = α(cid:0)W(Z (l))−1ζ(cid:1) gg∗α(cid:0)W(Z (l))−1ζ(cid:1)∗ . bζG∗ m  Xη∈Al+j Gm(cid:16)θbηj ⊗ IH(cid:17) G∗ ∞Xl=0 0(cid:17)∗ . 0(cid:17) gg∗α(cid:16)W(Z (l))−1ηj α(cid:16)W(Z (l))−1ηj ηj =ξ m = 0 = ∞Xl=0  Xη∈Al+j ηj =ξ If l ∈ N0 and η ∈ Al+j with ηj = ξ, by Lemma 4.2(2) and Lemma 3.5(2), (4.14) β(cid:16)WX 1/2 j ξ(cid:17) α(cid:16)W(Z (l))−1ηj 0(cid:17) g = β(cid:18)W(cid:16)X 1/2 j ⊗(Z (l))−1(cid:17)η(cid:19) g. 24 JENNIFER R GOOD Putting together equations (4.12), (4.13), and (4.14), j j j X 1/2 X 1/2 ξ(cid:19)∗ mβ(cid:18)W mβ(cid:18)W ξ(cid:19) Gm (ϕ∞hξ, ξi ⊗ IH) G∗ ξ(cid:19)∗ ξ(cid:19) α(cid:16)W(Z (l))−1ηj β(cid:18)W 0(cid:17) gg∗α(cid:16)W(Z (l))−1ηj 0(cid:17)∗ j ⊗(Z (l))−1(cid:17)η(cid:19)∗ j ⊗(Z (l))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 β(cid:18)W(cid:16)X 1/2 j j X 1/2 . ξ(cid:19)∗ j X 1/2 ξ(cid:19)∗ mβ(cid:18)W β(cid:18)W(cid:16)X 1/2 j ⊗(Z (l))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 j ⊗(Z (l))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 j ⊗(Z (l))−1(cid:17)η(cid:19)∗  j ⊗(Z (l))−1(cid:17)η(cid:19)∗ . Summing over ξ ∈ Aj, interchanging sums, and using Lemma 3.5(3), j j = = ηj =ξ ηj =ξ X 1/2 X 1/2 X 1/2 X 1/2 ∞Xl=0 β(cid:18)W ξ(cid:19) GmG∗ = β(cid:18)W  Xη∈Al+j β(cid:18)W  Xη∈Al+j ∞Xl=0 β(cid:18)W ξ(cid:19) GmG∗ Xξ∈Aj   Xη∈Al+j ∞Xl=0 = Xξ∈Aj  Xη∈Al+j β(cid:18)W(cid:16)X 1/2 ∞Xl=0 Xξ∈Aj β(cid:18)W ∞Xj=1  Xη∈Al+j  ∞Xj=1 ∞Xl=0  Xη∈Ak kXj=1 ∞Xk=1 X 1/2 ηj =ξ = = = j Finally, summing over j ∈ N, re-indexing, and interchanging sums, j X 1/2 ξ(cid:19)∗ ξ(cid:19) GmG∗ β(cid:18)W(cid:16)X 1/2 β(cid:18)W(cid:16)X 1/2 mβ(cid:18)W j ⊗(Z (l))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 j ⊗(Z (k−j))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 j ⊗(Z (l))−1(cid:17)η(cid:19)∗ j ⊗(Z (k−j))−1(cid:17)η(cid:19)∗ INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 25 as desired. Finally, to prove equation (4.11) it suffices to show that j ⊗(Z (k−j))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 β(cid:18)W(cid:16)X 1/2 β(cid:0)W(Z (k))−1η(cid:1) gg∗β(cid:0)W(Z (k))−1η(cid:1)∗ j ⊗(Z (k−j))−1(cid:17)η(cid:19)∗! (4.15) kXj=1 Xη∈Ak = Xη∈Ak k = Dk(cid:16)Pξ∈Ak ξ(cid:17) D∗ kg∗β(Wξ)∗xk2 = Xξ∈Ak Xξ∈Ak ≤ Xξ∈Ak k(W ∗ for any k ∈ N. Let x ∈ Jm+1. By definition of Dk and Lemma 3.3(3), DkD∗ ξ . Since kgk ≤ 1, WξW ∗ TξT ∗ k =Pξ∈Ak kg∗V ∗ m(W ∗ ξ ⊗ IH)Vm+1xk2 ξ ⊗ IH)xk2 = Xξ∈Ak(cid:10)(WξW ∗ ξ ⊗ IH)x, x(cid:11) = h(DkD∗ k ⊗ IH)x, xi = k(D∗ k ⊗ IH)xk2 . Using the notation and results from Lemma 3.6, it follows from Lemma 4.2(6) that for every ξ ∈ Ak, g∗β(Wξ)∗x belongs to Hξ. The preceding computation shows that (g∗β(Wξ)∗x)ξ∈Ak is an element in Hk; let ex SS∗⊗IH =Pξ∈Ak be its image in E⊗k ⊗σ H under the isomorphism γ∗ k. If S is any ele- ment in L(E⊗k), then using Lemma 3.3(1) and applying σE⊗k, we have (θSξ ⊗ IH). For each ξ ∈ Ak, the matrix for θSξ ⊗IH, when viewed as an operator in B(Hk), is [σhµ, SξiσhSξ, νi]µ,ν∈Ak by Lemma 3.6(2), so by Lemma 4.2 parts (6) and (8), a matrix computa- tion reveals, (4.16) hex, (SS∗ ⊗ IH)exi hσhSξ, µig∗β(Wµ)∗x, σhSξ, νi(g∗β(Wν)∗x)i!! = Xξ∈Ak Xµ∈Ak Xν∈Ak = Xξ∈Ak Xµ∈Ak Xν∈Ak(cid:10)g∗β(Wµhµ,Sξi)∗x, g∗β(Wνhν,Sξi)∗x(cid:11)!! hg∗β(WSξ)∗x, g∗β(WSξ)∗xi = Xξ∈Ak = Xξ∈Ak k andPk Using the identities, Z (k)∗Z (k) = R−2 as two applications of equation (4.16), first when S = X 1/2 hβ(WSξ)gg∗β(WSξ)∗x, xi . j=1 Xj ⊗R2 k, as well j ⊗(Z (k−j))−1 k−j = R2 26 JENNIFER R GOOD and second when S = (Z (k))−1, we obtain * kXj=1 Xη∈Ak = j ⊗(Z (k−j))−1(cid:17)η(cid:19)∗ x, x+ β(cid:18)W(cid:16)X 1/2 j ⊗(Z (k−j))−1(cid:17)η(cid:19) gg∗β(cid:18)W(cid:16)X 1/2 kXj=1 k−j) ⊗ IH(cid:1)exi = hex, (R2 hex,(cid:0)(Xj ⊗ R2 k ⊗ IH)exi = Xη∈AkD(cid:16)β(cid:16)W(Z (k))−1η(cid:17) gg∗β(cid:16)W(Z (k))−1η(cid:17)∗(cid:17) x, xE . β(cid:0)W(Z (k))−1η(cid:1) gg∗β(cid:0)W(Z (k))−1η(cid:1)∗ converges in B(Jm+1) and equation (4.15) holds, which completes the proof of inequality (4.8). Momentarily, we define Gm+1 ∈ B(K, Jm+1) as a 2 × 2 operator- valued matrix with respect to the decompositions K = (K ⊖ K0) ⊕ K0 and Jm+1 = Jm ⊕ (Jm+1 ⊖ Jm). One of the columns of Gm+1 is determined by the adjoint of a contraction F : Jm+1 → K ⊖ K0 that we define as follows at x ∈ Jm+1. By inequality (4.8), ThusPη∈Ak (4.17) ∞Xk=1 Xξ∈Ak =* ∞Xk=1 Xξ∈Ak xk2! kg∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ β(cid:0)W(Z (k))−1ξ(cid:1) gg∗β(cid:0)W(Z (k))−1ξ(cid:1)∗! x, x+ ≤ kxk2. It follows that(cid:16)δk≥1 (cid:0)g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ ement of P∞ is a well-defined el- k=0 ⊕Hk where δk≥1 is 1 when k ≥ 1 and 0 when k = 0. k=0 We define F (x) to be its image under Γ∗ in K, namely x(cid:1)ξ∈Ak(cid:17)∞ (4.18) F (x) = ∞Xk=1 Xξ∈Akbξ ⊗ g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ x! . M, and h ∈ H, The operator F is linear, and by inequality (4.17), F is contractive. If a ∈ M, k ∈ N, and ξ ∈ Ak, thenDbξ,baE = 0. Thus, for x ∈ Jm+1, a ∈ x, σDbξ,baE hE! = 0. hF (x),ba ⊗ hi = ∞Xk=1 Xξ∈AkDg∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ It follows that F (x) ∈ K ⊖ K0, as desired. For organizational purposes, we temporarily adopt the following no- ) denote tation: if L is a Hilbert space with closed subspace L0, let (V L L0 INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 27 the inclusion isometry in B(L0, L), and let (P L ) denote the projection L0 )∗ is the range restriction of in B(L) onto L0. (P L ) to L0, (V L ) is the identity on L0 L0 L0. In this notation, let us show that F and Gm satisfy the following relationship: It follows that (V L L0 ), and (V L L0 )∗ = (P L L0 )∗(V L L0 )(V L L0 (4.19) )∗F ∗ = Gm(V K If x ∈ Jm then by equations (4.18) and (4.5), (V Jm+1 Jm K⊖K0 ). F (V Jm+1 Jm )x = = ∞Xk=1 Xξ∈Akbξ ⊗ g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ (V Jm+1 )x! ∞Xk=1 Xξ∈Akbξ ⊗ g∗α(cid:0)W(Z (k))−1ξ(cid:1)∗ x! = (V K Jm )∗G∗ mx. K⊖K0 Therefore, F (V Jm+1 Jm ) = (V K K⊖K0 )∗G∗ m, and equation (4.19) follows. In our 2 × 2 matricial definition of Gm+1, the following maps, R, S, and T , will comprise three of the four corners. Define R = (V Jm+1 Jm S = (V Jm+1 T = Gm(V K K0 Jm+1⊖Jm ) )∗F ∗ = Gm(V K K⊖K0 ) in B(K ⊖ K0, Jm); )∗F ∗ in B(K ⊖ K0, Jm+1 ⊖ Jm); and in B(K0, Jm); noting that R is well-defined by equation (4.19). Matricially, By Parrott's lemma [11], there exists U ∈ B(K0, Jm+1 ⊖ Jm) such F ∗ =(cid:20)R S(cid:21) and Gm =(cid:2)R T(cid:3) . (cid:13)(cid:13)(cid:13)(cid:13)(cid:20)R T S U(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) = max {kF k , kGmk} , Gm+1 =(cid:20)R T S U(cid:21) (4.20) that (4.21) (4.22) (4.23) so at last we define in B(K, Jm+1). By equation (4.20), Gm+1(V K (V Jm+1 Jm K⊖K0 ) = F ∗ and )∗Gm+1 = Gm. i=1 , {Ji}m+1 i=1 , {Gi}m+1 Our goal is to show that(cid:0){ni}m+1 triple. We have already shown that the first four (m+1)-level conditions are satisfied. Condition (7, m + 1) readily follows from equation (4.21). i=1 (cid:1) is an (m + 1)- 28 JENNIFER R GOOD Towards proving Condition (5, m + 1), recall that σF(E) ◦ ϕ∞ : M → B(K) is the normal, unital ∗-homomorphism that maps a ∈ M to ϕ∞(a) ⊗ IH . Let each of ψ1 : M → B(K ⊖ K0), ψ2 : M → B(K0), π1 : M → B(Jm), and π2 : M → B(Jm+1 ⊖ Jm) be the compression of σF(E) ◦ ϕ∞ to the indicated space; for instance, for every a ∈ M, ψ1(a) = (V K ). The spaces K0, Jm, and Jm+1 reduce σF(E) ◦ ϕ∞, from which it follows that K ⊖ K0 and Jm+1 ⊖ Jm also reduce σF(E) ◦ ϕ∞. Therefore, ψ1, ψ2, π1, and π2 are normal, unital ∗-homomorphisms. We want to show that (4.24) )∗(ϕ∞(a)⊗IH)(V K K⊖K0 K⊖K0 R ∈ I(ψ1, π1); S ∈ I(ψ1, π2); T ∈ I(ψ2, π1); and U ∈ I(ψ2, π2). Note that Vm = (V K Jm σF(E) ◦ ϕ∞, if a ∈ M, then by Condition (5, m), ) and Vm+1 = (V K Jm+1 ). Since K ⊖ K0 reduces Rψ1(a) = Gm(P K K⊖K0 )(ϕ∞(a) ⊗ IH)(V K ) K⊖K0 = Gm(ϕ∞(a) ⊗ IH)(V K ) K⊖K0 = V ∗ m(ϕ∞(a) ⊗ IH)VmGm(V K K⊖K0 ) = π1(a)R, Thus R ∈ I(ψ1, π1); analogously T ∈ I(ψ2, π1). To show that S ∈ I(ψ1, π2), it suffices to show that for every a ∈ M, x ∈ Jm+1 ⊖ Jm, η ∈ E⊗j with j ∈ N, and h ∈ H, (4.25) Thus by Lemma 4.2, parts (6) and (8), (4.26) hbη ⊗ h, ψ1(a)S∗xi = Xξ∈Aj(cid:10)h, σhη, a · ξi(cid:0)g∗β(cid:0)W(Z (j))−1ξ(cid:1)∗ x(cid:1)(cid:11) = Xξ∈Aj(cid:10)h, g∗β(cid:0)W(Z (j))−1ξ·hξ,a∗ηi(cid:1)∗ x(cid:11) =(cid:10)h, g∗β(cid:0)W(Z (j))−1(a∗η)(cid:1)∗ x(cid:11) . Since K ⊖ K0 reduces σF(E) ◦ ϕ∞, by equation (4.18), ψ1(a)S∗x = (ϕ∞(a) ⊗ IH)F x hbη ⊗ h, ψ1(a)S∗xi = hbη ⊗ h, S∗π2(a)xi. = (ϕ∞(a) ⊗ IH) ∞Xk=1 Xξ∈Akbξ ⊗ g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ x!! ∞Xk=1 Xξ∈Akda · ξ ⊗ g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ x! . = INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 29 On the other hand, since Jm+1 ⊖ Jm reduces σF(E) ◦ ϕ∞, by Lemma 4.2(3), S∗π2(a)x = F (V K Jm+1 )∗(ϕ∞(a) ⊗ IH)(V K Jm+1 )x = = Jm+1 )∗(ϕ∞(a) ⊗ IH )(V K Jm+1 ∞Xk=1 Xξ∈Akbξ ⊗(cid:16)g∗β(cid:0)W(Z (k))−1ξ(cid:1)∗ (V K )x(cid:17)! ∞Xk=1 Xξ∈Akbξ ⊗ g∗β(cid:0)Wa∗·(Z (k))−1ξ(cid:1)∗ x! . hbη ⊗ h, S∗π2(a)xi = Xξ∈Aj(cid:10)h, σhη, ξi(cid:0)g∗β(cid:0)Wa∗·(Z (j))−1ξ(cid:1)∗ x(cid:1)(cid:11) = Xξ∈Aj(cid:10)h, g∗β(cid:0)Wa∗(Z (j))−1ξhξ,ηi(cid:1)∗ x(cid:11) . x(cid:11) =(cid:10)h, g∗β(cid:0)W(Z (j))−1(a∗η)(cid:1)∗ Therefore by Lemma 4.2, parts (6) and (8), since (Z (j))−1 ∈ ϕj(M)c, (4.27) Combining equations (4.26) and (4.27), we obtain equation (4.25), so S ∈ I(ψ1, π2). Examining the proof of Parrott's Lemma in [11], we see that U is given as the weak operator limit of the sequence (cid:8)−cnS(I − c2 nR∗R)−1R∗T(cid:9)∞ n=0 for some sequence of numbers {cn}∞ n=0. Since R ∈ I(ψ1, π1), S ∈ I(ψ1, π2), and T ∈ I(ψ2, π1), it follows that U ∈ I(ψ2, π2), which gives (4.24). To establish Condition (5, m+1), by Conditions (4, m+1) and (5, m), it suffices to show that (4.28) (V ∗ m+1(Y ⊗ IH)Vm+1)Gm+1 = Gm+1(Y ⊗ IH) in only two cases: when Y = ϕ∞(a) for some a ∈ M and when Y = Wξ for some ξ ∈ E. Using properties (4.24), we compute V ∗ 0 m+1(ϕ∞(a) ⊗ IH)Vm+1Gm+1 =(cid:20)π1(a) π2(a)(cid:21)(cid:20)R T S U(cid:21) =(cid:20)π1(a)R π1(a)T π2(a)S π2(a)U(cid:21) =(cid:20)Rψ1(a) T ψ2(a) Sψ1(a) Uψ2(a)(cid:21) =(cid:20)R T S U(cid:21)(cid:20)ψ1(a) ψ2(a)(cid:21) = Gm+1(ϕ∞(a) ⊗ IH). 0 0 0 This establishes equation (4.28) when Y = ϕ∞(a). For the case when Y = Wξ, it suffices to show that for every j ∈ N0, η ∈ E⊗j, h ∈ H, and 30 JENNIFER R GOOD x ∈ Jm+1, (4.29) (4.18), (cid:10)V ∗ m+1(Wξ ⊗ IH )Vm+1Gm+1 (bη ⊗ h) , x(cid:11) = hGm+1(Wξ ⊗ IH) (bη ⊗ h) , xi . With Wξ(bη) ⊗ h = \Zj+1(ξ ⊗ η) ⊗ h ∈ K ⊖ K0, by equations (4.22) and hGm+1(Wξ ⊗ IH) (bη ⊗ h) , xi m+1xE =D \Zj+1(ξ ⊗ η) ⊗ h, F xE =D \Zj+1(ξ ⊗ η) ⊗ h, (V K xE! . ∞Xk=1 Xµ∈AkD \Zj+1(ξ ⊗ η) ⊗ h,bµ ⊗ g∗β(cid:0)W(Z (k))−1µ(cid:1)∗ For k ∈ N and µ ∈ Ak, D \Zj+1(ξ ⊗ η),bµE is hZj+1(ξ ⊗ η), µi when k = j + 1 and is otherwise zero. Recalling that (Z (j+1))−1Zj+1 = I1 ⊗ (Z (j))−1, we continue the preceding computation using Lemma 4.2, parts (6) and (8), )∗G∗ K⊖K0 = (4.30) hGm+1(Wξ ⊗ IH) (bη ⊗ h) , xi = Xµ∈Aj+1(cid:10)h, σhZj+1(ξ ⊗ η), µi(cid:0)g∗β(cid:0)W(Z (j+1))−1µ(cid:1)∗ x(cid:1)(cid:11) = Xµ∈Aj+1Dh, g∗β(cid:16)W(Z (j+1))−1µhµ,Zj+1(ξ⊗η)i(cid:17)∗ xE =Dh, g∗β(cid:16)W(Z (j+1))−1Zj+1(ξ⊗η)(cid:17)∗ xE =(cid:10)h, g∗β(cid:0)Wξ⊗(Z (j))−1η(cid:1)∗ x(cid:11) =(cid:10)β(cid:0)Wξ⊗(Z (j))−1η(cid:1) g(h), x(cid:11) . )∗ and V ∗ On the other hand, we have that PmVm+1 = Vm(V Jm+1 Jm IH) = V ∗ and Lemma 4.2, parts (4) and (2), m+1(Wξ⊗ m+1(Wξ ⊗ IH)Pm by equation (4.1). Thus, by equation (4.23) V ∗ = V ∗ m+1(Wξ ⊗ IH )Vm+1Gm+1 (bη ⊗ h) = V ∗ m+1(Wξ ⊗ IH)Vm(V Jm+1 Jm m+1(Wξ ⊗ IH )PmVm+1Gm+1Lbηh = β (Wξ) α(cid:0)W(Z (j))−1η(cid:1) g(h) = β(cid:0)Wξ⊗(Z (j))−1η(cid:1) g(h). )∗Gm+1Lbηh = β(Wξ)GmLbηh This fact together with equation (4.30) yields equation (4.29), which completes the proof of Condition (5, m + 1). All that remains is Condition (6, m + 1), and we need only consider the case when j = m + 1. i VjGj = Gm+1 = Gi, as desired. Suppose 1 ≤ i ≤ m. Since Ji ⊆ Jm ⊆ Jm+1 ⊆ K, it follows that V ∗ ). Taking If i = m + 1, then V ∗ ) and V ∗ j Vi = (V Jm+1 Jm )(V Jm Ji mVi = (V Jm Ji INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 31 i=1 , {Ji}m+1 i=1 , {Gi}m+1 i=1 (cid:1) is an (m + 1)-triple. adjoints, applying equation (4.23), and using Condition (6, m), we have i VjGj = (V Jm V ∗ i VmGm = Gi. Thus, Ji Condition (6, m + 1) is satisfied. Having shown all seven conditions, )∗Gm+1 = (V Jm Ji )∗Gm = V ∗ )∗(V Jm+1 Jm we conclude that(cid:0){ni}m+1 i=1, {Gi}m i=1, {Ji}m i=1, {Gi}1 Recall that we are considering the case when there is no m ∈ N such that an m-triple ({ni}m i=1) exists with Jm = K. Since we know ({ni}1 i=1, {Ji}1 i=1) is a 1-triple, the result we have just shown, together with an inductive argument, guarantees the existence of three sequences {ni}∞ i=1 such that ({ni}m i=1) is an m-triple for every m ∈ N. As before, for every i ∈ N, let Vi ∈ B(Ji, K) be the inclusion map, and let Pi ∈ B(K) be the projection map onto Ji. i=1, and {Gi}∞ i=1, {Ji}∞ i=1, {Gi}m i=1, {Ji}m Towards defining eG, note that {Ji}∞ S∞ Therefore S∞ i=1 is an increasing family of sub- spaces of K because Condition (2, m) holds for every m, so J∞ := k=1 Jk is a linear subspace of K. An inductive argument using Con- dition (1, m) for m ≥ 2 implies that for all k ∈ N0, k ≤ nk+2. Thus for each k ∈ N0, Condition (3, k + 2) implies that Kk ⊆ Knk+2 ⊆ Jk+2. k=0 Kk ⊆ J∞. It follows that K = J∞. Suppose x ∈ J∞ and x ∈ Ji ∩ Jj for some i, j ∈ N. Assuming without loss of general- ity that i ≤ j, Ji ⊆ Jj by Condition (2, j), and by Condition (8, j), V ∗ i VjGj = Gi. Therefore, G∗ j x. Using the Condi- tions (7, m) for m ≥ 1, it follows that there is a well-defined contraction C ∈ B(K) such that for every k ∈ N and x ∈ Jk, C(x) = G∗ j Vix = G∗ i x = G∗ j V ∗ k(x). k, (4.31) ∀k ∈ N. kx. Thus 1 since V = V1. Since G1 = GV ∗ and P = stated in the conclusion of the theorem. First, note that for any k ∈ N Let us define eG := C ∗ and show that eG satisfies the four properties and x ∈ Jk, eG∗Vkx = Cx = G∗ eG∗Vk = G∗ In particular, eG∗V = G∗ V V ∗ is the projection map onto J in B(K), we have eG∗P = V G∗V ∗. k (Y ⊗ IH )eG = k eG(Y ⊗ IH). k (Y ⊗ IH)PkeG = V ∗ Properties (1) and (2) follow readily. Let Y ∈ H ∞(E, Z) and k ∈ N. By Condition (4, k), V ∗ k (Y ⊗ IH)Pk; by Condition (5, k) and two applications of the adjoint of equation (4.31), V ∗ V ∗ k (Y ⊗ IH )VkGk = Gk(Y ⊗ IH) = V ∗ Property (3) follows since K = J∞. Finally, C is a contraction, but also kCk ≥ 1 since kG1k = 1. This gives property (4), completing the proof of the theorem in the case when kGk = 1. If G is the zero operator in B(J), then the zero operator in B(K) satisfies the desired k (Y ⊗ IH) = V ∗ properties for eG. If G 6= 0, then a straightforward scaling argument utilizing the case treated above produces the desired result. (cid:3) 32 JENNIFER R GOOD The following corollary generalizes Theorem 4.1 to the case where the W ∗-algebra is represented on two Hilbert spaces. The proof makes use of the so-called Putnam trick. It is this corollary that we use to prove the weighted Nevanlinna-Pick interpolation theorem in Section 6. Corollary 4.3. Let σ1 : M → B(H1) and σ2 : M → B(H2) be faith- ful, normal, unital ∗-homomorphism for Hilbert spaces H1 and H2, and let Z be a sequence of weights associated with X. For j = 1, 2, sup- pose that Jj is a closed linear subspace of F(E) ⊗σj Hj such that for the inclusion map of Jj into F(E) ⊗σj Hj. Suppose there exists G ∈ B(J1, J2) such that for every Y ∈ H ∞(E, Z), G (V ∗ 1 (Y ⊗ IH1)V1) = (V ∗ H2) such that every Y ∈ H ∞(E, Z), (cid:0)Y ∗ ⊗ IHj(cid:1) (Jj) ⊆ Jj. For j = 1, 2, let Vj be 2 (Y ⊗ IH2)V2) G. Then there exists eG ∈ B(F(E) ⊗σ1 H1, F(E) ⊗σ2 (1) eG∗(J2) ⊆ J1, 2 eGV1 = G, (3) eG(Y ⊗ IH1) = (Y ⊗ IH2)eG for all Y ∈ H ∞(E, Z), and (4) keGk = kGk. (2) V ∗ Proof. We will use the "Putnam trick" to translate the two-space prob- lem into a one-space problem where we may apply the original result, Theorem 4.1. We then return to the two-space setting by picking off the lower left-hand entries in the 2 × 2 matricial expressions for certain operators. Let H = H1 ⊕ H2. Let σ = σ1 ⊕ σ2 : M → B(H). Define the induced representation spaces K1 = F(E) ⊗σ1 H1, K2 = F(E) ⊗σ2 H2, and K = F(E) ⊗σ H. We identify K with K1 ⊕ K2 in the usual way and let J be the image of J1 ⊕ J2 in K. For notational simplicity we will omit the implied isomorphisms in our computations. Let V be the inclusion map of J into K. Let P1, P2, and P be the projection maps in B(K1), B(K2), and B(K) onto J1, J2, and J, respectively. Define (4.32) G0 :=(cid:20) 0 0 G 0(cid:21) ∈ B(J). Towards applying Theorem 4.1, let us show that for all Y ∈ H ∞(E, Z), (4.33) (4.34) (Y ∗ ⊗ IH)(J) ⊆ J and G0(V ∗(Y ⊗ IH)V ) = (V ∗(Y ⊗ IH)V )G0. INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 33 By hypothesis, Pj(Y ⊗ IHj ) = Pj(Y ⊗ IHj )Pj for j = 1, 2. Thus, 0 P (Y ⊗ IH) =(cid:20)P1 0 P2(cid:21)(cid:20)Y ⊗ IH1 =(cid:20)P1 0 P2(cid:21)(cid:20)Y ⊗ IH1 Y ⊗ IH2(cid:21) Y ⊗ IH2(cid:21)(cid:20)P1 0 0 0 0 0 0 0 P2(cid:21) = P (Y ⊗ IH )P. Containment (4.33) follows. Since V V ∗ = P and V ∗V is the identity on J, (4.35) G0V ∗(Y ⊗ IH )V = V ∗V G0V ∗P (Y ⊗ IH)P V = V ∗(cid:20) 0 V2GV ∗ 1 0 0(cid:21)(cid:20)P1 0 0 P2(cid:21)(cid:20)Y ⊗ IH1 = V ∗(cid:20) 0 0 0 Y ⊗ IH2(cid:21)(cid:20)P1 0 P2(cid:21) V 1 (Y ⊗ IH1)P1 0(cid:21) V. 0 0 V2GV ∗ On the other hand, (4.36) V ∗(Y ⊗ IH)V G0 = V ∗P (Y ⊗ IH)P V G0V ∗V = V ∗(cid:20)P1 0 P2(cid:21)(cid:20)Y ⊗ IH1 0 0 0 Y ⊗ IH2(cid:21)(cid:20)P1 = V ∗(cid:20) 0 0 P2(cid:21)(cid:20) 0 V2GV ∗ 1 0 P2(Y ⊗ IH2)V2GV ∗ 1 0 0(cid:21) V 0(cid:21) V. 0 By our hypothesis, V2GV ∗ (4.34) now follows from equations (4.35) and (4.36). 1 (Y ⊗ IH1)P1 = P2(Y ⊗ IH2)V2GV ∗ 1 . Equation Having shown containment (4.33) and equation (4.34), we apply The- ∗ (J) ⊆ J, orem 4.1 and conclude that there exists fG0 ∈ B(K) such that: (1′) fG0 (2′) V ∗fG0V = G0, (3′) fG0(Y ⊗ IH) = (Y ⊗ IH)fG0 for all Y ∈ H ∞(E, Z), and (4′) kfG0k = kG0k. With A ∈ B(K1), B ∈ B(K2, K1), C ∈ B(K2), and eG ∈ B(K1, K2) such that (4.37) eG C(cid:21) , fG0 =(cid:20)A B 34 JENNIFER R GOOD 0 0 0 0 let us show that eG satisfies properties (1)-(4). By property (1′), PfG0 = PfG0P , so eG C(cid:21) = PfG0 = PfG0P 0 P2(cid:21)(cid:20)A B P2eG P2C(cid:21) =(cid:20)P1 (cid:20)P1A P1B P2eGP1 P2CP2(cid:21) . eG C(cid:21)(cid:20)P1 0 P2(cid:21) =(cid:20)P1AP1 P1BP2 =(cid:20)P1 0 P2(cid:21)(cid:20)A B from property (2′) that PfG0P = V G0V ∗, so by similar computations, P2eGP1 P2CP2(cid:21) =(cid:20) 0(cid:21) . (cid:20)P1AP1 P1BP2 Property (2) follows by equating the lower left-hand entries. Property (3′) implies Equating the lower left-hand entries, property (1) holds. It follows V2GV ∗ 1 0 for Y ∈ H ∞(E, Z). Again, equating the lower left-hand entries, we (Y ⊗ IH2)eG (Y ⊗ IH2)C(cid:21) eG(Y ⊗ IH1) C(Y ⊗ IH2)(cid:21) =(cid:20)(Y ⊗ IH1)A (Y ⊗ IH1)B (cid:20)A(Y ⊗ IH1) B(Y ⊗ IH2) obtain property (3). Finally, it follows from equation (4.37) that keGk ≤ kfG0k. By property (4′) and equation (4.32), kfG0k = kG0k = kGk. Since kGk ≤ keGk by property (2), we conclude that kGk = keGk. This gives property (4) and completes the proof. (cid:3) 5. The Commutant and Double Commutant In this section we identify the commutant of an induced image of the weighted Hardy algebra of the dual correspondence and the double commutant of an induced image of the weighted Hardy algebra of the original correspondence. For the remainder of the paper, we will assume that E is full. Let σ : M → B(H) be a fixed faithful, normal, unital ∗-homomor- phism for a Hilbert space H, and let Z be a sequence of weights asso- ciated with X. In Section 7 of [9], Muhly and Solel use X to construct an admissible sequence for Eσ, X ′ = {X ′ k=0, and Z to construct a sequence of weights associated with X ′, which we will denote by Z ′ = {Z ′ k=0. Our proof of the weighted Nevanlinna-Pick theorem will involve the commutant of the algebra {Y ⊗ IH Y ∈ H ∞(Eσ, Z ′)} in B(F(Eσ) ⊗ι H); our next goal is to identify it with H ∞(E, Z) in a certain way. A note of caution is in order, for what we will call Z ′ is referred to as C in [9]. Let us summarize the construction of X ′ and Z ′, simultaneously clarifying our notation. For k ∈ N0, we k}∞ k}∞ k :=(Iσ(M )′ C ′ if k = 0 if k ≥ 1 . INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 35 k for ϕ′ shall write Ak for ϕk(M)c, the commutant of ϕk(M) in L(E⊗k), and A′ k(σ(M)′)c, the commutant of the image of σ(M)′ under ϕ′ k in L((Eσ)⊗k). Using Lemma 7.2 of [9], for each k ∈ N0 we define X ′ k to be the unique element in A′ k ⊗ IH. For k ∈ N0, define k such that U σ∗ k (Xk ⊗ IH)U σ k = X ′ Ck =(IM Z (k)(Z (k−1) ⊗ I1)−1 if k = 0 if k ≥ 1 . k = Z ′ k (Ck ⊗ IH)U σ k to be unique element in A′ Since Ck ∈ Ak, we define Z ′ k such that U σ∗ k ⊗ IH ([9], Lemma 7.2). Here is where we have modified the notation used in [9]; we have interchanged the roles of C and Z ′ as they were introduced in Lemma 7.4 of [9]. By Theorem 7.6 of [9], X ′ = {X ′ k=0 is a sequence of weights associated with X ′. k=0 is an admissible sequence and Z ′ = {Z ′ k}∞ k}∞ We now repeat the construction with Eσ replacing E and Eσι re- k(σ(M))c, the commutant of k denote ϕ′′ placing Eσ. For k ∈ N0, let A′′ ϕ′′ k(σ(M)) in L((Eσι)⊗k). For k ∈ N0 define Z ′(k)(Z ′(k−1) ⊗ I ′ k, there exist unique X ′′ 1)−1 k , Z ′′ k = Z ′′ k }∞ k }∞ k (C ′ k (X ′ k ⊗ IH)U ι k, C ′ k = X ′′ k such that U ι∗ k ∈ A′ k ⊗ IH and U ι∗ k ∈ A′′ k ⊗ k ⊗ IH ([9], Lemma 7.2). It k=0 is an admissible k=0 is a sequence of weights associated Since X ′ IH)U ι follows from Theorem 7.6 of [9] that X ′′ = {X ′′ sequence for Eσι and Z ′′ = {Z ′′ with X ′′. The following fact will be useful in future computations. Lemma 5.1. For all k ∈ N0, Proof. The case when k = 0 follows from the fact that Z0 is the identity 0 is the identity on σ(M)′. Suppose k ≥ 1. By the discus- on M and C ′ sion preceding Theorem 7.6 in [9], Z ′(k) ⊗IH = U σ∗ k ; more- over, by Lemma 7.2 of [9], Z ′(k−1)⊗I ′ k . k (Z (k)⊗IH )U σ k (Zk ⊗ IH)U σ 1⊗IH = U σ∗ k = C ′ k ⊗ IH. U σ∗ k (cid:0)I1 ⊗ Z (k−1) ⊗ IH(cid:1) U σ Therefore, since Z (k)(cid:0)I1 ⊗ Z (k−1)(cid:1)−1 k ⊗ IH = (Z ′(k) ⊗ IH)(Z ′(k−1) ⊗ I ′ C ′ = Zk, 1 ⊗ IH)−1 = U σ∗ k (Z (k) ⊗ IH)(cid:0)I1 ⊗ Z (k−1) ⊗ IH(cid:1)−1 as desired. U σ k = U σ∗ k (Zk ⊗ IH)U σ k , (cid:3) We already know how to naturally identify M with σ(M), E with Eσι, and σ with . To identify X with X ′′ and Z with Z ′′, that is to show that (5.1) ωkZkω∗ k = Z ′′ k , and ωkXkω∗ k = X ′′ k , 36 JENNIFER R GOOD k U ι note that for all k ∈ N0, U σ E⊗k ⊗σ H. By Lemma 5.1, U ι∗ k (C ′ ωkZkω∗ ωkXkω∗ k ⊗ IH )U ι k ⊗ IH , so by the uniqueness of Z ′′ k = X ′′ k . k(ωk ⊗ IH) is the identity operator on k = k . Analogously, k = U ι∗ k , ωkZkω∗ k (Zk ⊗ IH)U σ k = Z ′′ k U σ∗ k U ι Lemma 5.2. Ad(ω∞)(H ∞(E, Z)) = H ∞(Eσι, Z ′′). (5.2) Proof. It suffices to show that ∞ = W Z ′′ ∞ = ϕ′′ ξ ω∗ ω∞ϕ∞(a)ω∗ ω∞W Z (5.3) ωk(ξ), ∞(σ(a)), ∀k ∈ N, ∀ξ ∈ E⊗k, and ∀a ∈ M. By straightforward computation with equation (5.1), ωiZ (i,i−k)ω∗ i = Z ′′(i,i−k) when i ≥ k ≥ 0, so ω∞DZ ∞ = Tωk(ξ), equation (5.2) follows. Equation (5.3) is simply the case when k = 0. (cid:3) k . Since ω∞Tξω∗ ∞ = DZ ′′ k ω∗ The following theorem gives two commutant results. The first plays a role at the crucial step of the weighted Nevanlinna-Pick theorem in the next section. The second is a weighted double commutant theorem. Theorem 5.3. In the notation established, (5.4) Proof. By Theorem 7.7 in [9], = πσ(H ∞(E, Z)), and = σF(E) (H ∞(E, Z)). (1) (cid:0)ιF(Eσ)(H ∞(Eσ, Z ′))(cid:1)′ (2) (cid:0)σF(E) (H ∞(E, Z))(cid:1)′′ That same theorem, now applied to Eσ, gives(cid:0)ιF(Eσ) (H ∞ (Eσ, Z ′))(cid:1)′ (cid:0)ιF(Eσ) (H ∞ (Eσ, Z ′))(cid:1)′ ρι (H ∞ (Eσι, Z ′′)). Now ρι ◦Ad(ω∞) = πσ since U σ ∞U ι identity operator on F(E) ⊗σ H, so by Lemma 5.2, = ρι (Ad(ω∞) (H ∞(E, Z))) = πσ (H ∞(E, Z)) , ∞) ◦ ιF(Eσ), part (1) now implies (cid:0)σF(E) (H ∞ (E, Z))(cid:1)′ = ∞ (ω∞ ⊗ IH ) is the = ρσ (H ∞ (Eσ, Z ′)) . which gives part (1). Since ρσ = Ad(U σ that (cid:0)ρσ (H ∞ (Eσ, Z ′))(cid:1)′ = Ad(U σ = Ad(U σ ∞)(cid:16)(cid:0)ιF(Eσ) (H ∞ (Eσ, Z ′))(cid:1)′(cid:17) ∞) (πσ (H ∞(E, Z))) . Since Ad(U σ ∞) ◦ πσ = σF(E) equation (5.4) gives (cid:0)σF(E) (H ∞(E, Z))(cid:1)′′ =(cid:0)ρσ (H ∞ (Eσ, Z ′))(cid:1)′ = Ad(U σ ∞) (πσ (H ∞(E, Z))) = σF(E) (H ∞(E, Z)) , which completes the proof of part (2). (cid:3) INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 37 The following corollary to Theorem 5.3(1) will be of use in the next section. For s, t ∈ N, H (s) denotes the direct sum of s copies of H, Ms×t(B(H)) denotes the collection of s × t matrices with entries in B(H), identified with B(H (t), H (s)) in the usual way, and Ms(B(H)) denotes the collection of s×s matrices with entries in B(H). We define ι(s) : σ(M)′ → Ms(B(H)) to be the direct sum of s copies of ι. Define πσ s×t : Ms×t (L(F(E))) → Ms×t (B(F(Eσ) ⊗ι H)) so that πσ s×t(cid:16)[Yaf ]s a=1 t f =1(cid:17) = [πσ(Yaf )]s t f =1 . a=1 We identify, as usual, F(Eσ) ⊗ι(s) H (s) with (F(Eσ) ⊗ι H)(s). Corollary 5.4. Let s, t ∈ N. The image of Ms×t (H ∞ (E, Z)) under πσ s×t is precisely the collection of elements A ∈ Ms×t (B(F(Eσ) ⊗ι H)) such that for every S ∈ H ∞ (Eσ, Z ′), A ◦ (ι(t))F(Eσ)(S) = (ι(s))F(Eσ)(S) ◦ A. 6. Weighted Nevanlinna-Pick Interpolation In this section we give our primary theorem, a weighted W ∗-version of the classic Nevanlinna-Pick interpolation theorem. While the key component of its proof, the weighted commutant lifting theorem, is already at our disposal, we begin this section with a series of technical lemmas that further facilitate the proof. Let σ : M → B(H) be a fixed faithful, normal, unital ∗-homomorphism for a Hilbert space H, and let Z be a sequence of weights associated with X. Let us establish some notation. If z ∈ D(X, σ), then the map Φz : k=1 z(k)(Xk ⊗ (k)∗ is completely positive, linear, and ultraweakly continuous by A)z Lemma 4.4 of [9]; in addition, kΦzk < 1. Therefore, in the Banach z converges in σ(M)′ → σ(M)′, defined for A ∈ σ(M)′ by Φz(A) = P∞ algebra of completely bounded maps on σ(M)′,P∞ norm. In fact, j=0 Φj (6.1) Φj z (A) = (k)(R2 k ⊗ A)z z (k)∗, A ∈ σ(M)′ ∞Xj=0 ∞Xk=0 by Theorem 4.5 of [9], where the summation on the left-hand side of the equation is with respect to the norm topology on σ(M)′, and the summation on the right-hand side is with respect to the ultraweak topology on σ(M)′. Let z ∈ I(σE ◦ ϕ, σ). For k ∈ N0, (cid:0)(Z (k))−1(cid:1)∗ Thus(cid:0)(cid:0)(Z (k))−1(cid:1)∗ ⊗ IH(cid:1) z(k)∗ belongs to(cid:0)E⊗k(cid:1)σ cz(k) to be its image in (Eσ)⊗k under the isomorphism Λσ∗ belongs to ϕk(M)c. z (k) = k . For w, z ∈ , and we define cZ 38 JENNIFER R GOOD D(X, σ) and A ∈ σ(M)′, we have hcw(k), A · cz(k)i = w(k)(R2 Define the Z-Cauchy kernel at z to be the tuple, k ⊗ A)z(k)∗. z = cz = (cz(k))∞ cZ k=0 . Lemma 6.1. If z ∈ D(X, σ), then cz ∈ F(Eσ). Proof. By equation (6.1),P∞ so for any N ∈ N, PN P∞ k=0 z(k)(R2 result follows. k ⊗ IH)z(k)∗ converges in σ(M)′, k ⊗ IH )z(k)∗ ≤ k ⊗ IH )z(k)∗. Since cz(k) ∈ (Eσ)⊗k for each k ∈ N0, the (cid:3) k=0 hcz(k), cz(k)i = PN k=0 z(k)(R2 k=0 z(k)(R2 Let w, z ∈ D(X, σ). We define K(w, z) : σ(M)′ → B(H) by K(w, z)(A) = hcw, A · czi = ∞Xk=0 (k)(R2 w k ⊗ A)z (k)∗, A ∈ σ(M)′. For s ∈ N, let K(s)(w, z) : σ(M)′ → Ms(B(H)) be the map defined at A ∈ σ(M)′ by K(s)(w, z)(A) = ι(s)(K(w, z)(A)) where, as in the preceding section, ι(s) is the direct sum of s copies of ι. Remark 6.2. While we will not explore the consequences here, it is read- ily shown that K : D(X, σ) × D(X, σ) → CB∗(σ(M)′, B(H)) is a "nor- mal completely positive kernel" in the sense of [4] and [6], independent of the choice of Z, where CB∗(σ(M)′, B(H)) denotes the completely bounded ultraweakly continuous maps from σ(M)′ to B(H). As such, by Proposition 41 of [6], K has an associated "Reproducing Kernel W ∗- Correspondence", a W ∗-analogue of the classical Reproducing Kernel Hilbert Space; in fact K, which we may refer to as the (X, σ)-Szego kernel, is a weighted W ∗-version of the classic Szego kernel. Recall thatcIH = (δi=0 IH )∞ i=0 belongs to F(Eσ) since IH is the multi- plicative identity of σ(M)′ = (Eσ)⊗0. If z ∈ D(X, σ), then cz belongs to F(Eσ) by Lemma 6.1. We write LI and Lz for the insertion operators LH cIH Lemma 6.3. For z ∈ D(X, σ) and Y ∈ H ∞(E, Z), , respectively. and LH cz (1) L∗ (2) L∗ (3) L∗ z πσ(Y )LIL∗ z = L∗ z πσ(Y ), z πσ(Y )LI = bY (z), and z πσ(Y ) = bY (z)L∗ z . Proof. Let us show that parts (1) and (2) hold when Y = ϕ∞(a) when a ∈ M. Since πσ(ϕ∞(a)) = I ′ ∞ ⊗ σ(a), hcz,cIHi = IH, and σ(a) = INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 39 (σ × z)(ϕ∞(a)), L∗ z πσ(ϕ∞(a))LI = L∗ z (I ′ ∞ ⊗ σ(a)) LI = L∗ z LIσ(a) = ιhcz,cIHiσ(a) = σ(a) = (σ × z)(ϕ∞(a)), which gives part (2). Part (1) now follows from the fact that if η ∈ F(Eσ) and x ∈ H, then L∗ z Lη = hcz, ηi belongs to σ(M)′, so using part (2), L∗ z πσ(ϕ∞(a))LIL∗ z (η ⊗ x) = σ(a)L∗ z Lηx = L∗ z Lησ(a)x = L∗ z (I ′ ∞ ⊗ σ(a))Lηx = L∗ z πσ(ϕ∞(a))(η ⊗ x). Now we show that parts (1) and (2) hold for Y = Wξ when ξ ∈ E. The ith diagonal entry of the diagonal operator πσ(D1) is zero when i = 0 and is otherwise U σ∗ i ⊗ IH (Lemma 5.1). It follows that i (Zi ⊗ IH)U σ i = C ′ πσ(Wξ) = πσ(D1)πσ(Tξ) (6.2) If Li denotes the insertion operator LH cially Lz is the column [Li]∞ i=0; hence L∗ By Lemma 5.1, C ′ U σ 1 (cz(1) ⊗ y) = cz(1)(y), when x, y ∈ H, 1 ⊗ IH = U σ∗ 1 (Z1 ⊗ IH)U σ cz(i) for each i ∈ N0, then matri- z πσ(Wξ)LI = L∗ 1 ⊗ IH )(ωξ). 1 ◦ (ωξ) = Lξ and 1 . Since U σ 1(C ′ (cid:10)L∗ z πσ(Wξ)LIx, y(cid:11) = h(C ′ 1 ⊗ IH)(ωξ)x, cz(1) ⊗ yi =(cid:10)(Z1 ⊗ IH )Lξx,(cid:0)(Z −1 1 )∗ ⊗ IH(cid:1) z ∗y(cid:11) = hzLξx, yi = h(σ × z)(Wξ)x, yi . We deduce that part (2) holds for Y = Wξ. Let us compare the row matrices for the operators on either side of the equality in part (1). By part (2), L∗ j=0. On the other hand, by equa- tion (6.2), L∗ . Therefore to obtain part (1), it suffices to show that for every j ∈ N0, z πσ(Wξ) =(cid:2)L∗ z πσ(Wξ)LIL∗ j+1(C ′ j=0 z = (cid:2)zLξL∗ j(cid:3)∞ j+1 ⊗ IH)(cid:0)I ′ j+1 ⊗ IH)(cid:0)I ′ j+1(C ′ j ⊗ ωξ(cid:1)(cid:3)∞ j ⊗ ωξ(cid:1) . zLξL∗ j = L∗ (6.3) 0 (C ′ 1 ⊗ IH)ωξ 0 0 0 0 (C ′ 2 ⊗ IH)(I ′ 1 ⊗ ωξ) 0 0 0 0 (C ′ 3 ⊗ IH)(I ′ . . . 2 ⊗ ωξ) . . . . . . =  .  40 JENNIFER R GOOD For any k ∈ N0, U σ so k Lk =(cid:0)((Z (k))−1)∗ ⊗ IH(cid:1) z(k)∗ by definition of cz(k), (6.4) L∗ kU σ∗ k = z (k)(cid:0)((Z (k))−1) ⊗ IH(cid:1) , k ∈ N0. Since Z (j+1) = Zj+1(I1 ⊗ Z (j)) and z of equation (6.4), first when k = j + 1 and then when k = j, yield (j+1) = z(I1 ⊗ z (j)), two applications (6.5) L∗ j+1U σ∗ j+1 (Zj+1 ⊗ IH) = z = z(cid:16)I1 ⊗ z (j+1)(cid:0)((Z (j+1))−1) ⊗ IH(cid:1) (Zj+1 ⊗ IH) (j)(cid:0)((Z (j))−1) ⊗ IH(cid:1)(cid:17) = z(cid:0)I1 ⊗ L∗ j (cid:1) j ⊗ ωξ(cid:1) = L(E⊗j ⊗σH) j+1(cid:0)I ′ j ⊗ ωξ(cid:1) = L∗ j ⊗ ωξ(cid:1) j+1(cid:0)I ′ j+1 (Zj+1 ⊗ IH) U σ j )L(E⊗j⊗σH) j = zLξL∗ j , = z(I1 ⊗ L∗ j+1U σ∗ j U σ∗ j U σ∗ U σ ξ ξ U σ j . Thus by As observed in Section 2.2, U σ Lemma 5.1 and equation (6.5), L∗ j+1(C ′ j+1 ⊗ IH)(cid:0)I ′ noting that the final equality follows readily from an elementwise com- putation. This gives equation (6.3), so part (1) holds for Y = Wξ when ξ ∈ E. To show that part (1) holds for arbitrary operators in H ∞(E, Z), suppose, inductively, that k ∈ N exists such that part (1) holds when- ever Y = Wξ for ξ ∈ E⊗j with 0 ≤ j ≤ k. For ξ ∈ E⊗k and η ∈ E, Wξ⊗η = Wξ ⊗ Wη, so L∗ z πσ(Wξ⊗η)LIL∗ z = L∗ z πσ(Wξ)LIL∗ = L∗ z πσ(Wξ)πσ(Wη)LIL∗ z πσ(Wη)LIL∗ z = L∗ z πσ(Wξ)πσ(Wη) = L∗ = L∗ z πσ(Wξ)LIL∗ z z πσ(Wη) z πσ(Wξ⊗η). Part (1) follows from routine approximation arguments. Towards ob- taining part (2) in the general case, define the linear and ultraweakly continuous map τ : H ∞(E, Z) → B(H) at Y ∈ H ∞(E, Z) by τ (Y ) = L∗ z πσ(Y )LI . By part (1), for Y1, Y2 ∈ H ∞(E, Z), τ (Y1Y2) = L∗ z πσ(Y2)LI = τ (Y1)τ (Y2). Thus τ is multiplicative. As both τ and (σ×z) are linear, multiplicative, and ultraweakly continuous and they agree on elements of the form Y = ϕ∞(a) when a ∈ M and Y = Wξ when ξ ∈ E, it follows that τ = (σ × z), which gives part (2). Finally, parts (1) and (2) together imply part (3). (cid:3) z πσ(Y1)πσ(Y2)LI = L∗ z πσ(Y1)LIL∗ Our final technical lemma concerns a weighted creation operator ξ = W ′ ξ ∈ H ∞(Eσ, Z ′). Recall that if z ∈ D(X, σ), then z∗ is W Z ′ INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 41 an element in Eσ and for any ξ ∈ Eσ, hξ, z∗i = ξ∗z∗ is an element of σ(M)′. Lemma 6.4. For z ∈ D(X, σ), ξ ∈ Eσ, and D ∈ σ(M)′, (cid:0)W ′ ξ(cid:1)∗ (D · cz) = hD∗ · ξ, z ∗i · cz. (Z ′ ξ Λσ z k+1(ζ). (6.6) (Z ′ ξ where T (k) k=0 (Z ′ = z = z Thus Λσ ξ ⊗ η, k+1(Z ′ k+1ζ ⊗ x) In particular, when ζ = ∗i · cz = (ξ∗ z ∗ · cz(k))∞ ∗ · cz(k), ηE. = (Ck+1 ⊗ IH )U σ k+1(ζ ⊗ x) = (Ck+1 ⊗ IH ) Λσ k+1(ζ)x. If ζ ∈ (Eσ)⊗k+1 and x ∈ H, then by the definition of Z ′ in terms of C, Proof. First, let us show that the result holds when D = IH. We have that (W ′ ξ maps η ∈ (Eσ)⊗k to ξ ⊗ η ∈ (Eσ)⊗k+1. Also, hξ, z k=0, so it suffices to show that for every k ∈ N0 and η ∈ (Eσ)⊗k, k+1)∗ (cz(k + 1))(cid:17)∞ k+1)∗ (cz(k + 1)) , ηE =Dξ∗ ξ)∗cz = (cid:16)T (k)∗ DT (k)∗ k+1ζ(cid:1) x = U σ k+1(cid:0)Z ′ k+1(cid:0)Z ′ k+1ζ(cid:1) = (Ck+1 ⊗ IH) Λσ (6.7) DT (k)∗ k+1)∗ (cz(k + 1)) , ηE =(cid:10)cz(k + 1), Z ′ k+1(ξ ⊗ η)(cid:11) (k+1)(cid:0)(Z (k+1))−1 ⊗ IH(cid:1) Λσ k+1(ξ ⊗ η)(cid:1) k+1(cid:0)Z ′ (k+1)(cid:0)(Z (k+1))−1Ck+1 ⊗ IH(cid:1) Λσ (k+1)(cid:0)(Z (k))−1 ⊗ I1 ⊗ IH(cid:1) Λσ (k)(cid:0)(Z (k))−1 ⊗ IH(cid:1) Λσ (k)(cid:0)(Z (k))−1 ⊗ IH(cid:1) (Ik ⊗ zξ) Λσ (k+1)(cid:0)(Z (k))−1 ⊗ I1 ⊗ IH(cid:1) Λσ ∞(D) = (W ′ D∗·ξ)∗ · cz = hD∗ · ξ, z∗i · cz, as desired. From equations (6.7) and (6.8), we obtain equation (6.6). Thus, the conclusion of the lemma holds when D = IH . For arbitrary D, we D∗·ξ)∗. Therefore, by our first case, observe that (W ′ (W ′ (cid:3) ξ)∗ϕ′ ξ)∗ (D · cz) = (W ′ We are ready to present our main result, a weighted Nevanlinna-Pick interpolation theorem. We phrase and prove a more-general, matricial version of the theorem that, aside from increased notational complex- ity, poses no additional difficulty; the main result, that which gener- alizes the classic Nevanlinna-Pick interpolation theorem, occurs when ∗ · cz(k), ηE = z (6.8) Dξ∗ On the other hand, k (zξ · η) k (η) k+1 (ξ ⊗ η) . = z = z = z k+1(ξ ⊗ η). k+1(ξ ⊗ η) ξ z 42 JENNIFER R GOOD we take s = t = 1 in the statement below. While we make necessary adjustments for the weights, our proof mirrors that of the unweighted Nevanlinna-Pick interpolation result given as Theorem 5.3 in [8]. Theorem 6.5 (Weighted Nevanlinna-Pick Interpolation). Fix s, t ∈ N. Let k ∈ N. Choose {zi}k i=1 ⊆ Ms (B (H)) and {Fi}k i=1 ⊆ Ms×t (B (H)). Define A : Mk(σ(M)′) → Mk (Ms (B (H))) at [Aij]k i=1 ⊆ D(X, σ) and two collections, {Bi}k i,j=1 ∈ Mk(σ(M)′) by A(cid:0)[Aij]k i,j=1(cid:1) = (cid:2)Bi · K(s)(zi, zj)(Aij) · B∗ Then the following conditions are equivalent: j − Fi · K(t)(zi, zj)(Aij) · F ∗ . i,j=1 (1) For every n ∈ N and for any {hpi 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ H (s) and {Api 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ σ(M)′, j(cid:3)k 2 , i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Api · czi ⊗ F ∗ Api · czi ⊗ B∗ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 2 i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 where on the left-hand side of the inequality, the norm occurs in the Hilbert space F(Eσ)⊗ι(t) H (t) and on the right in F(Eσ)⊗ι(s) H (s); (2) A is completely positive; and t (3) there exists Y = [Yaf ]s a=1 f =1 ∈ Ms×t (H ∞(E, Z)) such that kY k ≤ 1 and BihcYaf (zi)is t f =1 = Fi a=1 for every i ∈ N with 1 ≤ i ≤ k. Remark 6.6. The map A defined in the statement of Theorem 6.5 does not depend on Z, so if the theorem is satisfied for one sequence of weights associated with X it holds for every sequence of weights asso- ciated with X. Proof. First let us perform a preliminary computation. Let n ∈ N and choose two collections {hpi 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ H (s) and p=1 in {Api 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ σ(M)′. Define eh := (cid:0)(hpi)k i=1(cid:1)n INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 43 ((H (s))(k))(n). Since K(zi, zj)(A∗ piAqj) = hApi · czi, Aqj · czj i, j hqj(cid:11)! Thus, condition (2) implies condition (1). = piAqj]k Api · czi ⊗ F ∗ 2 , piAqj) · B∗ p,q=1 piAqj) · F ∗ p,q=1 where the norm occurs in the Hilbert space F(Eσ) ⊗ι(t) H (t) and the in- ner product is taken in ((H (s))(k))(n). Since an analogous result follows with each Fi replaced with Bi we compute, (eh)(cid:29) i,j=1in j(cid:3)k i hpi, ι(t)(cid:0)K(zi, zj)(A∗ piAqj)(cid:1) F ∗ i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:28)eh,h(cid:2)Fi · K(t)(zi, zj)(A∗ nXp,q=1 kXi,j=1(cid:10)F ∗ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 Deh, An(cid:16)(cid:2)[A∗ p,q=1(cid:17) (eh)E i,j=1(cid:3)n =(cid:28)eh,h(cid:2)Bi · K(s)(zi, zj)(A∗ (eh)(cid:29) i,j=1in j(cid:3)k −(cid:28)eh,h(cid:2)Fi · K(t)(zi, zj)(A∗ i,j=1in j(cid:3)k =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 nXp=1 kXi=1 completely positive. Let A :=(cid:2)δp=1 [δi=1 Aqj]k i,j=1(cid:3)n i,j=1(cid:3)n when i = 1 and is otherwise equal to 0. Then A∗A =(cid:2)[A∗ An(A∗A) is positive in B(cid:0)((H (s))(k))(n)(cid:1). Taking eh := (cid:0)(hpi)k i=1(cid:1)n 0 ≤Deh, An(A∗A)(eh)E =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Let us show that condition (2) implies condition (1). Suppose A is p,q=1 where δi=1 is 1 p,q=1 is positive in Mn(Mk(σ(M)′)). Since An is positive, we have that in ((H (s))(k))(n) and using equation (6.9) we have that nXp=1 kXi=1 2 i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 piAqj) · F ∗ p,q=1 Api · czi ⊗ B∗ Api · czi ⊗ F ∗ Api · czi ⊗ B∗ Api · czi ⊗ F ∗ 2 . p=1 2 . (eh)(cid:29) piAqj]k (6.9) 2 44 JENNIFER R GOOD piAqj]k Now let us assume condition (1) and show that condition (2) holds. Let n ∈ N and choose a collection {Api 1 ≤ i ≤ k, 1 ≤ p ≤ n} of ele- p=1 for some {hpi 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ H (s). Since we are assuming condition (1), we deduce from equation (6.9) that ments in σ(M)′. An arbitrary element of B(cid:0)((H (s))(k))(n)(cid:1) may be writ- ten eh = (cid:0)(hpi)k i=1(cid:1)n Deh, An(cid:16)(cid:2)[A∗ p,q=1(cid:17) (eh)E i,j=1(cid:3)n =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 nXp=1 kXi=1 is nonnegative. Therefore An(cid:16)(cid:2)[A∗ i,j=1(cid:3)n i,j=1(cid:3)n a sum of elements of the form(cid:2)[A∗ an arbitrary positive element in Mn(Mk(σ(M)′)) can be expressed as , it follows that An is positive. Therefore, A is completely positive; thus condition (1) implies condition (2). i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p,q=1(cid:17) is positive. Since Api · czi ⊗ B∗ Api · czi ⊗ F ∗ piAqj]k piAqj]k p,q=1 2 2 i , and Lt Let us pause for some notational simplifications. i denote the insertion operators LH czi If 1 ≤ i ≤ k, then Li, Ls , and LH (t) , respectively. Observe that Li is Lzi in the notation of Lemma czi , LH (s) 6.3. Also, LI , Ls , cIH and LH (t) , respectively. Since the identity operators on E⊗s and E⊗t cIH do not occur in our computations for the remainder of the proof, we temporarily write Is in place of IH (s) and It in place of IH (t). Note that when A ∈ σ(M)′, 1 ≤ i ≤ k, and h ∈ H (s), I denote the insertion operators LH cIH I, and Lt , LH (s) czi (6.10) A · czi ⊗ h = (ϕ′ ∞(A) ⊗ Is) (czi ⊗ h) = (ι(s))F(Eσ)(ϕ′ ∞(A))Ls i h. A similar result holds with t in place of s. Suppose Y = [Yaf ] satisfies the properties in condition (3), using a ∈ {1, . . . , s} to indicate indicate rows and f ∈ {1, . . . , t} to indicate columns for an s×t matrix. We show condition (1). Let i ∈ {1, . . . , k}. Since Ls i are diagonal operators with Li in each diagonal entry, by Lemma 6.3(3), i and Lt FiLt∗ i = BihcYaf (zi)i Lt∗ ii i = BihcYaf (zi)L∗ = Bi [L∗ i πσ(Yaf )] = BiLs∗ i πσ s×t (Y ) . INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 45 Taking adjoints, we have Lt (6.10) and Corollary 5.4, for 1 ≤ p ≤ n, iF ∗ i B∗ i . Thus by equation Api · czi ⊗ F ∗ Since kY k ≤ 1 and πσ nXp=1 kXi=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞(Api))Lt iF ∗ i hpi (ι(s))F(Eσ)(ϕ′ s×t is a linear isometry, i =(cid:0)πσ i hpi = (ι(t))F(Eσ)(ϕ′ = (ι(t))F(Eσ)(ϕ′ s×t(Y )(cid:1)∗ Ls ∞(Api))(cid:0)πσ s×t (Y )(cid:1)∗ =(cid:0)πσ =(cid:0)πσ s×t (Y )(cid:1)∗ (Api · czi ⊗ B∗ i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) s×t (Y )(cid:13)(cid:13)2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 ≤(cid:13)(cid:13)πσ ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXp=1 kXi=1 Api · czi ⊗ F ∗ 2 Api · czi ⊗ B∗ s×t (Y )(cid:1)∗ Ls ∞(Api))Ls i hpi) . i B∗ i B∗ i hpi i hpi 2 i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Api · czi ⊗ B∗ 2 . i hpi!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Therefore condition (3) implies condition (1). Finally, we show that condition (1) implies condition (3). Let JB be the norm-closure of the subspace of F(Eσ) ⊗ι(s) H (s) comprised of the elements of the form nXp=1 kXi=1 Api · czi ⊗ B∗ i hpi! i=1 Api · czi ⊗ F ∗ for some n ∈ N and some collections {hpi 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ H (s) and {Api 1 ≤ i ≤ k, 1 ≤ p ≤ n} ⊆ σ(M)′. Similarly, define JF to be the norm-closure of the subspace of F(Eσ) ⊗ι(t) H (t) consisting i hpi(cid:17). As we are of the elements of the formPn p=1(cid:16)Pk assuming condition (1), it follows that there is a well-defined, contrac- tive linear map R : JB → JF such that for any 1 ≤ i ≤ k, A ∈ σ(M)′, and h ∈ H (s), R (A · czi ⊗ B∗ i h) = A · czi ⊗ F ∗ i h. We aim to apply Corollary 4.3, corollary to the weighted commutant lifting theorem, with σ(M)′, Eσ, X ′, and Z ′ replacing M, E, X, and Z. In the statement of that corollary, we take σ1 = ι(t) on H (t), σ2 = ι(s) on H (s), J1 = JF in F(Eσ) ⊗ι(t) H (t), and J2 = JB in F(Eσ) ⊗ι(s) H (s). Let VF and VB denote the inclusion maps of JF into F(Eσ) ⊗ι(t) H (t) 46 JENNIFER R GOOD and JB into F(Eσ) ⊗ι(s) H (s), respectively. Let G = R∗ in B(JF , JB). Let us show that for any S ∈ H ∞(Eσ, Z ′), (i) (S∗ ⊗ It) (JF ) ⊆ JF , (ii) (S∗ ⊗ Is) (JB) ⊆ JB, and (iii) G (V ∗ F (S ⊗ It) VF ) = (V ∗ B (S ⊗ Is) VB) G. It suffices to show that each of these properties holds when S = ϕ′ ∞(A) ξ for ξ ∈ Eσ. When D ∈ σ(M)′, for A ∈ σ(M)′ and when S = W ′ ∞(A)∗ ⊗ It) (D · czi ⊗ F ∗ i h) = A∗D· 1 ≤ i ≤ k, and h ∈ H (s), we have (ϕ′ czi ⊗ F ∗ i h) = hD∗ · ξ, z∗ i h. Property (i) now follows from the definition of JF ; property (ii) holds by similar reasoning. To obtain property (iii), we observe that if D ∈ σ(M)′, 1 ≤ i ≤ k, and h ∈ H (s), i h. Moreover, by Lemma 6.4, (cid:0)(W ′ ξ)∗ ⊗ It(cid:1) (D · czi ⊗ F ∗ i i · czi ⊗ F ∗ (V ∗ B(ϕ′ ∞(A) ⊗ Is)VBG)∗(D · czi ⊗ B∗ i h) = RV ∗ ∞(A∗) ⊗ Is)VB(D · czi ⊗ B∗ i h) = (A∗D) · czi ⊗ F ∗ B(ϕ′ i h = V ∗ F (ϕ′ = (GV ∗ ∞(A) ⊗ It)VF )∗ = (V ∗ F (ϕ′ B(ϕ′ F (ϕ′ Thus (GV ∗ adjoints, we obtain property (iii) for S = ϕ′ 6.4, ∞(A∗) ⊗ It)VF R(cid:16)D · czi ⊗ B∗ i h(cid:17) ∞(A) ⊗ It)VF )∗(D · czi ⊗ B∗ ∞(A) ⊗ Is)VBG)∗, and by taking ∞(A). Finally, by Lemma i h) i h) i h) = V ∗ (D · czi ⊗ B∗ (D · czi ⊗ B∗ ∗ i i · czi ⊗ F ∗ i h (cid:0)V ∗ B(cid:0)W ′ = RV ∗ i h) = hD∗ · ξ, z i h) adjoints, we obtain property (iii) for S = W ′ Having satisfied the hypothesis of Corollary 4.3, we conclude that ξ ⊗ Is(cid:1) VBG(cid:1)∗ B(cid:0)(W ′ ξ)∗ ⊗ Is(cid:1) VB (D · czi ⊗ B∗ F(cid:0)(W ′ ξ)∗ ⊗ It(cid:1) VF R (D · czi ⊗ B∗ F(cid:0)W ′ =(cid:0)GV ∗ ξ ⊗ It(cid:1) VF(cid:1)∗ F(cid:0)W ′ Therefore(cid:0)GV ∗ B(cid:0)W ′ ξ ⊗ It(cid:1) VF(cid:1)∗ =(cid:0)V ∗ ξ ⊗ Is(cid:1) VBG(cid:1)∗, and by taking there exists an operator eG ∈ B(cid:0)F(Eσ) ⊗ι(t) H (t), F(Eσ) ⊗ι(s) H (s)(cid:1) such (I) eG∗(JB) ⊆ JF , BeGVF = R∗, (III) eG (S ⊗ It) = (S ⊗ Is)eG for all S ∈ H ∞(Eσ, Z ′), and (IV) keGk = kRk. Y = [Yaf ] in Ms×t (H ∞ (E, Z)) such that eG = πσ Property (III) and Corollary 5.4 together imply the existence of some s×t is an isometry, property (IV) implies that kY k = kπσ s×t(Y ). Since πσ ξ, as desired. (II) V ∗ that s×t(Y )k = keGk = (II), eG∗Ls iF ∗ Lt (6.11) i = eG∗PBLs i . Taking adjoints, we obtain F )eG∗(VBV ∗ i eG = FiLt∗ i . BiLs∗ By Lemma 6.3(2), (6.12) i πσ (Yaf ) LI] = Ls∗ hcYaf (zi)i = [L∗ BihcYaf (zi)i = BiLs∗ i eGLt i Lt I. i eGLt I = FiLt∗ i Lt I = Fi. Since czi(0) = IH, we have Lt∗ (6.11), I = It. Thus, by equations (6.12) and INTERPOLATION AND COMMUTANT LIFTING WITH WEIGHTS 47 kRk ≤ 1. To complete the proof, let PF ∈ B(F(Eσ) ⊗ι(t) H (t)) denote the projection map onto JF , and let PB ∈ B(F(Eσ) ⊗ι(s) H (s)) denote the projection map onto JB. Then VF V ∗ B = PB. For h ∈ H (s), Ls i h. It follows that Ls i = PBLs i . Thus by properties (I) and i B∗ i = F = PF and VBV ∗ i h = czi ⊗ F ∗ iF ∗ i h = czi ⊗ B∗ i and VF RV ∗ i h and Lt BLs iF ∗ i = Lt i = VF RV ∗ i = (VF V ∗ i B∗ i B∗ B)Ls i B∗ BLs i B∗ i B∗ i B∗ i B∗ Thus condition (1) implies condition (3), which completes the proof. (cid:3) Remark 6.7. In the scalar case, when M = E = H = C, our kernel K is simply the reproducing kernel to a weighted Hardy space. Be- cause one of the hypothesis of the admissible sequence X is that each Xk is positive, it follows from Theorem 7.33 of [1] that the weighted Hardy spaces under consideration are those that satisfy the complete Pick Property that is described in Definition 5.13 in [1]. Thus, for example, our theorem applies to the Hardy and Dirichlet spaces, but not the Bergman space, by Corollary 7.37, Corollary 7.41, and Exam- ple 5.17 of [1]. The original complete Pick property involves certain matrix-valued multipliers, and while we will not include the discussion in the present paper, notions such as reproducing kernel Hilbert spaces and their spaces of multipliers can be generalized to the W ∗-setting, which produces interesting examples of noncommutative functions. A promising topic of future work involves the formulation of a W ∗-version of the complete Pick property and an investigation of its implications for the representation theory of H(E, Z), as was begun in [9]. Acknowledgements Thank you to Paul Muhly and Baruch Solel for their support, en- couragement, and helpful feedback in this endeavor. 48 JENNIFER R GOOD References [1] Jim Agler and John E. McCarthy. Pick interpolation and Hilbert function spaces, volume 44 of Graduate Studies in Mathematics. American Mathemati- cal Society, Providence, RI, 2002. [2] Alvaro Arias. Projective modules on Fock spaces. J. Operator Theory, 52(1):139 -- 172, 2004. [3] Michel Baillet, Yves Denizeau, and Jean-Fran¸cois Havet. Indice d'une esp´erance conditionnelle. Compositio Math., 66(2):199 -- 236, 1988. [4] Stephen D. Barreto, B. V. Rajarama Bhat, Volkmar Liebscher, and Michael Skeide. Type I product systems of Hilbert modules. J. Funct. Anal., 212(1):121 -- 181, 2004. [5] E. C. Lance. Hilbert C ∗-modules, volume 210 of London Math. Soc. Lecture Note Ser. Cambridge University Press, 1995. [6] Jonas R. Meyer. Noncommutative Hardy Algebras, Multipliers, and Quotients. PhD thesis, The University of Iowa, 2010. [7] Paul S. Muhly and Baruch Solel. Tensor algebras over C ∗-correspondences: representations, dilations, and C ∗-envelopes. J. Funct. Anal., 158(2):389 -- 457, 1998. [8] Paul S. Muhly and Baruch Solel. Hardy algebras, W ∗-correspondences and interpolation theory. Math. Ann., 330(2):353 -- 415, 2004. [9] Paul S. Muhly and Baruch Solel. Matricial function theory and weighted shifts. Integral Equations Operator Theory, 84(4):501 -- 553, 2016. [10] R. Nevanlinna. Uber beschrankte funktionen, die in gegebenen punkten vorgeschrieben werte annehmen. Ann. Acad. Sci. Fenn. Ser. A, 13(1), 1919. [11] Stephen Parrott. On a quotient norm and the Sz.-Nagy - Foia¸s lifting theorem. J. Funct. Anal., 30(3):311 -- 328, 1978. [12] William L. Paschke. Inner product modules over B ∗-algebras. Trans. Amer. Math. Soc., 182:443 -- 468, 1973. [13] Georg Pick. Uber die Beschrankungen analytischer Funktionen, welche durch vorgegebene Funktionswerte bewirkt werden. Mathematische Annalen, 77(1):7 -- 23, 1915. [14] Gelu Popescu. Operator theory on noncommutative domains. Mem. Amer. Math. Soc., 205(964):vi+124, 2010. [15] Marc A. Rieffel. Induced representations of C ∗-algebras. Advances in Math., 13:176 -- 257, 1974. [16] Donald Sarason. Generalized interpolation in H ∞. Trans. Amer. Math. Soc., 127:179 -- 203, 1967. [17] Allen L. Shields. Weighted shift operators and analytic function theory. In Topics in operator theory, pages 49 -- 128. Math. Surveys, No. 13. Amer. Math. Soc., Providence, R.I., 1974. [18] B´ela Sz.-Nagy and Ciprian Foia¸s. Dilatation des commutants d'op´erateurs. C. R. Acad. Sci. Paris S´er. A-B, 266:A493 -- A495, 1968. Department of Mathematics, University of Wisconsin - Platteville, Platteville, WI 53818 E-mail address: [email protected]
1701.08922	1	1701	2017-01-31T05:32:09	Hardy-Littlewood inequalities on compact quantum groups of Kac type	[ "math.OA" ]	The Hardy-Littlewood inequality on $\mathbb{T}$ compares the $L^p$-norm of a function with a weighted $\ell^p$-norm of its Fourier coefficients. The approach has recently been studied for compact homogeneous spaces and we study a natural analogue in the framework of compact quantum groups. Especially, in the case of the reduced group $C^$-algebras and free quantum groups, we establish explicit $L^p-\ell^p$ inequalities through inherent information of underlying quantum group, such as growth rate and rapid decay property. Moreover, we show sharpness of the inequalities in a large class, including $C(G)$ with compact Lie group, $C_r^(G)$ with polynomially growing discrete group and free quantum groups $O_N^+$, $S_N^+$.	math.OA	math	HARDY-LITTLEWOOD INEQUALITIES ON COMPACT QUANTUM GROUPS OF KAC TYPE SANG-GYUN YOUN Abstract. The Hardy-Littlewood inequality on T compares the Lp-norm of a function with a weighted ℓp-norm of its Fourier coefficients. The approach has recently been studied for compact homogeneous spaces and we study a natu- ral analogue in the framework of compact quantum groups. Especially, in the case of the reduced group C ∗-algebras and free quantum groups, we establish explicit Lp − ℓp inequalities through inherent information of underlying quan- tum group, such as growth rate and rapid decay property. Moreover, we show sharpness of the inequalities in a large class, including C(G) with compact Lie group, C ∗ r (G) with polynomially growing discrete group and free quantum N , S+ groups O+ N . 1. Introduction Hardy and Littlewood [14] showed that, for each 1 < p ≤ 2, there exists a constant Cp such that (1.1) 1 1 (Xn∈Z (1 + n)2−p bf (n)p) p ≤ Cpf ∼Xn∈Zbf (n)znLp(T) for all f ∈ Lp(T). This implies that the multiplier Fw : Lp(T) → lp(Z), f 7→ (w(n)bf (n))n∈Z, with is bounded. Moreover, this is a stronger form of the Sobolev 2−p p w(n) := 1 (1 + n) embedding theorem 1 p − 1 q H p (T) ⊆ Lq(T), ∀1 < p < q < ∞, where H s p(T) :=(cid:8)f ∈ Lp(T) : (1 − ∆) s 2 (f ) ∈ Lp(T)(cid:9) is the Bessel potential space. "The Hardy-Littlewood inequality (1.1)" had been studied for compact abelian groups by Hewitt and Ross [15], and was recently extended to compact homogeneous manifolds by Akylzhanov, Nursultanov and Ruzhansky [[1] and [2]]. For compact Lie groups G with the real dimension n, the result of [2] is written as follows: For each 1 < p ≤ 2, there exists a constant Cp > 0 such that (1.2) p ≤ CpfLp(G) for all f ∈ Lp(G). Here, bG denotes a maximal family of mutually inequivalent irreducible unitary satisfies ∆ : πi,j 7→ −κππi,j for all π = (πi,j )1≤i,j≤nπ ∈ bG and all 1 ≤ i, j ≤ nπ. The above inequality (1.2) can be reduced to a more familiar form whose left representations of G, AHS := tr(A∗A) hand side is a natural weighted ℓp-norm of its Fourier coefficients: 2 and the Laplacian operator ∆ on G bf (π)p (Xπ∈ bG (1 + κπ) 2− p HS) n(2−p) n 1 2 2 π 1 1 (1.3) (Xπ∈ bG 1 (1 + κπ) n(2−p) 2 nπbf (π)p Sp nπ ) 1 p ≤ CpfLp(G). Key words and phrases. Hardy-Littlewood inequality, Quantum groups, Fourier analysis. 2010 Mathematics Subject Classification. Primary 20G42, Secondary 46L52 The author is supported by TJ Park Science Fellowship. 1 A notable point is that "the Hardy-Littlewood inequalities on compact Lie groups (1.2)" are determined by inherent geometric information, namely the real dimension and the natural length function on bG. Indeed, π 7→ √κπ is equivalent to the natural length · on bG (See Remark 6.1). The main purpose of this paper is to establish new Hardy-Littlewood inequal- ities on compact quantum groups of Kac type through geometric information of underlying quantum groups. As a part of efforts, we will present some explicit inequalities in important examples. The reduced group C∗-algebras C∗r (G) with discrete groups G, the free orthogonal quantum groups O+ N and the free permu- tation quantum groups S+ N are the main targets. Of course, non-commutative Lp analysis on quantum groups is widely studied from various perspectives ([7], [9], [16], [17] and [26]). For details of operator algebraic approach to quantum group itself, see [19], [20], [24] and [30]. To clarify our intention, let us show the main results of this article on "compact matrix quantum groups" which admit the natural length function · : Irr(G) → {0} ∪ N (See Definition 3.3 and Proposition 3.4). The following inequalities are determined by inherent information of underlying quantum group, namely growth rate and rapid decay property. Theorem 1.1. Let G be a compact matrix quantum group of Kac type and denote by · the natural length function on Irr(G). (1) Let G have a polynomial growth with Xα∈Irr(G):α≤k n2 α ≤ (1 + k)γ and γ > 0. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) such that ( Xα∈Irr(G) ≤( Xα∈Irr(G) for all f ∼ Xα∈Irr(G) that 1 1 p 2 α 1 ) 1 nα Sp HS ) 2− p (1 + α)(2−p)γ nαbf (α)p bf (α)p (1 + α)(2−p)γ n nαtr(bf (α)uα) ∈ Lp(G). fL∞(G) ≤ C(1 + k)βfL2(G) p ≤ KfLp(G) (2) Let G have the rapid decay property with universal constants C, β > 0 such (1.4) (1.5) i,j : α = k, 1 ≤ i, j ≤ nα(cid:9)). Define sk := Xα∈Irr(G):α=k n2 α. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) such that for all f ∈ span((cid:8)uα (Xk≥0 Xα∈Irr(G):α=k ≤(Xk≥0 for all f ∼ Xα∈Irr(G) 1 1 (2−p) 2 s k (1 + k)(2−p)(β+1) 1 p ) Sp nα nαbf (α)p p HS ) 2 ) 1 p ≤ KfLp(G) (1 + k)(2−p)(β+1) ( Xα∈Irr(G):α=k nαbf (α)2 nαtr(bf (α)uα) ∈ Lp(G). In particular, it was known that the rapid decay property of FN can be strength- ened in general holomorphic setting [18]. The improved result is called the strong Haagerup inequality. Based on this data, we can improve the Hardy-Littlewood inequality for C∗r (FN ) by focusing our attention on holomorphic forms. Theorem 2 5.3 justifies the claim and it seems appropriate to call the improved one "strong Hardy-Littlewood inequality". A natural view of the Hardy-Littlewood inequalities on compact Lie groups is multiplier that a properly chosen weight function w : bG → (0,∞) makes the corresponding Fw : Lp(G) → ℓp(bG) given by f 7→ (w(π)bf (π))π∈ bG be bounded for each 1 < p ≤ 2. Indeed, newly derived Hardy-Littlewood inequali- ties on compact quantum groups will give a decay pair (r, s) whose corresponding multiplier Fwr,s is bounded, where wr,s(α) := rα(1 + α)s . Moreover, in Section 6, we will show that there is no slower decay pair such that Fr,s is bounded when G is one of the followings: C(G) with compact Lie groups, C∗r (G) with polynomially growing discrete group and free quantum groups O+ N . See Theorem 6.6. N , S+ 1 This approach is quite natural because it is strongly related to Sobolev embedding properties. Indeed, for the case G = Td, Fw0,s : Lp(Td) → lp(Zd) is bounded if and only if ps 2−p ( 1 q q − 1 r ) H (Td) ⊆ Lr(Td) ∀1 < q < r < ∞, where H s p(Td) is the Bessel potential space. Lastly, in Section 7, we present some remarks that are by-products of this re- search. We show that most of free quantum groups do not admit an infinite (central) sidon set, and also give a Sobolev embedding theorem type interpretation of our results for C(G) with compact Lie group and C∗r (G) with polynomially growing discrete group G. Also, we present an explicit inequality on quantum torus Td θ. 2. Preliminaries 2.1. Compact quantum groups. A compact quantum group G is given by a unital C∗-algebra A and a unital ∗-homomorphism ∆ : A → A ⊗min A satisfying (1) (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆; (2) span{∆(a)(b ⊗ 1A) : a, b ∈ A} , span{∆(a)(1A ⊗ b) : a, b ∈ A} are dense in A. Every compact quantum group admits the unique Haar state h on A such that (h ⊗ id)(∆(x)) = h(x)1A = (id ⊗ h)(∆(x)) for all x ∈ A. A finite dimensional corepresentation of G is given by an element u = (ui,j)1≤i,j≤n ∈ Mn(A) such that ∆(ui,j) = nXk=1 i,j)1≤i,j≤nα(cid:9)α∈Irr(G) ui,k ⊗ uk,j for all 1 ≤ i, j ≤ n. We say that the corepresentation u is unitary if u∗u = uu∗ = Idn ⊗ 1A ∈ Mn(A) and irreducible if {X ∈ Mn : Xu = uX} = C · Idn where Idn is the identity matrix in Mn. be a maximal family of mutually inequivalent finite dimensional unitary irreducible corepresentations of G. It is well known that, for each α ∈ Irr(G), there is a unique positive invertible matrix Qα ∈ Mnα such that tr(Qα) = tr(Q−1 Let(cid:8)uα = (uα h((uβ s,t)∗uα i,j) = α ) and δα,βδj,t(Q−1 tr(Qα) α )i,s , ∀α, β ∈ Irr(G), 1 ≤ i, j ≤ nα, 1 ≤ s, t ≤ nβ, h(uβ s,t(uα i,j)∗) = δα,βδi,s(Qα)j,t tr(Qα) , ∀α, β ∈ Irr(G), 1 ≤ i, j ≤ nα, 1 ≤ s, t ≤ nβ. case, the Haar state h is tracial. We say that G is of Kac type if Qα = Idnα ∈ Mnα for all α ∈ Irr(G). In this Lastly, we define Cr(G) as the image of A in the GNS representation of the Haar state h and L∞(G) := Cr(G)′′. The Haar state h has a normal faithful extension to L∞(G). 3 2.2. Non-commutative Lp-spaces. Let M be a von Neumann algebra with a normal faithful tracial state ϕ. Note that the von Neumann algebra M admits the unique predual M∗. We define L1(M, ϕ) := M∗ and L∞(M, ϕ) := M, then consider a contractive injection j : M → M∗, given by [j(x)](y) := h(yx) for all y ∈ M. The map j has dense range. Now (M,M∗) is a compatible pair of Banach spaces and we can define non- commutative Lp-space Lp(M, ϕ) := (M,M∗) 1 is the complex interpolation space. For any x ∈ L∞(M, ϕ), its Lp-norm is given by xLp(M,ϕ) = ϕ(xp) In particular, for all 1 ≤ p ≤ ∞, we denote by Lp(G) the non-commutative Lp-space associated to the von Neumann algebra L∞(G) of a Kac type compact quantum group G and the tracial Haar state h. Then the space of polynomials for all 1 < p < ∞, where (·,·) 1 p for all 1 ≤ p < ∞. p p 1 i,j : α ∈ Irr(G), 1 ≤ i, j ≤ nα(cid:9)) P ol(G) := span((cid:8)uα (Aα)α∈Irr(G) ∈ Yα∈Irr(G) is dense in Cr(G) and Lp(G) for all 1 ≤ p < ∞. Under the assumption that G is of Kac type, for 1 ≤ p < ∞, nαtr(Aαp)) (Aα)α∈Irr(G)ℓ∞(bG) := sup Mnα : Xα∈Irr(G) (Aα)α∈Irr(G) ∈ Yα∈Irr(G) ℓp(bG) := (Aα)α∈Irr(G)ℓp(bG) := ( Xα∈Irr(G) ℓ∞(bG) := nαtr(Aαp) < ∞ and the natural ℓp-norm of (Aα)α∈Irr(G) ∈ ℓp(bG) is defined by p = ( Xα∈Irr(G) α∈Irr(G)Aα < ∞ and the ℓ∞-norm of (Aα)α∈Irr(G) ∈ ℓ∞(bG) is defined by It is known that ℓ1(bG) = (ℓ∞(bG))∗ and ℓp(bG) = (ℓ∞(bG), ℓ1(bG)) 1 group G, the Fourier transform F : L1(G) → ℓ∞(bG), ψ 7→ bψ, is defined by from L2(G) onto l2(bG) ([26], Proposition 3.1 and 3.2). Then, by the interpolation (bG) for each 1 ≤ p ≤ 2, where p′ is the conjugate of p. It is also known that F is an injective contractive map and it is an isometry theorem, we induce the Hausdorff-Young inequality again, i.e. F is a contractive map from Lp(G) into lp′ ∞. 2.3. Fourier analysis on compact quantum groups. For a compact quantum 2.4. The reduced group C∗-algebras. The reduced group C∗-algebra C∗r (G), can be defined for all locally compact groups, but we only consider discrete groups in this paper since we want to understand it as a compact quantum group. Definition 2.1. Let G be a discrete group and define λg ∈ B(l2(G)) for each g ∈ G by j,i)∗) for all α ∈ Irr(G), 1 ≤ i, j ≤ nα. (bψ(α))i,j := ψ((uα α∈Irr(G)Aα. for all 1 < p < p Also, 1 nαAαp Sp nα 1 p . ) Mnα : sup [(λg)(f )](x) := f (g−1x) for all x ∈ G. Then the reduced group C∗-algebra, C∗r (G), is defined as the norm-closure of the space span({λg : g ∈ G}) in B(l2(G)). Moreover, if we define a comultiplication ∆ : C∗r (G) → C∗r (G) ⊗min C∗r (G) by λg 7→ λg ⊗ λg for all g ∈ G, then (C∗r (G), ∆) forms a compact quantum group. 4 Note that, for G = (C∗r (G), ∆) of a discrete group G, L∞(G) is nothing but the group von Neumann algebra V N (G) and Irr(G) = {λg}g∈G is identified with G. 2.5. Free quantum groups of Kac type. Definition 2.2. (Free orthogonal quantum group [28]) Let N ≥ 2 and A be the universal unital C∗-algebra which is generated by the N 2 self-adjoint elements ui,j with 1 ≤ i, j ≤ N satisfying the relations: NXk=1 NXk=1 uk,iuk,j = ui,kuj,k = δi,j for all 1 ≤ i, j ≤ N. NXk=1 NXk=1 Also, we define a comultiplication ∆ : A → A ⊗min A by ui,j 7→ ui,k ⊗ uk,j. Then (A, ∆) forms a compact quantum group called the Free orthogonal quantum group. We denote it by O+ N . Definition 2.3. (Free permutation quantum group [29]) Let N ≥ 2 and A be the universal unital C∗-algebra which is generated by the N 2 self-adjoint elements ui,j with 1 ≤ i, j ≤ N satisfying the relations: u2 i,j = ui,j = u∗i,j and ui,k = uk,j = 1A for all 1 ≤ i, j ≤ N. NXk=1 NXk=1 Also, we define a comultiplication ∆ : A → A ⊗min A by ui,j 7→ ui,k ⊗ uk,j. Then (A, ∆) forms a compact quantum group called the Free permutation quantum group. We denote it by S+ N . These free quantum groups are of Kac type, so that the Haar states are tracial N +2) can be identified with {0} ∪ N. states. Also, for all N ≥ 2, Irr(O+ Moreover, N ) and Irr(S+ nk = k + 1 2r0−N rk+1 r0 0 + N−r0 2r0−N (N − r0)k+1 if G = O+ if G = O+ 2 or S+ 4 N or SN +2 with N ≥ 3 where r0 is the largest solution of the equation X 2−N X +1 = 0. Note that nk ≈ rk if N ≥ 3. 2.6. The noncommutative Marcinkiewicz interpolation theorem. The clas- sical Marcinkiewicz interpolation theorem has a natural non-commutative analogue for semi-finite von Neumann algebras. 0 Theorem 2.4. (The non-commutative Marcinkiewicz interpolation theorem [32]) Let M be equipped with a normal semifinite faithful trace φ and 1 ≤ p1 < p < p2 < ∞. Assume that a sub-linear map A : M → L1(N ) satisfies the following: There exists C1, C2 > 0 such that for any T1 ∈ Lp1(M ), T2 ∈ Lp2(M ) and for any y > 0, (2.1) φ(1(y,∞)(AT1)) ≤ ( C1 y C2 y Then A : Lp(M ) → Lp(N ) is a bounded map. Proof. The proof of [Theorem 1.22, [32]] is still valid under a natural modification. )p1T1p1 )p2T2p2 φ(1(y,∞)(AT2)) ≤ ( Lp1 (M), Lp2 (M). Also, the direct sum K := M ⊕ N and natural extension eA : K → L1(K) gives another proof. (cid:3) 5 If the sub-linear operator A satisfies the inequality (2.1), then we say that A is of weak type (p1, p1). Also, the boundedness of A : Lp(M ) → lp(N ) implies that A is of weak type (p, p). Now denote by c(Irr(G), ν) the space of all functions on the discrete space Irr(G) with a positive measure ν. Then the above theorem is written as follows: Corollary 2.5. Let G be a Kac type compact quantum group and let 1 ≤ p1 < p < p2 < ∞. Assume that A : L∞(G) → c(Irr(G), ν) is sub-linear and satisfies the following: There exists C1, C2 > 0 such that for any T1 ∈ Lp1(G), T2 ∈ Lp2 (G) and for any y > 0, Xα: (AT1)(α)≥y Xα: (AT2)(α)≥y ν(α) ≤ ( C1 y )p1T1p1 Lp1 (G), ν(α) ≤ ( C2 y )p2T2p2 Lp2 (G). Then A : Lp(G) → ℓp(Irr(G), ν) is a bounded map. 3. Paley-type inequalities 3.1. General Approach. In this subsection, we derive a Paley-type inequality for Kac type compact quantum group G via fundamental techniques such as Hausdorff- Young inequality, Plancherel theorem and the non-commutative Marcinkiewicz in- terpolation theorem. We prove the following theorem by adapting techniques used in [2]. Theorem 3.1. Let G be a Kac type compact quantum group and let w : Irr(G) → < ∞. Then, for each (0,∞) be a function such that Cw := sup 1 < p ≤ 2, there exists a universal constant K = K(p) > 0 such that t · Xα: w(α)≥t n2 α t>0 bf (α)p HS ) 1 p ≤ KfLp(G) 2− p 2 α (3.1) w(α)2−pn for all f ∼ Xα∈Irr(G) ( Xα∈Irr(G) nαtr(bf (α)uα) ∈ Lp(G). L1(G) → c(Irr(G), ν), f 7→ (bf (α)HS Proof. Put ν(α) := w(α)2n2 √nαw(α) from Lp(G) into ℓp(Irr(G), ν) for all 1 < p ≤ 2. First, α. We will show that the sub-linear operator A : )α∈Irr(G), is a well-defined bounded map Xα∈Irr(G) (Af )(α)2 HS ν(α) = Xα∈Irr(G) HS = f2 L2(G). nαbf (α)2 This implies that A is of (strong) type (2, 2) with C2 = 1. 6 y √nα Second, for all y > 0, since bf (α)HS ν(α) ≤ Xα: w(α)≤ f 1 Xα: Af (α)HS≥y = Xα: w(α)≤ f 1 =Z ( f 1 = 2Z f 1 ≤ 2Cw f1 )2 y 0 0 . y y αdx 0 n2 y Z w(α)2 [ Xx t( Xα: t≤w(α)≤ f 1 2 ≤w(α)≤ f 1 1 y y = ( tr(bf (α)∗bf (α)) nα w(α)2n2 α ) 1 2 ≤ bf (α) ≤ fL1(G), n2 α]dx by the Fubini theorem n2 α)dt by substituting x to t2 This says that A is of weak type (1, 1) with C1 = 2Cw. Now, by Corollary 2.5, w(α)2−pn p( 2 α p − 1 2 ) [ Xα∈Irr(G) 1 p . fLp(G). HS ] bf (α)p (cid:3) The left hand side of the inequality (3.1) can be reduced to a more familiar form. Recall that the natural non-commutative ℓp-norm on ℓ∞(bG) = ℓ∞ − ⊕α∈Irr(G)Mnα is given by 1 p (Aα)α∈Irr(G)ℓp(bG) = ( Xα∈Irr(G) nαAαp Sp nα ) under the condition that G is of Kac type. Corollary 3.2. Let 1 < p ≤ 2 and w be a function satisfying the condition of Theorem 3.1. Then we have that ( Xα∈Irr(G) w(α)2−pnαbf (α)p nαtr(bf (α)uα) ∈ Lp(G). for all f ∼ Xα∈Irr(G) Proof. First, Sp nα )) 1 p ≤ KfLp(G) Put 1 r = 1 p − 1 2 . Then 2 < r ≤ ∞ and tr(bf (α)p) ≤ bf (α)p This completes proof easily. . Sp nα tr(bf (α)p) = bf (α)p HSIdnαp Sr nα = n 1− p 2 α HS . bf (α)p (cid:3) Now we discuss an important subclass of compact quantum groups, namely com- pact matrix quantum groups which allows the natural length function on Irr(G). Definition 3.3. A compact matrix quantum group is given by a pair (A, ∆, u) with a unital C∗-algebra A, a ∗-homomorphism ∆ : A → A ⊗min A and a unitary u = 7 (ui,j)1≤i,j≤n ∈ Mn(A) such that (1) ∆ : ui,j 7→ is invertible in Mn(A) and (3) {ui,j}1≤i,j≤n generates A as a C∗-algebra. ui,k ⊗ uk,j, (2) u = (u∗i,j)1≤i,j≤n nXk=1 By definition, free orthogonal quantum groups O+ N and free permutation quantum groups S+ N are compact matrix quantum groups. Also, in the class of compact quantum groups, the subclass of compact matrix quantum groups is characterized by the following proposition. The conjugate α ∈ Irr(G) of α ∈ Irr(G) is determined by uα := Q α uαQ− 1 α = ((Qα) i,j)∗(Qα)− 1 i,i(uα 1 2 1 2 2 2 j,j )1≤i,j≤nα . Proposition 3.4. ([24]) A compact quantum group is a compact matrix quantum group if and only if there exists a finite set S := {α1,··· , αn} ⊆ Irr(G) such that any α ∈ Irr(G) is contained in some iterated tensor product of elements α1, α1,··· , αn, αn and the trivial corepresentation. Then there is a natural way to define a length function on Irr(G) ([25]). For non-trivial α ∈ Irr(G), the natural length α is defined by min{m ∈ {0} ∪ N : ∃β1,··· , βm such that α ⊆ β1 ⊗ ··· ⊗ βm, βj ∈ {αk, αk}n The length of the trivial corepresentation is defined by 0. k=1} . Then we can extract explicit inequalities from Theorem 3.1 and Corollary 3.2 by inserting geometric information of underlying quantum group, namely growth rate that is estimated by the quantities bk := Xα≤k n2 α [6]. Corollary 3.5. Let a Kac type compact matrix quantum group G satisfy n2 α ≤ C(1 + k)γ for all k ≥ 0 with C, γ > 0 bk = Xα∈Irr(G):α≤k with respect to the natural length function. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) such that HS ) 1 p ≤ KfLp(G) 1 p ) nα 1 1 2 α Sp 2− p (3.2) ( Xα∈Irr(G) ≤( Xα∈Irr(G) for all f ∼ Xα∈Irr(G)  t · Xα:α≤t (1 + α)(2−p)γ nαbf (α)p bf (α)p (1 + α)(2−p)γ n nαtr(bf (α)uα) ∈ Lp(G).  α Proof. Consider a weight function w(α) := t · Xα:α≤t = sup 0<t≤1 sup t>0 γ −1 γ −1 n2 − 1 − 1 1 (1 + α)γ . Then n2 α ≤ C sup 0<t≤1 t · (t− 1 γ )γ = C. (cid:3) Now the conclusion is obtained by Theorem 3.1 and Proposition 3.2. 3.2. A paley-type inequality under the rapid decay property. In this sub- section, we still assume that G is a compact matrix quantum group of Kac type. One of the main observations of this paper is that the more detailed geometric infor- mation actually improves Theorem 3.1 and Corollary 3.2 in various "exponentially growing" cases. A more refined paley-type inequality can be obtained under the condition that G has the rapid decay property in the sense of [25]. 8 Definition 3.6. ([25]) Let G be a Kac type compact matrix quantum group. Then we say that G has the rapid decay property with respect to the natural length function on Irr(G) if there exist C, β > 0 such that (3.3) Xα∈Irr(G):α=k nαXi,j=1 aα i,juα i,jL∞(G) ≤ C(1 + k)β Xα∈Irr(G):α=k nαXi,j=1 aα i,juα i,jL2(G) for any k ≥ 0 and scalars aα i,j ∈ C. Notation 1. (1) When the natural length function on Irr(G) is given, we will (2) We denote by pk the orthogonal projection from L2(G) to the clousre of use the notations Sk := {α ∈ Irr(G) : α = k} and sk :=Pα∈Sk span((cid:8)uα i,j : α ∈ Sk, 1 ≤ i, j ≤ nα(cid:9)). Proposition 3.7. Suppose that a Kac type compact matrix quantum group G has the rapid decay property with respect to the natural length function on Irr(G) and with inequality (3.3). Then we have that n2 α. (3.4) sup k≥0 (Pα∈Irr(G):α=k nαbf (α)2 (k + 1)β HS ) 1 2 ≤ CfL1(G) for all f ∈ L1(G). Proof. Since L1(G) is isometrically embedded into the dual space M (G) := Cr(G)∗ and P ol(G) is dense in Cr(G), we have that fL1(G) = = sup = ≥ (3.5) sup x∈P ol(G):xL∞(G)≤1hf, xiL1(G),L∞(G) x∈P ol(G):xL∞(G)≤1hf, x∗iL1(G),L∞(G) x∈P ol(G):xL∞(G)≤1 Xα∈Irr(G) sup sup nαtr(bf (α)bx(α)∗) 2 ≤1 Xα∈Irr(G):α=k HS) 1 1 x∈P ol(G):Pk≥0 C(k+1)β (Pα:α=k nαbx(α)2 ≥ sup k≥0 = sup k≥0 sup (Pα:α=k nαbx(α)2 HS) (Pα∈Irr(G):α=k nαbf (α)2 C(k + 1)β HS ) 1 2 . 2 ≤1 Xα∈Irr(G) nαtr(bf (α)bx(α)∗) C(k + 1)β tr(bf (α)bx(α)∗) nα (cid:3) Theorem 3.8. Let a Kac type compact matrix quantum group G have the rapid de- cay property with respect to the natural length function on Irr(G) and with inequality (3.3). Also, suppose that a weight function w : {0} ∪ N → (0,∞) satisfies (3.6) (Xk≥0 for all f ∼ Xα∈Irr(G) (k + 1)2β < ∞. w(k) ≤ 1 y Cw := sup y>0  y · Xk≥0: (k+1)β w(k)2−p( Xα∈Irr(G):α=k nαtr(bf (α)uα) ∈ Lp(G). 9 HS ) nαbf (α)2 Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) > 0 such that (3.7) 2 ) p 1 p ≤ KfLp(G) Proof. Put ν(k) := w(k)2. We will show that the sub-linear operator A : L1(G) → c({0} ∪ N, ν), f 7→ (pk(f )L2(G) )k≥0, is a well-defined bounded map from Lp(G) into lp({0} ∪ N, ν) for all 1 < p ≤ 2. w(k) Firstly, (pk(f )L2(G) w(k) Xk≥0 )2ν(k) =Xk≥0 pk(f )2 L2(G) = f2 L2(G). Therefore, A is of (weak) type (2, 2) with C2 = 1. Secondly, for all y > 0, Xk≥0:(Af )(k)>y ν(k) ≤ Xk: w(k) (k+1)β < Cf L1 (G) y w(k)2 by Proposition 3.7. w(k) (k + 1)β . Then Now put ew(k) := = Cf Xk: ew(k)< ≤Z ( = 2Z Cf 0 0 Cf L1 (G) y L1(G) y )2 L1 (G) y 0 Z ew(k)2 Xk:√x≤ ew(k) t · Xk:t≤ ew(k) 2CwCfL1(G) . y ≤ (k + 1)2βdx (k + 1)2βdx (k + 1)2βdt by substituting x = t2 Therefore, by Corollary 2.5, we obtain that (Xk≥0 w(k)2−ppk(f )p L2(G)) 1 p . fLp(G). (cid:3) Corollary 3.9. Let a Kac type compact matrix quantum group G have the rapid de- cay property with respect to the natural length function on Irr(G) and with inequality (3.3). Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) > 0 such that 1 (3.8) (Xk≥0 for all f ∼ Xα∈Irr(G) (1 + k)(2−p)(β+1) ( Xα∈Irr(G):α=k nαtr(bf (α)uα) ∈ Lp(G). 10 HS ) nαbf (α)2 p 2 ) 1 p ≤ KfLp(G) Proof. Take w(k) := (1 + k)β+1 and E = Irr(G). Then 1 Cw = sup y y>0 (1 + k)2β y · Xk≥0:(1+k)2β+1≤ 1 0<y≤1 t2βdt y ·Z ( 1 ) 0<y≤1(y · (2 · ( 1 y ) 2β + 1 2β+1 )2β+1 2β+1 +1 y ) 1 1 1 ≤ sup ≤ sup 2β+1 2β + 1 = < ∞. (cid:3) Corollary 3.10. Let a Kac type compact matrix quantum group G have the rapid decay property with respect to the natural length function on Irr(G) and with inequal- ity (3.3). Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) > 0 such that (3.9) (Xk≥0 Xα∈Irr(G):α=k for all f ∼ Xα∈Irr(G) 1 (1 + α)(2−p)(β+1)(Pβ∈Sk nαtr(bf (α)uα) ∈ Lp(G). Proof. Since = + 1 p 1 2 2 − p 2p and n− 1 Xα∈Sk Sp nαbf (α)p nα ≤ Xα∈Sk 2−p 2 n2 β) nαbf (α)p Sp nα ) 1 p ≤ KfLp(G) nα ≤ n− 1 2 2 2 p α 1 2 n HS 2− p α bf (α)Sp bf (α)p αbf (α)HS · n αbf (α)HS )α∈Skp · (Xα∈Sk p −1 α n2 α) 2−p 1 2 2 = (n ≤ (n = (Xα∈Sk α bf (α)HS , we have that ℓp(Sk) )α∈Skp ℓ2(Sk) · (Xα∈Sk nαbf (α)2 HS ) 2−p 2 n2 α) p 2 . (cid:3) Then we obtain the conclusion. 4. Hardy-Littlewood inequalities This section is devoted to establish explicit Hardy-Littlewood inequalities for the main targets: the reduced group C∗-algebras C∗r (G) with finitely generated discrete group G, free orthogonal quantum groups O+ N and free permutation quantum groups S+ N . 4.1. The reduced group C∗-algebras C∗r (G). In this subsection, we treat finitely generated discrete groups G. As expected, we found clear evidence that the geo- metric information of the underlying group is of significant importance for under- standing non-commutative Lp-spaces Lp(V N (G)). Definition 4.1. A discrete group with a fixed finite symmetric generating set S is said to be polynomially growing if there exist C > 0 and k > 0 such that #{g ∈ G : g ≤ n} ≤ C(1 + n)k for all n ≥ 0. 11 In this case, the polynomial growth rate k0 is defined as the minimum of such k. Then k0 becomes a natural number and is independent of the choice of generating set S. Theorem 4.2. (1) Let G be a finitely generated discrete group, which has the polynomial growth rate k0. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) such that (4.1) (Xg∈G 1 (1 + g)(2−p)k0 f (g)p) 1 p ≤ Kλ(f )Lp(V N (G)) for all λ(f ) ∼Xg∈G f (g)λg ∈ Lp(V N (G)). (2) Let G be a finitely generated discrete group with bk = #{g ∈ G : gS ≤ k} ≤ Crk for all k ≥ 0, where · S is the natural length function with respect to a finite symmetric generating set S. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p, S) > 0 such that (4.2) (Xg∈G for all λ(f ) ∼Xg∈G 1 r(2−p)g f (g)p) 1 p ≤ Kλ(f )Lp(V N (G)) f (g)λg ∈ Lp(V N (G)). Proof. (1) Clear from Corollary 3.5. (2) Consider w(g) := . Then 1 rg sup t>0 t · Xg≤logr ( 1 t ) 1 = sup 0<t≤1 t · Xg≤logr( 1 t ) 1 ≤ C. (cid:3) Then the conclusion follows from Theorem 3.1 Remark 4.3. (1) For every finitely generated discrete group, there exist C, r > 0 such that bk ≤ Crk for all k ≥ 0 by the Fekete's subadditivity lemma. Therefore, Theorem 4.2 covers all finitely generated discrete group. (2) In fact, the inequality (4.1) is sharp by Theorem 6.6. Although we can always find inequality (4.2) for every finitely generated discrete group, we can get a much better result by adding more detailed geometric infor- mation of underlying group. Indeed, if we assume hyperbolicity of group, then the inequality is considerably improved. Theorem 4.4. Let G be any non-elementary word hyperbolic group with bk ≤ Crk for all k ≥ 0 with respect to a finite symmetric generating set S. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(S, p) such that (4.3) 2 ) 1 1 p p (Xg∈G (2−p)g 2 r (1 + g)4−2p f (g)p) 1 p ≤ (Xk≥0 1 (k + 1)4−2p ( Xg∈G:g=k ≤ Kλ(f )Lp(V N (G)). f (g)2) for all λ(f ) ∼Xg∈G f (g)λg ∈ Lp(V N (G)). Proof. The conclusion follows from Corollary 3.10 and [13]. (cid:3) 12 4.2. Free quantum groups. Let us begin the investigation of 'genuine' quantum examples: Free orthogonal quantum groups O+ N and free permutation quantum groups S+ N becomes an equiv- alence under centrality, positivity, monotonic decrease and a non-oscillation type condition of Fourier coefficients, as for the result for SU (2) [Theorem 2.10 [2]]. This is considered as a way to prove sharpness of Hardy-Littlewood inequalities. N +2. Moreover, the Hardy-Littlewood inequality for O+ Theorem 4.5. permutation quantum group S+ universal constant K = K(p) such that (1) Let G be the free orthogonal quantum group O+ 2 or the free 4 . Then, for each 1 < p ≤ 2, there exists a (4.4) (Xk≥0 1 (1 + k)6−3p nkbf (k)p for all f ∼Xk≥0 nktr(bf (k)uk) ∈ Lp(G). Sp k+1 ) 1 p ≤ (Xk≥0 1 (1 + k)6−3p n 2− p 2 k 1 p ≤ KfLp(G) HS ) bf (k)p (2) Let G be a free orthogonal quantum group O+ N or a free permutation quantum N +2 with N ≥ 3. Then, for each 1 < p ≤ 2, there exists a universal group S+ constant K = K(p) such that (4.5) (Xk≥0 1 r(2−p)k 0 (1 + k)4−2p ) nkbf (k)p Sp nk 1 p ≤ (Xk≥0 r(2−p)k 0 ≤ KfLp(G) 1 (1 + k)4−2p n 2− p 2 k 1 p HS) bf (k)p for all f ∼Xk≥0 nktr(bf (k)uk) ∈ Lp(G), where r0 = N + √N 2 − 4 2 . Proof. (1) In this case, nk = k + 1 for all k, so that the conclusion follows from Corollary 3.5. (2) It is known that free orthogonal quantum groups O+ N and free quantum groups SN +2 with N ≥ 3 have the rapid decay property with β = 1 ([25] and [5]). Also, sk = n2 for all k ∈ {0} ∪ N. Therefore, Corollaries 3.9 and 3.10 complete proof. (cid:3) k ≈ r2k 0 Remark 4.6. All results of this paper for S+ N can be extended to quantum auto- morphism group Gaut(B, ψ) with a δ-trace ψ and dim(B) = N via the same proofs. An important observation for the free orthogonal quantum groups O+ N is that the inequalities (4.4) and (4.5) become actually equivalences in a large class. Essentially, this fact is based on the result for SU(2) [Theorem 2.10, [2]] and the following lemma moves the result to O+ N . Lemma 4.7. Let G = O+ for each cases. Then, for f ∼Xn≥0 N or S+ cnχ1 n ∈ Lp(G), N +2 with N ≥ 2 and consider G = SU (2) or SO(3) fLp(G) = Φ(f ) ∼Xn≥0 cnχ2 nLp(G) for all 1 ≤ p ≤ ∞. Here, χ1 representations of G and G respectively. n = tr(un) and χ2 n = tr(vn) where un and vn are the n-th irreducible unitary 13 Proof. In the above cases, it is known that G and G share the fusion rule. For k ∈ P ol(G) and m ∈ N, details, see [Proposition 6.7, [26]]. Now, for any x =Xk≥0 ckχ1 h((x∗x)m) = h((Xk,l≥0 ckclχ1 kχ1 l )m) = Xk1,l1,··· ,km,lm≥0 = Xk1,l1,··· ,km,lm≥0 =ZG (Xk,l≥0 =ZG x′2m, ckclχ2 kχ2 l )m ck1 ··· ckmcl1 ··· clmh(χ1 ck1 ··· ckmcl1 ··· clmZG k1 χ1 l1 ··· χ1 km χ1 lm) k1 χ2 χ2 l1 ··· χ2 kmχ2 lm where x′ =Xk≥0 proof. ckχ2 k ∈ P ol(G). Then the Stone-Weistrass theorem completes this (cid:3) 3 2 < p ≤ 2 and fix D > 0. Also, assume that f ∼ Corollary 4.8. Let N ≥ 2, N ) satisfies 2 (O+ 3 ckχk ∈ L Xk≥0 (4.6) Then we have (4.7) ck ≥ ck+1 ≥ 0 and Xm≥k N ) ≈ (Xk≥0 fLp(O+ cm m + 1 ≤ D · ck for all k ≥ 0. (1 + k)2p−4cp k) 1 p . 5. A strong Hardy-Littlewood inequality The studies of Hardy-Littlewood inequalities in [2], [14] and [15] deal with general Lp-functions, but a plenty of classical results of harmonic analysis on T shows that a theorem on a function space can have a stronger form when restricted to "holomorphic" setting [18]. An evidnce on non-commutative setting is "the strong Haagerup inequality" on the reduced group C∗-algebras C∗r (FN ). More precisely, it was shown that the rapid decay property can be strengthen in general holomorphic setting [18]. Let g1,··· , gN be canonical generators of FN and denote by F+ N the the set of elements of the form gi1gi2 ··· gim with m ∈ {0} ∪ N and 1 ≤ ik ≤ N for all 1 ≤ k ≤ m. Theorem 5.1. (Strong Haagerup inequality for C∗r (FN )) Consider a subset E := F+ N and Ek := {g ∈ E : g = k}. Then, for any k ∈ {0} ∪ N, we have that Xg∈Ek f (g)λgC ∗ r (FN ) ≤ √e√k + 1(Xg∈Ek f (g)2) 1 2 . Based on this information, we can modify the inequality (3.4) as follows. Proposition 5.2. Let N ≥ 2. Then we have that (5.1) (Pg∈Ek f (g)2) (k + 1) 1 fA(FN ) ≥ 1 √e sup k≥0 2 1 2 14 for all f ∈ A(FN ). Proof. We can repeat the proof of Proposition 3.7. The only difference is the im- provement of (1 + k)β to (1 + k) 2 in inequality (3.5). Then we are able to get conclusion by restricting support of x ∈ Cc(G) to F+ (cid:3) Theorem 5.3. Let N ≥ 2. Then, for each 1 < p ≤ 2, there exists a universal constant K = K(p) > 0 such that N in the proof. 1 (5.2) 1 3 2 p 1 p 1 2 ) (2−p)g 2 (2−p)N (1 + g) f (g)p) f (g)2) 1 (1 + k) 3 p ≤ (Xk≥0 ( Xg∈FN :g=k 2 (2−p) ≤ Kλ(f )Lp(V N (FN )) f (g)λg ∈ Lp(V N (FN )) with supp(f ) ⊆ F+ N . (Xg∈FN for all λ(f ) ∼ Xg∈FN 3.10. The only difference is to replace the operator A with λ(f ) 7→ (f · χEkl2(G)) 2 . Also, we choose a weight function w on {0}∪N by w(k) := . Then we can derive new inequality for general λ(f ) ∈ Lp(V N (FN )), but (cid:3) (1 + k) 3 our consideration is in the case supp(f ) ⊆ F+ N . Proof. It can be also obtained by repeating the proof of Theorem 3.8 and Corollary and (1+k)β with (1+k) w(k)2 1 1 2 )k≥0 The studied Hardy-Littlewood inequalities give a decay pair (r, s) such that the multiplier 6. Sharpness is bounded for each cases, where wr,s(α) = Fwr,s : Lp(G) → ℓp(bG), f 7→ (wr,s(α)bf (α))α∈Irr(G), 1 rα(1 + α)s with respect to the natural n(2 − p) ) for C(G) with compact Lie p length · on Irr(G). Here is the list: (0, group G, (0, 3(2 − p) p ) for O+ 2 or S+ k0(2 − p) 2−p , p ) for C∗r (G) with polynomially growing discrete group G, p 4 and (r 0 (0, Remark 6.1. ([21] and [27]) If G is a compact Lie group, then √κπ is equivalent to the natural length function · S generated by the fundamental generating set S N or SN +2 with N ≥ 3. p β 2(2 − p) ) for O+ of bG. Equivalently, (1 + κπ) 2 ≈ (1 + πS)β. To assert that the established inequalities are sharp, we will show that there is no slower decay pair (r, s) such that Fwr,s is bounded for the above cases. This viewpoint is different from the spirit of [Theorem 2.10 ,[2]] or Theorem 4.8, which requires finding an equivalence on a subclass. However, our approach is quite natural since it is strongly related to Sobolev embedding theorem. For example, Fw0,s : Lp(Td) → lp(Zd) is bounded if and only if H s (Td) if and only if H p is the Bessel potential space. In this direction, we will provide a Sobolev embedding type interpretation for results of this section in subsection 7.2. (Td) ⊆ Lr(Td) for all 1 < q < r < ∞, where H s p (Td) ⊆ Lp′ 2−p ( 1 q q − 1 r ) ps In addition, this view has a definite advantage over looking for equivalence be- cause we can cover much larger class. Our first strategy is handling an ultracontractivity problem on C(G) with com- pact Lie groups, C∗r (G) with polynomially growing discrete group. Actually ultra- contractivity problem is strongly related to Sobolev embedding property [31]. 15 Let M be the von Neumann subalgebra generated by {χα}α∈Irr(G) in L∞(G) and consider Lp(M ) as the non-commutative Lp-space associated to the restriction of the Haar state on M . Now suppose that l : Irr(G) → (0,∞) is a positive function and there exist 1 < p < 2 and a universal constant C > 0 such that (6.1) J(f ) ∼ Xα∈Irr(G) 1 β p l(α) cαχαLp′ (M) ≤ Cf ∼ Xα∈Irr(G) cαχαLp(M), where J is a densely defined positive operator on L2(M ) which maps χα 7→ 1 for all α ∈ Irr(G), 1 ≤ i, j ≤ nα. Indeed, J = K∗K where K : χα 7→ χα. β 2p l(α) Now take φ(t) := t 2β 2−p , ψ(z) := z 2β 2−p and L := J− p 2β . Then [Theorem 1.1, [31]] says that there exists a universal constant C′ > 0 such that 1 β p l(α) χα (6.2) e−tL(f ) ∼ Xα∈Irr(G) for all 0 < t < ∞ and all f ∼ Xα∈Irr(G) cα etl(α) χαL∞(M) ≤ C′ fL1(M) 2−p 2β t . 1 2 cαχα ∈ L1(M ). Our claim is that we get the following observations in the above situation: (6.3) sup 0<t<∞ t 2β 2−p Xα∈Irr(G) n2 α etl(α) =: C < ∞. 1 2 if G is given by C(G) with compact Lie group or C∗r (G) with polynomially growing discrete group. Lemma 6.2. ([Lemma 4.1, [8]] and [Proposition 5.7, [21]]) n2 π (1 + κπ) s 2 < ∞ Lemma 6.3. (1) Let G be a compact Lie group. Then there exist probability (2) Let G be a finitely generated discrete group with polynomial growth rate k0. 1 if and only if s > n. (1 + g)s < ∞ if and only if s > k0. (1) Let G be a compact Lie group with dimension n. Then Xπ∈ bG Then Xg∈G measures {νt}t>0 such that bνt(π) = (2) Let G be a compact Lie group and let f ∼ Xπ∈ bG that bf (π) ≥ 0 for all π ∈ bG. Then fL∞(G) = Xπ∈ bG nπtr(bf (π)). {νt}t>0 ⊆ L1(G). 1 etκπ Idnπ for all π ∈ bG. Moreover, nπtr(bf (π)π) ∈ L∞(G) such (3) Let G be a discrete group. Then A(G) has a bounded approximate identity if and only if G is amenable. In this case, the bounded approximate identity can be chosen as positive and compactly supported functions on G. (4) If G is an amenable discrete group, we have that λ(f ) ∼Xg∈G f (g)λgV N (G) =Xg∈G f (g) for any positive function f ∈ l1(G). 16 Proof. (1) Since Xπ∈ bG n2 π etκπ < ∞ by Lemma 6.2, we know that νt ∈ A(G) ⊆ C(G) ⊆ L1(G). The family {νt}t>0 is called the Heat semigroup of measures. convolution product, we have (2) Since f 7→ µt ∗ f is a contractive map on L∞(G) for all t > 0 where ∗ is the fL∞(G) ≥ sup t>0 ft ∼ Xπ∈ bG tr(bf (π)) ≤ fL1(G)Xπ t>0Xπ∈ bG ft(1) = sup t>0 nπ etκπ Here, sinceXπ nπ etκπ fL∞(G) ≥ sup The other direction is trivial. series of ft uniformly converges to ft ∈ C(G). Therefore, tr(bf (π)) = Xπ∈ bG nπtr(bf (π)). nπ etκπ n2 π etκπ tr(bf (π)π)C(G) < ∞ by Lemma 6.2, the Fourier (3) See [Theorem 7.1.3, [23]] and its proof. We also may assume the compact supportness by considering fǫ := f · χ{g∈G:f (g)>ǫ} for positive f ∈ l1(G). (4) This is the Kesten's condition that is equivalent to amenability. (cid:3) Now we can show that the claim is true. Proposition 6.4. Let G be C(G) with compact Lie group or C∗r (G) with polyomially growing discrete group. Also, suppose that the inequality (6.1) holds. Then (6.4) Proof. (1) By Lemma 6.3, =: C < ∞. t sup 0<t<∞ Xπ∈ bG n2 π etl(π) 1 2 2β n2 α etl(α) 2−p Xα∈Irr(G) r>0Xπ∈ bG 2 erκπ r>0 e−tL(νr)L∞(G) C′ = sup = sup etl(π) n2 π 1 1 2 ≤ 2β 2−p t sup r>0 νrL1(G) ≤ C′ 2β 2−p t for all 0 < t < ∞. (2) There exists a bounded approximate identity (ei)i in A(G) that consists of positive and compactly supported functions since polynomially growing discrete group is always amenable. Then inequality (6.2) says that Xg∈G 1 1 2 etl(g) = sup i Xg∈G ei(g) 1 2 etl(g) = sup i Xg∈G ei(g) 1 2 etl(g) λgC ∗ r (G) ≤ C′′ 2β 2−p t since lim i ei(g) = 1 for all g ∈ G. (cid:3) Proposition 6.4 allows us to extract an interesting quantitive observation. Proposition 6.5. Let G be C(G) with compact Lie group or C∗r (G) with polyomially growing discrete group. Also, suppose that the inequality (6.1) holds. Then we have that (6.5) Xα∈Irr(G) n2 α l(α) m 2 < ∞ for all natural numbers m > 2β 2 − p . 17 1 =: C0 < ∞ t 1 2 sup 1 xγ dx n2 α etl(α) n2 α exl(α) from (6.4), so that for all 0 < t ≤ 1. This implies that Proof. Choose γ ∈ (maxn 2β 2−p , m − 1o , m). Then we have 0<t≤1 tγ Xα∈Irr(G) Z 1 t Xα∈Irr(G) 0<t≤1 0<t≤1 tγ−(m−1) Xα∈Irr(G) Xα∈Irr(G) 2 dx ≤Z 1 2 2 tγ−1 Xα∈Irr(G) Then there exist D1, D2 > 0 such that so that we can inductively see that l(α) m 2 etl(α) l(α) 2 etl(α) 2 etl(α) l(α) n2 α 1 n2 α m−1 sup sup n2 α 1 1 =: C1 < ∞, 2 ≤ D1tm−γ + D2 for all 0 < t ≤ 1 1 =: Cm−1 < ∞. via the same way. Lastly, taking the limit t → 0 completes this proof. (cid:3) Theorem 6.6. Let 1 < p ≤ 2. (1) Let G be a compact Lie group with dimension n. Then (Xπ∈ bG 1 (1 + κπ) s 2 nπbf (π)p Sp nπ ) 1 p . fLp(G) (2) Let G be a finitely generated discrete group with polynomial growth rate k0. holds if and only if s ≥ n(2 − p). Then 1 (Xg∈G (1 + g)sf (g)p) holds if and only if s ≥ k0(2 − p). (3) Let G be O+ 2 or S+ 4 . Then 1 p . λ(f )Lp(V N (G)) (Xk≥0 1 (1 + k)s nkbf (k)p Sp k+1 ) 1 p . fLp(G) holds if and only if s ≥ 3(2 − p). (4) Let G be O+ N or S+ N +2 with N ≥ 3. Then (Xk≥0 1 r(2−p)k 0 (1 + k)s nkbf (k)p Sp nk holds if and only if s ≥ 4 − 2p, where r0 = 18 1 ) p . fLp(G) N + √N 2 − 4 . 2 Proof. "If" parts are obtained from (1.3), (4.1), (4.4) and (4.5). To prove the 2 and (2) 1 + g respectively. converse direction, firstly, define l(α) by (1) (1 + κπ) Then the assumed inequality 1 ( Xα∈Irr(G) nα l(α)s bf (α)p Sp nα ) 1 p . fLp(G) implies the inequality (6.1) for β = s. Then, by Proposition 6.5 and Lemma 6.2, we get that (1) 2n ≤ , (2) 2k0 ≤ respectively. In (3) and (4), consider G as SU (2) if G = O+ N +2 for each cases. Also, denote by χ′k the character corresponding to χk for each cases. Define l(k) := 1 + k. N and SO(3) if G = S+ 2β 2 − p 2β 2 − p Firstly, in (3), for each f ∼Xk≥0 ckχk ∈ Lp(G), Xk≥0 (1 + k)− s p ckχkLp′ (G) ≤ (Xk≥0 1 1 p (1 + k)s+p−2 ckp) . fLp(G) = f′ ∼Xk≥0 ckχ′kLp(G). by Proposition 4.7 and Hausdorff-Young inequality. On the other hand, in (4), for each f ∼Xk≥0 Xk≥0 ckχk ∈ Lp(G), (1 + k)− s+2−p p ckχkLp′ (G) ≤ (Xk≥0 1 1 p (1 + k)sckp) . fLp(G) = f′ ∼Xk≥0 ckχ′kLp(G) by the similar way. Now we can apply Proposition 6.5 and Lemma 6.2 for compact Lie groups again, (cid:3) so that (3) s ≥ 6 − 3p and (4) s − p + 2 ≥ 6 − 3p(⇔ s ≥ 4 − 2p) respectively. 7. Some remarks about Sidon sets, Sobolev embedding theorem and quantum torus As by-products of this study, we refer to an interesting lacunarity result for com- pact quantum groups, and present a sobolev embedding theorem type interpretation for C(G) with compact Lie group and for C∗r (G) with polynomially growing group. Also, we show an explicit inequality on quantum torus Td θ. 7.1. Sidon set on compact quantum groups. The study of Lacunarity, espe- cially on Sidon sets, is one of the major subject in harmonic analysis, and recently the notion has been extended to the setting of compact quantum groups [26]. Definition 7.1. Let G be a compact quantum group. (1) A subset E ⊆ Irr(G) is called a Sidon set if there exists K > 0 such that (2) A subset E ⊆ Irr(G) is called a central Sidon set if there exists K > 0 such bfℓ1(bG) ≤ KfL∞(G) for all f ∈ P olE(G), where P olE(G) :=nf ∈ P ol(G) : bf (α) = 0 for all α /∈ Eo. bfℓ1(bG) ≤ KfL∞(G) for all f = Xα∈Irr(G) that 19 cαχα ∈ P olE(G). Let G = (A, ∆) be of Kac type and E ⊆ Irr(G) be a central sidon set. Then [Proposition 6.4, [26]] implies that there exists µ ∈ M (G) = Cr(G)∗ such that j,i)∗))1≤i,j≤nα = Idnα for all α ∈ E. Since P ol(G) is dense in Cr(G), Proposition 3.7 still holds for µ ∈ M (G). Now, if G satisfies the assumptions of Proposition 3.7 and if E ⊆ Irr(G) is a central sidon set, then we get bµ(α) = (µ((uα 1 2 (Pα∈Ek n2 α) (1 + k)β 2 minα∈Ek nα (1 + k)β Ek 1 , ∞ > sup k≥0 ≥ sup k≥0 where Ek := {α ∈ E : α = k}. Irr(G) with r > 1. Therefore, it can not happen simultaneously that E = ∞ and nα > rα ∀α ∈ Remark 7.2. (1) The above argument shows that there is no an infinite (cen- tral) Sidon set in U + N with N ≥ 3, which are not explained in [26]. (2) Shortly after this research, the author of [26] personally informed me another 2 , the N cases. Under the identification Irr(U + N )∼= F+ simple idea to explain U + fact that χα4 = (1 + α) 1 4 ∀α ∈ F+ 2 implies that there is no an infinite Λ(4) set, so that there is no an infinite Sidon set on U + N with N ≥ 2. 7.2. Sobolev embedding properties. The contents of Section 6 can be inter- preted in terms of Sobolev embeddings properties by [Theorem 1.1, [31]]. For C(G) with compact Lie group whose real dimension is n, the computations in Section 6 says that (1 − ∆)− β 2 (f ) ∼ Xπ∈ bG nπ (1 + κπ) β 2 tr(bf (π)π)Lp′ (G) . fLp(G) if and only if β ≥ n(2 − p) p for each 1 < p ≤ 2. Moreover, it is equivalent to that 2 ( 1 p − 1 q )(f )Lq(G) . fLp(G) if and only if β ≥ n for each 1 < p < q < ∞. If we define the space H s the above result is interpreted as (1 − ∆)− β 2 (f ) ∈ Lp(G)(cid:9) as an analogue of Bessel potential space, then p(G) ⊆ Lq(G) if and only if s ≥ n( H s (cid:8)f ∈ Lp(G) : (1 − ∆) s p(G) := 1 p − (7.1) 1 q ) for each 1 < p < q < ∞. On the other hand, if G is a finitely generated discrete group with polynomial growth rate k0, then we define infinitesimal generator L on C∗r (G) by λg 7→ −gλg for all g ∈ G. Then we can derive the Sobolev embedding property of non- commutative spaces Lp(V N (G)) as follows. p − 1 q )(λ(f ))Lq (V N (G)) . λ(f )Lp(V N (G)) if and only if β ≥ k0 (7.2) (1 − L)−β( 1 for each 1 < p < q < ∞. : λg 7→ −g2λg, but it does not make an essential difference in replacing (1 − L) with (1 − L′) The reader may consider another natural infinitesimal generator L′ 2 . 1 20 7.3. Hardy-Littlewood inequality on Quantum torus. Quantum torus Td θ is a widely studied example of "quantum space". In this case, we can establish Hardy- θ, which is the same form as for Td. A proof can be given Littlewood inequality for Td by repeating the proof of Theorem 3.1. Remark 7.3. For a quantum torus Td θ, for each 1 < p ≤ 2, we have that (7.3) (Xm∈Zd 1 (1 + m1)d(2−p) bx(m)p) 1 p . xLp(Td θ ). Acknowledgement. The author is grateful to Hun Hee Lee for helpful com- ments on this work. Also, he would like to thank Simeng Wang for discussing the Sidon set of compact quantum groups. References [1] Akylzhanov, R., Nursultanov, E., Ruzhansky, M. (2016). Hardy-Littlewood inequalities and Fourier multipliers on SU (2). Studia Math., 234 (2016), 1-29. [2] Akylzhanov, R., Nursultanov, E., Ruzhansky, M. (2015). Hardy-Littlewood, Hausdorff-Young- Paley inequalities, and Lp-Lq Fourier multipliers on compact homogeneous manifolds. arXiv preprint arXiv:1504.07043. [3] Brannan, M. (2012). Quantum symmetries and strong Haagerup inequalities. Communications in Mathematical Physics, 311(1), 21-53. [4] Brannan, M. P. (2012). On the Reduced Operator Algebras of Free Quantum Groups, Ph.D Thesis, Queen's University. [5] Brannan, M. (2013). Reduced operator algebras of trace-preserving quantum automorphism groups. Documenta Mathematica, 18, 1349-1402. [6] Banica, T., Vergnioux, R. (2009). Growth estimates for discrete quantum groups. Infinite Dimensional Analysis, Quantum Probability and Related Topics, 12(02), 321-340. [7] Caspers, M. (2013). The Lp-Fourier transform on locally compact quantum groups. Journal of Operator Theory, 69(1), 161-193. [8] Dasgupta, A., Ruzhansky, M. (2014). Gevrey functions and ultradistributions on compact Lie groups and homogeneous spaces. Bulletin des Sciences Mathmatiques, 138(6), 756-782. [9] Franz, U., Hong, G., Lemeux, F., Ulrich, M. (2015). Hypercontractivity of heat semigroups on free quantum groups. Journal of Operator Theory, to appear. arXiv:1511.02753. [10] Folland, G. B. (2013). Real analysis: modern techniques and their applications. John Wiley & Sons. [11] Haagerup, U. (1978). An example of a non nuclearC*-algebra, which has the metric approxi- mation property. Inventiones Mathematicae, 50(3), 279-293. [12] Haagerup, U. (1979). Lp-spaces associated with an arbitrary von Neumann algebra. In Al- gebres doprateurs et leurs applications en physique mathmatique (Proc. Colloq., Marseille, 1977) (Vol. 274, pp. 175-184). [13] de la Harpe, P. (1988). Groupes hyperboliques, algebres doprateurs et un thoreme de Jolis- saint. CR Acad. Sci. Paris Sr. I Math, 307(14), 771-774. [14] Hardy, G. H., Littlewood, J. E. (1927). Some new properties of Fourier constants. Mathema- tische Annalen, 97(1), 159-209. [15] Hewitt, E., Ross, K. A. (1974). Rearrangements of Lr Fourier series on compact abelian groups. Proceedings of the London Mathematical Society, 3(2), 317-330. [16] Junge, M., Mei, T., Parcet, J. (2014). Smooth Fourier multipliers on group von Neumann algebras. Geometric and Functional Analysis, 24(6), 1913-1980. [17] Junge, M., Palazuelos, C., Parcet, J., Perrin, M. (2013). Hypercontractivity in group von Neumann algebras. Memoirs of American Mathematical Society, to appear. arXiv:1304.5789. [18] Kemp, T., Speicher, R. (2007). Strong Haagerup inequalities for free R-diagonal elements. Journal of Functional Analysis, 251(1), 141-173. [19] Kustermans, J., Vaes, S. (2000, November). Locally compact quantum groups. In Annales Scientifiques de lEcole Normale Suprieure (Vol. 33, No. 6, pp. 837-934). No longer published by Elsevier. [20] Kustermans, J., Vaes, S. (2003). Locally compact quantum groups in the von Neumann algebraic setting. Mathematica scandinavica, 68-92. [21] Lee, H. H., Youn, S. (2015). New deformations of Convolution algebras and Fourier algebras on locally compact groups. Canadian Journal of Mathematics, to appear. [22] Pisier, G., Xu, Q. (2003). Non-commutative Lp-spaces. Handbook of the geometry of Banach spaces, 2, 1459-1517. [23] Runde, V. (2002). Lectures on amenability (No. 1774). Springer Science & Business Media. 21 [24] Timmermann, T. (2008). An invitation to quantum groups and duality: From Hopf algebras to multiplicative unitaries and beyond. European Mathematical Society. [25] Vergnioux, R. (2007). The property of rapid decay for discrete quantum groups. Journal of Operator Theory, 57(2), 303-324. [26] Wang, S. (2016). Lacunary Fourier series for compact quantum groups. Communications in Mathematical Physics, 1-51. [27] Wallach.N.R. (1973). Harmonic analysis on homogeneous spaces. Pure and Applied Mathe- matics, No. 19. Marcel Dekker, Inc., New York. [28] Wang, S. (1995). Free products of compact quantum groups. Communications in Mathemat- ical Physics, 167(3), 671-692. [29] Wang, S. (1998). Quantum symmetry groups of finite spaces. Communications in mathemat- ical physics, 195(1), 195-211. [30] Woronowicz, S. L. (1987). Twisted SU (2) group. An example of a non-commutative dif- ferential calculus. Publications of the Research Institute for Mathematical Sciences, 23(1), 117-181. [31] Xiong, X. (2016). Noncommutative harmonic analysis on semigroup and ultracontrativity. arXiv preprint arXiv:1603.04247. [32] Xu, Q. (2007). Operator spaces and noncommutative Lp. The part on non-commutative Lp-spaces. Lectures in the Summer School on Banach spaces and Operator spaces, Nankai University-China. Sang-Gyun Youn : E-mail address: [email protected] 22
1712.03436	1	1712	2017-12-09T20:19:00	Derivations on ternary rings of operators	[ "math.OA", "math.FA" ]	To each projection $p$ in a $C^*$-algebra $A$ we associate a family of derivations on $A$, called $p$-derivations, and relate them to the space of triple derivations on $p A (1-p)$. We then show that every derivation on a ternary ring of operators is spatial and we investigate whether every such derivation on a weakly closed ternary ring of operators is inner.	math.OA	math	DERIVATIONS ON TERNARY RINGS OF OPERATORS ROBERT PLUTA AND BERNARD RUSSO Abstract. To each projection p in a C ∗-algebra A we associate a family of derivations on A, called p-derivations, and relate them to the space of triple derivations on pA(1 − p). We then show that every derivation on a ternary ring of operators is spatial and we investigate whether every such derivation on a weakly closed ternary ring of operators is inner. 1. S-derivations on C-algebras If A is a C ∗-algebra, we let D(A) denote the Banach Lie algebra of derivations on A. To be more precise D(A) consists of all operators δ ∈ B(A) that satisfy δ(xy) = δ(x)y + xδ(x) for every x, y in A. B(A) denotes the bounded linear operators on A. A derivation δ ∈ D(A) is called self-adjoint if δ = δ∗, where δ∗ is the derivation defined by δ∗(x) = δ(x∗)∗ for every x in A. The space of all self-adjoint derivations on A is a real Banach Lie subalgebra of D(A) and is denoted D∗(A). Derivations on C ∗-algebras have suitable counterparts in a more gen- eral setting of ternary rings of operators, or TROs for short, where they are sometimes termed triple derivations. However, in this paper we shall use the term triple derivation to denote a derivation of a Jordan triple system. For example, if X is a Banach subspace of a C ∗-algebra and xy∗z +zy∗x ∈ X for every x, y, z in X, then X is called a JC∗-triple and a triple derivation on X is an operator τ ∈ B(X) satisfying τ ({xy∗z}) = {τ (x)y∗z} + {xτ (y)∗z} + {xy∗τ (z)} for every x, y, z in X, where {xyz} = (xy∗z + zy∗x)/2. We shall use the term TRO-derivation, as follows: If X is a Banach subspace of a C ∗-algebra and xy∗z ∈ X for every x, y, z in X, then X is called a TRO and a TRO-derivation on X is an operator τ ∈ B(X) satisfying τ (xy∗z) = τ (x)y∗z + xτ (y)∗z + xy∗τ (z) Date: September 18, 2018. Key words and phrases. C-algebra, ternary ring of operators, TRO, derivation, linking algebra, W-TRO. 1 2 ROBERT PLUTA AND BERNARD RUSSO for every x, y, z in X. It is clear that a TRO (resp. JC∗-triple) can also be defined as a Ba- nach subspace of B(H, K), the bounded operators from Hilbert space H to Hilbert space K, which is closed under the triple product xy∗z (resp. (xy∗z + zy∗x)/2). If a TRO is weakly closed, it is called a W∗- TRO. In this section we will introduce the class of S-derivations on a C ∗- algebra A associated with a subspace S ⊆ A. Of particular interest will be the case S = pAp for a projection p in A. We will seek to determine the relationship between the class of pAp derivations (which we call p-derivations for short) on A and the class of TRO-derivations on pA(1 − p). Definition 1.1. Let A be a C ∗-algebra and let S be a subspace of A. We say that a derivation δ ∈ D(A) is associated with S, or simply that δ is an S-derivation, if δ leaves S invariant in the sense that δ(S) ⊆ S. We use DS(A) to denote the set of all S-derivations. In order to simplify the notation, we write De(A) for DeAe(A) in case S = eAe, for some idempotent e ∈ A, and we abuse the terminology slightly by referring to the elements of De(A) simply as e-derivations. To repeat, given an arbitrary idempotent e in a C ∗-algebra A, which in particular may be a projection, by an e-derivation on A we mean a derivation δ ∈ D(A) satisfying δ(eAe) ⊆ eAe. This condition is easily seen to be equivalent to the requirement that δ(e) = 0. Example 1.2. Let A be a C ∗-algebra and let e ∈ A be an idempotent. Fix a ∈ eAe and b ∈ (1 − e)A(1 − e) = {x − xe − ex + exe : x ∈ A}. Then δ : A → A defined by δ(x) = (a + b)x − x(a + b) is an e-derivation. Lemma 1.3. Let A be a C ∗-algebra and let S be a subalgebra with an identity element 1S (possibly different from the identity element of A if A is unital). Let δ ∈ D(A) be a derivation. The following state- ments hold. (1) If δ(S) ⊆ S then δ(1S) = 0. (2) If δ(1S) = 0 then δ(S) ⊆ 1SA1S. Proof. A straightforward consequence of the derivation property. (cid:3) Lemma 1.4. Let A be a C ∗-algebra and let e ∈ A be an idempotent. Let δ ∈ D(A) be a derivation. The following statements hold. (1) If δ(e) = 0, then δ leaves invariant the following subspaces eAe, eA(1 − e), (1 − e)Ae, (1 − e)A(1 − e). DERIVATIONS ON TERNARY RINGS OF OPERATORS 3 (2) If δ leaves invariant eAe or (1 − e)A(1 − e), then δ(e) = 0. Additionally, let δ = δ∗ and e = e∗. Then the following statement holds. (3) If δ leaves invariant eA(1 − e) or (1 − e)Ae, then δ(e) = 0. Proof. The assertions (1) and (2) are straightforward consequences of the derivation property. To prove (3), assume that eA(1−e) is invariant for δ = δ∗, and e = e∗. Since δ(e) = δ(e)e + eδ(e), we have eδ(e)e = 0 and hence δ(e) = eδ(e)(1 − e) + (1 − e)δ(e)e. This shows that both eδ(e) and δ(e)(1 − e) are equal to eδ(e)(1 − e), and so both eδ(e) and δ(e)(1−e) are elements of the subspace eA(1−e) which is invariant under δ. We will show that δ(e) = 0 by showing that δ(e)2 = 0. For this, we identify A with (cid:16) eAe (1−e)Ae (1−e)A(1−e)(cid:17) and write δ(e) and δ2(e) as eA(1−e) δ(e) = (cid:16) 0 eδ(e)(1−e) (1−e)δ(e)e 0 δ2(e) = ( a b c d ) . (cid:17) , Then δ(e)2 = (cid:16) eδ(e)(1−e)δ(e)e 0 0 (1−e)δ(e)eδ(e)(1−e)(cid:17) and since δ(eδ(e)) = (cid:16) eδ(e)(1−e)δ(e)e+ea δ(δ(e)(1 − e)) = (cid:16) −eδ(e)(1−e)δ(e)e 0 0 eb (1−e)δ(e)eδ(e)(1−e)(cid:17) ∈ (cid:0) 0 eA(1−e) (cid:1) , d(1−e)−(1−e)δ(e)eδ(e)(1−e)(cid:17) ∈ (cid:0) 0 eA(1−e) b(1−e) 0 0 0 0 it follows that (1 − e)δ(e)eδ(e)(1 − e) = 0 = eδ(e)(1 − e)δ(e)e. Thus δ(e)2 = 0, as desired. (cid:3) (cid:1) , If A is a C ∗-algebra and p ∈ A is a projection, we let D∗ p(A) denote the (real) Banach Lie algebra of self-adjoint p-derivations on A. To be more precise D∗ p(A) consists of all derivations δ ∈ D(A) that satisfy δ(p) = 0 and δ = δ∗. If X is a TRO, we use DT RO(X) to denote the (real) Banach Lie algebra of all TRO-derivations on X. Remark 1.5. Let A be a unital C ∗-algebra and let p ∈ A be a projection. Then the map ∆ : D∗ p(A) → DT RO(pA(1 − p)), ∆(δ) = δpA(1−p) is a homomorphism of Banach Lie algebras. Example 1.6. Let A = M2(C), p = ( 1 0 on A is: 0 0 ) . The set of all p-derivations Dp(A) = {δ ∈ D(A) : δ(p) = 0} ≃ (cid:8)(cid:0) α 0 0 β(cid:1) : α, β ∈ C(cid:9) = a complex Banach Lie algebra. 4 ROBERT PLUTA AND BERNARD RUSSO The set of all self-adjoint p-derivations is: D∗ p(A) = {δ ∈ Dp(A) : δ = δ∗} ≃ (cid:8)(cid:0) α 0 0 β(cid:1) : α, β ∈ C with ℜ(α) = ℜ(β)(cid:9) = a real Banach Lie algebra. The mapping ∆ : D∗ p(A) → DT RO(X), ∆(δ) = δX defines a linear surjection between the self-adjoint p-derivations on A and the TRO-derivations on X = pA(1 − p) = ( 0 C 0 0 ) (see Lemma 2.1). The kernel of ∆ is isomorphic to the center of A, i.e., ker ∆ = Z(A) = {( α 0 0 α ) : α ∈ C} . In other words, the TRO-derivations on X = pA(1−p) = ( 0 C cisely the self-adjoint p-derivations on the linking algebra ( XX ∗ X A = M2(C). 0 0 ) are pre- X ∗ X ∗X ) = Example 1.7. Let A = M5(C), and let p ∈ A be the projection matrix with 1 in the (1, 1) and (2, 2) position and zero's elsewhere. The set of all p-derivations on A is: Dp(A) = {δ ∈ D(A) : δ(p) = 0} 0 B ) : A ∈ M2(C), B ∈ M3(C)} ≃ {( A 0 = a complex Banach Lie algebra. p(A) = {δ ∈ Dp(A) : δ = δ∗} The set of all self-adjoint p-derivations is D∗ and it can be identified with the real Banach Lie algebra consisting of all matrices of the form ( A 0 0 B ) where A ∈ M2(C), B ∈ M3(C), and (cid:0) A+A∗ 0 0 B+B∗(cid:1) is in the center of A. 2. Derivations on TROs If A is a unital C-algebra and e is a projection in A, then X := eA(1 − e) is a TRO. Conversely if X ⊂ B(K, H) is a TRO, then with X ∗ = {x∗ : x ∈ X} ⊂ B(H, K), XX ∗ = span {xy∗ : x, y ∈ X} ⊂ B(H), X ∗X = span {z∗w : z, w ∈ X} ⊂ B(K), Kl(X) = XX ∗n , Kr(X) = X ∗X , we let1 n AX = (cid:20)Kl(X) + C1H X ∗ X Kr(X) + C1K(cid:21) ⊂ B(H ⊕ K) 1If Kl(X) and Kr(X) are unital subalgebras of B(H) and B(K) (resp.), and X is nondegenerate, that is, XX ∗X is dense in X, then we take AX to beh Kl(X) X X ∗ Kr (X)i DERIVATIONS ON TERNARY RINGS OF OPERATORS 5 denote the (unital) linking C-algebra of X. Then we have a TRO- isomorphism X ≃ eAX (1 − e), where e = [ 1 0 0 0 ]. Lemma 2.1. Let X be a TRO and let D : X → X be a TRO-derivation of X. If A0 = ( XX ∗ X X ∗ X ∗X ), then the map δ0 : A0 → A0 given by (cid:18)Pi xiy∗ i y∗ Pj z∗ x j wj(cid:19) 7→ (cid:18)Pi(xi(Dyi)∗ + (Dxi)y∗ (Dy)∗ i ) Dx j (Dwj) + (Dzj)∗wj)(cid:19) Pj(z∗ is well defined and a bounded -derivation of A0, which extends D (when X is embedded in AX via x 7→ ( 0 x 0 0 )), and which itself extends to a -derivation δ of AX. Thus, the Lie algebra homomorphism ∆ : δ 7→ δX given in Remark 1.5 is onto. Proof. If Pi xiy∗ i = 0, then for every z ∈ X, xiy∗ i z) 0 = D(Xi = Xi = (Xi ((Dxi)y∗ i z + xi(Dyi)∗z + xiy∗ i (Dz)) ((Dxi)y∗ i + xi(Dyi)∗))z. i + xi(Dyi)∗) = 0 (see [5, Lemma 2.3(iv)]) and it follows that δ0 is well defined. Since this is true for every z, we have Pi((Dxi)y∗ The map δ0 is self-adjoint since if a = (cid:16) Pi xiy∗ δ0(a∗) = δ0(cid:18)Pi yix∗ j zj(cid:19) = (cid:18)Pi(yi(Dxi)∗ + (Dyi)x∗ x∗ Pj w∗ (Dx)∗ i ) y i i x y∗ Pj z ∗ j wj(cid:17), then Dy j (Dzj) + (Dwj)∗zj)(cid:19) Pj(w∗ = δ0(a)∗. It is easy to verify that δ0(a2) = δ0(a)a+aδ0(a) so that δ0 is a Jordan -derivation of A0. (We omit that calculation.) To see that δ0 is bounded, we first note that D is bounded, since it is a Jordan triple derivation on the JB-triple X with the Jordan triple product {xyz} = (xy∗z + zy∗x)/2, and hence bounded by the theorem i ) by of Barton and Friedman [1]. Now denoting Pi(xi(Dyi)∗ + (Dxi)y∗ 6 ROBERT PLUTA AND BERNARD RUSSO α, we have (by [5, Lemma 2.3(iv)] again) kPi(xi(Dyi)∗ + (Dxi)y∗ kαk = sup kαzk i )k = kzk≤1,z∈X sup kαz + kzk≤1,z∈X sup kzk≤1,z∈X sup kzk≤1,z∈X kXi kDXi = = = z Xi D(xiy∗ xiy∗ =0 xiy∗ } i (Dz) −Xi i z) −Xi i z − kDkXi xiy∗ xiy∗ i (Dz) k { i (Dz)k xiy∗ i Dz kDk k ≤ 2kDkkXi xiy∗ i k. Thus δ0 is bounded and therefore extends to a bounded Jordan - derivation δ of A0 and hence to AX by setting δ(e) = 0, where e = (cid:2) 1H 0 0 0(cid:3). By the theorem of Sinclair ([16, Theorem 3.3]), δ is a derivation of AX.2 (cid:3) n For any C-algebra A ⊂ B(H), the Lie algebra homomorphism A ∋ ) is onto (theorem of Kadison and Sakai ([15, 4.1.6])) w w z 7→ ad z ∈ D(A and so we have the Lie algebra isomorphism w ) ≃ D(A w /Z(A A w ). It follows (cf. [15, 4.1.7]) that w {t ∈ A w : ad t(A) ⊂ A}/Z(A ) ≃ D(A), and w {t ∈ A w : t∗ = −t, ad t(A) ⊂ A}/Z(A ) ≃ D∗(A). Further, for a projection e in A, we have w {t ∈ A w : et = te, t∗ = −t, ad t(A) ⊂ A}/Z(A ) ≃ D∗ e(A). Using these facts in the setting of Lemma 2.1, and noting that, by [10, page 268], AX following theorem. w = A′′ X = h Kl(X)′′ X Kr(X)′′i, we can now prove the X ∗w w Theorem 2.2. Every TRO-derivation of a TRO X is spatial in the sense that there exist α ∈ Kl(X)′′ and β ∈ Kr(X)′′ such that α∗ = −α, β∗ = −β, and Dx = αx + xβ for every x ∈ X. 2It is also easy to verify directly, by (a more involved) calculation, that δ0 is a derivation, thereby avoiding the use of Sinclair's theorem DERIVATIONS ON TERNARY RINGS OF OPERATORS 7 Proof. If D ∈ DT RO(X), choose δ = ad t for some t ∈ AX t∗ = −t, te = et and w with Moreover The conditions on t imply that t = (cid:2) α 0 δ(cid:20)0 x 0 0(cid:21) = (cid:20)α 0 0 0(cid:21) −(cid:20)0 x 0 β(cid:21)(cid:20)0 x (cid:20)0 Dx 0 0 (cid:21) = δ(cid:20)0 x 0 0(cid:21) . 0 β(cid:3) with α∗ = −α and β∗ = −β. 0 β(cid:21) = (cid:20)0 αx + x(−β) (cid:21) . 0 0(cid:21)(cid:20)α 0 0 0 (cid:3) j=1(c∗ A TRO derivation D of a TRO X is said to be an inner TRO deriva- tion if there exist α = −α∗ ∈ XX ∗ and β = −β∗ ∈ X ∗X such that Dx = αx + xβ for x ∈ X. Note that there exist ai, bi, cj, dj ∈ X, i ) and β = i − bia∗ 1 ≤ i ≤ n, 1 ≤ j ≤ m such that α = Pn Pm j dj − d∗ Corollary 2.3. Every TRO derivation of a C∗-algebra A is of the form with α∗ = −α, β∗ = −β. A ∋ x 7→ αx + xβ with elements α, β ∈ A In particular, every TRO derivation of a von Neumann algebra is an inner TRO derivation i=1(aib∗ j cj). w Thus, every W∗-TRO which is TRO-isomorphic to a von Neumann algebra has only inner TRO derivations. For example, this is the case for the stable W∗-TROs of [13] (see subsection 3.2) and the weak- closed right ideals in certain continuous von Neumann algebras acting on separable Hilbert spaces (see Theorem 3.3). Theorem 2.2 is an improvement of [18], in which, although proved for the slightly more general case of derivation pairs, it is assumed that the TRO (called B-triple system in [18]) contains the finite rank operators. For the extension of Zalar's result to unbounded operators, see [17]. A triple derivation δ of a JC∗-triple X is said to be an inner triple derivation if there exist finitely many elements ai, bi ∈ X, 1 ≤ i ≤ n, i=1({aibix} − {biaix}) for x ∈ X, where {xyz} = (xy∗z + zy∗x)/2. For convenience, we denote the inner triple derivation x 7→ {abx} − {bax} by δ(a, b). Thus such that δx = Pn δ(a, b)(x) = (ab∗x + xb∗a − ba∗x − xa∗b)/2. Let X be a TRO. As noted in the proof of Lemma 2.1, X is a JC∗-triple in the triple product (xy∗z + zy∗x)/2, and every TRO- derivation of X is obviously a triple derivation. On the other hand, Indeed, if every inner triple derivation is an inner TRO-derivation. 8 ROBERT PLUTA AND BERNARD RUSSO δ(x) = {abx} − {bax}, for some a, b ∈ X, then δ(x) = Ax + xB, where A = ab∗ −ba∗ ∈ XX ∗, B = b∗a−a∗b ∈ X ∗X with A, B skew-hermitian. Moreover, since by [1, Theorem 4.6], every triple derivation δ on X is the strong operator limit of a net δα of inner triple derivations, hence TRO-derivations, we have (i) and (ii) in the following proposition. Proposition 2.4. Let X be a TRO. (i): Every TRO-derivation is the strong operator limit of inner TRO-derivations. (ii): The triple derivations on X coincide with the TRO-derivations. (iii): The inner triple derivations on X coincide with the inner TRO-derivations (iv): All TRO derivations of X are inner, if and only if, all triple derivations of X are inner. Proof. Since (iv) is immediate from (ii) and (iii), we only need to show part of (iii), that is, that every inner TRO-derivation is an inner triple derivation. If D is an inner TRO-derivation, then Dx = αx + xβ, with α∗ = −α ∈ XX ∗ and β∗ = −β ∈ X ∗X. We must show that there k=1 δ(ak, bk)x where δ(ak, bk) is i=1 xiy∗ i j wj, then it suffices to take p = m + n and choose ai = xi/2, bi = yi for 1 ≤ i ≤ n and an+i = wi, bn+i = zi/2 for 1 ≤ i ≤ m. (cid:3) exist elements ak, bk such that Dx = Pp the inner triple derivation x 7→ {akbkx} − {bkakx}. If α = Pn and β = Pm j=1 z∗ 3. Derivations on W-TROs A von Neumann algebra M is an example of a unital reversible JW ∗- algebra, and as such, by [7, Theorem 2 and the first sentence in its proof], every triple derivation on M is an inner triple derivation. Hence we see that the last statement in Corollary 2.3 follows also from this and Proposition 2.4(iv). For completeness, we include a proof of the former result which avoids much of the Jordan theory, starting with the following lemma, the first part of which is straightforward. Lemma 3.1. Let A be a unital Banach ∗-algebra equipped with the ternary product given by {a, b, c} = 1 2 (ab∗c + cb∗a) and the Jordan product a ◦ b = (ab + ba)/2. • Let D be an inner derivation, that is, D = ad a : x 7→ ax − xa, for some a in A. Then D = ad a is a -derivation whenever a∗ = −a. Conversely, if D is a -derivation, then a∗ = −a + z for some z in the center of A. • Every triple derivation is the sum of a Jordan -derivation and an inner triple derivation. DERIVATIONS ON TERNARY RINGS OF OPERATORS 9 Proof. To prove the second statement, we modify the proof in [8, Sec- tion 3] which is in a different context. We note first that for a triple derivation δ, δ(1)∗ = −δ(1). Next, for a triple derivation δ, the map- ping δ1(x) = δ(1) ◦ x is equal to the inner triple derivation − 1 4 δ(δ(1), 1) so that δ0 := δ − δ1 is a triple derivation with δ0(1) = 0. Finally, any triple derivation which vanishes at 1 is a Jordan -derivation. (cid:3) Theorem 3.2. Every triple derivation on a von Neumann algebra is an inner triple derivation. Proof. It suffices, by the second statement in Lemma 3.1, to show that every self-adjoint Jordan derivation is an inner triple derivation. If δ is a self-adjoint Jordan derivation of M, then δ is an associative derivation (by the theorem of Sinclair, [16, Theorem 3.3])) and hence by the theorem of Kadison and Sakai ([15, 4.1.6]) and the first statement in Lemma 3.1, δ(x) = ax − xa where a∗ + a = z is a self adjoint element of the center of M. Since for every von Neumann algebra, we have M = Z(M) + [M, M], where Z(M) denotes the center of M (see [12, Section 3] for a discussion of this fact), we can therefore write a = z′ +Xj = z′ +Xj [bj + icj, b′ j + ic′ j] ([bj, b′ j] − [cj, c′ ([cj, b′ j] + [bj, c′ j]), j]) + iXj where bj, b′ j, cj, c′ It follows that j are self adjoint elements of M and z′ ∈ Z(M). 0 = a∗ + a − z = (z′)∗ + z′ − z + 2iXj ([cj, b′ j] + [bj, c′ j]) j] + [bj, c′ j]) belongs to the center of M. We now have so that Pj([cj, b′ δ = ad a = ad Xj ([bj, b′ j] − [cj, c′ j]). Pj(cid:0)δ(bj, 2b′ A direct calculation shows that δ is equal to the inner triple derivation (cid:3) j) − δ(cj, 2c′ j)(cid:1), completing the proof. 3.1. Weakly closed right ideals in von Neumann algebras. In this subsection, we shall consider the TRO pM where M is a von Neumann algebra and p is a projection in M. A TRO of the form pM, with M a continuous von Neumann algebra, is classified into four types in [9] as follows. • II a • II a 1 if M is of type II1 and p is (necessarily) finite. ∞,1 if M is of type II∞ and p is a finite projection. 10 ROBERT PLUTA AND BERNARD RUSSO • II a ∞ if M is of type II∞ and p is a properly infinite projection. • III a if M is of type III and p is a (necessarily) properly infinite projection. Similarly, we also define types for pM for M of type I: 1 if M is finite of type I and p is (necessarily) finite. ∞,1 if M is of type I∞ and p is a finite projection. ∞ if M is of type I∞ and p is a properly infinite projection. • I a • I a • I a The following theorem involves the cases II a ∞, III a and when M is a factor, the cases I a 1 , I a 1,∞, and I a ∞. Theorem 3.3. Let X = pM be a TRO, where M is a von Neumann algebra and p is a projection in M. (i): If X is of type II a ∞ or III a, and has a separable predual, then every TRO-derivation of X is an inner TRO-derivation. (ii): If M is of type III and countably decomposable, then every TRO-derivation of X = pM is an inner TRO-derivation. (iii): If M = B(H) is a factor of type I, then (1) If dim H < ∞, then every TRO-derivation of X = pM is an inner TRO-derivation. (2) If dim pH = dim H, then every TRO-derivation of X = pM is an inner TRO-derivation. (3) If dim pH < dim H = ∞, then X = pM admits outer TRO-derivations. Proof. If M is a continuous von Neumann algebra with a separable predual and p is a properly infinite projection in M, then it is shown in [9, Theorem 5.16] that pM is triple isomorphic to a von Neumann algebra, and hence by Theorem 3.2, every triple derivation is an inner triple derivation in this case. Consequently, by Proposition 2.4(iv), every TRO-derivation is an inner TRO-derivation. (Another way to see this latter fact is to note that by [9, Lemma 5.15], pM is actually TRO- isomorphic to a von Neumann algebra, and to apply Corollary 2.3.) This proves (i). To prove (ii), we note first that if A is a von Neumann algebra with a projection p ∼ 1, then pA is TRO-isomorphic to A. Indeed, If u is a partial isometry in A with uu∗ = p and u∗u = 1, then x 7→ u∗x is a TRO-isomorphism from pA onto A. Now if A is of type III, then A := c(p)A is of type III, c(p) is the identity of A and pA = p A. Further, if A is countably decomposable, then by [15, 2.2.14], since in A, c(p) = 1 A = c(1 A), we have p ∼ 1 A, so A is TRO-isomorphic to p A = pA. DERIVATIONS ON TERNARY RINGS OF OPERATORS 11 Finally, let M = B(H). The first statement in (iii) follows from the fact that every finite dimensional semisimple Jordan triple system has only inner derivations. This result first appeared in [11, Chapter 11] (for an outline of a proof, see [14, Theorem 2.8,p. 136] and for the definitions of Jordan triple system and semisimple, see [3, Section 1.2]). If dim pH = dim H, then pM ≃ B(H) has only inner triple derivations by Theorem 3.2. On the other hand, if dim pH < dim H = ∞, then pM ≃ B(H, pH) has outer triple derivations, as shown in [7, Corollary 3]. By Proposition 2.4(iv), this proves (iii) (cid:3) Remark 3.4. Although it follows from Theorem 3.3, it is worth pointing out that the TROs B(C, H) and B(H, C) support outer TRO deriva- tions if and only if dim H = ∞. According to [9, Lemma 5.15], if B is a von Neumann algebra of type II∞ or III, and H is a separable Hilbert space, then B and B⊗B(C, H) are TRO-isomorphic. Corol- lary 2.3 shows that B⊗B(C, H) has only inner TRO-derivations and only inner triple derivations, although, as just noted, B(C, H) can have an outer TRO derivation and an outer triple derivation. This contrasts the situation of derivations on tensor products of C∗-algebras, as in [2, Proposition 3.2]). 3.2. W-TROs of types I,II,III. We begin by recalling some con- cepts from [13]. If R is a von Neumann algebra and e is a projection in R, then V := eR(1 − e) is a W-TRO. Conversely if V ⊂ B(K, H) is a W-TRO, then with V ∗ = {x∗ : x ∈ V } ⊂ B(H, K), M(V ) = XX ∗sot ⊂ B(H), N(V ) = X ∗X ⊂ B(K), we let sot RV = (cid:20)M(V ) V ∗ V N(V )(cid:21) ⊂ B(H ⊕ K) denote the linking von Neumann algebra of V . Then we have a SOT- continuous TRO-isomorphism V ≃ eRe⊥, where e = (cid:2) 1H 0 (cid:2) 0 0 0 1K(cid:3). A W-TRO V is stable if it is TRO-isomorphic to B(ℓ2)⊗V . A W- TRO is of type I,II,or III, by definition, if its linking von Neumann algebra is of that type as a von Neumann algebra. There is a further classification of the types I and II depending on the types of M(V ) and N(V ) leading to the types Im,n, IIα,β where m, n are cardinal numbers and α, β ∈ {1, ∞}. See [13, Section 4] for detail. 0 0(cid:3) and e⊥ = In what follows, for ultraweakly closed subspaces A ⊂ M and B ⊂ N, where M and N are von Neumann algebras, A⊗B denotes the ultraweak closure of the algebraic tensor product A ⊗ B. We shall use the following results from [13], which we summarize as a theorem. 12 ROBERT PLUTA AND BERNARD RUSSO Theorem 3.5 (Ruan [13]). Let V be a W∗-TRO acting on separable Hilbert spaces. (i) [13, Theorem 3.2] If V is a stable W-TRO, then V is TRO- isomorphic to M(V ) and to N(V ). (ii) [13, Corollary 4.3] If V is a W-TRO of one of the types I∞,∞, II∞,∞ or III, then V is a stable W-TRO, and hence TRO-isomorphic to a von Neumann algebra. (iii) [13, Theorem 4.4] If V is a W-TRO of type II1,∞ (respec- tively II∞,1), then V is TRO-isomorphic to B(H, C)⊗M (respectively B(C, H)⊗N), where M (respectively N) is a von Neumann algebra of type II1. Because taking a transpose is a triple isomorphism, we have the following consequence of Theorem 3.5(iii). Lemma 3.6. A W-TRO of type II1,∞ is triple isomorphic to a W- TRO of type II∞,1. More precisely, B(H, C)⊗M is triple isomorphic to B(C, H)⊗M t, where xt = Jx∗J, for x ∈ M ⊂ B(H) and J is a conjugation on H. Proposition 3.7. Let V be a W-TRO. (i): If V acts on a separable Hilbert space and is of one of the types I∞,∞, II∞,∞ or III, then every triple derivation of V is an inner triple derivation and every TRO derivation of V is an inner TRO-derivation. (ii): If every TRO-derivation of any W∗-TRO of type II1,∞ has only inner TRO-derivations, then every TRO-derivation of any W∗-TRO of type II∞,1 has only inner TRO-derivations. The converse also holds. Proof. (i) is an immediate consequence of Theorem 3.5(ii), Theorem 3.2 and Proposition 2.4(iv). (ii) is an immediate consequence of Lemma 3.6 and Proposition 2.4(iv). (cid:3) It follows from Remark 3.4 that if M is a von Neumann algebra of type II∞ or III and H is a separable Hilbert space, then B(C, H)⊗M is triple isomorphic to a von Neumann algebra and hence has only inner TRO-derivations, giving alternate proofs of parts of Proposition 3.7(i). By [13, Theorem 4.1], if V is a W-TRO of type I, then V is TRO- isomorphic to ⊕αL∞(Ωα)⊗B(Kα, Hα). In the next two results, we con- sider the related TRO C(Ω, B(H, K)), where Ω is a compact Hausdorff space. Lemma 3.8. Let E be a TRO and Ω a compact Hausdorff space. DERIVATIONS ON TERNARY RINGS OF OPERATORS 13 (i): If every TRO derivation of V := C(Ω, E) is an inner TRO derivation, then the same holds for E. (ii): If every triple derivation of V := C(Ω, E) is an inner triple derivation, then the same holds for E. Proof. By Proposition 2.4(iv), it is sufficient to prove (i). If D is a TRO derivation of E, then δf (ω) := D(f (ω)) is a TRO derivation of V , as is easily checked. Suppose every TRO derivation of V is an inner TRO derivation. Then δf = αf + f β, where α = j wj for some i for some xi, yi ∈ V , and β = −β∗ = Pi z∗ −α∗ = Pi xiy∗ zj, wj ∈ V . For a ∈ E, let 1 ⊗ a ∈ V be the constant function equal to a. Then D(a) = D((1 ⊗ a)(ω)) = δ(1 ⊗ a)(ω) = α(ω)a + aβ(ω) for all ω ∈ Ω. Since α(ω)∗ = −α(ω) ∈ EE∗ and β(ω)∗ = −β(ω) ∈ E∗E, D is an inner TRO derivation of E. (cid:3) Recall from Theorem 3.3(iii) that the TRO B(H, K) supports outer TRO derivations if and only if it is infinite dimensional and dim H 6= dim K. Proposition 3.9. If V = ⊕αC(Ωα, Eα), where Eα = B(Kα, Hα) and if every triple derivation of V is an inner triple derivation, then for every α, either dim Eα < ∞ or dim Kα = dim Hα. Proof. Let δ be a triple derivation of V , and let δα = δC(Ωα,Eα), which is a triple derivation of the weak-closed ideal C(Ωα, Eα). Then δ({fα}) = {δαfα}. Moreover if δ is an inner triple derivation, say δ = Pi δ(ai, bi) α) is an inner triple α, bi for ai = {ai derivation of C(Ωα, Eα). α}, bi = {bi α} ∈ V , then δα = Pi δ(ai Now suppose that every triple derivation of V is an inner triple derivation, and that for some α0, Eα0 is infinite dimensional and dim Kα0 does not equal dim Hα0. Then, as noted above, there is an outer triple derivation D of C(Ωα0, Eα0). Then the triple derivation on V which is zero on C(Ωα, Eα) for α 6= α0 and equal to D on C(Ωα0, Eα0), cannot be inner by the preceding paragraph, which is a contradiction. (cid:3) 4. Some questions left open Questions 1. It remains to complete the results of Theorem 3.3 to include the cases where p is a finite projection in a continuous von Neumann algebra, or when p is arbitrary and M is a general von Neu- mann algebra of type I. As a possible tool for the first question, we note that there is an alternate proof of Proposition 2.4 (ii), in the case X = pM, p finite, using the technique in [6, Section II.B]. 14 ROBERT PLUTA AND BERNARD RUSSO Questions 2. Besides the problem of extending the known cases to non separable Hilbert spaces, the cases left open in Proposition 3.7 for arbitrary W-TROs are those of types II1,1 and II1,∞ (the latter being equivalent to II∞,1). Questions 3. Let E be a W-TRO, and let V = ⊕αL∞(Ωα)⊗B(Kα, Hα) be a W-TRO of type I. • If every derivation of the W-TRO L∞(Ω)⊗E is inner, does it follow that every derivation of E is inner? • If every derivation of V is inner, does it follow that dim B(Kα, Hα) < ∞, for all α; or supα dim B(Kα, Hα) < ∞? • If supα dim B(Kα, Hα) < ∞, does it follow that every derivation of V is inner? Remark 4.1. With respect to Questions 3, (i) In the first bullet, if E had a separable predual, then a variant of [15, 1.22.13] would state that L∞(Ω)⊗E = L∞(Ω, E) and the technique in Lemma 3.8 could be used. (ii) In the first bullet, suppose that E = pM, with M a von Neumann algebra in B(H) and p a projection in M, and let D is a derivation of E. Then δ := id ⊗ D is a derivation of V = L∞(Ω)⊗E. Assuming that δ is inner, there exist α = −α∗ ∈ V V ∗ = L∞(Ω)⊗(EE∗) (EE∗ denoting the weak closure) and β ∈ V ∗V = L∞(Ω)⊗(E∗E), such that 1 ⊗ Dx = α(1 ⊗ x) + (1 ⊗ x)β, (x ∈ E). We have EE∗ = pMp ⊂ B(pH), E∗E ⊂ B(H), and L∞(Ω) ⊂ B(L2(Ω)). For each ϕ ∈ B(L2(Ω))∗, let Rϕ : B(L2(Ω)⊗pH) → B(pH) be the slice map of Tomiyama defined by Rϕ(f ⊗ x) = ϕ(f )x ([4, Lemma 7.2.2]). Since V V ∗ is the ultraweak closure of L∞(Ω) ⊗ EE∗, by the weak- continuity of Rϕ, we have ϕ(1)Dx = Rϕ(α)x + xRϕ(β) with Rϕ(α) ∈ EE∗ and Rϕ(β) ∈ E∗E. Thus, if dim H = dim pH, or if E is finite dimensional, then Rϕ(β) ∈ E∗E, so that D is an inner TRO-derivation, (take ϕ to be a normal state so that Rϕ is self-adjoint and Rϕ(α)∗ = −Rϕ(α) and Rϕ(β)∗ = −Rϕ(β).) In general, D could be called a "quasi-inner" TRO-derivation. (iii) In the second bullet, if each B(Kα, Hα) had a separable predual, then a variant of [15, 1.22.13] would state that L∞(Ωα)⊗B(Kα, Hα) = L∞(Ωα, B(Kα, Hα)) and the technique in Proposition 3.9 could be used. DERIVATIONS ON TERNARY RINGS OF OPERATORS 15 References [1] T. J. Barton and Y. Friedman, Bounded derivations of JB∗-triples, Quart. J. Math. Oxford Ser. (2) (1990), 255 -- 268. [2] C. J. K. Batty, Derivations of tensor products of C-algebras, J. London Math. Soc. (2) 17 (1978), 129 -- 140. [3] Cho-Ho Chu, Jordan structures in geometry and analysis, Cambridge Tracts in Mathematics, vol. 190, Cambridge University Press, Cambridge, 2012. [4] E. G. Effros and Z. J. Ruan, Operator Spaces, London Mathematical Society Monographs, vol. 23, Clarendon Press Oxford, 2000. [5] M. Hamana, Triple envelopes and Silov boundaries of operator spaces, Math. J. Toyama Univ. 22 (1999), 77 -- 93. [6] J. Hamhalter, K. K. Kudaybergenov, A. M. Peralta, and B. Russo, Bounded- ness of completely additive measures with application to 2-local triple deriva- tions, J. Math. Physics 57 (2016), no. 2, 22 pp. [7] T. Ho, J. Mart´ınez-Moreno, A. M. Peralta, and B. Russo, Derivations on real and complex JB∗-triples, J. London Math. Soc. (2) 65 (2002), no. 1, 85 -- 102. [8] T. Ho, A. M. Peralta, and B. Russo, Ternary weakly amenable C-algebras and JB-triples, Quarterly J. Math. 64 (2013), 1109 -- 1139. [9] G. Horn and E. Neher, Classification of continuous JBW∗-triples, Trans. Amer. Math. Soc. 306 (1988), 553 -- 578. [10] Manmohan Kaur and Zhong-Jin Ruan, Local properties of ternary rings of operators and their linking C-algebras, J. Functional Analysis 195 (2002), 262 -- 305. [11] Kurt Meyberg, Lectures on algebras and triple systems, The University of Vir- ginia, Charlottesville, Va., 1972. [12] Robert Pluta and Bernard Russo, Triple derivations on von Neumann algebras, Studia Math. 226 (2015), no. 1, 57-73. [13] Zhong-Jin Ruan, Type decomposition and the rectangular AFD property for W∗-TROs, Canad. J. Math. 36 (2004), no. 4, 843 -- 870. [14] Bernard Russo, Derivations and projections on Jordan triples: an introduc- tion to nonassociative algebra, continuous cohomology, and quantum functional analysis, In: Advanced courses of mathematical analysis V (2016), 118 -- 227. World Sci. Publ., Hackensack, NJ. [15] S. Sakai, C-algebras and W*-algebras, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 60, Springer-Verlag, New York Heidelberg Berlin, 1971. [16] A. M. Sinclair, Jordan homomorphisms and derivations on semisimple Banach algebras, Proc. Amer. Math. Soc. 24 (1970), 209 -- 214. [17] W. Timmermann, Remarks on automorphism and derivation pairs in ternary rings of unbounded operators, Arch. Math. (Basel) 74 (2000), no. 5, 379 -- 384. [18] Borut Zalar, On the structure of automorphism and derivation pairs of B ∗- triple systems, Topics in operator theory, operator algebras and applications (Timisoara 1994), Rom. Acad., Bucharest (1995), 265 -- 271. 16 ROBERT PLUTA AND BERNARD RUSSO Department of Mathematics, University of California, Irvine, CA 92697-3875, USA E-mail address: [email protected] Department of Mathematics, University of California, Irvine, CA 92697-3875, USA E-mail address: [email protected]
1804.01930	1	1804	2018-04-05T16:06:28	Operator-valued Triebel-Lizorkin spaces	[ "math.OA", "math.FA" ]	This paper is devoted to the study of operator-valued Triebel-Lizorkin spaces. We develop some Fourier multiplier theorems for square functions as our main tool, and then study the operator-valued Triebel-Lizorkin spaces on $\mathbb{R}^d$. As in the classical case, we connect these spaces with operator-valued local Hardy spaces via Bessel potentials. We show the lifting theorem, and get interpolation results for these spaces. We obtain Littlewood-Paley type, as well as the Lusin type square function characterizations in the general way. Finally, we establish smooth atomic decompositions for the operator-valued Triebel-Lizorkin spaces. These atomic decompositions play a key role in our recent study of mapping properties of pseudo-differential operators with operator-valued symbols.	math.OA	math	OPERATOR-VALUED TRIEBEL-LIZORKIN SPACES RUNLIAN XIA AND XIAO XIONG 8 1 0 2 r p A 5 ] Abstract. This paper is devoted to the study of operator-valued Triebel-Lizorkin spaces. We develop some Fourier multiplier theorems for square functions as our main tool, and then study the operator-valued Triebel-Lizorkin spaces on Rd. As in the classical case, we connect these spaces with operator-valued local Hardy spaces via Bessel potentials. We show the lifting the- orem, and get interpolation results for these spaces. We obtain Littlewood-Paley type, as well as the Lusin type square function characterizations in the general way. Finally, we establish smooth atomic decompositions for the operator-valued Triebel-Lizorkin spaces. These atomic decompositions play a key role in our recent study of mapping properties of pseudo-differential operators with operator-valued symbols. Contents . A O h t a m [ 1 v 0 3 9 1 0 . 4 0 8 1 : v i X r a Introduction and preliminaries 0. Noncommutative Lp-spaces Fourier analysis 1. Operator-valued local Hardy spaces Calder´on-Zygmund theory Characterizations Atomic decomposition 2. Multiplier theorems 2.1. Global multipliers 2.2. Hilbert-valued multipliers 2.3. Multipliers on hc p 3. Operator-valued Triebel-Lizorkin spaces 3.1. Definitions and basic properties 3.2. 3.3. Triebel-Lizorkin spaces with α > 0 4. Characterizations 4.1. General characterizations 4.2. Characterizations via Lusin functions 5. Smooth atomic decomposition 5.1. Smooth atomic decomposition of hc 5.2. Atomic decomposition for F α,c References Interpolation 1(Rd,M) 1 (Rd,M) 1 3 3 4 6 6 7 8 8 9 15 18 18 22 23 24 24 31 32 32 36 43 0. Introduction and preliminaries 2 ≤ ξ ≤ 2}, ϕ > 0 on {ξ : 1 Let ϕ be a Schwartz function on Rd such that supp ϕ ⊂ {ξ : 1 ξ < 2}, and Pk∈Z ϕ(2−kξ) = 1 for all ξ 6= 0. For each k ∈ N, let ϕk be the function whose Fourier transform is equal to ϕ(2−k·), and let ϕ0 be the function whose Fourier transform is equal to 1 −Pk>0 ϕ(2−k·). Then {ϕk}k≥0 gives a Littlewood-Paley decomposition on Rd. The classical (inhomogeneous) Triebel-Lizorkin spaces F α p,q(Rd) for 0 < p < ∞, 0 < q ≤ ∞ and α ∈ R are 2 < 2000 Mathematics Subject Classification: Primary: 46L52, 42B30. Secondary: 46L07, 47L65. Key words: Noncommutative Lp-spaces, operator-valued Triebel-Lizorkin spaces, operator-valued Hardy spaces, Fourier multipliers, interpolation, characterizations, atomic decomposition. 1 2 R. Xia and X. Xiong defined as with the (quasi-)norm p,q(Rd) = {f ∈ S′(Rd) : kfkF α F α p,q < ∞} kfkF α p,q =(cid:13)(cid:13)(Xj≥0 2qjαϕj ∗ fq) . 1 q(cid:13)(cid:13)p We refer the reader to Triebel's books [34] and [35] for more concrete definition and properties of Triebel-Lizorkin spaces on Rd. This kind of function spaces is closely related to other function spaces, such as Sobolev and Besov spaces. In particular, Triebel-Lizorkin spaces can be viewed as generalizations of Hardy spaces, since the Bessel potential J α is known to be an isomorphism between F α p,2(Rd) and hp(Rd) (local Hardy spaces introduced in [8]). All these spaces are ba- sic for many branches of mathematics such as harmonic analysis, PDE, functional analysis and approximation theory. This paper is devoted to the study of operator-valued Triebel-Lizorkin spaces. As in the classical case, it can be viewed as an extension of our recent work [36] on operator-valued local Hardy spaces. On the other hand, the Triebel-Lizorkin spaces studied here are Euclidean counterparts of those on usual and quantum tori studied in [40]. Our main motivation is to build a kind of function spaces where we can carry out the investigation of pseudo-differential operators with operator-valued symbols. 1 Due to noncommutativity, there are several obstacles on our route, which do not appear in the classical case. First of all, in the noncommutative integration, the simple replacement of the usual absolute value by the modulus of operators in the formula (cid:13)(cid:13)(Pj≥0 2qjαϕj ∗ fq) does not give a norm except for q = 2. Even though one could use Pisier's definition of ℓq- valued noncommutative Lp-spaces by complex interpolation (see [24]), we will not study that kind of spaces and will focus only on the case q = 2. The reason for this choice is that, for q = 2, the Triebel-Lizorkin spaces of operator-valued distributions are isomorphic to the Hardy spaces developed in [36], as mentioned above. Another difficulty is the lack of pointwise maximal functions in the noncommutative case. As is well known, the maximal functions play a crucial role in the classical theory; but they are no longer at our disposal in the noncommutative setting. In [40], when studying the Triebel-Lizorkin spaces on quantum tori, we use Calder´on-Zygmund and Fourier multiplier theory as substitution. In this paper, we will still rely heavily on this theory. However, we have to consider its local (or inhomogeneous) counterpart, since the theory used in [40] for quantum tori are nonlocal (or homogeneous) ones. Besides the local nature, we also develop Hilbert space valued Fourier multiplier theory, which will be used to deduce general characterizations of operator-valued Triebel-Lizorkin spaces by the Lusin type square function. q(cid:13)(cid:13)p Our definition of (column) operator-valued Triebel-Lizorkin spaces is (Rd,M) = {f ∈ S′(Rd; L1(M) + M) : kfkF α,c F α,c p p < ∞}, where kfkF α,c p =(cid:13)(cid:13)(Xj≥0 22jαϕj ∗ f2) . 1 2(cid:13)(cid:13)p p Here the norm k · kp is the norm of the semi-commutative Lp-space Lp(L∞(Rd)⊗M). Different from the classical case, we have also row and mixture versions; see section 3 for concrete definitions. We present here two major results of this paper. The first one gives general characterizations of F α,c (Rd,M) by any reasonable convolution kernels in place of the Littlewood-Paley decom- position {ϕj}j≥0. These characterizations can be realized either by the Littlewood Paley type g-function or by the Lusin type integral function, with the help of the Calder´on-Zygmund theory and Fourier multiplier theory mentioned above. The second major result is the atomic decom- position of F α,c (Rd,M). When α = 0, in [36], we deduce from the h1-bmo duality an atomic decomposition of hc 1(Rd,M), which does not require any smooth condition on each atom. In this paper, we refine the smoothness of that atomic decomposition by the Calder´on reproducing identity, via tent spaces. Using the same trick, we extend that refinement to F α,c (Rd,M); but compared with the case of local Hardy spaces, subatoms enter in the game. These smooth atomic decom- positions will play a crucial role in the study of pseudo-differential operators in the forthcoming paper [37]. 1 1 Inhomogeneous Triebel-Lizorkin spaces 3 In the following, let us recall some notation and background in the interface between harmonic analysis and operator algebras that we will need throughout the paper, although they are probably well-known to experts. Noncommutative Lp-spaces. We start with a brief introduction of noncommutative Lp spaces. Let M be a von Neumann algebra equipped with a normal semifinite faithful trace τ ; for 1 ≤ p ≤ ∞, let Lp(M) be the noncommutative Lp-space associated to (M, τ ). The norm of Lp(M) will be often denoted simply by k · kp. But if different Lp-spaces appear in a same context, we will sometimes precise the respective Lp-norms in order to avoid possible ambiguity. The reader is referred to [26] and [41] for more information on noncommutative Lp-spaces. We will also need Hilbert space- valued noncommutative Lp-spaces (see [14] for more details). Let H be a Hilbert space and v ∈ H with kvk = 1. Let pv be the orthogonal projection onto the one-dimensional subspace generated by v. Define Lp(M; H r) = (pv ⊗ 1M)Lp(B(H)⊗M) and Lp(M; H c) = Lp(B(H)⊗M)(pv ⊗ 1M). These are the row and column noncommutative Lp-spaces. Like the classical Lp-spaces, noncom- mutative Lp-spaces form an interpolation scale with respect to the complex interpolation method: For 1 ≤ p0 < p1 ≤ ∞ and 0 < η < 1, we have (cid:0)Lp0(M), Lp1 (M)(cid:1)η = Lp(M) with equal norms, p = 1−η where 1 for the same indicies, we have + η p1 p0 . Since Lp(M; H c) and Lp(M; H r) are 1-complemented subspaces of Lp(B(H)⊗M), (cid:0)Lp0 (M; H c), Lp1(M; H c)(cid:1)η = Lp(M; H c) with equal norms. Fourier analysis. Fourier multipliers will be one of the most important tools of this paper. Let us give some Fourier multipliers that will be frequently used. They are all very well known in the classical harmonic theory. First, we recall the symbols of Littlewood-Paley decomposition on Rd. Fix a Schwartz function ϕ on Rd satisfying: (0.1)  2 ≤ ξ ≤ 2}. 2 < ξ < 2}, supp ϕ ⊂ {ξ : 1 ϕ > 0 on {ξ : 1 Pk∈Z ϕ(2−kξ) = 1,∀ ξ 6= 0. For each k ∈ N, let ϕk be the function whose Fourier transform is equal to ϕ(2−k·), and let ϕ0 be the function whose Fourier transform is equal to 1 −Pk>0 ϕ(2−k·). Then {ϕk}k≥0 gives a Littlewood-Paley decomposition on Rd such that suppbϕk ⊂ {ξ ∈ Rd : 2k−1 ≤ ξ ≤ 2k+1}, ∀ k ∈ N, suppbϕ0 ⊂ {ξ ∈ Rd : ξ ≤ 2} (0.2) and that (0.3) ∞Xk=0bϕk(ξ) = 1 ∀ ξ ∈ Rd. Xk∈Zbϕk(ξ) = 1 ∀ ξ 6= 0. The homogeneous counterpart of the above decomposition is given by { ϕk}k∈Z. This time, for (0.4) every k ∈ Z, these functions are given by bϕk(ξ) = ϕ(2−kξ). We have The Bessel potential and the Riesz potential are J α = (1− (2π)−2∆) 2 and I α = (−(2π)−2∆) 2 , respectively. If α = 1, we will abbreviate J 1 as J and I 1 as I. We denote also Jα(ξ) = (1 + ξ2) α on Rd and Iα(ξ) = ξα on Rd\{0}. Then Jα(ξ) and Iα(ξ) are the symbols of the Fourier multipliers J α and I α, respectively. Given a Banach space X, let S(Rd; X) be the space of X-valued rapidly decreasing functions on Rd with the standard Fr´echet topology, and S′(Rd; X) be the space of continuous linear maps from S(Rd) to X. All operations on S(Rd) such as derivations, convolution and Fourier transform transfer to S′(Rd; X) in the usual way. On the other hand, Lp(Rd; X) naturally embeds into S′(Rd; X) for 1 ≤ p ≤ ∞, where Lp(Rd; X) stands for the space of strongly p-integrable functions α α 2 4 R. Xia and X. Xiong from Rd to X. By this definition, Fourier multipliers on Rd, in particular the Bessel and Riesz potentials, extend to vector-valued tempered distributions in a natural way. We denote by H σ 2 (Rd) the potential Sobolev space, consisting of all tempered distributions f such that J σ(f ) ∈ L2(Rd). If σ > d 2 , we have (cid:13)(cid:13)F−1(f )(cid:13)(cid:13)1 =Zs≤1(cid:12)(cid:12)F−1(f )(s)(cid:12)(cid:12)ds +Xk≥0Z2k<s≤2k+1(cid:12)(cid:12)F−1(f )(s)(cid:12)(cid:12)ds 22kσZ2k<s≤2k+1(cid:12)(cid:12)F−1(f )(s)(cid:12)(cid:12)2 2 ds(cid:17) 1 , ds +Xk≥0 ≤ C1(cid:16)Zs≤1(cid:12)(cid:12)F−1(f )(s)(cid:12)(cid:12)2 ≤ C2(cid:13)(cid:13)f(cid:13)(cid:13)Hσ kφ ∗ gkLp(Rd;X) ≤ kφk1kgkLp(Rd;X) ≤ C2kbφkHσ 2 where C1 and C2 are uniform constants. Therefore, ifbφ ∈ H σ (0.5) holds for any g ∈ Lp(Rd; X) with 1 ≤ p ≤ ∞. Here X is an arbitrary Banach space. Inequality (0.5) indicates that functions in H σ 2 (Rd) are the symbols of bounded Fourier multipliers, even in the vector-valued case. 2 kgkLp(Rd;X) 2 (Rd), the following Young's inequality In the sequel, we will mainly consider the case X = L1(M) + M, i.e., consider operator- valued functions or distributions on Rd. We will frequently use the following Cauchy-Schwarz type inequality for operator-valued square function, (0.6) where φ : Rd → C and f : Rd → L1(M) + M are functions such that all integrations of the above inequality make sense. We also require the operator-valued version of the Plancherel formula. For sufficiently nice functions f : Rd → L1(M) + M, for example, for f ∈ L2(Rd) ⊗ L2(M), we have (0.7) (cid:12)(cid:12)ZRd ≤ZRd φ(s)2dsZRd f (s)2ds, φ(s)f (s)ds(cid:12)(cid:12)2 ZRd f (s)2ds =ZRd bf (ξ)2dξ. Throughout, we will use the notation A . B, which is an inequality up to a constant: A ≤ cB for some constant c > 0. The relevant constants in all such inequalities may depend on the dimension d, the test function Φ or p, etc, but never on the function f in consideration. The equivalence A ≈ B will mean A . B and B . A simultaneously. The layout of this paper is the following. In the next section, we briefly introduce the definition of local Hardy spaces, and the main results in [36]. In section 2, we develop several Fourier multiplier theorems: the first one is the inhomogeneous version of the Fourier multiplier theorem proved in [40], fitted to local Hardy spaces; the second is a Hilbertian Fourier multiplier theorem, in order to deal with the Lusin area square functions. In section 3, we give the definition of Triebel- Lizorkin spaces, and some immediate properties. Section 4 is devoted to different characterizations of Triebel-Lizorkin spaces. The proofs in this section are technical and tedious, based on Calder´on- Zygmund theory and Fourier multiplier theorems. In the last section, we demonstrate the smooth atomic decompositions of F α,c p(Rd,M), and then extend the result to general α by a similar argument. (Rd,M): we begin with the space F 0,c p (Rd,M) = hc p 1. Operator-valued local Hardy spaces Let us review the operator-valued local Hardy spaces studied in [36], and collect some of the main results there that will be useful in this paper. We keep the following notation: (M, τ ) is trace, and N = L∞(Rd)⊗M is equipped with the tensor a von Neumann algebra with n.s.f. trace; letters s, t are used to denote variables of Rd, while letters x, y are reserved for operators in noncommutative Lp-spaces. Let P be the Poisson kernel on Rd: P(s) = cd 1 (s2 + 1) d+1 2 Inhomogeneous Triebel-Lizorkin spaces 5 with cd the usual normalizing constant and s the Euclidean norm of s. Let Pε(s) = ) = cd 1 εd P( s ε ε . d+1 2 (s2 + ε2) For any function f on Rd with values in L1(M) + M, its Poisson integral, whenever it exists, will be denoted by Pε(f ): The truncated Lusin area square function of f by Pε(f )(s) =ZRd sc(f )(s) =(cid:16)ZeΓ(cid:12)(cid:12) ∂ where eΓ is the truncated cone {(t, ε) ∈ Rd+1 p(Rd,M) = {f ∈ L1(M; Rc hc L2(Rd, ∂ε dt + where the hc p(Rd,M)-norm of f is defined by Pε(s − t)f (t)dt, (s, ε) ∈ Rd+1 + . εd−1(cid:17) 1 Pε(f )(s + t)(cid:12)(cid:12)2 dtdε , s ∈ Rd, 2 1+td+1 ). For 1 ≤ p < ∞, define the column local Hardy space hc d) : kfkhc d) + L∞(M; Rc p(Rd,M) to be < ∞}, p : t < ε < 1}. Denote by Rd the Hilbert space kfkhc The row local Hardy space hr p = kf∗khc with the norm kfkhr p(Rd,M) = ksc(f )kLp(N ) + kP ∗ fkLp(N ). p(Rd,M) is the space of all f such that f∗ ∈ hc p. Moreover, define the mixture space hp(Rd,M) as follows: p(Rd,M), equipped equipped with the sum norm hp(Rd,M) = hc p(Rd,M) + hr p(Rd,M) for 1 ≤ p ≤ 2 kfkhp(Rd,M) = inf{kgkhc p + khkhr p : f = g + h, g ∈ hc p(Rd,M), h ∈ hr p(Rd,M)}, and hp(Rd,M) = hc equipped with the intersection norm p(Rd,M) ∩ hr p(Rd,M) for 2 < p < ∞ The local analogue of the Littlewood-Paley g-function of f is defined by kfkhp = max{kfkhc p ,kfkhr p}. gc(f )(s) =(cid:0)Z 1 ∂ ∂ε 0 Pε(f )(s)2εdε(cid:1) 1 p ≈ kgc(f )kp + kP ∗ fkp kfkhc 2 , s ∈ Rd. It is proved in [36] that for all 1 ≤ p < ∞. The dual of hc any cube Q ⊂ Rd, we denote its volume by Q. Let f ∈ L∞(M; Rc is denoted by fQ := 1 1(Rd,M) is characterized as a local version of bmo space, defined as follows. For d). The mean value of f over Q QRQ f (s)ds. Set kfkbmoc(Rd,M) = maxn sup Q<1(cid:13)(cid:13)( The local version of bmo spaces are defined as (1.1) 1 QZQ f − fQ2dt) 1 2(cid:13)(cid:13)M , sup Q=1(cid:13)(cid:13)(ZQ f2dt) 1 2(cid:13)(cid:13)Mo. bmoc(Rd,M) = {f ∈ L∞(M; Rc d) : kfkbmoc < ∞}. Define bmor(Rd,M) to be the space of all f ∈ L∞(M; Rr d) such that kf∗kbmoc(Rd,M) is finite, with the norm kfkbmor = kf∗kbmoc . And bmo(Rd,M) is defined as the intersection of bmoc(Rd,M) and bmor(Rd,M), equipped with the intersection norm. The above Hardy and bmo type spaces are local analogues of the spaces studied by Mei [18]. They turn out to have similar properties with their non-local versions, such as duality and interpolation. The following two theorems are quoted from [36]. Theorem 1.1. We have hc is its conjugate index, then hc 1(Rd,M)∗ = bmoc(Rd,M) with equivalent norms. If 1 < p < 2 and q p(Rd,M)∗ = hc q(Rd,M) with equivalent norms. 6 R. Xia and X. Xiong Theorem 1.2. Let 1 < p < ∞. We have (1) (cid:0)bmoc(Rd,M), hc (2) (cid:0)X, Y(cid:1) 1 p 1(Rd,M)(cid:1) 1 p = hc p(Rd,M). = Lp(N ), where X = bmo(Rd,M) or L∞(N ), and Y = h1(Rd,M) or L1(N ). Calder´on-Zygmund theory. The usual Calder´on-Zygmund operators which satisfy the Hormander condition are not necessarily bounded on local Hardy spaces. In order to guarantee the bound- p(Rd,M), an extra decay at infinity is imposed on edness of a Calder´on-Zygmund operator on hc the kernel in [36]. Let K ∈ S′(Rd; L1(M) + M) coincide on Rd \ {0} with a locally integrable L1(M) + M-valued function. We define the left singular integral operator K c associated to K by K(s − t)f (t)dt, and the right singular integral operator K r associated to K by K c(f )(s) =ZRd K r(f )(s) =ZRd f (t)K(s − t)dt. Both K c(f ) and K r(f ) are well-defined for sufficiently nice functions f with values in L1(M)∩M, for instance, for f ∈ S ⊗ (L1(M) ∩ M ). 0(Rd,M) denote the subspace of bmoc(Rd,M) consisting of compactly supported func- tions. The extra decay of the kernel K given in [36] is condition (2) in the following lemma. Let bmoc Lemma 1.3. Assume that (1) the Fourier transform of K is bounded: supξ∈Rd kbK(ξ)kM < ∞; (2) K satisfies a size estimate: there exist C1 and ρ > 0 such that kK(s)kM ≤ , ∀s ≥ 1; C1 sd+ρ (3) K has the Lipschitz regularity: there exist a constant C2 and γ > 0 such that Then K c is bounded on hc A similar statement also holds for K r and the corresponding row spaces. kK(s − t) − K(s)kM ≤ C2 p(Rd,M) for 1 ≤ p < ∞ and from bmoc s − td+γ , ∀s > 2t. 0(Rd,M) to bmoc(Rd,M). tγ Characterizations. Next, we are going to present the characterizations of local Hardy spaces obtained in [36], which will play an important role when studying the characterizations of Triebel- Lizorkin spaces in this paper. The main idea of these characterizations is to replace the Poisson kernel by good enough Schwartz ε ) for functions. Let Φ be a Schwartz function on Rd of vanishing mean, and set Φε(s) = ε−dΦ( s positive ε. Φ is said to be nondegenerate if: 0 = 1, (1.4) (1.2) (1.3) dε ε Then there exists a Schwartz function Ψ of vanishing mean such that ∀ξ ∈ Rd \ {0} ∃ ε > 0 s.t. bΦ(εξ) 6= 0. Z ∞ bΦ(εξ)bΨ(εξ) ∀ξ ∈ Rd \ {0} . Furthermore, we can find two functions φ, ψ such that bφ,bψ ∈ H σ bφ(ξ)bψ(ξ) = 1 −Z 1 0 bΦ(εξ)bΨ(εξ) Φ(f )(s) =(cid:16)ZZeΓ Φε ∗ f (s + t)2 dtdε εd+1(cid:17) 1 Φ(f )(s) =(cid:16)Z 1 ε (cid:17) 1 0 Φε ∗ f (s)2 dε , s ∈ Rd. For any f ∈ L1(M; Rc functions of f associated to Φ by d) + L∞(M; Rc dε ε sc gc 2 2 , s ∈ Rd, 2 (Rd), bφ(0) > 0,bψ(0) > 0 and . d), we define the local versions of the conic and radial square Fix the four test functions Φ, Ψ, φ, ψ as above. The following theorem is proved in [36]. Inhomogeneous Triebel-Lizorkin spaces 7 p(Rd,M) if and only if sc Theorem 1.4. Let 1 ≤ p < ∞ and φ, Φ be as above. For any f ∈ L1(M; Rc Φ(f ) ∈ Lp(N ) and φ ∗ f ∈ Lp(N ) if and only if gc f ∈ hc φ ∗ f ∈ Lp(N ). If this is the case, then (1.5) Φ(f )kp + kφ ∗ fkp ≈ kgc with the relevant constants depending only on d, Φ and φ. Φ(f )kp + kφ ∗ fkp p ≈ ksc kfkhc d) + L∞(M; Rc d), Φ(f ) ∈ Lp(N ) and We have a discrete version of Theorem 1.4. The square functions sc Φ and gc Φ can be discretized as follows: f = λj aj, ∞Xj=1 where the aj's are hc the sense of distribution. We equip hc 1-atoms and λj ∈ C such thatP∞j=1 λj < ∞. The above series converges in 1,at(Rd,M) with the following norm: kfkhc 1,at = inf{ λj : f = λj aj; aj's are hc 1 -atoms, λj ∈ C}. ∞Xj=1 ∞Xj=1 Similarly, we can define the row and mixture versions. The following theorem is also proved in [36]. gc,D Φ (f )(s) =(cid:16)Xj≥1 Φ (f )(s) =,(cid:16)Xj≥1 sc,D 2 , Φj ∗ f (s)2(cid:17) 1 2djZB(s,2−j ) Φj ∗ f (t)2dt(cid:17) 1 2 . Here Φj is the inverse Fourier transform of Φ(2−j·). This time, to get a resolvent of the unit on Rd, we need to assume that Φ satisfies Then adapting the proof of [32, Lemma V.6] , we can find a Schwartz function Ψ of vanishing mean such that ∀ ξ ∈ Rd \ {0} ∃ 0 < 2a ≤ b < ∞ s.t. bΦ(εξ) 6= 0, ∀ ε ∈ (a, b]. (1.6) +∞Xj=−∞bΦ(2−jξ)bΨ(2−jξ) = 1, Again, there exist two functions φ and ψ such that bϕ,bψ ∈ H σ ∞Xj=1bΦ(2−jξ)bΨ(2−jξ) +bφ(ξ)bψ(ξ) = 1, (1.7) ∀ξ ∈ Rd \ {0}. 2 (Rd), bφ(0) > 0,bψ(0) > 0 and ∀ξ ∈ Rd. Now we fix the pairs (Φ, Ψ) and (φ, ψ) satisfying (1.6) and (1.7). Theorem 1.5. Let φ and Φ be test functions as in (1.7) and 1 ≤ p < ∞. Then for any f ∈ L1(M; Rc Φ (f ) ∈ Lp(N ) and φ∗ f ∈ Lp(N ) if and only if gc,D d) + L∞(M; Rc Φ (f ) ∈ Lp(N ) and φ ∗ f ∈ Lp(N ). Moreover, p(Rd,M) if and only if sc,D d), f ∈ hc kfkhc p ≈ ksc,D Φ (f )kLp(N ) + kφ ∗ fkp ≈ kgc,D Φ (f )kp + kφ ∗ fkp with the relevant constants depending only on d, Φ and φ. Atomic decomposition. Finally, let us include the atomic decomposition of the local Hardy space hc 1-atom associated with Q is a function a ∈ L1(M; Lc 1(Rd,M). Let Q be a cube in Rd with Q ≤ 1. If Q = 1, an hc • supp a ⊂ Q; 2(Rd)) such that 2 ≤ Q− 1 2 . If Q < 1, we assume additionally: • τ(cid:0)RQ a(s)2ds(cid:1) 1 • RQ a(s)ds = 0. Let hc 1,at(Rd,M) be the space of all f admitting a representation of the form 8 R. Xia and X. Xiong Theorem 1.6. We have hc 1,at(Rd,M) = hc 1(Rd,M) with equivalent norms. Remark 1.7. In the above definition of atoms, we can replace the support of atoms Q by any bounded multiple of Q. 2. Multiplier theorems We are going to develop some Fourier multiplier theorems in this section. They can be viewed as a special case of Calder´on-Zygmund theory, and will be used to investigate various square funtions that characterize the Triebel-Lizorkin spaces. Our presentation follows closely the argument in Section 4.1 of [40]. Recall again that ϕ is a fixed function satisfying (0.1), ϕ0 is the inverse Fourier transform of 1 −Pk>0 ϕ(2−k·), and ϕk is the inverse Fourier transform of ϕ(2−k·) when k > 0. Moreover, we denote by ϕ(k) the Fourier transform of ϕk for every k ∈ N0 (N0 being the set of nonnegative integers). 2.1. Global multipliers. Firstly, let us state the following homogeneous version of [40, Theo- rem 4.1]. Theorem 2.1. Let σ ∈ R with σ > d functions on Rd\{0} such that 2 . Assume that (φj )j∈Z and (ρj )j∈Z are two sequences of and supp φj ρj ⊂ {ξ : 2j−1 ≤ ξ ≤ 2j+1}, j ∈ Z kφj(2j+k·)ϕkHσ 2 (Rd) < ∞. sup j∈Z −2≤k≤2 Let 1 < p < ∞. Then for any f ∈ S′(Rd; L1(M) + M), we have kφj(2j+k·)ϕkHσ 22jα φj ∗ ρj ∗ f2) 1 where the constant depends on p, σ, d and ϕ. (cid:13)(cid:13)(Xj∈Z (cid:13)(cid:13)(Xj≥K 2(cid:13)(cid:13)p . sup j∈Z −2≤k≤2 1 2(cid:13)(cid:13)p . sup j∈Z −2≤k≤2 2(cid:13)(cid:13)(Xj∈Z 22jαρj ∗ f2) , 1 2(cid:13)(cid:13)p 2(cid:13)(cid:13)(Xj≥K , 1 2(cid:13)(cid:13)p Proof. Without loss of generality, we may take α = 0. It suffices to show that for any integer K, (2.1) φj ∗ ρj ∗ f2) kφj(2j+k·)ϕkHσ ρj ∗ f2) with the relevant constant independent of K ∈ Z. To this end, we set By easy computation, we have ψj−K = φj(2K·), ηj−K = ρj(2K·), and bg = bf (2K·). supp ψjηj ⊂ {ξ : 2j−1 ≤ ξ ≤ 2j+1}, ∀ j ≥ 0, φj ∗ ρj ∗ f = 2dK ψj−K ∗ ρj−K ∗ g(2K·). φj ∗ ρj ∗ f2) ρj ∗ f2) (cid:13)(cid:13)(Xj≥K (cid:13)(cid:13)(Xj≥K (p−1)dK p (p−1)dK p 1 2(cid:13)(cid:13)p = 2 2(cid:13)(cid:13)p = 2 1 (cid:13)(cid:13)(Xj≥0 (cid:13)(cid:13)(Xj≥0 . 1 2(cid:13)(cid:13)p ψj ∗ ηj ∗ g2) ηj ∗ g2) . 1 2(cid:13)(cid:13)p and This ensures (2.2) Similarly, (2.3) Moreover, since ψj(2j+k·) = φj+K (2j+k+K·), we have 2 = sup j≥K −2≤k≤2 ≤ sup j∈Z −2≤k≤2 kψj(2j+k·)ϕkHσ −2≤k≤2 sup j≥0 (2.4) kφj(2j+k·)ϕkHσ 2 kφj(2j+k·)ϕkHσ 2 . Inhomogeneous Triebel-Lizorkin spaces 9 Now applying [40, Theorem 4.1] to ψj, ρj and g defined above, we obtain (cid:13)(cid:13)(Xj≥0 ψj ∗ ηj ∗ g2) 1 2kp . sup j≥0 −2≤k≤2 kψj(2j+k·)ϕkHσ ηj ∗ g2) 1 2 kp. 2 ((cid:13)(cid:13)(Xj≥0 Putting (2.2), (2.3) and (2.4) into this inequality, we then get (2.1), which yields Theorem 2.1 by approximation. (cid:3) Theorem 2.1 is developed to deal with the multiplier problem of square functions, and also the p(Rd,M) by virtue of their characterizations (see [38]). In multiplier problem of the Hardy spaces Hc order to deal with the corresponding problems on the inhomogeneous versions of square functions or Hardy spaces, we need the following global version of Theorem 2.1. The main difference is that in the inhomogeneous case, we need a careful analysis of the convolution kernel near the origin. Theorem 2.2. Let 1 < p < ∞, α ∈ R and σ > d sequences of functions on Rd such that 2 . Assume that (φj)j≥0 and (ρj)j≥0 are two supp (φj ρj) ⊂ {ξ ∈ Rd : 2j−1 ≤ ξ ≤ 2j+1}, j ∈ N, supp (φ0ρ0) ⊂ {ξ ∈ Rd : ξ ≤ 2}, and (2.5) kφj(2j+k·)ϕkHσ 2 (Rd) < ∞ and kφ0(ϕ(0) + ϕ(1))kHσ 2 (Rd) < ∞. sup j≥1 −2≤k≤2 Then for any L1(M) + M-valued distribution f , 22jα φj ∗ ρj ∗ f2) (cid:13)(cid:13)(Xj≥0 kφj(2j+k·)ϕkHσ 2 2(cid:9) ,kφ0(ϕ(0) + ϕ(1))kHσ Proof. This theorem follows easily from its homogeneous version, i.e., Theorem 2.1. Indeed, we where the constant depends only on p, σ, d and ϕ. can divide(cid:13)(cid:13)(Pj≥0 22jα φj ∗ ρj ∗ f2) 22jα φj ∗ ρj ∗ f2) 1 (cid:13)(cid:13)(Xj≥0 (cid:13)(cid:13)(Xj≥1 22jα φj ∗ ρj ∗ f2) , 1 2(cid:13)(cid:13)p j≥1 −2≤k≤2 22jαρj ∗ f2) 1 1 2(cid:13)(cid:13)p . max(cid:8) sup ·(cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)p into two parts 2(cid:13)(cid:13)p ≈(cid:13)(cid:13)(Xj≥1 2(cid:13)(cid:13)p . sup −2≤k≤2 j≥1 1 1 22jα φj ∗ ρj ∗ f2) 2(cid:13)(cid:13)p + k φ0 ∗ ρ0 ∗ fkp 2(cid:13)(cid:13)(Xj≥1 2(cid:13)(cid:13)p φ0 ∗ ρ0 ∗ f = F−1(cid:0)φ0(ϕ(0) + ϕ(1))(cid:1) ∗ ρ0 ∗ f. kφj(2j+k·)ϕkHσ ρj ∗ f2) . 1 and treat them separately. Applying Theorem 2.1 to the sequences (φj )j∈Z, (ρj)j∈Z with φj = 0 and ρj = 0 for j ≤ 0, we get the estimate of the first term on the right hand side. The result is The second term k φ0 ∗ ρ0 ∗ fkp is also easy to handle. By the support assumption on φ0ρ0, we have Hence, k φ0 ∗ ρ0 ∗ fkp ≤ kF−1(cid:0)φ0(ϕ(0) + ϕ(1))(cid:1)k1k ρ0 ∗ fkp . kφ0(ϕ(0) + ϕ(1))kHσ The assertion is proved. 2 k ρ0 ∗ fkp. (cid:3) 2.2. Hilbert-valued multipliers. In fact, both theorems above deal with Fourier multipliers acting on Hilbert-valued noncommutative Lp spaces (the Hilbert space being ℓ2). In this subsection titled "Hilbert-valued multipliers", our target is to extend Theorem 2.2 to the general case where ℓ2 is replaced with more complicated Hilbert spaces. Assume that we have a sequence of Hilbert spaces Hj for every j ∈ N0, and denote H = ⊕∞j=0Hj. Then an element f ∈ Lp(N ; H c) has the form f = (fj)j≥0 with fj ∈ Lp(N ; H c j ) for every j. In this case, it still makes sense to consider the action of the Calder´on-Zygmund operator k = ( φj )j≥0. 10 R. Xia and X. Xiong Since it will be frequently used in the following, we introduce an elementary inequality (see [40, Lemma 4.2]): (2.6) kf gkHσ 2 (Rd;ℓ2) ≤ kfkHσ 2 (Rd;ℓ2)ZRd (1 + s2)σF−1(g)(s)ds, where σ > d 2 , and the functions f : Rd → ℓ2 and g : Rd → C satisfy f ∈ H σ 2 (Rd; ℓ2) and ZRd (1 + s2)σF−1(g)(s)ds < ∞. 2 (Rd; ℓ2) is the ℓ2-valued Potential Sobolev space of order σ. Note also that ℓ2 could be an Here H σ ℓ2-space on an arbitrary index set, depending on the problems in consideration. The following lemma is an analogue of Lemma 4.3 in [40]. The main difference is that in order to get a Calder´on-Zygmund operator which is bounded on local Hardy or bmo spaces, we need to consider the Littlewood-Paley decomposition covering the origin. Lemma 2.3. Let φ = (φj )j≥0 be a sequence of continuous functions on Rd, viewed as a function from Rd to ℓ2. For σ > d 2 , we assume that (2.7) kφk2,σ def= max(cid:8) sup k≥1 kφ(2k·)ϕkHσ 2 (Rd;ℓ2), kφϕ(0)kHσ 2 (Rd;ℓ2)(cid:9) < ∞. precisely, Let k = (kj)j≥0 with kj = F−1(φj). Then k is a Calder´on-Zygmund kernel with values in ℓ2, more (1) kbkkL∞(Rd;ℓ2) . kφk2,σ; (2) Rs≥ 1 (3) supt∈RdRs>2t kk(s − t) − k(s)kℓ2 ds . kφk2,σ. The relevant constants depend only on ϕ, σ and d. 2 kk(s)kℓ2 ds . kφk2,σ; Proof. For any ξ ∈ Rd and k ≥ 1, by the Cauchy-Schwarz inequality, we have kφ(2kξ)ϕ(ξ)kℓ2 =(cid:13)(cid:13)Z F−1(φ(2k·)ϕ)(s)e−2πis·ξds(cid:13)(cid:13)ℓ2 2 (Rd;ℓ2)(Z (1 + s2)−σds) ≤ kφ(2k·)ϕkHσ 1 2 . kφk2,σ. In other words, we have kφϕ(2−k·)kL∞(Rd;ℓ2) . kφk2,σ. Likewise, kφϕ(0)kL∞(Rd;ℓ2) . kφk2,σ also holds. Thus, by (0.2) and (0.3), we easily deduce that kbkkL∞(Rd;ℓ2) . kφk2,σ. To show the third property of k, we decompose φ into φ =Xk≥0 φϕ(k). The convergence of the above series can be proved by a limit procedure of its partial sums, which is quite formal. By (0.2) and (0.3), we write φϕ(k) = φ(ϕ(k−1) + ϕ(k) + ϕ(k+1))ϕ(k) def= φ(k)ϕ(k), k ≥ 0. Here we make the convention that ϕ(k) = 0 if k < 0. Then for s ∈ Rd, F−1(φϕ(k))(s) = F−1(φ(k)) ∗ F−1(ϕ(k))(s) = 2kdF−1(φ(k)(2k·)) ∗ F−1(ϕ)(2ks), k ≥ 0. By (2.6), we have (ZRd (1 + 2ks2)σkF−1(φϕ(k))(s)k2 ℓ2 ds) 1 2 . 2 kd 2 kφ(k)(2k·)kHσ 2 (Rd;ℓ2). Inhomogeneous Triebel-Lizorkin spaces 11 Notice that if k ≥ 1, we have ϕ(k)(2k·) = ϕ. Thus, if k ≥ 2, kφ(k)(2k·)kHσ 2 (Rd;ℓ2) ≤ . = 1Xj=−1 1Xj=−1 1Xj=−1 kφ(2k·)ϕ(k−j)(2k·)kHσ 2 (Rd;ℓ2) kφ(2k−j·)ϕ(k−j)(2k−j·)kHσ 2 (Rd;ℓ2) kφ(2k−j·)ϕkHσ 2 (Rd;ℓ2) ≤ 3kφk2,σ. For k = 0, 1, we treat φ(k)(2k·) in the same way: kφ(1)(2·)kHσ kφ(0)kHσ 2 (Rd;ℓ2) . kφϕ(0)kHσ 2 (Rd;ℓ2) . kφϕ(0)kHσ 2 (Rd;ℓ2) + kφ(2·)ϕkHσ 2 (Rd;ℓ2) + kφ(2·)ϕkHσ 2 (Rd;ℓ2) + kφ(4·)ϕkHσ 2 (Rd;ℓ2) ≤ 3kφk2,σ. 2 (Rd;ℓ2) ≤ 3kφk2,σ; In summary, we obtain (ZRd (1 + 2ks2)σkF−1(φϕ(k))(s)k2 ℓ2 ds) 1 2 . 2 kd 2 kφk2,σ. Thus, by the Cauchy-Schwarz inequality, for any t ∈ Rd \ {0} and k ≥ 0, we have (1 + 2ks2)−σds) kF−1(φϕ(k))(s)kℓ2 ds . 2 kd 1 2 (2.8) Zs>t 2 kφk2,σ(Zs>t 2 −σkφk2,σ. d . (2kt) Consequently, Zs>2t kF−1(φϕ(k))(s) − F−1(φϕ(k))(s − t)kℓ2 ds . (2kt) d 2 −σkφk2,σ. We notice that d another estimate 2 − σ < 0, so the estimate above is good only when 2kt ≥ 1. Otherwise, we need where et(ξ) = e2πiξ·t. Thus, F−1(φϕ(k))(s) − F−1(φϕ(k))(s − t) = F−1(φ(k)ϕ(k)(1 − et))(s) = 2kdF−1(φ(k)(2k·)) ∗ [F−1(ϕ) − F−1(ϕ)(· − 2kt)](2ks), (ZRd (1 + 2ks2)σkF−1(φϕ(k))(s) − F−1(φϕ(k))(s − t)k2 2 kφk2,σ2ktZ (1 + s2)σF−1(ϕ)(s − θ2kt)ds 2 kφk2,σ2kt(Z J σ[ϕ(s)e2πis·θ2k t]2ds) 2 kφk2,σ2kt, . 2 . 2 . 2 ℓ2 kd kd kd 1 2 1 2 ds) where θ ∈ [0, 1]. Then as before, for 2kt < 1, we have kF−1(φϕ(k))(s) − F−1(φϕ(k))(s − t)kℓ2 ds . 2ktkφk2,σ. Combining the previous estimates, we obtain sup Zs>2t t∈RdZs>2t t∈RdXk≥0Zs>2t t∈RdXk≥0 . kφk2,σ sup ≤ sup kk(s − t) − k(s)kℓ2 ds kF−1(φϕ(k))(s) − F−1(φϕ(k))(s − t)kℓ2 ds min(2kt, (2kt) d 2 −σ) . kφk2,σ. 12 R. Xia and X. Xiong Finally, the second estimate of k can be deduced from (2.8) by letting t = 1 2 : Zs≥ 1 2 kk(s)kℓ2 ds ≤Xk≥0Zs≥ 1 2 (2k−1) ≤Xk≥0 kF−1(φϕ(k))(s)kℓ2 ds d 2 −σkφk2,σ . kφk2,σ. The proof is complete. (cid:3) We keep the notation H = ⊕∞j=0Hj. By the above lemma, we can apply the (local) Calder´on- Zygmund theory introduced in section 1, to deduce the following lemma: Lemma 2.4. Let 1 < p < ∞ and φ = (φj)j≥0 be a sequence of continuous functions on Rd satisfying (2.7). For any f = (fj)j≥0 ∈ Lp(N ; H c), we have k( φj ∗ fj)j≥0kLp(N ;Hc) . kφk2,σk(fj)j≥0kLp(N ;Hc), where the relevant constant depends only on ϕ, σ, p and d. Proof. Consider k as a diagonal matrix with diagonal entries (kj)j≥0 determined by bkj = φj and f = (fj)j≥0 as a column matrix. The associated Calder´on-Zygmund operator is defined on Lp(B(H)⊗N ) by k(f )(s) =ZRd k(s − t)f (t)dt. Now it suffices to show that k is a bounded operator on Lp(N ; H c). We claim that k is bounded from L∞(N ; H c) into bmo(Rd, B(H)⊗M). Put K(s) = k(s) ⊗ 1M ∈ B(H)⊗M, for any s ∈ Rd. Then we have kk(s)kℓ2 ≥ kk(s)kℓ∞ = kK(s)kB(H)⊗M and kfkL∞(N ;Hc) = kfkB(H)⊗N . Thus, the claim is equivalent to saying that K is bounded from L∞(N ; H c) into bmo(Rd, B(H)⊗M), if we regard L∞(N ; H c) as a subspace of B(H)⊗N . First, we show that K is bounded from L∞(N ; H c) into bmoc(Rd, B(H)⊗M). Let Q be a cube in Rd centered at c. We decompose f as f = g + h with g = f 1 eQ, where eQ = 2Q is the cube which has the same center as Q and twice the side length of Q. Set a =ZRd\ eQ K(c − t)f (t)dt. Then where whence K(f )(s) − a = K(g)(s) +Z [K(s − t) − K(c − t)]h(t)dt. Thus, for Q such that Q < 1, we have 1 A = 1 1 B = QZQ K(f ) − a2ds ≤ 2(A + B), QZQ K(g)2ds, QZQ Z [K(s − t) − K(c − t)]h(t)dt2ds. QA ≤Z bK(ξ)bg(ξ)2dξ =Z bg(ξ)∗bK(ξ)∗bK(ξ)bg(ξ)dξ ≤Z kbK(ξ)k2 .Z kbk(ξ)k2 ≤ eQkφk2 2,σZ eQ f (s)2ds 2,σkfk2 ℓ2bg(ξ)2dξ . kφk2 L∞(N ;Hc), 2,σkfk2 B(H)⊗N = eQkφk2 kAkB(H)⊗M . kφk2 2,σkfk2 L∞(N ;Hc). The term A is easy to estimate. By Lemma 2.3 and the Plancherel formula (0.7), B(H)⊗Mbg(ξ)2dξ Inhomogeneous Triebel-Lizorkin spaces 13 To estimate B, writing h = (hj)j≥0, by Lemma 2.3, we get (cid:12)(cid:12)Z [K(s − t) − K(c − t)]h(t)dt(cid:12)(cid:12)2 .ZRd\ eQ kK(s − t) − K(c − t)kB(H)⊗MdtZRd\ eQ kK(s − t) − K(c − t)kB(H)⊗Mh(t)2dt .ZRd\ eQ kk(s − t) − k(c − t)kℓ2 dtZRd\ eQ kk(s − t) − k(c − t)kℓ2h(t)2dt . kφk2 L∞(N ;Hc). 2,σkfk2 2,σkfk2 . kφk2 B(H)⊗N Hence, kBkB(H)⊗M ≤ B(H)⊗M Combining the previous inequalities, we deduce that, for any Q < 1 ds . kφk2 2,σkfk2 L∞(N ;Hc). 1 1 QZQ(cid:13)(cid:13)Z [K(s − t) − K(c − t)]h(t)dt(cid:13)(cid:13)2 QZQ K(f ) − a2ds) (cid:13)(cid:13)(cid:13)( QZQ K(f )2ds ≤ 2 2(cid:13)(cid:13)(cid:13)B(H)⊗M QZQ K(g)2ds + 2 1 1 1 . kφk2,σkfkL∞(N ;Hc). 1 QZQ K(h)2ds. Now we consider the case when Q = 1. We have The first term on the right hand side of the above inequality is equal to the term A, so it remains 2 . Then by to estimate the second term. When t ∈ Rd\eQ, s ∈ Q and Q = 1, we have s − t ≥ 1 (2) in Lemma 2.3 and the Cauchy-Schwarz inequality (0.6), we easily deduce that K(h)(s)2 =(cid:12)(cid:12)Z K(s − t)h(t)dt(cid:12)(cid:12)2 ≤ZRd\ eQ kK(s − t)kB(H)⊗MdtZRd\ eQ kK(s − t)kB(H)⊗Mh(t)2dt L∞(N ;Hc)(ZRd\ eQ kk(s − t)kℓ2 dt)2 . kfk2 . kφk2 2,σkfk2 Thus, we have, for any Q = 1, L∞(N ;Hc). Next we show that K is bounded from L∞(N ; H c) into bmor(Rd, B(H)⊗M). We still use the QZQ K(f )2ds(cid:1) 1 (cid:13)(cid:13)(cid:13)(cid:0) 1 2(cid:13)(cid:13)(cid:13)B(H)⊗M . kφk2,σkfkL∞(N ;Hc). 1 1 A′ = where Therefore, K is bounded from L∞(N ; H c) into bmoc(Rd, B(H)⊗M). same decomposition f = g + h, then we obtain QZQ [K(f ) − a]∗2ds ≤ 2(A′ + B′), QZQ K(g)∗2ds, QZQ(cid:12)(cid:12)Z [(K(s − t) − K(c − t))h(t)]∗dt(cid:12)(cid:12)2 QZQ(cid:13)(cid:13)Z [(K(s − t) − K(c − t))h(t)]∗dt(cid:13)(cid:13)2 QZQ(cid:13)(cid:13)Z [K(s − t) − K(c − t)]h(t)dt(cid:13)(cid:13)2 The estimate of B′ can be reduced to that of B. Indeed, kB′kB(H)⊗M ≤ B′ = = 1 1 1 . kφk2 2,σkfk2 L∞(N ;Hc). ds. ds B(H)⊗M ds B(H)⊗M 14 R. Xia and X. Xiong However, for A′, we need a different argument. A′ can be viewed as a bounded operator on H ⊗ L2(M). So kA′kB(ℓ2)⊗M = sup b { 1 QZQ kk(g)(s) bk2 H⊗L2(M)ds}, where the supremum runs over all b in the unit ball of H ⊗ L2(M). By the Plancherel formula (0.7), we have ZQ kk(g)(s) bk2 Let diag(fj)j be the diagonal matrix in B(H)⊗N with entries in B(Hj)⊗N . By the Cauchy- Schwarz inequality, the Plancherel formula (0.7) and Lemma 2.3, we continue the estimate above as H⊗L2(M)ds =ZQhk(g)(s) b, k(g)(s) biH⊗L2(M)ds ≤Z hbk(ξ)bg(ξ) b,bk(ξ)bg(ξ) biH⊗L2(M)dξ. ℓ2Z hbg(ξ) b,bg(ξ) biH⊗L2(M)dξ ξ kbk(ξ)k2 2,σZ eQ k diag(fj)j(s) bk2 . kφk2 2,σk diag(fj)jk2 . Qkφk2 2,σkfk2 ≤ Qkφk2 L∞(N ;Hc), H⊗L2(M)ds B(H)⊗Nkbk2 H⊗L2(M) Z hbk(ξ)bg(ξ) b,bk(ξ)bg(ξ) biH⊗L2(M)dξ ≤ sup whence, 1 L∞(N ;Hc). Following the estimate of 1 kA′kB(ℓ2)⊗M . kφk2 QZQ K(f )∗2ds ≤ 2A′ + 2 ≤ 2A′ + 2 = 2A′ + 2 2,σkfk2 QRQ K(f )(s)2ds, we get, when Q = 1, QZQ K(h)∗2ds QZQ kK(h)∗k2 QZQ kK(h)k2 2,σkfk2 Therefore, K is bounded from L∞(N ; H c) into bmor(Rd, B(H)⊗M). In summary, we have proved that k is bounded from L∞(N ; H c) into bmo(Rd, B(H)⊗M). It is also clear that k is bounded from L2(N ; H c) into L2(B(H)⊗N ), then by the interpolation in Theorem 1.2, k is bounded from Lp(N ; H c) into Lp(B(H)⊗N ) for any 2 ≤ p < ∞. The case 1 < p < 2 is obtained by duality. (cid:3) L∞(N ;Hc). . kφk2 B(H)⊗M B(H)⊗M 1 1 1 ds ds Note that when all Hj degenerate to one dimensional Hilbert space, then H = ℓ2, the above lemma gives a sufficient condition for (φj )j≥0 being a bounded Fourier multiplier on Lp(N ; ℓc 2). So we can also use Lemmas 2.3 and 2.4 to prove Theorem 2.2 by an argument similar to the proof of [40, Theorem 4.1]; details are left to the reader. But here our target is to extend Theorem 2.2 to a more general setting. Theorem 2.5. Let p, α, σ, (φj)j≥0 and (ρj )j≥0 be the same as in Theorem 2.2. Then, for any f ∈ S′(Rd; L1(M) + M), 2j(2α+d)ZB(0,2−j ) φj ∗ ρj ∗ f (· + t)2dt) kφj(2j+k·)ϕkHσ 2 2(cid:9) ,kφ0(ϕ(0) + ϕ(1))kHσ −2≤k≤2 2j(2α+d)ZB(0,2−j ) ρj ∗ f (· + t)2dt) (cid:13)(cid:13)(cid:13)(Xj≥0 . max(cid:8) sup ·(cid:13)(cid:13)(cid:13)(Xj≥0 j≥1 1 2(cid:13)(cid:13)(cid:13)p 2(cid:13)(cid:13)(cid:13)p 1 , Inhomogeneous Triebel-Lizorkin spaces 15 where the constant depends only on p, σ, d and ϕ. Proof. Set Hj = L2(cid:0)B(0, 2−j), 2jddt(cid:1) and H = ⊕∞j=0Hj. So we have 2j(2α+d)ZB(0,2−j ) φj ∗ ρj ∗ f (· + t)2dt) 1 (cid:13)(cid:13)(Xj≥0 Let 2(cid:13)(cid:13)p = k(2jα φj ∗ ρj ∗ f (· + ·))jkLp(N ;Hc). ζj = φj(ϕ(j−1) + ϕ(j) + ϕ(j+1)), j ≥ 2, ζ1 = φ1(ϕ + ϕ(1) + ϕ(2)), ζ0 = φ0(ϕ(0) + ϕ) and ζj = 0 if j < 0. By the support assumption on φjρj, we have that φjρj = ζjρj. So for any f ∈ S′(Rd; L1(M)+M), φj ∗ ρj ∗ f = ζj ∗ ρj ∗ f, j ∈ N0. Now we show that ζ = (ζj )j≥0 satisfies (2.7) with ζ instead of φ. Indeed, by the support assumption of ϕ, the sequence ζ(2k·)ϕ = (cid:0)ζj(2k·)ϕ(cid:1)j≥0 has at most five nonzero terms of indices j with k − 2 ≤ j ≤ k + 2. Thus for any k ∈ N0, kζ(2k·)ϕkHσ 2 (Rd;ℓ2) ≤ k+2Xj=k−2 kζj(2k·)ϕkHσ 2 . Moreover, by (2.6), we have kζj(2k·)ϕkHσ Therefore, the condition (2.5) yields 2 . kφj(2k·)ϕkHσ 2 , k − 2 ≤ j ≤ k + 2. 2 (Rd;ℓ2) . sup j≥1 −2≤k≤2 kφj(2j+k·)ϕkHσ 2 + kφ0(ϕ(0) + ϕ(1))kHσ 2 < ∞, where the relevant constant depends only on σ, ϕ and d. In a similar way, we have kζjϕ(0)kHσ 2 kφj(2j+k·)ϕkHσ 2 + kφ0(ϕ(0) + ϕ(1))kHσ 2 < ∞. . sup j≥1 −2≤k≤2 Now applying Lemma 2.4 to fj = 2jα ρj∗f (·+·), and ζj instead of φj , we conclude the theorem. (cid:3) The above theorem will be useful when we consider the conic square function characterizations of local Hardy spaces and inhomogeneous Triebel-Lizorkin spaces in section 4. p. Note that both Theorem 2.2 and Theorem 2.5 do not deal with the case p with 1 ≤ p ≤ 2 in the 2.3. Multipliers on hc p = 1. So we include the corresponding Fourier multiplier results for hc following. When the Hilbert space H degenerates to ℓ2, we have Lemma 2.6. Let 1 ≤ p ≤ 2 and φ = (φj)j≥0 be a sequence of continuous functions on Rd satisfying (2.7). For f ∈ hc p(Rd,M), sup k≥1 kζ(2k·)ϕkHσ 2 (Rd;ℓ2) ≤ X0≤j≤2 kζϕ(0)kHσ The relevant constant depends only on ϕ, σ and d. Proof. Now we view k = (kj)j≥0 = ( φj )j≥0 as a column matrix and the associated Calder´on- Zygmund operator k is defined on Lp(N ): (cid:13)(cid:13)(Xj≥0 φj ∗ f2) 1 2(cid:13)(cid:13)p . kφk2,σkfkhc . p k(f )(s) =ZRd k(s − t)f (t)dt, ∀s ∈ R. Thus k maps function with values in Lp(M) to sequence of functions. Then we have to show that k is bounded from hc 2) for 1 ≤ p ≤ 2. The case p = 2 is trivial, so by interpolation, it suffices to consider the case p = 1. To prove that k is bounded from hc 1(Rd,M) to L1(N ; ℓc 2), passing to the dual spaces, it is equal to proving that the adjoint of k is bounded from p(Rd,M) to Lp(N ; ℓc 16 R. Xia and X. Xiong L∞(N ; ℓc sequence f = (fj)j≥0 (viewed as a column matrix), the adjoint of k is defined by 2) to bmoc(Rd,M). We keep all the notation in the proof of Lemma 2.4. For any finite k∗(f )(s) =ZRdXj ekj(s − t)fj(t)dt, whereek(s) = k(−s)∗ (so it is a row matrix). Put eK(s) =ek(s)⊗1M. In this case, keK(f )kbmoc(Rd,M) = keK(f )kbmoc(Rd,B(ℓ2)⊗M). Then we apply the estimates used in Lemma 2.4 by replacing K with eK. It follows that k∗ is bounded from L∞(N ; ℓc 2) into bmoc(Rd,M), so the desired assertion is (cid:3) The next theorem is a complement of Theorem 2.2 for the case p = 1, which relies heavily on proved. 1(Rd,M) given in Theorem 1.5. the characterization of hc Theorem 2.7. We keep the assumption in Theorem 2.2. Assume additionally that for any j ≥ 1, ρj = ρ(2−j·) for some Schwartz function ρ with supp ρ ⊂ {ξ : 2−1 ≤ ξ ≤ 2} and ρ(ξ) > 0 for any 2−1 < ξ < 2, and that supp ρ0 ⊂ {ξ : ξ ≤ 2} and ρ0(ξ) > 0 for any ξ < 2. Then for f ∈ S′(Rd; L1(M) + M), we have 22jα φj ∗ ρj ∗ f2) (cid:13)(cid:13)(Xj≥0 kφj(2j+k·)ϕkHσ 2 2(cid:9) ,kφ0(ϕ(0) + ϕ(1))kHσ 1 2(cid:13)(cid:13)1 . max(cid:8) sup ·(cid:13)(cid:13)(Xj≥0 j≥1 −2≤k≤2 22jαρj ∗ f2) . 1 2(cid:13)(cid:13)1 ∀ξ ∈ Rd. Proof. By the assumptions of ρ and ρ0, we can select a Schwartz functioneρ with the same properties as ρ and a Schwartz function eρ0 satisfying the same conditions as ρ0, such that Let Ψj = (I−αρ)(2−j·), eΨj = (Iαρ)(2−j·) for j ≥ 1 and Ψ0 = J−αρ0, eΨ0 = Jαρ0. We have Applying Theorem 1.5 (the equivalence kgc,D functions in the above identity, we get kgkhc ∞Xj=1 ρ(2−jξ)eρ(2−jξ) + ρ0(ξ)eρ0(ξ) = 1, ∞Xj=1 Ψj(ξ)eΨj(ξ) + Ψ0(ξ)eΨ0(ξ) = 1, 2(cid:13)(cid:13)1 Φ (f )kp + kφ ∗ fkp ≈ kfkhc p) to g = J αf with the text Ψj ∗ g2) ∀ ξ ∈ Rd. . 1 Now let us show the following equivalence: 1 Ψj ∗ g2) It is easy to see that Ψ0 ∗ g = ρ0 ∗ f and 2jα ρj ∗ f = Ψj ∗ I αf , so it suffices to prove (2.9) . 1 First, let us consider the case α ≥ 0. By [31, Lemma 3.2.2], there exists a finite measure µα on Rd such that (cid:13)(cid:13)(Xj≥0 (cid:13)(cid:13)(Xj≥1 (cid:13)(cid:13)(Xj≥1 1 Ψj ∗ I αf2) 22jαρj ∗ f2) Ψj ∗ J αf2) 1 ≈(cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)1 ≈(cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)1 ≈(cid:13)(cid:13)(Xj≥1 ξα =bµα(ξ)(1 + ξ2) Ψj ∗ I αf = µα ∗ Ψj ∗ J αf, .(cid:13)(cid:13)(Xj≥1 2(cid:13)(cid:13)1 Ψj ∗ I αf2) 2 =bνα(ξ) + ξ−αbλα(ξ). (1 + ξ2)− α 2 . ∀ j ≥ 1. Ψj ∗ J αf2) α 1 1 . 2(cid:13)(cid:13)1 2(cid:13)(cid:13)1 . 1 2(cid:13)(cid:13)1 Thus, we have This implies that Then, we move to the case α < 0. Also by [31, Lemma 3.2.2], there exist two finite measures να and λα on Rd such that Inhomogeneous Triebel-Lizorkin spaces 17 Let ( ϕk)k∈Z be the homogeneous resolution of the unit defined in (0.4). It follows that 2 ξ−α Xk≥0 (1 + ξ2)− α ξ−α Xk≥0 ϕk(ξ) = bνα(ξ) Thus, by the support assumption of bρ, we have Ψj ∗ I αf = ωα ∗ Ψj ∗ J αf, ϕk(ξ) +bλα(ξ)Xk≥0 ϕk(ξ). with ωα = να ∗Xk≥0 F−1(Iα ϕk) + λα ∗ F−1(Xk≥0 ϕk). Both F−1(Pk≥0 ϕk) and Pk≥0 F−1(Iα ϕk) are finite measures. Since Pk≥0 ϕk = 1 −Pk<0 ϕk, andPk<0 ϕk is a Schwartz function, we know that F−1(Pk≥0 ϕk) = δ0−F−1(Pk<0 ϕk) is a finite measure, where δ0 denotes the Dirac measure at the origin. Moreover, it is known in [40, Lemma 3.4] that kF−1(Iα ϕk)k1 . 2kα. Then we have kF−1(Xk≥0 2kα < ∞. Therefore, ωα is a finite measure on Rd. Thus, 1 Iα ϕk)k1 .Xk≥0 2(cid:13)(cid:13)1 .(cid:13)(cid:13)(Xj≥1 2(cid:13)(cid:13)1 .(cid:13)(cid:13)(Xj≥1 1 ≈(cid:13)(cid:13)(cid:0)Xj≥0 1 . . 1 1 Ψj ∗ I αf2) Ψj ∗ J αf2) 2(cid:13)(cid:13)1 2(cid:13)(cid:13)1 2(cid:13)(cid:13)1 22jαρj ∗ f2(cid:1) 1 . Ψj ∗ I αf2) Ψj ∗ J αf2) (cid:13)(cid:13)(Xj≥1 (cid:13)(cid:13)(Xj≥1 kgkhc 1 = kJ αfkhc Similarly, for α ∈ R, we can prove that In summary, we have proved (2.9), which yields that Now define a new sequence ζ = (ζj)j≥0 by setting ζj = 2jαI−αφjρj for j ≥ 1 and ζ0 = J−αφ0ρ0. Then ζj ∗ g = 2jα φj ∗ ρj ∗ I−αg and ζ0 ∗ g = φ0 ∗ ρ0 ∗ f. Repeating the argument for (2.9) with ζ = (ζj )j≥0 instead of Ψ = (Ψj)j≥0, we get ζj ∗ g2) 22jα φj ∗ ρj ∗ f2) ζj ∗ I αf2) 1 1 2(cid:13)(cid:13)1 ≈(cid:13)(cid:13)(Xj≥0 . 1 2(cid:13)(cid:13)1 Then, we apply Lemma 2.6 to g with this new ζ instead of φ to get (cid:13)(cid:13)(Xj≥0 (cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)1 =(cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)1 . kζk2,σkgkhc 1 ζj ∗ g2) 1 ≈ kζk2,σ(cid:13)(cid:13)(Xj≥0 22jαρj ∗ f2) . 1 2(cid:13)(cid:13)1 It suffices to estimate the term kζk2,σ. By the definition of ζ = (ζj )j≥, we have −2≤k≤2 kζj(2j+k·)ϕkHσ sup j≥1 kζ0(ϕ(0) + ϕ(1))kHσ 2 2 kφj(2j+k·)ϕkHσ 2 , . sup j≥1 −2≤k≤2 . kφ0(ϕ(0) + ϕ(1))kHσ 2 . So we can use the same argument at the end of the proof of Theorem 2.5, to get kζk2,σ . max(cid:8) sup j≥1 −2≤k≤2 kφj(2j+k·)ϕkHσ 2 2(cid:9). ,kφ0(ϕ(0) + ϕ(1))kHσ Combining the above inequalities, we get the desired assertion. (cid:3) In the setting where ℓ2 is replaced by H = ⊕∞j=0Hj with Hj = L2(cid:0)B(0, 2−j), 2jddt(cid:1), the coun- terpart of Lemma 2.6 is the following: 18 R. Xia and X. Xiong Lemma 2.8. Let φ = (φj )j≥0 be a sequence of continuous functions on Rd satisfying (2.7). Then for 1 ≤ p ≤ 2 and f ∈ hc p(Rd,M), 2djZB(0,2−j ) φj ∗ f (· + t)2dt) (cid:13)(cid:13)(Xj≥0 1 2(cid:13)(cid:13)p . kφk2,σkfkhc p . The relevant constant depends only on ϕ, σ and d. Proof. The proof of this lemma is similar to Lemma 2.6; let us point out the necessary change. Consider the H-valued Calder´on-Zygmund operator k defined on Lp(N ) given by k(f )j(· + t) = φj ∗ f (· + t). p(Rd,M) to Lp(N ; H c) for 1 ≤ The lemma is then reduced to showing that k is bounded from hc p < 2. Since each Hj is a normalized Hilbert space, such that the constant function 1 has Hilbert norm one, the kernel estimates of our k here are the same as the ones in Lemma 2.4. So we can repeat the proof in Lemma 2.4 and Lemma 2.6. The desired assertion follows. (cid:3) Combining the above lemma with Theorem 1.5 (ksc,D deduce the analogue of Theorem 2.7 in the setting H = ⊕∞j=0Hj with Hj = L2(cid:0)B(0, 2−j), 2jddt(cid:1). Its proof is similar to that of Theorem 2.7, and is left to the reader. Theorem 2.9. Keep the assumption in Theorem 2.5 and assume additionally that for any j ≥ 1, ρj = ρ(2−j·) for some Schwartz function with supp ρ ⊂ {ξ : 2−1 ≤ ξ ≤ 2} and ρ(ξ) > 0 for any 2−1 < ξ < 2, and that supp ρ0 ⊂ {ξ : ξ ≤ 2} and ρ0(ξ) > 0 for any ξ < 2. Then for any f ∈ S′(Rd; L1(M) + M), Φ (f )kLp(N ) + kφ ∗ fkp ≈ kfkhc p), we can 2j(2α+d)ZB(0,2−j ) φj ∗ ρj ∗ f (· + t)2dt) kφj(2j+k·)ϕkHσ 2 2(cid:9) ,kφ0(ϕ(0) + ϕ(1))kHσ −2≤k≤2 2j(2α+d)ZB(0,2−j ) ρj ∗ f (· + t)2dt) (cid:13)(cid:13)(cid:13)(Xj≥0 . max(cid:8) sup ·(cid:13)(cid:13)(cid:13)(Xj≥0 j≥1 1 2(cid:13)(cid:13)(cid:13)1 2(cid:13)(cid:13)(cid:13)1 1 . This theorem fills the gap of p = 1 left by Theorem 2.5. Both of them will be useful when we consider the conic square functions of inhomogeneous Triebel-Lizorkin spaces in section 4. 3. Operator-valued Triebel-Lizorkin spaces In this section, we give the definition of operator-valued Triebel-Lizorkin spaces, and then prove some basic properties of them. Among the others, we connect operator-valued Triebel-Lizorkin spaces with local Hardy spaces introduced in [36] via Bessel potentials. By this connection, we are able to deduce the duality and the complex interpolation of Triebel-Lizorkin spaces. We also show that for α > 0 the F α,c (Rd,M)-norm is the sum of two homogeneous norms. 1 3.1. Definitions and basic properties. Recall that ϕ is a Schwartz function satisfying (0.1). For each j ∈ N, ϕj is the function whose Fourier transform is equal to ϕ(2−j·), and ϕ0 is the function whose Fourier transform is equal to 1 −Pj≥1 ϕ(2−j·). Moreover, the Fourier transform of ϕj is denoted by ϕ(j) for j ∈ N0. Definition 3.1. Let 1 ≤ p < ∞ and α ∈ R. (1) The column Triebel-Lizorkin space F α,c p (Rd,M) is defined by F α,c p where (Rd,M) = {f ∈ S′(Rd; L1(M) + M) : kfkF α,c p < ∞}, kfkF α,c p =(cid:13)(cid:13)(Xj≥0 22jαϕj ∗ f2) . 1 2(cid:13)(cid:13)p (2) The row space F α,r p norm kfkF α,r p = kf∗kF α,c p . (Rd,M) consists of all f such that f∗ ∈ F α,c p (Rd,M), equipped with the Inhomogeneous Triebel-Lizorkin spaces 19 (3) The mixture space F α p (Rd,M) is defined to be equipped with F α p F α,c p (Rd,M) =(F α,c p =(inf{kgkF α,c p kfkF α (Rd,M) + F α,r (Rd,M) ∩ F α,r p p (Rd,M) (Rd,M) if if 1 ≤ p ≤ 2 2 < p < ∞, p + khkF α,r ,kfkF α,r p } p max{kfkF α,c p : f = g + h} if 1 ≤ p ≤ 2 if 2 < p < ∞. In the sequel, we focus on the study of the column Triebel-Lizorkin spaces. All results obtained in -norm the rest of this paper also admit the row versions. The following proposition shows that F α,c is independent of the choice of the function ϕ satisfying (0.1). p Proposition 3.2. Let ψ be another Schwartz function satisfying the same condition (0.1) as ϕ. For each j ∈ N, let ψj be the function whose Fourier transform is equal to ψ(2−j·), and let ψ0 be the function whose Fourier transform is equal to 1 −Pj≥1 ψ(2−j·). Then Proof. For any f ∈ S′(Rd; L1(M) + M), by the support assumption of ψ and ϕ, we have, for any j ≥ 0, 22jαψj ∗ f2) 2(cid:13)(cid:13)p kfkF α,c . 1 p ≈(cid:13)(cid:13)(Xj≥0 1Xk=−1 ψj ∗ ϕk+j ∗ f, with the convention ϕ−1 = 0. Thus by Theorems 2.2 and 2.7, ψj ∗ f = 22jαψj ∗ f2) 1 2(cid:13)(cid:13)p ≤ (cid:13)(cid:13)(Xj≥0 1Xk=−1(cid:13)(cid:13)(Xj≥0 . max(cid:8) sup .(cid:13)(cid:13)(Xj≥0 2 −2≤k≤2 kψ(2k·)ϕkHσ 2(cid:13)(cid:13)p 22jαϕj ∗ f2) . 1 22jαψj ∗ ϕk+j ∗ f2) 1 2(cid:13)(cid:13)p ,kψ0(ϕ(0) + ϕ(1))kHσ p p (Rd,M) is a Banach space. (Rd,M) ⊂ F β,c Changing the role of ϕ and ψ, we get the reverse inequality. Proposition 3.3. Let 1 ≤ p < ∞ and α ∈ R. Then (1) F α,c (2) F α,c (3) F 0,c Proof. (1) Let {fi} be a Cauchy sequence in F α,c (ϕ0 ∗ fi, . . . , 2jαϕj ∗ fi, . . .) is also a Cauchy sequence in Lp(N ; ℓc function f = (f 0, . . . , f j, . . .) in Lp(N ; ℓc 2(N0)). Formally we take (3.1) p(Rd,M) with equivalent norms. (Rd,M) if α > β. p (Rd,M) = hc f j. p p f =Xj≥0 2(cid:9)(cid:13)(cid:13)(Xj≥0 22jαϕj ∗ f2) 1 2(cid:13)(cid:13)p (cid:3) (Rd,M). Then, the sequence {ai} with ai = 2(N0)). Thus, ai converges to a Since for each j ∈ N, suppcf j ⊂ {ξ : 2j−1 ≤ ξ ≤ 2j+1} and suppcf 0 ⊂ {ξ : ξ ≤ 2}, the series (3.1) converges in S′(Rd; Lp(M)). Let ϕj = 0 if j < 0. By the support assumption of ϕ, when i → ∞, we get ϕj ∗ fi = ϕk ∗ ϕj ∗ fi → ϕj ∗ f k = ϕj ∗ f, j+1Xk=j−1 j+1Xk=j−1 which implies that f j = 2jαϕj ∗ f , for any j ≥ 0. Thus, f ∈ F α,c in F α,c p (Rd,M) and {fi} converges to f p (Rd,M). (2) is obvious. 20 R. Xia and X. Xiong terization of hc (3) It is easy to see that any ϕ satisfying (0.1) also satisfies (1.6). Then by the discrete charac- (cid:3) p(Rd,M) given in Theorem 1.5, we get the desired assertion. Given a ∈ R+, we define Di,a(ξ) = (2πiξi)a for ξ ∈ Rd, and Da i to be the Fourier multiplier (Rd,M). We set Da = D1,a1 ··· Dd,ad and with symbol Di,a(ξ) on Triebel-Lizorkin spaces F α,c Da = Da1 i = ∂a i , so there does not exist any conflict of notation. The operator Da can be viewed as a fractional extension of partial derivatives. The following is the so-called lifting (or reduction) property of Triebel-Lizorkin spaces. +. Note that if a is a positive integer, Da for any a = (a1,··· , ad) ∈ Rd 1 ··· Dad d p Proposition 3.4. Let 1 ≤ p < ∞ and α ∈ R. (1) For any β ∈ R, J β is an isomorphism between F α,c (Rd,M) and hc J α is an isomorphism between F α,c p p (Rd,M) and F α−β,c p(Rd,M). p (2) Let β > 0. Then f ∈ F α,c p (Rd,M) if and only if ϕ0 ∗ f ∈ Lp(N ) and Dβ i f ∈ F α−β,c p (Rd,M) for all i = 1, . . . , d. Moreover, in this case, (Rd,M). In particular, kfkF α,c p ≈ kϕ0 ∗ fkp + dXi=1 kDβ i fkF α−β,c p . Proof. (1) Let f ∈ F α,c p (Rd,M). Applying Theorems 2.2 and 2.7 with ρ = ϕ, we obtain 2−jβkJβ(2j+k·)ϕkHσ 2 ,kJβ(ϕ(0) + ϕ(1))kHσ 2 } kJ βfkF α−β,c p (3.2) 22j(α−β)ϕj ∗ J βf2) 1 2(cid:13)(cid:13)p =(cid:13)(cid:13)(Xj≥0 ·(cid:13)(cid:13)(Xj≥0 . max{ sup j≥1 −2≤k≤2 22jαϕj ∗ f2) . 1 2(cid:13)(cid:13)p It is easy to check that all partial derivatives of 2−jβJβ(2j+k·)ϕ of order less than or equal to [σ] + 1 are bounded uniformly in j ≥ 1 and −2 ≤ k ≤ 2, and that Jβ(ϕ(0) + ϕ(1)) ∈ H σ 2 (Rd). Thus (Rd,M), and its inverse kJ βfkF α−β,c J−β is also continuous from F α−β,c . So J β is continuous from F α,c (Rd,M) to F α−β,c . kfkF α,c p p p p p (Rd,M) to F α,c p (Rd,M). (2) If we take σ ∈ ( d 2 , β + d 2 ), then we have kDi,βϕ0kHσ 2 < ∞ and kDi,βϕkHσ 2 < ∞. Replacing J β by Dβ i in (3.2), we obtain that, for any i = 1, . . . , d, kDβ i fkF α−β,c p . kfkF α,c p , which implies immediately that kϕ0 ∗ fkp + dXi=1 kDβ i fkF α−β,c p . kfkF α,c p . To show the reverse inequality, we choose a nonnegative infinitely differentiable function χ on R such that χ(s) = 0 if s < 1 . For i = 1, . . . , d, we define χi on Rd as 2√d follows: and χ(s) = 1 if s ≥ 1√d χi(ξ) = 1 χ(ξ1)ξ1β + . . . + χ(ξd)ξdβ χ(ξi)ξiβ (2πiξi)β , whenever the first denominator is positive, say, when ξ ≥ 1. Then for any j ≥ 1, χiϕj is a well-defined infinitely differentiable function on Rd\{ξ : ξi = 0} and ϕ(j) = dXi=1 χiDi,βϕ(j). Inhomogeneous Triebel-Lizorkin spaces 21 Then by Theorem 2.1, we have kfkF α,c p ≤ kϕ0 ∗ fkp + . kϕ0 ∗ fkp + 22jα χi ∗ ϕj ∗ Dβ 2(cid:13)(cid:13)p i f2(cid:1) 1 2(cid:13)(cid:13)(cid:0)Xj≥1 2jβkχi(2j+k·)ϕkHσ dXi=1(cid:13)(cid:13)(cid:0)Xj≥1 dXi=1 2jβkχi(2j+k·)ϕkHσ −2≤k≤2 sup j≥1 2 (Rd) = kφi(2k·)ϕkHσ 2 (Rd), However, where 22j(α−β)ϕj ∗ Dβ 2(cid:13)(cid:13)p i f2(cid:1) 1 . φi(ξ) = χ(2jξ1)ξ1β + . . . + χ(2jξd)ξdβ 1 χ(2jξi)ξiβ (2πiξi)β . Since all partial derivatives of φiϕ(2k·), of order less than a fixed integer, are bounded uniformly in 2 (Rd) are bounded from above by a constant independent j, k and i, and the norm of φiϕ(2k·) in H σ of j, k and i. Then we deduce kfkF α,c p . kϕ0 ∗ fkp + ≤ kϕ0 ∗ fkp + The assertion is proved. dXi=1(cid:13)(cid:13)(cid:0)Xj≥1 dXi=1 kDβ 22j(α−β)ϕj ∗ Dβ 2(cid:13)(cid:13)p i f2(cid:1) 1 i fkF α−β,c p . (cid:3) Definition 3.5. For α ∈ R, we define F α,c ∞ (Rd,M) as the space of all f ∈ S′(Rd;M) such that kϕ0 ∗ fkN + sup Q<1(cid:13)(cid:13)(cid:13) We endow the space F α,c ∞ (Rd,M) with the norm: = kϕ0 ∗ fkN + sup kfkF α,c ∞ 1 QZQ Xj≥− log2(l(Q)) Q<1(cid:13)(cid:13)(cid:13) 1 QZQ Xj≥− log2(l(Q)) 22jαϕj ∗ f (s)2ds(cid:13)(cid:13)(cid:13) 1 2 M < ∞. 22jαϕj ∗ f (s)2ds(cid:13)(cid:13)(cid:13) 1 2 . M p Proposition 3.6. Let 1 ≤ p < ∞, α ∈ R and q be the conjugate index of p. Then the dual space of F α,c (Rd,M) coincides isomorphically with F −α,c (Rd,M). Proof. First, we show that J α is an isomorphism between F α,c end, we use the discrete Carleson characterization of bmoc(Rd,M) in [36, Corollary 5.13]: (3.3) ∞ (Rd,M) and bmoc(Rd,M). To this 1 1 2 q kfkbmoc ≈ kφ ∗ fkN + sup where Φ ∈ S(Rd) and φ ∈ H σ (3.3) to J αf : 2 (Rd) satisfying (1.7). By taking φ = ϕ0 and Φ = J−αϕ, we apply , M QZQ Xj≥− log2(l(Q)) Q<1(cid:13)(cid:13)(cid:13) QZQ Xj≥− log2(l(Q)) Q<1(cid:13)(cid:13)(cid:13) QZQ Xj≥− log2(l(Q)) Q<1(cid:13)(cid:13)(cid:13) Φj ∗ f (s)2ds(cid:13)(cid:13)(cid:13) (J−αϕ)j ∗ (J αf )(s)2ds(cid:13)(cid:13)(cid:13) 22jαϕj ∗ f (s)2ds(cid:13)(cid:13)(cid:13) M 1 1 1 2 1 2 M kJ αfkbmoc ≈ kϕ0 ∗ fkN + sup = kϕ0 ∗ fkN + sup = kfkF α,c ∞ . Since J α is also an isomorphism between F α,c 1-bmoc duality and the hc hc with equivalent norms. p-hc p (Rd,M) and hc p(Rd,M) for any 1 < p < ∞, by the (Rd,M) (cid:3) (Rd,M)∗ = F −α,c q q duality in Theorem 1.1, we see that F α,c p 22 R. Xia and X. Xiong 3.2. Interpolation. Now we indicate a complex interpolation result of Triebel-Lizorkin spaces. It is deduced from the interpolation of local Hardy and bmo spaces in Theorem 1.2, and the boundedness of complex order Bessel potentials on them. Proposition 3.7. Let α0, α1 ∈ R and 1 < p < ∞. Then (cid:0)F α0,c ∞ (Rd,M), F α1,c 1 (Rd,M)(cid:1) 1 p = F α,c p (Rd,M), α = (1 − 1 p )α0 + α1 p . Proof. Let f ∈ F α,c p(Rd,M). Therefore, according to Theorem 1.2 (1), there exists a continuous function on the strip {z ∈ C : 0 ≤ Rez ≤ 1}, analytic in the interior, such that J αf = F ( 1 (Rd,M). By Proposition 3.4, we have J αf ∈ hc p p ) ∈ hc p(Rd,M) and sup sup t∈R kF (it)kbmoc < ∞ and t∈R kF (1 + it)khc 1 < ∞. We consider Bessel potentials of complex order. For z ∈ C, define Jz(ξ) = (1 + ξ2) z be the associated Fourier multiplier. We set 2 , and J z to For any t ∈ R, and keF (it)kF keF (1 + it)kF t∈R keF (it)kF This will imply that f = eF ( 1 F α,c sup p J−(1−z)α0−zα1 F (z). p )2 eF (z) = e(z− 1 ∞ ≈ e−t2+ 1 1 ≈ e−t2+(1− 1 α0 ,c α1,c p2 kJ it(α0−α1)F (it)kbmoc p )2 kJ it(α0−α1)F (1 + it)khc 1 . α1,c 1 sup α0,c ∞ < ∞ and ∞ (Rd,M), F α1,c t∈R keF (1 + it)kF (Rd,M)(cid:1) 1 ∞ (Rd,M), F α1,c p ) ∈(cid:0)F α0,c (Rd,M) ⊂(cid:0)F α0,c 1 1 p (Rd,M)(cid:1) 1 p . Hence, . < ∞. We claim that J it is a bounded Fourier multiplier on hc bmoc(Rd,M) too. Therefore, we will have 1(Rd,M), so by duality, it is bounded on By duality, we will get the reverse inclusion for the Calder´on's second interpolation (·,·) the inclusion (·,·) 1 we will obtain the desired assertion. 1 p . Then by 1 p between two kinds of complex interpolations (see [2, Theorem 4.3.1]), p ⊂ (·,·) Now, we prove the claim. First, we easily check that Jit is d-times differentiable on Rd \ {0}, and for any m ∈ Nd 0 and m1 ≤ d, we have Next, we check that (with Jit(2kξ) = (1 + 2kξ2) max −2≤k≤2kJit(2k·)ϕkHd 2 sup(cid:8)ξm1DmJit(ξ) : ξ 6= 0(cid:9) . (1 + t)d. and kJit(ϕ(0) + ϕ(1))kHd . (1 + t)d it 2 ), 2 . (1 + t)d. By (3) in Proposition 3.3, if we take (ϕj )j≥0 to be the Littlewood-Paley decomposition on Rd satisfying (0.2) and (0.3), we have and kJ itfkhc 1 ≈(cid:13)(cid:13)(Xj≥0 1 ≈(cid:13)(cid:13)(Xj≥0 −2≤k≤2 kJit(2k·)ϕkHd kfkhc 2 1 2(cid:13)(cid:13)1 Jit ∗ ϕj ∗ f2) ϕj ∗ f2) . 1 2(cid:13)(cid:13)1 Then, we apply Theorem 2.7 with ρj = ϕj, φj (2j·) = Jit, and α = 0, σ = d, 2(cid:9)kfkhc . max(cid:8) max ,kJit(ϕ(0) + ϕ(1))kHd kJ itfkhc 1 1 The claim is proved. . (1 + t)dkfkhc 1 . (cid:3) Inhomogeneous Triebel-Lizorkin spaces 23 Remark 3.8. The real interpolation of the couple(cid:0)F α,c of(cid:0)F α1,c that of Hardy spaces (see Theorem 1.2) and Proposition 3.4. But if α1 6= α2, the real interpolation in this paper, and refer the reader to [40] for similar results on homogeneous Triebel-Lizorkin (and Besov) spaces. (Rd,M)(cid:1) follows easily from (Rd,M)(cid:1) will give Besov type spaces. We will not consider this problem ∞ (Rd,M), F α2,c ∞ (Rd,M), F α,c 1 1 1 3.3. Triebel-Lizorkin spaces with α > 0. The following result shows that when α > 0, the F α,c (Rd,M)-norm can be rewritten as the sum of two homogeneous norms. Recall that for a fixed Schwartz function ϕ in (0.1), the functions ϕj's determined by bϕj(ξ) = ϕ(2−jξ), j ∈ Z give a homogeneous Littlewood-Paley decomposition on Rd satisfying (0.4). Proposition 3.9. Let α > 0. If 1 ≤ p < ∞, then 22jα ϕj ∗ f2) , ∀ f ∈ F α,c p (Rd,M). If 1 ≤ p ≤ 2, kfkF α,c +∞Xj=−∞ p ≈ kϕ0 ∗ fkp +(cid:13)(cid:13)( +∞Xj=−∞ p ≈ kfkp +(cid:13)(cid:13)( 1 2(cid:13)(cid:13)p 2(cid:13)(cid:13)p , 1 ∀ f ∈ F α,c Proof. Firstly, we prove the first equivalence. By the definition of the F α,c 22jα ϕj ∗ f2) kfkF α,c p (Rd,M). -norm, it is obvious that To prove the reverse inequality, it suffices to show: 22jα ϕj ∗ f2) kfkF α,c p . kϕ0 ∗ fkp +(cid:13)(cid:13)( +∞Xj=−∞ +∞Xj=1 2(cid:13)(cid:13)p . kϕ0 ∗ fkp +(cid:13)(cid:13)( 1 22jα ϕj ∗ f2) 22jα ϕj ∗ f2) . 1 2(cid:13)(cid:13)p 0Xj=−∞ (cid:13)(cid:13)( p . 1 2(cid:13)(cid:13)p By the support assumption of ϕ, we have ϕ(0) = 1 for any ξ ≤ 1. Thus, when j < 0, Then (3.4) ϕ(2j·) = ϕ(2j·)ϕ(0). ϕj ∗ f = ϕj ∗ ϕ0 ∗ f. By the triangle inequality, (3.4) and [39, Lemma 1.7], we obtain 0Xj=−∞ (cid:13)(cid:13)( 2jα ϕj ∗ f2) 1 2(cid:13)(cid:13)p . 2jαk ϕj ∗ ϕ0 ∗ fkp + k ϕ0 ∗ fkp . −1Xj=−∞ −1Xj=−∞ 0Xj=−∞ +∞Xj=1 . kϕ0 ∗ fkp +(cid:13)(cid:13)( . 2jαk ϕjk1kϕ0 ∗ fkp +(cid:13)(cid:13)ϕ(ϕ0 + ϕ1 + ϕ2) ∗ f(cid:13)(cid:13)p +∞Xj=1 2jαkϕ0 ∗ fkp +(cid:13)(cid:13)( 22jαϕj ∗ f2) 2(cid:13)(cid:13)p 1 22jαϕj ∗ f2) 1 . 2(cid:13)(cid:13)p 2(cid:13)(cid:13)p gives rise to an equivalent 1 Therefore, we have proved that kϕ0 ∗ fkp +(cid:13)(cid:13)(P+∞j=1 22jαϕj ∗ f2) (Rd,M) when α > 0. norm on F α,c p Now let us deal with the second equivalence. For any 1 ≤ p ≤ 2 and α > 0, we have (Rd,M) ⊂ hc . Combined with the equiva- p(Rd,M) ⊂ Lp(N ). Therefore kfkp . kfkF α,c F α,c lence obtained above, we see that p p kfkp +(cid:13)(cid:13)( +∞Xj=−∞ 22jα ϕj ∗ f2) 1 2(cid:13)(cid:13)p . kfkF α,c p . The reverse inequality can be easily deduced by the fact that kϕ0 ∗ fkp ≤ kϕ0k1kfkp. (cid:3) 24 R. Xia and X. Xiong We also have a continuous counterpart of Proposition 3.9. For any ε ≥ 0, we define ϕε = F−1(ϕ(ε·)). Corollary 3.10. Let 1 ≤ p ≤ 2 and α > 0. Then, for any f ∈ F α,c ε−2α ϕε ∗ f2 dε ) kfkF α,c ε p p ≈ kfkp +(cid:13)(cid:13)(cid:13)(Z ∞ 0 4. Characterizations (Rd,M), . 1 2(cid:13)(cid:13)(cid:13)p In this section we give two kinds of characterizations of the Triebel-Lizorkin spaces defined previously: one is done by directly replacing the function ϕ in Definition 3.1 by more general convolution kernels; the other is described by Lusin square functions. Since the local Hardy spaces are included in the family of inhomogeneous Triebel-Lizorkin spaces, these two characterizations can be seen as extensions as well as improvements of those in [36] for local Hardy spaces, listed in Theorems 1.4 and 1.5. The multiplier theorems in section 2 will play a crucial role in this section. 4.1. General characterizations. We have seen in section 3.1 that the definition of Triebel- Lizorkin spaces is independent of the choice of ϕ satisfying (0.1). In this section, we will show that this kernel is not even necessarily a Schwartz function. Let σ > d 2 and Φ(0), Φ be two complex-valued infinitely differentiable functions defined respec- tively on Rd and Rd\{0}, which satisfy Φ(0)(ξ) > 0 if ξ ≤ 2, 2−kα0kΦ(0)(2k·)ϕkHσ sup k∈N0 2 < ∞,  1 2 ≤ ξ ≤ 2, 2−kα0kΦ(2k·)ϕkHσ (1 + s2)σF−1(Φϕ(0)I−α1 )(s)ds < ∞. < ∞, 2 Φ(ξ) > 0 if sup k∈N0  ZRd (4.1) and (4.2) write Then (4.4) Let Φ(j) = Φ(2−j·) for j ≥ 1, and Φj be the function whose Fourier transform is equal to Φ(j) Recall that here I−α1 (ξ) is the symbol of the Riesz potential I−α1 = (−(2π)−2∆) −α1 for any j ∈ N0. Theorem 4.1. Let 1 ≤ p < ∞ and α ∈ R. Assume that α0 < α < α1, α1 ≥ 0 and Φ(0), Φ satisfy conditions (4.1) and (4.2) respectively. Then for any f ∈ S′(Rd; L1(M) + M), we have (4.3) . , 1 2 kfkF α,c p ≈(cid:13)(cid:13)(Xj≥0 22jαΦj ∗ f2) 2(cid:13)(cid:13)p where the relevant constants are independent of f . Proof. We follow the pattern of the proof of [40, Theorem 4.17]. Denote the norm on the right hand side of (4.3) by kfkF α,c Step 1. Let ϕk = 0 (and so is ϕ(k)) if k < 0. Given a positive integer K, for any j ∈ N0, we p,Φ . Φ(j) = K−1Xk=−∞ Φj ∗ f = Xk≤K−1 Φ(j)ϕ(j+k) + ∞Xk=K Φj ∗ ϕj+k ∗ f + Xk≥K Φ(j)ϕ(j+k). Φj ∗ ϕj+k ∗ f. Temporarily we take for granted that the second series is convergent not only in S′(Rd; L1(M)+M) but also in F α,c (Rd,M), which is to be settled up in the last step. Then we obtain p kfkF α,c p,Φ ≤ I + II + III, Inhomogeneous Triebel-Lizorkin spaces 25 where The term II is easy to deal with. By (0.5) and (4.1), we obtain K−1Xk=0 kΦ0 ∗ ϕk ∗ fkp = kΦ0 ∗ (ϕk−1 + ϕk + ϕk+1) ∗ ϕk ∗ fkp 22jαΦj ∗ ϕj+k ∗ f2) kΦ0 ∗ ϕk ∗ fkp, 22jαΦj ∗ ϕj+k ∗ f2) , 1 2(cid:13)(cid:13)p 2(cid:13)(cid:13)p . 1 I = Xk≤K−1(cid:13)(cid:13)(Xj≥1 II = Xk≤K−1 III = Xk≥K(cid:13)(cid:13)(Xj≥0 K−1Xk=0 K−1Xk=0 . kϕk ∗ fkpkΦ0 ∗ (ϕk−1 + ϕk + ϕk+1)k1 . sup k∈N0 2−kα0kΦ(0)(2k·)ϕkHσ 2 . CKkfkF α,c p . K−1Xk=0 2k(α0−α)k2kαϕk ∗ fkp Let us treat the terms I and III separately. By the support assumption of ϕ(k) and the property that it is equal to 1 when ξ ≤ 1, for k ≤ K − 1, we have Φ(ξ)ϕ(0)(2−Kξ) Φ(ξ)ϕ(k)(ξ) = (4.5) ξα1 = 2kα1 η(ξ)ρ(k)(ξ), ξα1 ϕ(k)(ξ) where η, ρ are defined by η(ξ) = Let η(j) = η(2−j·), j ∈ Z. For j ≥ 1, define ηj = F−1(η(j)). Then for any j ≥ 1, we have Φ(ξ)ϕ(0)(2−Kξ) ξα1 and ρ(ξ) = ξα1 ϕ(ξ). Φj ∗ ϕj+k ∗ f = 2kα1 ηj ∗ ρj+k ∗ f. Now we are ready to estimate I. Applying Theorems 2.2 and 2.7 twice, we get (4.6) 1 1 1 , max 22(j+k)αηj ∗ ρj+k ∗ f2) I = Xk≤K−1 2k(α1−α)(cid:13)(cid:13)(Xj≥1 2(cid:13)(cid:13)p = Xk≤K−1 2k(α1−α)(cid:13)(cid:13)( Xj≥k+1 2(cid:13)(cid:13)p 22jαηj−k ∗ ρj ∗ f2) . Xk≤K−1 2(cid:9) 2k(α1−α) max(cid:8)kη(−k)(ϕ(0) + ϕ(1))kHσ −2≤ℓ≤2kη(−k−ℓ)ϕkHσ ·(cid:13)(cid:13)(Xj≥0 2(cid:13)(cid:13)p 22jαρj ∗ f2) . Xk≤K−1 2k(α1−α) max(cid:8)kη(−k)(ϕ(0) + ϕ(1))kHσ 2(cid:9)(cid:13)(cid:13)(Xj≥0 · max(cid:8)kIα1 (ϕ(0) + ϕ(1))kHσ = Xk≤K−1 2k(α1−α) max(cid:8)kη(−k)(ϕ(0) + ϕ(1))kHσ 2(cid:9)kfkF α,c · max(cid:8)kIα1 (ϕ(0) + ϕ(1))kHσ 2(cid:9) −2≤ℓ≤2kη(−k−ℓ)ϕkHσ 2(cid:13)(cid:13)p 22jαϕj ∗ f2) 2(cid:9) −2≤ℓ≤2kη(−k−ℓ)ϕkHσ ,kIα1 ϕkHσ ,kIα1 ϕkHσ , max , max . p 2 2 2 2 2 1 Let us deal with all the factors in the last term of the above inequality. Firstly, when α1 = 0, it 2 (Rd). Secondly, we treat the case α1 > 0. < ∞. Next, we deal with the term Iα1 (ϕ(0) + ϕ(1)), which can is obvious that Iα1 ϕ ∈ H σ First, it is easy to see that kIα1 ϕkHσ 2 (Rd) and Iα1 (ϕ(0) + ϕ(1)) ∈ H σ 2 26 R. Xia and X. Xiong be reduced to Iα1 ϕ(0) by dilation. Let N be a positive integer such that α1 > 1 N . If the dimension d is odd, we consider the function F (z) = e(z− N +2 N )zϕ(0), which is continuous on the strip {z ∈ C : 0 ≤ Re(z) ≤ 1}, and analytic in the interior. A direct computation shows that supt∈R kF (it)kH 2 , we have < ∞ and supt∈R kF (1 + it)kH < ∞. Then for θ = ξα1− 1 N + 1 2 1 N +1 N +(1+ 1 2N +2 )2 2− 1 > 1 2 − 1 + 1 2 d 2 2 d 2 1 2 F (θ) = Iα1 ϕ(0) ∈ H σ (Rd), H 2 . On the other hand, if d is even, set F (z) = e(z− 1 < ∞, and that supt∈R kF (1 + it)kH 2 (Rd) =(cid:0)H 2 − 1 2 d 2 2 d 2 + 1 2 2N )2 d 2 2 +1 d 2 2 for some σ > d check that supt∈R kF (it)kH have F (θ) = Iα1 ϕ(0) ∈ H d 2 + 1 2 2N (Rd) =(cid:0)H d 2 2 (Rd), H d 2 +1 2 Thus, for any α1 > 0, we can always choose a positive σ > d 2 (Rd). Finally, we have to estimate kη(−k)ϕkHσ 2 uniformly in k, which will yield the convergence of the last sum in (4.6) by dilation again. To this end, note that by (4.2), η is integrable on Rd, then we use the Cauchy-Schwarz inequality in the following way: 2 such that Iα1 ϕ(0) ∈ H σ 2 and kη(−k)ϕ(0)kHσ 2 ϕ(0). We can also 2N , we , (Rd)(cid:1)θ ξN α1z+ α1 < ∞. Then for θ = 1 (Rd)(cid:1)θ . F−1(η(−k)ϕ)(s)2 = ZRd η(t)F−1(ϕ)(s − 2kt)dt2 ≤ kηk1ZRd η(t) · F−1(ϕ)(s − 2kt)2dt. (1 + s − 2kt2)σF−1(ϕ)(s − 2kt)2dsdt Hσ 2 For k ≤ K − 1, we have kη(−k)ϕk2 ≤ kηk1ZRd . kηk1ZRd ≤ 2Kσkηk1ZRd ≤ Cϕ0,σ,K(cid:0)ZRd (4.7) The other term kη(−k)ϕ(0)kHσ (4.8) (1 + s2)σF−1(η(−k)ϕ)(s)2ds =ZRd (1 + s2)σZRd η(t) · F−1(ϕ)(s − 2kt)2dtds (1 + 2kt2)ση(t)ZRd (1 + t2)ση(t)dtZRd (1 + t2)ση(t)dt(cid:1)2 I . CΦ,ϕ(0),α1,α,KZRd η = I−α1 Φ[ϕ(0)(2−K·) − ϕ(0)] + I−α1 Φϕ(0). 2 is dealt with in the same way. (1 + t2)ση(t)dtkfkF α,c . p . (1 + s2)σF−1(ϕ)(s)2ds Going back to the estimate of I, by the previous inequalities, we obtain In order to return from η back to ϕ0, we write Since I−α1 Φ(ϕ(0)(2−K·) − ϕ(0)) is an infinitely differentiable function with compact support, we have (1 + t2)σF−1(I−α1 Φ(ϕ(0)(2−K·) − ϕ(0)))(t)dt = C′Φ,ϕ(0) ,α1,α,K < ∞. Then (4.2) implies that ZRd (1 + t2)ση(t)dt . C′Φ,ϕ(0),α1,α,K +ZRd ZRd Therefore, (1 + s2)σF−1(I−α1 Φϕ(0))(s)ds < ∞. Step 2. Now it remains to estimate the third term III. Let H be a Schwartz function such that I . kfkF α,c p . (4.9) supp H ⊂ {ξ ∈ Rd : 1 4 ≤ ξ ≤ 4} and H(ξ) = 1 if 1 2 ≤ ξ ≤ 2. Inhomogeneous Triebel-Lizorkin spaces 27 Let H (k) = H(2−k·). For k ≥ K, we have Φ(ξ)ϕ(k)(ξ) = (4.10) and (4.11) Φ(0)(ξ)ϕ(k)(ξ) = Φ(ξ) ξα0 H (k)(ξ)ϕ(k)(ξ)ξα0 , Φ(0)(ξ) ξα0 H (k)(ξ)ϕ(k)(ξ)ξα0 . For any j ∈ N0, we keep using the notation Φj = F−1(Φ(j)) and Hj = F−1(H (j)). Thus, we have Φj ∗ ϕj+k ∗ f = 2kα0 (I−α0 Φ)j ∗ Hj+k ∗ (Iα0 ϕ)j+k ∗ f. Since both H and ϕ vanish near the origin, by Theorems 2.2 and 2.7, we obtain Therefore, III = Xk≥K = Xk≥K Xk≥K 2k(α0−α)(cid:13)(cid:13)(cid:0)Xj≥k max(cid:8)2−kα0 max · Xk≥K 2k(α0−α)kfkF α,c . sup k∈N0 . p 2(cid:13)(cid:13)p 22(j+k)α(I−α0 Φ)j ∗ Hj+k ∗ (Iα0 ϕ)j+k ∗ f2(cid:1) 1 2(cid:13)(cid:13)p 22jα(I−α0 Φ)j−k ∗ Hj ∗ (Iα0 ϕ)j ∗ f2(cid:1) 1 2k(α0−α)(cid:13)(cid:13)(cid:0)Xj≥0 2k(α0−α)(cid:13)(cid:13)(cid:0)Xj≥k 2(cid:13)(cid:13)p 22jα(I−α0 Φ)j−k ∗ Hj ∗ (Iα0 ϕ)j ∗ f2(cid:1) 1 −2≤ℓ≤2kI−α0 Φ(2k+ℓ·)H(2ℓ·)ϕkHσ . 2 2(cid:9) , 2−kα0kI−α0 Φ(0)(2k·)H(ϕ(0) + ϕ(1))kHσ Then by (2.6), (4.1) and (4.2), we have, for any −2 ≤ ℓ ≤ 2, 2−kα0kI−α0 Φ(2k+ℓ·)H(2ℓ·)ϕkHσ 2 ZRd ≤ 2−kα0kΦ(2k+ℓ·)ϕkHσ . 2−kα0kΦ(2k+ℓ·)ϕkHσ 2 ≤ sup k∈N0 2 (1 + t2)σF−1(I−α0 H(2ℓ·))(t)dt 2−kα0kΦ(2k·)ϕkHσ 2 < ∞, (4.12) and (4.13) 2−kα0kI−α0 Φ(0)(2k·)H(ϕ(0) + ϕ(1))kHσ = 2−kα0kI−α0Φ(0)(2k·)H ϕ(2−ℓ′·)kHσ 2 2 1Xℓ′=−2 . 2−kα0 ≤ 2−kα0 kI−α0Φ(0)(2k+ℓ′·)H(2ℓ′·)ϕkHσ 2 1Xℓ′=−2 1Xℓ′=−2 2−kα0kΦ(0)(2k·)ϕkHσ 2 ZRd kΦ(0)(2k+ℓ′·)ϕkHσ < ∞. 2 . sup k∈N0 Then we get (1 + t2)σF−1(I−α0 H(2ℓ·))(t)dt III ≤ CΦ,α0,α,KkfkF α,c Combining this estimate with those of I and II, we finally get . p kfkF α,c p,Φ . kfkF α,c p . Step 3. We turn to the reverse inequality. Note that ϕ(0)(ξ) = 1 when ξ ≤ 1, then by (4.1) and (4.2), for any j ∈ N0, we write (4.14) ϕ(j)(ξ) = ϕ(j)(ξ) ϕ(0)(2−j−M ξ) = ϕ(j)(ξ) Φ(j)(ξ) ϕ(0)(2−j−M ξ)Φ(j)(ξ), 28 R. Xia and X. Xiong where M is a positive integer to be chosen later. By Theorems 2.2 and 2.7, kfkF α,c p =(cid:13)(cid:13)(cid:0)Xj≥0 . max(cid:8) max ·(cid:13)(cid:13)(cid:0)Xj≥0 .(cid:13)(cid:13)(cid:0)Xj≥0 2(cid:13)(cid:13)p 22jαϕj ∗ f2(cid:1) 1 −2≤ℓ≤2kΦ−1(2ℓ·)ϕ(2ℓ·)ϕkHσ 2(cid:13)(cid:13)p 22jα(ϕ0)j+M ∗ Φj ∗ f2(cid:1) 1 2(cid:13)(cid:13)p 22jα(ϕ0)j+M ∗ Φj ∗ f2(cid:1) 1 , 2 2(cid:9) ,k(Φ(0))−1ϕ(0)(ϕ(0) + ϕ(1))kHσ where (ϕ0)j+M is the Fourier inverse transform of ϕ(0)(2−j−M·). Let h = 1 − ϕ(0). Write ϕ(0)(2−j−M ξ)Φ(j)(ξ) = Φ(j)(ξ) − h(j+M)(ξ)Φ(j)(ξ). Then, we have kfkF α,c p . kfkF α,c p,Φ +(cid:13)(cid:13)(cid:0)Xj≥0 2(cid:13)(cid:13)p 22jαhj+M ∗ Φj ∗ f2(cid:1) 1 , where the relevant constant depends only on p, σ, d and ϕ(0). Applying the arguments in the estimate of III, (4.10) with h(M)Φ in place of Φ and (4.11) with h(M)Φ(0) in place of Φ(0), we deduce ≤ C1 sup k≥M (cid:13)(cid:13)(cid:0)Xj≥0 2(cid:13)(cid:13)p 22jαhj+M ∗ Φj ∗ f2(cid:1) 1 2−kα0 max(cid:8) max −2≤ℓ≤2kh(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ · Xk≥M 2k(α0−α)kfkF α,c 2−kα0 max(cid:8) max 2M(α0−α) 1 − 2α0−αkfkF α,c −2≤ℓ≤2kh(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ = sup k≥M · C1 , p p 2 2 2(cid:9) ,kh(2k−M·)Φ(0)(2k·)ϕkHσ 2(cid:9) ,kh(2k−M·)Φ(0)(2k·)ϕkHσ where C1 is a constant which depends only on p, σ, d, H and α0. Now we replace h in the above Sobolev norm by 1 − ϕ(0): kh(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ 2 ≤ kΦ(2k+ℓ·)ϕkHσ 2 + kϕ(0)(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ 2 . The support assumptions of ϕ(0) and ϕ imply that when k ≥ M , ϕ(0)(2k−M+ℓ·)ϕ 6= 0 if and only if k + ℓ = M or k + ℓ = M + 1. Then by (2.6), we have kϕ(0)(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ 2 ≤ C2kΦ(2k+ℓ·)ϕkHσ 2 , where C2 depends only on ϕ(0), σ and d. Thus, kh(2k−M+ℓ·)Φ(2k+ℓ·)ϕkHσ 2 ≤ (1 + C2)kΦ(2k+ℓ·)ϕkHσ 2 . Similarly, we have Putting all the estimates that we have obtained so far together, we get kh(2k−M·)Φ(0)(2k·)ϕkHσ 2 ≤ (1 + C2)kΦ(0)(2k·)ϕkHσ 2 . kfkF α,c p ≤ C3(cid:16)C1(1 + C2) p,Φ(cid:17), + kfkF α,c 2M(α0−α) 1 − 2α0−α sup k≥M 2−kα0 max{kΦ(2k·)ϕkHσ 2 ,kΦ(0)(2k·)ϕkHσ 2 }kfkF α,c p where the three constants C1, C2, C3 are independent of M , so we could take M large enough to make sure the multiple of kfkF α,c 2 , so that we have above is less than 1 p kfkF α,c p . kfkF α,c p,Φ . Inhomogeneous Triebel-Lizorkin spaces 29 Step 4. We now settle the convergence issue of the second series in (4.4). For every j ≥ 0, Φj ∗ ϕj+k ∗ f is an L1(M) + M-valued tempered distribution on Rd. We now show that the series converges in S′(Rd; L1(M) + M). By (4.12) and (4.13), for any L > K, we have kΦj ∗ ϕj+k ∗ fkp 2jα LXk=K . kIα0 ϕkHσ . kfkF α,c . p 2 Xk≥K 2k(α0−α) sup k∈N0 max(cid:8)2−kα0kΦ(2k·)ϕkHσ 2 , 2−kα0kΦ(0)(2k·)ϕkHσ 2(cid:9)kfkF α,c p Therefore, for any j ≥ 0, Pk≥K+1 Φj ∗ ϕj+k ∗ f converges in Lp(N ), so in S′(Rd; L1(M) + M) too. In the same way, we can show that the series also converges in F α,c the proof. (Rd,M), which completes (cid:3) p The following is the continuous analogue of Theorem 4.1. We use similar notation for continuous parameters: given ε > 0, Φε denotes the function whose Fourier transform is Φ(ε) = Φ(ε·). Theorem 4.2. Keep the assumption of the previous theorem. Then for f ∈ S′(Rd; L1(M) + M), we have (4.15) kfkF α,c p ≈ kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(cid:0)Z 1 0 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 ε−2αΦε ∗ f2 dε . Proof. This proof is very similar to the previous one. We keep the notation there and only point out the necessary modifications. First, we need to discretize the integral on the right hand side of (4.15). There exist two constants C1, C2 such that C1 ∞Xj=0 22jαZ 2−j 2−j−1 Φε ∗ f2 dε ε ≤Z 1 0 ε−2αΦε ∗ f2 dε ε ≤ C2 ∞Xj=0 22jαZ 2−j 2−j−1 Φε ∗ f2 dε ε . By approximation, we can assume that f is good enough so that each integral over the interval (2−j−1, 2−j) can be approximated uniformly by discrete sums. Instead of Φ(j)(ξ) = Φ(2−jξ), we have now Φ(ε)(ξ) = Φ(εξ) with 2−j−1 < ε ≤ 2−j. We transfer the split (4.5) into: Φ(ε)(ξ)ϕj+k(ξ) = Φ(2−j · 2jεξ)ϕ(0)(2−Kξ) 2−jξα1 2−jξα1 ϕj+k(ξ). Thus, with Φε ∗ ϕj+k ∗ f = 2kα1 ηj ∗ ρj+k ∗ f η(ξ) = Φ(2jεξ)ϕ(0)(2−K ξ) ξα1 and ρ(ξ) = ξα1 ϕ(ξ). We proceed as in step 1 of the previous theorem, where we transfer (4.7) to the present setting: kη(−k)ϕkHσ 2 (1 + t2)ση(t)dt . Cϕ(0) ,σ,kZRd = Cϕ(0) ,σ,kZRd (1 + t2)σ(cid:12)(cid:12)F−1(cid:0)I−α1 Φ(δj·)ϕ(0)(2−K·)(cid:1)(t)(cid:12)(cid:12)dt j ZRd j 2−K·)(cid:1)(t)(cid:12)(cid:12)dt, (1 + t2)σ(cid:12)(cid:12)F−1(cid:0)I−α1 Φϕ(0)(δ−1 ≤ Cϕ(0) ,σ,kδα1 30 R. Xia and X. Xiong where δj = 2jε and 1 2 < δj ≤ 1. The last integral is estimated as follows: ZRd j 2−K·)(cid:1)(t)(cid:12)(cid:12)dt (1 + t2)σ(cid:12)(cid:12)F−1(cid:0)I−α1 Φϕ(0)(δ−1 ≤ZRd (1 + t2)σF−1(I−α1 Φϕ(0))(t)dt +ZRd (1 + t2)σ(cid:12)(cid:12)F−1(cid:0)I−α1 Φ[ϕ(0) − ϕ(0)(δ−1 ≤ZRd (1 + t2)σF−1(I−α1 Φϕ(0))(t)dt 2 <δ≤1ZRd + sup 1 j 2−K·)](cid:1)(t)(cid:12)(cid:12)dt (1 + t2)σ(cid:12)(cid:12)F−1(cid:0)I−α1 Φ[ϕ(0) − ϕ(0)(δ−12−K·)](cid:1)(t)(cid:12)(cid:12)dt. 0 . Xk≤K−1 Xk≤K−1(cid:13)(cid:13)(cid:13)(cid:0)Z 1 Note that the above supremum is finite since I−α1 Φ[ϕ(0)−ϕ(0)(δ−12−K·)] is a compactly supported and infinitely differentiable function whose inverse Fourier transform depends continuously on δ. Then it follows that for 2−j−1 ≤ ε ≤ 2−j, 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 ε−2αΦε ∗ f2 dε kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(cid:0)Z 1 We make similar modifications in step 2 of the previous theorem and then establish the third part. Moreover, by the previous theorem, kΦ0 ∗ fkp . kfkF α,c 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 ε−2αΦε ∗ f2 dε there exists 2 < a ≤ 2√2 such that Φ(ξ) > 0 on {ξ : a−1 ≤ ξ ≤ a}. Then for j ≥ 1, Rj = {ε : a−12−j+1 < ε ≤ a2−j−1} are disjoint sub intervals on (0, 1], and ϕ(j) Φ(ε) is well-defined for any ε ∈ Rj. We slightly modify (4.14) as follows: for any ε ∈ Rj , we have For the reverse inequality, we follow the argument in step 3 in the previous proof. By (4.2), 2k(α1−α)kfkF α,c p . kfkF α,c . Thus, we have proved . kfkF α,c p 0 . . p p ϕ(j)(ξ) = ϕ(j)ϕ(0)(2−j−K ξ) = ϕ(j)(ξ) Φ(ε)(ξ) ϕ(0)(2−j−K ξ)Φ(ε)(ξ), j ∈ N0. Since for any −2 ≤ ℓ ≤ 2, kΦ−1(2−jε−12ℓ·)ϕ(2ℓ·)ϕkHσ 2 ≤ sup 2a−1≤δ≤ a 2 and k(Φ(0))−1(2−jε−1·)ϕ(0)(ϕ(0) + ϕ(1))kHσ 2 ≤ sup 2a−1≤δ≤ a 2 kΦ−1(δ2ℓ·)ϕ(2ℓ·)ϕkHσ 2 < ∞ k(Φ(0))−1(δ·)ϕ(0)(ϕ(0) + ϕ(1))kHσ 2 < ∞, we follow the argument in step 3 in the previous theorem to get kfkF α,c p .(cid:13)(cid:13)(cid:0)Xj≥0 . kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(cid:0)Z 1 22jαZRj (ϕ0)j+k ∗ Φε ∗ f2(cid:1) 1 2(cid:13)(cid:13)p 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 ε−2αΦε ∗ f2 dε 0 +(cid:13)(cid:13)(cid:16)Xj≥0 22jαZRj hj+k ∗ Φε ∗ f2 dε 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 . (cid:3) The remaining of the proof follows step 3 with necessary modifications. We now concretize the general characterization in the previous theorem to the case of Poisson kernel. Recall that P denotes the Poisson kernel of Rd and Pε(f )(s) =ZRd Pε(s − t)f (t)dt, (s, ε) ∈ Rd+1 + . The following theorem improves [35, Section 2.6.4] even in the classical case: [35, Section 2.6.4] requires k > d + max{α, 0} for the Poisson characterization while we only need k > max{α, 0}. The proof of this theorem is similar to but easier than that of [40, Theorem 4.20], since we assume k > 0 here; we omit the details. The key ingredient is the improvement of the characterization of Hardy spaces in terms of Poisson kernel given in [38, Theorem 1.5] Inhomogeneous Triebel-Lizorkin spaces 31 Theorem 4.3. Let 1 ≤ p < ∞, α ∈ R, and k ∈ N such that k > max{α, 0}. Assume that Φ(0) satisfies (4.1). Then for f ∈ S′(Rd; L1(M) + M), we have kfkF α,c p ≈ kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(cid:0)Z 1 0 ε2(k−α)(cid:12)(cid:12) ∂k 2(cid:13)(cid:13)(cid:13)p ∂εk Pε(f )(cid:12)(cid:12)2 dε ε (cid:1) 1 . 4.2. Characterizations via Lusin functions. We are going to give some characterizations for Triebel-Lizorkin spaces via Lusin square functions. As what we did in the previous part of this section, we still use Fourier multiplier theorems as our main tool. But now we have to rely on the Hilbertian (instead of ℓ2) versions of the Fourier multiplier theorems. The following characterization, via Lusin square functions associated to ϕ given by the condition (0.1), is a special case of the characterization in Theorem 1.5. We keep using the notation ϕj being the function whose Fourier transform is equal to ϕ(2−j·) for j ∈ N, and ϕ0 being the function whose Fourier transform is equal to 1 −Pj≥1 ϕ(2−j·). Proposition 4.4. For 1 ≤ p < ∞ and f ∈ F α,c (4.16) p kfkF α,c p ≈ kϕ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(Xj≥1 (Rd,M), we have 2j(2α+d)ZB(0,2−j ) ϕj ∗ f (· + t)2dt) . 1 2(cid:13)(cid:13)(cid:13)p (Rd,M), by the lifting property in Proposition 3.4, we have J αf ∈ Proof. For any f ∈ F α,c p(Rd,M). Then, we apply the discrete characterization in Theorem 1.5 with φ = J−αϕ0 and hc Φ = I−αϕ to J αf , p Following the argument in the proof of (2.9), we can prove kfkF α,c p ≈ kJ αfkhc p ≈ kϕ0 ∗ fkp + ksc,D I−αϕ(J αf )kp. Moreover, we can easily check that . I−αϕ(I αf )(cid:13)(cid:13)p (cid:13)(cid:13)sc,D I−αϕ(J αf )(cid:13)(cid:13)p ≈(cid:13)(cid:13)sc,D I−αϕ(I αf )(cid:13)(cid:13)p =(cid:13)(cid:13)(cid:13)(Xj≥1 (cid:13)(cid:13)sc,D p ≈ kϕ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(Xj≥1 2j(2α+d)ZB(0,2−j ) ϕj ∗ f (· + t)2dt) 2(cid:13)(cid:13)(cid:13)p 2j(2α+d)ZB(0,2−j ) ϕj ∗ f (· + t)2dt) 2(cid:13)(cid:13)(cid:13)p . 1 1 kfkF α,c Therefore, we conclude The assertion is proved. . (cid:3) From the above Lusin square function by ϕ, we can deduce Lusin type characterizations with general convolution kernels by the aide of Theorems 2.5 and 2.9. Theorem 4.5. Let 1 ≤ p < ∞ and α ∈ R. Assume that α0 < α < α1, α1 ≥ 0 and Φ(0), Φ satisfy conditions (4.1), (4.2). Then for any f ∈ S′(Rd; L1(M) + M), we have kfkF α,c p (Rd,M) ≈ kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(Xj≥1 where the equivalent constant is independent of f . 2j(2α+d)ZB(0,2−j ) Φj ∗ f (· + t)2dt) , 1 2(cid:13)(cid:13)(cid:13)p Proof. This proof is very similar to that of Theorem 4.1. The main target is to replace the standard test functions ϕ0 and ϕ in Proposition 4.4 with Φ0 and Φ satisfying (4.1) and (4.2). This time we need to use the Lusin type multiplier theorem i.e. Theorem 2.5, instead of Theorem 2.2. For the special case p = 1, we apply Theorem 2.9 instead of Theorem 2.7. (cid:3) Using a similar argument as in Theorem 4.2, we also have the following continuous analogue of the above theorem. This is the general characterization of Triebel-Lizorkin spaces by Lusin square functions. Recall thateΓ = {(t, ε) ∈ Rd+1 + : t < ε < 1}. 32 R. Xia and X. Xiong Theorem 4.6. Keep the assumption in the previous theorem. Then for any L1(M) + M-valued tempered distribution f on Rd, we have kfkF α,c p (Rd,M) ≈ kΦ0 ∗ fkp +(cid:13)(cid:13)(cid:13)(cid:0)ZeΓ 2(cid:13)(cid:13)(cid:13)p εd+1(cid:1) 1 ε−2αΦε ∗ f (· + t)2 dtdε . 5. Smooth atomic decomposition 1 This section is devoted to the study of atomic decomposition of F α,c (Rd,M). We aim to decompose F α,c (Rd,M) into atoms which have good enough size, smooth and moment conditions. To proceed in an orderly way step by step, we begin with the special case α = 0, i.e., the space hc 1(Rd,M) below does not lead to the one for general 1(Rd,M). Even though the result for hc F α,c (Rd,M) are already contained in those for hc 1(Rd,M). The main results in this section will be very useful in our forthcoming paper [37] on mapping properties of pseudo-differential operators. (Rd,M) directly, the main ingredients to obtain smooth decomposition for F α,c 1 1 1 5.1. Smooth atomic decomposition of hc 1(Rd,M). In the classical theory, the smooth atoms have been widely studied and have played a crucial role when studying the mapping properties of pseudo-differential operators acting on local Hardy spaces, or more generally, on Triebel-Lizorkin spaces. Details can be found in [3], [4] and [35]. In this subsection, we will show that in our operator-valued case, the atoms in Theorem 1.6 can also be refined to be infinitely differentiable. As in the classical case, the theory of tent spaces will be of great service in our proof of smooth atomic decomposition theorem. Tent spaces in the operator-valued setting have been introduced in [18] and [19] first; see also [38] for further complement. For our use, we study the local version of tent spaces defined in [36]. For any function defined on the strip S = Rd × (0, 1) with values in L1(M) + M, whenever it exists, we define First, we introduce a lemma concerning the atomic decomposition of the tent space T c 1 (Rd,M). , s ∈ Rd. 2 For 1 ≤ p < ∞, we define equipped with the norm kfkT c Ac(f )(s) =(cid:16)ZeΓ f (t + s, ε)2 dtdε εd+1(cid:17) 1 p (Rd,M) = {f : Ac(f ) ∈ Lp(N )} T c p (Rd,M) = kAc(f )kp. A function a ∈ L1(cid:0)M; L2(S, dsdε ε )(cid:1) is called a T c • τ(cid:16)RT (Q) a(s, ε)2 dsdε ε (cid:17) 1 ≤ Q− 1 2 . • supp a ⊂ T (Q) for some cube Q in Rd with Q ≤ 1; 1 -atom if Let T c 2 (5.1) f = λj aj, 1,at(Rd,M) be the space of all f : S → L1(M) admitting a representation of the form ∞Xj=1 ∞Xj=1 where the aj's are T c the following norm 1 -atoms and λj ∈ C such that P∞j=1 λj < ∞. We equip T c 1,at(Rd,M) with kfkT c 1,at = inf{ λj : f = λjaj; aj's are T c 1 -atoms, λj ∈ C}. ∞Xj=1 Lemma 5.1. We have T c Proof. In order to prove T c satisfies kakT c 1 1,at(Rd,M) = T c 1,at(Rd,M) ⊂ T c . 1. By the support assumption of a, we have 1 (Rd,M) with equivalent norms. 1 (Rd,M), it is enough to show that any T c 1 -atom a kakT c 1 =(cid:13)(cid:13)Ac(a)(cid:13)(cid:13)1 = τZRd(cid:16)Z 1 0 ZB(t,ε) a(s, ε)2 dsdε εd+1(cid:17) 1 2 τ(cid:16)ZRdZ 1 0 ZB(t,ε) a(s, ε)2 dsdε 2 τ(cid:16)ZT (Q) a(t, ε)2 dtdε ε (cid:17) 1 . Q d Q εd+1 = c 1 2 2 1 1 2 dt . 1. 2 dt(cid:17) 1 Inhomogeneous Triebel-Lizorkin spaces 33 Then by the duality T c 1 (Rd,M)∗ = T c ∞(Rd,M) (see [36]), we have T c ∞(Rd,M) ⊂ T c 1,at(Rd,M)∗. Now let Q be a cube in Rd with Q ≤ 1. If f ∈ L1(cid:0)M; Lc 2(T (Q), dsdε ε )(cid:1), then is a T c 2(T (Q), dsdε 2kfk−1 a = Q− 1 1 -atom supported in T (Q). Hence, ε )(cid:1)f ε )(cid:1). 1,at induces a continuous functional on L1(cid:0)M; Lc L1(cid:0)M;Lc 2kfkL1(cid:0)M;Lc 1,at ≤ Q 1,at(Rd,M) for every cube Q. Therefore, every continuous func- Thus, L1(cid:0)M; Lc tional ℓ on T c than or equal to Q 1 1,at)∗ . Let Q0 be the cube centered at the origin with side length 1 and Qm = Q0 + m for each m ∈ Zd. Then Rd = ∪m∈Zd Qm. Consequently, we can choose a sequence of functions (gm)m∈Zd such that ε )(cid:1) with norm smaller ε )(cid:1) ⊂ T c 2(T (Q), dsdε 2(T (Q), dsdε 2kℓk(T c kfkT c 2(T (Q), dsdε 1 ℓ(a) = τZT (Qm) a(s, ε)g∗m(s, ε) dsdε ε , ∀ T c 1 -atom a with supp a ⊂ T (Qm), and 1,at)∗ . 2(T (Qm), dsdε kgmkL∞(cid:0)M;Lc ℓ(a) = τZS Let g(s, ε) = gm(s, ε) for (s, ε) ∈ T (Qm). Then, we have dsdε ε )(cid:1) ≤ kℓk(T c ∀ T c It follows that, for any cube Q with Q ≤ 1, gT (Q) ∈ L∞(cid:0)M; Lc ε )(cid:1) ≤ Q 2(T (Qm), dsdε 1,at(Rd,M)∗ ⊂ T c which implies g ∈ T c 1,at(Rd,M)∗ with equivalent norms. Finally, by the density of T c T c get the desired equivalence. kgT (Q)kL∞(cid:0)M;Lc ∞(Rd,M). Hence, T c a(s, ε)g∗(s, ε) ε , 1 1 -atom a. 2(T (Q), dsdε ε )(cid:1) and 1,at)∗ , 2kℓk(T c ∞(Rd,M). Therefore, T c 1,at(Rd,M) in T c ∞(Rd,M) = 1 (Rd,M), we (cid:3) The following Lemma shows the connection between T c modelled on the classical argument of [3, Theorem 6]. p (Rd,M) and hc p(Rd,M). The proof is Lemma 5.2. Fix a Schwartz function Φ on Rd satisfying: Φ is supported in the cube with side length 1 and centered at the origin; (5.2)  RRd Φ(s)ds = 0; Let πΦ be the map given by Φ is nondegenerate in the sense of (1.2). Then πΦ is bounded from T c πΦ(f )(s) =Z 1 0 ZRd p (Rd,M) to hc dtdε , Φε(s − t)f (t, ε) p(Rd,M) for any 1 ≤ p < ∞. ε s ∈ Rd. Proof. For any 1 < p < ∞, let q be its conjugate index. By Theorem 1.1, it suffices to estimate τR πΦ(f )(s)g∗(s)ds, for any g ∈ hc q(Rd,M). Note that τZRd πΦ(f )(s)g∗(s)ds = τZRdZ 1 = τZRdZ 1 0 0 Φε(s − t)f (t, ε) dtdε ε g∗(s)ds f (t, ε)(eΦε ∗ g)∗(t) dεdt ε , 34 R. Xia and X. Xiong where eΦ(s) = Φ(−s). Then by the Holder inequality, cd(cid:12)(cid:12)(cid:12)τZRdZ 1 cd(cid:12)(cid:12)(cid:12)τZRdZeΓ πΦ(f )(s)g∗(s)ds(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)τZRd 1 1 = . kAc(f )kpksc pkgkhc . kfkT c q eΦ . 0 ZB(s,ε) f (t, ε)(eΦε ∗ g)∗(t) f (s + t, ε)(eΦε ∗ g)∗(s + t) (g)kq dεdt εd+1 dεdt εd+1 ds(cid:12)(cid:12)(cid:12) ds(cid:12)(cid:12)(cid:12) Now we deal with the case p = 1. The argument below is based on the atomic decompositions 1(Rd,M) and T c of hc 1 -atom to a bounded multiple of an hc 1 based on some cube Q with Q ≤ 1. Since Φ is supported in the unit cube, it follows from the definition of πΦ that πΦ(a) is supported 1 (Rd,M). By Lemma 5.1, it is enough to show that πΦ maps a T c in 2Q. Moreover, it satisfies the moment cancellation thatR πΦ(a)(s)ds = 0 since bΦ(0) = 0. So it remains to check that πΦ(a) satisfies the size estimate. To this end, we use the Cauchy-Schwarz inequality and the Plancherel formula (0.7), 1-atom. Let a be an atom in T c 2 dε kπΦ(a)kL1(M;Lc 2(Rd)) = τ(cid:0)ZRd \πΦ(a)(ξ)2dξ(cid:1) 1 = τ(cid:16)ZRd Z 1 ε 2dξ(cid:17) 1 0 bΦ(εξ)ba(ξ, ε) ≤ τ(cid:16)ZRdZ 1 ε Z 1 dξ(cid:17) 1 0 bΦ(εξ)2 dε 0 ba(ξ, ε)2 dε ≤ τ(cid:16)ZT (Q) a(s, ε)2 dsdε ε (cid:17) 1 ≤ Q− 1 2 . ε 2 2 2 (5.3) (5.4) where (5.5) (5.6) (5.7) Therefore we obtain the boundedness of πΦ from T c 1,at(Rd,M). Now we are able to refine the smoothness of the atoms given in Theorem 1.6. 1,at(Rd,M) to hc (cid:3) Theorem 5.3. For any f ∈ L1(M; Rc can be represented as d) + L∞(M; Rc d), f belongs to hc 1(Rd,M) if and only if it • the bj's are infinitely differentiable atoms supported in 2Q0,j with Q0,j = 1. For any 0, there exists a constant Cγ which depends on γ satisfying multiple index γ ∈ Nd • the gj's are infinitely differentiable atoms supported in 2Qj with Qj < 1, and such that f = ∞Xj=1 (µjbj + λjgj), τ(cid:0)Z2Q0,j Dγ bj(s)2ds(cid:1) 1 2 ≤ Cγ; 2 . Qj− 1 2 and gj(s)ds = 0; Z2Qj • the coefficients µj and λj are complex numbers such that (µj + λj) < ∞. τ(cid:0)Z2Qj gj(s)2ds(cid:1) 1 ∞Xj=1 1(Rd,M). Moreover, the infimum of (5.7) with respect to all admissible representations gives rise to an equivalent norm on hc Proof. Since the bj's and gj's are atoms in hc can be represented as in (5.4) and 1(Rd,M), it suffices to show that any f ∈ hc 1(Rd,M) ∞Xj=1 (µj + λj) . kfkhc 1 . Inhomogeneous Triebel-Lizorkin spaces 35 To begin with, we construct a smooth resolution of the unit on Rd. Let κ be a radial, real and infinitely differentiable function on Rd which is supported in the unit cube centered at the origin. Moreover, we assume thatbκ(0) > 0. We take bΦ = · 2bκ, which can be normalized as: = 1, ξ ∈ Rd\{0}. And we define (5.8) 0 ε Z ∞ bΦ(εξ)2 dε bφ(ξ) = 1 −Z 1 0 bΦ(εξ)2 dε ε , ξ ∈ Rd. By the Paley-Wiener theorem, bΦ can be extended to an analytic function bΦ(z) of d complex variables z = (z1, . . . , zd), and for any λ > 0, there exists a constant Cλ such that √d holds for any z = ξ1 + iξ2. Therefore, Z 1 0 bΦ(εz)2 dε 4 + eε( λ 2 )ξ2(ξ12 + ξ22) bΦ(z) ≤ Cλe( λ λZ 1 2 +√d)ξ2ε3dε · (ξ12 + ξ22)2 ε ≤ C2 λZ 1 2 +√d)ξ2(ξ12 + ξ22)2 ε3dε · e( λ ≤ C2 2 +√d)ξ2(1 + ξ12)2(1 + ξ22)2 λe( λ ≤ C2 λe(λ+2√d)ξ2(1 + ξ1)4. ≤ C2 0 0 Now applying the Paley-Wiener-Schwartz theorem to distributions, we obtain that φ is a distribu- tion with support in {s ∈ Rd : s ≤ 2√d}. On the other hand, if we define its value at the origin as 0, the function R 1 ε is an infinitely differentiable function on Rd, which ensures that φ is a Schwartz function. Thus, supp φ ⊂ {s ∈ Rd : s ≤ 2√d}. By (5.8), we arrive at the following 0 bΦ(ε·)2 dε decomposition of f : (5.9) f = φ ∗ f +Z 1 0 Φε ∗ Φε ∗ f dε ε . We first deal with φ ∗ f . By Theorem 1.6, we obtain an atomic decomposition of f : (5.10) where the aj's are hc 1-atoms andPj eµj . kfkhc f =Xj eµjaj, φ ∗ f =Xj eµj φ ∗ aj. 1. Thus, 2Q0 (with Q0 the unit cube centered at the origin), andPk∈Zd X0(s − k) = 1 for every s ∈ Rd. We now show that every φ∗ aj can be decomposed into smooth atoms supported in cubes with side length two. Let X0 be a nonnegative infinitely differentiable function on Rd such that suppX0 ⊂ See [32, Section VII.2.4] for the existence of such X0. Set Xk = X0(· − k). Then Xk is supported in the cube 2Qk = k + 2Q0, and all Xk's form a smooth resolution of the unit: (5.11) 1 = Xk∈Zd Xk(s), ∀ s ∈ Rd. Take a to be one of the atoms in (5.10) supported in Q. Since φ has compact support, i.e. there exists a constant C such that supp φ ⊂ CQ0, then φ ∗ a is supported in (C + 1)Q0. Thus, we get the decomposition φ ∗ a = NXk=1 bk with bk = Xk · (φ ∗ a), 36 R. Xia and X. Xiong where N is a positive integer depending only on the dimension d and C. For any β, γ ∈ Nd β ≤ γ if βj ≤ γj for every 1 ≤ j ≤ d. Then, by the Cauchy-Schwarz inequality, for any k, 0, denote 2 2 τ(cid:0)ZRd Dγ bk(s)2ds(cid:1) 1 2 .Xβ≤γ .Xβ≤γ ≤Xβ≤γ(cid:0)ZQZ2Qk Dβφ(s − t)2dsdt(cid:1) 1 τ(cid:0)Z2Qk Dβφ ∗ a(s) · Dγ−βXk(s)2ds(cid:1) 1 τ(cid:0)Z2Qk ZRd Dβφ(s − t)a(t)dt2ds(cid:1) 1 2 τ(cid:0)ZQ a(t)2dt(cid:1) 1 φ ∗ f =Xj with bj as desired. Furthermore,Pj µj . kfkhc 2 ≤ 1, . Q µjbj, 1. 1 definition of the tent space and Theorem 1.4 that Φε ∗ f ∈ T c . kfkhc kφ ∗ fk1 + kΦε ∗ fkT c 1 1 . 1 (Rd,M) and 2 · τ(cid:0)ZQ a(t)2dt(cid:1) 1 2 where the relevant constants depend only on γ, φ and X0. Thus, we have proved that φ ∗ f can be decomposed as follows: Now it remains to deal with the second term on the right hand side of (5.9). It follows from the 1 -atoms based on cubes with side length less than or equal to 1. For each By Lemma 5.1, we decompose Φε ∗ f as follows: (5.12) Φε ∗ f (s) = ∞Xj=1 where the eaj's are T c eaj(s, ε) based on Qj in (5.12), we set gj(s) =Z 1 (5.13) 0 λjeaj(s, ε) with dε ε ∞Xj=1 λj . kΦε ∗ fkT c 1 , We observe from the proof of Lemma 5.2 that gj is a bounded multiple of an hc 1-atom supported in 2Qj with vanishing mean. Moreover, gj is infinitely differentiable. Thus, gj satisfies (5.6) with relevant constant depending only on Φ. Combining (5.12) and (5.13), we obtain the decomposition Φε ∗eaj(s, ε) = πΦeaj(s), ∀s ∈ Rd. Z 1 0 Φε ∗ Φε ∗ f dε ε = λj gj, ∞Xj=1 1. The proof is complete. withP∞j=1 λj . kfkhc 5.2. Atomic decomposition for F α,c (Rd,M). For every l = (l1,··· , ld) ∈ Zd, µ ∈ N0, we define Qµ,l in Rd to be the cubes centered at 2−µl, and with side length 2−µ. For instance, Q0,0 = [− 1 2 )d is the unit cube centered at the origin. Let Dd be the collection of all the cubes Qµ,l defined above. We write (µ, l) ≤ (µ′, l′) if (Rd,M). Now we turn to the general space F α,c 2 , 1 (cid:3) 1 1 µ ≥ µ′ and Qµ,l ⊂ 2Qµ′,l′ . For a ∈ R, let a+ = max{a, 0} and [a] the largest integer less than or equal to a. Recall that 0 and J α is the Bessel potential γ1 = γ1 + ··· + γd for γ ∈ Nd of order α. Definition 5.4. Let α ∈ R, and let K and L be two integers such that K ≥ ([α] + 1)+ and L ≥ max{[−α],−1}. for s ∈ Rd, β ∈ Nd 1 ··· sβd 0, sβ = sβ1 d • supp b ⊂ 2Q0,k; (1) A function b ∈ L1(cid:0)M; Lc • τ(cid:0)RRd Dγ b(s)2ds(cid:1) 1 (2) Let Q = Qµ,l ∈ Dd, a function a ∈ L1(cid:0)M; Lc 2 ≤ 1, ∀γ ∈ Nd 2(Rd)(cid:1) is called an (α, 1)-atom if • supp a ⊂ 2Q; 0 , γ1 ≤ K. 2(Rd)(cid:1) is called an (α, Q)-subatom if Inhomogeneous Triebel-Lizorkin spaces 37 • τ(cid:0)RRd Dγ a(s)2ds(cid:1) 1 2 ≤ Q α • RRd sβa(s)ds = 0, ∀β ∈ Nd 2(Rd)(cid:1) is called an (α, Qk,m)-atom if (3) A function g ∈ L1(cid:0)M; Lc τ(cid:0)ZRd J αg(s)2ds(cid:1) 1 d − γ1 0 , β1 ≤ L. 2 . Qk,m− 1 d , ∀γ ∈ Nd 0 , γ1 ≤ K; and g = X(µ,l)≤(k,m) (5.14) 2 dµ,laµ,l, for some k ∈ N0 and m ∈ Zd, where the aµ,l's are (α, Qµ,l)-subatoms and the dµ,l's are complex numbers such that (cid:0) X(µ,l)≤(k,m) dµ,l2(cid:1) 1 2 ≤ Qk,m− 1 2 . Remark 5.5. If L < 0, the third assumption of an (α, Q)-subatom means that no moment cancellation is required. In the second assumption of an (α, 1)-atom b and that of an (α, Q)- subatom a, it is tacitly assumed that b and a have derivatives up to order K. For such a, we can define a norm by kak∗ = sup γ1≤K kDγakL1(cid:0)M;Lc 2(Rd)(cid:1). Then the convergence in (5.14) is understood in this norm, and we will see that the atom g in (5.14) belongs to F α,c 1 (Rd,M). Remark 5.6. In the classical case, the first size estimate in (5.14) is not necessary. words, if g =P(µ,l)≤(k,m) dµ,laµ,l with the subatoms aµ,l's and the complex numbers dµ,l's such that(cid:0)P(µ,l)≤(k,m) dµ,l2(cid:1) 1 2 , then g satisfies that estimate in (5.14) automatically. We refer the readers to [35] for more details. Unfortunately, in the current setting, we are not able to prove this estimate, so we just add it in (5.14) for safety. 2 ≤ Qk,m− 1 In other The following is our main result on the atomic decomposition of F α,c (Rd,M). The idea comes from [35, Theorem 3.2.3], but many techniques used are different from those of [35, Theorem 3.2.3] due to noncommutativity. Theorem 5.7. Let α ∈ R and K, L be two integers fixed as in Definition 5.4. Then any f ∈ F α,c 1 1 (Rd,M) can be represented as (5.15) f = (5.16) where the bj's are (α, 1)-atoms, the gj's are (α, Q)-atoms, and µj, λj are complex numbers with ∞Xj=1(cid:0)µjbj + λj gj(cid:1), ∞Xj=1 (µj + λj) < ∞. ∞Xj=1 (µj + λj) . kfkF α,c 1 . Moreover, the infimum of (5.16) with respect to all admissible representations is an equivalent norm in F α,c Proof. Step 1. First, we show that any f ∈ F α,c (Rd,M) admits the representation (5.15) and (Rd,M). 1 1 The proof of this part is similar to the proof of Theorem 5.3. Let κ be the Schwartz function that N ≥ max{L, α}, then Φ can be normalized as follows: defined in the proof of Theorem 5.3. We take bΦ = · Nbκ, where N is a positive even integer such Z ∞ bΦ(εξ)2 dε (J−αbΦ)(2−jξ)2 andP∞j=−∞ Since −α+N ≥ 0, bothP∞j=−∞ (J−αbΦ)(2−jξ)2 are rapidly decreasing ∞Xj=−∞ (J−αbΦ)(2−jξ)2 < ∞ and infinitely differentiable functions on Rd \ {0}. So we have (5.17) ∀ ξ ∈ Rd \ {0}. = 1, ε 0 38 and (5.18) with Applying the Paley-Wiener-Schwartz theorem, we get a compactly supported function Φ0 ∈ S such that Denote by Φε the Fourier inverse transform of Φ(ε·). For any f ∈ F α,c (5.19) 1 . (Rd,M), we have Let us deal with the two terms on the right hand side of (5.19) separately. The term Φ0∗f is easy to treat. If α ≥ 0, Proposition 3.3 ensures that F α,c Then we can repeat the first part of the proof of Theorem 5.3: for any f ∈ F α,c admits the decomposition 1 (Rd,M) ⊂ hc 1(Rd,M). (Rd,M), Φ0 ∗ f 1 R. Xia and X. Xiong (I−αbΦ)(2−jξ)2 < ∞. 0 bΦ(εξ)2 dε ∞Xj=−∞ bΦ0(ξ) = 1 −Z 1 f = Φ0 ∗ f +Z 1 Φε ∗ Φε ∗ f ε 0 . dε ε Φ0 ∗ f =Xj µjbj, Xj µj . kfkhc 1 . kfkF α,c 1 , µjbj, . Then J [α]Φ0 ∗ f =Xj Φ0 ∗ f =Xj µjJ−[α]bj. where the bj's, together with their derivatives Dγ bj's, satisfy (5.5) with some constants Cγ de- pending on γ. When K is fixed, we can normalize the bj's by maxγ1≤K Cγ, then the new bj's are (α, 1)-atoms. If α < 0, by Propositions 3.3 and 3.4, we have J [α]f ∈ F α−[α],c 1. Then J [α]Φ0 ∗ f admits the decomposition ⊂ hc 1 withPj µj . kJ [α]fkhc 1 . kfkF α,c 1 If −[α] is even, it is obvious that supp J−[α]bj ⊂ supp bj. Moreover, for any γ ∈ Nd γ1 ≤ K, we have 0 such that τ (ZRd Dγ J−[α]bj(s)2ds) 1 2 . Xγ′1≤K−2[α] τ (ZRd Dγ′bj(s)2ds) 1 2 ≤ CK . We normalize J−[α]bj by this constant CK depending on K, then we can make it an (α, 1)-atom. When −[α] is odd, it suffices to replace [α] in the above argument by [α] − 1, and then we get the desired decomposition. It follows from Step 2. Now we turn to the second term on the right hand side of (5.19). 1 (Rd,M) and Theorem 4.6 and the definition of the tent space that ε−αΦε ∗ f ∈ T c By Lemma 5.1, we have (5.20) 1 kε−αΦε ∗ fkT c ∞Xj=1 ε−αΦε ∗ f (s) = . kfkF α,c 1 . λj bj(s, ε), Φε ∗ f (s) = ∞Xj=1 λjaj(s, ε) where the bj's are T c εαbj(s, ε), we obtain 1 -atoms based on the cubes Qj's with Qj ≤ 1. Then, if we set aj(s, ε) = and (5.21) In particular, (5.22) For every aj, we set (5.23) supp aj ⊂ T (Qj) and τ(cid:16)ZT (Qj ) gj(s) = πΦ(aj)(s) =Z 1 0 ε (cid:17) 1 ε−2αaj(s, ε)2 dsdε 2 ≤ Qj− 1 2 . Φε ∗ aj(s, ε) dε ε . Then supp gj ⊂ 2Qj. We arrive at the decomposition = Φε ∗ Φε ∗ f dε ε Z 1 0 λj gj. ∞Xj=1 Now we show that every gj is an (α, Qkj ,mj )-atom. Firstly, for any Qj, there exist kj ∈ N0 and s ∈ Rd such that 2−kj−1 ≤ l(Qj) ≤ 2−kj and cQj = l(Qj)s. Take mj = [s] ∈ Zd, where [s] = ([s1],··· , [sd]). Then, we easily check that (5.24) Qj ⊂ 2Qkj ,mj , Qkj ,mj ∈ Dd. Next, by the argument similar to that in (5.3) and by (5.22), we have Inhomogeneous Triebel-Lizorkin spaces 39 ∞Xj=1 λj . kε−αΦε ∗ fkT c 1 . kfkF α,c 1 . If α ≤ 0, it is clear that τ(cid:0)ZRd I απΦ(aj)(s)2ds(cid:1) 1 τ(cid:0)ZRd J απΦ(aj)(s)2ds(cid:1) 1 τ(cid:0)ZRd J απΦ(aj)(s)2ds(cid:1) 1 If α > 0, we have 2 . 2 ≤ Qj− 1 2 . Qkj ,mj− 1 2 . Qkj ,mj− 1 2 . τ(cid:16)ZT (Qj ) ε (cid:17) 1 ε−2αaj(t, ε)2 dtdε 2 ≤ τ(cid:0)ZRd I απΦ(aj)(s)2ds(cid:1) 1 2 . Qj− 1 2 . τ(cid:0)ZRd πΦ(aj)(s)2ds(cid:1) 1 2 + τ(cid:0)ZRd I απΦ(aj)(s)2ds(cid:1) 1 . τ(cid:16)ZT (Qj ) aj(t, ε)2 dtdε ε (cid:17) 1 + Qj− 1 . τ(cid:16)ZT (Qj ) ε (cid:17) 1 ε−2αaj(t, ε)2 dtdε ≤ 2Qj− 1 2 . Qkj ,mj− 1 + Qj− 1 2 . 2 . 2 2 2 2 2 Then we get, for any α ∈ R, (5.25) τ(cid:0)ZRd J αgj(s)2ds(cid:1) 1 2 = τ(cid:0)ZRd J απΦ(aj )(s)2ds(cid:1) 1 2 . Qkj ,mj− 1 2 . Finally, we decompose the slice T (Qj) ∩ {2−µ−1 ≤ ε ≤ 2−µ} into (d + 1)-dimensional dyadic cubes whose projections on Rd belong to Dd, and with side length 2−µ, µ ∈ N0. Let bQ be one of those dyadic cubes with side length 2−µ and Q be its projection on Rd. Let a(s) =Z bQ Φε(s − t)aj(t, ε) dtdε . ε By the support assumption of Φ, it follows that Then supp a ⊂ 2Q, supp a ⊂ 2Qj ⊂ 4Qkj ,mj . ba(ξ) =Z 2−µ+1 2−µ bΦ(εξ)F(cid:0)aj(·, ε)1Q(cid:1)(ξ) dε ε . 40 R. Xia and X. Xiong τ(cid:0)Z a(s)2ds(cid:1) 1 Since DβbΦ(0) = 0 for any β1 ≤ N , we obtain Furthermore, by the Cauchy-Schwarz inequality, we have ∀β1 ≤ L. ZRd (−2πis)βa(s)ds = Dβba(0) = 0, 2 = τ(cid:16)Z5Q(cid:12)(cid:12)Z 2−µ+1 ZQ ε (cid:12)(cid:12)(cid:12) · τ(cid:16)Z 2−µ+1 2(cid:16)Z 2−µ+1 ZQ ε (cid:17) 1 . τ(cid:16)Z 2−µ+1 ZQ aj(s, ε)2 dsdε ε (cid:17) 1 Φε(s − t)aj(t, ε) ε−2d dtdε . Q dtdε 2−µ 2−µ 2−µ 2−µ 1 2 2 2 2 ds(cid:17) 1 ZQ aj(t, ε)2 dtdε ε (cid:17) 1 2 α d τ(cid:16)Z 2−µ+1 2−µ . Q Similarly, we have γ τ(cid:0)Z D a(s)2ds(cid:1) 1 2 ≤ C′γQ The above discussion gives ε (cid:17) 1 ε−2αaj(s, ε)2 dsdε 2 . ZQ d τ(cid:16)Z 2−µ+1 2−µ α d − γ1 ZQ ε (cid:17) 1 ε−2αaj(s, ε)2 dsdε 2 . (5.26) where each aj µ,l, dj µ,laj gj = X(µ,l)≤(kj ,mj ) γ1≤K{C′γ}τ(cid:16)Z 2−µ+1 ZQµ,l 2 ≤ Cτ(cid:16)ZT (Qj ) µ,l2(cid:1) 1 dj µ,l is an (α, Qµ,l)-subatom. The normalizing factor is given by ε (cid:17) 1 ε−2αaj(s, ε)2 dsdε ε (cid:17) 1 ε−2αaj(s, ε)2 dsdε (cid:0) X(µ,l)≤(kj ,mj) dj µ,l = max 2), we get 2−µ 2 2 By the elementary fact that ℓ2(L1(M)) ⊃ L1(M; ℓc (5.27) . ≤ CQkj ,mj− 1 2 , where C is independent of f . We may assume C = 1, otherwise, we can put C in (5.20) in the numbers λj , which does not change (5.21). In summary, (5.24), (5.25), (5.26) and (5.27) ensure that gj is an (α, Qkj ,mj )-atom. Step 3. Now we show the reverse assertion: if f is given by (5.15), then f ∈ F α,c 1 (Rd,M) and kfkF α,c 1 . (µj + λj). ∞Xj=1 To this end, we have to show that every (α, 1)-atom b and every (α, Q)-atom g belong to F α,c and 1 (Rd,M) Let b be an (α, 1)-atom in F α,c α ≤ 0, by Proposition 3.3, hc 3.4, we have 1 ⊂ F α,c 1 1 1 . 1 kbkF α,c (Rd,M). We observe that b is also an atom in hc and kgkF α,c . 1. 1 . Then, we have kbkF α,c 1 . kbkhc 1 1(Rd,M). For . 1. If α > 0, by Proposition Note that for any 1 ≤ i ≤ d, DK we have 1(Rd,M). Since α − K < 0, by Proposition 3.3, 1 ≈ kϕ0 ∗ bk1 + kbkF α,c i b is an atom in hc dXi=1 kDK i bkF α−K,c 1 . kbkF α,c 1 . kϕ0 ∗ bk1 + dXi=1 kDK i bkhc 1 . 1. Inhomogeneous Triebel-Lizorkin spaces 41 On the other hand, let g be an (α, Qk,m)-atom in the sense of Definition 5.4. We may use the discrete general characterization of F α,c 1 (Rd,M) given in Theorem 4.1, i.e. 1 kgkF α,c 22jαΦj ∗ g2) ∞Xj=0 1 ≈(cid:13)(cid:13)( j=0 andP∞j=k. When j ≥ k, by the support assumption of Φ, we 2(cid:13)(cid:13)1 . have supp Φj ∗ g ⊂ 5Qk,m. If α ≥ 0, by (5.18), (5.14) and the Plancherel formula (0.7), we obtain We splitP∞j=0 into two partsPk−1 τ(cid:0)Z5Qk,m ∞Xj=k 22jαΦj ∗ g(s)2ds(cid:1) 1 If α < 0, by (5.17), (5.14) and the Plancherel formula (0.7) again, we have τ(cid:0)Z5Qk,m ∞Xj=k 22jαΦj ∗ g(s)2ds(cid:1) 1 2 2 2 2 . 2 ≤ Qm,k− 1 2 = τ(cid:0)Z5Qk,m ∞Xj=k (I−αΦ)j ∗ I αg(s)2ds(cid:1) 1 ≤ τ(cid:0)ZRd ∞Xj=k (I−αbΦ)(2−jξ)2Iαbg(ξ)2dξ(cid:1) 1 . τ(cid:0)ZRd Iαbg(ξ)2dξ(cid:1) 1 2 = τ(cid:0)ZRd I αg(s)2ds(cid:1) 1 ≤ τ(cid:0)ZRd J αg(s)2ds(cid:1) 1 2 ≤ τ(cid:0)Z5Qk,m ∞Xj=k 22jαJ−αΦj ∗ J αg(s)2ds(cid:1) 1 ≤ τ(cid:0)ZRd ∞Xj=k (J−αbΦ)(2−jξ)2Jαbg(ξ)2dξ(cid:1) 1 . τ(cid:0)ZRd Jαbg(ξ)2dξ(cid:1) 1 2 = τ(cid:0)ZRd J αg(s)2ds(cid:1) 1 2(cid:13)(cid:13)1 . 1. eg = 2k(α−d)g(2−k·). ≤ Qm,k− 1 2 . 2 2 2 j=0 , we begin with a technical modification of g. Let It follows that 1 22jαΦj ∗ g2) ∞Xj=k (cid:13)(cid:13)( In order to estimate the sumPk−1 Then it is easy to see thateg is an (α, Q0,m)-atom. Moreover, we have which implies that Φj ∗ g = 2k(d−α)Φj−k ∗eg(2k·), 2(cid:13)(cid:13)1 ≤(cid:13)(cid:13)( 22jαΦj ∗eg2) −1Xj=−∞ 1 (5.28) k−1Xj=0 (cid:13)(cid:13)( 22jαΦj ∗ g2) where (Φ0)−k denotes the inverse Fourier transform of the function Φ(0)(2k·). In other words, we can assume, by translation, that the atom g is based on a cube Q with side length 1 and centered 2(cid:13)(cid:13)1 + 2−kαk(Φ0)−k ∗egk1, at the origin. Then, let us estimate the right hand side of (5.28) with g instead ofeg. By the triangle inequality, we have 1 −1Xj=−∞ (cid:13)(cid:13)( 22jαΦj ∗ g2) 2jατZRd Φj ∗ g(s)ds 1 2(cid:13)(cid:13)1 ≤ ≤ −1Xj=−∞ −1Xj=−∞ X(µ,l)≤(0,0) dµ,l 2jατZRd Φj ∗ aµ,l(s)ds. 42 R. Xia and X. Xiong M + α > 0 and L ≥ M − 1. By the moment cancellation of aµ,l, we have Φj ∗ aµ,l(s) Now we estimate 2jατRRd Φj ∗ aµ,l(s)ds for every (µ, l) ≤ (0, 0). Let M = [−α] + 1. Then = 2jdZ2Qµ,l(cid:2)Φ(2js − 2jt) − Φ(2js − 2j2−µl)(cid:3)aµ,l(t)dt = 2j(d+M) Xβ1=M Φj ∗ aµ,l(s)2 . Xβ1=M (1 − θ)M DβΦ(cid:0)2js − 2j(θt + (1 − θ)2−µl)(cid:1)aµ,l(t)dθ dt. (1 − θ)2MDβΦ(cid:0)2js − 2j(θt + (1 − θ)2−µl)(cid:1)2dθdt Z2Qµ,l (2−µl − t)βZ 1 22j(d+M)Z2Qµ,lZ 1 It follows that M + 1 β! 0 0 ·Z2Qµ,l t − 2−µl2Maµ,l(t)2dt. 2 . If Φj ∗ aµ,l(s) 6= 0, then we have 2js − 2jt ≤ 1 for some t ∈ 2Qµ,l. Hence, Φj ∗ aµ,l(s) = 0 if s − 2−µl > 3 · 2−j−1√d. Consequently, 2jατZRd Φj ∗ aµ,l(s)ds −1Xj=−∞ 2j(d+M+α)τ(cid:0)Z2Qµ,l t − 2−µl2Maµ,l(t)2dt(cid:1) 1 −1Xj=−∞ · Xβ1=MZs−2−µl≤3·2−j−1√d(cid:16)Z2Qµ,lZ 1 −1Xj=−∞ −1Xj=−∞ = 2−µ(α+M) −1Xj=−∞ (1 − θ)2MDβΦ(2js − 2j(θt + (1 − θ)2−µl))2dθdt(cid:17) 1 2 τ(cid:0)Z2Qµ,l aµ,l(t)2dt(cid:1) 1 2Zs−2−µl≤3·2−j−1√d 2j(d+M+α) · 2−jd · 2−µ(α+M)Qµ,l 2j(d+M+α) · 2−µMQµ,l 2 . 2−µ(α+M)Qµ,l 2j(M+α)Qµ,l 1 2 . ds ds . . 1 2 0 1 1 2 Similarly, we also have and Therefore, kgkF α,c 1 . 1. The proof is complete. Thus, by the Cauchy-Schwarz inequality, we get 2 ≤ 2−µ(α+M)Qµ,l 1 1 22jαΦj ∗ g2) 2(cid:13)(cid:13)1 ≤ −1Xj=−∞ (cid:13)(cid:13)( 2−kατZRd (Φ0)−k ∗ aµ,l(s)ds . 2−k(M+α)2−µ(α+M)Qµ,l 22jατZRd Φj ∗ g(s)ds −1Xj=−∞ ∞Xµ=0 2−µ(α+M)(cid:0)Xl 2(cid:0)Xl dµ,l2(cid:1) 1 ∞Xµ=0 2−kαk(Φ0)−k ∗ gk1 . 2−µ(α+M) < ∞. 2−µ(α+M) < ∞, . . ∞Xµ=0 2 Qµ,l(cid:1) 1 1 2 . (cid:3) Inhomogeneous Triebel-Lizorkin spaces 43 We close this section by a very useful result of pointwise multipliers, which can be deduced from the above atomic decomposition. Let k ∈ N and Lk(Rd,M) be the collection of all M-valued functions on Rd such that Dγ h ∈ L∞(N ) for all γ with 0 ≤ γ1 ≤ k. Corollary 5.8. Let α ∈ R and let k ∈ N be sufficiently large and h ∈ Lk(Rd,M). Then the map f 7→ hf is bounded on F α,c Proof. First, consider the case α > 0. We apply the atomic decomposition in Theorem 5.7 with K = k and L = −1. In this case, no moment cancellation of subatoms is required. We can easily check that, multiplying every (sub)atom in Definition 5.4 by h, we get another (sub)atom. Moreover, (Rd,M) 1 (5.29) khfkF α,c 1 ≤ Xγ≤k sup s∈Rd kDγh(s)kM · kfkF α,c 1 . The case α ≤ 0 can be deduced by induction. Assume that (5.29) is true for α > N ∈ Z. Let and can be represented as f = J 2g = (1 − (2π)−2∆)g with g ∈ F α+2,c α > N − 1. Any f ∈ F α,c kfkF α,c . Since hf = (1 − (2π)−2∆)(hg) + ((2π)−2∆h)g + (2π)−2∇h · ∇g, 1 ≈ kgkF α+2,c 1 1 1 we deduce khfkF α,c 1 . k(1 − (2π)−2∆)(hg)kF α,c 1 + k(∆h)gkF α,c 1 + (5.30) . kgkF α+2,c 1 + k(∆h)gkF α+2,c 1 + dXi=1 k∂ih · ∂igkF α+1,c 1 . dXi=1 k∂ih · ∂igkF α,c 1 If k ∈ N is sufficiently large, we have . kgkF α+2,c Continuing the estimate in (5.30), we obtain k(∆h)gkF α+2,c 1 1 khfkF α,c 1 . kgkF α+2,c 1 +Xi which completes the induction procedure. , k∂ih · ∂igkF α+1,c 1 . k∂igkF α+1,c 1 . k∂igkF α+1,c 1 . kgkF α+2,c 1 . kfkF α,c 1 , (cid:3) Acknowledgements. The authors are greatly indebted to Professor Quanhua Xu for having suggested to them the subject of this paper, for many helpful discussions and very careful reading of this paper. The authors are partially supported by the the National Natural Science Foundation of China (grant no. 11301401). References [1] T. Bekjan, Z. Chen, M. Perrin and Z. Yin. Atomic decomposition and interpolation for Hardy spaces of non- commutative martingales. J. Funct. Anal. 258 (2010), no. 7, 2483-2505. [2] J. Bergh and J. Lofstrom. Interpolation Spaces: An Introduction. Springer, Berlin, 1976. [3] R. Coifman, Y. Meyer and E. M. Stein. Some new function spaces and their applications to Harmonic analysis. J. Funct. Anal. 62 (1985), no. 2, 304-335. [4] M. Frazier and B. Jawerth. The ϕ-transform and applications to distribution spaces. In Function spaces and applications, M. Cwikel et al. eds., Springer Lect. Notes in Math. 1302 (1988), 223-246. [5] M. Frazier and B. Jawerth. A discrete transform and decompositions of distribution spaces. J. Funct. Anal. 93 (1990), no. 1, 34-170. [6] M. Frazier, R. Torres and G. Weiss. The boundedness of Caldern-Zygmund operators on the spaces F α,q p . Rev. Mat. Iberoam. Vol. 4 (1988), 41-72. [7] J. Garc´ıa-Cuerva and J.L. Rubio de Francia. Weighted norm inequalities and related topics. North-Holland Mathematics Studies, 116. Notas de Matemtica, 104. North-Holland Publishing Co., Amsterdam, 1985. x+604 pp. [8] D. Goldberg. A local version of real Hardy spaces. Duke Math. J. 46 (1979), no. 1, 27-42. [9] L. Grafakos. Classical Fourier analysis. Second Edition. Springer-Verlag, New York, 2008. [10] G. Hong, L. D. L´opez-Snchez, J. M. Martell, and J. Parcet. Caldern-Zygmund operator associated to matrix- valued kernels. Int. Math. Res. Not. 2014, no. 5, 1221-1252. [11] G. Hong and T. Mei. John-Nirenberg inequality and atomic decomposition for noncommutative martingales. J. Funct. Anal. 263 (2012), no. 4, 1064-1097. 44 R. Xia and X. Xiong [12] Y. Jiao, F. Sukochev, D. Zanin and D. Zhou. Johnson-Schechtman inequalities for noncommutative martingales. J. Funct. Anal. 272 (2017), no. 3, 976-1016. [13] M. Junge. Doob's inequality for non-commutative martingales. J. Reine Angew. Math. 549 (2002), 149-190. [14] M. Junge, C. Le Merdy and Q. Xu. H∞-functional calculus and square functions on noncommutative Lp-spaces. Ast´erisque No. 305 (2006), vi+138 pp. [15] M. Junge and Q. Xu. Non-commutative Burkholder/Rosenthal Inequalities, I and II. Ann. Prob. 31 (2003), no. 2, 948-995 and Israel J. Math. 167 (2008), 227-282. [16] M. Junge and Q. Xu. Noncommutative maximal ergodic theorems. J. Amer. Math. Soc. 20 (2007), 385-439. [17] M. Junge and T. Mei. Noncommutative Riesz transforms - a probabilistic approach. Amer. J. Math. 132 (2010), 611-681. [18] T. Mei. Operator-valued Hardy Spaces. Mem. Amer. Math. Soc. 188 (2007), no. 881, vi+64 pp. [19] T. Mei. Tent spaces associated with semigroups of operators. J. Funct. Anal. 255 (2008), no. 12, 3356-3406. [20] M. Musat. Interpolation between non-commutative BMO and non-commutative Lp-spaces. J. Funct. Anal. 202 (2003), no. 1, 195-225. [21] Yu. V. Netrusov. Embedding theorems for Lizorkin-Triebel spaces. (Russian) Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 159 (1987), 103-112. translation in J. Soviet Math. 47 (1987), no. 6, 2896-2903. [22] J. Parcet. Pseudo-localization of singular integrals and noncommutative Calder´on-Zygmund theory. J. Funct. Anal. 256 (2009), no. 2, 509-593. [23] J. Parcet and N. Randrianantoanina. Gundy's decomposition for non-commutative martingales and applications. Proc. London Math. Soc. 93 (2006), 227-252. [24] G. Pisier. Noncommutative vector valued Lp spaces and completely p-summing maps. Ast´erisque. 247 (1998), vi+131 pp. [25] G. Pisier and Q. Xu. Non-commutative Martingale Inequalities. Comm. Math. Phys. 189 (1997), 667-698. [26] G. Pisier and Q. Xu. Noncommutative Lp-spaces. Handbook of the geometry of Banach spaces, Vol. 2, 1459- 1517, North-Holland, Amsterdam, 2003. [27] N. Randrianantoanina. Noncommutative martingale transforms. J. Funct. Anal. 194 (2002), 181-212. [28] N. Randrianantoanina. Conditional square functions for noncommutative martingales. Ann. Prob. 35 (2007), 1039-1070. [29] N. Randrianantoanina and L. Wu. Noncommutative Burkholder/Rosenthal inequalities associated with convex functions. Ann. Inst. Henri Poincar´e Probab. Stat. 53 (2017), no. 4, 1575-1605. [30] N. Randrianantoanina, L. Wu and Q. Xu. Noncommutative Davis type decompositions and applications. arXiv: 1712.01374. [31] E. M. Stein. Singular integrals and differentiability properties of functions. Princeton Mathematical Series, No. 30 Princeton University Press, Princeton, N.J. 1970. [32] E. M. Stein. Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals. Princeton Math- ematical Series, 43. Monographs in Harmonic Analysis, III. Princeton University Press, Princeton, NJ, 1993. [33] E. M. Stein and G. Weiss. Introduction to Fourier analysis on Euclidean spaces. Princeton Mathematical Series, No. 32. Princeton University Press, Princeton, N.J., 1971. [34] H. Triebel. Theory of function spaces. Modern Birkhuser Classics. Birkhuser/Springer Basel AG, Basel, 2010. [35] H. Triebel. Theory of function spaces. II. Monographs in Mathematics, 84. Birkhuser Verlag, Basel, 1992. [36] R. Xia and X. Xiong. Operator-valued local Hardy spaces Arxiv: 1803.10321. [37] R. Xia and X. Xiong. Pseudo-differential operators with operator-valued kernel. Preprint. [38] R. Xia, X. Xiong and Q. Xu. Characterizations of operator-valued Hardy spaces and applications to harmonic analysis on quantum tori. Adv. Math. 291 (2016), 183-227. [39] X. Xiong, Q. Xu and Z. Yin. Function spaces on quantum tori. C. R. Math. Acad. Sci. Paris 353 (2015), no. 8, 729-734. [40] X. Xiong, Q. Xu and Z. Yin. Sobolev, Besov and Triebel-Lizorkin spaces on quantum tori. Mem. Amer. Math. Soc. 252 (2018), no. 1203, vi+118 pp. [41] Q. Xu. Noncommutative Lp-spaces and martingale inequalities. Book manuscript, 2007. Laboratoire de Math´ematiques, Universit´e de Franche-Comt´e, 25030 Besanc¸on Cedex, France, and Instituto de Ciencias Matem´aticas, 28049 Madrid, Spain E-mail address: [email protected] Department of Mathematics and Statistics, University of Saskatchewan, Saskatoon, Saskatchewan, S7N 5E6, Canada E-mail address: [email protected]
1807.08371	1	1807	2018-07-22T21:31:07	Column extreme multipliers of the Free Hardy space	[ "math.OA", "math.FA" ]	The full Fock space over $\mathbb C ^d$ can be identified with the free Hardy space, $H^2 (\mathbb B ^d _\mathbb N)$ - the unique non-commutative reproducing kernel Hilbert space corresponding to a non-commutative Szeg\"{o} kernel on the non-commutative, multi-variable open unit ball $\mathbb B ^d _\mathbb N := \bigsqcup _{n=1} ^\infty \left( \mathbb C^{n\times n} \otimes \mathbb C ^d \right) _1$. Elements of this space are free or non-commutative functions on $\mathbb B ^d _\mathbb N$. Under this identification, the full Fock space is the canonical non-commutative and several-variable analogue of the classical Hardy space of the disk, and many classical function theory results have faithful extensions to this setting. In particular to each contractive (free) multiplier $B$ of the free Hardy space, we associate a Hilbert space $\mathcal H(B)$ analogous to the deBranges-Rovnyak spaces in the unit disk, and consider the ways in which various properties of the free function $B$ are reflected in the Hilbert space $\mathcal H(B)$ and the operators which act on it. In the classical setting, the $\mathcal H(b)$ spaces of analytic functions on the disk display strikingly different behavior depending on whether or not the function $b$ is an extreme point in the unit ball of $H^\infty(\mathbb D)$. We show that such a dichotomy persists in the free case, where the split depends on whtether or not $B$ is what we call {\it column extreme}.	math.OA	math	COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE MICHAEL T. JURY AND ROBERT T.W. MARTIN Abstract. The full Fock space over Cd can be identified with the free Hardy space, H 2(Bd N) - the unique non-commutative reproducing kernel Hilbert space corresponding to a non-commutative Szego kernel on the non-commutative, multi-variable open unit ball Bd n=1 (cid:0)Cn×n ⊗ Cd(cid:1)1. N := F∞ Elements of this space are free or non-commutative functions on Bd N. Under this identification, the full Fock space is the canonical non-commutative and several-variable analogue of the classical Hardy space of the disk, and many classical function theory results have faithful extensions to this setting. In particular to each contractive (free) multiplier B of the free Hardy space, we associate a Hilbert space H(B) analogous to the deBranges-Rovnyak spaces in the unit disk, and consider the ways in which various properties of the free function B are reflected in the Hilbert space H(B) and the operators which act on it. In the classical setting, the H(b) spaces of analytic functions on the disk display strikingly different behavior depending on whether or not the function b is an extreme point in the unit ball of H∞(D). We show that such a dichotomy persists in the free case, where the split depends on whtether or not B is what we call column extreme. 1. Introduction The classical Hardy space, H 2(D), can be defined as the Hilbert space of all analytic functions on D whose Taylor series at 0 have square summable coefficients (and with inner product equal to the ℓ2 inner product of these Taylor coefficients). Equivalently, H 2(D) = H(k), is the unique reproducing kernel Hilbert space (RKHS) of functions on D corresponding to the positive sesqui-analytic kernel function k : D × D → C: k(z, w) := 1 1 − zw∗ ; z, w ∈ D, the Szego kernel. The operator of multiplication by z on H 2(D) is called the shift, and it is easily seen to be isomorphic to the unilateral shift on ℓ2(N0), where N0 denotes the non-negative integers. Proofs of many deep results in classical Hardy space theory ultimately appeal to the fact that S is the universal cyclic pure isometry (recall the Wold decomposition says that any isometry is unitarily equivalent to a direct sum of shifts and a unitary operator). From the viewpoint of reproducing kernel theory and operator theory, the canonical (commutative) multi-variable analogue of the Hardy space is then the Drury-Arveson space, H 2 d := H(k), where now k : Bd × Bd → C is: 1 − zw∗ ; the multi-variable Szego kernel, and zw∗ := z1w∗ d = (w, z)Cd . (Here, Bd := (Cd)1, the multi-variable open unit ball.) The appropriate analogue of the shift in this setting 1 + ...zdw∗ k(z, w) := z, w ∈ Bd, 1 Date: July 24, 2018. 1 2 MICHAEL T. JURY AND ROBERT T.W. MARTIN is the Arveson d−shift, S = (S1, ..., Sd) : H 2 d , (Sjh)(z) := zj h(z); z = (z1, ..., zd) ∈ Bd. This is a (row) partial isometry (from d copies of H 2 d into itself), but no longer an isometry, and this defect is the source of several differences between the single and several-variable theories. Faithful analogues of classical Hardy space results typically seem to exist, but often new (and often more complicated) proof techniques and approaches are required [23]. d ⊗ Cd → H 2 An alternative approach to extending Hardy space theory from one to several variables would be to seek analogues of Hardy space results for a several-variable shift. Namely, the natural multi-variable analogue of ℓ2(N0) is ℓ2(Fd), where Fd is the free monoid (unital semi-group) of all words in the d letters {1, ..., d}, and with unit equal to the empty word, ∅, containing no letters. This monoid can be identified with a simple directed tree starting at single node and with d branches at each node (clearly F1 ≃ N0). There is a natural d−tuple of shifts, L := (L1, ..., Ld) on ℓ2(Fd) which are defined by Lkeα := ekα, where {eα}α∈Fd is the canonical orthonormal basis of ℓ2(Fd). It is easy to see that each Lk is a pure isometry and the Lk have pairwise orthogonal ranges L∗ In particular, the row L = (L1, ..., Ld) : ℓ2(Fd) ⊗ Cd → ℓ2(Fd) is an isometry from d copies of ℓ2(Fd) into itself which we call the left free shift. The Popescu-Wold decomposition for row isometries shows that L has the same universal property as the shift S: any row isometry (an isometry from d copies of a Hilbert space into itself) is isomorphic to the direct sum of several copies of L and a row unitary (an onto row isometry). kLj = δk,jI. The left free shifts Lk are of course non-commuting, and it would appear that one loses the analytic function theory interpretation of the shift as acting as multiplication by the independent variable on a space of analytic functions. Surprisingly, this is not the case: the fields of non-commutative function theory [13, 1, 18, 20, 19], and the recently developed theory of non-commutative reproducing kernel Hilbert spaces (NC-RKHS) [3] have shown that ℓ2(Fd) is canonically isomorphic to the free Hardy space, H 2(Bd N) of non-commutative or free holomorphic functions on a certain non-commutative multi-variable open unit ball, Bd N (we will introduce these objects and this theory in an upcoming subsection). The Drury-Arveson space H 2 d can be identified with a subspace of H 2(Bd N) which is co-invariant and cyclic for both the left and right free shifts: H 2 d ≃ _z∈Bd Kz ⊆ H 2(Bd N), the span of all the kernel vectors at level one. This subspace is the orthogonal complement of the range of both a right inner and a left inner free multiplier. For example, if d = 2, H 2 d ≃ Ran(cid:18) 1 √2 (L1L2 − L2L1)(cid:19)⊥ , and this shows that the theory of H 2 d should be closer in analogy to that of the theory of model subspaces of H 2(D). In particular, commutative Drury-Arveson space analogues of all of the results of this paper (and those of [8]) can be easily obtained by compression. In recent work, we have extended Hardy space results including the concept of Aleksandrov- Clark measure, the theory of Clark's unitary perturbations, and equivalent characteriza- tions of extreme points from one to several variables. In particular, the reference [8] extends the theory of Clark measures and Clark peturbations to the non-commutative setting of the full Fock space over Cd (which can be identified with ℓ2(Fd)) using the theory of free formal reproducing kernel Hilbert spaces [4]. The goal of this paper is to develop non-commuative analogues of our recent results on extreme points of the closed unit ball of the multiplier algebra of Drury-Arveson space [11, 12]. We will also extend and re-cast the main results of [8] in the modern language COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 3 of NC-RKHS. In particular we give a number of equivalent characterizations of so-called column extreme multipliers of the free Hardy space. 2. Preliminaries All Hilbert space inner products will be conjugate linear in their first argument. If X is a Banach space, (X)1 and [X]1 denote the open and closed unit balls of X, respectively. 2.1. The full Fock space. Recall that the full Fock space over Cd, F 2 of all tensor powers of Cd: d , is the direct sum F 2 d := C ⊕(cid:16)Cd ⊗ Cd(cid:17) ⊕(cid:16)Cd ⊗ Cd ⊗ Cd(cid:17) ⊕ ··· = ∞Mk=0(cid:16)Cd(cid:17)k·⊗ . Fix an orthonormal basis {e1, ..., ed} of Cd. The left creation operators L1, ..., Ld are the operators which act as tensoring on the left by these basis vectors: and similarly the right creation operators Rk; 1 ≤ k ≤ d are defined by tensoring on the right Lkf := ek ⊗ f ; f ∈ F 2 d , Rkf := f ⊗ ek. d ⊗ Cd into F 2 The left and right free shifts are the row operators L := (L1, ..., Ld) and R := (R1, ..., Rd) d . Both L, R are in fact row isometries: L∗L = IF 2 ⊗ Id = R∗R. which map F 2 It follows that the component shifts are also isometries with pairwise orthogonal ranges. The orthogonal complement of the range of L or R is the vacuum vector 1 which spans the the subspace C =: (Cd)0·⊗ ⊂ F 2 d is then {eα}α∈Fd where eα = Lα1 = Rα1 and Fd is the free unital semigroup or monoid on d letters. Here, if α = i1 ·· · in ∈ Fd, we use the standard notation Lα = Li1 Li2 ··· Lin . Recall here that the free monoid, Fd, on d ∈ N letters, is the multiplicative semigroup of all finite products or words in the d letters {1, ..., d}. That is, given words α := i1...in, β := j1...jm, ik, jl ∈ {1, ..., d}; 1 ≤ k ≤ n, 1 ≤ l ≤ m, their product αβ is defined by concatenation: d . A canonical orthonormal basis for F 2 αβ = i1...inj1...jm, and the unit is the empty word, ∅, containing no letters. Given α = i1 ··· in, we use the standard notation α = n for the length of the word α. The transpose map † : Fd → Fd, defined by i1 ··· id = α 7→ α† := id ··· i1, is an involution. Define L∞ d := Alg(I, L)−W OT , R∞ d right) free analytic Toeplitz algebra (W OT denotes weak operator topology). The transpose unitary, U† : F 2 d , and it is easy to verify that d , defined by eα 7→ eα† is a unitary involution of F 2 := Alg(I, R)−W OT , the left (resp. d → F 2 U†LkU ∗ † = Rk, so that adjunction by U† implements a unitary isomorphism between L∞ d and R∞ d . 2.2. The free Hardy space. It will be convenient to view F 2 d as a non-commutative reproducing kernel Hilbert space (NC-RKHS) [3] of freely non-commutative (holomorphic) functions on the non-commutative open unit ball [13]: Bd N := Bd n; ∞Gn=1 Bd n :=(cid:16)Cn×n ⊗ Cd(cid:17)1 . 4 MICHAEL T. JURY AND ROBERT T.W. MARTIN Elements of Bd vector space V , n are viewed as strict row contractions on Cn. Recall that for any complex The NC unit ball Bd under direct sums, and one writes: N is an example of a NC set: A set Ω ⊆ Vnc is an NC set if it is closed Vn := V ⊗ Cn×n =: V n×n. Vnc :=G Vn; Ω =:G Ωn; f : Ωn → Cn×n; A function f : Ω → Cnc is called a NC or free function if: Ωn := Ω ∩ Vn. f respects the grading, and if X ∈ Cn×m, Z ∈ Xn, W ∈ Xm obey ZX = XW , then, f respects intertwinings. As shown in [3], F 2 variable NC unit ball Bd to the NC-Szego kernel: K : Bd f (Z)X = Xf (W ); d = H 2(Bd N, i.e. H 2(Bd N × Bd Z αP (W ∗)α† K(Z, W )[P ] := Xα∈Fd N) can be viewed as the free Hardy space of the multi- N) = Hnc(K) is the unique NC-RKHS corresponding N → L(Cnc) defined by: ; Z ∈ Bd n, W ∈ Bd m, P ∈ Cn×m. See [3] for the full definition and theory of NC kernels. respects the grading and intertwinings in both arguments [3, Section 2.3]. In particular, any NC kernel One can show that elements of H 2(Bd automatically) holomorphic free functions on Bd is Fr´echet and Gateaux differentiable at any point Z ∈ Bd series expansion (Taylor-Taylor series) about any point. N) := Hnc(K) are locally bounded (and hence N [13, Chapter 7]. That is, any f ∈ H 2(Bd N) N and f has a convergent power (Generally any) Hnc(K) is formally defined as the Hilbert space completion of the linear span: _Z∈Bd K{W, x, u}(Z)y := K(Z, W )[yu∗]x; K{Z, y, v}, n, y,v∈Cn Completion is with respect to the inner product: where the K{W, x, u} are the free functions on Bd N, K{W, x, u} : Bd n → Cn×n, defined by: W ∈ Bd m, Z ∈ Bd n; u, x ∈ Cm, y ∈ Cn. hK{Z, y, v}, K{W, x, u}i := (y, K(Z, W )[vu∗]x)Cn ; Z ∈ Bd n, v, y ∈ Cn; W ∈ Bd m, u, x ∈ Cm. These point evaluation vectors have a familiar reproducing property: K(Z, y, v) is the unique vector in H 2(Bd N) such that for any f ∈ H 2(Bd N), (2.1) hK{Z, y, v}, fi = (y, f (Z)v)Cn . For any Z ∈ Bd N)) as follows: Any A ∈ Cn×n can be written as a linear combination of the rank one outer products n one can also define a natural kernel map KZ ∈ L(Cn×n, H 2(Bd yv∗ = y1 ... yn (cid:2)v1, ··· , vn(cid:3) ; y ∈ Cn, v∗ ∈ (Cn)∗. Then we define KZ on rank one matrices yv∗ by the formula KZ(yv∗) := K{Z, y, v} ∈ H 2(Bd N). (2.2) Let us check that KZ is well defined: the vectors y and v determining a rank one matrix yv∗ are unique up to the scaling y → λy, v → λ v where λ is any nonzero complex number. From the reproducing formula (2.1), it is evident that the vector K{Z, y, v} is invariant −1 COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 5 under such a scaling, and so the formula (2.2) is unambiguous. If we view Cn×n as a Hilbert space equipped with the normalized trace inner product, then KZ : Cn×n → H 2(Bd N) extends to a bounded linear map, and its Hilbert space adjoint is the point evaluation map at Z: The free Hardy space and the full Fock space are canonically isomorphic: Define U : F 2 d → H 2(Bd N) by: K∗ Z F = F (Z) ∈ Cn×n. x := Xα∈Fd fx(Z) := Xα∈Fd xαLα1 Z αxα; U 7→ fx ∈ H 2(Bd N), Z ∈ Bd N. The inverse, U −1, acts on kernel vectors as: (2.3) K{Z, y, v} U −1 7→ x[Z, y, v] := Xα∈Fd hZ αv, yi Lα1 ∈ F 2 d . 2.3. Left and Right free multipliers. As in the classical setting, given a NC-RKHS Hnc(K) on an NC set Ω (e.g. Bd N), it is natural to consider the left and right multiplier algebras MultL(Hnc(K)), MultR(Hnc(K)) N) := MultL(H 2(Bd of NC functions on Ω which left or (resp.) right multiply Hnc(K) into itself. Namely, N is said to be a left free multiplier if, for any f ∈ H 2(Bd a free function F on Bd N), N). Similarly, G is called a right free multiplier if f G ∈ H 2(Bd F f ∈ H 2(Bd N) for all f ∈ H 2(Bd N). As in the classical setting, the left and right free multiplier algebras, H ∞ L (Bd N) are weak operator toplogy (WOT)-closed unital operator algebras. Moreover, adjunction by the canonical unitary U defines a unitary ∗−isomorphism of the left and right free analytic Toeplitz algebras L∞ d onto these N). As in the classical setting of H ∞(D), the left and right free multiplier algebras of H 2(Bd multiplier norm of any F ∈ H ∞ N) can be computed as the supremum norm on the NC unit ball: N)), H ∞ d , R∞ R (Bd L (Bd kFk := sup Z∈Bd N kF (Z)k. The left and right Schur classes, Ld, Rd are then defined as the closed unit balls of these left and right multiplier algebras (equivalently as the closed unit balls of L∞ d , R∞ d ). Observe that if F is a left free multiplier then, hK{Z, y, v}, F fi = hy, F (Z)f (Z)vi = hK{Z, F (Z)∗y, v}, fi , so that (2.4) (2.5) and similarly, if G is a right free multiplier, (M L F )∗K{Z, y, v} = K{Z, F (Z)∗y, v}, (M R G )∗K{Z, y, v} = K{Z, y, G(Z)v}. Alternatively, using the kernel maps KZ , we can write: (M L F )∗KZ(yv) = KW (F (Z)∗yv∗), and (M R G )∗KZ = KZ(yv∗G(Z)∗). One can check that if, e.g., right multiplication by G(Z) is a right free multiplier then ((M R G )∗KZ )∗((M R G )∗KW ) = K(Z, W )[G(Z) · G(W )∗]. In particular, free holomorphic F (Z), G(Z) belong to the left or right Schur classes if and only if K F (Z, W )[·] := K(Z, W ) − F (Z)K(Z, W )[·]F (W )∗ 6 or MICHAEL T. JURY AND ROBERT T.W. MARTIN K G(Z, W )[·] := K(Z, W ) − K(Z, W )[G(Z)[·]G(W )∗] are CPNC kernels, respectively. These NC kernels are called the left or right free deBranges- Rovnyak kernels of F, G (resp.) and in this case the corresponding NC-RKHS Hnc(K F ) =: H L(F ), Hnc(K G) =: H R(G) are the left and right free deBranges-Rovnyak spaces of F, G. 2.4. Left vs. Right. Any element F ∈ L∞ series: FαLα; F ∼ F (L) := Xα∈Fd Fα := hLα1, F 1i . d can be identified with the left free Fourier That is, F is identified with its symbol : f := F 1 = Xα∈Fd FαLα1 ∈ F 2 d , and we say that F (L) = M L f acts as left multiplication by f = F 1. In general the free Fourier series does not converge in SOT or WOT, but the Ces`aro sums converge in the strong operator toplogy (SOT) to F [6]. Similarly, in the operator valued setting, any F ∈ L∞ f , where the symbol, f ∈ F 2 d ⊗ L(H, J ) is defined by M L d ⊗ L(H,J ) is written F = F (L) = f := F (1 ⊗ IH) =Xα Lα1 ⊗ Fα; Fα ∈ L(H,J ). In this case the operator-valued free holomorphic function F (Z) takes values in (C)nc ⊗ L(H,J ). We can also identify any G ∈ R∞ d with its symbol: g := G1 = Xα∈Fd GαLα1, then we can view G as right multiplication by g(Z), Alternatively, we can write G = M R g(Z). g = Xα∈Fd Gα† Rα1, so that G = M R g(Z) = g†(R), where g†(Z) := Xα∈Fd Gα† Z α. That is, if G ∈ R∞ G(Z), then M R sums converge SOT to G). d acts as right multiplication by the free NC holomorphic function G(Z) is identified with the right free Fourier series G†(R) (whose Ces`aro Remark 2.5. In the right operator-valued setting, suppose that G(R) := g(R) ⊗ X ∈ R∞ d ⊗ L(H, J ) and F := f (R) ⊗ Y ∈ R∞ d , X ∈ L(H,J ), and Y ∈ L(J ,K). If H = F G, then observe that d ⊗ L(J ,K) with f, g ∈ R∞ H †(Z) = g†(Z)f †(Z) ⊗ Y X. This extends to a 'right product' for arbitrary operator-valued free holomorphic functions on Bd N, H(Z) = F (Z)•R G(Z). In the scalar-valued setting this simply reduces to F (Z)•R G(Z) = G(Z)F (Z). COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 7 2.6. Operator-valued free multipliers. It will also be convenient to consider operator- valued (left and right) free multipliers between vector-valued free Hardy spaces. Namely, if H is an auxiliary Hilbert space, one can consider the NC-RKHS H 2(Bd N )⊗H of H-valued NC functions on Bd N. This NC-RKHS has the operator-valued CPNC kernel: K(Z, W ) ⊗ IH, and is spanned by the elements with inner product defined by K{Z, y, v}h := K{Z, y, v} ⊗ h, h ∈ H, n, W ∈ Bd hK{Z, y, v}h, K{W, x, u}gi := (y, K(Z, W )[vu∗]x)Cn · hh, giH , m, v, y ∈ Cn and u, x ∈ Cm. We will write H ∞ L (Bd for h, g ∈ H, Z ∈ Bd N)⊗ L(H, J ) in place of MultL(H 2(Bd N)⊗J ), the W OT−closed left multiplier space between these vector-valued free Hardy spaces. (That is, we write H ∞ N) ⊗ L(H, J ) in place of the weak operator topology closure of this algebraic tensor product). The operator-valued Schur classes, Ld(H,J ), Rd(H,J ) are then the closed unit balls of these operator-valued left and right (resp.) multiplier spaces. N)⊗H, H 2(Bd L (Bd If A ∈ Ld(H,J ), consider the operator-valued CPNC kernel: K A(Z, W )[·] = K(Z, W )[·] ⊗ IJ − A(Z)(K(Z, W )[·] ⊗ IH)A(W )∗. Here, K is the free Szego kernel. This is the left free deBranges-Rovnyak CPNC ker- nel of A, and the corresponding NC-RKHS, Hnc(K A) =: H L(A) is called the left free deBranges-Rovnyak space of A. Namely, H L(A) is the closed linear span of vectors of the form K A{Z, y, v}g whose inner product is defined by: DK A{Z, y, v}g, K A{W, x, u}fEH L(A) := (y ⊗ g, (K(Z, W )[vu∗] ⊗ IJ − A(Z)(K(Z, W )[vu∗] ⊗ IH)A(W )∗) x ⊗ f )Cn⊗J . Z : Cn×n ⊗ J → H L(A) for the On the other hand, if B ∈ Rd(H, J ), then the right free deBranges-Ronvyak space In this vector-valued setting, for Z ∈ Bd Z (yv∗ ⊗ h) = K A{Z, y, v}h. kernel map K A n we write K A H R(B) is spanned by the vectors K B{Z, v, y}g with inner product: DK B{Z, y, v}g, K B{W, x, u}fEH R(B) := (cid:16)y ⊗ g, K B(Z, W )[vu∗ ⊗ IH]x ⊗ f(cid:17)Cn⊗J K B(Z, W ) = K(Z, W ) ⊗ IJ − (K(Z, W ) ⊗ IJ )[B†(Z)(· ⊗ IH)B†(W )∗]. It is not difficult to see that free operator-valued holomorphic functions A, B on Bd N belong to the left or right free Schur classes if and only if the above NC deBranges-Rovnyak kernels are (completely) positive. Remark 2.7. (Right Product) If F ∈ MultR(K1, K2)⊗ L(H, J ) is a right operator-valued multiplier between vector-valued NC-RKHS on (say) the open unit NC ball Bd N, then one can easily verify that for any g ∈ J , (M R F )∗K2 {Z, y, v} g = K1{Z, y, F (Z) •R v}g, where for any f ∈ Hnc(K1) we define: hK1{Z, y, F (Z) •R v}g, fi = hy ⊗ h, F (Z) •R f (Z)vi . Also, in the above, given any element f of a H−valued NC-RKHS Hnc(K), note that f (Z) ∈ Cn×n ⊗ H, so that f (Z)v is to be interpreted as an element of Cn ⊗ H. 8 MICHAEL T. JURY AND ROBERT T.W. MARTIN 2.8. Coefficient evaluation and free formal RKHS. Let K be an operator-valued CPNC kernel on Bd 1 = Bd converges absolutely on Bd N, and uniformly on compacta: N whose Taylor-Taylor series about 0 ∈ Bd K(Z, W )[P ] = Xα,β∈Fd Kα,βZ α[P ](W ∗)β† ; Kα,β ∈ L(H). Any right or left (operator-valued) deBranges-Rovnyak kernel has this property, for ex- ample. The coefficient kernel function K(·,·) : Fd × Fd → L(H) is then an operator-valued free formal kernel in the sense of [4, 3] (see also [8] which develops free Aleksandrov-Clark theory using the free formal RKHS setup). If F (Z) :=Pα FαZ α ∈ Hnc(K), then for any α ∈ Fd, the linear H−valued map defined α)∗ : H → Hnc(K) will be called the is bounded. The Hilbert space adjoint Kα := (K∗ coefficient kernel map, and one always has α(F ) := Fα ∈ H, by coefficient evaluation: K∗ Kα,β = K∗ αKβ ∈ L(H), and Kβ(Z) = Xα∈Fd Kα,βZ α. Observe that Kα,β is a positive kernel function in the classical sense on the discrete set Fd. 2.9. The free Herglotz-Schur classes. Definition 2.10. The left free Herglotz-Schur class, L + valued free NC functions on Bd N. d , is the set of all accretive matrix- If H ∈ L + follows that d , then H(Z) is an accretive matrix (positive semi-definite real part), and it is a contractive free function on Bd N, i.e., B ∈ Ld. Conversely, given such a B, BH (Z) := (H(Z) − I)(H(Z) + I)−1, HB(Z) := (I + B(Z))(I − B(Z))−1, has non-negative real part. That is, this fractional linear transformation, the Cayley Transform, defines a bijection between Ld and L + d . The right free Herglotz-Schur class, R + d under the transpose map, †, and we similarly define the operator- valued free Herglotz-Schur classes L + d (H, H) as the image of Ld(H) under Cayley transform. d is the image of L + d (H) = L + Remark 2.11. By [10, Lemma 3.2, Lemma 3.3], if B ∈ Rd or Ld, then 1 − B is outer (and necessarily invertible on Bd N). Moreover, the free Herglotz-Schur classes are contained in the free Smirnov classes, which can be identified with closed, densely-defined (generally unbounded) right and left multipliers of H 2(Bd N) [10]. In particular, by [10, Corollary 3.13, Corollary 3.15], if B ∈ Ld or A ∈ Rd, then the free polynomials belong to the domains of and are cores for both HB(L)∗ and HA(R)∗. 3. Free Aleksandrov-Clark measures Let Ad :=(cid:0)Wα∈Fd Lα(cid:1)−k·k, the free disk algebra. In the case where d = 1, L = S (the shift), and we recover the classical disk algebra A1 = A(D) of bounded analytic functions on D with continous extensions to the unit circle T = ∂D. COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 9 Recall the Herglotz representation formula for Herglotz functions (analytic functions with non-negative real part) on the disk: If H is a Herglotz function, then there is a unique finite positive Borel measure µ on T so that: H(z) = iIm (H(0)) +ZT 1 + zζ 1 − zζ µ(dζ). As discussed above, any such H ∈ S + := L + 1 is the Cayley transform of some contractive analytic Schur class function b ∈ S := [H ∞(D)]1 = L1, H = Hb = 1+b 1−b , and the measure µ =: µb is called the Aleksandrov-Clark measure of b. This defines bijections (modulo imaginary constants) between S , S +, and the set of all finite positive Borel measures on T. Any finite positive Borel measure µ on T can be identified with a positive linear func- tional on the disk algebra operator system: A1 + A∗ 1: µ(Sn) :=ZT ζ nµ(dζ). The theory of closed, densely-defined operators affiliated to the shift [24, 22], implies that if b ∈ S , then multiplication by Hb ∈ S + is a closed, densely-defined (and accretive) operator on H 2(D), and that W Sn1 is a core for M ∗ It is then easy to verify the following formula for µb: Hb . (3.1) µb(Sn) = 1 2(cid:0)hHb(S)∗1, Sn1iH 2 + h1, Hb(S)∗Sn1iH 2(cid:1) , where m denotes normalized Lebesgue measure on T. Equivalently, using that any kernel vector kz, for z ∈ D, is necessarily an eigenvector for Hb(S)∗ with eigenvalue Hb(z)∗, one can check that (3.2) Hb(z) = iIm (Hb(0)) + µb(cid:0)(I + zS∗)(I − zS∗)−1(cid:1) . 1 1 = d µB(Lα) := This Clark functional formula (3.1) extends verbatim to the non-commutative several- variable setting: Definition 3.1. Let (A, B) ∈ Rd × Ld be a transpose-conjugate (A = B†) free Schur class pair. The Clark functional of (A, B) is the self-adjoint linear functional: µA = µB : W Lα1 +W(Lα)∗ → C defined by: In the above, recall that the free monomials always belong to the domain of HA(R)∗, d in place of its norm closure: 2(cid:16)hHA(R)∗1, Lα1iF 2 2(cid:16)HB(0)δα,∅ + h1, HA(R)∗Lα1iF 2 d(cid:17) . see Remark 2.11. To simplify notation, we will write Ad +A∗ Ad + A∗ Lemma 3.2. µB extends by continuity to a positive, bounded linear functional on the norm-closed operator system Ad + A∗ d. Proof. That Lα1 belongs to the domain of HA(R)∗ (and that, in fact, free polynomials are a core for HA(R)∗) follows from [10, Corollary 3.9, Corollary 3.10, Remark 3.12]. It d(cid:17) + h1, HA(R)∗Lα1iF 2 is easy to check that µB is positive on Wα Lα +Wα(Lα)∗. Since this is a positive linear functional, its norm is given by d = (Ad + A∗ d)−k·k. kµBk = µB(I) = Re ((HB)∅) = Re (HB(0)) < ∞. Hence, µB extends by continuity to a bounded positive linear functional on the free disk operator system. (cid:3) 10 MICHAEL T. JURY AND ROBERT T.W. MARTIN Theorem 3.3. The map B 7→ µB is a bijection from Ld onto the set of positive linear functionals on the free disk system, and one has the free Herglotz formula: For any Z ∈ Bd n, HB(Z) = iIm (HB(0n)) + (idn ⊗ µB)(cid:0)(In×F 2 + ZL∗)(In×F 2 − ZL∗)−1)(cid:1) . In the above, and Also note that: ZL∗ := (Z ⊗ IF 2)(In ⊗ L)∗ = Z1 ⊗ L∗ 1 + ... + Zd ⊗ L∗ d, In×F 2 := In ⊗ IF 2 d . (I − ZL∗)−1 = ∞Xk=0 (ZL∗)k = Xα∈Fd Z α ⊗ (L∗)α. Proof. Let HB(Z) = Pα HαZ α be the Taylor-Taylor series of HB(Z) about Z = 0 ∈ Bd = Bd 1. Consider (idn ⊗ µB)(cid:0)(I − ZL∗)−1(cid:1): (idn ⊗ µB)(cid:0)(I − ZL∗)−1(cid:1) = Xα = Xα Z α 1 Z αµB((L∗)α) 1, 1E + δα†,∅H ∗ ∅(cid:17) hLγ , 1i Hβ† 1 2(cid:16)DHA(R)∗Lα† 2Xα Z α Xγβ=α† 2Xα Z αHα 1 InH ∗ ∅ + HB(0n)∗ + HB(0n)∗ + HB(Z). 1 2 = = = 1 2 1 2 1 2 In the above, note that HA(R) = U†HB(L)U ∗ the fact that if † , so that HA(R) = M R . We also used H † B (Z) is any free holomorphic function so that M R Lα1 ∈ Dom((M R F (Z))∗), and F (Z) is densely-defined, then the monomials F (Z) =Xα Z αFα, (M R F (Z))∗Lα1 = Xγβ=α Lγ 1F ∗ β , see for example [10, Corollary 3.13] and [8, Lemma 2.3]. Using that (I+ZL∗)(I−ZL∗)−1 = 2(I − ZL∗)−1 − I, we obtain: (idn ⊗ µB)(cid:0)(I + ZL∗)(I − ZL∗)−1(cid:1) = HB(0n)∗ + HB(Z) − Re (HB(0n)) , and the formula follows. Conversely, starting with a positive linear functional on the free disk system, this Herglotz formula defines a free Herglotz function, and by Cayley transform we obtain a free Schur function whose Clark functional is the original functional. This shows B 7→ µB is surjective. (cid:3) Remark 3.4. Replacing F 2 d ⊗ H where H is a separable or finite-dimensional Hilbert space, the above results are easily extended to the operator-valued setting of B ∈ Ld(H). In this operator-valued setting we define the Clark map µB : Ad +A∗ d → L(H) by the formula: d with F 2 µB(Lα) := 1 2 H∅δα,∅IH + 1 2 (h1, ·1i ⊗ idH) (HA(R)∗(Lα ⊗ IH)) . COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 11 Again this defines a bijection between the operator-valued free Schur classes, the operator- valued free Herglotz-Schur classes, and L(H)−valued completely positive maps on the free disk system. Remark 3.5. Theorem 3.3 was first obtained by Popescu in [19, Section 5, Theorem 5.3] (with a different, but equivalent formula for the Clark functional). Given B ∈ Ld, the Clark functional µB can also be defined in terms of the Fourier series coefficients of the free Herglotz function HB, as in [8, Section 4]. 4. Free Cauchy transforms In his seminal paper on unitary perturbations of the shift (see [21] for the fully general, non-inner case), D.N. Clark showed that there is a canonical isometry, the weighted Cauchy Transform, Fb, from H 2(µb), the closure of the analytic polynomials in L2(µb) (the Hilbert space of functions on T which are square-integrable with respect to µb), onto H (b), the deBranges-Rovnyak space of b ∈ S = L1 [5]: For any polynomial, p ∈ H 2(µb), (Fbp) (z) := (I − b(z))ZT 1 1 − zζ ∗ p(ζ)µb(dζ). One can also define an unweighted Cauchy Transform, Cb, from H 2(µb) onto H+(Hb) := H(K Hb ), the Herglotz space of b, the unique RKHS corresponding to the positive sesqui- analytic Herglotz kernel: K Hb (z, w) := 1 2 Hb(z) + Hb(w)∗ 1 − zw∗ With a bit of algebra, one can verify that ; z, w ∈ D. K Hb (z, w) = (I − b(z))−1kb(z, w)(I − b(w)∗)−1, where kb is the deBranges-Rovnyak kernel of b. The theory of RKHS then implies that the multiplier Ub := M(I−b) : H+(Hb) → H (b), 1 is an isometry of the Herglotz space onto the deBranges-Rovnyak space of b. The Cauchy Transform Cb : H 2(µb) → H+(Hb) is the linear map defined by: 1 − zζ ∗ p(ζ)µb(dζ), (Cb(p)) (z) :=ZT and this extends to an isometry of H 2(µb) onto the Herglotz space of b so that Fb = UbCb. In [8], we extended the notions of Cauchy Transform and weighted Cauchy Transform to the (operator-valued and) free setting using the theory of free formal RKHS. Here we describe Cauchy transforms in the setting of NC-RKHS: Assume that B ∈ Ld(H) or that A ∈ Rd(H) are in the left or right operator-valued free Schur classes and that A = B† so that µA = µB. The free left Herglotz space H L + (HB) = Hnc(K L) is the NC-RKHS corresponding to the free left Herglotz kernel: K L(Z, W )[P ] := 1 2 (HB(Z)(K(Z, W )[P ] ⊗ IH) + (K(Z, W )[P ] ⊗ IH)HB(W )∗) , where K is the free Szego kernel. As in the classical theory, it is straightforward to verify that is an onto isometric left free multiplier. If A ∈ Rd(H), then the right free Herglotz space, H R + (HA) onto H R(A). + (HA) is defined similarly, and UA = M R + (HB) → H L(B), (I−A†(Z)) is an isometric multiplier of H R UB := M L (I−B) : H L We can expand this kernel in a formal power series (actually a convergent Taylor-Taylor series about 0): K L(Z, W )[P ] :=Xα,β K L α,β Z αP (W ∗)β† , 12 MICHAEL T. JURY AND ROBERT T.W. MARTIN where K L α,β := µB((Lα† )∗Lβ† ), is the free left formal Herglotz kernel defined in [8, Proposition 4.5]. In the right case, if A ∈ Rd(H) one simply defines K R α,β := µA((Lα)∗Lβ). As described in [8], given a transpose-conjugate pair (A, B) ∈ Rd(H) × Ld(H), the appropriate generalization of the (analytic part of the) Clark measure space H 2(µb) is the Stinespring-Gelfand-Naimark-Segal (S-GNS) space or free Hardy space of µB : Ad +A∗ d → L(H), F 2(µB). This is defined as the Hilbert space completion of Ad ⊗H (modulo vectors of zero length) with respect to the pre-inner product: DLα ⊗ h, Lβ ⊗ gEµB :=Dh, µB(cid:16)(Lα)∗Lβ(cid:17) gEH . The semi-Dirichlet property: (Ad)∗Ad = Ad + A∗ defined inner product, and the left regular representation: Lα 7→ πB(L)α where d (norm closure) ensures this is a well- πB(Lα)p(L) ⊗ h := Lαp(L) ⊗ h, is completely isometric, unital, and extends to a ∗-representation of the Cuntz-Toeplitz C ∗−algebra. We will set ΠB k := πB(Lk), so that ΠB = πB(L) is the S-GNS row isometry on F 2(µB). This also provides a S-GNS formula for µB: where [I⊗]B : H → F 2(µB) is the bounded embedding: µB(Lα) = [I⊗]∗ BπB(L)α[I⊗]B, [I⊗]Bh := I ⊗ h. The left and right Cauchy transforms, CL : F 2(µB) → H L + (HA) are then defined by H R + (HB) and CR : F 2(µA) → (4.1) CL(Lα ⊗ h) := K L α† h, and, CR(Lα ⊗ h) := K R α h. Observe that if HB(Z) =Pα HαZ α, then HA(Z) =Pα Hα† Z α. One can then calcu- 1 late that: In the above, we write β ≥ α if β = αγ, and β > α if β ≥ α and β 6= α. Similarly, K L α,β = α,β = K R 1 (β†\α†)† 2 H ∗ β† > α† 2 H(α†\β†)† α† > β† Re (H∅) α = β else. 0 1 (β\α)† 2 H ∗ β > α 1 2 H(α\β)† α > β α = β Re (H∅) else. 0 These formulas follow easily from the Clark map formula. For example if β > α (and H = C) then K R α,β = µA(Lβ\α) = = 0 + 1 2(cid:16)DHA(R)∗1, Lβ\α1E +D1, HA(R)∗Lβ\α1E(cid:17) 2 Xγλ=β\α h1, Lγ 1i H ∗ λ† 1 = (H)∗ (β\α)†. 1 2 The above formulas allow one to alternatively define the Clark map of B in terms of the Fourier series coefficients of the Herglotz functions HB, as was done in [8, Section 4]. (cid:3) . COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 13 Lemma 4.1. The left free Cauchy transform acts as: (CLp(L)1)(Z) = (idn ⊗ µB)(cid:0)(In×F 2 − ZL∗)−1(In ⊗ p(L))(cid:1) . Proof. It suffices to check on monomials, so take p(L) = Lβ† 1. Then, the above becomes: Z αµB((Lα† )∗Lβ† ) = K L α (Z), Xα as claimed. Lemma 4.2. Given any Z ∈ Bd , (I − Z∗L)−1(y ⊗ 1)(cid:17)Cn n and v, y ∈ Cn, (Z αv, y)Cn Lα† 1 K L{Z, y, v} = CLXα d DK L{Z, y, v}, FE = (y, F (Z)v)Cn = CL(cid:16)v ⊗ IF 2 Proof. For any F ∈ H L,+(HB), we have that = Xα = Xα = *Xα (y, Z αv)Cn Fα (y, Z αv)CnDK L (Z αv, y)Cn K L The above proves that K L {Z, y, v} = Pα (Z αv, y)Cn K L CLLα† 1. Hence we have that: α , FE α , F+ . K L{Z, y, v} = CLX (Z αv, y)Cn Lα† 1 , (Z∗)α† = CLX(cid:16)v ⊗ IF 2 = CL(cid:16)v ⊗ IF 2 d d ⊗ Lα† (y ⊗ 1)(cid:17)Cn , (I − Z∗L)−1(y ⊗ 1)(cid:17)Cn . α , and by definition, K L α = (cid:3) )(cid:17) Remark 4.3. We also have the formula: K L {W, x, u} (Z) = Xα (W αu, x)Cn (idn ⊗ µB)(cid:16)(I − ZL∗)−1(In ⊗ Lα† ⊗ Lα†! = ((·u, x)Cn ⊗ µB) (I − ZL∗)−1Xα = ((·u, x)Cn ⊗ µB)(cid:0)(I − ZL∗)−1(I − LW ∗)−1(cid:1) . µb((I − zL∗)−1(I − Lw∗)−1) = K b(z, w), (W ∗)α† The above is the free version of the commutative Cauchy transform formula, from [11, Proposition 2.6, Subsection 2.8]. Here K b(z, w) is the positive Herglotz kernel for b in the Schur class of contractive Drury-Arveson space multipliers. If A = B† so that µA = µB and F 2(µA) = F 2(µB), then the weighted free Cauchy transforms F L, F R, are defined as: F L = M L (I−B)CL, and, F R = M R (I−A†)CR, and these are isometries of F 2(µB) onto H L(B) and H R(A), respectively. 14 MICHAEL T. JURY AND ROBERT T.W. MARTIN 4.4. Cauchy Transform of the Stinespring-GNS representation. As in the com- mutative setting, if B ∈ Ld(H) we define V B := CLπµB (L)(CL)∗, a row isometry on the left Herglotz space H L Proposition 4.5. The range R of the row isometry V B is: + (HB). and for any Z ∈ Bd (V B R :=_(cid:16)K HB{Z, y, v} − K HB{0n, y, v}(cid:17) = _α6=∅ n, v, y ∈ Cn, and j = 1, . . . , d, j )∗(cid:16)K HB{Z, y, v} − K HB{0n, y, v}(cid:17) = K HB{Z, y, Zj v} K HB α , (so that the span of all such vectors is dense in H L The image of Ran(cid:0)V B(cid:1) under (CL)∗ is F 2 non-constant free monomials in F 2(µB). If F ∈ H L then there is a f ∈ H so that for any Z ∈ Bd n, + (HB) ⊗ Cd). 0 (µB) = Wα6=∅ Lα ⊗ H, the closure of the + (HB) is orthogonal to Ran(cid:0)V B(cid:1), F (Z) = In ⊗ f, i.e. F ≡ f is constant-valued. Proof. By the proof of [10, Lemma 3.14], for any α ∈ Fd, one can find jointly nilpotent Z ∈ Bd n and v, y ∈ Cn with n = α + 1 so that (cid:16)Z β v, y(cid:17)Cn = δα,β . It then follows from Lemma 4.2 and the definition of left free Cauchy transform that K L{Z, v, y} = K L α . This shows that the two formulas for R above are the same. By definition, CL(Lα ⊗ h) = K L α† h, and it follows that the image of R under inverse Cauchy transform is F 2 0 (µB) = Ran (πB(L)). Since V B and ΠB are unitarily equivalent under Cauchy transform, it follows that R = Ran(cid:0)V B(cid:1). + (HB) is orthogonal to Ran(cid:0)V B(cid:1), set f := (K HB If F ∈ H L 0 = DK HB {Z, y, v}, FE −DK HB{0n, y, v}, FE )∗F ∈ H. Then, for any Z ∈ Bd n and v∗, y ∈ Cn, 0 = (y ⊗ IH, F (Z)v) − (In ⊗ f ) (y, v)Cn , and it follows that F (Z) = In ⊗ f . The second claim is a straightforward calculation: for each j = 1, . . . , d, (V B j )∗(cid:16)K L{Z, y, v} − K L{0n, y, v}(cid:17) = CLπ(Lj)∗Xα6=∅ (Z αv, y)Cn Lα† 1 = CLXβ (cid:16)Z βjv, y(cid:17)Cn = K L{Z, y, Zj v}. Lβ† 1 Since V B is an isometry, the above shows that the closed span of ⊕d all of H L j=1K L{Z, y, Zj v} is (cid:3) + (HB) ⊗ Cd. 5. Gleason solutions 5.1. The free setting. Fix A ∈ Rd(H, J ). Exactly as in the commutative setting, we define: COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 15 Definition 5.2. A linear map X : H R(A) → H R(A) ⊗ Cd is called a Gleason solution for H R(A) if: (5.1) Such an X is contractive if Z(Xf )(Z) = f (Z) − f (0n) ∀ Z ∈ Bd n. (5.2) and extremal if equality holds. X ∗X ≤ I − K A 0 (K A 0 )∗, Similarly, a linear map A : H → H R(A) ⊗ Cd is called a Gleason solution for A if ZA(Z) = A†(Z) − A(0)In ∀ Z ∈ Bd N. and extremal if equality holds. A∗A ≤ IH − A(0)∗A(0), (5.3) A is contractive if (5.4) Define: (5.5) X := L∗ ⊗ IJH R(A), and A := (L∗ ⊗ IJ )A. The right free deBranges-Rovnyak space, H R(A), is left shift co-invariant, L∗ j A ∈ H R(A), and X, A obey the contractivity conditions of Gleason solutions for H R(A), A, respec- tively [2, Proposition 4.2]. It is also easy to check that X, A are Gleason solutions. For example, given f =Pα fαLα1, it is clear that ( Xj f )(Z) =Xα fαLα\j1, where we set Lα\j = Lβ if β = jα, and = 0 else. It follows that fαZ α = f (Z) − f (0n). Z( Xf )(Z) =Xα6=∅ Also note that the defining formula (5.1) for a Gleason solution for H R(A) is equivalent to: which can be re-arranged to: (L∗K A Z )∗(Xf ) = (K A Z )∗f − (K A 0n )∗f, (5.6) Z = K A 0n . Given A ∈ Rd(H,J ), the support of A is defined to be (I − X ∗L∗)K A n; v∗,y∈Cn(cid:16)y ⊗ IH, A†(Z)∗v∗ ⊗ J(cid:17) ⊆ H. _Z∈Bd (5.7) supp(A) := Proposition 5.3. Suppose that A ∈ Rd(H,J ). A linear map X : H R(A) → H R(A)⊗Cd is a contractive Gleason solution for H R(A) if and only if there is a contractive Gleason solution A : H → H R(A) ⊗ Cd for A so that, XK A{W, x, u} = K A{W, W ∗x, u} − A(cid:16)u, A†(W )∗x(cid:17)Cm ∈ L(J , H R(A) ⊗ Cd). X is extremal if A is extremal. Conversely A is extremal if X is extremal and supp(A) = H. This is a free analogue of [11, Theorem 4.4]. Since the proof is (formally) analogous, we prove only the sufficiency. Remark 5.4. The expression (cid:0)u, A†(W )∗x(cid:1)Cm is to be interpreted as taking values in m, A†(W )∗ ∈ Cm×m ⊗ L(J ,H). Namely, for any g ∈ J , the L(J ,H) since, for W ∈ Bd above formula can be written: XK A{W, x, u}g = K A{W, W ∗x, u}g − A(cid:16)u ⊗ IH, A†(W )∗x ⊗ g(cid:17) . 16 MICHAEL T. JURY AND ROBERT T.W. MARTIN Proof. Let A be a contractive Gleason solution for A. We wish to show that the formula in the proposition statement defines a contractive Gleason solution for H R(A). To prove this, it is sufficient to check that Formula (5.1) holds on kernel vectors. Namely, it suffices to show that (5.8) DK A{Z, Z∗y, v}, XK{W, x, u}E =(cid:16)y, Z(XK A{W, x, u})(Z)v(cid:17)Cn In the above we have used the compact notation: = (cid:16)y,(cid:16)K A(Z, W )[vu∗] − K A(0n, W )[vu∗](cid:17) x(cid:17)Cn ∈ L(J ). DK A{Z, Z∗y, v}, XK{W, x, u}E := dXj=1DK A{Z, Z∗ and we will continue to use this throughout. Calculate: j y, v}, Xj K{W, x, u}E , DK A{Z, Z∗y, v}, XK {W, x, u}E =DK A{Z, Z∗y, v}, K A{W, W ∗x, u}E −DK A{Z, Z∗y, v}, AE(cid:16)u, A†(W )∗x(cid:17)Cm = (cid:16)y, ZK A(Z, W )[v∗u]W ∗x(cid:17)Cn − (y, ZA(Z)v∗)Cn(cid:16)u∗, A†(W )∗x(cid:17)Cm = (cid:16)y,(cid:16)K A(Z, W )[vu∗] − vu∗ + A†(Z)vu∗A†(W )∗(cid:17) x(cid:17)Cn = (cid:16)y,(cid:16)K A(Z, W )[vu∗] − vu∗ + A†(Z)vu∗A†(W )∗(cid:17) x(cid:17)Cn +(cid:16)y, A(0n)vu∗A†(W )∗x(cid:17)Cn = (cid:16)y,(cid:16)K A(Z, W )[vu∗] − K A(0n, W )[vu∗](cid:17) x(cid:17)Cn ∈ L(J ). 1 = Bd, since A† is In the above, note that A†(0n) = A(0n) = A(0)In = A∅In where 0 ∈ Bd a free function. This proves that X is a Gleason solution. To see that X is contractive, again calculate on kernel vectors: kXK A{Z, y, v}k2 = DK A{Z, Z∗y, v}, K A{Z, Z∗y, v}E −(cid:16)y, (A†(Z) − A†(0n))v(cid:17)Cn(cid:16)u, A†(W )∗x(cid:17)Cm −(cid:16)y, A†(Z)vu∗A†(W )∗x(cid:17)Cn 2 −DK A{Z, Z∗y, v}, AE(cid:16)A†(Z)v, y(cid:17)Cn −(cid:16)y, A†(Z)v(cid:17)CnDA, K A{Z, Z∗y, v}E + kAk2(cid:12)(cid:12)(cid:12)(cid:16)A†(Z)v, y(cid:17)Cn(cid:12)(cid:12)(cid:12) ≤ (cid:16)y, ZK A(Z, Z)[vv∗]Z∗y(cid:17)Cn −(cid:16)y, (A†(Z) − A(0n))v(cid:17)Cn(cid:16)A†(Z)v, y(cid:17)Cn −(cid:16)y, A†(Z)v(cid:17)Cn(cid:16)(A†(Z) − A(0n))v, y(cid:17)Cn +(I − A(0)∗A(0))(cid:16)y, A(Z)vv∗A†(Z)∗y(cid:17)Cn −2(cid:16)y, A†(Z)vv∗A†(Z)∗y(cid:17)Cn +(I − A(0)∗A(0))(cid:16)y, A†(Z)vv∗A†(Z)∗y(cid:17)Cn = (cid:16)y, K A(Z, Z)[vv∗]y(cid:17)Cn − (y, vv∗y)Cn +(cid:16)y, A(0n)vv∗A†(Z)∗y(cid:17)Cn +(cid:16)A(0n)vv∗A†(Z)∗y, y(cid:17)Cn − A(0)∗A(0)(cid:16)y, A†(Z)vv∗A†(Z)∗y(cid:17)Cn = (cid:16)y, K A(Z, Z)[vv∗]y(cid:17)Cn −(cid:16)y, (vv∗ − A†(Z)vv∗A†(Z)∗)y(cid:17)Cn +(cid:16)y, A(0n)vv∗A†(Z)∗y(cid:17)Cn (5.9) . +(cid:16)A(0n)vv∗A†(Z)∗y, y(cid:17)Cn COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 17 Observe that equality holds in the above if A is extremal. Compare this to: 0 (K A 0 )∗)K A{Z, y, v}E DK A{Z, y, v}, (I − K A =(cid:16)y, K A(Z, Z)[vv∗]y(cid:17)Cn − kK A{Z, y, v}(0)k2 = kK A{Z, y, v}k2 −(cid:12)(cid:12)(cid:12)(cid:16)y, K A(0n, Z)v(cid:17)Cn(cid:12)(cid:12)(cid:12) = kK A{Z, y, v}k2 −(cid:12)(cid:12)(cid:12)(cid:16)y,(cid:16)In − A(0n)A†(Z)∗(cid:17) v(cid:17)Cn(cid:12)(cid:12)(cid:12) = kK A{Z, y, v}k2 − (y, vv∗y)Cn + 2Re(cid:16)(y, v)Cn(cid:16)y, A(0)A†(Z)∗∗(cid:17)Cn(cid:17) −A(0)∗A(0)(cid:16)y, A†(Z)∗v(cid:17)Cn(cid:16)A†(Z)∗v, y(cid:17)Cn 2 2 , which is the same (up to elementary manipulations) as Equation (5.9) above. This proves that X ∗X ≤ I − K A 0 )∗ so that X is a contractive Gleason solution (which will be extremal if A is). (cid:3) 0 (K A Theorem 5.5. Suppose that A ∈ Rd(H). Then A := (L∗ ⊗ IH)A and X := (L∗ ⊗ IH)H R(A) are the unique contractive Gleason solutions for A and H R(A), respectively. The proof uses similar arguments to those of [11, Section 4]. Lemma 5.6. ([8, Proposition 6.2]) The contractive Gleason solution A = (L∗ ⊗ IH)A is given by the formula A = F RΠ∗ A[I⊗]A(I − A(0)) = UA(V A)∗K HA 0 (I − A(0)), where ΠA = πA(L) is the row isometry obtained from the S-GNS representation of the free Clark measure µA. Recall that UA := M R V A is the row isometry on H L (I−A†(Z)) is the unitary multiplier of H R + (HA) onto H R(A), and + (HA) defined in Subsection 4.4. + (HA) → H R + (HA) ⊗ Cd by: Proof. (of Theorem 5.5) Let A be any contractive Gleason solution for A. Define a linear map D∗ : H R (5.10) D∗K HA {Z, y, v} := K HA{Z, Z∗y, v} + M R Recall U ∗ construction, (I−A†(Z))−1 , the unitary right multiplier of H R(A) onto H R (I−A†(Z))−1 A(I − A(0))−1 (v, y)Cn . + (HA). By A := M R D∗ (K {Z, y, v} − K{0n, y, v}) = K HA{Z, Z∗y, v} = (V A)∗ (K {Z, y, v} − K{0n, y, v}) . We claim that D∗ is a contraction: kD∗K HA {Z, y, v}k2 = kK HA{Z, Z∗y, v}k2 AAE (I − A(0))−1 (v∗, y)Cn + c.c. +DK HA{Z, Z∗y, v},U ∗ + (y, v)Cn (I − A(0)∗)−1 hA, Ai (I − A(0))−1 (v, y)Cn ≤ kK HA{Z, Z∗y, v}k2 +DK HA{Z, Z∗y, (I − A†(Z))−1v}, AE (I − A(0))−1 (v, y)Cn +c.c. + K HA (0, 0) (y, vv∗y)Cn , (5.11) 18 MICHAEL T. JURY AND ROBERT T.W. MARTIN where c.c. denotes the complex conjugate of the previous term. The cross-term becomes: 1 = DK HA{Z, Z∗y, (I − A†(Z))−1v}, AE (I − A(0))−1 (v, y)Cn = (cid:16)y, (I − A(0n))−1(A†(Z) − A(0n))(I − A†(Z))−1v(cid:17)Cn 2(cid:16)y, (H † kK HA{Z, Z∗y, v}k2 + A(0n))(vv∗)y(cid:17)Cn 2(cid:16)y,(cid:16)H † A(Z)vv∗ + vv∗H † A(Z) − H † 1 . A(Z)∗(cid:17) y(cid:17)Cn (v, y)Cn It follows that Equation (5.11 ) becomes: = kK HA{Z, y, v}k2. This proves that D∗ is a contractive extension of (V A)∗ so that D is a row contractive extension of V A (by, for example, [11, Lemma 2.3]). However V A is a row isometry and has no non-trivial extensions. Hence, D = V A, and Equation (5.10) and Lemma 5.6 then imply that A = (L∗ ⊗ IH)A = A. It follows also that X is unique, by Proposition 5.3, so that X = (L∗ ⊗ IH)H R(A). (cid:3) If B ∈ Ld(H), the formula is similar: (5.12) B = M L (I−B(Z))(V B)∗K HB 0 (I − B(0)). 6. Column Extreme Recall that B ∈ Ld(H) or A ∈ Rd(H) is said to be quasi-extreme if F 2 0 (µB) = F 2(µB), i.e. if and only if I ⊗ H ⊆ F 2 0 (µB), see [8] and [11, Definition 3.19]. This concept of quasi-extreme was first introduced for contractive scalar multipliers of the Drury-Arveson space in [9], extended to operator- valued multipliers b ∈ Sd(H) in [11], and to the 'rectangular setting' of arbitrary b ∈ Sd(H,J ) in [15]. (Here Sd(H,J ) denotes the Schur class of contractive operator-valued multipliers between vector-valued Drury-Arveson spaces.) The main result of [12] shows that a more descriptive name for this property could be column extreme (CE), and we will use this new terminology for the remainder of this paper. Definition 6.1. A Schur class B ∈ Ld(H,J ) is column extreme (CE) if there is no non-zero A ∈ Ld(H, J ) so that the column: A(cid:19) ∈ Ld(H,J ⊗ C2), (cid:18)B is also Schur class. Column extreme for the right Schur class is defined analogously. Remark 6.2. Observe that the definition of column extremity can be recast as follows: B is column extreme if and only if the only multiplier A satisfying the inequality (6.1) M L∗ A M L A ≤ I − M L∗ B M L B is A = 0. The existence of such A for given B was considered by Popescu [17], who showed that a nonzero A exists if and only if e(I − M L∗ B ) > −∞, where e(·) is the so-called entropy of a multi-analytic Toepliz operator as defined in [17]. However, it seems to be difficult to compute the entropy for arbitrary B (or even to decide if it is finite or not). Regarding the equivalences in Theorem 6.4 below, it is not hard to see from the definition of the entropy invariant, that e(I − M L∗ B ) = −∞ is equivalent to our condition (5), so that the equivalence of (1) and (5) is essentially contained in [17, Corollary 1.2]. B M L B M L COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 19 In this general 'rectangular' setting, it will often be convenient to consider the square completion [B], of B: The above column-extreme property is clearly invariant under conjugation by isometries; a given B ∈ Ld(H,J ) is CE if and only if B′ = V BW ∗ is CE, where W : H → H′, and V : J → J ′ are fixed onto isometries. It follows that we can assume, without loss of generality, that H ⊆ J or J ( H, and complete B to a 'square' [B] ∈ Ld(J ) or in Ld(H), respectively by adding columns or rows of zeros: [B] := (cid:2)B 0J ⊖H(cid:3) H ( J 0H⊖J(cid:21) (cid:20) B J ( H. Remark 6.3. Observe that if H ( J then H L(B) = H L([B]) so that the unique contractive Gleason solution for B is given by [B]H, where [B] is the unique contractive Gleason solution for [B]. It is clear that B is extremal if and only if [B] is extremal in this case. In the second case where J ( H we have that H L([B]) = H L(B)L(cid:0)F 2 and the unique contractive Gleason solution for [B] is given by: d ⊗ (H ⊖ J )(cid:1), [B] = (L∗ ⊗ IH)(cid:18) B 0H⊖J(cid:19) =(cid:18) B 0H⊖J(cid:19) , where 0H⊖J : H ⊖ J → H L(B) ⊗ Cd maps everything to the zero element. In this case it is clear that and it follows as before that B is extremal if and only if [B] is extremal. Theorem 6.4. Given B ∈ Ld(H, J ), the following are equivalent: B =(cid:0)I, 0(cid:1) [B], (1) B is column extreme. (2) The unique contractive Gleason solution B = R∗B : H → H L(B) ⊗ Cd, for B is (3) The unique contractive Gleason solution X = R∗H R(B) for H R(B) is extremal, extremal. and H = supp(B). (I − B(0))H ⊆ Ran(cid:16)V [B](cid:17). H[B] (4) K 0 (5) B has the Szego extremal property: I ⊗ (I − B(0))H ⊆ F 2 0 (µ[B]). (6) There is no non-zero H-valued constant function H ≡ h ∈ H L (7) There is no non-zero h ∈ H so that Bh ∈ H L(B). + (H[B]). If B = [B] is square, then the above are equivalent to: (8) πµB (L) (equivalently V B) is a Cuntz row isometry. In the above, recall that F 2 defined in Equation (5.7). 0 (µB) =Wα6=∅ Lα ⊗ IH ⊆ F 2(µB), and the support of B was Remark 6.5. In the classical (single-variable, scalar-valued) setting, the equivalent state- ments in the above theorem recover several characterizations of extreme points of the Schur class, S , of contractive analytic functions on D: Theorem. Given b ∈ S , the following are equivalent: (0) b is an extreme point. (1) b is column extreme. (2) S∗b is extremal, i.e. kS∗bkH (b) = 1 − b(0)2. (3) X := S∗H (b) is extremal, i.e., X ∗X = I − kb (5) H 2(µb) = H 2 (7) b does not belong to H (b). (8) All Clark perturbations of S∗H (b) are unitary. 0 (µb). 0(kb 0)∗. 20 MICHAEL T. JURY AND ROBERT T.W. MARTIN Conditions (1) and (2) are characterizations of extreme points of S which follow from results of Sarason, [21, Chapter III, Chapter IV]. The spaces F 2(µB), and F 2 0 (µB) are multi-variable non-commutative analogues of the spaces H 2(µb), and H 2 0 (µb), the clo- sure of the analytic polynomials and closed span of the non-constant analytic monomials 0 (µb) = H 2(µb) (resp.) in L2(µb), for b ∈ S . In the classical setting the condition that H 2 is equivalent to b being an extreme point of the Schur class. This follows from the Szego-Kolmogoroff-Kreın distance formula for the distance from the constant function 1 to H 2 0 (µb), the formula for the Radon-Nikodym derivative of µb with respect to normal- ized Lebesgue measure, and other classical facts, see [7, Chapter 4, Chapter 9], and the discussion in [8, Section 1]. Item (7) is again a result of Sarason [21, Chapter IV], and the final item is equivalent to the well-known fact that b is extreme if and only if all of the Clark perturbations of the restricted backward shift S∗H (b) are unitary, see e.g. [14]. Corollary 6.8 will prove that any column-extreme B ∈ Ld or Rd is necessarily an extreme point, whether the converse holds is an open problem. We will need the following free or NC analogue of a result from vector-valued RKHS theory, [16, Theorem 10.17] (see the proof of [12, Proposition 5.1]): Lemma 6.6. Let Hnc(K) be a vector-valued NC-RKHS on a NC set Ω. A (vector-valued) free NC function f on Ω belongs to Hnc(K) if and only if λ2K(Z, W )− f (Z)(·)f (W )∗ ≥ 0 is a (operator-valued) CPNC Kernel on Ω for some λ2 > 0. The norm of f is the infimum of all such λ. Lemma 6.7. If B ∈ Ld(H,J ) and C := [B] ∈ Ld(K), then Bh ∈ H L(B) for h ∈ H if and only if F ≡ (I − C(0)∗)−1h ∈ H is a constant function in H L Proof. If Bh ∈ H L(B) set f := (I − C(0)∗)−1h ∈ H. Then, + (HC). (K B 0 f )(Z) = In ⊗ f − B(Z)(In ⊗ B(0)∗(I − C(0)∗)−1h) = In ⊗ f − B(Z)C(0)∗(I − C(0)∗)−1h = In ⊗ f − B(Z)(I − C(0)∗)−1h + B(Z)h = (In − B(Z))f + B(Z)h. Since B(Z)h = B(Z)(In⊗h) ∈ H L(B), we conclude that F ≡ f ∈ H belongs to H L (it is the image of (I − B)f under the canonical unitary multiplier). In ⊗ h, then M L (I−B(Z))H = F ∈ H L(C), and also + (HC) is such that H ≡ h ∈ H, i.e. H(Z) = This argument is reversible: if H ∈ H L + (HC) so that 0 (Z)h = In ⊗ h − B(Z)(In ⊗ B(0)∗h), K B (K B 0 − F )(Z) = B(Z)(I − B(0)∗), and we conclude that B(I − C(0)∗)h = B(I − B(0)∗)h ∈ H L(C). If J ⊇ H so that [B] ∈ Ld(J ), then H L(C) = H L(B). Otherwise if H ) J then H L(C) = H L(B) ⊕ d ⊗(H⊖J ). Since C(Z)h = B(Z)h for any h ∈ H, and B(Z) ∈ Cn×n⊗ L(H,J ), Lemma F 2 6.6 implies that there is a λ2 > 0 so that (cid:18)B(Z)g(B(W )g)∗ 0 0 0H⊖J(cid:19) = C(Z)g(C(W )g)∗ ≤ λ2K C (Z, W ) = λ2(cid:18)K B(Z, W ) where g := (I − B(0)∗)h = (I − C(0)∗)h ∈ H. Comparing top left entries, Lemma 6.6 again implies that B(Z)(I − B(0)∗)h = Bg ∈ H L(B). K(Z, W ) ⊗ (H ⊖ J )(cid:19) , 0 0 (cid:3) The proof of equivalence of the first two items is the most involved, so we will first establish the equivalence of the remaining items. COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 21 Proof. (of equivalence of items (2) − (10) in Theorem 6.4) To simplify notation, we will write C = [B] ∈ Ld(K) (with K = H or J ). (2) ⇔ (3). This was proven as part of Proposition 5.3 which relates Gleason solutions for H L(B) and Gleason solutions for B. Also note that if supp(B) ( H, then B will have a matrix representation of the form: (cid:0)B′, so that for any Schur A′ ∈ Ld, A(cid:19) ∈ Ld(H ⊕ C,J ⊕ C), 0 A′(cid:19) =:(cid:18)B (cid:18)B′ 0(cid:1) , 0 is Schur and B is not CE in this case. (2) ⇔ (4). There are two cases to consider. If C = [B] ∈ Ld(H) then H L(B) ⊕ F 2 (H ⊖ J ) = H L(C) and B is extremal if and only if C is. If C is extremal then d ⊗ I − C(0)∗C(0) = C∗C 0 = (I − C(0)∗)(K HC )∗V C (V C )∗K HC ≤ (I − C(0)∗)K HC (0, 0)(I − C(0)) = I − C(0)∗C(0), 0 (I − C(0)) and this happens if and only if V C is a co-isometry. This establishes the equivalence in this case. Alternatively, if H ⊆ J then B = CH so that B will be extremal if and only if I − B(0)∗B(0) = B∗B = PHC∗CPH = PH(I − C(0)∗)(K HC ≤ PH(I − C(0)∗C(0))PH = I − B(0)∗B(0). )∗V C(V C )∗K HC 0 0 (I − C(0))PH 0 0 (I − C(0))H = K HC 0 (µC ) onto Ran(cid:0)V C(cid:1), and which maps I ⊗ g ∈ F 2(µC ) to K HC (I − B(0))H ⊆ Ran(cid:0)V C(cid:1), and The above holds if and only if K HC this proves the equivalence in the second case. (4) ⇔ (5). This follows immediately from the fact that the Cauchy transform CL : F 2(µC ) → H L + (HC) is an onto isometry which intertwines ΠC = πC (L) and V C , which takes F 2 g, see Proposition 4.5. (4) ⇒ (6). Assume that (6) does not hold so that there is a constant H−valued function + (HC). Set f := (I − B(0)∗)h ∈ H so that h = (I − C(0)∗)−1f . If (4) H ≡ h ∈ H in H L (I − C(0))f ∈ Ran(cid:0)V C(cid:1) ⊥ H so that holds then K HC 0 =DK HC (I − B(0))f , HE =(cid:10)(I − C(0))f , (I − C(0)∗)−1f(cid:11) = kfk2. We conclude that f = 0 so that h = 0 (since B(0)∗ is a pure contraction). This shows that (4) cannot hold. (6) ⇔ (7). This equivalence is an immediate consequence of Lemma 6.7. (7) ⇒ (4). Our proof will be a bit circuitous: First consider the following condition (4)′: If F ∈ H L 0 0 0 + (HC) is constant valued, F ≡ f ∈ K, then PH(I − B(0)∗)f = 0. We claim that (4)′ ⇒ (4). Condition (4)′ implies that PH(I − B(0)∗)(K HC and taking the adjoint of this expression gives: 0 )∗(I − V V ∗) = 0, (I − V V ∗)K HC 0 (I − B(0))PH = 0, which is condition (4). It remains to show that (7) ⇒ (4)′, and this will be accomplished If (4)′ does not hold then there is a f ∈ K so by demonstrating the contrapositive. + (HC), and PH(I − B(0)∗)f 6= 0. By Lemma 6.7, that F (Z) = In ⊗ f belongs to H L 22 MICHAEL T. JURY AND ROBERT T.W. MARTIN (I − C(0)∗)f = g is such that Cg ∈ H L(C). There are two cases: If H ⊆ J then In this case Cg = BPHg ∈ H L(B), and by assumption, PHg = H L(C) = H L(B). PH(I − B(0)∗)f 6= 0. We conclude that (7) does not hold in this case. In the second case J ⊆ H, and g = (I− B(0)∗)f ∈ H is such that Cg = Bg ∈ H L(C) = H L(B)⊕ F 2 d ⊗(H⊖ J ). Since B(Z) ∈ Cn×n ⊗ L(H,J ), one can apply Lemma 6.6 (as in the proof of Lemma 6.7) to show that Bh ∈ H L(B), and again (7) does not hold. Hence (7) ⇒ (4)′ ⇒ (4). Assuming now that B = [B] = C ∈ Ld(H) is square, item (6) is equivalent to the statement that V C is a Cuntz (onto) row isometry, and since V C is unitarily equivalent to ΠC = πC (L) via Cauchy transform, it follows that (6) ⇔ (8). (cid:3) The proof of (1) ⇔ (2) is the free and operator-valued extension of the main result of [12], and the argument is formally analogous. Proof. ((1) ⇒ (2) of Theorem 6.4) Suppose B ∈ Ld(H,J ) is not column-extreme so that A(cid:19) is Schur. there is a non-zero A ∈ Ld(H, J ) so that the two-component column C :=(cid:18)B Without loss of generality, we can assume that 0 6= A∅ = A(0) ∈ L(H,J ). The argument is as in [12, Lemma 5.2]: If A∅ = 0 choose α ∈ Fd of minimal length so that Aα 6= 0, and set: eA := (Lα)∗ ⊗ IJ A. Then (cid:18)B eA(cid:19) =(cid:18)I C =(cid:0)R∗ ⊗ IJ , R∗ ⊗ IJ(cid:1)(cid:18)B is also Schur and satisfies eA∅ = Aα 6= 0. The unique contractive Gleason solution, C for A(cid:19) , (Lα)∗ ⊗ IJ(cid:19)(cid:18)B A(cid:19) , A(cid:19) =(cid:18)B where B, A are the unique contractive Gleason solutions for B, A. Observe that C∗C ≤ I − C(0)∗C(0) = I − B(0)∗B(0) − A(0)∗A(0) < I − B(0)∗B(0), C is: 0 0 since we can assume A(0) 6= 0. H L(C) for any h ∈ H, where 1 ≤ j ≤ d, there is a tj > 0 so that as CPNC kernels, Now we apply the argument of [12, Proposition 5.1]: By Lemma 6.6, since each Cj h ∈ ∗ ∗(cid:19) j(cid:18)K B(Z, W ) ∗ ∗ ∗(cid:19) , (Cj(Z)h)(Cj(W )h)∗ = (cid:18)(Bj(Z)h)(Bj (W )h)∗ ∗ ≤ t2 j K C (Z, W ) = t2 and one can take tj := kCj hk. It follows that for any h ∈ H, kBhk2 = hh, B∗Bhi t2 j dXj=1 ≤ = hh, C∗Chi ≤ hh, (I − C(0)∗C(0))hi = hh, (I − B(0)∗B(0) − A(0)∗A(0))hi . This proves that B∗B < I − B(0)∗B(0), so that B is not extremal. (cid:3) COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 23 The proof of (2) ⇒ (1) will employ the colligation and transfer-function theory of [2]. We briefly recall the pertinent facts: A colligation is any contractive linear map: U := A1 B1 ... ... Ad Bd C D  =:(cid:18)A B C D(cid:19) :(cid:18)K H(cid:19) →(cid:18)K ⊗ Cd J (cid:19) . The transfer-function of the contractive colligation U is the function BU defined on the free unit ball by: BU (Z) := D + C(I − ZA)−1ZB ∈ Cn×n ⊗ L(H, J ); Z ∈ Bd n, where ZA := Z1A1 + ...ZdAd. The theory of [2] shows that a free function B on Bd belongs to the left free Schur class if and only if B = BU is the transfer function of some contractive colligation U (see [2, Theorem 3.1]). Moreover, any B ∈ Ld(H,J ) always has the (left) canonical deBranges-Rovnyak colligation N CdBR DdBR(cid:19) UdBR :=(cid:18)AdBR BdBR AdBR := R∗H L(B), BdBR := R∗B, CdBR := (K B constructed by choosing K := H L(B) and 0 )∗, and DdBR := B(0), (so that B is recovered as the transfer function of this colligation) see [2, Theorem 4.3]. Similarly, if A ∈ Rd(H,J ) then A(R) = M R A†(Z) is such that the free holomorphic func- tion A†(Z) can be recovered as the transfer function of the (right) canonical deBranges- Rovnyak colligation given by choosing K := H R(A), and AdBR := L∗H R(A), BdBR := L∗A, CdBR := (K A 0 )∗, and DdBR := A(0). Proof. (of (2) ⇒ (1)) We give the proof for right free multipliers A ∈ Rd(H, J ). Assuming A is not extremal, we choose 0 ≤ a∅ ∈ L(H) satisfying: ∅ = I − A(0)∗A(0) − A∗A. a2 As in the proof of Proposition 5.3 one can calculate that, (6.2) = (cid:16)y, K A(Z, W )[vu∗]x(cid:17)Cn − (y, ZA(Z)v)Cn(cid:16)u, A†(W )∗x(cid:17)Cm DX ∗XK A{Z, y, v}, K A {W, x, u}E −(cid:16)A†(Z)∗y, v(cid:17)Cn (W A(W )u, x)Cm +(cid:16)y, A†(Z) (vu∗ ⊗ A∗A) A†(W )∗x(cid:17)Cn . ∅, the above becomes: In the above K A {Z, y, v} is a bounded linear map from J into H R(A), so that the above inner product is L(J )−valued (we have omitted vectors to simplify the notation). Applying the definition of a2 = DK A {Z, y, v}, K A {W, x, u}E −(cid:16)y, (vu∗ − A†(Z)vu∗A†(W )∗)x(cid:17)Cn −(cid:16)y, (A†(Z) − A†(0n))vu∗A†(W )∗x(cid:17)Cn −(cid:16)y, A†(Z)vu∗(A†(W )∗ − A†(0m)∗)x(cid:17)Cn +(cid:16)y, A†(Z)(cid:0)vu∗ ⊗ (I − A(0)∗A(0) − a2 ∅)(cid:1) A†(W )∗x(cid:17)Cn = DK A {Z, y, v}, K A {W, x, u}E −(cid:16)y, A†(Z)(vu∗ ⊗ a2 − (y, vu∗x)Cn +(cid:16)y, A(0n)vu∗A†(W )∗x(cid:17)Cn +(cid:16)y, A†(Z)vu∗A(0m)∗x(cid:17)Cn −(cid:16)y, A†(Z)(vu∗ ⊗ A(0)∗A(0))A†(W )∗x(cid:17)Cn ∅)A†(W )∗x(cid:17)Cn (6.4) (6.3) . 24 MICHAEL T. JURY AND ROBERT T.W. MARTIN On the other hand, one can calculate that (up to a change of sign) line (6.3) + line (6.4) in the above are equal to: 0 (K A DK A 0 )∗K A {Z, y, v}, K A {W, x, u}E , and it follows that (6.5) DX ∗XK A {Z, y, v}, K A {W, x, u}E =D(I − K A −(cid:16)y, A†(Z)(vu∗ ⊗ a2 ∅)A†(W )∗x(cid:17)Cn GA(Z, W )[P ] := A†(Z)[·] ⊗ a2 If we define the L(J )−valued CPNC kernel: . ∅A†(W )∗, 0 (K A 0 )∗)K A {Z, y, v}, K A {W, x, u}E then Equation (6.5) implies that GA ≤ K A as CPNC kernels so that, by Lemma 6.6, (where (Aa∅)(Z) = A†(Z)a∅). Moreover, Equation (6.5) further implies that Aa∅ : H → H R(A), (6.6) I − X ∗X = K A 0 (K A 0 )∗ + Aa∅(Aa∅)∗. This is the appropriate analogue of the formula from [12, Proposition 3.2]. To complete the proof, we apply the transfer function theory of [2]. We define: U := X 0 )∗ (K A −(Aa∅)∗ H (cid:19) →(cid:18)H R(A) ⊗ Cd J ⊗ C2 (cid:19) . A a∅ A(0)  :(cid:18)H R(A) a(cid:19) ∈ Rd(H,J ⊗ C2), (cid:18)A The top 2 × 2 block of U is the canonical deBranges-Rovnyak colligation with transfer function equal to A. It follows that if the above U is contractive, then its transfer function will have the form: for some non-zero a ∈ Rd(H, J ). We will prove that, in fact, U is an isometry: (6.7) 0 )∗ + Aa∅(Aa∅)∗ X ∗A + K A 0 (K A ∗ ∅(cid:19) . 0 A(0) − Aa2 ∅ A∗A + A(0)∗A(0) + a2 U ∗U =(cid:18)X ∗X + K A By previous formulas the diagonal entries are equal to IH R(A) and IH, respectively, and it remains to show that the top right (and hence also the bottom left) component vanishes. This can be verified as follows: DK A {Z, y, v} f , X ∗AhE =DK A{Z, Z∗y, v}f − A(cid:16)v ⊗ IH, A†(Z)∗y ⊗ f(cid:17), AhE = (cid:16)y ⊗ f, (A†(Z) − A(0n))v ⊗ h(cid:17) −(cid:16)(v ⊗ IH, A†(Z)∗y ⊗ f ), (I − A(0)∗A(0) − a2 ∅)h(cid:17) = (cid:16)y ⊗ f,(cid:16)A†(Z) − A(0n) − A†(Z) + A†(Z) ⊗ (A(0)∗A(0) + a2 ∅)(cid:17) v ⊗ h(cid:17) . On the other hand, (6.8) DK A {Z, y, v} f , K A ∅hE 0 A(0)h − Aa2 = (y ⊗ f, v ⊗ A(0)h) −(cid:16)y ⊗ f, A†(Z)v ⊗ A(0)∗A(0)h(cid:17) −(cid:16)y ⊗ f, A†(Z)v ⊗ a2 ∅h(cid:17) . Adding these expressions together gives 0, which proves the off-diagonal component van- ishes. (cid:3) Corollary 6.8. If B ∈ Ld(H, J ) is column extreme, then it is an extreme point. COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 25 Proof. This is the same contrapositive proof as in [12, Corollary 1.2]: If B is not extreme then there is a non-zero A ∈ Ld(H,J ) so that both B ± A are Schur class, which implies: (M L B−A)∗M L B−A ≤ I, and (M L B+A)∗M L B+A ≤ I. Averaging these inequalities gives: (M L A )∗M L A + (M L B )∗M L B ≤ I, i.e., (cid:18)B A(cid:19) ∈ Ld(H,J ⊗ C2), so that B is not column-extreme. (cid:3) The next two corollaries were established in the proof of (2) ⇒ (1) of Theorem 6.4 above: Corollary 6.9. Given A ∈ Rd(H, J ), define a∅ ∈ L(H)+ by the formula: ∅ := I − A(0)∗A(0) − A∗A ≥ 0. a2 Then for any h ∈ H, Aa∅h ∈ H R(A), so that Aa∅ : H → H R(A). Given A ∈ Rd(H, J ), define DA ⊆ H as the linear space of all h ∈ H such that Ah ∈ H R(A), and let A : DA → H R(A) be the linear transformation Ah := Ah ∈ H R(A). Here, we write A = A(R) so that (Ah)(Z) = A†(Z)h. The previous Corollary 6.9 shows that Ran (a∅) ⊆ Dom( A). Corollary 6.10. Given A ∈ Rd(H,J ), we have the identity: I − X ∗X = K A 0 (K A 0 )∗ + ( Aa∅)( Aa∅)∗. Claim 6.11. The linear transformation A is closed. Proof. Suppose (hn) ⊂ DA, hn → h ∈ H, and Ahn → F ∈ H R(A). It is easy to see that for any Z ∈ Bd n, F (Z) = lim n A†(Z)hn = A†(Z)h, so that F = Ah, proving that h ∈ DA and Ah = F . (cid:3) Since A is a closed linear transformation, it follows by general facts that A∗ A is densely- defined in the Hilbert space DA and positive semi-definite on a domain Dom( A∗ A) ⊆ DA (which is a dense in DA). Proposition 6.12. Given any h ∈ Dom( A∗ A), ∅(I + A∗ A)h = h, a2 and Dom( A) = DA = Ran (a∅). Viewing A as a closed linear transformation from DA ⊆ DA → H, ∅ = (IDA + A∗ A)−1. a2 This is a free and operator-valued analogue of [12, Lemma 3.3]. Proof. Given h ∈ Dom( A∗ A) ⊆ DA, we calculate (I − X ∗X)Ah ∈ H R(A) in two different ways: First, since Ah ∈ H R(A), XAh = Ah, and DK A {Z, y, v} g, (I − X ∗X)AhE =(cid:16)y ⊗ g, A†(Z)v ⊗ h(cid:17) −DK A{Z, Z∗y, v}g, AhE +DA(cid:16)v ⊗ IH, A†(Z)∗y ⊗ g(cid:17), AhE = (y ⊗ g, A(0n)v ⊗ h) +(cid:16)y ⊗ g, A†(Z)v ⊗ A∗Ah(cid:17) . ((I − X ∗X)Ah) (Z) = A(0)(In ⊗ h) + A†(Z)(In ⊗ A∗Ah). Equivalently, for any Z ∈ Bd n, 26 MICHAEL T. JURY AND ROBERT T.W. MARTIN Secondly, apply the identity (6.6) to obtain: (cid:16)K A 0 (K A 0 )∗Ah + Aa∅(Aa∅)∗Ah(cid:17) (Z) Equating these two expressions yields: ∅ A∗ Ah) = K A(Z, 0)A(0)h + A†(Z)(In ⊗ a2 = A(0n)(I ⊗ h)A†(Z)(cid:16)In ⊗ a2 A†(Z)(cid:16)In ⊗ a2 A∗ Ah − In ⊗ A(0)∗A(0)h(cid:17) . A∗ Ah − In ⊗ A(0)∗A(0)h(cid:17) = A†(Z)(In ⊗ A∗Ah). ∅ ∅ Using that A∗A = I − A(0)∗A(0) − a2 ∅ yields: a2 ∅ A∗ Ah − A(0)∗A(0)h = (I − A(0)∗A(0) − a2 ∅)h, and solving for h gives: ∅(I + A∗ A)h = h. a2 On the other hand, Corollary 6.9 shows that Ran (a∅) ⊆ Dom( A) = DA. For any g ∈ Dom( A∗ A) and any h ∈ H, This can be re-arranged as: = (cid:10)g, a2 = (cid:10)g, a2 ∅hE =(cid:10)g, (I − a2 hg, hi = D(I + A∗ A)g, a2 ∅hE ∅h(cid:11) +D A∗ Ag, a2 ∅hE ∅h(cid:11) +D Ag, Aa∅hE . ∅)h(cid:11) , ∅h ∈ Dom( A∗), and that ∅)h, D Ag, Aa2 for any g ∈ Dom( A∗ A) and h ∈ H. Since A is a closed linear transformation, Dom( A∗ A) is a core for A, and the above then implies that Aa2 ∅h = (I − a2 A∗ Aa2 which is equivalent to (I + A∗ A)a2 ∅h = h, and we conclude that a2 ∅ = (I + A∗ A)−1. This proves that Ran(cid:0)a2 ∅(cid:1) = Dom( A∗ A), and by the polar decomposition for closed operators, Ran (a∅) = Dom(pI + A∗ A) = Dom(p A∗ A) = Dom( A). Corollary 6.13. Given A ∈ Rd(H,J ), and F ∈ H R(A), we have that Lj F ∈ H R(A) if j F ∈ Dom( A). and only if A∗ (cid:3) Proof. The proof is formally identical to that of the commutative analogue of this result in [9, Corollary 4.5]. Consider, for g ∈ J , DK A {Z, y, v} g, X ∗ j FE = DK A{Z, Z∗ j y, v}g, FE −DAj(cid:16)v ⊗ IH, A†(Z)∗y ⊗ g(cid:17), FE Now A∗ j F ∈ H R(A) so that also Zj F (Z) ∈ H R(A). Conversely if Zj F (Z) ∈ H R(A), then the above formula shows that AA∗ = (y ⊗ g, Zj F (Z)v) −(cid:16)y ⊗ g, A†(Z)v ⊗ A∗ j F(cid:17) . j F ∈ Ran (a∅), then AA∗ j F ∈ H, so by the previous theorem, if A∗ (cid:3) j F ∈ H R(A). Corollary 6.14. Given A ∈ Rd(H, J ), if Ah ∈ H R(A) for every h ∈ H, then H R(A) is L-invariant. COLUMN EXTREME MULTIPLIERS OF THE FREE HARDY SPACE 27 Corollary 6.15. Given any A ∈ Rd(H, J ), we have that H R(A) is L−invariant if and only if Ran (A∗) ⊆ Ran (a∅) = Dom( A). This happens if and only if A is densely-defined and there is a 0 < r < 1 such that or, equivalently, there is a 0 < ρ < 1 so that r(I − A(0)∗A(0)) ≤ (I + A∗ A)−1, A∗A ≤ ρ(I − A(0)∗A(0)). Proof. By Corollary 6.13, H R(A) is L−invariant if and only if Ran (A∗) ⊆ Ran (a∅). By the Douglas Factorization Lemma, this happens if and only if there is a λ2 > 0 so that Re-arranging gives A∗A∗ ≤ λ2a2 ∅ = λ2(I − A∗A − A(0)∗A(0)). A∗A ≤ λ2 1 + λ2 (I − A(0)∗A(0)). If Dom( A) = Ran (a∅) is not dense, then since Ran (A∗) = Ran (A∗A), we have that there is a non-zero h ∈ H so that a∅h = 0 (recall that a∅ ≥ 0 so that Ran (a∅)⊥ = Ker (a∅)). It then follows that and one cannot have A∗A ≤ λ2a2 dense range. In this case, ∅ in this case. Hence A∗A ≤ λ2a2 ∅ implies Dom( A) has A∗Ah = (I − A(0)∗A(0) − a2 ∅)h 6= 0, I − a2 ∅ − A(0)∗A(0) ≤ λ2a2 ∅ ⇒ I − (I + A∗ A)−1 − A(0)∗A(0) ≤ λ2(I + A∗ A)−1 ⇒ 1 + λ2 (I − A(0)∗A(0)) ≤ (I + A∗ A)−1. 1 (cid:3) 6.16. Clark Intertwining. Fix A ∈ Rd(H). As in [8, Theorem 6.3], one can verify that the weighted Cauchy transform F R : F 2(µA) → H R(A) intertwines the adjoint of the Stinespring-GNS row isometry ΠA = πA(L) with a perturbation of the restricted backward left free shift, L∗ ⊗ IHH R(A) = X (and this is a rank-one perturbation in the case where + (HA) onto H R(A) H = C). Equivalently, the unitary multiplier UA := M R (I−A†(Z)) of H R intertwines the adjoint of the isometry V A with a perturbation of X: 0 )∗ = F RΠ∗ L∗ ⊗ IHH R(A) + A(I − A(0))−1(K A A(F R)∗ = UA(V A)∗U ∗ A. Any U ∈ L(H) yields a different free Clark functional µAU ∗ . Since H R(AU ∗) = H R(A), it follows that every U ∈ L(H) gives a different perturbation of the restricted backward left free shift. In particular, if A is column extreme, each of these perturbations will be a Cuntz unitary (an onto row isometry). In the classical (d = 1, H = C) case, one recovers Clark's perturbations of the backward shift. References [1] J. Agler and J.E. McCarthy. Global holomorphic functions in several non-commuting vari- ables. arXiv:1305.1636, 2013. [2] Joseph A. Ball, Vladimir Bolotnikov, and Quanlei Fang. Schur-class multipliers on the Fock space: de Branges-Rovnyak reproducing kernel spaces and transfer-function realizations. In Operator theory, structured matrices, and dilations, volume 7 of Theta Ser. Adv. Math., pages 85 -- 114. Theta, Bucharest, 2007. [3] Joseph A. Ball, Gregory Marx, and Victor Vinnikov. Noncommutative reproducing kernel Hilbert spaces. J. Funct. Anal., 271(7):1844 -- 1920, 2016. 28 MICHAEL T. JURY AND ROBERT T.W. MARTIN [4] Joseph A. Ball and Victor Vinnikov. Formal reproducing kernel Hilbert spaces: the commu- tative and noncommutative settings. In Reproducing kernel spaces and applications, volume 143 of Oper. Theory Adv. Appl., pages 77 -- 134. Birkhauser, Basel, 2003. [5] Douglas N. Clark. One dimensional perturbations of restricted shifts. J. Analyse Math., 25:169 -- 191, 1972. [6] Kenneth R. Davidson, Jiankui Li, and David R. Pitts. Absolutely continuous representations and a Kaplansky density theorem for free semigroup algebras. J. Funct. Anal., 224(1):160 -- 191, 2005. [7] Kenneth Hoffman. Banach spaces of analytic functions. Dover Publications, Inc., New York, 1988. Reprint of the 1962 original. [8] M. T. Jury and R. T. W. Martin. Non-commutative Clark measures for the free and abelian Toeplitz algebras. J. Math. Anal. Appl., 456(2):1062 -- 1100, 2017. [9] Michael T. Jury. Clark theory in the Drury-Arveson space. J. Funct. Anal., 266(6):3855 -- 3893, 2014. [10] M.T. Jury and R.T.W. Martin. The free Smirnov classes. 2018. In Preparation. [11] M.T. Jury and R.T.W. Martin. Aleksandrov-Clark theory for Drury-Arveson space. Integral Equations and Operator Theory, to appear. arXiv:1608.04325. [12] M.T. Jury and R.T.W. Martin. Extremal multipliers of the Drury-Arveson space. Proceedings of the American Mathematical Society, to appear. arXiv:1608.04327. [13] Dmitry S. Kaliuzhnyi-Verbovetskyi and Victor Vinnikov. Foundations of free noncommuta- tive function theory, volume 199 of Mathematical Surveys and Monographs. American Math- ematical Society, Providence, RI, 2014. [14] R. T. W. Martin. Representation of simple symmetric operators with deficiency indices (1, 1) in de Branges space. Complex Anal. Oper. Theory, 5(2):545 -- 577, 2011. [15] R.T.W. Martin and A. Ramanantoanina. A Gleason solution model for row contractions. arXiv:1612.07972, 2016. [16] Vern I. Paulsen and Mrinal Raghupathi. An introduction to the theory of reproducing ker- nel Hilbert spaces, volume 152 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2016. [17] Gelu Popescu. Entropy and multivariable interpolation. Mem. Amer. Math. Soc., 184(868):vi+83, 2006. [18] Gelu Popescu. Free holomorphic functions on the unit ball of B(H )n. J. Funct. Anal., 241(1):268 -- 333, 2006. [19] Gelu Popescu. Noncommutative transforms and free pluriharmonic functions. Adv. Math., 220(3):831 -- 893, 2009. [20] Gelu Popescu. Free holomorphic functions on the unit ball of B(H )n. II. J. Funct. Anal., 258(5):1513 -- 1578, 2010. [21] Donald Sarason. Sub-Hardy Hilbert spaces in the unit disk, volume 10 of University of Arkansas Lecture Notes in the Mathematical Sciences. John Wiley & Sons, Inc., New York, 1994. A Wiley-Interscience Publication. [22] Donald Sarason. Unbounded operators commuting with restricted backward shifts. Oper. Matrices, 2(4):583 -- 601, 2008. [23] O. Shalit. Operator theory and function theory in Drury -- Arveson space and its quotients. In Handbook of Operator Theory, pages 1125 -- 1180. Springer, 2015. [24] Daniel Su´arez. Closed commutants of the backward shift operator. Pacific J. Math., 179(2):371 -- 396, 1997. University of Florida E-mail address: [email protected] University of Cape Town E-mail address: [email protected]
1312.5195	2	1312	2015-03-30T02:36:20	Strong pure infiniteness of crossed products	[ "math.OA", "math.DS" ]	Consider an exact action of discrete group $G$ on a separable $C^$-algebra $A$. It is shown that the reduced crossed product $A\rtimes_{\sigma, \lambda} G$ is strongly purely infinite - provided that the action of $G$ on any quotient $A/I$ by a $G$-invariant closed ideal $I\neq A$ is element-wise properly outer and that the action of $G$ on $A$ is $G$-separating (cf. Definition 4.1). This is the first non-trivial sufficient criterion for strong pure infiniteness of reduced crossed products of $C^$-algebras $A$ that are not $G$-simple. In the case $A=\mathrm{C}_0(X)$ the notion of a $G$-separating action corresponds to the property that two compact sets $C_1$ and $C_2$, that are contained in open subsets $C_j\subseteq U_j \subseteq X$, can be mapped by elements of $g_j\in G$ onto disjoint sets $\sigma_{g_j}(C_j)\subseteq U_j$, but we do not require that $\sigma_{g_j}(U_j)\subseteq U_j$. A generalization of strong boundary actions on compact spaces to non-unital and non-commutative $C^*$-algebras $A$ (cf. Definition 6.1) is also introduced. It is stronger than the notion of $G$-separating actions by Proposition 6.6, because $G$-separation does not imply $G$-simplicity and there are examples of $G$-separating actions with reduced crossed products that are stably projection-less and non-simple.	math.OA	math	STRONG PURE INFINITENESS OF CROSSED PRODUCTS E. KIRCHBERG AND A. SIERAKOWSKI Abstract. Consider an exact action of discrete group G on a separable C -algebra A. It is shown that the reduced crossed product A ⋊σ,λ G is strongly purely infinite -- provided that the action of G on any quotient A/I by a G-invariant closed ideal I 6= A is element-wise properly outer and that the action of G on A is G-separating (cf. Definition 4.1). This is the first non-trivial sufficient criterion for strong pure infiniteness of reduced crossed products of C -algebras A that are not G-simple. In the case A = C0(X) the notion of a G-separating action corresponds to the property that two compact sets C1 and C2, that are contained in open subsets Cj ⊆ Uj ⊆ X, can be mapped by elements of gj ∈ G onto disjoint sets σgj (Cj ) ⊆ Uj, but we do not require that σgj (Uj) ⊆ Uj. A generalization of strong boundary actions [18] on com- pact spaces to non-unital and non-commutative C -algebras A (cf. Definition 6.1) is also introduced. It is stronger than the notion of G-separating actions by Proposi- tion 6.6, because G-separation does not imply G-simplicity and there are examples of G-separating actions with reduced crossed products that are stably projection-less and non-simple. 5 1 0 2 r a M 0 3 ] . A O h t a m [ 2 v 5 9 1 5 . 2 1 3 1 : v i X r a Contents 1. Introduction 2. Preliminaries 3. Proper outerness and ideal structure 4. Strongly purely infinite crossed products 5. The case of commutative C -algebras 6. Strong boundary actions versus G-separating actions Acknowledgments Appendix A. Date: October 11, 2018. 2010 Mathematics Subject Classification. Primary: 46L35; Secondary: 19K99, 46L80, 46L55. 1 2 4 5 11 14 17 27 27 References 29 1. Introduction In this paper we pursue the study of C -dynamical systems with applications in classification via equivariant KK-theory. It was shown by the first named author that for any two separable nuclear strongly purely infinite C -algebras, both with primitive ideal space isomorphic to a T0-space X, the algebras are isomorphic if and only if they are KKX -equivalent. It is however far from understood when C -algebra crossed products A ⋊σ,λ G associated to C -dynamical systems are strongly purely infinite in terms of properties of the action σ, in particular in the non-simple case. Our main focus of this work is such characterisation for crossed products that are either simple or more generally contain ideals coming from arbitrary G-invariant ideals of the algebra A on which the group G acts. We begin (in Section 2) by introducing crossed products and by proving the notation used throughout the paper. In Section 3 we look at results related to the ideal structure of crossed products. It was shown in [23] that residually properly outer (Definition 4.1) and exact (Definition 3.5) actions σ : G → Aut(A) on a separable C -algebra A have the property that the lattice of (closed) ideals of the reduced crossed product A⋊σ,λ G is naturally isomorphic to the lattice of G-invariant ideals of A (by the map I 7→ A ∩ I). We refine this result by showing that for any exact and residually properly outer action σ of a discrete group G on a separable or commutative C -algebra A the set A+ is a filling family (Definition A.3) for A ⋊σ,λ G (which implies that I 7→ A ∩ I is injective, see Remark A.5 for details). In Section 4 we introduce the notion of G-separating actions (Definition 4.1). We show in Theorem 4.3 that for any exact and residually properly outer action σ of a discrete group G on a separable or commutative C -algebra A and for any filling family F ⊆ A+, the crossed product A ⋊σ,λ G is strongly purely infinite if and only if F has the diagonalization property (Definition A.9) in A ⋊σ,λ G. Applying the work [15] we obtain (in Proposition 4.5) an equivalent characterisation of G-separating actions, from which we can deduce that A+ has the diagonalization property whenever the action on A is G-separating. By evoking [15] once again we prove our main result: 2 Theorem 1.1. Suppose that (A, G, σ) is a C-dynamical system, where G is discrete and A is separable or commutative. If the action σ of G on A is exact (Def. 3.5), residually properly outer (Def. 3.1) and G-separating (Def. 4.1), then A ⋊σ,λ G is strongly purely infinite. In Section 5 we look at actions on commutative C -algebras. Here we characterise the notion of G-separating action purely in terms of the underlying geometry. More specifically we consider actions α of a discrete group G on a locally compact Hausdorff space X, and denote by σ the induced action on A := C0(X). We show (in Lemma 5.1) that the action of G on A is G-separating if and only if the following holds: For every open U1, U2 ⊆ X and compact K1, K2 ⊆ X with K1 ⊆ U1, K2 ⊆ U2, there exist g1, g2 ∈ G such that αg1(K1) ⊆ U1, αg2(K2) ⊆ U2, αg1(K1) ∩ αg2(K2) = ∅. This result is what motives the choice of our terminology "G-separating". As a consequence of this characterisation we obtain in Corollary 5.2 a characterisation of when a crossed product C0(X) ⋊σ,λ G is a strongly purely infinite C -algebra in terms of condition on α. In the final Section 6 we consider actions that produce simple and strongly purely infi- nite crossed products. We introduce (Definition 6.2 and 6.1) the notion of n-majorizing (n ≥ 1) and n-covering actions (n ≥ 2), the later for actions on unital C -algebras. These two notions aim to refine results on simple purely infinite crossed products in [18, 11] where the notion of strong boundary actions (Definition A.1) and n-fillings ac- tions (Definition A.2) was introduced. We prove in Remark 6.7 our notions are weaker: Any n-filling action on a unital C -algebra A is n-covering, and for any action α on a compact spaces X with more than two points (on which strong boundary actions are defined) the action α is a strong boundary action if and only if its adjoint action σ on C(X) is 1-majorizing. Both our notions are G-simple. Therefore we call the 1-majorizing actions on not-necessarily unital or commutative C-algebras also strong boundary actions. Despite our weaker assumptions we are able to prove: Theorem 1.2. Suppose that the C-dynamical system (A, G, σ) with discrete G is n- majorizing (Def. 6.1) for some n ≥ 1 or n-covering (Def. 6.2) for some n ≥ 2, the latter if A is unital. If the action σ is element-wise properly outer (Def. 3.1), and A is separable or commutative, then A ⋊σ,λ G is simple strongly purely infinite. In Section 6 we also look at how the different properties relate to each other. In Lemma 6.3 we show that each n-covering action (for n ≥ 2) on a unital C -algebra 3 A action is n-majorizings, and the latter properly (for n ≥ 1) implies that the action is (n + 1)-covering. In Proposition 6.6 we prove that that any 1-majorizing action on a non-unital C -algebra A is automatically G-separating. In Remark 6.7 we prove that any action on a unital commutative C -algebra A is n-filling if and only if it is n- covering, and for n = 2 this is again equivalent to a strong boundary (i.e., 1-majorizing) action. We end with a number of remarks, including a proof of the fact that our notions of G-separating, n-majorizing and n-covering actions can be expressed in terms of projections when A has real rank zero (see Remark 6.9). We hope that the study of crossed products -- even those for actions of amenable discrete groups on locally compact Polish spaces -- can help to detect possible differences between strong and weak pure infiniteness. This paper is a very first step in this direction, and gives a sufficient criterium by conditions on the action that implies strong pure infiniteness of reduced crossed products. 2. Preliminaries We let A+ denote the set of positive elements in a C -algebra A. We denote the positive and the negative part of a selfadjoint element a ∈ A by a+ := (a + a)/2 ∈ A+ and a− := (a − a)/2) ∈ A+ , where a := (a∗a)1/2. If a ∈ A+, then (a − ε)+, the positive part of a − ε1 in M(A), is again in A+. Here M(A) is the multiplier algebra of A. This notation will be used also for functions f : R → R, then e.g. (f − ε)+(ξ) = max(f (ξ) − ε , 0) . A subset F ⊆ A+ is invariant under ε-cut-downs if for each a ∈ F and ε ∈ (0,kak) we have (a − ε)+ ∈ F . The minimal unitalisation of A is denoted A. Restriction of a map f to X is denoted fX. We let Cc(0,∞]+ denote the set of all non-negative continuous functions ϕ on [0,∞) with ϕ[0, η] = 0 for some η ∈ (0,∞), such that limt→∞ ϕ(t) exists. Remarks 2.1. (i) Suppose that a, b ∈ A+ and ε > 0 satisfy k a − bk < ε . Then the positive part (b−ε)+ ∈ A of (b−ε·1) ∈ M(A) can be decomposed into d∗ad = (b−ε)+ with some contraction d ∈ A ([14, lem. 2.2]). (ii) Let τ ∈ [0,∞) and 0 ≤ b ≤ a + τ · 1 (in M(A)), then for every ε > τ there is a contraction f ∈ A such that (b − ε)+ = f ∗a+f . (See [14, lem. 2.2] and [3, sec. 2.7].) We abbreviate C -dynamical systems by (A, G, σ) with discrete groups G. We de- note by e the unit of G, and consider only closed and two-sided ideals of A. The reduced 4 (resp. the full) crossed product associated to (A, G, σ) is denoted by A ⋊σ,λ G (resp. A ⋊σ G). The norm on A ⋊σ,λ G will be sometimes written as k · kλ if it is necessary to distinguish it from other norms. Let I(A) denote the lattice of ideals in a C -algebra A. The map η : A → A ⋊σ G means the natural embedding into the full crossed prod- uct. Let πλ : A ⋊σ G → A ⋊σ,λ G be the natural epimorphism. We will sometimes suppress the canonical inclusion maps η : A → A ⋊σ G and πλ ◦ η : A ⊆ A ⋊σ,λ G. Let U denote the canonical unitary representation U : G → M(A ⋊σ G). Notice here that the linear span of η(A)U(G) is is a dense -subalgebra of A ⋊σ G. We denote by Uλ : G → M(A ⋊σ,λ G) the regular representation for some more precise explanations. The same happens with ηλ := πλ ◦ η. The set Cc(G, A) consists of the maps f : G → A with finite support F := G\f −1(0). There is a natural linear embedding of Cc(G, A) into A ⋊σ G by canonical identification of f : G → A (of finite support) with an element of A⋊σG : Let F ⊆ G be a finite subset, with f (g) = 0 for g 6∈ F . Then f will be identified with the element Pg∈F η(ag)U(g) of A ⋊σ G , where ag := f (g). In this way Cc(G, A) becomes a -subalgebra of A ⋊σ G that contains A. The natural C -morphism πλ : A ⋊σ G → A ⋊σ,λ G is faithful on Cc(G, A), and we do not distinguish between πλ(X η(ag)U(g)) = X g∈F g∈F ηλ(ag)Uλ(g) and Pg∈F η(ag)U(g). In particular, η(a) ∈ A⋊σ G and ηλ(a) ∈ A⋊σ,λ G will be denoted simply by a ∈ A, and Uλ(g) might be denoted U(g). We now recall the conditional expectation E : A ⋊σ G → η(A) ∼= A : The map Ealg : Cc(G, A) → A , Pg∈F agU(g) 7→ ae , extends by continuity to a faithful con- ditional expectation Eλ : A ⋊σ,λ G → A. In particular Eλ is a completely positive contraction, Eλ(A ⋊σ,λ G) = A, and Eλ(b) = 0 imply b = 0 for b ∈ (A ⋊σ,λ G)+. Since A is also contained in its full crossed product A ⋊σ G, we can use the natural epimorphism A ⋊σ G → A ⋊σ,λ G to define E by E := Eλ ◦ πλ as a (not necessarily faithful) conditional expectation from A ⋊σ G onto its C -subalgebra A. 3. Proper outerness and ideal structure In this section we look at conditions on a C -dynamical system (A, G, σ) ensuring that the set A+ is a filling family for A ⋊σ,λ G in the sense of Definition A.3. This in particular implies that the is a on-to-one correspondence between ideals of A⋊σ,λ G and G-invariant ideals of A, but (as wee shall see) also applies to the verification of when 5 a crossed product is strongly purely infinite. Proper outerness of the automorphisms σt of A defining the action σ turns out also to be a crucial ingredient. We recall the definition below. Definition 3.1. Suppose that (A, G, σ) is a C -dynamical system and that G is discrete. The action σ will be called element-wise properly outer if, for each g ∈ G\{e}, the automorphism σg of A is properly outer in the sense of [6, def. 2.1], i.e., k σgI − Ad(U)k = 2 for any σg-invariant non-zero ideal I of A and any unitary U in the multiplier algebra M(I) of I. See also [19, thm. 6.6(ii)]. We call here an action σ residually properly outer if, for every G-invariant ideal J 6= A of A, the induced action [σ]J of G on A/J is element-wise properly outer. Remarks 3.2. (i) Notice that element-wise proper outerness passes to subgroups, i.e., for each subgroup H ⊆ G the system (A, H, σH) is element-wise properly outer on A if (A, G, σ) is element-wise properly outer. But residual proper outerness does not necessarily pass to subgroups. The system (A, H, σH) is not necessarily residually properly outer, if (A, G, σ) is residual proper outer, because possibly there could be more H-invariant ideals than G-invariant ideals of A. (ii) If A is non-commutative, then topological freeness of (A, G, σ) in sense of [1, def. 1] is -- at least formally -- stronger than the assumption of element-wise proper outerness of (A, G, σ) in Definition 3.1 (cf. [1, prop. 1]). We do not know examples where they actually differ. Thus, for non-commutative A, "essential freeness" of the corresponding action of G on bA in the sense of [23, def. 1.17] (inspired by [22, def. 4.8]) is -- formally -- stronger than our residual proper outerness of (A, G, σ). (iii) If G is countable and acting on C0(X), one can show -- using the Baire property of X -- that elementwise proper outerness is the same as the requirement (for the action α of G on X with σg(f ) := f ◦ αg−1) that points with trivial fix-point subgroup (trivial isotropy) are dense in X, i.e., Definition [23, def. 1.17] holds. We can reformulate this as: stability subgroups of non-empty open subsets are trivial. Remark 3.3. We recall [19, lem. 7.1.] (cf. also [17, lem. 3.2]): If α1, α2, ..., αn are properly outer automorphisms of a separable C-algebra A, there is, for each a0, a1, a2, ..., an ∈ eA, with 0 6= a0 ≥ 0, and each ε > 0, an element x ∈ A+ with kxk = 1 such that kxa0xk > ka0k − ε , kxaiαj(x)k < ε , 6 1 ≤ i, j ≤ n . If A is commutative, i.e., A ∼= C0(X) for X = bA ⊆ A∗, then it is not necessary to suppose that A is separable in the quoted lemma of D. Olesen and G. Pedersen (Compare also [7]): An automorphism σ ∈ Aut(A) is properly outer, if and only if, for every open subset ∅ 6= U ⊆ X there exists y ∈ U with bσ(y) 6= y. Thus, for every finite set S ⊆ Aut(A) of properly outer automorphisms, every non-empty open subset U ⊆ X contains a non-empty open subset V ⊆ U with bσ(V ) ∩ V = ∅ for all σ ∈ S. If one takes U := a−1 0 (ka0 k−ε,∞) and non-empty V ⊆ U as above, then each x ∈ C0(X) with kxk = 1 and support in V satisfies kxa0xk > ka0k − ε and xσ(x) = 0 for σ ∈ S. The following Lemma 3.4 is a suitable modification of proofs of [19, lem.7.1,thm.7.2]. It has been proved in [1] under the stronger assumption that the action σ is topologically free, and part (iii) has been shown in [12, thm. 4.1] even to be equivalent to the topological freeness of the action if A is commutative and unital and G is amenable. Compare also Remark 5.3 for a "residual" version. Lemma 3.4. Suppose that A is separable or commutative, and that the action of G on A is element-wise properly outer. (i) For every b ∈ (A ⋊σ G)+ with E(b) 6= 0 and ε > 0 there exist x ∈ A+ satisfying that kxk = 1 , kxbx − xE(b)xk < ε , kxE(b)xk > kE(b)k − ε . This holds also for b ∈ (A ⋊σ,λ G)+ and Eλ in place of E. (ii) If h : A ⋊σ G → L(H) is a -representation such that hA is faithful, then kh(b)k ≥ kE(b)k for all b ∈ (A ⋊σ G)+. In particular, the kernel of h is contained in the kernel Iλ of the natural epimorphism πλ : A ⋊σ G → A ⋊σ,λ G. (iii) Every non-zero ideal of A ⋊σ,λ G, has non-zero intersection with A. Proof. (i): Let b ∈ (A ⋊σ G)+ with E(b) 6= 0, and ε > 0. Let a0 := E(b). Since Cc(G, A) is dense in A ⋊σ G, there exists a′ = c0 + Pm j=1 ajU(gj) ∈ Cc(G, A) with g−1 6= e for i 6= j ∈ {1, . . . , m}, i gj and ka′ − bk < ε/6. Since E is a contraction, it follows that kb − ak < ε/3 and E(a) = a0 = E(b) for g0 := e and a := a0 + (a′ − c0) = Pm j=0 ajU(gj). By [19, lem. 7.1] and Remark 3.3 there exists x ∈ A+ with kxk = 1, kxE(a)xk > kE(a)k − ε/3m, and kxajσgj (x)k < ε/3m for gj 6= e, j = 1, . . . , m . In particular, 6= e and gj kxE(b)xk = kxE(a)xk > kE(a)k − ε = kE(b)k − ε . 7 Since kxajU(gj)xk = kxajσgj (x)k we get in A ⋊σ G that kxajσgj (x)k ≤ ε/3 . kx(a − E(a))xk ≤ X gj6=e Thus, in the full crossed product A ⋊σ G we have kxbx − xE(b)xk ≤ kx(b − a)xk + kx(a − E(a))xk + kx(E(a) − E(b))xk < ε . The same arguments work for b ∈ (A ⋊σ,λ G)+ and Eλ in place of E. (ii): The restriction of h to A ⊆ A ⋊σ G is faithful, hence kh(a)k = kak for all a ∈ A. Let b ∈ (A ⋊σ G)+ be given. If E(b) = 0 then kh(b)k ≥ kE(b)k. If E(b) 6= 0, then select x ∈ A+ as in (i). It follows that kh(xE(b)x)k = kxE(b)xk ≥ kE(b)k − ε . On the other hand, kh(b)k ≥ kh(x)h(b)h(x)k = kh(xbx)k and ε > kxbx − xE(b)xk ≥ kh(xbx) − h(xE(b)x)k. Thus kh(b)k + ε ≥ kh(xE(b)x)k, and kh(b)k + 2ε ≥ kE(b)k for all ε > 0. Since E = Eλ ◦ πλ, we have b ∈ (A ⋊σ G)+ and E(b) = 0 implies that b is contained in the kernel of πλ : A ⋊σ G → A ⋊σ,λ G . In particular, if h : A ⋊σ G → L(H) is any -representation with kh(b)k ≥ kE(b)k for all b ∈ (A ⋊σ G)+, then the kernel h−1(0) of h is contained in the kernel of the natural epimorphism πλ : A ⋊σ G → A ⋊σ,λ G. (iii): Let I a closed ideal of A ⋊σ,λ G with I∩A = {0}, consider (A ⋊σ,λ G)/I as a C - subalgebra of some L(H), and let h : A ⋊σ G → L(H) the corresponding representation with kernel h−1(0) = J := π−1 λ (0). Then h is faithful on A and, therefore, satisfies π−1 (cid:3) λ (0) ⊇ h−1(0). It follows I = πλ(h−1(0)) = {0}. λ (I) ⊇ π−1 Definition 3.5 ([23, def. 1.2]). Suppose that (A, G, σ) is a C -dynamical system with locally compact G. The action σ of G on A is exact, if, for every G-invariant ideal J in A, the sequence 0 → J ⋊σJ,λ G → A ⋊σ,λ G → A/J ⋊[σ]J ,λ G → 0 is short-exact. Remarks 3.6. (i) Recall that a locally compact group G is exact, if and only if, every action σ : G → Aut(A) is exact. If G is discrete, then this is equivalent to the exactness of the C -algebra C ∗ λ(G), cf. [16]. This applies to all amenable groups G, e.g. G = Z. Under Definition 3.5 each minimal (= G-simple) action is exact. In particular, non- exact discrete groups can have exact (and faithful) actions. (ii) Let F denote the (small) Thompson group and ρ : F → Homeo(R) the minimal action of F (or only of its commutator subgroup F ′) on the real line R as described by Haagerup and Picioroaga in [10, rem. 2.5.]. One can use ρ to construct a F - separating, non-minimal and exact action α of F (or F ′) on the disjoint union of two 8 lines X := R ∪ (i + R) ⊆ C if one considers the restriction α(g) := β(g)X to X of the action g ∈ F → β(g) on C given by β(g)(s + it) := ρ(g)(s) + it ( 1 ). It is at present unknown whether the Thompson group F is exact or not, cf. [2, 8, 9]. (iii) It is not known if Gromov's examples of non-exact groups can have non-exact actions on commutative C -algebras. It is likely that it has to do with still missing non-trivial geometric conditions for G-actions on locally compact spaces X that are equivalent to the exactness of the adjoint action σ : G → Aut(C0(X)) given by σg(f ) := f ◦ αg−1. Therefore we use the trivial and non-geometric definition and define α to be exact if its adjoint action σ on C0(X) is exact. Remark 3.7. Combination of Lemma 3.4(iii) and of the exactness of an action σ : G → Aut(A) on a separable or commutative C -algebra A shows that the lattice of (closed) ideals of the reduced crossed product A ⋊σ,λ G is naturally isomorphic to the lattice of G-invariant ideals of A (by the map J 7→ A ∩ J), if σ is exact and residual properly outer. (See [23, Remark 2.23] for details.) Theorem 3.8. Let (A, G, σ) a C-dynamical system, with discrete G and separable or commutative A. If the action σ of G on A is exact and residually properly outer, then the elements of A+ build a filling family for A ⋊σ,λ G in the sense of Definition A.3. Proof. We show that for every hereditary C -subalgebra D of A ⋊σ,λ G and every (closed) ideal I of A ⋊σ,λ G with D 6⊆ I there exist f ∈ A+ and z ∈ A ⋊σ,λ G such that z∗z ∈ D and zz∗ = f 6∈ I. Suppose that D is a hereditary C * -- subalgebra of A ⋊σ,λ G and that I is an ideal of A ⋊σ,λ G with D 6⊆ I. Let J := I ∩ A, then J is a G-invariant ideal of A with J ⋊σJ,λ G ⊆ I and g ∈ G 7→ [σg]J is an element-wise properly outer action on A/J by our assumptions on σ. We denote this action by α, i.e., αg(a + J) := σg(a) + J. By Remark 3.7, the exactness and residual proper outerness of σ : G → Aut(A) allow a natural identification (A ⋊σ,λ G)/I = (A/J) ⋊α,λ G . Since D 6⊆ I implies D+ 6⊆ I, there exists d ∈ D+, d /∈ I. The epimorphism πI : A ⋊σ,λ G → (A ⋊σ,λ G)/I is equal to the quotient map πJ from A ⋊σ,λ G onto (A/J)⋊α,λ G ∼= (A⋊σ,λ G)/I (under natural identifications). We denote the conditional 1 This action is not topologically free. 9 expectation Eλ : (A/J) ⋊α,λ G → A/J (temporarily) by E and define b := πI(d) , and ε := 1 4 kE(b)k > 0 . Lemma 3.4(i) gives an element x ∈ (A/J)+ such that kxk = 1, kxbx − xE(b)xk < ε, kxE(b)xk > kE(b)k − ε = By Remark 2.1(i), there is a contraction y ∈ (A/J) ⋊α,λ G such that 3 4 kE(b)k . y∗xbxy = (xE(b)x − ε)+ ∈ (A/J)+ . Note that y∗xbxy 6= 0 , because k(xE(b)x − ε)+k = kxE(b)xk − 1 4 kE(b)k > 1 2 kE(b)k = 2ε . Since πIA = πJA and (xE(b)x − ε)+ ∈ (A/J)+, there is exists c ∈ A+ such that πJ(c) = (xE(b)x − ε)+. Since πJ (= πI ) is surjective, there exists a contraction w ∈ A ⋊σ,λ G with πJ (w) = xy. We obtain that c = w∗dw + v for some v ∈ I. The set Cc(G, J) is dense in I, because I = J ⋊σ,λ G and G is discrete. This allows us to see, that JIJ is dense in I. It follows that {e ∈ J+ ; kek < 1} is an approximate unit of I. In particular, there exists e ∈ J+ with kv − evk < ε. Let 1 denote the unit of eA ⋊σ,λ G, then A ⋊σ,λ G is an ideal of eA ⋊σ,λ G. With g := (1 − e) ∈ eA+, kgk ≤ 1 we get kgw∗dwg − gcgk = kgvgk ≤ kv − evk < ε = 1 4 kE(b)k . Since gzg = z + eze − (ze + ez) and πI (e) = πJ (e) = 0, we have πI(gzg) = πI(z) for all z ∈ A ⋊σ,λ G. By Remark 2.1(i), there exists a contraction h ∈ A ⋊σ,λ G such that h∗(gw∗dwg)h = (gcg − ε)+ ∈ A+ . With z := (d1/2wgh)∗ we have that z∗z ∈ D and zz∗ = (gcg − ε)+ =: f ∈ A+ . Finally, we see from πI (gcg) = πI(c) that kπI(f )k = kπI((gcg) − ε)+)k = (kπI(gcg)k − ε)+ = (kπI(c)k − ε)+ = (k(xE(b)x − ε)+k − ε)+ = kxE(b)xk − Hence, f 6∈ I. 10 1 2kE(b)k > 1 4kE(b)k > 0 . (cid:3) 4. Strongly purely infinite crossed products In this section we prove out main result Theorem 1.1. We start with the definition of an G-separating action. Definition 4.1. Suppose that (A, G, σ) is a C -dynamical system with discrete group G. The action of G on A is G-separating if for every a, b ∈ A+ , c ∈ A, ε > 0, there exist elements s, t ∈ A and g, h ∈ G such that k s∗a s − σg(a)k < ε , k t∗b t − σh(b)k < ε and k s∗c tk < ε . (1) Remarks 4.2. (i) Notice that Definition A.6 and Remark A.7 immediately implies that every action σ : G → Aut(A) is G-separating if A itself is strongly purely infinite: Take h = g = e ∈ G. If the contractions s, t ∈ A satisfy the defining inequalities (7) of strongly p.i. algebras A then they also satisfy the inequalities (1). (ii) G-separating actions on a locally compact space X are not necessarily minmal. One can show that above mentioned example of an exact and non-minimal action of the (small) Thompson group F on two parallel lines R∪ (i + R) ⊆ C is also F -separating. (iii) The existence of a G-separating action on A imposes requirement on A itself, e.g. the cases a = b = c = p and a = b = c = 1 with ε = 1/3 in inequalities (1) show that A can not contain minimal non-zero projections p ∈ A and that 1A must be properly infinite in A if A is unital. Therefore, C -algebras, that are commutative and unital, can not have any G-separating actions. (iv) Further variations of the concepts that we introduce here are possible, e.g. one could start with conditions that are weaker than conditions for G-separating actions. Also one could require the existence of n ∈ N such that for a, b ∈ A+ and ε > 0 there is a solution d1, . . . , dn ∈ A and g1, . . . , gn ∈ G of the inequality (3) in Definition 6.1 of n- majorizing actions whenever b is in the smallest G-invariant closed ideal that contains a. Or one could attempt to replace the filling family F := A+ by smaller filling families F ⊆ A+ and require more elaborate local matrix diagonalization formulas involving also G-translates, cf. Definition A.8. Combing Theorem 3.8 with Theorem A.13 we obtain the following result Theorem 4.3. Let G be a discrete group acting by σ : G → Aut(A) on a separable or commutative C-algebra A. Suppose that the action is residually properly outer 11 (cf. Def. 3.1) and exact (cf. Def. 3.5). Let F ⊆ A+ be a filling family for A. Then the following are equivalent: (i) The crossed product A ⋊σ,λ G is strongly purely infinite. (ii) The family F has the diagonalization property in A ⋊σ,λ G. Proof. Let B := A ⋊σ,λ G. The assumptions ensure that A+ is a filling family for B by Theorem 3.8. Since F is filling for A, F is also filling for B by Lemma A.4. (i)⇒(ii): If B is s.p.i., then B+ has the diagonalization property (see Definition A.9) in B, cf. [14, lem. 5.7]. This implies that our family F ⊆ A+ ⊆ B+ has the diagonalization property in B. (ii)⇒(i): Since our family F ⊆ A+ is filling for B, and since F has the diagonalization property in B, we get that B is s.p.i. by Theorem A.13. (cid:3) Remark 4.4. Let (A, G, σ) a C -dynamical system. (i) For each a1, a2 ∈ A+, x, d1, d2 ∈ A, g0, g1, g2 ∈ G and s1 := d1U(g1), s2 := 0 g2g2), c := xU(g0), b1 := a1, and b2 := σg0(a2) the following (d2)U(g−1 σg−1 equalities hold: 0 ks∗ j ajsj − ajk = kd∗ j bjdj − σgj (bj)k and ks∗ 1cs2k = kd∗ 1xd2k . (ii) With g0 = e in (i) the equalities reduce to: ks∗ j ajsj − ajk = kd∗ j ajdj − σgj (aj)k and ks∗ 1cs2k = kd∗ 1cd2k . Proposition 4.5. Suppose that (A, G, σ) is a C-dynamical system with discrete G. The following properties (i) -- (ii) are equivalent: (i) The action of G on A is G-separating in the sense of Definition 4.1. (ii) For every a1, a2 ∈ A+, c ∈ A ⋊σ,λ G and ε > 0 there exist d1, d2 ∈ A and g1, g2 ∈ G such that the elements sj = djU(gj) of Cc(G, A) satisfy, for j = 1, 2, (2) k s∗ jajsj − aj k < ε , and k s∗ 1cs2 k < ε . Proof. (ii)⇒(i): If we take c ∈ A, a1 := a and a2 := b for a, b ∈ A+ and ε > 0, then (ii) implies, using Remark 4.4, that there exist d1, d2 ∈ A and g1, g2 ∈ G such that kd∗ 1cd2k < ε, so the inequalities (1) of Definition 4.1 are satisfied with d1, d2, g1, g2 in place s, t, g, h. j ajdj − σgj (aj)k < ε and kd∗ 12 (i)⇒(ii): Define C := {dU(g) ; d ∈ A , g ∈ G} and S := C . Select any ε0 > 0. Clearly, the closed linear span of C is equal to A ⋊σ,λ G. If we can show that F := A+ , C and S satisfy the assumptions (i)-(iii) of Lemma A.11 -- with A ⋊σ,λ G in place of A -- , then it follows from Lemma A.11 that for every a1, a2 ∈ A+, c ∈ A ⋊σ,λ G and ε > 0 there exist d1, d2 ∈ A and g1, g2 ∈ G such that sj = djU(gj) ∈ S fulfil (2), which in turn gives (ii). It is evident that our C and S satisfy properties (ii) and (iii) in Lemma A.11. Since A+ is closed under ε-cut-downs, property (i) becomes automatic if each pair (a1, a2), with a1, a2 ∈ A+, has the matrix diagonalization property of Definition A.8 with respect to S and C: If a1, a2 ∈ A+, c = xU(g0) ∈ C with x ∈ A, g0 ∈ G, and ε > 0 are given, then we define b1 := a1, b2 := σg0(a2). (If we instead of ε are given positive ε1, ε2 and τ , set ε := min(ε1, ε2, τ ).) Since the action σ is G-separating, we can find d1, d2 ∈ A and g1, g2 ∈ G with kd∗ 1xd2k all strictly below ε. Remark 4.4 provides elements sj ∈ C satisfying (2). Thus (a1, a2) has the matrix diagonalization property with respect to S and C. 2b2d2 − σg2(b2)k and kd∗ 1b1d1 − σg1(b1)k, kd∗ (cid:3) Theorem 4.6. Let (A, G, σ) a C-dynamical system, with discrete G. Suppose that A+ is a filling family for A ⋊σ,λ G and that the action of G on A is G-separating. Then A ⋊σ,λ G is strongly purely infinite. Proof. By Theorem A.13 it remains to show that A+ has the diagonalization property in A ⋊σ,λ G. Since A+ is closed under ε-cut-downs Lemma A.10 applies, and therefore it is enough to show that each pair (a1, a2) with a1, a2 ∈ A+, has the matrix diagonal- ization property in A ⋊σ,λ G. But this follows from the G-separation property of the action σ by Proposition 4.5. (cid:3). Proof of Theorem 1.1: By Theorems 3.8 and 4.6 the assumptions imply that A ⋊σ,λ G is strongly purely infinite. (cid:3) Remark 4.7. Suppose that (A, G, σ) is a C -dynamical system and that G is discrete. Then a family F ⊆ A+ ⊆ A ⋊σ,λ G which is invariant under ε-cut-downs has the diagonalization properly in A ⋊σ,λ G , if and only if, for every a1, a2 ∈ F , c ∈ Cc(G, A) and > 0 , there exist s1, s2 ∈ Cc(G, A) such that, for j = 1, 2, k s∗ jajsj − aj k < ε and 13 k s∗ 1cs2 k < ε . This follows from Lemma A.10, Lemma A.11 and the fact that Cc(G, A) is dense in A ⋊σ,λ G . Remark 4.8. Notice that for an exact locally compact group G the reduced group C - algebra C ∗ λ(G) is an exact C -algebra (cf. [16, p. 171]). By Theorem A.12, the minimal C -tensor product A ⊗min B of a s.p.i. C -algebra A with an exact C -algebra B is again s.p.i. Hence, if G is an exact locally compact group, σ(g) := idA is the trivial action on a s.p.i. C -algebra A then A ⋊σ,λ G ∼= A ⊗min C ∗ λ(G) is s.p.i. This shows that there is room for refinements of our sufficient conditions on the actions that imply strong pure infiniteness of the reduced crossed products: Here the action σ is even not element-wise properly outer, but satisfies the G-separation property and is an exact action by C -exactness of C ∗ λ(G). 5. The case of commutative C -algebras The case of G-actions on commutative C -algebras allows some topological interpre- tation. The next lemma has inspired our choice of the name G-separating in Definition 4.1. Notice that we do not require αgj (Uj) ⊆ Uj in its part (ii). Lemma 5.1. Suppose that (A, G, σ) is a C-dynamical system, that A ∼= C0(X) is commutative, and that the action σ of G on C0(X) is induced by the action α of G on for f ∈ A, g ∈ G). Then the following properties are X ∼= bA (i.e., σg(f ) := f ◦ α−1 g equivalent: (i) The action of G on A is G-separating, i.e., for every a, b ∈ A+ , c ∈ A, ε > 0, there exist elements d1, d2 ∈ A and g1, g2 ∈ G such that k d∗ 1ad1 − σg1(a)k < ε , k d∗ 2bd2 − σg2(b)k < ε and k d∗ 1cd2 k < ε . (ii) For every open U1, U2 ⊆ X and compact K1, K2 ⊆ X with K1 ⊆ U1, K2 ⊆ U2, there exist g1, g2 ∈ G such that αg1(K1) ⊆ U1, αg2(K2) ⊆ U2, αg1(K1) ∩ αg2(K2) = ∅ . Proof. (ii)⇒(i): Let a, b ∈ A+, c ∈ A and ε > 0. We use assumption (ii) on U1 := a−1(ε/4,∞) = {x ∈ X ; a(x) > ε/4} , U2 := {x ∈ X ; b(x) > ε/4} , K1 := {x ∈ X ; a(x) ≥ ε/2} , K2 := {x ∈ X ; b(x) ≥ ε/2} , 14 and find g1, g2 ∈ G such that αg1(K1) ⊆ U1 , αg2(K2) ⊆ U2 , αg1(K1) ∩ αg2(K2) = ∅ . Since a, b ∈ C0(X)+, we have that U1 ⊆ a−1[ε/4,∞) and U2 ⊆ b−1[ε/4,∞) are compact subsets of X. Since the compact sets αg1(K1) and αg2(K2) are disjoint, applications of Tietze extension theorem gives elements e1, e2 ∈ A+ with kejk ≤ 2/√ε and a contraction f = f ∗ ∈ A such that e1U1 = a−1/2U1 , e2U2 = b−1/2U2 , fαg1(K1) = −1 , fαg2(K2) = 1 . Let f+, f− ∈ A+ be the canonical decomposition f = f+ − f− with f+f− = 0. We define d1 := e1(σg1(a) − ε/2)1/2 1cd2 = 0 because f+f− = 0. Then d∗ + f− and d2 := e2(σg2(b) − ε/2)1/2 + f+ . Since (σg1(a) − ε/2)+(x) 6= 0 implies a(αg−1 (x) ∈ K1, and x ∈ αg1(K1) ⊆ U1 ⊆ U1. It implies that f−(x) = 1 and e1(x) = a−1/2(x). We obtain that (x)) > ε/2, we get αg−1 1 1 1ad1 = e2 d∗ 1a(σg1(a) − ε/2)+(f−)2 = (σg1(a) − ε/2)+ . In a similar way we see that d∗ 2bd2 = (σg2(b) − ε/2)+. (i)⇒(ii): Let U1, U2 ⊆ X be open and K1, K2 ⊆ X compact subsets such that K1 ⊆ U1 and K2 ⊆ U2. We can assume that the intersection K1 ∩ K2 is non-empty. There exists an open set W with a compact closure W such that K1 ∪ K2 ⊆ W ⊆ W ⊆ U1 ∪ U2 . By the Tietze extension theorem, there are contractions a, b, c ∈ A+ such that aK1 = 1 , bK2 = 1 , cW = 1 , supp(a) ⊆ U1 ∩ W, supp(b) ⊆ U2 ∩ W, supp(c) ⊆ U1 ∪ U2 . We apply assumption (i) to a, b, c and ε := 1/4, and get elements d1, d2 ∈ A and g1, g2 ∈ G such that kd∗ 1ad1 − σg1(a)k < 1/4 , kd∗ 2bd2 − σg2(b)k < 1/4 , kd∗ 1cd2k < 1/4 . If x ∈ W , then c(x) = 1 and d1(x)d2(x) ≤ kd∗ 1cd2k < 1/4. Thus, V1 := {x ∈ W ; d1(x) ≥ 1/2} , V2 := {x ∈ W ; d2(x) ≥ 1/2} 15 1 1 (x) ∈ K1 and σg1(a)(x) = a(αg−1 are disjoint sets. If x ∈ αg1(K1), then αg−1 (x)) = 1. It follows d1(x)2a(x) ≥ 3/4 and d1(x) > 1/2 (the latter because 1 ≥ a(x) > 0). Thus, x ∈ U1∩ W and x ∈ V1. It shows αg1(K1) ⊆ U1∩ V1. In a similar way we get αg2(K2) ⊆ U2 ∩ V2. It implies αg1(K1) ⊆ U1 , αg2(K2) ⊆ U2 and that αg1(K1) ∩ αg2(K2) = ∅ . (cid:3) The following condition (i) in Corollary 5.2 is satisfied if the action α has the residual version of the topological freeness in sense of [1, def. 1], see e.g. the essential freeness defined in [23, def. 1.17] (inspired by [22, def. 4.8]) when G is countable. Corollary 5.2. Let G be a discrete group, α : G → Homeo(X) an action of G on a locally compact Hausdorff space X. Suppose that (i) for every closed G-invariant subset Y of X and every e 6= g ∈ G the set {y ∈ Y : αg(y) = y} has empty interior, (ii) the action σ : G → Aut(C0(X)), given by σg(f ) := f ◦ (αg)−1, is exact on the C-algebra C0(X), and (iii) the action α is G-separating, i.e., by Lemma 5.1, for every U1, U2 ⊆ X open and K1, K2 ⊆ X compact such that K1 ⊆ U1, K2 ⊆ U2, there exist g1, g2 ∈ G such that αg1(K1) ⊆ U1, αg2(K2) ⊆ U2, αg1(K1) ∩ αg2(K2) = ∅ . Then C0(X) ⋊σ,λ G is a strongly purely infinite C-algebra. Proof. Let A := C0(X). It is easy to see that property (i) implies that the action on any quotient A/I by a G-invariant closed ideal I 6= A is element-wise properly outer. Now Theorem 1.1 applies to A ⋊σ,λ G. (cid:3) The following remark shows that in case of commutative A and discrete amenable G several of the previously considered properties are equivalent. Remark 5.3. If A is commutative and G is a discrete amenable group that acts on A by σ, then the following properties are equivalent: (i) A separates the ideals of A ⋊σ G, i.e., I 7→ A ∩ I is an injective map from I(A ⋊σ G) into I(A) (see [23, def. 1.9]). (ii) The action σ : G → Aut(A) is residually properly outer (Definition 3.1). (iii) The family F := A+ is filling for A ⋊σ G (Definition A.3). Proof. (i)⇒(ii): By [12, thm. 4.1] (in case of unital A, and [1, thm. 2] for the general -- non-unital -- case) the separation property implies that the adjoint action of G on the 16 Gelfand spectrum of A is topologically free, which is equivalent to element-wise proper outerness by [1, prop. 1]. This applies also to the quotients (A/J) ⋊[σ]J ,λ G, because the property (i) passes to quotients by amenability of G. See also [23, thm. 1.13]. (ii)⇒(iii): Since amenable G are exact, the residual proper outerness of the action implies that F := A+ is filling for A ⋊σ,λ G by Theorem 3.8. (iii)⇒(i): By Remark A.5, the subalgebra A separates the ideals of B := A ⋊σ,λ G if (cid:3) F := A+ is filling for B. Asking G to be amenable can be weakened to exactness of σ and A ⋊σ,λ G ∼= A ⋊σ G. One might also expect nuclearity of A ⋊σ,λ G would suffice in place of amenability of G (this is know at least in the unital case). 6. Strong boundary actions versus G-separating actions In this section we prove our Theorem 1.2. We state with the definition of n- majorizing and n-covering actions. Definition 6.1. Let n ∈ N and A a non-zero C -algebra, that is not isomorphic to a subalgebra of Mn+1(C) if A is unital. An action σ : G → Aut(A) will be called an n-majorizing action of G on A, if, for every non-zero a ∈ A+, every non-invertible b ∈ A+ and every ε > 0, there exist d1, . . . , dn ∈ A and g1, . . . , gn ∈ G such that nX j=1 k d∗ j σgj (a) dj − bk < ε . (3) Definition 6.2. Let n ∈ N, n ≥ 2 ( 2 ). Suppose that A is a unital C -algebra, that is not isomorphic to a -subalgebra of Mn(C). An action σ of G on A is an n-covering action if, for every non-zero a ∈ A+, and every ε > 0, there exist d1, . . . , dn ∈ A and g1, g2, . . . , gn ∈ G and such that nX (4) d∗ j σgj (a) dj − 1 k < ε . k j=1 The following lemma denies the existence of non-zero "socles" in C -algebras A that admit n-majorizing or n-covering actions considered in Definitions 6.1 and 6.2. 2 If n = 1 then property (4) holds if and only if A is a unital simple purely infinite C -algebra. 17 Lemma 6.3. Let (A, G, σ) C-dynamical system. Consider the following properties (α) or (β) of (A, G, σ) depending on n ∈ N: (α) There is n ≥ 1 such that, for each non-zero a ∈ A+, non-invertible b ∈ A+ and ε > 0, there exist d1, . . . , dn ∈ A and g1, . . . , gn ∈ G that satisfy the inequality (3) in Definition 6.1. (β) A is unital, and there is n ≥ 2 such that, for each non-zero a ∈ A+ and ε > 0, there exist d1, . . . , dn ∈ A and g1, . . . , gn ∈ G that satisfy the inequality (4) in Definition 6.2. If A is unital and (A, G, σ) satisfies (α) then it satisfies (β) with n replaced by n+1, and if (A, G, σ) satisfies (β) then it satisfies (α) -- with same n ∈ N. If (A, G, σ) satisfies (α) or (β), then the algebra A is G-simple, i.e., A and {0} are the only G-invariant closed ideals of A. If A contains a projection p 6= 0 with pAp = C· p, then A is a C-subalgebra of Mn+1 (respectively of Mn) if (A, G, σ) has property (α) (respectively has property (β)). The shift action σ of the cyclic group Zn+1 on A := C(Zn+1) satisfies (α) for n ∈ N and is element-wise properly outer. Proof. If A is unital and σ has property (α), then let b := 1 − (kak−1a)3, and find d1, . . . , dn and g1, . . . , gn that satisfy the inequality (3). If we let gn+1 := e and dn+1 := kak−3/2a, then a and g1, . . . , gn, gn+1 satisfy (4). It shows that actions on unital A that satisfy property (α) are also actions that satisfy (β) with n + 1 in place of n. If (A, G, σ) satisfies (β) and non-zero elements a, b ∈ A+ are given with kbk = 1, then d1b1/2, . . . , dnb1/2 and g1, . . . , gn is a solution of the inequality (3) in Definition 6.1 if d1, . . . , dn and g1, . . . , gn satisfy the inequality (4) of Definition 6.2. The properties (α) and (β) imply that {0} and A are the only G-invariant closed ideals of A : If A is non-unital in case (α), then the approximation, as expressed by the inequalities (3), shows that for each non-zero a ∈ A+ the smallest closed G-invariant ideal of A containing a contains each b ∈ A+. If A is unital and the actions has property (β) then each G-invariant closed ideal of A contains 1. If A is unital and the C-dynamical system (A, G, σ) satisfy property (α) then it satisfies property (β) with n replaced by n + 1. Thus, again, A and {0} are the only closed G-invariant ideals. 18 From now on, we suppose that there exists a projection 0 6= p ∈ A+ with pAp = Cp. We call those projections "minimal", even if minimal non-zero projections of a C - algebra A do not have the property pAp = Cp in general, e.g. the unit of the Jiang-Su algebra Z is a minimal projection. We show that this assumption of the existence of such p ∈ A, together with the assumption that σ satisfies (α), implies that A is unital. Thus A satisfies (β) with n + 1 in place of n . Then we derive that property (β) and the existence of such p ∈ A imply that A is a C -subalgebra of Mn. It is obvious that the ideal socle(A) generated by those "minimal" projections is (universally) invariant under all automorphisms of A, i.e., α(socle(A)) = socle(A) for all α ∈ Aut(A). This happens also for the closure J of socle(A). Thus, J must be G-invariant. It follows that J = A using socle(A) 6= ∅. It is not difficult to see, that J is isomorphic to the c0-direct sum of a family of algebras K(Hτ ) of compact operators on suitable Hilbert spaces Hτ , and that p is a rank-one projection on some Hτ0. Let H denote the Hilbert space sum of the Hilbert spaces Hτ . Then A becomes isomorphic to a non-degenerate C -subalgebra of the alge- bra of compact operators K(H) on H, in a way that each minimal non-zero projection p ∈ A becomes a rank one projection on H. This happens also for all σg(p). Recall that every projection in A ⊆ K(H) has finite rank in L(H). Since A is a C -subalgebra of K(H), A is in particular an AF-algebra, and -- therefore -- contains an approximate unit (qτ ) consisting of an upward directed net of projections of finite rank in H. We show that A must be unital if (A, G, σ) satisfies (α) in addition: Suppose that A is not unital, then none of the projections (qτ ) are invertible in A. Therefore, we can take b := qτ , a := p and ε = 1/2 in (3). It follows that each qτ has linear rank ≤ n. This implies that the approximate unit (qτ ) must be constant qτ = e for all τ ≥ τ1 with suitable τ1. Then e ∈ A is necessarily the unit element of A, in contradiction to our assumption that A is not unital. It follows that A must be unital, and -- as above observed -- the action σ satisfies property (β) with n replaced by n + 1. If A is unital and (A, G, σ) satisfies property (β), then we take a := p and ε := 1/2 in inequalities (4). It shows that the rank of 1A in its representation is ≤ n. Thus A is a C -subalgebra of Mn in case (β). The crossed product C(Zn+1) ⋊ Zn+1 is naturally isomorphic to Mn+1. Hence, by If a ∈ Remark 5.3, the action of Zn+1 on C(Zn+1) is element-wise properly outer. 19 A+ := C(Zn+1)+ is non-zero and b ∈ A+ is not invertible, then there are non-zero minimal projections p, q ∈ A+ and δ > 0 such that a ≥ δp and b ≤ kbk · (1 − q). Select g1, . . . , gn ∈ Zn+1 with Pn j=1 σgj (a) ≥ 1 − q. Thus, there exists T ∈ (1 − q)A+(1 − q) with T (Pn j=1 σgj (a))T = 1 − q. Then a, b, j=1 σgj (p) = 1 − q. It implies that δ−1Pn dj := T b1/2, j = 1, . . . , n and g1, . . . , gn satisfy the inequality (3) for each ε > 0. (cid:3) Remark 6.4. Let B be a non-zero simple C -algebra. In preperation for the proof of Theorem 1.2 we display here a number of properties equivalent to strong pure infinite- ness of B. (i) B is strongly purely infinite. (ii) Each non-zero element of B+ is properly infinite in sense of [13]. (iii, n) There exists n ∈ N such that, for each non-zero elements a, b ∈ B+ and ε > 0, there exists n elements d1, . . . , dn ∈ B with kd∗ 1ad1 + · · · + d∗ nadn − bk < ε , (5) and B is not isomorphic to Mk for any k ≤ n. (iv) B is locally purely infinite in sense of [3, def. 1.3], i.e., each non-zero hereditary C -subalgebra of B contains a non-zero stable C -subalgebra. (v) B is purely infinite in the sense of J. Cuntz [5, p. 186], i.e., each non-zero hereditary C -subalgebra contains an infinite projection. Proof. Property (ii) implies (i) by [3, thm. 5.8] and (i) implies (ii) by [14, prop. 5.4]. Property (iii,n = 1) is equivalent to (ii) by [13, thm. 4.16]. The properties (iii,n = 1), (iv) and (v) are equivalent by [3, prop. 3.1]. (iii)⇒(ii): Suppose that B is elementary, i.e., that B ∼= K(H) for some Hilbert space H. By (iii), applied on some rank-one projection a := p ∈ B+, b ∈ B+, and ε = 1/2, it follows that each element of b ∈ B+ has rank ≤ n. Thus, H has dimension k ≤ n, and B ∼= Mk. But the latter case was excluded in (iii,n). Therefore B is non-elementary and hence has the global Glimm halving property in sense of [3, def. 2.6]. This is easy to see for non-elementary simple B (or use Glimm halving [20, lem. 6.7.1]). Since B is simple property (iii, n) ensures that B satisfies property (i) of [3, def. 1.2] of pi(n). Therefore, [3, prop. 4.14] says that B is pi(1). Since B is simple and B 6= C, there are no non-zero characters on B. In particular pi(1) ensures that B is purely infinite in the sense of [13, def. 4.1]. By [13, thm. 4.1] we obtain property (ii). (cid:3) 20 Proof of Theorem 1.2: By Lemma 6.3, A is G-simple. Thus the action σ is automatically exact by Definition 3.5. Since σ is element-wise properly outer by as- sumption, it is now also residually properly outer, and Theorem 3.8 applies. It says that F := A+ is a filling family in A ⋊σ,λ G. In particular, A separates the ideals of the reduced crossed product. It shows that B := A ⋊σ,λ G is simple. If b ∈ B+ with kbk = 1, then there exists non-zero z ∈ B such that z∗z ≤ b and zz∗ ∈ A, because A+ is filling for B: One way to see this is to use property (i) of Definition A.3 on a′ := (3b−2)+, b′ := (3b−1)+−a′ and c′ := 3b−a′−b′ to get elements zj, d ∈ B. This imply that z∗ i zi = c′ec′ ≤ kc′kkekc′ ≤ 3kc′kkek b, where e 6= 0 because d∗ed = a′ 6= 0. j zj ≤ e = Pi z∗ Take δ ∈ (0,kzk2/2). Then 0 6= z(z∗z − δ)+z∗ = ϕ(zz∗) ∈ A+ for some suitable ϕ ∈ Cc(0,∞]+. We consider three cases: (i) the action is n-majorizing and A is non-unital, (ii) the action is n-majorizing and A is unital and (iii) the action is n-covering and A is unital. Since G is discrete, A is a non-degenerate C -subalgebra of B. In particular, A+ contains an approximate unit (eν) of positive contractions in A+ for B, which we will use for case (i). In case (ii)-(iii) let eν := 1, where 1 is the unit of A. Define m := n + 1 for case (ii) and m := n for case (i) and (iii). By each of the Definitions 6.1 and j (z(z∗z − δ)+z∗)fjk < ε for fj := U(g−1 j )dj in B. 6.2 (and using Lemma 6.3 to get property (β) in case (ii) with n replaced by m), for each ε > 0 and eν ∈ A+ there are d1, . . . , dm ∈ A and g1, . . . , gm ∈ G such that keν − Pj f ∗ By Remark 2.1(ii), there exists a contraction d0 ∈ B with d∗ the elements yj := d0z∗fj ∈ B satisfy keν−Pm to a ∈ B+, we get that B has the following property: For any two non-zero elements b, a ∈ B+ and ε > 0 there exists m elements d1, . . . , dm ∈ B such that k a − Pm j bdj k < ε. A simple C -algebra B with this property is strongly purely infinite by Remark 6.4 if B is not isomorphic to Mk for some k ≤ m. The case that B is isomorphic to a C -subalgebra of Mm has been ex- cluded by the definitions: For case (i) we know A is non-unital, hence B is non-unital. For case (ii)-(iii) we know A is not isomorphic to a C -subalgebra of Mm, hence B is not isomorphic to a C -subalgebra of Mm. (cid:3) 0bd0 = (z∗z − δ)+. Then j byjk < ε. Since a1/2eνa1/2 converges j=1 y∗ j=1 d∗ Lemma 6.5. The following are equivalent for C-algebras B. (i) B does not contain a non-zero projection p ∈ B+ with pBp = C · p. 21 (ii) For every non-zero hereditary C-subalgebra D of B, each maximal commutative C-subalgebra C of D has perfect Gelfand spectrum bC. i.e., bC does not contain an isolated point. (iii) For every a1, a2 ∈ B+\{0} and c ∈ B there are b1, b2 ∈ B+\{0} with b1cb2 = 0, and bj ≤ aj (j = 1, 2). Proof. (iii)⇒(i): Let p∗ = p = p2 ∈ B\{0}, put ak := c := p for k = 1, 2. Then there are non-zero b1, b2 ∈ (pBp)+ with b1b2 = 0. Thus pBp 6= C · p. (i)⇒(ii): Let D 6= {0} a hereditary C -subalgebra of B, and C a maximal commuta- tive C -subalgebra of D. Suppose that bC is not perfect. Then bC contains an isolated point χ. The point χ corresponds to a projection 0 6= p ∈ C+ ⊆ D+ with pBp = C · p, see [21, lem. 7.14]. (ii)⇒(iii): It is easy to see that a commutative C -algebra C with a perfect spectrum bC contains non-zero contractions e1, e2 ∈ C+ with e1e2 = 0, because the locally compact Hausdorff space bC must in particular contain two different points (6= ∞). Given c ∈ B and non-zero a1, a2 ∈ B+, let dj := (aj − kajk/2)+ and x := d1/2 1 cd1/2 . Notice that 0 6= dj ∈ B+, and that 0 6= d1/2 j ≤ aj for all non-zero contractions 0 ≤ yj ∈ djBdj. If x 6= 0, consider the hereditary C -subalgebra D := x∗Bx = x∗xBx∗x that is generated by x∗x, and is contained in 2Bd2. Let C be a maximal commutative C -subalgebra of D with x∗x ∈ C. Then C d∗ contains non-zero contractions e1, e2 ∈ D+ with e1e2 = 0 = e1(x∗x)1/2e2. If x = 0 then take bj j yjd1/2 := dj. 2 It is well-known (and easy to see) that the polar decomposition x = v(x∗x)1/2 in B∗∗ 1Bd1. Thus f := ve1v∗ ∈ xBx∗ and has the 2 e2d1/2 has the property that vDv∗ = xBx∗ ⊆ d∗ property f xe2 = ve1(x∗x)1/2e2 = 0. It follows that b1 := d1/2 satisfy b1cb2 = d1/2 and b2 := d1/2 1 f d1/2 1 f xe2d1/2 2 = 0 and 0 6= bj ≤ aj. 1 2 (cid:3) Proposition 6.6. Let (A, G, σ) a C-dynamic system with non-zero non-unital A. If the action σ of G on A is an 1-majorizing action in the sense of Definition 6.1, then A is G-simple and σ is G-separating for A. Proof. The algebra A is G-simple and A does not contain a projection p 6= 0 with pAp = C · p by Lemma 6.3. To show that σ is G-separating let a1, a2 ∈ A+\{0}, c ∈ A and ε > 0. By Lemma 6.5, there exist b1, b2 ∈ A+\{0} with b1cb2 = 0 and bj ≤ aj (j = 1, 2). Using (twice) that the action is 1-majorizing we find ej ∈ A, hj ∈ G for 22 j = 1, 2 such that k e∗ we get kd∗ j ajdj − σgj (aj)k < ε and d∗ j σhj(cid:0)bjajbj(cid:1) ej − aj k < ε . With gj := h−1 j 1cd2 = 0, i.e., σ is G-separating. and dj := bjσgj (ej) (cid:3) Remark 6.7. Suppose that (A, G, σ) is a C -dynamical system, A is unital and com- mutative, and G is discrete. Then the following properties (i) -- (iv) of the action σ are equivalent: (i) The action is 2-covering in sense of Definition 6.2. (ii) The corresponding (adjoint) action bσ, on bA is a strong boundary action in the sense of Definition A.1. (iii) The action is 1-majorizing in sense of Definition 6.1. (iv) The action is 2-filling in sense of Definition A.2, and A is not isomorphic to a subalgebra of M2(C). We do not know if, also for non-commutative and unital A, every 2-covering action is a 1-majorizing action, or a 2-filling action. Proof. We show more general implications, except for (i)⇒(ii). In particular we show that an action σ on a unital abelian C -algebra A is n-filling if and only if it is n- covering provided that the (linear) dimension of A is greater than n. g (i)⇒(iv) (for A unital, commutative, any n): Suppose that A ∼= C(X), and take any n ≥ 2. Let α denote the action of G on X inducing σ, i.e., σg(f ) = f ◦ α−1 for f ∈ A and g ∈ G. Since the action α is minimal by Lemma 6.3, Remark [11, rem. 0.4] shows it suffices to prove that for each non-empty subset U of X there exist g1, . . . , gn ∈ G, such that αg1(U)∪ αg2(U)∪· · ·∪ αgn(U) = X. Let such U be given. Select non-zero a ∈ A+ with support contained in U. By (i) there exist g1, . . . , gn ∈ G and d1, . . . , dn ∈ A such that Pn 2 . In particular for each x ∈ X, σgj (a)(x) is non-zero for some j, so x ∈ αgi(U). j σgj (a)dj ≥ 1 j=1 d∗ (iv)⇒(i) (for A unital, commutative/non-commutative, any n): Suppose that A is unital, and take any n ≥ 2. Let 0 6= a ∈ A+. Using (iv) there are g1, . . . , gn ∈ G and δ > 0 such that D := Pn j=1 σgj (a) ≥ δ1 , and A is not isomorphic to a C -subalgebra of Mn. Thus, D is invertible in A and Pn j σgj (a)dj = 1 for dj := D−1/2 in A. j=1 d∗ (iii)⇒(i) (for A unital, commutative/non-commutative, any n): Each n-majorizing actions on unital A is an (n + 1)-covering action by Lemma 6.3. 23 (i)⇒(ii) (for A unital, commutative, one n): Suppose that A ∼= C(X) and let α denote the action of G on X inducing σ. The equivalence of (i) and (iv), shows that for given 0 6= a ∈ A+, there exists g ∈ G and δ > 0 such that a + σg(a) ≥ δ1. Let U ⊆ X open and non-empty. There is 0 6= a ∈ C(X)+ with support a−1(0,∞) ⊆ U. Choose h ∈ G and δ > 0 with a + σh(a) ≥ δ1. It implies that σh(a)(x) > 0 for all x ∈ X\U. Thus, αg(x) ∈ U for all x ∈ X\U and g := h−1, i.e., there exists g ∈ G with αg(X\U) ⊆ U. Given non-empty open subsets U and V of X. We let W := U ∩ V if U ∩ V 6= ∅. Then g ∈ G with αg(X\W ) ⊆ W satisfies αg(X\U) ⊆ V . If U ∩ V = ∅ then we find g, h ∈ G with αg(X\U) ⊆ U ⊆ X\V and αh(X\V ) ⊆ V . Then αhg(X\U) ⊆ V . The space X contains more than two points because A is not isomorphic to a C - subalgebra of M2(C). Thus, (X, G, α) satisfies the conditions of Definition A.1 of a strong boundary action. (ii)⇒(iii) (for A unital/non-unital, commutative, one n): We show (iii) using possibly less than (ii): Let X be a locally compact space that is not necessarily compact and contains more than 2 points. Let α an action of G on X with the property that, for every compact subset K ⊆ X with K 6= X and each non-empty open subset U ⊆ X, there exists g ∈ G with αg(K) ⊆ U. Then the adjoint action σ of α on A := C0(X) is an 1-majorizing action of G on A. Indeed: Let 0 6= a ∈ A+, b ∈ A+ non-invertible, and ε > 0. Put δ := ε/3. Then, considered as functions on X, they have the property that U := a−1(kak/2,∞) is non- empty and open and K := b−1[δ,∞) is compact. Find h ∈ G with αh(K) ⊆ U. Then x ∈ K ⇔ b(x) ≥ δ implies that for g := h−1 we get σg(a)(x) = a(αh(x)) > kak/2 . It follows kd∗σg(a)d − bk < ε with d ∈ A+ given by d(x) := σg(a)(x)−1/2(b(x) − 2δ)1/2 + for x ∈ K and d(x) := 0 for x ∈ X\K. (cid:3) Remarks 6.8. (i) Let α be an action of a discrete group G on a locally compact Hausdorff space X with more than two points, and σ the induced action on A := C0(X). By Remark 6.7 the following properties are equivalent for X compact: (1) The action α is a strong boundary action (Definition A.1) in the sense of [18]: For each pair of non-empty open subsets U and V of X there exists g ∈ G with U ∩ αg(V ) = X. 24 (2) For any compact set K 6= X and any open set U 6= ∅ there exist t ∈ G such that αt(K) ⊆ U. (3) For every non-zero a ∈ A+, every non-invertible b ∈ A+ and every ε > 0, there exist d ∈ A and g ∈ G such that kd∗σg(a)d − bk < ε. Clearly, this can not be the case if X is locally compact but is not compact. In general (i.e., when X is compact or non-compact) we know (1)⇒(2)⇒(3). Properties (2)-(3) are both candidates for a generalised notion of a strong boundary action, however only (3) applies when A is non-commutative. (ii) The notion of a strong boundary action (Definition A.1) is defined on compact Hausdorff spaces with more than two points. In view of Remark 6.8(i) and Remark 6.7, we call the 1-majorizing actions on not-necessarily unital or commutative C -algebras also strong boundary actions. (iii) Suppose that a discrete group G acts by a topologically free action α on a compact Hausdorff space X, and that X contains more than two points. It was shown in [18, thm. 5] that the crossed product C(X) ⋊σ,λ G is purely infinite provided that the action -- in addition -- is a strong boundary action. Since topological freeness implies σ is element-wise properly outer (by [1, prop. 1]) we conclude that, with the terminology of Remark 6.8(ii), [18, thm. 5] is a special case of Theorem 1.2. (cid:3) (iv) Let α be an action on a non-compact locally compact Hausdorff space X with more than two points and σ the induced action. It was shown in Proposition 6.6 that σ is G-separating if σ is a strong boundary (i.e., 1-majorizing) action. A simpler argument applies if we assume that for any compact set K 6= X and any non-empty open set U ⊆ X there exist g ∈ G such that αg(K) ⊆ U: Proof. Since any finite subset M of X is compact, it can be moved by suitable αg into any non-empty open subset U of X. In particular X is perfect and each non-empty open set V ⊆ X contains at least two non-empty open disjoint subsets V1 and V2. Let K1 ⊆ U1 and K2 ⊆ U2 with Kj compact (hence Kj 6= X) and Uj open. If U1 and U2 are disjoint, then we can take g1 = g2 = e in Lemma 5.1(ii). If V := U1 ∩ U2 6= ∅, then consider the above disjoint non-empty open subsets Vj ⊆ V . By assumption, there exist g1, g2 ∈ G with αgj (Kj) ⊆ Vj ⊆ Uj. Thus, the adjoint action σ of α is G-separating. (cid:3) (v) Suppose that a discrete group G acts by a topologically free action α on a non- compact locally compact Hausdorff space X, and that X contains more than two points. 25 Then the crossed product C0(X) ⋊σ,λ G is purely infinite provided that the following property holds: for any compact set K 6= X and any non-empty open set U ⊆ X there exist t ∈ G such that αt(K) ⊆ U. This follows as a corollary of Theorem 1.1, also of Theorem 1.2 or of Corollary 5.2. Proof. We must verify the following properties according to each of the listed results: (1.1) The action σ is exact, residually properly outer and G-separating. (1.2) The action σ is 1-majorizing, and element-wise properly outer. (5.2) The action σ is exact, G-separating and fulfills (): For every closed G-invariant subset Y of X and every g 6= e the set {y ∈ Y : αg(y) = y} has empty interior. By Remark 6.8(i) we know the action σ is a strong boundary (i.e., 1-majorizing) action. Hence A is G-simple, cf. Lemma 6.3. In particular it follows that the action α on X is minimal. This reduces property () to the definition of topological freeness, cf. [1, p.120]. The minimality of the action α implies that the corresponding adjoint action σ : G → Aut(C0(X)) is exact, and that it becomes residually properly outer if it is element-wise properly outer. But σ is element-wise properly outer if and only if α is a topological free action (see [1, p.120] or [23, cor 2.22]). It remains to show σ is G-separating, but this is already contained in Remark 6.8(iv). (cid:3) (vii) It is an important point that a strong boundary action is often G-separating and (in fact always) minimal, but the notion of a G-separating action is not typically related to minimality. Consequently, working with G-separating actions allows us to consider ideal-related classification of non-simple strongly purely infinite crossed products. Remark 6.9. If A has real rank zero, then one can restrict the conditions in Definitions 4.1, 6.1 and 6.2 to projections p, q ∈ A in place of the elements a, b ∈ A+. Proof. Case of Definition 4.1: Let a1, a2 ∈ A+, c ∈ A, ε > 0 and define δ := ε/(1 + ka1k + ka2k) . By [4], Dj := ajAaj contains an approximate unit consisting of non-zero projec- tions. Thus, there are projections pj ∈ Dj such that kaj − a1/2 j k < δ. Use [13, prop. 2.7(i)] and the comment following [13, prop. 2.6] to select zj ∈ Dj satisfying j aj zj = pj. Let c′ := z∗ z∗ Suppose that there exists ej ∈ A, gj ∈ G such that and ke∗ j pjej − σgj (pj)k < δ 1c′ e2k < δ . pj a1/2 j 1cz2. ke∗ 26 j ) and dj j := zjejσgj (a1/2 j ) . They satisfy d∗ j ajdj = v∗ j e∗ j pja1/2 1c′e2v2, we get kd∗ 1e∗ j pjejvj, j ajdj − σgj (aj)k < (1 + kajk)δ ≤ ε . Since ). Thus, kd∗ 1cd2k < δ(ka1k · ka2k)1/2 ≤ δ(ka1k + ka2k) ≤ ε . := σgj (a1/2 Define vj j σgj (pj)vj = σgj (a1/2 v∗ 1cd2 = v∗ d∗ Case of Definitions 6.1 and 6.2: Let a1, a2 ∈ A+ and ε > 0, with a1 6= 0 and a2 not invertible in A (respectively a2 = 1). We can assume ε ≤ 1. Define δ := ε/(1 + ka2k) . Choose pj, zj ∈ Dj := ajAaj as above with ka1/2 j ajzj = pj. Then p1 6= 0 and p2 is not invertible in A (i.e., p2 6= 1) if a2 is not invertible, otherwise p2 = 1: If p2 is invertible then 1 ∈ a2Aa2, so a2 is invertible. Conversely, if a2 is invertible then kp2 − 1k < ε/2, so p2 is invertible. If there are e1, . . . , en ∈ A and g1, . . . , gn ∈ G with k p2 − Pj e∗ dj := σgj (z1)eja1/2 j σgj (p1)ej k < δ , then j −ajk < δ and z∗ j pja1/2 (cid:3) satisfies k a2 − Pj d∗ 2 j σgj (a1)dj k < (1 + ka2k)δ ≤ ε . Acknowledgments Parts of this work were conducted while the second named author was at the Fields Institute from 2009 to 2012. It is with great pleasure we forward our thanks to the Fields Institute and in particular Professor George Elliott for all the support. This research was also supported by the Australian Research Council. Appendix A. In this appendix we have included a few recent definitions and results that are frequently cited throughout this paper. The results quoted from [15] are available as preprint. Definition A.1 ([18]). Let α be an action of a discrete group G on a compact spaces X with at least three points. The action α is as strong boundary action if for every pair U and V of non-empty open subsets of X there exists t ∈ G such that αt(X \ U) ⊆ V . Definition A.2 ([11]). An action σ of a discrete group G on a unital C -algebra A is n-filling (n ≥ 2) if, for all b1, . . . , bn ∈ A+, with kbjk = 1 for each j, and all ε > 0, there exist g1, . . . , gn ∈ G such that Pn Definition A.3 ([15]). Let F be a subset of A+. The set F is a filling family for A, if F satisfies the following equivalent conditions (i) and (ii). j=1 σgj (bj) ≥ 1 − ε. 27 (i) For every a, b, c ∈ A with 0 ≤ a ≤ b ≤ c ≤ 1, with ab = a 6= 0 and bc = b, there exists z1, z2, . . . , zn ∈ A and d ∈ A with zj(zj)∗ ∈ F , such that ec = e and d∗ed = a for e := z∗ 1z1 + . . . + z∗ nzn. (ii) For every hereditary C -subalgebra D of A and every primitive ideal I of A with D 6⊆ I there exist f ∈ F and z ∈ A with z∗z ∈ D and zz∗ = f 6∈ I. Lemma A.4 ([15]). Suppose that A ⊆ B is a C-subalgebra of B and F ⊆ A+ is a subset of A+. If F is filling for A, and A+ is filling for B, then F is a filling family for B. Remark A.5 ([15]). Let A ⊆ B be C -algebras and F ⊆ A+. If F := A+ ⊆ B is filling for B, then the map I ∈ I(B) 7→ I ∩ A ∈ I(A) is injective, i.e., A separates the closed ideals of B. Definition A.6 ([15]). A C -algebra A is strongly purely infinite (for short: s.p.i. ) if, for every a, b ∈ A+ and ε > 0, there exist elements s, t ∈ A such that k s∗a2s − a2 k < ε , k t∗b2t − b2 k < ε and k s∗abtk < ε . (6) Remark A.7 ([15]). A C -algebra A is strongly purely infinite if and only if for every a, b ∈ A+, c ∈ A and ε > 0, there exist contractions s, t ∈ A such that k s∗as − ak < ε , k t∗bt − bk < ε and k s∗ctk < ε . (7) Definition A.8 ([15]). Let S ⊆ A be a multiplicative sub-semigroup of a C -algebra A and C ⊆ A a subset of A. An n-tuple (a1, . . . , an) of positive elements in A has the matrix diagonalization property with respect to S and C, if for every [aij] ∈ Mn(A)+ with ajj = aj and aij ∈ C (for i 6= j) and εj > 0, τ > 0 there are elements s1 , . . . , sn ∈ S with ks∗ j ajj sj − ajjk < εj , (8) If S = C = A then this is the matrix diagonalization property of (a1, . . . , an) as defined in [14, def. 5.5], and we say that (a1, . . . , an) has matrix diagonalization (in A). i aijsjk < τ for i 6= j . and ks∗ Definition A.9 ([15]). Let F be a subset of A+. The family F has the (matrix) diagonalization property (in A) if each finite sequence a1, . . . , an ∈ F has the matrix diagonalization property (in A) of Definition A.8. Lemma A.10 ([15]). Suppose that F ⊆ A+ is invariant under ε-cut-downs, i.e., that for each a ∈ F and ε ∈ (0,kak) we have (a − ε)+ ∈ F . Then the family F has the matrix diagonalization property, if and only if, each pair of elements in F has the matrix diagonalization property of Definition A.8. 28 Lemma A.11 ([15]). Let ε0 > 0 and non-empty subsets F ⊆ A+, C ⊆ A be given, and let S ⊆ A be a (multiplicative) sub-semigroup of A that satisfies s∗ 2Cs1 ⊆ C for all s1, s2 ∈ S. Suppose that the following properties hold: (i) For every ε0 > δ > 0, the pair ((a1 − δ)+, (a2 − δ)+) the matrix diagonalization property with respect to S and C of Definition A.8. (ii) ϕ(a1)cϕ(a2) ∈ C for each c ∈ C and ϕ ∈ Cc(0,∞]+. (iii) ϕ(a1)s, ϕ(a2)s ∈ S for each s ∈ S and ϕ ∈ Cc(0,∞]+. Then, for every c ∈ span(C), a1, a2 ∈ F , ε0/2 ≥ ε > 0, and τ > 0, there exists s1, s2 ∈ S that fulfil ksjk2 ≤ 2kajk/ε and ks∗ 1a1s1 − a1k < ε , ks∗ 2a2s2 − a2k < ε and ks∗ 1cs2k < τ . (9) Theorem A.12 ([15]). The minimal tenor product of a strongly purely infinite and an exact C-algebra is strongly purely infinite. Theorem A.13 ([15]). Suppose that A+ contains a filling family F that has the diag- onalization property (in A). Then A is strongly purely infinite. References [1] R.J. Archbold and J.S. Spielberg, Topologically free actions and ideals in discrete C-dynamical systems, Proc. Edinburgh Math. Soc. (2) 37 (1994), 119 -- 124. 6, 7, 16, 17, 25, 26 [2] G. Arzhantseva, V. Guba, and M. Sapir, Metrics on diagram groups and uniform embeddings in Hilbert space, Comment. Math. Helv. 81 (2006), 911 -- 929. 9 [3] E. Blanchard and E. Kirchberg, Non-simple purely infinite C-algebras: the Hausdorff case, J. Funct. Anal. 207 (2004), 461 -- 513. 4, 20 [4] L.G. Brown and G.K. Pedersen, C-algebras of real rank zero, J. Funct. Anal. 99 (1991), 131 -- 149. 26 [5] J. Cuntz, K-theory for certain C-algebras, Ann. of Math. 113 (1981), 181 -- 197. 20 [6] G.A. Elliott, Some simple C-algebras constructed as crossed products with discrete outer auto- morphism groups, Publ. Res. Inst. Math. Sci. 16 (1980), 299 -- 311. 6 [7] R. Exel, M. Laca, and J. Quigg, Partial dynamical systems and C-algebras generated by partial isometries, J. Operator Theory 47 (2002), 169 -- 186. 7 [8] D.S. Farley, Proper isometric actions of Thompson's groups on Hilbert space, Int. Math. Res. Notes 45 (2003), 2409 -- 2414. 9 [9] E. Guentner and J. Kaminker, Exactness and uniform embeddability of discrete Groups, J. London Math. Soc.(2) 70 (2004), 703 -- 718. 9 [10] U. Haagerup and G. Picioroaga, New presentations of Thompson's groups and applications, J. Operator Theory 66 (2011), 217 -- 232. 8 29 [11] P. Jolissaint and G. Robertson, Simple purely infinite C-algebras and n-filling actions, J. Funct. Anal. 175 (2000), 197 -- 213. 3, 23, 27 [12] S. Kawamura and J. Tomiyama, Properties of topological dynamical systems and corresponding C-algebras, Tokyo J. Math. 13 (1990), 251 -- 257. 7, 16 [13] E. Kirchberg and M. Rørdam, Non-simple purely infinite C-algebras, Amer. J. Math. 122 (2000), 637 -- 666. 20, 26 [14] , Infinite non-simple C-algebras: absorbing the Cuntz algebras O∞, Adv. Math. 167 (2002), no. 2, 195 -- 264. 4, 12, 20, 28 [15] E. Kirchberg and A. Sierakowski, Filling families and strong pure infiniteness, preprint 2014. 2, 27, 28, 29 [16] E. Kirchberg and S. Wassermann, Exact groups and continuous bundles of C-algebras, Math. Ann. 315 (1999), 169 -- 203. 8, 14 [17] A. Kishimoto, Outer automorphisms and reduced crossed products of simple C-algebras, Comm. Math. Phys. 81 (1981), 429 -- 435. 6 [18] M. Laca and J. Spielberg, Purely infinite C-algebras from boundary actions of discrete groups, J. Reine Angew. Math. 480 (1996), 125 -- 139. 1, 3, 24, 25, 27 [19] D. Olesen and G.K. Pedersen, Applications of the Connes spectrum to C-dynamical systems. III, J. Funct. Anal. 45 (1982), 357 -- 390. 6, 7 [20] G.K. Pedersen, C-algebras and their automorphism groups, LMS Monographs, vol. 14, Academic Press Inc., London, 1979. 20 [21] M. Rørdam, F. Larsen, and N. Laustsen, An introduction to K-theory for C-algebras, London Mathematical Society Student Texts 49, Cambridge University Press, Cambridge, 2000. 22 [22] J. Renault, The ideal structure of groupoid crossed product C-algebras, With an appendix by Georges Skandalis. J. Operator Theory 25 (1991), 3 -- 36. 6, 16 [23] A. Sierakowski, The ideal structure of reduced crossed products, Munster J. Math. 3 (2010), 223 -- 248. 2, 6, 8, 9, 16, 17, 26 Institut fur Mathematik, Humboldt Universitat zu Berlin, Unter den Linden 6, D -- 10099 Berlin, Germany [email protected] School of Mathematics & Applied Statistics, University of Wollongong, Faculty of Engineering & Information Sciences, 2522 Wollongong, Australia [email protected] 30
1012.4435	2	1012	2011-04-13T16:18:04	Integrable representations of involutive algebras and Ore localization	[ "math.OA", "math.RA", "math.RT" ]	Let $\mathcal A$ be a unital algebra equipped with an involution $(\cdot)^\dagger$, and suppose that the multiplicative set $\mathcal S\subseteq \mathcal A$ generated by the elements of the form $1 + a^\dagger a$ satisfies the Ore condition. We prove that: (i) Cyclic representations of $\mathcal A$ admit an integrable extension (acting on a possibly larger Hilbert space), and (ii) Integrable representations of $\mathcal A$ are in bijection with representations of the Ore localization $\mathcal A\mathcal S^{-1}$ (which we prove to be an involutive algebra). This second result is a limited converse to a theorem by Inoue asserting that representations of symmetric involutive algebras are integrable.	math.OA	math	Integrable representations of involutive algebras and Ore localization Rodrigo Vargas Le-Bert∗ Instituto de Matem´atica y F´ısica, Universidad de Talca Casilla 747, Talca, Chile April 5, 2011 Abstract Let A be a unital algebra equipped with an involution (·)†, and suppose that the multiplicative set S ⊆ A generated by the elements of the form 1 + a†a contains only regular elements and satisfies the Ore condition. We prove that: • Ultracyclic representations of A admit an integrable extension (acting on a possibly larger Hilbert space). • Integrable representations of A are in bijection with representations of the Ore localization AS−1 (which we prove to be an involutive algebra). This second result can be understood as a restricted converse to a theorem by Inoue asserting that representations of symmetric involutive algebras are integrable. 2010 MSC: 16S (primary); 46L (secondary). 1 Introduction Unbounded operator algebras appear in several important domains, such as quantum field theory, representations of Lie algebras and quantum groups. Consequently, there has long been an interest in developing their theory, which despite that has grown rich in technical details and relatively poor in applica- tions (see [14, 6, 1, 7] for complete expositions, [2] for a physical applications survey, and [12] for applications in other fields). Among the causes for this fact ∗[email protected]. Supported by Fondecyt Postdoctoral Grant No3110045. 1 lie the inherent difficulties in the representation theory of general involutive algebras, which we summarize as follows. Let A be a unital algebra equipped with an involution (·)† and H a Hilbert space. 1. Speaking of a representation presupposes that sum, product and involu- tion are well-defined; however, unbounded operators are not defined all over H and cannot be blindly added or composed. Hence, one usually postulates the existence of an invariant domain V ⊆ H for the operators which will represent A, thus allowing for a pointwise definition of their sum and product. Letting L†(V) = { a : V → V V ⊆ D (a∗) and a∗V ⊆ V } , one obtains an algebra with involution a† = a∗V, and then can define a representation to be a morphism π : A → L†(V). 2. Operators in L†(V) are closable, for their adjoint is densely defined. Therefore, the natural algebraic operations with them are the so-called strong sum and strong product, see Definition 2.3. The problem arises from this, together with the fact that each a ∈ L†(V) may admit more than one closed extension. As a consequence, pointwise operations become ambiguous in a sense. They are too weak a version of the strong ones. For a representation π, this means that ¯π(a + b) ⊆ ¯π(a) + ¯π(b), ¯π(ab) ⊆ ¯π(a) ¯π(b), ¯π(a†) ⊆ ¯π(a)∗, and equality does not hold in general. Simply put, L†(V) is not an object that truly captures the algebraic structure of closed operators, thus allowing, even in the simplest cases, for the appearance of ill-behaved representations. The following two examples are archetypical of good and bad behaviour. Example 1.1. Let G be a Lie group with Lie algebra g. Representations of the universal enveloping algebra A = U(g) that arise by differentiation from unitary representations of G are called integrable (a complete exposition is found in [14]). They are characterized by the fact that they respect the involution: ¯π(a†) = ¯π(a)∗, ∀a ∈ A, a property which is taken as the definition of integrability in the general case. Integrability is also very important when A is the observable algebra of a quantum system. Example 1.2. Let M be any properly infinite von Neumann algebra acting on a separable Hilbert space. Then, there exists a representation π of A = C[x, y], the algebra of polynomials in two commuting hermitian variables, such that π(x) 2 and π(y) are essentially self-adjoint and their spectral projections generate M. Such representations are not integrable. For more details, see [13, Section 5] and [14, Section 9.4]. The second example shows that conditions must be imposed on represen- tations, in order to exclude pathological behaviour and make them useful in practice. Now, during the last decade, two general approaches to the well- behaved representations have been developed. In what follows we provide a brief description of both (a complete discussion is found in [1]). After that, we will be in position to comment on one important limitation that they share, and on how it is overcome in an approach that we propose in this paper. The first one, of an algebraic flavor, is due to Schm udgen [15]. It generalizes the notion of integrable representations of universal enveloping algebras U(g), integrable representations of U(g) determine unitary in the following way: representations of the corresponding simply connected Lie group G. Those, in turn, are equivalent to representations of the Banach algebra L1(G). Now, it turns out that an integrable representation of U(g) can be recovered from the corresponding representation of C∞ 0 (G) ⊆ L1(G) by making use of the natural action of U(g) on C∞ 0 (G). The generalization to arbitrary involutive algebras goes as follows: say that the involutive algebra A and the normed involutive algebra A0 are compatible if there is an action of A on A0 such that (a · x)†y = x†(a† · y), ∀a ∈ A, ∀x, y ∈ A0. Then, non-degenerate, continuous representations π0 of A0 determine what we define to be the well-behaved representations π of A by π(a)π0(x) = π0(a · x), a ∈ A, x ∈ A0. The second approach, of an analytic flavor, was proposed by Bhatt, Inoue and Ogi [3]. It is based on unbounded C-seminorms. Given an involutive algebra A, an unbounded C-seminorm is a C-seminorm p : A0 → R, where A0 is a subalgebra of A. The kernel I of such a seminorm is a bilateral ideal of A0, and the completion with respect to p of A0/I is a C-algebra which we denote by A. Now, let N = { x ∈ A0 ax ∈ A0, ∀a ∈ A } , which is a left ideal of A. Each representation π0 : A → B(H ) induces a representation π : A → L†(V), where V = Span(cid:8) π0(x + I)ξ (cid:12)(cid:12)(cid:12) x ∈ N, ξ ∈ H (cid:9) , by the simple formula π(a)π0(x + I)ξ = π0(ax + I)ξ. Observe, however, that V might not be dense in H and, when this is the case, we say that the repre- sentation is well-behaved. One sees that well-behaved representations satisfy k ¯π(a)k = p(a), for all a ∈ A0 (in general, k ¯π(a)k ≤ p(a)). 3 This second approach is more general than the first one, and can be fur- ther generalized to the case of partial involutive algebras [17]. The precise relationship between the two approaches has been worked out in [11]. The limitation referred to above is more apparent in the first one: obtaining well- behaved representations of A = U(g) requires knowledge of representations of A0 = C∞ 0 (G). If one is to obtain representations of G from representations of U(g), this is clearly going in the wrong direction. The same happens with the second approach, as is best seen with an example: take A = C[x], the commu- tative free algebra on one hermitian generator, and consider its well-behaved representation by multiplication operators on L2(R). There is no subalgebra of A on which the corresponding norm is finite. Thus, obtaining such a simple representation by the second approach requires enlarging C[x] to contain at least one function of exponential decay -- a procedure which is, again, taking as given something that, in several practical cases, should come as a result. Our approach is precisely based on enlarging the algebra A to contain inverses to the elements of the form 1 + a†a, thus obtaining a so-called symmetric involutive algebra. It is known that representations of the latter are integrable [9]. Our main result is Theorem 3.15, which says that when the multiplicative set S generated by { 1 + a†a a ∈ A } satisfies the Ore condition (see Section 3.2), integrable representations of A are in bijection with representations of the Ore localization AS−1. The paper is organized as follows: the second section is a short reminder of necessary background, and our results are proved in the third section. For convenience, an elementary appendix on closable operators is included. 2 Short reminder on representation theory This section contains no new results. It is included only for the convenience of non-expert readers, and it will serve to introduce our notations, too. 2.1 Ultracyclic representations and GNS construction Let V be a complex vector space and a ∈ L(V). We denote their corresponding algebraic duals by V† = (cid:8) f : V → C (cid:12)(cid:12)(cid:12) f is complex linear (cid:9) , a† : f ∈ V† 7→ f a ∈ V†. Now, let A be a unital involutive algebra with involution (·)†, to be represented by operators in L(V). Allowing for a correspondence between involution and algebraic duality requires a choice of antilinear inclusion V ֒→ V†, which we assume given by an embedding V ֒→ H as a dense subspace in a Hilbert space. 4 Definition 2.1. Given a dense subspace V of a Hilbert space H , denote by L†(V) = { a : V → V V ⊆ D (a∗) , a∗V ⊆ V } , which is an involutive algebra with involution a† = a∗V. A representation of the involutive algebra A is a morphism π : A → L†(V) of involutive algebras. We say that π is ultracyclic if it admits an ultracyclic vector, that is, an Ω ∈ V such that V = π(A)Ω. Definition 2.2. Let A be an involutive algebra and Ah the real vector subspace of its hermitian elements. We say that x ∈ Ah is positive if it belongs to the cone Π(A) =  n X i=1 λia† i ai (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) λi > 0, ai ∈ A  ⊆ Ah. The cone of positive elements allows one to define an order relation on A. This, in turn, gives an order relation on its algebraic dual: we say that f ∈ A† is positive if f (x) ≥ 0, ∀x ∈ Π(A). Recall that positivity implies: 1. f (a†b) = f (b†a). 2. The Cauchy-Schwartz inequality f (a†b)2 ≤ f (a†a) f (b†b). We say that f ∈ A† is a state if it is positive and f (1) = 1. We denote the set of states of A by Σ(A). Given a representation π : A → L†(V), one obtains a state f ∈ A† from any vector Ω ∈ V with kΩk = 1 by the formula f (a) = hΩ, π(a)Ωi. The GNS construction allows a recovery of both ππ(A)Ω (modulo unitary con- jugation) and Ω (modulo a phase factor) from f , thus establishing a correspon- dence between ultracyclic representations and states. We proceed to describe it; for a complete treatment, see [14]. Let f ∈ Σ(A), and consider the set I = { a ∈ A f (a†a) = 0 } . Using the Cauchy-Schwartz inequality it is easily seen that I is a left ideal. The quotient V = A/I is densely embedded in its Hilbert space completion H with respect to the scalar product h[a], [b]i = f (a†b), a, b ∈ A, where [·] denotes the equivalence class in V of its argument. The GNS repre- sentation π is simply given by the canonical left A-module structure of V. It admits the ultracyclic vector Ω = [1] ∈ V. 5 2.2 Integrability As mentioned in the introduction, the algebraic operations of L†(V) are not satisfactory from an analytic point of view. Indeed, every a ∈ L†(V) is closable because its adjoint a∗ is densely defined. Now, while closed operators cannot always be added or composed, the natural operations when they can are the following. Definition 2.3. Let A, B ∈ C(H ), the set of closed, densely defined operators on a Hilbert space H . If D = D (A) ∩ D (B) is dense and AD + BD is closable, we define their strong sum A + B = AD + BD ∈ C(H ). Analogously, if D = B−1D (A) is dense and ABD is closable, we define their strong product AB = ABD ∈ C(H ). Remark 2.4. When operating with closed operators we will always mean strong operations. This should cause no confusions, because we will always write elements of C(H ) in uppercase, and elements of L†(V) in lowercase. Proposition 2.5. Let a, b ∈ L†(V). If A and B are their respective closures, then A + B and AB exist. Proof. Let D = D (A) ∩ D (B). One has that h(A + B)ξ, ηi = hξ, (a† + b†)ηi, ∀ξ ∈ D, ∀η ∈ V, whence a† + b† ⊆ (AD + BD)∗ and A + B exists. Analogously, if D = B−1D (A), hABξ, ηi = hξ, b†a†ηi, ∀ξ ∈ D, ∀η ∈ V, whence b†a† ⊆ (ABD)∗ and AB exists. (cid:3) Remark 2.6. The analytic procedure of closure breaks down the algebraic struc- ture of L†(V). Indeed, a + b ⊆ ¯a + ¯b, ab ⊆ ¯a¯b, a† ⊆ a∗, ∀a, b ∈ LH (V), and equalities do not hold in general. Thus, given a representation π : A → L†(V), one should not expect that ¯π be a representation too, where ¯π : a ∈ A 7→ π(a) ∈ C(H ). Integrable representations are, almost by definition, those for which this is actually the case. 6 Definition 2.7. We say that a representation π : A → L†(V) is integrable if ¯π(a†) = ¯π(a)∗, for all a ∈ A. Proposition 2.8. Let π : A → L†(V) be a representation. One has that π is integrable if, and only if, ¯π(A) ⊆ C(H ) is an involutive algebra and ¯π : A → ¯π(A) is a morphism. Proof. Sufficiency is obvious. For necessity, let a, b ∈ A and A = ¯π(a), B = ¯π(b). From the proof of Proposition 2.5, A + B = (AD + BD)∗∗ ⊆ π(a† + b†)∗ = ¯π(a + b), where D = D (A) ∩ D (B) and the last equality follows from integrability. Anal- ogously, AB = (ABD)∗∗ ⊆ π(b†a†)∗ = ¯π(ab), where D = B−1D (A). (cid:3) 2.3 Symmetric involutive algebras We have seen that representations which are not integrable do not really deserve their name from an analytic point of view, and that the problem originates in the fact that L†(V) is not a good replacement for B(H ) in generalizing the theory of C-algebras to unbounded operator algebras. Now, the need for such a generalization has long been recognized and, over time, the notion of partial involutive algebras [1] has emerged as the safest candidate (because of its generality). In that approach, one actually gives up the structure of algebra, fully acknowledging the fact that strong sum and strong product are not defined all over C(H ) × C(H ). Here, we will limit ourselves to work with subsets of C(H ) which are involutive algebras with respect to strong sum, strong product and operator involution. Two known examples are: the set of measurable operators affiliated with a von Neumann algebra admitting a normal, semifinite trace [16]; and the so-called symmetric involutive algebras [9], on which we base our approach to the well-behaved representations of involutive algebras. Definition 2.9. A unital, involutive algebra is said to be symmetric if 1 + a†a is invertible, for all a ∈ A. There is a generalization to non-unital algebras, see [7], for instance. Repre- sentations π : A → L†(V) of symmetric involutive algebras enjoy several good properties, among which we mention: • They are integrable [10]. We will revisit this in the course of this paper. • They are direct sums of cyclic representations. We remark that this is not always the case if A is any involutive algebra, see [14, Corollary 11.6.8]. 7 3 Integrable representations and Ore localization 3.1 Integrability and symmetry All over this subsection, π : A → L†(V) will be a representation and n Y i=1 S =  (cid:16)1 + a† i ai(cid:17) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ai ∈ A  . If π is integrable, it follows immediately from Propositions 2.8 and 3.21 that ¯π(s) is invertible, for all s ∈ S. Here we show that these two properties are actually equivalent, and then give an alternative proof of Inoue's result asserting that representations of symmetric involutive algebras are integrable. Lemma 3.1. Let s = Qn i=1(1 + a† i ai) ∈ S and S = n Y i=1 (1 + A∗ i Ai), Ai = ¯π(ai). If ¯π(s) is surjective, then ¯π(s) = S. In particular, ¯π(s) is invertible and has a bounded inverse. Proof. Observe that S = (1 + A∗ 1A1)(1 + A∗ 2A2)D2 · · · (1 + A∗ i+1Ai+1)−1Di and D1 = D(cid:16)1 + A∗ nAn)Dn 1A1(cid:17). It follows that S is where Di+1 = (1 + A∗ closed (for it is invertible with bounded inverse). Since π(s) ⊆ S, we conclude that ¯π(s) ⊆ S, too. But, by hypothesis, ¯π(s) is surjective, whence it does not admit any injective, strict extension and must be equal to S. (cid:3) Corollary 3.2. If ¯π(s) is surjective for all s ∈ S, then ¯π(st) = ¯π(s) ¯π(t), ∀s, t ∈ S. Proof. It is a straightforward consequence of Lemma 3.1. (cid:3) Lemma 3.3. Let a ∈ A and s ∈ S. If ¯π(t) is surjective for all t ∈ S, then ¯π(s)V is a core for both ¯π(a) and ¯π(a)∗. Proof. Let t = 1+a†a ∈ S. From Lemma 3.1 and Proposition 3.21 in the appendix, we see that ¯π(t)−1 : (H , k k) → (cid:16)D ( ¯π(a)) , k k ¯π(a)(cid:17) 8 is continuous and has dense range. Therefore, ¯π(t)−1D will be a core for ¯π(a), for any dense D ⊆ H . Now, consider D = ¯π(ts)V. It is a dense subspace, for ¯π(ts) is surjective and V is a core for it. Using Corollary 3.2, we conclude that ¯π(t)−1 ¯π(ts)V = ¯π(s)V is a core for ¯π(a). The fact that it is also a core for ¯π(a)∗ follows from the same argument, applied this time to t = 1 + aa†. Indeed, now ¯π(t)−1 : (H , k k) → (cid:16)D ( ¯π(a)∗) , k k ¯π(a)∗(cid:17) is continuous and has dense range. (cid:3) Corollary 3.4. π is integrable if, and only if, ¯π(s) is surjective, for all s ∈ S. Proof. Indeed, given a ∈ A, ¯π(a†) ⊆ ¯π(a)∗. But, being V a core for both of them, they must actually coincide. (cid:3) We finish this subsection with the following result. As mentioned before, our main result is a partial converse. Proposition 3.5. Suppose that A ⊆ B, where B is an involutive algebra such that the elements of S are invertible in it. If π admits an extension π : B → L†(V), then it is integrable. Proof. Given s = Qn the proof of Lemma 3.1, i=1(1 + a† i ai), we have to prove that ¯π(s) is surjective. As in ¯π(s) ⊆ S = n Y i=1 (1 + A∗ i Ai), Ai = ¯π(ai). Now, let ξ ∈ H and consider a sequence { vn } ⊆ V converging to ξ. Since S−1 is bounded, un = S−1vn converges to, say, η ∈ H . But, by hypothesis, s is invertible in B, and un = S−1vn = S−1π(s) π(s−1)vn = π(s−1)vn, so that un → η and π(s)un = vn → ξ. By closedness, ¯π(s)η = ξ, as needed. (cid:3) Remark 3.6. As a consequence, we recover the following result due to Inoue [9]: if A is symmetric, then its representations are integrable. 9 3.2 Ore localization for involutive algebras Let A be a unital ring and S ⊆ A a multiplicative subset (that is, such that SS = S. Note that, in particular, 1 ∈ S). Proposition 3.5 leads to study the problem of adjoining inverses to the elements of S. It is not hard to see that there exists a ring AS, called universal localization of A at S, which is a universal solution to this. Now, when localizing, unexpected things can happen (such as ending up with a trivial ring) and it is good to have conditions controlling AS and, above all, enabling one to make calculations in it. From our point of view, having a manageable localization is important to investigate two problems: 1. The extension of states from A to AS. By GNS construction, this amounts to obtaining integrable extensions of cyclic representations. 2. The possibility that integrable representations actually extend to repre- sentations of AS, providing a converse to Proposition 3.5. In this section we revisit the Ore construction, which deals with the particular case in which AS is the non-commutative analogue of a ring of fractions. A complete introduction can be found in [8]. The novelty here is that we show that the Ore construction carries over smoothly to the case of involutive algebras. Definition 3.7. Let A be a ring. We say that a subset S ⊆ A is a right (resp. left) Ore set if it is multiplicative and ∀(a, s) ∈ A × S, ∃(b, t) ∈ A × S, at = sb (resp. ta = bs), and we say that it is an Ore set if it is both a left and a right Ore set. Remark 3.8. If s and t are invertible, the equation at = sb can be rewritten as s−1a = bt−1. In intuitive terms, the Ore property is establishing the possibility of writing "non-commutative fractions" indifferently from the right or from the left. We say that a ring morphism f : A → B is S-inverting if f (s) is invertible in B, for all s ∈ S. Then, consider the following universal problem for the S-inverting morphism ℓ: A ℓ f AS f B, where f is any S-invertible morphism. If S ⊆ A is an Ore set, the solution can be constructed as follows. On the set A × S define the equivalence relation (a, s) ∼ (b, t) ⇔ ∃u, v ∈ A, (au, su) = (bv, tv) ∈ A × S, 10 and denote by [a, s] the equivalence class of (a, s). Linear combination and multiplication of elements [a1, s1], [a2, s2] ∈ (A × S)/∼ are defined as follows: • Choose (b, t) ∈ A × S such that s1t = s2b. Then, λ[a1, s1] + [a2, s2] = [λa1t, s1t] + [a2b, s2b] = [λa1t + a2b, s1t]. • Choose (b, t) ∈ A × S such that a2t = s1b. Then, [a1, s1] · [a2, s2] = [a1b, s2t] (cid:16) = [a1b, s1b][a2t, s2t]. (cid:17) The last equality is enclosed in parentheses because it does not necessarily make sense; we include it because it does motivate the definition. It can be checked that this definitions equip AS−1 = (A × S)/∼ with a unital ring structure, whose unit is [1, 1]. The morphism ℓ : A → AS−1 is defined by a 7→ [a, 1]. This turns out to give an inclusion if S contains uniquely regular elements -- that is, elements s such that 0 ∈ {as, sa} ⇒ a = 0, ∀a ∈ A. An analog construction to that of AS−1 can be done if S ⊆ A is a left Ore set, and in the "bilateral" case those two localizations coincide, modulo an isomorphism whose restriction to (the image of) A is trivial. Remark 3.9. Before going further, we note a simple property of the multipli- cation. Given (a, s) ∈ A × S, let (b, t) ∈ A × S be such that at = sb. Then, by definition, [1, s][a, 1] = [b, t]. Now, given any u ∈ A such that us ∈ S, one still has that uat = usb, and therefore [1, us][ua, 1] = [b, t] = [1, s][a, 1]. We are ready to treat the case of a unital involutive algebra A, which is not dealt with in ring theory and constitutes our humble contribution to the subject. Suppose that S ⊆ A is an Ore subset such that S† = S (observe that, in this case, the left and right Ore conditions are equivalent). On AS−1 define the following operation: [a, s]† = [1, s†][a†, 1]. By Remark 3.9, this depends uniquely on the equivalence class of (a, s) because, given u ∈ A such that su ∈ S, [au, su]† = [1, u†s†][u†a†, 1] = [1, s†][a†, 1] = [a, s]†. Proposition 3.10. (·)† : AS−1 → AS−1 is an involution, and ℓ : A → AS−1 is a morphism of unital involutive algebras. 11 Proof. In fact, given (a1, s1), (a2, s2) ∈ A × S: • Let (b, t) ∈ A × S be such that s1t = s2b. We have that (cid:16)λ[a1, s1] + [a2, s2](cid:17)† = [λa1t + a2b, s1t]† = [1, t†s† 1][ ¯λt†a† 1 + b†a† 2, 1]. On the other hand, using Remark 3.9, ¯λ[a1, s1]† + [a2, s2]† = [1, s† 2][b†a† 2, 1] 1, 1] + [1, s† 1][ ¯λa† 1][ ¯λt†a† 1][ ¯λt†a† 2][a† 2, 1] 1, 1] + [1, b†s† + b†a† 2, 1]. 1 = [1, t†s† = [1, t†s† Hence, (·)† is antilinear. • Let (b, t) ∈ A × S be such that a2t = s1b. We have that (cid:16)[a1, s1][a2, s2](cid:17)† On the other hand, = [a1b, s2t]† = [1, t†s† 2][b†a† 1, 1]. [a2, s2]†[a1, s1]† = [1, s† = [1, s† 2, 1][1, s† 1][a† 2][a† 2][1, t†][b†, 1][a† 1, 1] = [1, s† 1, 1] = [1, t†s† 2, s† 2][a† 2][b†a† 1][a† 1, 1]. 1, 1] • Finally, [a, s]†† = (cid:16)[1, s†][a†, 1](cid:17)† = [a†, 1]†[1, s†]† = [a, 1][1, s] = [a, s]. (cid:3) From now on, we will write as−1 instead of [a, s] for the elements of AS−1. 3.3 Integrable representations and representations of AS−1 i=1(1 + a† that S = n Qn All over this subsection A will be a unital involutive algebra and we will assume ai ∈ A o contains only regular elements and satisfies the Ore condition. We start by proving that every ultracyclic representation of A admits an integrable extension, which is a simple consequence of the following fact. i ai) (cid:12)(cid:12)(cid:12) Lemma 3.11. Π(A) is cofinal in Π(AS−1). Proof. Given (a, s) ∈ A × S, we will prove that (s−1a)†s−1a ≤ a†a (recall that all the elements of AS−1 can be written in the form s−1a). In order to do so, note the following order property of involutive algebras, which is a direct consequence of the definition: if x ≤ y, then a†xa ≤ a†ya. Therefore, it suffices to show that (s−1)†s−1 ≤ 1, for all s ∈ S -- which we do next. 12 Let b ∈ A. We have that 1 ≤ 1 + 2b†b + (b†b)2 = (1 + b†b)2, from which (cid:16)(1 + b†b)−1(cid:17)† (1 + b†b)−1 = (1 + b†b)−2 ≤ 1. Now, suppose that s = (1 + b† 1b1)(1 + b† 2b2) · · · (1 + b† nbn) ∈ S, bi ∈ A. Inductively, we see that (s−1)†s−1 = (1 + b† 1b1)−1 · · · (1 + b† nbn)−1(1 + b† nbn)−1 · · · (1 + b† 1b1)−1 ≤ 1. (cid:3) Corollary 3.12. Every ultracyclic representation π : A → L†(V) admits an inte- grable extension π : A → L†( V). Proof. Let f ∈ Σ(A) be the state producing, by GNS construction, the represen- tation π. Since Ah is cofinal in the real ordered vector space (AS−1)h, the restric- tion fh = f Ah admits a positive extension fh to all of (AS−1)h (see, for instance, [5, Theorem 9.8]) which, in turn, uniquely determines a state f ∈ Σ(AS−1) by f (x + iy) = fh(x) + i fh(y), x, y ∈ Ah. By GNS construction, we get an ultracyclic representation π : AS−1 → L†( V), where V = (AS−1)/ { a ∈ AS−1 f (a†a) = 0 } . The inclusion A ⊆ AS−1 induces an inclusion V ⊆ V, which is well defined because { a ∈ A f (a†a) = 0 } ⊆ { a ∈ AS−1 f (a†a) = 0 } . The restriction of π to A gives the desired extension, which is integrable thanks to Proposition 3.5. (cid:3) Remark 3.13. Results of this kind are not unknown in some important cases. For instance, in [4] Borchers and Yngvason work out the case of Borchers algebras, appearing naturally in quantum field theory. Next we prove our main result. We start with the following crucial lemma. Lemma 3.14. Suppose that π is integrable, and let (a, s), (b, t) ∈ A × S be such that ta = bs. Then, ¯π(a) ¯π(s)−1 ⊇ ¯π(t)−1 ¯π(b). In particular, ¯π(s)−1V ⊆ D ( ¯π(a)) , for all (a, s) ∈ A × S. Proof. Indeed, starting from π(t)π(a) = π(b)π(s) we find, pre-multiplying by ¯π(t)−1 and post-multiplying by ¯π(s)−1π(s)V, that ¯π(a) ¯π(s)−1π(s)V = ¯π(t)−1 ¯π(b)π(s)V. 13 Now, given ξ ∈ D ( ¯π(b)), by Lemma 3.3 there exists a sequence [scn] → ξ such that [bscn] → ¯π(b)ξ. Since ¯π(t)−1 is bounded, we see that ¯π(a)[cn] = ¯π(a) ¯π(s)−1[scn] = ¯π(t)−1[bscn] converges. Therefore, ¯π(s)−1ξ ∈ D ( ¯π(a)) and ¯π(a) ¯π(s)−1ξ = ¯π(t)−1 ¯π(b)ξ, as needed. (cid:3) Theorem 3.15. There exists a bijective correspondence between integrable representa- tions of A and representations of AS−1. Proof. By Proposition 3.5, representations of AS−1 induce, by restriction, inte- grable representations of A. In the other direction, let π : A → L†(V) be an integrable representation. We extend it to AS−1 as follows. Define V = span (cid:18)[ s∈S ¯π(s)−1V(cid:19) ⊇ V. Note that, by Lemma 3.14, π(a) := ¯π(a) V belongs to L†( V), which means verifying that: V ⊆ D ( ¯π(a)), for all a ∈ A. We will show that 1. ¯π(a) V ⊆ V. Indeed, given s ∈ S, let (b, t) ∈ A × S be such that ta = bs. One has that ¯π(a) ¯π(s)−1V = ¯π(t)−1 ¯π(b)V ⊆ ¯π(t)−1V ⊆ V. V ⊆ D(cid:16)( ¯π(a) V)∗(cid:17) = D ( ¯π(a)∗). By integrability, ¯π(a)∗ = ¯π(a†) and we already know that V ⊆ D(cid:16) ¯π(a†)(cid:17). 2. 3. ( ¯π(a) V)∗ V ⊆ V. Again, follows by integrability and the already proved fact that ¯π(a†) V ⊆ V. By the universal property of the Ore localization, the morphism π will factor through AS−1 if we prove, given s ∈ S, that π(s) is invertible in L†( V). This means verifying that ¯π(s)−1 V ∈ L†( V). But, by integrability, V is invariant under ¯π(s)−1: ¯π(s)−1 V = [ t∈S ¯π(s)−1 ¯π(t)−1V = [ t∈S ¯π(ts)−1V ⊆ V. The same calculation shows that V is invariant under ( ¯π(s)−1)∗ = ¯π(s†)−1, and the conclusion follows. (cid:3) Remark 3.16. An important question arises: do integrable representations al- ways admit an extension to the universal localization AS? Our proof makes essential use of the Ore condition, which conceivable might not always hold for subsets of C(H ) which are algebras. 14 Appendix: short reminder on unbounded operators Let H be a Hilbert space and a : D (a) ⊆ H → H a linear operator, not necessarily bounded. We say that a is closed if its graph G(a) = { ξ ⊕ aξ ξ ∈ H } ⊆ H ⊕ H is closed, and that a is closable if G(a) is a graph, namely if 0 ⊕ η ∈ G(a) ⇒ η = 0, ∀η ∈ H . In this case, there exists a unique operator ¯a such that G(a) = G( ¯a), which is called the closure of a. Remark 3.17. If a closed operator was defined all over H , by the closed graph theorem it would be bounded. While this is not always the case, it is expected that closed operators inherit some of the good behaviour of bounded operators. In order to define the adjoint a∗ of the unbounded operator a, we consider the linear function fξ = hξ, a(·)i : D (a) → C. If fξ had an extension ¯fξ ∈ H ∗, then a∗ξ should be its inverse image via the antilinear isomorphism ι : H → H ∗ induced by the inner product. Define, whence, D (a∗) = (cid:8) ξ ∈ H (cid:12)(cid:12)(cid:12) ∃C(ξ) > 0, ∀η ∈ D (a) , hξ, aηi ≤ C(ξ)kηk (cid:9) . By Riesz's theorem, if ξ ∈ D (a∗), then a∗ξ = ι−1( ¯fξ) exists. Remark 3.18. D (a∗) has no reason to be dense in H . A geometrical interpretation for the adjoint operator can be obtained through the isometry J : ξ ⊕ η ∈ H ⊕ H 7→ (−η) ⊕ ξ ∈ H ⊕ H . Indeed, it is easily shown that G(a∗) = (cid:16)J G(a)(cid:17)⊥ Another consequence is the following proposition. . In particular, a∗ is closed. Proposition 3.19. D (a∗) is dense in H if, and only if, a is closable. Moreover, in this case a∗∗ = ¯a. An operator with dense domain a : D (a) ⊆ H → H is hermitian if hξ, aηi = haξ, ηi, ∀ξ, η ∈ D (a) . This means that a∗ is an extension of a, which we denote by a ⊆ a∗. In particular, D (a) ⊆ D (a∗) and therefore a is closable. Since the closure of a is a minimal closed extension, we have that a ⊆ ¯a ⊆ a∗. 15 If ¯a = a∗, we say that a is essentially self-adjoint and that ¯a is self-adjoint -- but, of course, this is not always the case. Now we consider closed operators. In order to distinguish them from arbitrary unbounded operators, we write them in capital letters. Definition 3.20. A subspace V ⊆ H is said to be a core for the closed operator A : D (A) ⊆ H → H if V ⊆ D (A) and AV = A. By definition, a core V for A is such that { ξ ⊕ Aξ ξ ∈ V } is dense in G(A). Now, the isomorphism of vector spaces ξ ∈ D (A) 7→ ξ ⊕ Aξ ∈ G(A) enables one to pass the inner product of H ⊕ H to D (A): we obtain hξ, ηiA = hξ, ηi + hAξ, Aηi. In terms of this space, V is a core for A if, and only if, it is dense in (D (A) , k kA). We conclude the appendix with the statement of an important result -- see [5], for example, for a proof. Proposition 3.21. Let A : D (A) ⊆ H → H be a closed operator. One has that: 1. 1 + A∗A : A−1D (A∗) → H is a bijection, and its inverse (1 + A∗A)−1 : H → A−1D (A∗) ⊆ D (A) is bounded as an operator (H , k k) → (D (A) , k kA). 2. D (A∗A) = A−1D (A∗) is a core for A. References [1] J. P. Antoine, A. Inoue, and C. Trapani, Partial ∗-algebras and their operator realizations, Kluwer Academic Publishers, 2002. [2] F. Bagarello, Algebras of unbounded operators and physical applications: a survey, Rev. Math. Phys. 19 (2007), 231 -- 271. [3] S. J. Bhatt, A. Inoue, and H. Ogi, Unbounded C-seminorms and unbounded C-spectral algebras, J. Operat. Theor. 45 (2001), no. 1, 53 -- 80. [4] H. J. Borchers and J. Yngvason, Partially conmmutative moment problems, Math. Nachr. 145 (1990), 111 -- 117. [5] J. B. Conway, A course in functional analysis, Springer -- Verlag, 1994. 16 [6] D. A. Dubin and M. A. Hennings, Quantum mechanics, algebras and distribu- tions, Pitman research notes in mathematics, vol. 238, Longman scientific and technical, 1990. [7] M. Fragoulopoulou, Topological algebras with involution, North-Holland mathematical studies, vol. 200, Elsevier, 2005. [8] K. R. Goodearl and R. B. Warfield, An introduction to noncommutative noethe- rian rings, second ed., Cambridge university press, 2004. [9] A. Inoue, On a class of unbounded operator algebras, Pacific J. Math. 65 (1976), no. 1, 77 -- 95. [10] [11] , Unbounded representations of symmetric ∗-algebras, J. Math. Soc. Japan 29 (1977), no. 2, 219 -- 232. , Well-behaved ∗-representations of ∗-algebras, Acta Univ. Oulu. Ser. A Sci. Rerum Natur. 408 (2004), 107 -- 117. [12] A. Mallios and M. Haralampidou, Topological algebras and applications, American mathematical society, 2005. [13] R. T. Powers, Self-adjoint algebras of unbounded operators, Commun. Math. Phys. 21 (1971), 85 -- 124. [14] K. Schm udgen, Unbounded operator algebras and representation theory, Birkhauser, 1990. [15] , On well-behaved unbounded representations of ∗-algebras, J. Operat. Theor. 48 (2002), no. 3, 487 -- 502. [16] I. Segal, A non-commutative extension of abstract integration, Ann. Math. 57 (1953), no. 3, 401 -- 457. [17] C. Trapani, Unbounded C-seminorms, biweights, and ∗-representations of ∗- algebras: a review, Int. J. Math. Math. Sci. 2006 (2006), 1 -- 34. 17
1106.5523	2	1106	2011-07-15T14:18:58	Divisibility properties for C*-algebras	[ "math.OA" ]	We consider three notions of divisibility in the Cuntz semigroup of a C-algebra, and show how they reflect properties of the C-algebra. We develop methods to construct (simple and non-simple) C-algebras with specific divisibility behaviour. As a byproduct of our investigations, we show that there exists a sequence $(A_n)$ of simple unital infinite dimensional C-algebras such that the product $\prod_{n=1}^\infty A_n$ has a character.	math.OA	math	Divisibility properties for C ∗-algebras Leonel Robert∗ and Mikael Rørdam∗ Abstract We consider three notions of divisibility in the Cuntz semigroup of a C ∗-algebra, and show how they reflect properties of the C ∗-algebra. We develop methods to construct (simple and non-simple) C ∗-algebras with specific divisibility behaviour. As a byproduct of our investigations, we show that there exists a sequence (An) of An has a simple unital infinite dimensional C ∗-algebras such that the product Q∞ character. n=1 1 Introduction A unital embedding of a matrix algebra Mm(C) into a unital C ∗-algebra A can exist only if the equation mx = [1A] has a solution x ∈ K0(A). Thus, only C ∗-algebras in which the class of the unit in K0 is m-divisible admit a unital embedding of Mm(C). Whereas all von Neumann algebras (with no central summand of type In for n finite) have this divisibility property for all m, the same is not true for C ∗-algebras, even for the simple ones. C ∗- algebras can fail to have non-trivial projections. Even if they have many projections, as in the real rank zero case, one cannot expect to solve the equation mx = [1A] exactly in K0(A). This paper is concerned with different weaker notions of divisibility, phrased in terms of the Cuntz semigroup of the C ∗-algebra, and with how they relate to embeddability properties of the C ∗-algebra. Instead of solving the equation mx = [1A] for x ∈ K0(A), one should look for less restrictive notions of divisibility. One can try, for example, to solve the inequalities mx ≤ h1Ai ≤ nx in the Cuntz semigroup of A for fixed positive integers m and n (typically with m < n). We say that A is (m, n)-divisible if one can solve this inequality. This is one of three divisibility properties we shall consider in this paper. We show that there is a full ∗-homomorphism from CMm(C), the cone over Mm(C), into A if and only if A is (m, n)-divisible for some n. Let us mention three embedding problems that served as motivation for this paper. Let A be a unital C ∗-algebra with no non-zero finite dimensional representations. Can one always find an embedding of some unital simple infinite dimensional C ∗-algebra into A? Can one always find an embedding of CM2(C) into A whose image is full in A? Can one ∗This research was supported by the Danish National Research Foundation (DNRF) through the Centre for Symmetry and Deformation 1 always find two positive mutually orthogonal full elements in A? An affirmative answer to the former problem will imply an affirmative anwer to the second problem, which is known as the "Global Glimm Halving problem". An affirmative answer to the Global Glimm Halving problem will imply an affirmative answer to the last mentioned problem. We suspect that all three problems may have negative answers. The second and the third problem led us to consider two new notions of divisibility properties. In more detail, we say that A is weakly (m, n)-divisible if there are elements x1, . . . , xn in Cu(A) such that mxj ≤ h1Ai ≤ x1 + · · · + xn. Weak divisibility measures the rank of A in the sense that A is weakly (m, n)-divisible for some n if and only if A has no non-zero representations of dimension < m. In particular, A has no non-zero finite dimensional representations if and only if for every m there is n such that A is weakly (m, n)-divisible. We say that A is (m, n)-decomposable if there are elements y1, . . . , ym in Cu(A) such that y1 + · · · + ym ≤ h1Ai ≤ nyj. For a given m, A is (m, n)-decomposable for some n if and only if A contains m pairwise orthogonal, pairwise equivalent full positive elements. It was shown in [DHTW09] that there exists a simple unital infinite dimensional C ∗-al- gebra which does not admit a unital embedding of the Jiang-Su algebra Z. This answered in the negative a question posed by the second named author. It is implicit in [DHTW09] that this simple C ∗-algebra has bad divisibility properties, cf. Remark 3.14. This leads us to a useful observation, which loosely can be formulated as follows: if A and B are unital C ∗-algebras, and if there is a unital ∗-homomorphism from A to B, then the divisibility properties of B are no worse than those of A. In other words, if A has better divisibility properties than B, then you can not unitally embed A into B. Comparability in the Cuntz semigroup is concerned with the extent to which one can conclude that x ≤ y if the "size" of x (e.g., measured in terms of states) is (much) smaller than the "size" of y. Comparability and divisibility are probably the two most fundamental properties of the Cuntz semigroup. Good comparability and divisibility properties are necessary and sufficient conditions in Winter's theorem, [Win], to conclude that a simple, separable, unital C ∗-algebra with locally finite nuclear dimension tensorially absorbs the Jiang-Su algebra. Also, good comparability and divisibility properties are both necessary and sufficient conditions to ensure that the Cuntz semigroup of a simple, separable, unital, exact C ∗-algebra A is (naturally) isomorphic to Aff(T (A)) ⊔ V (A), cf. [PT07], [BPT08], and [ERS]. The existence of simple C ∗-algebras with bad comparability properties was discovered by Villadsen, [Vil98], in the mid 1990's. This discovery was the first indication that the Elliott conjecture could be false (in general), and it was also the first example of a simple C ∗-algebra exhibiting "infinite dimensional" behaviour. Villadsen's example in [Vil98] has been generalized extensively by several authors (including Villadsen himself) to exhibit simple C ∗-algebras with various kinds of unexpected behaviour, including many ways of failing to have good comparability properties. However, little work has been done to construct simple C ∗-algebras with bad divisibility behaviour, and the literature does not contain systematic ways of producing such examples. In this paper we show that there is a duality between comparability and divisibility (Lemma 6.1), and we use this duality to 2 construct examples of simple and non-simple C ∗-algebras with bad divisibility behaviour. We use Lemma 6.1 to obtain a result that concerns the structure of C ∗-algebras that arise as the tensor product of a sequence of unital (simple non-elementary) C ∗-algebras. Each such C ∗-algebra will of course have non-trivial central sequences. Dadarlat and Toms n=1 A of a fixed unital C ∗-algebra A contains a unital copy of an AHS-algebra without characters, then it automatically absorbs the Jiang-Su algebra. It is not known if this condition always is satisfied, even n=1 A has the Corona Factorization Property for every unital A without characters (and in particular for every unital simple C ∗-algebra A 6= C). In the other direction we give, in Section 7, an example of n=1 An, proved in [DT09] that if the infinite tensor power N∞ when A is simple and non-elementary. We show in Section 6 that N∞ a sequence of simple unital infinite dimensional C ∗-algebras whose tensor power, N∞ does not absorb (or admit an embedding of) the Jiang-Su algebra. sional C ∗-algebras such that Q∞ Non-divisibility of a C ∗-algebra can be interpreted as a degree of inhomogeneity (or "lumpiness") of the C ∗-algebra. Simple C ∗-algebras are sometimes thought of as being very homogeneous, as for example in [KOS03]. From this point of view it may at first be surprising that a simple infinite dimensional C ∗-algebra can fail to have good divisibility properties. We show that there exists a sequence (An) of simple, unital, infinite dimen- n=1 An (and also the associated ultrapowers of (An)) has a character. None of the C ∗-algebras An can have a character (being simple and not equal to C), however we can show that they posses "almost characters" as defined in Section 8. In Section 9 we consider what one might call "super-divisibility", which leads to a (new) notion of infiniteness of positive elements (and which implies that a multiple of the given element is properly infinite). We use this to reformulate the Corona Factorization Property of semigroups considered in [OPR]. We study variations of examples, originally due to Dixmier and Douady, and answer in this way two questions from [KR00] in the negative: The sum of two properly infinite positive elements need not be properly infinite, and the multiplier algebra of a C ∗-algebra which has a properly infinite strictly positive element need not be properly infinite. 2 Preliminaries Let A be a C ∗-algebra and let Cu(A) denote the Cuntz semigroup of A, i.e., the set of Cuntz equivalence classes of positive elements in A ⊗ K endowed with suitable order and addition structures. Given a positive element a in A ⊗ K, we shall denote by hai the Cuntz class of a. In [CEI08], Coward, Elliott and Ivanescu give an alternative picture of the Cuntz semigroup where Cu(A) consists of suitable equivalence classes of countably generated Hilbert C ∗-modules over A. When using the Hilbert modules picture of Cu(A), we shall denote the equivalence class of a countably generated Hilbert module H by [H]. We present here some well-known definitions and facts about the Cuntz semigroup. First of all, we shall frequently use the axioms of the category Cu, of which Cu(A) is always an object (see [CEI08]). An ordered abelian semigroup S is an object in the category Cu if 3 (A1) every increasing sequence in S has a supremum, (A2) for every u ∈ S there exists a sequence (ui) in S such that ui ≪ ui+1 and supi ui = u, (A3) if u′ ≪ u and v′ ≪ v, then u′ + v′ ≪ u + v, (A4) if (ui) and (vi) are increasing sequences then supi ui + supi vi = supi(ui + vi). Recall that u ≪ v in S if whenever v = supi vi for some increasing sequence (vi) in S, then u ≤ vi for some i. An element u ∈ S is called compact if u ≪ u. We also note the following two additional properties of the Cuntz semigroup of a C ∗- algebrawhich are note listed among the axioms of Cu. The first of them asserts that the Cuntz semigroup of a C ∗-algebra almost has the Riesz Decomposition Property, and the second states that its order relation is almost the algebraic order. (P1) if u′ ≪ u ≤ v + w, then there exist v′ and w′, with v′ ≤ u, v and w′ ≤ u, w, and such that u′ ≪ v′ + w′. (P2) if u′ ≪ u ≤ v, then there exists w such that u′ + w ≤ v ≤ u + w. For the proofs of these facts, see [Rob11, Proposition 5.1.1] for the first and [RW10, Lemma 7.1 (i)] for the second. We will also make use of the sequential continuity with respect to inductive limits of the functor Cu(·) proved in [CEI08] (see also the proof of [ERS, Theorem 4.8]). It can be stated as follows: Proposition 2.1 ([CEI08]). Let A = lim−→(Ai, ϕi,j) be a sequential inductive limit of C ∗-al- gebras. (i) For every u ∈ Cu(A) there exists an increasing sequence (ui)∞ i=1 with supremum u and such that each ui belongs to Sj Im(Cu(ϕj,∞)). (ii) If u, v ∈ Cu(Ai) are such that Cu(ϕi,∞)(u) ≤ Cu(ϕi,∞)(v), then for every u′ ≪ u there exists j such that Cu(ϕi,j)(u′) ≤ Cu(ϕi,j)(v). Remark 2.2 (The cone over a matrix algebra). Let m be a positive integer, and let CMm(C) denote the cone over Mm(C), i.e., the C ∗-algebra of all continuous functions f : [0, 1] → Mm(C) that vanish at 0. For each i, j = 1, 2, . . . , m, let eij denote the (i, j)th matrix unit in Mm(C), and denote by eij ⊗ ι the function t 7→ teij in CMm(C). Then (eii ⊗ ι)m i=1 are positive contractions in CMm(C) which are pairwise equivalent and orthogonal. (We say here that two positive elements a and b are equivalent, denoted a ∼ b, if a = xx∗ and b = x∗x for some element x in the ambient C ∗-algebra.) We recall the following well-known universal property of CMm(C) (see for example [RW10, Propositions 2.3 and 2.4]): Let A be any C ∗-algebra and let a1, a2, . . . , am be pos- itive contractions in A. Then there exists a ∗-homomorphism ϕ : CMm(C) → A satisfying ϕ(ejj ⊗ ι) = aj if and only if a1, a2, . . . , am are pairwise orthogonal and pairwise equivalent in A. 4 The following lemma is well-known: Lemma 2.3. Let A be a C ∗-algebra and let a, b1, b2, . . . , bn be positive elements in A. Then: i=1hbii if and only if for each ε > 0 there exist d1, d2, . . . , dn ∈ A such that i=1hbii ≤ hai if and only if for each ε > 0 there exist mutually orthogonal positive elements a1, a2, . . . , an in aAa such that ai ∼ (bi − ε)+ for all i. i=1 dibid∗ i . (a − ε)+ = Pn (i) hai ≤ Pn (ii) Pn Proof. (i). If hai ≤ Pn i=1hbii, then a - b1 ⊕ b2 ⊕ · · · ⊕ bn, whence (a − ε)+ = d∗(b1 ⊕ b2 ⊕ · · · ⊕ bn)d = n Xi=1 d∗ i bidi for some d = (d1, d2, . . . , dn)t ∈ Mn,1(A). The converse statement is trivial. (ii). Suppose that Pn for some d = (d1, d2, . . . , dn) in M1,n(A). Thus d∗ all i. Put ai = a1/2did∗ The converse statement is trivial. i=1hbii ≤ hai. Then (b1 − ε)+ ⊕ (b2 − ε)+ ⊕ · · · ⊕ (bn − ε)+ = d∗ad i adi = (bi − ε)+ for i a1/2. It is now straightforward to verify that the ai's are as desired. j adi = 0 if j 6= i and d∗ Here is another lemma that we will use frequently: Lemma 2.4. Let a and b be positive elements in A ⊗ K with kak ≤ 1, and let m ∈ N. The following are equivalent: (i) mhai ≤ hbi, (ii) for each ε > 0 there exist mutually orthogonal positive elements b1, b2, . . . , bm in b(A ⊗ K)b such that hbii = h(a − ε)+i for all i. (iii) for each ε > 0 there exists a ∗-homomorphism ϕ : CMm(C) → b(A ⊗ K)b such that hϕ(e11 ⊗ ι)i = h(a − ε)+i. Proof. The implications (iii) ⇒ (ii) ⇒ (i) are clear, cf. Remark 2.2 and Lemma 2.3. Let us show that (i) implies (iii). Let ε > 0 be given. By Lemma 2.3 (ii) there are mutually orthogonal positive elements b1, b2, . . . , bm in b(A ⊗ K)b such that each bj is equivalent to (a − ε)+. By the universal property of the cone CMm(C), see Remark 2.2, there is a ∗-ho- momorphism ϕ : CMm(C) → b(A ⊗ K)b satisfying ϕ(ejj ⊗ ι) = bj. Hence (iii) holds. 3 Three divisibility properties Definitions and basic properties Definition 3.1. Let A be a C ∗-algebra and fix u ∈ Cu(A). Let m, n ≥ 1 be integers. Then: 5 (i) u is (m, n)-divisible if for every u′ ∈ Cu(A) with u′ ≪ u there exists x ∈ Cu(A) such that mx ≤ u and u′ ≤ nx. The least n such that u is (m, n)-divisible is denoted by Divm(u, A), with Divm(u, A) = ∞ if no such n exists. (ii) u is (m, n)-decomposable if for every u′ ∈ Cu(A) with u′ ≪ u there exist elements x1, x2, . . . , xm ∈ Cu(A) such that x1 + x2 + · · · + xm ≤ u and u′ ≤ nxj for all j = 1, 2, . . . , m. The least n such that u is (m, n)-decomposable is denoted by ∂ivm(u, A), with ∂ivm(u, A) = ∞ if no such n exists. (iii) u is weakly (m, n)-divisible if for every u′ ∈ Cu(A) with u′ ≪ u there exist elements x1, x2, . . . , xn ∈ Cu(A) such that mxj ≤ u for all j = 1, 2, . . . , m and u′ ≤ x1 + x2 + · · · + xn. The least n such that u is weakly (m, n)-divisible is denoted by divm(u, A), with divm(u, A) = ∞ if no such n exists. Remark 3.2. In the case that u in Definition 3.1 is compact (e.g., when A is unital and u = h1Ai), the conditions above read a little easier: (i) u is (m, n)-divisible if there exists x ∈ Cu(A) such that mx ≤ u ≤ nx. (ii) u is (m, n)-decomposable if there exist elements x1, x2, . . . , xm ∈ Cu(A) such that x1 + x2 + · · · + xm ≤ u ≤ nxj for all j = 1, 2, . . . , m. (iii) u is weakly (m, n)-divisible if there exist elements x1, x2, . . . , xn ∈ Cu(A) such that mxj ≤ u ≤ x1 + x2 + · · · + xn. The three divisibility properties above are related as follows: Proposition 3.3. Let m, n ∈ N and u ∈ Cu(A). Then divm(u, A) ≤ Divm(u, A), ∂ivm(u, A) ≤ Divm(u, A), divm(u, A) ≤ ∂ivm(u, A)m. Proof. The two first inequalities are clear (take xi = x in both cases). To prove the last inequality, suppose that u is (m, n)-decomposable. We show that u is weakly (m, nm)-divisible. Let u′ ≪ u and find u′′ such that u′ ≪ u′′ ≪ u. There exist i=1 xi ≤ u and u′′ ≤ nxi for all i. We proceed to elements x1, . . . , xm in Cu(A) such that Pm find elements y(i1, . . . , ik), y(i1, . . . , ik) ∈ Cu(A), k = 1, . . . , m, ij = 1, . . . , n, satisfying (a) y(i1, . . . , ik) ≪ y(i1, . . . , ik), 6 (b) y(i1, . . . , ik−1) ≪ Pn (c) y(i1, . . . , ik−1) ≤ Pn i=1 y(i1, . . . , ik−1, i) if k ≥ 2, and u′ ≪ Pn i=1 y(i1, . . . , ik−1, i) if k ≥ 2, and u′ ≤ Pn (d) y(i1, . . . , ik) ≤ xk, i=1 y(i), i=1 y(i), (e) y(i1, . . . , ik−1, ik) ≤ y(i1, . . . , ik−1) if k ≥ 2, and y(i) ≤ u′′. The elements above are constructed inductively after k using the following fact: (∗) if x′ ≪ x ≤ nz in Cu(A), then there exist y1, . . . , yn ∈ Cu(A) such that x′ ≪ Pn yi ≤ x, and yi ≤ z, i=1 yi, which follows from Property (P1) of the Cuntz semigroup stated in the previous section. Take first k = 1. The existence of y(i), with i = 1, . . . , n, satisfying (b), (d) and (e) follows from (∗) applied to u′ ≪ u′′ ≤ nx1. The existence of y(i) ≪ y(i) satisfying (a) and (c) then follows from Axiom (A2) of the Cuntz semigroup from the previous section. Assume that 2 ≤ k ≤ m and that y(i1, i2, . . . , ik−1) and y(i1, i2, . . . , ik−1) have been found. The existence of y(i1, . . . , ik−1, i), with i = 1, . . . , n, satisfying (b), (d) and (e) follows from (∗) applied to y(i1, . . . , ik−1) ≪ y(i1, . . . , ik−1) ≤ nxk. (To see that the latter inequality holds, note that y(i1, . . . , ik−1) ≪ u′′, which follows by repeated use of (e).) The existence of y(i1, i2, . . . , ik) satisfying (a) and (c) follows from Axiom (A2). We claim that the nm elements (cid:0)y(i1, . . . , im)(cid:1) witness the weak (m, nm)-divisibility of u. Indeed, it follows from (d) and (e) that y(i1, . . . , im) ≤ xj for all j = 1, . . . , m, whence m·y(i1, . . . , im) ≤ x1 + x2 + · · · + xm ≤ u. It follows from (b) and (c) that the sum of the elements y(i1, . . . , im) is greater than or equal to u′. If any of the divisibility numbers Divm(u, A), ∂ivm(u, A), and divm(A) is less than m, then u (or a multiple of u) must be properly infinite, as shown below. We shall pursue this and related questions in more detail in Section 9. Proposition 3.4. Let A be a C ∗-algebra and let u ∈ Cu(A). (i) If u is properly infinite, then Divm(u, A) = 1 for all integers m ≥ 1. (ii) If 1 ≤ n < m are integers and if u is either (m, n)-divisible, (m, n)-decomposable or weakly (m, n)-divisible, then nu is properly infinite, i.e., nu = 2nu. (iii) If 1 ≤ n < m are integers and if u is compact and (m, n)-divisible, then u is properly infinite. 7 Proof. (i). If u is properly infinite, then mu ≤ u for all m, whence Divm(u, A) = 1 . (ii). Assume that u is weakly (m, n)-divisible and take u′ ≪ u. Then there exist x1, . . . , xn such that mxi ≤ u for all i, and u′ ≤ Pn n i=1 xi. Thus, mu′ ≤ Xi=1 mxi ≤ nu. As this holds for all u′ ≪ u, we get ((m − n) + n)u = mu ≤ nu. This entails that (k(m − n) + n)u ≤ nu for all positive integers k, whence ℓu ≤ nu for all ℓ ≥ n. In particular, 2nu ≤ nu, which implies that nu is properly infinite. Next, suppose that u is (m, n)-decomposable and let u′ ≪ u. Then there exist x1, . . . , xm such that Pm i=1 xi ≤ u and u′ ≤ nxi for all i. Thus, mu′ ≤ n m Xi=1 xi ≤ nu. Arguing as before, we conclude that nu is properly infinite. Finally note that if u is (m, n)-divisible, then it is both (m, n)-decomposable and weakly (m, n)-divisible, whence nu is properly infinite. (iii). Since Divm(u, A) = n < m and u ≪ u, there exists x such that mx ≤ u ≤ nx. Arguing as above this implies that nx is properly infinite. Remark 3.5. By functoriality, each ∗-homomorphism ϕ : A → B between C ∗-algebras A and B induces a morphism Cu(ϕ) : Cu(A) → Cu(B) which preserves order, addition, and the relation of compact containment. Thus, for each u ∈ Cu(A), and with v = Cu(ϕ)(u), we have: Divm(v, B) ≤ Divm(u, A), ∂ivm(v, B) ≤ ∂ivm(u, A), divm(v, B) ≤ divm(u, A). In particular, if A and B are unital C ∗-algebras, and if Divm(h1Bi, B) > Divm(h1Ai, A) for some m (or if the corresponding inequality holds for one of the other two divisibility numbers), then one can not find a unital embedding of A into B. Divisibility numbers thus serve as an obstruction for embedding a unital C ∗-algebra with nice divisibility properties into a unital C ∗-algebra with less nice divisibility properties. The three divisibility properties behave well with respect to inductive limits thanks to the sequential continuity of the functor Cu(·): Proposition 3.6. Let A = lim−→(Ai, ϕi,j) be a sequential inductive limit of C ∗-algebra. Let u ∈ Cu(A1) and, for each i, denote by ui ∈ Cu(Ai) and u∞ ∈ Cu(A) the images of u in Cu(Ai) and Cu(A), respectively. Then: Divm(u∞, A) ≤ inf i Divm(ui, Ai), ∂ivm(u∞, A) ≤ inf i ∂ivm(ui, Ai), If u is compact (i.e., if u ≪ u), then the above inequalities are equalities. divm(u∞, A) ≤ inf i divm(ui, Ai). 8 Proof. We will only prove the statements above in the former case; the proofs for the two other cases are similar. The inequalities Divm(u∞, A) ≤ Divm(ui, Ai), with i = 1, 2, . . . , follow from Re- mark 3.5. Suppose now that u is compact. Set Divm(u∞, A) = n. Then there exists x ∈ Cu(A) such that mx ≤ u∞ ≤ nx. By Proposition 2.1 (i) and compactness of u∞, it follows that x is the image of some y ∈ Cu(Ai) for some i. By Axiom (A2) of the Cuntz semigroup and by compactness of u∞ there exists y′ ≪ y in Cu(Ai) such that u∞ ≤ n Cu(ϕi,∞)(y′). Since the ui's are compact, Proposition 2.1 (ii) implies that there exists j > i such that m Cu(ϕi,j)(y′) ≤ uj ≤ n Cu(ϕi,j)(y′). Thus uj is (m, n)-divisible in Cu(Aj). Definition 3.7. Let A be a σ-unital C ∗-algebra. Then A contains a strictly positive element. This element represents a class in Cu(A), which is independent of the choice of the strictly positive element, and which we shall denote by hAi. If A is unital, then hAi = h1Ai. We shall write divm(A), ∂ivm(A), and Divm(A) for divm(hAi, A), ∂ivm(hAi, A), and Divm(hAi, A), respectively. If A and B are unital C ∗-algebras such that there exist unital ∗-homomorphisms A → B and B → A, then, by Remark 3.5, we must have Divm(A) = Divm(B), ∂ivm(A) = ∂ivm(B), divm(A) = divm(B) for all m. This applies in particular to the situation where A is any unital C ∗-algebra and B = A ⊗ D for some unital C ∗-algebra D which has a character. In general, if D is any unital C ∗-algebra, possibly without characters, the divisibility numbers associated with A ⊗ D are smaller than or equal to those of A. Examples and remarks Let us first examine the divisibility numbers for matrix algebras: Example 3.8. Let m ≥ 2 and k ≥ 2 be integers. Using that an elementary algebraic argument yields that (cid:0)Cu(Mk(C)), h1i(cid:1) ∼= (cid:0){0, 1, 2, 3, . . . , ∞}, k(cid:1), divm(Mk(C)) = ∂ivm(Mk(C)) = Divm(Mk(C)), and Here ⌈·⌉ and ⌊·⌋ are the "ceiling" and "floor" functions. In particular, Divm(Mk(C)) = m if and only if m k, and Divm(Mk(C)) = m + 1 if m∤ k and m(m − 1) ≤ k. m ⌋(cid:25) , if m ≤ k, if m > k. (3.1) (cid:24) k Divm(Mk(C)) =   ⌊ k ∞, 9 Definition 3.9 (The rank of a C ∗-algebra). Let A be a C ∗-algebra. Let rank(A) denote the smallest positive integer n for which A has an irreducible representation on a Hilbert space of dimension n, and set rank(A) = ∞ if A has no finite dimensional (irreducible) representation. Note that rank(A) = 1 if and only if A has a character. We remind the reader about the following classical result due to Glimm: Proposition 3.10 (Glimm). Let A be a (not necessarily unital) C ∗-algebra. Then there is a non-zero ∗-homomorphism CMn(C) → A if and only if A admits at least one irreducible representation on a Hilbert of dimension ≥ n. It follows from Remark 3.5 that divm(A), ∂ivm(A), and Divm(A) are greater than or equal to Divm(Mn(C)) if rank(A) = n. In particular, these three quantities are infinite when m > rank(A). Example 3.11 (Simple C ∗-algebras). If A is a simple, unital, infinite dimensional C ∗-al- gebra, then Divm(A), ∂ivm(A), and divm(A) are finite for all positive integers m. Indeed, by the assumption that A is infinite dimensional, it follows that there is a non-zero x ∈ Cu(A) such that mx ≤ h1Ai. As every simple unital C ∗-algebra is algebraically simple, it follows that h1Ai ≤ nx for some positive integer n, i.e., h1Ai is (m, n)-divisible. Hence Divm(A) ≤ n, which entails that also ∂ivm(A) ≤ n and divm(A) ≤ n. Example 3.12. The dimension drop C ∗-algebra Zp,q, associated with the positive integers p and q, is defined to be Zp,q = {f ∈ C([0, 1], Mp ⊗ Mq f (0) ∈ Mp ⊗ C1q, f (1) ∈ C1p ⊗ Mq}. Note that rank(Zp,q) = min{p, q}. It was shown in [Rør04, Lemma 4.2] (and its proof) that Divm(Zm,m+1) = m + 1. By Remark 3.5, it follows that if Zm,m+1 maps unitally into A, then Divm(A) ≤ m + 1. Moreover, as shown in [RW10, Proposition 5.1], if A is a unital C ∗-algebra of stable rank one, then Divm(A) ≤ m + 1 if and only if Zm,m+1 maps unitally into A. Remark 3.13 (Almost divisibility). The property "almost divisibility" of a C ∗-algebra is expressed by saying that Divm(A) ≤ m + 1 for all integers m ≥ 1. If every dimension drop algebra Zm,m+1 maps unitally into A, or if the Jiang-Su algebra maps unitally into A, then A is almost divisible. Remark 3.14 (Non-embeddability of the Jiang-Su algebra). It was shown in [DHTW09] that there is a simple unital infinite dimensional nuclear C ∗-algebra A such that the dimen- sion drop C ∗-algebra Z3,4, and hence the Jiang-Su algebra Z, do not embed unitally into A. The divisibility properties of A were not explicitly mentioned in [DHTW09], but it is easily seen (using Lemma 6.1, that is paraphrased from [Rør04, Lemma 4.3]) that Div3(A) > 4. We shall in Section 7 give further examples of simple unital infinite dimensional C ∗-algebras where the divisibility numbers attain non-trivial values. 10 Remark 3.15 (Real rank zero C ∗-algebras). It was shown in [PR04, Proposition 5.7] that if A is a unital C ∗-algebra of real rank zero then A admits a unital embedding of a finite dimensional C ∗-algebra of rank at least n if and only if rank(A) ≥ n. Combining this with Remark 3.8 we see that Divm(A) ≤ m + 1 whenever A is a unital C ∗-algebra of real rank zero and with rank(A) ≥ m(m − 1). In particular, every unital C ∗-algebra A of real rank zero and with rank(A) = ∞ is almost divisible. Kirchberg considered in [Kir06] a covering number of a unital C ∗-algebra B. Let us recall the definition: Definition 3.16 (Kirchberg). Let m ∈ N. The covering number of a unital C ∗-alge- bra B, denoted by cov(B, m), is the least positive integer n such that there exist finite dimensional C ∗-algebras F1, F2, . . . , Fn with rank(Fi) ≥ m, ∗-homomorphisms ϕi : CFi → B, and d1, d2, . . . , dn ∈ B such that 1B = Pn i=1 d∗ i ϕi(1Fi ⊗ ι)di. Kirchberg's covering number cov(B, m) relates to our divm(B) as follows. Proposition 3.17. Let B be a unital C ∗-algebra and let m be a positive integer. (i) cov(B, m) is the least n for which there exist x1, x2, . . . , xn ∈ Cu(B) such that xi ≤ h1Bi ≤ x1 + x2 + · · · + xn, xi = ki Xj=i mijyij (3.2) for some integers mij ≥ m, some positive integers ki, and some yij ∈ Cu(A). (ii) cov(B, m) ≤ divm(B) ≤ (2m − 1) cov(B, m). Proof. (i). Assume that n ≥ cov(B, m) and let Fi, ϕi : CFi → B, and di ∈ B be as in j=1 Mmij (C) with mij ≥ m. Let e(ij) be a one-dimensional projection in Mmij (C). It then follows from Lemma 2.3 that the elements Definition 3.16. Write Fi = Lki xi = hϕi(1Fi ⊗ ι)i, yij = hϕi(e(ij) ⊗ ι)i satisfy the relations in (3.2). Cu(B) satisfying (3.2). Put Fi = Lki Suppose, conversely, that n ≥ 1 is chosen such that there are elements xi and yij in j=i mijyij ≤ h1Bi it follows from Lemma 2.3 (ii) that there are mutually orthogonal positive elements aijr in B, where 1 ≤ i ≤ n, 1 ≤ j ≤ ki, 1 ≤ r ≤ mij, such that haijri = yij. We can further assume that the r positive elements aij1, . . . , aijr are pairwise equivalent. It then follows from the universal property of the cone over a matrix algebra (see Remark 2.2) that there are ∗-homomorphisms ϕi : CFi → B such that hϕi(e(ij) ⊗ ι)i = yij, where e(ij) is a one-dimensional projection in the summand Mmij (C) of Fi. The existence of di ∈ B j=1 Mmij (C). By the assumption that Pki with 1B = Pn i=1 d∗ i ϕi(1Fi ⊗ ι)di follows from Lemma 2.3 (i). Thus cov(B, m) ≤ n. (ii). To prove the first inequality, assume that divm(B) = n < ∞ and take y1, . . . , yn such that myj ≤ h1Bi ≤ y1 + · · · + yn. Then (3.2) holds with ki = 1 and xi = myi. 11 Assume next that cov(B, m) = n < ∞, and find elements xi and yij satisfying the relations in (3.2). Upon replacing yij with an integral multiple of yij we can assume that m ≤ mij < 2m for all i and j. Let zik, 1 ≤ k ≤ 2m − 1, be the sum of a suitable subset of j=i mijyij = xi. The (2m − 1)n elements (zik) will then k=1 zik = Pki the yij's such that P2m−1 witness that divm(B) ≤ (2m − 1)n. 4 The asymptotic divisibility numbers One can collect the sequence of divisibility numbers (cid:0)Divm(A)(cid:1)∞ A into a single divisibility number as follows: m=2 of a unital C ∗-algebra Div∗(A) = lim inf m→∞ Divm(A) m . In a similar way one can define ∂iv∗(A) and div∗(A). Propositions 4.1 and 4.2 below hold verbatim for those quantities as well. However, to keep the exposition bounded, we only treat the case of "Div". It follows from Proposition 3.4 that Div∗(A) = 0 if and only if A is properly infinite and that Div∗(A) ≥ 1 if A is not properly infinite. Proposition 4.1. Let A be a unital C ∗-algebra. (i) Divm(A) ≤ m Div∗(A) + 1 for all integers m ≥ 2. (ii) Div∗(A) = limm→∞ Divm(A)/m (the limit always exists, but is possibly equal to ∞). (iii) If A is not properly infinite, then Div∗(A) = 1 if and only if Divm(A) ≤ m + 1 for all integers m ≥ 2. It follows from Proposition 3.4 and from (iii) above, that A is almost divisible if and only if Div∗(A) ≤ 1 (i.e., if and only if Div∗(A) = 0 or Div∗(A) = 1). Proof. (i). If Div∗(A) = ∞ there is nothing to prove. Assume that 1 ≤ Div∗(A) < ∞. Let m ≥ 2 be given. Let L be the smallest integer strictly greater than m Div∗(A). We show that Divm(A) ≤ L. Choose α > 1 and a positive integer r0 such that α r0 + 1 r0 m Div∗(A) ≤ L. By the definition of Div∗(A) there is k ≥ r0m such that ℓ := Divk(A) ≤ αk Div∗(A). Take x ∈ Cu(A) such that kx ≤ h1Ai ≤ ℓx. Write k = rm+ d, with 0 ≤ d < m and r ≥ r0. Also, write ℓ = tr − d′, with 0 ≤ d′ < r and t ≥ 1. Put y = rx ∈ Cu(A). Then my ≤ h1Ai ≤ ty. 12 With ⌈ · ⌉ denoting the ceiling function, we have Divm(A) ≤ t = ⌈ ℓ r ⌉ k − d k = l ℓ ≤ lα ≤ lα ≤ lα mm m Div∗(A)m m Div∗(A)m m Div∗(A)m ≤ L. k − d r + 1 r0 + 1 r r0 (ii). It follows from (i) that lim supm→∞ Divm(A) m ≤ Div∗(A) = lim inf m→∞ Divm(A) m , and so the claims follows. (iii). The "if" part is trivial, and the "only if" part follows from (i). We proceed to discuss how Div∗( · ) behaves under forming matrix algebras: Proposition 4.2. Let A be a unital C ∗-algebra. (i) Div∗(Mn(A)) ≤ Div∗(A) for all integers n ≥ 2. (ii) If Cu(A) is almost unperforated, then Div∗(Mn(A)) = Div∗(A) for all integers n ≥ 2. Proof. (i) follows from Remark 3.5 (as A embeds unitally into Mn(A)). (ii). Assume that Cu(A) is almost unperforated. We show first that Div∗(A) ≤ n + 1 n − 1 Div∗(Mn(A)) (4.1) for all n ≥ 2. To see this take any integer m ≥ 2, and use Proposition 4.1 (i) to see that ℓ := Divm(Mn(A)) ≤ m Div∗(Mn(A)) + 1. Write m = r(n + 1) + d and ℓ = t(n − 1) − d′, where r and t are positive integers, 0 ≤ d < n + 1, and 0 ≤ d′ < n − 1. Identify Cu(Mn(A)) with Cu(A) in the canonical way, where h1Mn(A)i ∈ Cu(Mn(A)) is identified with nh1Ai. Under this identification we can find x ∈ Cu(A) such that mx ≤ nh1Ai ≤ ℓx. In particular, (n + 1)rx ≤ nh1Ai ≤ (n − 1)tx, which by the assumption that Cu(A) is almost unperforated implies that rx ≤ h1Ai ≤ tx. This shows that Divr(A) r ≤ t r ℓ n − 1 n − 1m ≤ r−1(cid:0) = r−1l ℓ ≤ r−1(cid:16)m Div∗(Mn(A)) + 1 + 1(cid:1) + 1(cid:17) Div∗(Mn(A)) + r−1 n n − 1 n + 1 n − 1 n − 1 ≤ 13 Div∗(Mn(A)) + r−1 n n − 1 . Now, r → ∞ as m → ∞, and so (4.1) follows by letting m tend to infinity. To complete the proof of (ii), take n ≥ 2. By (i) and (4.1) we have: Div∗(A) ≤ kn + 1 kn − 1 Div∗(Mnk(A)) ≤ kn + 1 kn − 1 Div∗(Mn(A)) for all k ≥ 1, which shows that Div∗(A) ≤ Div∗(Mn(A)). We have previously remarked that Divm(A) = ∞ whenever m > rank(A). It follows that Div∗(A) = ∞ whenever rank(A) < ∞, i.e., whenever A admits a non-zero finite dimensional representation. Remark 4.3. It can happen that Div∗(Mn(A)) < Div∗(A). Take for example A such that Mn(A) is properly infinite, but A itself is not properly infinite, cf. [Rør03]. Then Div∗(Mn(A)) = 0 and Div∗(A) ≥ 1. It is an important open problem if Div∗(A) ≤ 1 (i.e., if A is almost divisible) for every (simple) unital infinite dimensional C ∗-algebra A for which Cu(A) is almost unperforated. 5 Finite-, infinite-, and ω-divisibility The property that any of the divisibility numbers Divm(A), ∂ivm(A), and divm(A) is finite, when A is a unital C ∗-algebra, has interpretations in terms of structural properties of the C ∗-algebra A. We have already noted that the divisibility numbers always are finite when A is a simple C ∗-algebra, and the corresponding structural properties of the C ∗-algebra are, as we shall see, trivially satisfied for simple C ∗-algebras. The correct definition of "finite divisibility" in the non-unital case is what we call (m, ω)-divisibility as defined below. Definition 5.1. Let A be a C ∗-algebra, let u ∈ Cu(A), and let m be a positive integer. Then: (i) u is (m, ω)-divisible if for all u′ ∈ Cu(A) with u′ ≪ u there exists x ∈ Cu(A) such that mx ≤ u and u′ ≤ nx for some positive integer n. (ii) u is (m, ω)-decomposable if for all u′ ∈ Cu(A) with u′ ≪ u there exist elements x1, x2, . . . , xm ∈ Cu(A) such that x1 + x2 + · · · + xm ≤ u and u′ ≤ nxj for some positive integer n and for all j. (iii) u is weakly (m, ω)-divisible if for all u′ ∈ Cu(A) with u′ ≪ u there exist elements x1, x2, . . . , xn in Cu(A) such that mxj ≤ u for all j and u′ ≤ x1 + x2 + · · · + xn. Remark 5.2. If u in Definition 5.1 is compact, then u is (m, ω)-divisible, (m, ω)-decompo- sable, respectively, weakly (m, ω)-divisible if and only if Divm(u, A) < ∞, ∂ivm(u, A) < ∞, respectively, divm(u, A) < ∞, cf. Remark 3.2. 14 In the next result we express (m, ω)-divisibility in terms of structural properties of the C ∗-algebra. Part (iii) is almost contained in [Kir06] (see [Kir06, Definition 3.1] and [Kir06, Remark 3.3 (7)] and compare with Definition 3.16 and Proposition 3.17). Recall the defi- nition of the rank of a C ∗-algebra from Definition 3.9. Theorem 5.3. Let A be a σ-unital C ∗-algebra and let e be a strictly positive element of A. (If A is unital, we can take e to be the unit of A.) Put u = hei = hAi. (i) u is (m, ω)-divisible if and only if for every ε > 0 there exists a ∗-homomorphism ϕ : CMm(C) → A such that (e − ε)+ belongs to the closed two-sided ideal generated by the image of ϕ. (ii) u is (m, ω)-decomposable if and only if for every ε > 0 there exist mutually orthogonal positive elements b1, b2, . . . , bm in A such that (e − ε)+ belongs to the closed two-sided ideal generated by bi for each i. (iii) The following are equivalent: (a) u is weakly (m, ω)-divisible, (b) rank(A) ≥ m, (c) there exist ∗-homomorphisms ϕi : CMm(C) → A, i = 1, 2, . . . , n, for some n, such that (e − ε)+ belongs to the closed two-sided ideal generated by the union of the images of the ϕi's. Proof. (i). Let us assume that u is (m, ω)-divisible. Let ε > 0. Find x ∈ Cu(A) and a positive integer n such that mx ≤ u and h(e − ε/2)+i ≤ nx. Choose a positive element a in A ⊗ K such that x = hai, and choose η > 0 such that h(e − ε)+i ≤ nh(a − η)+i. By Lemma 2.4 there exists ϕ : CMm(C) → A such that hϕ(e11 ⊗ ι)i = h(a − η)+i. Then h(e − ε)+i ≤ nhϕ(e11 ⊗ ι)i which implies that (e − ε)+ belongs to the closed two-sided ideal generated by the image of ϕ. Suppose conversely that for every ε > 0 there exists ϕ : CMm → A such that (e − ε)+ is in the closed two-sided ideal generated by ϕ(e11 ⊗ ι). Set hϕ(e11 ⊗ ι)i = x. Then mx ≤ u by Lemma 2.4, while h(e − 2ε)+i ≤ nx for some positive integer n. This shows that u is (m, ω)-divisible. (ii). "Only if". Let ε > 0 and suppose that b1, b2, . . . , bm in A exist with the stipulated properties. Set hbji = xj ∈ Cu(A). Then x1 + x2 + · · · + xm = hb1 + b2 + · · · + bmi ≤ u. Since (e − ε)+ belongs to the closed two-sided ideal generated by bj, h(e − 2ε)+i ≤ nxj for some integer n ≥ 1. It follows that u is (m, ω)-decomposable. "If". If u = hei is (m, ω)-decomposable and if ε > 0, then there are positive elements a1, a2, . . . , am in A ⊗ K such that ha1i + ha2i + · · · + hami ≤ u and h(e − ε/2)+i ≤ nhaji for some positive integer n. Choose η > 0 such that h(e − ε)+i ≤ nh(aj − η)+i for all j. By Lemma 2.3 (ii) there are pairwise orthogonal positive elements b1, b2, . . . , bm in A such 15 that bj ∼ (aj − η)+. Then the closed two-sided ideal generated by bj contains (e − ε)+ for each j. (iii). (a) ⇒ (b). Assume that u is weakly (m, ω)-divisible. Suppose that A has an irreducible representation π : A → B(Ck) = Mk(C) of finite positive dimension k. Then π is necessarily surjective. Since (m, ω)-divisibility is preserved by ∗-homomorphisms (cf. Remark 3.5), we conclude that Mk(C) is weakly (m, ω)-divisible. But then k ≥ m, cf. Example 3.8. Hence (b) holds. (b) ⇒ (c). Assume that (b) holds. Let (ϕi)i∈I be the family of all non-zero ∗-ho- momorphisms ϕi : CMm(C) → A and let I be the closed two-sided ideal in A generated by the images of all ϕi's. Thus each ϕi maps CMm(C) into I. We claim that I = A. Assume, to reach a contradiction, that I 6= A. By the assumption that rank(A) ≥ m, all irreducible representations of A/I have dimension at least m. It follows from Glimm's lemma (Proposition 3.10) that there is a non-zero ∗-homomorphism CMm(C) → A/I, which by projectivity lifts to a ∗-homomorphism ϕ : CMm(C) → A. But the image of ϕ is not contained in I, which is a contradiction. For each finite subset F of I consider the closed two-sided ideal IF of A generated by Si∈F ϕi(CMm(C)). Then A is the closure of the union of the upwards directed family of ideals (IF ). Hence, for each ε > 0, there is a finite subset F of I such that (e − ε)+ belongs to IF . Thus (c) holds. (c) ⇒ (a). Assume that (c) holds. Set zi = hϕi(e11 ⊗ ι)i for i = 1, 2, . . . , n. Then mzi ≤ u for all i. Moreover, (a − ε)+ belongs to the algebraic ideal generated by the n elements ϕi(e11 ⊗ ι), whence h(a − ε)+i ≤ Pn j=1 njzj for suitable positive integers nj. Put N = P nj and let x1, x2, . . . , xN be a listing of the elements z1, . . . , zn, with zj repeated nj times. Then mxj ≤ u and (a − ε)+ ≤ x1 + x2 + · · · + xN . This shows that u is weakly (m, ω)-divisible. The theorem above can be simplified in the case where u is compact, and in particular in the case where A is unital: Corollary 5.4. Let A be a unital C ∗-algebra, and let m be a positive integer. Then: (i) Divm(A) < ∞ if and only if there exists a ∗-homomorphism ϕ : CMm(C) → A whose image is full in A. (ii) ∂ivm(A) < ∞ if and only if there exist full, pairwise orthogonal positive elements b1, b2, . . . , bm in A. (iii) The following are equivalent: (a) divm(A) < ∞, (b) rank(A) ≥ m, (c) there exist ∗-homomorphisms ϕi : CMm(C) → A, i = 1, 2, . . . , n for some n, such that the union of their images is full in A. Propositions 3.10 (Glimm), Proposition 3.3, and Corollary 5.4 (i) immediately imply: 16 Corollary 5.5. If A is a unital, infinite dimensional, simple C ∗-algebra, then the three divisibility numbers Divm(A), ∂ivm(A), and divm(A) are finite for every integer m ≥ 1. Let us also note what it means to have infinite divm( · ) numbers: Corollary 5.6. Let A be a unital C ∗-algebra. (i) A admits a character if and only if div2(A) = ∞. (ii) A admits a finite dimensional representations if and only if divm(A) = ∞ for some integer m ≥ 2. Remark 5.7 (The Global Glimm Halving Problem). Glimm's lemma (Proposition 3.10) says that there exists a non-zero ∗-homomorphism from CMn(C) into a C ∗-algebra A if and only if A admits an irreducible representation of dimension at least n. It is not known how "large" one can make the image of such a ∗-homomorphism. In particular, it is not known for which C ∗-algebras A one can find a ∗-homomorphism CMn(C) → A whose image is full in A (i.e., the image is not contained in any proper closed two-sided ideal in A). For n = 2 this problem is known as the "Global Glimm Halving Problem" (see [BK04a], [BK04b] and [KR02]). A unital C ∗-algebra A is said to have the Global Glimm Halving Property if there is a ∗-homomorphism CM2(C) → A with full image. More specifically, one can ask if any (unital) C ∗-algebra, which admits no finite di- mensional representation, satisfies the Global Glimm Halving Property. In view of Corol- lary 5.4, this problem for unital C ∗-algebras A may be restated as follows: Does Div2(A) = ∞ imply that divm(A) = ∞ for some positive integer m? For a non-unital C ∗-algebra A, the one can restate the problem in the following way: Does Div2(A) = ∞ imply that hAi fails to be (m, ω)-divisible for some positive integer m. It is shown in [KR02] that if A is a weakly purely infinite C ∗-algebra, then A is purely infinite if and only if all hereditary sub-C ∗-algebras of A have the Global Glimm Halving (It is easy to see that the rank of any weakly purely infinite C ∗-algebra is Property. infinite.) It is an open problem if all weakly purely infinite C ∗-algebras are purely infinite. Remark 5.8. Let A be a unital C ∗-algebra. It follows from Proposition 3.3 (and also from Corollary 5.4) that Divm(A) < ∞ =⇒ ∂ivm(A) < ∞ =⇒ divm(A) < ∞ for all positive integers m. None of the two reverse implications hold in general. For each integer m ≥ 2 consider the Bott projection p in C(S2m) ⊗ K. This projection has (complex) dimension m and it has no non-trivial sub-projections. The unital C ∗-al- gebra A = p(C(S2m) ⊗ K)p is a homogeneous C ∗-algebra of rank m. Hence divm(A) < ∞. Suppose that ∂ivm(A) < ∞. Then, by Corollary 5.4 (ii), there would exist full, pairwise orthogonal, positive elements b1, . . . , bm in A. This would entail that each bj is one-dimensional in each fiber of A, and hence that fj.bj is a one-dimensional projection for some fj ∈ C(S2m). But this contradicts the fact that p has no proper subprojections. 17 To see that ∂ivm(A) < ∞ does not imply Divm(A) < ∞, consider the C ∗-algebra B = C(S2)⊗K, and take the (one-dimensional) Bott projection p ∈ B and a trivial (m−1)- dimensional projection q ∈ B such that p and q are orthogonal. Put A = (p + q)B(p + q). It follows from a K-theoretical argument that p + q cannot be written as the sum of m pairwise orthogonal and equivalent projections (because [p + q] is not divisible by m in K0(A)). The unit of A can be written as the sum of m (necessarily full) projections, so ∂ivm(A) < ∞. Assume that Divm(A) < ∞. Then, by Corollary 5.4 (i), there is a ∗-ho- momorphism ϕ : CMm(C) → A whose image is full in A. As explained in Remark 2.2, this entails that there exist full, pairwise orthogonal, pairwise equivalent, positive elements b1, b2, . . . , bm in A. Arguing as in the paragraph above, we can assume that each bj is in fact a projection. But that contradicts the fact that 1A = p+q is not the sum of m pairwise equivalent projections. 6 Divisibility and comparability Let A and B be C ∗-algebras. Then there is a natural bi-additive map Cu(A) × Cu(B) → Cu(A ⊗ B), (x, y) 7→ x ⊗ y, defined as follows: If x = hai and y = hbi with a a positive element in A ⊗ K and b a positive element in B ⊗ K, then x ⊗ y = ha ⊗ bi, where we identify (A ⊗ K) ⊗ (B ⊗ K) with A ⊗ B ⊗ K. Note that x1 ⊗ y1 ≤ x2 ⊗ y2 if x1 ≤ x2 and y1 ≤ y2. Part (i) of the following result was (implicitly) proved in [Rør04, Lemma 4.3], and was used to prove that Cu(A ⊗ Z) is almost unperforated for all unital C ∗-algebras A. Lemma 6.1. Let A and B be unital C ∗-algebras and let 1 ≤ m < n be integers. (i) Let x, y ∈ Cu(A) be such that nx ≤ my. If B is (m, n)-divisible, then x ⊗ h1Bi ≤ y ⊗ h1Bi. (ii) Let x1, x2, . . . , xm, y ∈ Cu(A) be such that nxj ≤ y for all j. If B is (m, n)-decompo- sable, then (x1 + x2 + · · · + xm) ⊗ h1Bi ≤ y ⊗ h1Bi. (iii) Let x, y1, y2, . . . , yn ∈ Cu(A) be such that x ≤ myj for all j. If B is weakly (m, n)- divisible, then x ⊗ h1Bi ≤ (y1 + y2 + · · · + yn) ⊗ h1Bi. Proof. (i). Take z ∈ Cu(B) such that mz ≤ h1Bi ≤ nz. Then x ⊗ h1Bi ≤ x ⊗ nz = nx ⊗ z ≤ my ⊗ z = y ⊗ mz ≤ y ⊗ h1Bi. 18 (ii). Take z1, z2, . . . , zm ∈ Cu(B) such that z1 + z2 + · · · + zm ≤ h1Bi ≤ nzj. Then (x1 + x2 + · · · + xm) ⊗ h1Bi ≤ x1 ⊗ nz1 + x2 ⊗ nz2 + · · · + xm ⊗ nzm = nx1 ⊗ z1 + nx2 ⊗ z2 + · · · + nxm ⊗ zm ≤ y ⊗ (z1 + z2 + · · · + zm) ≤ y ⊗ h1Bi. (iii). Take z1, z2, . . . , zn ∈ Cu(B) such that mzj ≤ h1Bi ≤ z1 + z2 + · · · + zn. Then (y1 + y2 + · · · + yn) ⊗ h1Bi ≥ y1 ⊗ mz1 + y2 ⊗ mz2 + · · · + yn ⊗ mzn = my1 ⊗ z1 + my2 ⊗ z2 + · · · + myn ⊗ zn ≥ x ⊗ (z1 + z2 + · · · + zn) ≥ x ⊗ h1Bi. The lemma above can loosely be paraphrased as follows: Good divisibility properties of B ensure good comparability properties of A ⊗ B, and bad comparability properties of A ⊗ B entail bad divisibility properties of B. We proceed to show that infinite tensor products of (suitable) unital C ∗-algebras cannot have very bad comparability properties. Lemma 6.2. Let (cid:0)Ak(cid:1)∞ ∞. Then k=1 be a sequence of unital C ∗-algebras such that N := supk div2(Ak) < divm( ∞ Ok=1 Ak) ≤ N n, for all integers m ≥ 2, where n is any integer satisfying 2n ≥ m. Proof. Take n such that 2n ≥ m. For each k, find z(k) 2z(k) i=1 z(k) i . Given a multi-index (i1, i2, . . . , in) ∈ {1, . . . , N}n, put i ∈ Cu(Ak), i = 1, 2, . . . , N, such that i ≤ h1Aki ≤ PN zi1,i2,...,in = z(1) i1 ⊗ z(2) i2 ⊗ · · · ⊗ z(n) in ∈ Cu(cid:16) n Ok=1 Ak(cid:17). Then m · zi1,i2,...,in ≤ 2nzi1,i2,...,in = (2z(1) ii ) ⊗ (2z(2) i2 ) ⊗ · · · ⊗ (2z(n) in ) ≤ h1A1 ⊗ 1A2 ⊗ · · · ⊗ 1Ani i (cid:17) ⊗(cid:16) z(1) N Xi=1 i (cid:17) ⊗ · · · ⊗(cid:16) z(2) N Xi=1 i (cid:17) z(n) zi1,i2,...,in. N ≤ (cid:16) Xi=1 Xi1,i2,...,in=1 = N 19 Recall that a C ∗-algebra A has the Corona Factorization Property if and only if all full projections in M(A ⊗ K) are properly infinite. It was shown in [OPR] that if A is a sepa- rable C ∗-algebra, then A and all its closed two-sided ideals have the Corona Factorization Property if and only if for every integer m ≥ 2, and for all x′, x, y1, y2, . . . in Cu(A) such that x′ ≪ x and x ≤ myj for all j, one has x′ ≤ y1 + y2 + · · · + yN for some integer N ≥ 1. Proposition 6.3. Let (cid:0)Ak(cid:1)∞ k=1 be a sequence of unital C ∗-algebras such that sup div2(Ak) < ∞. k It follows that the C ∗-algebra N∞ N∞ Proof. Put B = N∞ k=1 Ak and all its closed two-sided ideals have the Corona Factorization Property. In particular, if A is a unital C ∗-algebra without characters then k=1 A and all its closed two-sided ideals have the Corona Factorization Property. We shall view (Bn)∞ is dense in B, and we shall identify B with Bn ⊗ Dn for all n. k=1 Ak and Dn = N∞ k=1 Ak, and for each n ≥ 1 put Bn = Nn n=1 as an increasing sequence of sub-C ∗-algebras of B such that S∞ k=n+1 Ak. n=1 Bn Let m ≥ 1 be an integer and let x′, x, y1, y2, y3, . . . in Cu(B) be such that x′ ≪ x and x ≤ myj for all j. By Lemma 6.2 there is a positive integer N such that divm(Dn) ≤ N for all n. We show that x′ ≤ y1 + y2 + · · · + yN . This will prove that B has the Corona Factorization Property. Repeated use of Proposition 2.1 (i) and (ii) shows that there exists a positive integer n, and elements x′′, y′ 1, y′ x′ ≤ x′′ ⊗ h1Dni ≤ x, 2, . . . , y′ N in Cu(Bn) such that y′ j ⊗ h1Dni ≪ yj in Cu(B); x′′ ≪ my′ j in Cu(Bn), where x 7→ x ⊗ h1Dni denotes the canonical embedding Cu(Bn) → Cu(B). We can now apply Lemma 6.1 (iii) to deduce that x′ ≤ x′′ ⊗ h1Dni ≤ (y′ 1 + y′ 2 + · · · + y′ N ) ⊗ h1Dni ≤ y1 + y2 + · · · + yN as desired. 7 Obstructions to Divisibility A trivial obstruction to (weak) divisibility of a C ∗-algebra is its rank: divm(A) < ∞ if and only if m ≤ rank(A) (by Corollary 5.4 (iii)). In this section we shall discuss ways of obtaining homogeneous C ∗-algebras with large rank and large weak divisibility constant. We use these techniques to construct unital simple C ∗-algebras with large weak divisibility constants. We remark first that Lemma 6.1 provides non-trivial obstructions to divisibility in B. Indeed, it follows by that lemma that if there exists a unital C ∗-algebra A and x, y ∈ Cu(A) such that nx ≤ my but x ⊗ h1Bi (cid:2) y ⊗ h1Bi, then Divm(B) > n. Similarly, if there exist x1, . . . , xm, y in Cu(A) such that nxj ≤ y for all j while (x1 + x2 + · · · + xm) ⊗ h1Bi (cid:2) y ⊗ h1Bi, 20 then ∂ivm(B) > n. Finally, if there exist x, y1, . . . , yn ∈ Cu(A) such that x ≤ myj for all j while x ⊗ h1Bi (cid:2) (y1 + y2 + · · · + yn) ⊗ h1Bi, then divm(B) > n. We introduce below another way to obtain bad divisibility behavior: Lemma 7.1. Let u, v ∈ Cu(A) be compact elements. If Div2(u + v, A) ≤ N then there exist x1, x2, . . . , xN in Cu(A) such that 2xi ≤ v for all i and 2v ≤ v + (2N + 1)u + N Xi=1 2xi. Proof. By assumption there exists x such that 2x ≤ u + v ≤ Nx. By compactness of u + v we can find x′′ ≪ x′ ≪ x such that u + v ≤ Nx′′. Since x′ ≪ x ≤ u + v, it follows from Property (P1) of the Cuntz semigroup (see Section 2) (leaving u unchanged) that there exists v1 such that x′ ≤ u + v1, 1 ≪ v1 such that x′′ ≤ u + v′ v1 ≤ x, v1 ≤ v. 1. Apply (P2) to the relation As x′′ ≪ u + v1 there is v′ v′ 1 ≪ v1 ≤ v to obtain v2 satisfying v′ 1 + v2 ≤ v ≤ v1 + v2. By compactness of v we can find v′ 2 ≪ v2 such that v ≤ v1 + v′ 2. Now, 2 ≪ v2 ≤ u + v ≤ Nx′′ ≤ Nu + Nv′ v′ 1, and so we can use (P1) (leaving Nu unchanged) to find x1, . . . , xN such that v′ 2 ≤ Nu + N Xj=1 xj, xj ≤ v′ 1, xj ≤ v2. It follows that 2xj ≤ v′ 1 + v2 ≤ v and that 2v ≤ 2v1 + 2v′ 2 ≤ 2x + 2Nu + 2 N Xj=1 xj ≤ v + (2N + 1)u + 2 N Xj=1 xj, as desired. The corollary below illustrates how the preceding lemma can be used to find elements with bad divisibility properties: Corollary 7.2. Let X be a compact Hausdorff space and suppose that p ∈ C(X) ⊗ K is a projection such that [1] (cid:2) (2N + 1)[p] in K0(C(X)), where 1 denotes the unit of C(X). Then Div2(h1i + hpi, C(X)) > N. 21 Proof. Consider the compact elements u = hpi and v = h1i of Cu(C(X)). Note that 2x ≤ h1i implies x = 0 and that 2h1i (cid:2) h1i + (2N + 1)hpi. The desired conclusion now follows from Lemma 7.1. The relation [1] (cid:2) (2N + 1)[p] in K0(C(X)) is satisfied whenever the (2N + 1)-fold direct sum of p with itself is a projection with non-trivial Euler class (as explained in more detail below). It is known that for each integer d ≥ 1 and for each positive integer N there exist X and p ∈ C(X) ⊗ K such that (2N + 1)[p] has non-trivial Euler class and p has rank d. The unital C ∗-algebra A = (p ⊕ 1)(cid:0)C(X) ⊗ K(cid:1)(p ⊕ 1) with this choice of X and p will then satisfy rank(A) = d and Div2(A) > N. We will now give a different method for constructing elements with large divisibility constants and large rank, where we get upper bounds and where we also can give sharper lower bounds for the weak divisibility constant. Let S2 denote the 2-dimensional sphere. Let p denote the "Bott-projection" in C(S2) ⊗ M2 ⊆ C(S2) ⊗ K, i.e., the projection associated to the Hopf line bundle over S2. For each 1 ≤ j ≤ N, let pj ∈ C((S2)N ) ⊗ K be given by pj(x1, x2, . . . , xN ) = p(xj), (x1, . . . , xN ) ∈ (S2)N . Since h1i ≤ 2hpi in Cu(C(S2)), we have Nh1i ≤ 2D N Mi=1 piE. As another obstruction to weak divisibility, we shall use the following corollary of Lemma 6.1 (iii), cf. the remarks at the beginning of this section, applied to the relations h1i ≤ 2hpii, i = 1, 2 . . . , N. Corollary 7.3. Let X be a locally compact Hausdorff space, and let q ∈ C0(X) ⊗ K be a projection. Let (pi)N i=1 be the projections in C((S2)N )⊗K defined in the preceding paragraph. Suppose that q ⊗ 1 - q ⊗ N Mi=1 pi. Then div2(hqi, C0(X)) > N. Let us now give examples of projections to which the corollary above can be applied. We will make use of characteristic classes of vector bundles. Recall that projections in C(X) ⊗ K, with X compact and Hausdorff, give rise to vector bundles over X: if p is a projection, then ηp = (Ep, X, π), with Ep = {(x, v) ∈ X × l2(N) p(x)v = v} is the vector bundle associated to p. Up to Murray-von Neumann equivalence of projections and isomorphism of vector bundles, this correspondence is a bijection. We denote by e(ηp) ∈ H ∗(X), or simply e(p), the Euler class of ηp. For the cartesian product of spheres (S2)N we have (e.g., by the Kunneth formula) that H ∗((S2)N ) ∼= C[z1, z2, . . . , zN ]/(z2 N ). 1, z2 2, . . . , z2 22 With this identification, the Euler classes of the projections pi ∈ C((S2)N ) ⊗ K defined earlier can be shown to be e(pi) = zi. Proposition 7.4. Let X be a compact Hausdorff space and let q ∈ C(X)⊗K be a projection such that e(q)N 6= 0. Then div2(h1 ⊕ qi, C(X)) > N. Proof. By Corollary 7.3 it suffices to show that (1X ⊕ q) ⊗ 1(S2)N - (1X ⊕ q) ⊗ N Mi=1 pi (7.1) in Cu(C(X) ⊗ C((S2)N )), where 1X denotes the unit in C(X). (In the formulation of the proposition above, we denoted 1X simply by 1.) Observe that the trivial rank 1 projection is a subprojection of the projection on the left-hand side of (7.1). Thus, it suffices to show that the right side of (7.1) has non-zero Euler class. Set rank(q) = k. For each positive integer i, let ci(q) ∈ H 2i(X) denote the ith char- acteristic class of q (so that ck(q) = e(q)). By the Kunneth Theorem ([MS74, Theorem A.6]), we can identify H ∗(X × (S2)N ) with H ∗(X) ⊗ H ∗((S2)N ). Then e(cid:0)(1X ⊕ q) ⊗ N Mi=1 pi(cid:1) = e(cid:0)1X ⊗ N N Mi=1 N N Mi=1 pi(cid:1) e(q ⊗ pi) pi(cid:1) e(cid:0)q ⊗ Yi=1 Xj=0 ck−j(q)e(pi)j e(pi) e(q)N 6= 0. e(1X ⊗ pi) e(pi) k N Yi=1 = = = N Yi=1 Yi=1 Yi=1 N In the above computation we have used that e(q⊗p) = Pk j=0 ck−j(q)e(p)j, for q a projection of rank k and p a projection of rank 1. To obtain the last equality we have used that e(pi)2 = 0 for all i. Let us now give examples of families of projections to which the above proposition can be applied. We shall here and in the following, whenever p is a projection (in a C ∗-algebra) and n is a positive integer, let n·p denote the n-fold direct sum, p ⊕ p ⊕ · · · ⊕ p, (in a matrix algebra over the given C ∗-algebra) of the projection p. Example 7.5. Let N be a positive integer, and let CPN denote the 2N-dimensional complex projective space. Let η denote the tautological line bundle over CPN and pη the rank 1 projection associated to it. It is known that e(pη) = z2 ∈ C[z2]/(z2N ), where we 23 have identified H ∗(CPN ) with C[z2]/(z2N ). Let d, d′ be positive integers such that dd′ < N. Then e(d·pη)d′ = z2dd′ 6= 0. It follows that (cid:22)N − 1 d (cid:23) < div2(h1 ⊕ d·pηi, C(CPN )) ≤ Div2(h1 ⊕ d·pηi, C(CPN )) ≤ (cid:24)N + d + 1/2 ⌊d/2⌋ (cid:25) . Indeed, the first inequality follows from Proposition 7.4 and the calculations made above. The second inequality follows from Proposition 3.3. The last inequality can be proved as follows: Put x = ⌊d/2 ⌋hpηi. Then 2x ≤ h1 ⊕ d · pηi. By a classical result about vector bundles (see [Hus94, Chapter 9, Proposition 1.1]) we have that 1 - k ·pη if 2N ≤ 2k − 1. It follows that or, equivalently, that h1 ⊕ d·pηi ≤ nx, if n ⌊d/2 ⌋ ≥ N + d + 1/2. 1 ⊕ d·pη - n ⌊d/2 ⌋ pη, Simple C ∗-algebras with bad divisibility In this and the following two subsections we give examples of unital simple C ∗-algebras with bad divisibility behaviour. They use the Euler class obstruction described in the following example. Example 7.6. Let d be a positive integer. Following the notation in [Rør03], for each set I = {i1, i2, . . . , ik} ⊆ {1, 2, . . . , d}, let pI be the one-dimensional projection in C((S2)d)⊗K given by pI(x) = pi1(x) ⊗ pi2(x) ⊗ · · · ⊗ pik (x), x ∈ (S2)d, where pi is as defined above Corollary 7.3. It is shown in [Rør03, Proposition 4.5] that if I1, I2 . . . , Ir are subsets of {1, 2, . . . , d} that admit a matching (i.e., Si∈F Ii ≥ F for all subset F of {1, 2, . . . , r}) then the Euler class of pI1 ⊕ pI2 ⊕ · · · ⊕ pIr is non-zero. The examples constructed in this and the following two subsectons are built on the same template described in the following lemma (which is a variation of one of Villadsen's con- structions). We retain the terminology from the example above throughout the rest of this section. j=1 be a sequence of pairwise disjoint subsets of N. Choose dn large enough so that all Jj, j = 1, 2, . . . , n, are contained in the set {1, 2, . . . , dn}. (This, of j=1 Jj.) Consider the projection qn of rank Lemma 7.7. Let (cid:0)Jj(cid:1)∞ course, can be accomplished by taking dn = Pn 2n in C((S2)dn) ⊗ K given by qn = 1 ⊕ pJ1 ⊕ 2·pJ2 ⊕ · · · ⊕ 2n−1·pJn. It follows that there is a simple unital AH-algebra A which is the inductive limit of the sequence q1(cid:0)C((S2)d1) ⊗ K(cid:1)q1 where the connecting mappings ϕn are unital. ϕ1 / q2(cid:0)C((S2)d2) ⊗ K(cid:1)q2 ϕ2 / · · · / A, 24 / / / Proof. Set Xn = (S2)dn and An = qn(cid:0)C(Xn) ⊗ K(cid:1)qn. Write Xn+1 = Xn × (S2)dn+1−dn, let πn : Xn+1 → Xn be the projection mapping, and let πm,n : Xn → Xm denote the composition map πm ◦ πm+1 ◦ · · · ◦ πn−1. Choose xn ∈ Xn for each n such that the set {πm,n(xn) n ≥ m} is dense in Xm for all m ≥ 1. Define a ∗-homomorphism ϕ0 n : C(Xn, K) → C(Xn+1, K) by ϕ0 n(f )(x) = f (πn(x)) ⊕ (cid:0)f (xn) ⊗ pJn+1(x)(cid:1), f ∈ C(Xn, K), x ∈ Xn+1, where we in a suitable way have identified K ⊕ (K ⊗ K) with a subalgebra of K. We also identify C(X) ⊗ K with C(X, K). We make another identification: if J ⊆ {1, 2, . . . , dn} ⊆ {1, 2, . . . , dn+1}, then the projection pJ is defined both in C(Xn) ⊗ K and in C(Xn+1) ⊗ K and pJn = pJn ◦ πn (where the former occurrence of pJn is viewed as an element in the former algebra, and the latter in the latter). We shall use the same notation for the two projections. Taking these identification a step further, we have qn = qn ◦ πn and that qn+1 = qn ⊕ 2n ·pJn+1. (These identification hold, strictly speaking, only up to conjugation with an inner automorphism on K.) In this notation we get ϕ0 n(qn)(x) = qn(πn(x)) ⊕ (cid:0)qn(xn) ⊗ pJn+1(x)(cid:1) n(qn) = qn+1 (possibly after composing ϕ0 = qn(πn(x)) ⊕ rank(qn)·pJn+1(x) = qn+1(x), for all x ∈ Xn+1, i.e., ϕ0 phism on K). This shows that ϕ0 ∗-homomorphism that arises in this way, i.e., ϕn is the restriction of ϕ0 co-restriction to An+1), and let A be the inductive limit of the sequence n with an inner automor- n maps An unitally into An+1. Let ϕn denote the unital n to An (and the A1 ϕ1 / A2 ϕ2 / A3 ϕ3 / · · · / A. Let ϕm,n : Am → An denote the composition map ϕn−1◦ϕn−2◦· · ·◦ϕm when m ≤ n. One can check that ϕm,n(f )(x) is non-zero for all x ∈ Xn if f is a function in Am = qmC(Xm, K)qm which is non-zero on at least one point in the set {πm,k(xk) m ≤ k ≤ n}. By the choice of the points xn, it follows that for each m and for each non-zero f in Am there is n ≥ m such that ϕm,n(f ) is full in An (i.e., that ϕm,n(f )(x) 6= 0 for all x ∈ Xn). This entails that A is simple. Lemma 7.8. Let N be a positive integer. In the notation of Lemma 7.7 choose the sequence (Jj)∞ j=1 such that Jj = N ·2n−1. It then follows that div2(cid:0)hqni, C((S2)dn)(cid:1) > N, Div2(cid:0)hq2i, C((S2)d2(cid:1) ≤ 3N + 4. for all n. 25 / / / / Proof. We use Proposition 7.4 to prove the first claim. It suffices to show that the Euler class of the projection N·pJ1 ⊕ 2N·pJ2 ⊕ · · · ⊕ 2n−1N·pJn is non-zero. But this follows from [Rør03, Proposition 4.5], cf. Example 7.6 above, and from the choice of the sets Jn. To prove the second claim, put x = hpJ2i and note that 2x ≤ hq2i. It follows from [Dup76, Proposition 1] that q2 - M ·pJ2 if M − 4 ≥ (2d2 − 1)/2 = 3N − 1/2. This shows that hq2i ≤ (3N + 4)x. Theorem 7.9. For each positive integer N there exists a simple unital infinite dimensional AH-algebra A such that N < div2(A) ≤ Div2(A) ≤ 3N + 4. Proof. Let A be the simple C ∗-algebra constructed in Lemma 7.7 based on the choice of j=1 made in Lemma 7.8. Then A is the inductive limit of the sequence of C ∗-algebras (Jj)∞ An = qn(cid:0)C((S2)dn) ⊗ K(cid:1)qn with unital connecting mappings. It follows from Lemma 7.8 that div2(An) > N for all n, and that Div2(A2) ≤ 3N + 4. By Proposition 3.6 and Remark 3.5, div2(A) = inf n∈N div2(An) > N, and Div2(A) ≤ Div2(A2) ≤ 3N + 4. Remark 7.10 (Initial objects). Suppose that C is a class of unital C ∗-algebra. An element A in C is an inital object in C if there exists a unital ∗-homomorphism A → B for every B in C. It is well-known that the Cuntz algebra O∞ is an initial object in the class of unital properly infinite C ∗-algebras. In fact, a unital C ∗-algebra is properly infinite if and only if it contains O∞ as a unital sub-C ∗-algebra. Every properly infinite unital sub-C ∗-algebra of O∞ is then also an initial object in the class of unital properly infinite C ∗-algebras. Hence the Cuntz-Toeplitz algebras, Tn, n ≥ 2, are initial objects and so are all unital Kirchberg algebras A for which the assignment [1A] 7→ 1 extends to a homomorphism K0(A) → Z. It was shown in [ER06] that also the class of unital C ∗-algebras of real rank zero and of infinite rank has initial objects. One can even find initial objects to this class which are simple AF-algebras (necessarily with infinite dimensional trace simplex). It follows in particular that the class of unital simple infinite dimensional C ∗-algebras of real rank zero has initial objects. Clearly, C is an initial object in the category of all unital C ∗-algebras, and so is any unital C ∗-algebra that admits a character. (Note that we do not require the unital ∗-ho- momorphism A → B to be injective.) The corollary below shows that initial objects do not exist in the general non-real rank zero case. Corollary 7.11. The class of unital simple infinite dimensional C ∗-algebras and the class of unital C ∗-algebras of infinite rank do not have initial objects. In fact, there is no unital C ∗-algebra without characters that maps unitally into every unital simple infinite dimen- sional C ∗-algebra. 26 Proof. If A is a unital C ∗-algebra that maps unitally into every unital simple infinite dimensional C ∗-algebra, then div2(A) ≥ div2(B) for all unital simple infinite dimensional C ∗-algebras B, cf. Remark 3.5, whence div2(A) = ∞ by Theorem 7.9. On the other hand, if A has no character, then div2(A) < ∞ by Corollary 5.6. The asymptotic divisibility numbers We can give a lower and an upper bound on the asymptotic divisibility constant (discussed in Section 4) for the C ∗-algebra considered above: Corollary 7.12. Let N be a positive integer, and let A be the simple AH-algebra con- structed in Theorem 7.9 associated with N. It follows that (N − 1)/2 < Div∗(A) ≤ 2N + 2. Proof. By Proposition 4.1 we get that Div∗(A) ≥ (Div2(A) − 1)/2 > (N − 1)/2. To prove the reserve inequality, take any positive integer n and put m = 2n−1. We show that Divm(An) ≤ (2N + 2)m, where An is as in the proof of Theorem 7.9. In the notation of Lemma 7.8, let x = hpJni, put u = hqni, and recall that qn is the unit of the C ∗-algebra An. By the definition of qn (in Lemma 7.8) it follows that mx ≤ u. As dim((2 + 2N)m·pJn) − dim(qn) = (2N + 2)m − 2n = 2nN ≥ dim(Xn) − 1 2 , if follows from [Dup76, Proposition 1] that (2 + 2N)m·pJn - qn, whence u ≤ (2N + 2)mx. This proves that Divm(An) ≤ (2N + 2)m. It follows from Remark 3.5 that Divm(A) ≤ (2N + 2)m whenever m is a power of 2, and this entails that Div∗(A) = lim inf m→∞ Divm(A)/m ≤ lim inf n→∞ Div2n−1(A)/2n−1 ≤ 2N + 2, as desired. We can use Lemma 7.7 to construct a unital, simple AH-algebra A such that Div∗(A) = ∞. The proof requires the following sharpening of Corollary 7.3 that may have independent interest. Corollary 7.13. Let X be a locally compact Hausdorff space, and let q ∈ C0(X) ⊗ K be a projection. Let m and N be positive integers, let I1, I2, . . . , IN be pairwise disjoint subsets of N with Ii = m − 1 for all i, and let (pIi)N i=1 be the associated projections in C((S2)(m−1)N ) ⊗ K defined in Example 7.6. Suppose that Then divm(hqi, C0(X)) > N. q ⊗ 1 - q ⊗ N Mi=1 pIi. 27 Proof. Apply Lemma 6.1 (iii) to x = h1i and yi = hpIii, and note that x ≤ myi, cf. Example 7.6. Lemma 7.14. Let (Jj)∞ Then, in the notation of Lemma 7.7, we have j=1 be a sequence of pairwise disjoint subsets of N with Jj = 22j−1 j. 2k k < div2k(cid:0)hqni, C((S2)dn)(cid:1) < ∞ if n ≥ k, and that div2k(cid:0)hqni, C((S2)dn)(cid:1) = ∞ if n < k. it suffices to show that qn ⊗LN Proof. We use Corollary 7.13 with N = 2k k and m = 2k to prove the first claim. As 1 - qn i=1 pIi has non-trivial Euler class, when I1, . . . , IN are as in Corollary 7.13. Write qn ⊗ N Mi=1 pIi = n Mj=0 N Mi=1 2max{0,j−1}·pJj ⊗ pIi = n Mj=0 N Mi=1 2max{0,j−1}·pJj∪Ii. qn ⊗ LN As explained in Example 7.6, to prove non-triviality of the Euler class of the projection i=1 pIi one needs to verify the combinatorial fact that the family of sets (Jj ∪ Ii), j = 0, . . . , n, i = 1, . . . , N, and with the set Jj ∪ Ii repeated 2max{0,j−1} times, satisfies the Marriage Lemma condition. By first exhausting the elements in the sets Ii, and using that Pk−1 j=0 2max{0,j−1} = 2k−1 < Ii, it suffices to show that the family of sets (Jj), j = k, . . . , n, with each set repeated 2j−1N = 2j+k−1k times, satisfies the Marriage Lemma condition. However, this holds because Jj = 22j−1j ≥ 2j+k−1k when j ≥ k. The second claim follows from the fact that the dimension of the projection qn is 2n and that divm(hqni, C((S2)dn)) = ∞ whenever m > dim(qn). Theorem 7.15. There is a simple unital infinite dimensional AH-algebra A which satisfies Div∗(A) = ∞. Proof. Let A be the simple AH-algebra contructed in Lemma 7.7 with respect to the choice of (Jj) from Lemma 7.14. Recall that A is an inductive limit of a sequence of unital C ∗- algebras A1 → A2 → · · · , where An = qn(cid:0)C((S2)dn) ⊗ K(cid:1)qn. It therefore follows from Lemma 7.14 that div2k (An) > 2k k, when n ≥ k, and div2k(An) = ∞ when n < k. This entails that Div2k(A) > 2k k by Proposition 3.6. Finally, by Propo- sition 4.1 (ii), Div∗(A) = lim sup k→∞ Div2k (A)/2k = ∞. As remarked in Section 4, if A is any unital C ∗-algebra, then Div∗(A) = 0 if and only if A is properly infinite, and Div∗(A) ∈ [1, ∞] otherwise. Moreover, Div∗(A) = 1 if and only if A is almost divisible (and not properly infinite). In other words, the range of the invariant 28 Div∗( · ) is contained in the set {0} ∪ [1, ∞], and Div∗(A) ≤ 1 if and only if A is almost divisible. We can easily produce examples of simple, unital, infinite dimensional C ∗-algebras A such that Div∗(A) = 0 (eg., A could be a Cuntz algebra), or such that Div∗(A) = 1 (eg., A is any simple, unital, infinite-dimensional C ∗-algebra of real rank zero, cf. Example 3.15). The theorem above provides an example of a simple, unital, infinite dimensional C ∗-algebra A where Div∗(A) = ∞. It follows from Corollary 7.12 that Div∗( · ) attains infinitely many values in the interval (1, ∞), when restricted to the class of unital, simple, infinite dimensional C ∗-algebras, and that the possible values of Div∗( · ) in this interval is upwards unbounded. We do not know if all values in the interval (1, ∞) are thus attained. For that matter we cannot exhibit any number in the interval (1, ∞) which for sure is the value of Div∗(A) for some simple unital infinite dimensional C ∗-algebra A. Divisibility of infinite tensor products C ∗-algebra (or a unital C ∗-algebra without characters), cf. [DT09]. We end this section by giving yet another class of examples of simple unital C ∗-algebras j=1 Aj, where the Aj's are unital simple infinite dimensional C ∗-algebras. In particular, such C ∗- algebras need not absorb the Jiang-Su algebra tensorially. It remains an open problem if j=1 A absorbs the Jiang-Su algebra whenever A is a simple unital infinite dimensional with bad divisibility properties. The ones we construct below are of the form N∞ N∞ neous C ∗-algebras of rank two such that N∞ tally into N∞ Jiang-Su algebra does not embed into N∞ It was shown in [HRW07, Example 4.8] that there exists a sequence (An) of homoge- n=1 An does not absorb the Jiang-Su algebra tensorially. (It is an easy conseqence of this that the Jiang-Su algebra cannot embed uni- n=k An for some k.) Of course, one can regroup the tensor factors An to get a new sequence (Bn) of unital C ∗-algebras each of which has infinite rank and where the n=1 Bn. It is not known if every unital C ∗-algebra of infinite rank admits an embedding of a unital simple infinite dimensional C ∗-algebra. If it were true, then Theorem 7.17 below would follow from [HRW07, Example 4.8] . We introduce some notation to keep track of the combinatorics. Define a total order on the set N × N0 by (k, j) ≤ (ℓ, i) ⇐⇒ k + j < ℓ + i or (k + j = ℓ + i and k < ℓ). For each (k, j) ∈ N × N0 and for each integer m ≥ k let S(m; k, j) denote the set of all m-tuples (i1, i2, . . . , im) ∈ Nm 0 such that ik = j and (ℓ, iℓ) < (k, j) for all ℓ 6= k. Lemma 7.16. Let N ≥ 1 be an integer. For each integer k ≥ 1, let (J (k) of subsets of N such that J (k) i = ∅ whenever (k, j) 6= (ℓ, i), and such that j ∩ J (ℓ) )∞ j=1 be a sequence j J (k) j = max m≥k X(i1,...,im)∈S(m;k,j) N m Yt=1 2max{it−1,0}. 29 (The quantity on the right-hand side is finite because S(m; k, j) is finite for all (k, j) and all m ≥ k, and S(m; k, j) = ∅ when m > k + j.) Let d(k) n ) ⊗ K be as defined in Lemma 7.7 associated with the sequence (J (k) n ∈ C((S2)d(k) n ∈ N and q(k) )∞ j=1. It then follows that j div2(cid:16)(cid:10)q(1) n ⊗ q(2) n ⊗ · · · ⊗ q(m) n (cid:11), C((S2)d(1) n ) ⊗ C((S2)d(2) n ) ⊗ · · · ⊗ C((S2)d(m) n )(cid:17) > N, for all positive integers n and m. Proof. Let T (n, m) be the set of all non-zero m-tuples (i1, i2, . . . , im) ∈ Nm such that 0 ik ≤ n for all k = 1, 2, . . . , m. Adopt the convention J (k) 0 = ∅ for all k and let p∅ denote the trivial (= constant) one-dimensional projection. We can then express the projection q(1) n ⊗ q(2) n ⊗ · · · ⊗ q(m) as follows: n 1 ⊕ X(i1,...,im)∈T (n,m)(cid:0) m Yt=1 2max{it−1,0}(cid:1) · pJ (1) i1 ⊗ pJ (2) i2 ⊗ · · · ⊗ pJ (m) im . By Proposition 7.4 it suffices to show that the Euler class of the projection X(i1,...,im)∈T (n,m)(cid:0)N m Yt=1 2max{it−1,0}(cid:1) · pJ (1) i1 ⊗ pJ (2) i2 ⊗ · · · ⊗ pJ (m) im is non-zero, or, equivalently, that the Euler class of the projection X(i1,...,im)∈T (n,m)(cid:0)N m Yt=1 2max{it−1,0}(cid:1) · pJ (1) i1 ∪J (2) i2 ∪···∪J (m) im i2 ∪ · · · ∪ J (m) i2 ∪ · · · ∪ J (m) i1 ∪ J (2) im is repeated N ·Qm i2 ∪ · · · ∪ J (m) i1 ∪ J (2) is non-zero. By [Rør03, Proposition 4.5], cf. Example 7.6, it suffices to show that the family of sets J (1) im , where (i1, . . . , im) ∈ T (n, m) and where the set i1 ∪ J (2) J (1) Construct a matching by selecting the matching elements for the N · Qm t=1 2max{it−1,0} copies of the set J (1) , where (k, j) is the largest of the elements (1, i1), (2, i2), . . . , (m, im). To check that this work, i.e., to see that J (k) is large enough, let T (n, m; k, j) be the set of those m-tuples (i1, i2, . . . , im) in T (n, m) for which t=1 2max{it−1,0} times, admits a matching. im inside the subset J (k) j j (k, j) = max{(1, i1), (2, i2), . . . , (m, im)}. Then T (n, m; k, j) ⊆ S(m; k, j), and so it follows by the assumption on J (k) j that J (k) j ≥ X(i1,...,im)∈T (n,m;k,j) N m Yt=1 2max{it−1,0}. The suggested matching is therefore possible. 30 C ∗-algebra A = N∞ The theorem below shows that an infinite tensor product of simple unital infinite di- mensional C ∗-algebras does not necessarily have good divisibility properties. Any such n=1 An will have (many) non-trivial central sequences, i.e., the central sequence algebra Aω ∩ A′ with respect to an ultrafilter ω on N is non-trivial. For example, CMm(C) embeds into Aω ∩ A′ for all m, albeit, not necessarily with full image. However, in the example contructed below, one cannot embed the Jiang-Su algebra into Aω ∩ A′. Theorem 7.17. For each integer N > 2, there exist a sequence (An) of unital simple infinite dimensional C ∗-algebras (in fact, AH-algebras) such that div2(cid:16) ∞ On=1 An(cid:17) > N. In particular, N∞ a unital embedding of Z. n=1 An does not absorb the Jiang-Su algebra Z, and it does not even admit j (cid:1)∞ the sequence (cid:0)J (k) Proof. Let Ak be the simple unital AH-algebra constructed in Lemma 7.7 associated with j=1 from Lemma 7.16. Then Ak is an inductive limit of a sequence, Ak(1) → Ak(2) → · · · , of unital homogenous C ∗-algebras with unital connecting maps, where Ak(n) = q(k) n (cid:16)C((S2)d(k) n ) ⊗ K(cid:17)q(k) n . n=1 An is the inductive limit of the sequence A1 → A1 ⊗ A2 → A1 ⊗ A2 ⊗ A3 → · · · The infinite tensor product N∞ with unital connecting maps. It therefore suffices to show that div2(cid:0)Nm every m, cf. Proposition 3.6. Now, Nm Ok=1 Ak(3) → · · · , Ak(1) → Ak(2) → Ok=1 k=1 Ak is the inductive limit of the sequence m Ok=1 m m k=1 Ak(cid:1) > N for with unital connecting mappings, and so, again by Proposition 3.6, it suffices to show that div2(cid:16) m Ok=1 Ak(n)(cid:17) > N for every m and n. The latter is precisely the content of Lemma 7.16. 8 Ultrapowers In this section we show that our divisibility properties behave well with respect to taking direct products and ultrapowers of sequences of unital C ∗-algebras. This has the surprising 31 consequence that such products and ultrapowers may admit characters even if all the C ∗- algebras in the ingoing sequence are unital, simple and infinite dimensional. We define the notion of "almost characters" and show that the existence of such is re- lated to the invariant div2( · ). It follows in particular that simple unital infinite dimensional C ∗-algebras can have almost characters. First we need some technical lemmas: Lemma 8.1. Let A be a unital C ∗-algebra and let (Iλ) be an upward directed family of ideals of A. Set S Iλ = I. It follows that Divm(A/I) = inf λ Divm(A/Iλ), ∂ivm(A/I) = inf λ ∂ivm(A/Iλ) for all positive integers m. divm(A/I) = inf λ divm(A/Iλ) Proof. The inequality "≤" in all three cases follows from Remark 3.5 since we have a unital ∗-homomorphism A/Iλ → A/I for each λ. We prove the reverse inequality "≥" only in the case of ∂ivm( · ); the proofs of the other two instances are similar. Set ∂ivm(A/I) = n, and let us show that ∂ivm(A/Iλ) ≤ n for some λ. Find x1 . . . , xm in Cu(A/I) be such that x1 + x2 + · · · + xm ≤ h1i ≤ nxj for all j. Find positive contractions a1, . . . , am in A ⊗ K such that xj = hbji, where bj ∈ A/I ⊗ K is the image of aj under the quotient mapping A → A/I. Find ε > 0 such that h1i ≤ nh(bj −ε)+i for all j. It follows from [KR00, Lemma 4.12] that there are positive elements c, c′ m in I ⊗ K such that 1, . . . , c′ h(a1 − ε/2)+i + · · · + h(am − ε/2)+i ≤ h1i + hci, h1i ≤ nh(aj − ε)+i + hc′ ji for all j. There is δ > 0 such that h(a1 − ε)+i + · · · + h(am − ε)+i ≤ h1i + h(c − δ)+i, h1i ≤ nh(aj − ε)+i + h(c′ j − δ)+i. Since S Iλ is dense in I, it follows that (c−δ)+ and (c′ j −δ)+ all belong to Iλ⊗K for some λ. Let zj ∈ Cu(A/Iλ) be the Cuntz class of the image of the element (aj − ε)+ under the quotient mapping A → A/Iλ. Then z1 +· · ·+zm ≤ h1i ≤ nzj, whence ∂ivm(A/Iλ) ≤ n. For each ε > 0, let hε : R+ → [0, 1] be a continuous functions such that hε(0) = 0 and hε(t) = 1 when t ≥ ε. Lemma 8.2. Let A be a unital C ∗-algebra. Let b1, b2, . . . , bn be positive elements in A such j=1hbji ≥ h1Ai. Then, for some ε > 0, there are contractions yj in A such that that Pn n Xj=1 y∗ j hε(bj)yj = 1A. 32 Proof. By assumption, and by compactness of h1Ai, there are elements vj ∈ A such that j (bj − ε)+vj is invertible for some ε > 0, and so there are + wj and notice j (bj − ε)+wj = 1A. Put yj = (bj − ε)1/2 j bjvj = 1A. Thus Pn Pn j=1 v∗ j=1 v∗ elements wj ∈ A such that Pn that hε(bj)(bj − ε)+ = (bj − ε)+ for all j. Thus j=1 w∗ n Xj=1 y∗ j hε(bj)yj = n Xj=1 y∗ j yj = 1A, which shows that the yj's are contractions with the desired properties. Lemma 8.3. Let A be a unital C ∗-algebra, and let m, n be positive integers. (i) A is weakly (m, n)-divisible if and only if there exist positive contractions aij and contractions yj in A, j = 1, 2, . . . , n and i = 1, 2, . . . , m, such that a1j, a2j, . . . , amj are pairwise equivalent and orthogonal for all j, and such that 1A = Pn (ii) A is (m, n)-decomposable if and only if there exist pairwise orthogonal positive con- tractions ai and contractions yij in A, j = 1, 2, . . . , n and i = 1, 2, . . . , m, such that j=1 y∗ j a1jyj. Pn j=1 y∗ ijaiyij = 1A for all i. (iii) A is (m, n)-divisible if and only if there exist pairwise equivalent and pairwise or- thogonal positive contractions ai and contractions yj in A, j = 1, 2, . . . , n and i = 1, 2, . . . , m, such that Pn j=1 y∗ j a1yj = 1A. Proof. We view A as being a sub-C ∗-algebra of A ⊗ K which is identified with the corner 1A(A ⊗ K)1A. (i). "If". Put xj = ha1ji = haiji ∈ Cu(A). Then mxj = h m Xi=1 n n y∗ Xj=1 aiji ≤ h1Ai = (cid:10) j a1jyj(cid:11) ≤ j ≪ xj such that h1Ai ≤ x′ Xj=1 1+x′ hy∗ j a1jyji ≤ n Xj=1 ha1ji = n Xj=1 xj. "Only if". Choose x′ n. Use Lemma 2.3 (ii) to find positive contractions bij in A, j = 1, 2, . . . , n and i = 1, 2, . . . , m, such that b1j, b2j, . . . , bmj are pairwise orthogonal and equivalent for all j and such that x′ j ≤ hbiji ≤ xj. It then follows that 1A - b11 ⊕b12 ⊕· · ·⊕b1n, and so it follows from Lemma 8.2 that there are ε > 0 j hε(b1j)yj = 1A. The contractions yj together 2+· · ·+x′ and contractions yj in A such that Pn with the positive contractions aij = hε(bij) are then as desired. j=1 y∗ (ii). "If". Put xi = haii ∈ Cu(A). Then x1 + · · · + xm = (cid:10)a1 + · · · + am(cid:11) ≤ h1Ai = (cid:10) n Xj=1 y∗ ijaiyij(cid:11) ≤ n Xj=1 hy∗ ijaiyiji ≤ nxi for all i. "Only if". Choose x′ i ≪ xi such that nx′ orthogonal and equivalent positive contractions b1, b2, . . . , bm in A such that x′ i ≥ h1Ai. Use Lemma 2.3 (ii) to find pairwise i ≤ hbii ≤ xi. 33 Proceed as in the proof of "only if" in part (i) to obtain elements ai = hε(bi) (for a suitable ε > 0) and yij with the desired properties. (iii). "If". Put x = ha1i = haii ∈ Cu(A). Then mx = (cid:10)a1 + a2 + · · · + ami ≤ h1Ai ≤ (cid:10) n Xj=1 y∗ j a1yj(cid:11) ≤ n Xj=1 hy∗ j a1yji ≤ nx. "Only if". Choose x′ ≪ x such that h1Ai ≤ nx′. By Lemma 2.3 (ii) there are pairwise orthogonal and pairwise equivalent positive contractions b1, b2, . . . , bm in A such that x′ ≤ hbji ≤ x. We can now follow the proof of "only if" in part (i) to obtain elements aj = hε(bj) (for a suitable ε > 0) and yj with the desired properties. bounded sequences (ak), with ak ∈ Ak. If (Ak) is a sequence of C ∗-algebras, then we denote by Q∞ cω({Ak}) the closed two-sided ideal in Q∞ limω kakk = 0. Finally, denote the quotient Q∞ k=1 Ak the C ∗-algebra of all If ω is a (free) filter on N, then denote by k=1 Ak consisting of those sequences (ak) for which k=1 Ak/cω({Ak}) by Qω Ak. Proposition 8.4. Let (An) be a sequence of unital C ∗-algebras. Then, for all integers m ≥ 2 and for any free filer ω on N we have: Divm(Ak). ∂ivm(Ak). divm(Ak). ω k=1 Ak(cid:17) = supk Divm(Ak), Divm(cid:16)Qω Ak(cid:17) = lim sup k=1 Ak(cid:17) = supk ∂ivm(Ak), ∂ivm(cid:16)Qω Ak(cid:17) = lim sup k=1 Ak(cid:17) = supk divm(Ak), divm(cid:16)Qω Ak(cid:17) = lim sup (i) Divm(cid:16)Q∞ (ii) ∂ivm(cid:16)Q∞ (iii) divm(cid:16)Q∞ We have unital ∗-homomorphisms Q∞ We show next that Divm(cid:16)Q∞ Proof. We only prove (i). The proofs of (ii) and (iii) are very similar. ω ω k=1 Ak → An for all n. Therefore the inequality k=1 Ak(cid:17) ≤ supk Divm(Ak). Let n be a positive integer such that Divm(Ak) ≤ n for all k. Then, by Lemma 8.3 (i), for each k we can find positive contractions a(k) in Ak, for j = 1, 2, . . . , n and i = 1, 2, . . . , m, such that a(k) mj are pairwise orthogonal and equivalent for all j, and such that ij and contractions y(k) 2j , . . . , a(k) 1j , a(k) j "≥" holds in the first identity in (i) (and also in (ii) and (iii)), cf. Remark 3.5. 1Ak = n Xj=1 (y(k) j )∗a(k) 1j y(k) j . Put aij = (a(k) ij ) ∈ ∞ Yk=1 Ak, yj = (y(k) j ) ∈ ∞ Yk=1 Ak. 34 Then a1j, a2j, . . . , amj are pairwise orthogonal and equivalent, and Pn to the unit of Q∞ whence Divm(cid:16)Q∞ ∗-homomorphism Qk∈I Ak → Qω Ak for each I ∈ ω. We can therefore use Remark 3.5 and k=1 Ak. By Lemma 8.3 (i), this shows that Q∞ k=1 Ak(cid:17) ≤ n. To prove the second part of (i) note first that we have a natural unital (surjective) j=1 y∗ j a1jyj is equal k=1 Ak is (m, n)-divisible, the first identity in (i) to conclude that Divm(cid:0)Yω Ak(cid:1) ≤ Divm(cid:0)Yk∈I Ak(cid:1) = sup k∈I Divm(Ak), which shows that Divm(cid:0)Qω Ak(cid:1) ≤ lim supω Divm(Ak). We proceed to prove the reverse inequality. For each I ∈ ω consider the ideal J(I) in Q∞ k=1 Ak consisting of those sequences (ak) for which ak = 0 for all k ∈ I. Then cω({Ak}) = [I∈ω J(I), (where ω is ordered by reverse inclusion). We can now use Lemma 8.1 and the first identity in (i) to conclude that Divm(cid:0)Yω Ak(cid:1) = inf I∈ω Divm(cid:16)(cid:0) ∞ Yk=1 Ak(cid:1)/J(I)(cid:17) = inf I∈ω Divm(cid:0)Yk∈I Ak(cid:1) = inf I∈ω sup k∈I Divm(Ak) = lim sup ω Divm(Ak). If we combine the proposition above with Corollary 5.6 (i) we obtain: Corollary 8.5. Let (Ak) be a sequence of unital C ∗-algebras such that limk→∞ div2(Ak) = ∞. Then Q∞ k=1 Ak has a character, and so does Qω Ak for each free filter ω on N. If we combine the corollary above with Theorem 7.9, then we obtain the following surprising fact: Corollary 8.6. There is a sequence (Ak) of unital simple infinite dimensional C ∗-algebras such that Q∞ k=1 Ak and Qω Ak have characters for each free filter ω on N. Clearly, none of the C ∗-algebras Ak in the corollary above can have a character. However, they have "almost characters" in the sense defined below. This is one way of understanding how the product C ∗-algebra can have a character when none of the individual C ∗-algebras has one. Definition 8.7. Let N ≥ 2 be an integer and let ε > 0. A unital C ∗-algebra A is said to have (N, ε)-characters if for every N-tuple u1, u2, . . . , uN of unitaries in A there exists a state ρ on A such that ρ(uj) ≥ 1 − ε for j = 1, 2, . . . , N. 35 A state ρ on a unital C ∗-algebra is a character if and only if ρ(u) = 1 for all unitary elements u ∈ A. Most simple C ∗-algebras that we know of do not have (2, ε)-characters for small ε > 0. For example, if A is a C ∗-algebra which contains unitaries u, v such that kuvu∗v∗ − λ1Ak < η for some λ ∈ T and for some η < 1 − λ, then A does not admit any (2, ε)-character for some small enough ε > 0. Indeed, if ρ is a state on A such that ρ(u) and ρ(v) are close to 1, then ρ(uvu∗v∗) is close to 1. Proposition 8.8. A unital C ∗-algebra has a character if and only if it has (N, ε)-characters for all pairs (N, ε), where N ≥ 2 is an integer and ε > 0. Proof. The "only if" part is trivial. Assume that A is a unital C ∗-algebra that has (N, ε)- characters for all pairs (N, ε). For each finite subset F of the unitary group of A and for each ε > 0, let S(F, ε) denote the set of states ρ on A such that ρ(u) ≥ 1 − ε for all u ∈ F . It follows that T(F,ε) S(F, ε) is non-empty, Then, by assumption, S(F, ε) is non-empty. and any state in this intersection is a character. Proposition 8.9. Let (Ak) be a sequence of unital C ∗-algebras, and let ω be a free ultrafilter ε > 0 there exists I ∈ ω such that Ak has (N, ε)-characters for each k ∈ I. on N. Then Qω Ak has a character if and only if for each integer N ≥ 2 and for each Proof. We prove first the "if" part. By Proposition 8.8 it suffices to show that Qω Ak has (N, ε)-characters for all (N, ε). Fix (N, ε) and find I ∈ ω such that Ak has (N, ε)-characters for each k ∈ I. Let u1, . . . , uN be unitaries in Qω Ak, and let (u(k) j ) ∈ Q∞ k=1 Ak be a lift of uj. Then for each k ∈ I there is a state ρk on Ak such that ρk(u(k) j ) ≥ 1 − ε for j = 1, 2, . . . , N. Choose arbitrary states ρk on Ak for k /∈ I and define a state ρ on Qω Ak by ρ(x) = limω ρk(xk), where (xk) ∈ Q∞ k=1 Ak is a lift of x. (A priori, ρ defines a state on Q∞ k=1 Ak, and one checks that it vanishes on the ideal cω({Ak}).) Then ρk(u(k) ρk(u(k) j ) ≥ 1 − ε, ρ(uj) = lim k→ω j ) ≥ inf k∈I for j = 1, 2, . . . , N, which shows that Qω Ak has (N, ε)-characters. Suppose next that Qω Ak has a character ρ. Fix (N, ε), and let J be the set of those k ∈ N for which Ak does not have (N, ε)-characters. For each k ∈ J choose unitaries u(k) j in Ak, j = 1, 2, . . . , N, such that there is no state ρ′ on Ak for which ρ′(u(k) j ) ≥ 1 − ε for all j = 1, 2, . . . , N. Choose arbitrary unitaries u(k) j ∈ Ak for k /∈ J, and let uj be the unitary element (u(k) j ) in Q∞ by the unitaries πω(uj), where πω is the quotient mapping Q∞ for all x = (xk) ∈ Q∞ k=1 Ak. Let B be the (separable) sub-C ∗-algebra of Qω Ak generated k=1 Ak → Qω Ak. By [Kir06, Lemma 2.5] there is a sequence ρk of pure states on Ak such that ρ(πω(x)) = limω ρk(xk) k=1 Ak with πω(x) ∈ B. Now, 1 = ρ(πω(uj)) = lim ω ρk(u(k) j ) = lim inf k→ω ρk(u(k) j ) = sup I∈ω ρk(u(k) j ). inf k∈I It follows that there exists I ∈ ω such that ρk(u(k) j ) ≥ 1 − ε for all k ∈ I and for all j = 1, 2, . . . , N. This entails that I ∩ J = ∅. Hence Ak has (N, ε)-characters for all k ∈ I. 36 We can relate the existence of (N, ε)-characters on a C ∗-algebra A to the divisibility quan- tity div2(A). Theorem 8.10. For each pair (N, ε), where N ≥ 2 is an integer and ε > 0, there exists an integer n ≥ 2 such that every unitalC ∗-algebra A which satisfies div2(A) ≥ n has (N, ε)- characters. Conversely, for every integer n ≥ 2 there exists a pair (N, ε), where N ≥ 2 is an integer and where ε > 0, such that every unital C ∗-algebra A which has (N, ε)-characters satisfies div2(A) ≥ n. Proof. Suppose that the first claim were false. Then there would exist a pair (N, ε) and a sequence (An) of unital C ∗-algebras such that div2(An) ≥ n and none of the An's have (N, ε)-characters. However, if ω is any free ultrafilter N, then Qω An has a character by Corollary 8.5, whence An has (N, ε)-characters for each n in some subset I ∈ ω, a contradiction. Suppose next that the second statement were false. Then there would exist an integer n ≥ 2 and a sequence (Ak) of unital C ∗-algebras such that Ak has (k, 1/k)-characters but div2(Ak) < n. Let ω be a free ultrafilter on N. It then follows from Proposition 8.9 that Qω Ak has a character. Hence div2(Qω Ak) = ∞, whence limω div2(Ak) = ∞ by Proposition 8.4, a contradiction. Corollary 8.11. For each pair (N, ε), where N ≥ 2 is an integer and ε > 0, there exists a unital simple infinite dimensional C ∗-algebra which has (N, ε)-characters. We end this section by giving several equivalent formulation of some well-known open problems for C ∗-algebras. Recall that a C ∗-algebra A has the Global Glimm Halving property if there is a ∗-homomorphism CM2(C) → A whose image is full in A. Proposition 8.12. The following statements are equivalent: (i) Every unital C ∗-algebra that has no finite dimensional representation has the Global Glimm Halving property. (ii) For all unital C ∗-algebras A, if divm(A) < ∞ for all m ≥ 2, then Div2(A) < ∞. (iii) For every sequence (Ak) of unital C ∗-algebras, if supk divm(Ak) < ∞ for all m, then supk Div2(Ak) < ∞. Proof. (i) ⇔ (ii). A has no finite dimensional representations if and only if rank(A) ≥ m for all m, which by Corollary 5.4 (iii) is equivalent to divm(A) < ∞ for all m. It was shown in Corollary 5.4 (i) that the Global Glimm Halving property holds for A if and only if Div2(A) < ∞. (ii) ⇒ (iii). Given a sequence (Ak) of unital C ∗-algebras such that supk divm(Ak) < ∞ k=1 Ak. Then divm(A) = supk divm(Ak) < ∞ by Proposition 8.4. Thus Div2(A) < ∞, which implies that supk Div2(Ak) = Div2(A) < ∞, again by Proposition 8.4. for all m. Consider the C ∗-algebra A = Q∞ (iii) ⇒ (ii) is trivial: Take Ak = A for all k. 37 Proposition 8.13. The following statements are equivalent: (i) All unital C ∗-algebras A that have no finite dimensional representation contain two positive full elements that are orthogonal to each other. (ii) For all unital C ∗-algebras A, if divm(A) < ∞ for all m ≥ 2, then ∂iv2(A) < ∞. Proof. (i) ⇔ (ii). As in the proof of Proposition 8.12, A has no finite dimensional rep- resentations if and only if divm(A) < ∞ for all m. It was shown in Corollary 5.4 (ii) that A contains two positive full elements that are orthogonal to each other if and only if ∂iv2(A) < ∞. (ii) ⇒ (iii) is similar to the proof of (ii) ⇒ (iii) in Proposition 8.12. (iii) ⇒ (ii) is trivial. 9 Infinite elements Following [KR00], a Cuntz class u in the Cuntz semigroup of a C ∗-algebra A is said to be properly infinite if it satisfies u = 2u (whence u = ∞ · u). Similarly, a countably generated Hilbert module over A is properly infinite if its Cuntz class is properly infinite. We saw in Proposition 3.4 how infiniteness of a Cuntz class can arise from a certain type of divisibility property. In this section we shall investigate this and related phenomena further with emphasis on the following property: Definition 9.1. Let A be a C ∗-algebra, let n ≥ 1 be an integer, and let u be an element in i=1 xi ≤ u Cu(A). We say that u is (ω, n)-decomposable if there exist x1, x2, . . . such that P∞ and u ≤ nxi for all i. If u is (ω, n)-decomposable, then u is (m, n)-decomposable for all m. Proposition 3.4 (ii), it follows that nu is properly infinite. In particular, by The condition in the definition above can be reformulated in several different ways: Lemma 9.2. Let A be a C ∗-algebra, let n ≥ 1 be an integer, and let u be an element of Cu(A). Then the following conditions are equivalent: (i) u is (ω, n)-decomposable, (ii) there exist x1, x2, . . . such that P∞ (iii) there exist x1, x2, . . . and y1, y2, . . . such that P∞ (iv) there exist x1, x2, . . . such that P∞ and ∞·u ≤ ∞·sup yi, i=1 xi ≤ u and nP∞ j=k xj = ∞·u for all k. i=1 xi ≤ u and nxi = ∞·u for all i, i=1 xi ≤ u, yi−1 ≤ yi ≤ nxi for all i, 38 Proof. (i) ⇒ (ii). Suppose that x1, x2, . . . satisfy P∞ partition of the natural numbers into infinite sets. Then the elements x′ that condition (ii) holds. i=1 xi ≤ u ≤ nxj. Let {Ii}∞ i = Pj∈Ii i=1 be a xj witness To get (ii) ⇒ (iii), set yi = ∞·u for all i and choose (xi)∞ (iii)⇒ (iv). Let (xi)∞ i=1 and (yi)∞ i=1 be as in (iii). Then P∞ the left side of this inequality decreases as k increases. Thus, P∞ k′ ≥ k. Taking the supremum over all k′ ≥ k we get P∞ i=k xi ≥ ∞·sup yi ≥ ∞·u. (iv) ⇒ (i). Suppose that x1, x2, . . . satisfy the condition in (iv). Let (ui)∞ i=1 be such that ui ≪ ui+1 for all i and supi ui = u. Then there exists a sequence 1 = k0 < k1 < · · · such that the elements x′ i=1 be a partition of the natural numbers into infinite sets. Then the elements x′′ x′ j satisfy the condition in Definition 9.1. i ≥ ui for all i. Let {Ii}∞ i = Pj∈Ii i = Pki−1 xj satisfy x′ j=ki−1 i=1 that satisfies (ii). i=k xi ≥ ∞·yk. Observe that i=k xi ≥ ∞ · yk′ for all It was shown in [OPR] that the Corona Factorization Property for a C ∗-algebra is equivalent to a condition for its Cuntz semigroup, that we here shall refer to as (CFP4S). A complete ordered abelian semigroup is said to have (CFP4S) if whenever (xi)∞ i=1 is a full sequence, x′ ≪ x1, and (yi)∞ i=1 is such that myi ≥ xi for all i and some m, then there exists n such i=1 yi ≥ x′. Recall that a full sequence is one that is increasing and such that sup xi is a full element, i.e., ∞ · sup xi is the largest element of the semigroup (which we shall denote by ∞). that Pn In Section 6 we discussed a related notion, called the strong Corona Factorization Property, and its analog for the Cuntz semigroup. The proposition below relates the (CFP4S) with the notion of (ω, m)-divisibility. In fact, it is a consequence of this proposition that a semigroup in the category Cu has a full elements which is (ω, m)-divisible and not properly infinite if and only if the semigroup does not satisfy (CFP4S). Proposition 9.3. The following four conditions are equivalent for any object S in the category Cu. (i) S has property (CFP4S). (ii) For every sequence (yi)∞ i=1 in S, if there is a full sequence (xi)∞ i=1 in S such that myi ≥ xi for some m and for all i, then P∞ i=1 yi = ∞. (iii) For every sequence (yi)∞ ∞. P∞ (iv) For every sequence (yi)∞ i=1 yi = ∞. i=1 in S, if myi = ∞ for some m and for all i, then P∞ i=1 in S, if P∞ i=n myi = ∞ for some m and for all n, then i=1 yi = (v) For every full element y in S, if y is (ω, m)-decomposable for some m, then y is properly infinite (whence y = ∞). 39 Proof. (i) ⇒ (ii). Apply the (CFP4S) to the tail sequences (xi)∞ i=n and (yi)∞ i=n. Then we i=n yi for all x′ ≪ xn, whence ∞ = supn xn ≤ P∞ (ii) ⇒ (i). If (xi) is a full sequence and if (yi) is another sequence such that xi ≤ myi, i=1 yi for some n by the i=1 yi = ∞ by (ii). In particular, if x′ ≪ x1, then x′ ≤ Pn i=1 yi. (ii) ⇒ (iii). Suppose that (ii) holds and that (yi) is a sequence in S such that myi = ∞ i=n yi ≤ P∞ definition of compact containment. get x′ ≤ PN then P∞ for all i. Then ∞ = P∞ i=1 yi ≤ y and myi = ∞ · y = ∞ for all i. But then P∞ P∞ i=n myi = ∞ for some m and for all n. Put y = P∞ P∞ (v) ⇒ (iv). Suppose that (v) holds and let (yi)∞ i=1 yi by (ii) with xi = ∞ for all i. whence y = ∞. (iii) ⇒ (v). Suppose that (iii) holds, and that y ∈ S is full and (ω, m)-decomposable. Then y satisfies condition (ii) of Lemma 9.2, so there exists a sequence (yi) such that i=1 yi = ∞ because (iii) holds, i=1 be a sequence in S such that i=1 yi. We must show that y = ∞. We know that my = ∞, so y is full. It is easy to see that y satisfies condition (iv) of Lemma 9.2, so y is (ω, m)-decomposable. Hence y = ∞ by the assumption that (v) holds. i=1 be a full sequence in S, let m ≥ 1 be a positive integer, and let (iv) ⇒ (ii). Let (xi)∞ (yi) be such that myi ≥ xi for all i. Then ∞ Xi=n myi ≥ ∞ Xi=n xi ≥ ∞ Xi=k xi ≥ ∞·xk for all k ≥ n. As ∞ = supk ∞·xk, we conclude that P∞ i=1 yi = ∞, and in particular that x1 ≤ P∞ entails that P∞ i=1 yi. i=n myi = ∞ for all n. By (iv) this In the following example we describe a Cuntz semigroup with an element u that is (ω, 2)- decomposable but not properly infinite. In particular, 2u is properly infinite while u is not. This example is well known in other contexts, and it was discussed in the paragraph preceding Corollary 7.3. Example 9.4. Let X = (S2)∞ be a countable cartesian product of 2-dimensional spheres, and let pi ∈ C(X, K) be the one-dimensional projection arising as the pull back of a non- trivial rank one projection p in C(S2) ⊗ K along the ith coordinate projection X → S2, cf. the comments above Corollary 7.3. Let e be a trivial one-dimensional projection. Then i=1 pj for all N, because the Euler class of the projection on the right-hand side is Put xi = hpii, v = hei, and put u = P∞ i=1 xi. Then v (cid:2) u, v ≤ 2xi, and u + u = ∞·v = j=i xi = ∞·v = ∞·u, and so u is (ω, 2)-decomposable; but u is not properly non-trivial. e - LN ∞·u. Hence 2P∞ infinite. We now look more closely at the properties of (ω, n)-decomposable elements. Proposition 9.5. Let (ai)∞ a C ∗-algebra A such that P∞ that Pj≥i nhaii = ∞ for all i. Then A ⊗ B is stable for every σ-unital C ∗-algebra B with i=0 be a sequence of mutually orthogonal positive elements in i=0 ai converges to a strictly positive element in A. Assume rank(B) ≥ n. 40 Proof. Set P∞ i=0 ai = a ∈ A and let b be a strictly positive element in B. Notice that a ⊗ b is a strictly positive element of A ⊗ B. In order to prove stability of A ⊗ B we will use the stability criterion obtained in [HR98]: A ⊗ B is stable if for every ε > 0 there exists a positive element c in A which is orthogonal to (a ⊗ b − ε)+ and satisfies h(a ⊗ b − ε)+i ≤ hci. Arguing as in the proof of (iv) ⇒ (i) in Lemma 9.2 we may assume that nhaii = ∞ for all i. By Theorem 5.3 (iii), rank(B) ≥ n is equivalent to weak (n, ω)-divisibility for hbi. Thus there exist a sequence (xi)∞ i=1 xi = ∞. We can form a new sequence (x′ i=1 in which each xi appears repeated infinitely often. i≥j xi = ∞ for all j. Find i=1 in Cu(B) such that nxi ≤ hbi for all i and P∞ In this way we may assume without loss of generality that P∞ positive elements bi in B such that xi = hbii and kbik ≤ 2−i. Then i)∞ (a ⊗ b − ε)+ = ∞ Xi=1 (ai ⊗ b − ε)+ = N (ai ⊗ b − ε)+, Xi=1 for some integer N ≥ 1. Set c = Pi>N ai ⊗ b. Then c is orthogonal to (a ⊗ b − ε)+. Also, hai ⊗ bi = haii ⊗ hbi ≥ n haii ⊗ hbii = ∞·hai ⊗ hbii for each i. Hence hci = Xi>N hai ⊗ bi = Xi>N ∞·hai ⊗ hbii = ∞. Thus, h(a ⊗ b − ε)+i ≤ ∞ = hci. This shows that A ⊗ B is stable. The proposition above can be applied to the C ∗-algebra A = P (C(X) ⊗ K)P arising from i=1 pi ∈ M(C(X) ⊗ K). The C ∗-algebra A is not stable (because e /∈ A while e ∈ M2(A)), but A ⊗ B is stable for every C ∗-algebra B that does not have a character by Proposition 9.5 and Example 9.4. Example 9.4 with P = L∞ The example obtained in [Rør97] of a simple C ∗-algebra A of stable rank 1 such that Mn(A) is stable, while Mn−1(A) is not stable, likewise satisfies the hypotheses of Proposi- tion 9.5. In fact, to the authors knowledge, every example of a C ∗-algebra that tensored with Mn(C) becomes stable also has the stronger property of becoming stable after being tensored with any C ∗-algebra that has no representations of dimension less than n. This raises the following question: Question 9.6. Is there a C ∗-algebra A such that M2(A) is stable but A ⊗ B is not stable for some C ∗-algebra B without characters? Proposition 9.7. The following statements are equivalent for every C ∗-algebra A with a strictly positive element a. (i) hai is (ω, n)-decomposable. (ii) A contains a full hereditary subalgebra B such that B ⊗C is stable for any C ∗-algebra C such that rank(C) ≥ n. 41 (iii) A contains a full hereditary subalgebra B such that Mn(B) is stable. Proof. (i) ⇒ (ii). Let (xi)∞ i=1 xi ≤ hai and nxi = ∞. Let b ∈ A ⊗ K be strictly positive and let bi ∈ A ⊗ K be mutually orthogonal elements such that hbii = xi. We can find mutually orthogonal positive elements ai in A such that haii ≤ hbii, nhaii ≥ i=1 ai is convergent. It then follows from Proposition 9.5 that i=1 be such that P∞ h(b − 1/i)+i, and such that P∞ (ii) holds when B is the hereditary sub-C ∗-algebra generated by P∞ (ii) ⇒ (iii) is clear. (iii) ⇒ (i). Since A is σ-unital, and B is stably isomorphic to A, B is σ-unital too. Let i=1 ai. b be a strictly positive element in B. Use [OPR, Lemma 5.3] to find a sequence (bk) of pairwise orthogonal positive elements in B such that h(b − 1/k)+i ≤ nh(b − 1/k)+i ≤ nhbki for all k. Then condition (iii) of Lemma 9.2 is satisfied with u = hbi, xk = hbki, and yk = h(b − 1/k)+i, whence hbi is (ω, n)-decomposable. Finally, by the fact that hbi ≤ hai ≤ ∞·hbi, it follows by the equivalence of (i) and (ii) in Lemma 9.2 that hai is (ω, n)-decomposable. Definition 9.8. Let n ∈ N and u ∈ Cu(A). We call u weakly (ω, n)-divisible if for every u′ ≪ u there exist xi ∈ Cu(A), i = 1, 2, . . . , n, such that ∞·xi ≤ u and u′ ≤ Pn Observe that if u is weakly (ω, n)-divisible, then u is weakly (m, n)-divisible for all m ∈ N, whence nu is properly infinite by Proposition 3.4. i=1 xi. We will next give an example of a Cuntz semigroup element that is weakly (ω, 2)- divisible but not properly infinite. This example needs some preparatory results. Let us first recall an example given by Dixmier and Doaudy in [DD63]. Example 9.9 (Dixmier -- Douady, [DD63, §17]). Let B∞ denote the closed unit ball of l2(N) endowed with the weak topology. Let l2(B∞) denote the C(B∞)-Hilbert module of continuous maps from B∞ to l2(N). We will construct a countably generated C(B∞)- Hilbert module D such that l2(B∞) ֒→ D ֒→ l2(B∞) but D ≇ l2(B∞). Let x : B∞ → l2(N) ⊕ Ce be given by x(z) = z +p1 − kzk2 · e, for z ∈ B∞. Consider the C(B∞)-module D0 of functions from B∞ to l2(N) ⊕ Ce that have the form y + xλ, with y ∈ l2(B∞) and λ ∈ C(B∞). The module D0 is a pre-Hilbert C-module over C(B∞) when endowed with the pointwise inner product. Indeed, if y1 + xλ1 and y2 + xλ2 are vectors in D0 then hy1 + xλ1, y2 + xλ2i = hy1, y2i + hy1, ziλ2 + λ1hz, yi + λ1λ2 ∈ C(B∞). Let D denote the completion of D0 with respect to the norm induced by its C(B∞)-valued inner product. Observe that l2(B∞) ֒→ D0 ⊆ D. Since D is countably generated, we also have that D ֒→ l2(B∞) by Kasparov's stabilization theorem. Let us see that D ≇ l2(B∞). Consider E ⊆ D, the orthogonal complement of {x}. Then E = E0, where E0 = {y + xλ ∈ D hy(z), zi + λ(z) = 0 for all z ∈ B∞}. (9.1) 42 It was implicitly shown by Dixmier and Douady, and explicitly pointed out by Blanchard and Kirchberg ([BK04a, Proposition 3.6]), that for any v ∈ E there exists z ∈ B∞ such that hv, vi(z) = 0. That is, every section of E vanishes at some point (we will reprove this fact in Proposition 9.10 below). Notice that D = E + x · C(B∞) ∼= E ⊕ C(B∞). It can be deduced from this that D ≇ l2(B∞) (see [DD63, Proposition 19]). Let B3 denote the unit ball in R3. Let f ∈ M2(B3)+ be defined as f (x, y, z) = 1 2 (cid:18) 1 + z x − iy 1 − z (cid:19) . x + iy (The function f is a homeomorphism from B3 to the set of positive elements of M2(C) with trace 1. On the boundary 2-sphere of B3 it agrees with the tautological rank 1 projection.) Consider the C(B3)-module associated to f : F := f (cid:18)C(B3) C(B3)(cid:19). (9.2) Proposition 9.10. Let B∞ and B3 be as before. Let X = Qi∈I Xi, where each Xi is either B∞ or B3 and the index set I is non-empty. For each i, let Hi be the pull-back along the projection map πi : X → Xi of either the module E defined in Example 9.9 or the module F defined in (9.2). Finally, let H be the C(X)-module defined by H = Li∈I Hi. Then C(X) does not embed in H as a C(X)-module (i.e., for every v ∈ H there exists z ∈ X such that hv, vi(z) = 0). Notice that if every Xi agrees with B3, the above proposition can be proven using standard methods in algebraic topology (e.g., characteristic classes). Indeed, it suffices to restrict to the boundary 2-sphere of each Xi and use that on that set F is the tautological rank 1 projective module. It is the inclusion of the spaces B∞ in the definition of X that forces us to use a different route in the proof. Proof. Let v ∈ H, and write v = Pi∈I vi, with vi ∈ Hi. In order to show hv, vi(z) = 0 for some z ∈ X, it suffices to prove this for v belonging to a dense submodule of H. For suppose that (v(n)) is a sequence in H such that v(n) → v and hv(n), v(n)i(zn) = 0 for some zn ∈ X. Then by the compactness of X there exists a subsequence (znk) such that znk → z ∈ X, and so hv, vi(z) = 0. Thus, we may assume that the index set I is finite. Furthermore, for the indices i such that Hi = π∗ i, where H ′ i ⊆ Hi is the pull back along πi of the dense submodule E0 defined in (9.1). i (E), we may assume that vi ∈ H ′ In the sequel, we assume that I = {1, 2, . . . , n}, Xi = B∞ for i = 1, 2, . . . , n1, and Xi = B3 for i = n1 + 1, . . . , n, where n1 ≤ n. We will argue by contradiction that hv, vi(z) = 0 for some z ∈ X. Suppose that hv, vi is invertible, and assume without loss of generality that hv, vi = 1. Observe that, for each 43 i ≤ n1, vi is a function from X into the unit ball of l2(N) ⊕ C, while for n1 < i ≤ n the entry vi is a function from X into the unit ball of C ⊕ C (let us denote it by B4). Let h0 : l2(N) ⊕ C → l2(N) denote the projection onto the first direct summand and let h1 : B4 → B3 denote the Hopf fibration (extended to the unit ball): h1(z0, z1) := (2z0z1, z02 − z12). Let λ : [0, 1] → [0, 1] be such that λ(0) = 0, λ(t) = 1 for t ∈ [ 1 Define h0, h1 : B4 → B3 by n, 1], and λ is linear in [0, 1 n ]. h0(w) = h0(cid:16) λ(w) w w(cid:17), h1(w) = −h1(cid:16)λ(w) w w(cid:17). Consider the continuous map Φ : X → X given by the vector of functions Φ := (h0 ◦ v1, h0 ◦ v2, . . . , h0 ◦ vn1, h1 ◦ vn1+1, . . . , h1 ◦ vn). Since X is a compact convex subset of the vector space (l2(N))n1 × (R3)n−n1, the map Φ has a fixed point by the Schauder fixed point theorem. Let z := (zi)n i=1 ∈ X be a fixed point of Φ. Since kv(z)k = 1, we must have kvi(z)k ≥ 1 n for at least one index i. Notice that both h0 and h1 map all vectors of norm at least 1/n into the unit sphere of either B∞ or B3. It follows that the fixed point z satisfies kzik = 1 and zj = 0 for all j 6= i. There are two cases to consider: i ≤ n1 and i > n1. Suppose that i ≤ n1. The general form of vi ∈ H ′ kzik = 1, we have i is f + (zi +p1 − kzik2e)α, for some f : X → l2(N) and α ∈ C(X). Since vi(z) = f (z) + α(z)zi = zi. But hf (z), zii + α(z) = 0. This contradicts that kzik = 1. Suppose that i > n1. Since zi 7→ vi(· · · , zi, · · · ), with zi ∈ S2, is a section of the tautological bundle on S2, we have h1 ◦ vi(z) = zi whenever zi ∈ S2. It follows that h1 ◦ vi(z) = −zi. But h1 ◦ vi(z) = zi, by the fixed point property of z. This again contradicts that kzik = 1. We are now prepared to give examples of weakly (ω, 2)-divisible elements which are not properly infinite. Example 9.11. Let X = B∞ × B3 and consider the Hilbert module H = π∗ 2(F ), described in the statement of the previous proposition. We have shown that [C(X)] (cid:2) [H]. In particular, [H] is not properly infinite (since it is full). Let us show that [H] is weakly (ω, 2)-divisible. Consider the open sets U := B∞ × B+ 3 and B− 3 are (open) upper and lower hemispheres of B3 that together cover B3. We claim that l2(U) ֒→ HC0(U) and l2(V ) ֒→ HC0(V ). Indeed, 3 and V := B∞ × B− 3 , where B+ 1(E) ⊕ π∗ HC0(U) = π∗ = π∗ 1(E)C0(U) ⊕ π∗ 1(E)C0(U) ⊕ π∗ 2(F )C0(U) 2(F C0(B+ 3 )). 44 But F C0(B+ 3 ) ∼= C0(B+ 3 ) ⊕ C0(B+ 3 \S2). Therefore, HC0(U) = π∗ = π∗ 1(E)C0(U) ⊕ C0(U) ⊕ F ′ 1(E ⊕ C(B∞))C0(U) ⊕ F ′. But l2(B∞) ֒→ D = E ⊕ C(B∞). Thus, l2(U) ֒→ HC0(U). Symmetrically, we have that l2(V ) ֒→ HC0(V ). It follows that [HC0(U)] and [HC0(V )] are properly infinite, and [H] ≤ [HC0(U)] + [HC0(V )]. Thus, [H] is weakly (ω, 2)-divisible. Remark 9.12. The previous example answers a question posed in [KR00, Question 3.10]: If a and b are properly infinite positive elements, is a + b properly infinite? In the language of Hilbert modules, this question asks whether H is properly infinite if H = H1 + H2, and H1, H2 ⊆ H are properly infinite submodules of H. We obtain a counterexample taking H as in the previous example, H1 = HC0(U) and H2 = HC0(V ). Example 9.13. In this example we answer (in the negative) the following question, posed in [KR00, Question 3.4]: if [H] is properly infinite, is the unit of B(H) a properly infinite projection? Let X = B∞ × (B3)∞ and consider the Hilbert C(X)-module H = C(X) ⊕ π∗ 1(E) ⊕ ∞ Mi=2 π∗ i (F ). The module C(X) ⊕ π∗ Since l2(C(B∞)) embeds in D, l2(C(X)) embeds in C(X) ⊕ π∗ 1(E) is the pull back along π1 of the Dixmier-Douady module D. 1(E). Thus, [H] is properly infinite. Also, the direct sum of the module L∞ i (F ) with itself gives l2(C(X)) (because F ⊕ F contains C(B3) as a direct summand). Therefore, H ⊕ H ∼= l2(C(X)). However, H is not isomorphic to l2(C(X)), because every section of π∗ i (F ) vanishes, and so adding the trivial rank 1 module to it cannot yield the trivial Hilbert module l2(C(X)) (see the proof of D ≇ l2(C(B∞)) in [DD63, Proposition 19]). It follows that H ⊕ H is not a direct summand of H, i.e., the unit of B(H) is not properly infinite. 1(E) ⊕L∞ i=2 π∗ i=2 π∗ References [BK04a] E. Blanchard and E. Kirchberg, Global Glimm halving for C ∗-bundles, J. Operator Theory 52 (2004), no. 2, 385 -- 420. [BK04b] , Non-simple purely infinite C ∗-algebras: the Hausdorff case, J. Funct. Anal. 207 (2004), no. 2, 461 -- 513. [BPT08] N. Brown, F. Perera, and A. S. Toms, The Cuntz semigroup, the Elliott conjecture, and dimension functions on C ∗-algebras, J. Reine Angew. Math. 621 (2008), 191 -- 211. [CEI08] K. T. Coward, G. A. Elliott, and C. Ivanescu, The Cuntz semigroup as an invariant for C ∗-algebras, J. Reine Angew. Math. 623 (2008), 161 -- 193. 45 [DD63] J. Dixmier and A. Douady, Champs continus d'espaces hilbertiens et de C ∗-alg`ebres, Bull. Soc. Math. France 91 (1963), 227 -- 284. [DHTW09] M. Dadarlat, I. Hirshberg, A. Toms, and W. Winter, The Jiang-Su Algebra does not always embed, Math. Res. Lett. 16 (2009), no. 1, 23 -- 26. [DT09] M. Dadarlat and A. Toms, Z-stability and infinite tensor powers of C ∗-algebras, Adv. Math. 220 (2009), no. 2, 341 -- 366. [Dup76] M. J. Dupr´e, Classifying Hilbert bundles. II, J. Funct. Anal. 22 (1976), no. 3, 295 -- 322. [ER06] [ERS] [HR98] G. A. Elliott and M. Rørdam, Perturbation of Hausdorff moment sequences, and an application to the theory of C ∗-algebras of real rank zero, Operator Algebras: The Abel Symposium 2004, Abel Symp., vol. 1, Springer, Berlin, 2006, pp. 97 -- 115. G. A. Elliott, L. Robert, and L. Santiago, The cone of lower semicontinuous traces on a C-algebra, American J. Math., to appear. J. Hjelmborg and M. Rørdam, On stability of C ∗-algebras, J. Funct. Anal. 155 (1998), no. 1, 153 -- 170. [HRW07] I. Hirshberg, M. Rørdam, and W. Winter, C0(X)-algebras, stability and strongly self-absorbing C ∗-algebras, Math. Ann. 339 (2007), no. 3, 695 -- 732. [Hus94] [Kir06] D. Husemoller, Fibre Bundles, 3rd. ed., Graduate Texts in Mathematics, no. 20, Springer Verlag, New York, 1966, 1994. E. Kirchberg, Central sequences in C ∗-algebras and strongly purely infinite C ∗- algebras, Operator Algebras (Berlin) (S. Neshveyev C. Skau O. Bratteli, ed.), Abel Symp., vol. 1, Springer, 2006, pp. 175 -- 232. [KOS03] A. Kishimoto, N. Ozawa, and S. Sakai, Homogeneity of the pure state space of a separable C ∗-algebra, Canad. Math. Bull. 46 (2003), no. 3, 365 -- 372. [KR00] E. Kirchberg and M. Rørdam, Non-simple purely infinite C ∗-algebras, American J. Math. 122 (2000), 637 -- 666. [KR02] , Infinite non-simple C ∗-algebras: absorbing the Cuntz algebra O ∞, Advances in Math. 167 (2002), no. 2, 195 -- 264. [MS74] [OPR] [PR04] [PT07] J.W. Milnor and J.D. Stasheff, Characteristic classes, no. 76, Princeton Univ Pr, 1974. E. Ortega, F. Perera, and M. Rørdam, The Corona Factorization property, Stability, and the Cuntz semigroup of a C*-algebra., Int. Math. Res. Not. IMRN. To appear. F. Perera and M. Rørdam, AF-embeddings into C ∗-algebras of real rank zero, J. Funct. Anal. 217 (2004), no. 1, 142 -- 170. F. Perera and A. S. Toms, Recasting the Elliott conjecture, Math. Ann. 338 (2007), no. 3, 669 -- 702. 46 [Rob11] L. Robert, The cone of functionals on the Cuntz semigroup, 2011. [Rør97] M. Rørdam, Stability of C ∗-algebras is not a stable property, Documenta Math. 2 (1997), 375 -- 386. [Rør03] , A simple C ∗-algebra with a finite and an infinite projection, Acta Math. 191 (2003), 109 -- 142. [Rør04] , The stable and the real rank of Z-absorbing C ∗-algebras, Internatinal J. Math. 15 (2004), no. 10, 1065 -- 1084. [RW10] M. Rørdam and W. Winter, The Jiang-Su algebra revisited, J. Reine Angew. Math 642 (2010), 129 -- 155. [Vil98] [Win] J. Villadsen, Simple C ∗-algebras with perforation, J. Funct. Anal. 154 (1998), no. 1, 110 -- 116. W. Winter, Nuclear dimension and Z-stability of pure C ∗-algebras, Invent. Math., to appear. Department of Mathematical Sciences, University of Copenhagen, Univer- sitetsparken 5, DK-2100 Copenhagen Ø, Denmark E-mail address: [email protected] Department of Mathematical Sciences, University of Copenhagen, Univer- sitetsparken 5, DK-2100 Copenhagen Ø, Denmark E-mail address: [email protected] 47
1102.4494	1	1102	2011-02-22T13:13:03	Noncommutative maximal ergodic inequality for non-tracial L1-spaces	[ "math.OA" ]	We extend the noncommutative L1-maximal ergodic inequality for semifinite von Neumann algebras established by Yeadon in 1977 to the framework of noncommutative L1-spaces associated with sigma-finite von Neumann algebras. Since the semifnite case of this result is one of the two essential parts in the proof of noncommutative maximal ergodic inequality for tracial Lp-spaces (1<p<infinity) by Junge-Xu in 2007, we hope our result will be helpful to establish a complete noncommutative maximal ergodic inequality for non-tracial Lp-spaces in the future.	math.OA	math	Noncommutative maximal ergodic inequality for non-tracial L1-spaces Qin Zhang ∗ Graduate School of Mathematical Sciences, The University of Tokyo Abstract We extend the noncommutative L1-maximal ergodic inequality for semifinite von Neumann algebras established by Yeadon in 1977 to the framework of noncommutative L1-spaces associated with σ-finite von Neumann algebras. Since the semifinite case of this result is one of the two essential parts in the proof of noncommutative maximal ergodic inequality for tracial Lp-spaces (1 < p < ∞) by Junge-Xu in 2007, we hope our result will be helpful to establish a complete noncommutative maximal ergodic inequality for non-tracial Lp-spaces in the future. 1 Introduction Theory of von Neumann algebras is regarded as the noncommutative measure and inte- gration theory (Chapter IX, [28]), so it is natural to consider extensions of classical ergodic theorems for spaces of measurable functions to the framework of noncommutative spaces associated with von Neumann algebras. Such an extension topic appeals to many math- ematicians and they had interesting results even from 1970's (for example, the pioneering works of [17] and [30]). Since then many classical mean ergodic theorems and ergodic theo- rems of other types were successfully transformed to the noncommutative context which is a semifinite von Neumann algebra or a non-commutative Lp-space associated with a semifinite von Neumann algebra, and some authors even considered a general von Neumann algebra or a non-tracial Lp-space associated with it. Among these noncommutative ergodic theorems, however, the problem of finding a noncommutative analogue of the famous Dunford-Schwartz maximal ergodic inequality ([7]) was left open until the appearance of Junge-Xu's prominent ∗Address: 3-8-1 Komaba, Meguro, Tokyo, 153-8914, Japan. The author is supported by the Japanese Government Scholarship (No.052111). Email Address: [email protected] 1 paper in 2007. The main obstacle in this problem is that it is difficult to define the supremum of a sequence of operators even in the finite-dimensional Hilbert space cases, although it is straightforward to take the supremum of a sequence of measurable functions. For this rea- son, many powerful techniques in classical ergodic theory involving maximal functions seem no longer available in the research of noncommutative ergodic results. This difficulty was overcome in Junge-Xu's work ([15]) by using the noncommutative vector valued Lp-space theory developed by Pisier and Junge (see [21] for the case of hyperfinite von Neumann alge- bras and [14] for that of general ones). Junge-Xu established noncommutative version of the Dunford-Schwartz maximal ergodic inequality first for non-commutaive Lp-spaces associated with a semifinite von Neumann algebra, or in other words, tracial Lp-spaces (Theorems 4.1, [15]). In order to state their result we need some notations as follows. Let M be a semifinite von Neumann algebra equipped with a normal semifinite faithful trace τ , and let Lp(M) be the associated noncommutative Lp-space (see, for example, [19] for a detailed definition). The symbol T : M → M denotes a linear map which satisfies the following conditions. (J1) T is a contraction on M: T x ≤ x for all x ∈ M, where · means the ∞-norm, or in other words, the usual operator norm. (J2) T is positive: T x ≥ 0 if x ≥ 0. (J3) τ ◦ T ≤ τ : τ (T (x)) ≤ τ (x) for all x ∈ L1(M) ∩ M+. Theorem 1.1. (Theorem 4.1 in [15]). Let T be a linear map on a semifinite von Neumann algebra M with (J1) -- (J3), then T extends naturally to a contraction on Lp(M) for 1 ≤ p < ∞. Put Sn ≡ Sn(T ) = 1 n + 1 then for every p, 1 < p < ∞, we have T k, nXk=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)sup n +Sn(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)p ≤ cp xp , for all x ∈ Lp(M), for some positive constant cp depending only on p. Note that sup+ n Sn(x)p is the notation for {Sn(x)}n∈NLp(M ; l∞) used in [15], where Lp(M; l∞) is the Pisier-Junge's noncommutative vector valued Lp-space defined by the space of all sequences x = {xn}n∈N in Lp(M) (1 ≤ p ≤ ∞) such that each sequence admits a 2 factorization of the following form: there are a, b ∈ L2p(M) and y = {yn}n∈N ⊂ L∞(M) = M such that xn = aynb for all n ∈ N and the norm is defined by xLp(M ; l∞) ≡ inf{a2p sup n∈N yn b2p}, where the infimum runs over all such factorizations as above. For details of such spaces, we refer the reader to [21] and [14]. But considering only semifinite von Neumann algebras is not enough sometimes. It was stated in [28] that most of factors arising from physics are of type III, which are of course not semifinite. Another fact is that it was shown by Pisier ([22]) that OH cannot completely embed in a semifinite L1(M). More and more recent works concern the type III case or need results on noncommutative Lp-spaces associated with a not necessarily semifinite von Neumann algebra. For a general von Neumann algebra, there are several equivalent constructions of non- commutative Lp-spaces associated with it, and the important ones include [2, 6, 9, 11, 13, 16, 29] and they are all based on the Tomita-Takesaki theory. Since a general von Neumann algebra does not necessarily admit a normal semifinite faithful trace, any of these (equivalent) constructions are called non-tracial Lp-spaces. The first non-tracial Lp-spaces are Haagerup's ones, and just for Haagerup's Lp-spaces, Junge-Xu established the non-commutative Dunford-Schwartz maximal ergodic inequality as a non-tracial extension of Theorem 1.1 above (see Theorem 7.4, [15]). Their method is to use an early result of Haagerup named reduction method to approximate Haagerup's Lp-spaces by simifinite ones. We state this theorem here. Theorem 1.2. (Theorem 7.4 in [15]). Suppose M is a (σ-finite) von Neumann algebra with a normal faithful state ϕ, let T be a linear map on M satisfying the following properties (H1) -- (H4). (H1) T is a contraction on M: T x ≤ x for all x ∈ M. (H2) T is completely positive. (H3) ϕ(T (x)) ≤ ϕ(x) for all x ∈ M+. (H4) T ◦ σϕ t = σϕ t ◦ T for all t ∈ R. 3 Then T extends naturally to a contraction on Haagerup's Lp(M) for 1 ≤ p < ∞. Put Sn ≡ Sn(T ) = 1 n + 1 nXk=0 T k. Then for every p, 1 < p < ∞, we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)sup n for some positive constant cp depending only on p. ≤ cp xp , for any x ∈ Lp(M), +Sn(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)p This theorem was also established in [10] (Theorem 7.9). We may compare the statement of this theorem with that of Theorem 1.1. It is clear that the assumption (J2) is reinforced to the assumption (H2), and (H4) is new, the others are natural extensions. Since the modular automorphism group στ t trivial, we do not think the condition (H4) is unnatural any more when we consider the induced by a trace τ is non-tracial cases. However, we feel that the "complete positivity" assumption of (H2) seems a little bit stronger. Although it is remarked in [15], page 425, that the complete positivity is unlikely to be really necessary, they did not give a proof to release this assumption, and they still used complete positivity assumption in [10]. We would like to remark that Haagerup's reduction method used in [10] is really pow- erful to extend the noncommutative martingale inequalities from the tracial cases to the non-tracial ones without any change in the statement form for almost all those results. This is because a conditional expectation in the framework of von Neumann algebras is certainly completely positive ([1] and [27]). In noncommutative ergodic theorems, "complete pos- itivity" seems a little restrictive, moreover, many important positive but not completely positive state-preserving transforms describing open quantum evolution are now considered by mathematical physicists (see, for example, [24, 25, 12, 5, 26]). So we do not think Theorem 1.2 is a complete answer for the problem of non-tracial extension of Theorem 1.1. Recalling the proof of Theorem 1.1 in [15], we know that this proof consists of two essential parts, one is the noncommutative maximal ergodic inequality for semifinite L1(M) established by Yeadon, and the other one is the noncommutative semifinite Marcinkiewicz interpolation theorem in [15]. A natural and possible way of thought for the problem of a complete non-tracial extension of Theorem 1.1 is to get non-tracial extensions of these two parts. The main result of our paper is just to provide the non-tracial extension of maximal ergodic inequality for the non-tracial L1(M). We have been unable to get a noncommutative Marcinkiewicz interpolation theorem for the non-tracial Lp-spaces so far 4 and we will explain the reason in the next section. We hope the result in this paper will be helpful to establish a complete noncommutative maximal ergodic inequality for the non- tracial Lp-spaces in the future. It is well-known that all the von Neumann algebras encountered in quantum statistical mechanics and quantum field theory are σ-finite ([4], p.84), probably for this reason, only the noncommutative Lp-spaces associated with a σ-finite von Neumann algebra are consid- ered in Theorem 1.2 instead of general non-tracial situation and we think this restriction to the σ-finite case is of full meaning. Our framework is the Hilsum spatial Lp-spaces associ- ated with a σ-finite von Neumann algebra which are certainly isometrically isomorphic to Haagerup's ones associated with a σ-finite von Neumann algebra. On the detailed construc- tion of the Hilsum spatial Lp-spaces and the close relation between those spaces and complex interpolation theory, we will give a short description on them in the beginning of the next section. Let M be a σ-finite von Neumann algebra represented on a Hilbert space H, hence we may assume M admits a normal faithful state ϕ, and ψ is a normal faithful state on the commutant M ′ of M, let Lp(M; ψ) denote the Hilsum spatial Lp-spaces with respect to ψ. Our main result is the following theorem. Theorem 1.3. (Theorem 2.1 in the next section). Assume M is a σ-finite von Neumann algebra which admits a normal faithful state ϕ and T is a linear map on M satisfying the following conditions. (1) T is a contraction of M, i.e., T x ≤ x for all x ∈ M. (2) T is positive, i.e., T y ≥ 0 if y ∈ M+. (3) ϕ(T (y)) ≤ ϕ(y), for all y ∈ M+, Then T extends to a positive linear contraction on L1(M; ψ). For any a ∈ L1 +(M; ψ) and any λ > 0, and any n ∈ N, there exists a projection en ∈ M such that enSr(a)en ≤ λenden for all r ∈ {0, 1, ..., n}, and where Sr(a) = 1 Furthermore, for any a ∈ L1 r+1Pr ϕ(1 − en) ≤ 2 λZ adψ, k=0 T k(a), d = dϕ dψ is the spatial derivative and 1 is the identity of M. +(M; ψ), there exists a projection e ∈ M such that for any r ∈ N, Z eSr(a)edψ ≤ 4λ, and ϕ(1 − e) ≤ 2 λZ adψ. 5 We recall that in the semifinite case, the L1-norm of the tracial L1-space L1(M) is defined by ·1 = τ ( · ), and in the spatial L1-space L1(M; ψ), the L1-norm is defined by ·1 = (R ·dψ), and we know that when M is semifinite, L1(M; ψ) is equivalent to the tracial L1(M), so this is the reason we useR ·dψ in the statement of our theorem. In the semifinite case, Yeadon's result ([30]) provided a bound on eSr(a)e, and this is because in semifinite case, the elements in Lp-spaces are unbounded (also including bounded) operators affiliated with the von Neumann algebra, hence we may cut such unbounded operators by using some projection in the von Neumann algebra. However, for the non-tracial case, elements in the Hilsum (or other equivalent) Lp-spaces are never affiliated with the von Neumann algebra M, so eSr(a)e cannot be majorized by any bounded operator, and the result in our theorem is the best one we can hope in the non-tracial case. Such a result also has its own value. In the commutative case we know that the following Chebyshev type inequality is quite important in measure theory and probability theory, µ({x : sup n∈N sn(f )(x) > λ}) ≤ c λ f 1 , and our result is a non-commutative extension of this inequality for σ-finite von Neumann algebras. 2 Main Part Now we introduce the definition of the Hilsum spatial Lp-spaces. We assume M is a general von Neumann algebra (because the Hilsum Lp-spaces are defined for general von Neumann algebras) which admits a normal semifinite faithful weight ϕ and furthermore, M is represented on a Hilbert space H and we have a normal semifinite faithful weight ψ on the commutant M ′ of M. A vector ξ ∈ H is said to be ψ-bounded if there exists a positive constant c such that yξ2 ≤ cψ(y ∗y) for any y ∈ nψ, where nψ = {y ∈ M ′ ψ(y ∗y) < ∞}. We let D(H, ψ) be the subspace of H consisting of all ψ-bounded vectors. Then for any ξ ∈ D(H, ψ), Rψ(ξ) is the unique bounded operator from Hψ (the GNS representation Hilbert space induced by ψ) to H such that Rψ(ξ)ηψ(y) = yξ, where ηψ is the canonical injection of nψ into Hψ, and θψ(ξ, ξ) = Rψ(ξ)Rψ(ξ)∗ ∈ M. The map ξ → φ(θψ(ξ, ξ)) defines a lower semicontinuous positive form on D(H, ψ), where φ is any normal semifinite weight on M. Then the positive self-adjoint operator associated with this form is called the spatial derivative dφ dψ defined by Connes ([6]). 6 The Hilsum spatial Lp-space Lp(M; ψ) is defined as (1 ≤ p < ∞) Lp(M; ψ) = = R apdψ < ∞ The Lp-norm is defined by ·p = (R · pdψ) operator norm. a is a closed densely defined operator on H with polar decomposition a = ua such that u ∈ M and ap = dφ dψ for some φ ∈ M + ∗ a is a closed densely defined operator on H and (− 1 p )-homogeneous with respect to ψ such that   1 p . If p = ∞, L∞(M; ψ) = M with the usual For the definition of homogeneity with respect to ψ and the detailed properties of dφ dψ and Lp(M; ψ), we refer the reader to [6, 11, 29]. Concerning the Hilsum spatial Lp-spaces, Terp's paper ([29]) deeply revealed the close relation between the complex interpolation theory on (M, M∗) and Lp(M; ψ). The subspace L of M consists of x ∈ M for which there exists a ϕx ∈ M∗ such that for any y, z ∈ nϕ : (z∗y, ϕx) = hJπϕ(x)∗Jηϕ(y), ηϕ(z)i, where J is the associated modular conjugation in Hϕ, (·, ·) denotes the duality between M and M∗, and h , i is the scalar inner product on Hϕ . For x ∈ L, we put xL = max{x , ϕx}, where the norm for ϕx means the functional norm. The normed space (L, ·L) is a Banach space and it can be embedded naturally into M and M∗ by x 7→ x : L → M and x 7→ ϕx : L → M∗. By transposition of the above two embeddings we have the injections M → L∗ and M∗ → L∗ given by for all y ∈ M and (x, y)(L,L∗) = (y, ϕx)(M,M∗), x ∈ L (x, φ)(L,L∗) = (x, φ)(M,M∗), x ∈ L for all φ ∈ M∗, where L∗ means the dual of (L, ·L). The following diagram commutes, DDDDDDDD ={{{{{{{{ L∗ M > AAAAAAAA L M∗ 7 ! ! > = and L = M ∩ M∗ when M and M∗ are considered as subspaces of L∗ (see Section 1, [29]), and L is σ-weakly dense in M, the embedding of L in M∗ is weakly and norm dense in M∗ (Corollary 5, [29]). Hence (M, M∗) is turned into a compatible pair of Banach spaces in the complex inter- polation sense (Section 2.3, [3]), and Terp proved that for 1 < p < ∞, the Hilsum spatial Lp-spaces Lp(M; ψ) are just the complex interpolation spaces of M∗ and M. Accurately speaking, Lp(M; ψ) = C 1 (M, M∗) (Theorem 36 in [29]). p Let T be a linear map on M satisfying the following conditions. (1) T is a contraction of M, i.e., T x ≤ x for all x ∈ M. (2) T is positive, i.e., T y ≥ 0 if y ∈ M+. (3) ϕ(T (y)) ≤ ϕ(y) for all y ∈ L+. From now on, we will concentrate on the σ-finite cases. Let M be a σ-finite von Neumann algebra acting standardly on the Hilbert space H, M admits a normal faithful state ϕ, ψ is a normal faithful state on M ′, and T is a linear transform satisfying the above conditions (1) − (3). We note that for any x ∈ M, we have x ∈ mϕ = span{x ∈ M+ ϕ(x) < ∞} as ϕ is a state, and we know from Note (2), p.329 in [29] that mϕ ⊆ L, combined with the fact that L is defined to be a linear subspace of M, we have mϕ = M = L in the σ-finite cases. That is to say, condition (3) above can be replaced by the following one when M is σ-finite. (3) ϕ(T (y)) ≤ ϕ(y) for all y ∈ M+. Moreover, the embedding of M = L into L1(M; ψ) is given by the map M → d 2 , where d = dϕ dψ is the spatial derivative of ϕ with respect to ψ (see Section 2.3 and Theorem 27 in [29]). Such an embedding is equivalent to the embedding M → M η ⊆ M∗ with η = 1 2 (the symmetric case) in Definition 7.1 of [16] and Lp(M; ψ) is equivalent to Kosaki's Lp-spaces 2 (Definition 7.2, [16]). Moreover, since the Hilbert space is standard, C 1 the Hilsum space Lp(M; ψ) is now as the same as the Araki-Masuda Lp-space in [2]. (M η, M∗) with η = 1 p 1 1 2 M d Then for a general p, 1 ≤ p < ∞, we may define the following map Tp as for any x ∈ M. Tp : d 1 2p M d 1 2p → d 1 2p M d 1 2p 1 d 2p xd 1 2p 7→ d 1 2p T (x)d 1 2p , 8 We claim that the map T1 defined above extends naturally to a positive contraction of L1(M; ψ) → L1(M; ψ), and it will still be denoted by T1. To show this claim, we need the Lemma 5.2 in [10]. Although this lemma is stated and proved for Haagerup's L1-spaces, it is still valid in the framework of the Hilsum spatial L1-spaces through isometric isomorphism. We state it here for L1(M; ψ). Let x ∈ M and x is self-adjoint, then = inf{ϕ(a) + ϕ(b) x = a − b, a, b ∈ M+}. 1 2 xd (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 The proof of this result for L1(M; ψ) is essentially the same as that of Lemma 5.2 in [10] when replacing D there by the spatial derivative d. Then we will use this result and follow the method of Lemma 5.3 in [10] to show our claim. Let y ∈ M+, then d T1(d 2 ≥ 0, hence T1 is also positive. By condition (3), 2 ≥ 0, so 2 T (y)d 2 ) = d 2 yd 2 yd 1 1 1 1 1 1 Now assume x ∈ M and x is self-adjoint, for any ε > 0, there exist a, b ∈ M+ such that x = a − b and It follows that 1 1 1 1 2 ad 2 T (y)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 =Z d (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d =ϕ(T (y)) ≤ ϕ(y) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2 )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 2 T (a)d 2 xd 2 bd 1 1 1 1 1 1 2 T (y)d 1 1 1 . 2 yd 2 dψ =Z T (y)ddψ 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 2 T (b)d + ε. 2 xd 2 xd 1 1 1 1 1 1 2 xd 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)T1(d 2 )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 1 + ε. 1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 1 1 1 1 1 2 xd 2 xd . And since d for any x ∈ M is self-adjoint since ε is arbitrary. Fi- That is (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)T1(d nally, decomposing any x ∈ M into its real and imaginary parts, we get (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)T1(d 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2 is norm dense in L1(M; ψ), T1 extends to a bounded pos- itive map on L1(M; ψ) with T1 ≤ 2. Thus it remains to reinforce the norm bound 2 to 1 which is a linear map on M = L1(M; ψ)∗. 1. To this end, we consider the adjoint map T ∗ 1 attains its norm at the identity 1 of M, i.e., Since T1 is positive, T ∗ T1 = T ∗ 1 (1) ≤ 1. Indeed, by condition (3), 1 (1). Hence we are reduced to showing T ∗ 1 is also positive and T ∗ 2 )(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 1 = T ∗ 2 M d 1 2 xd ≤ 1 1 1 (1)d 1 2 yd 1 2 dψ =Z T1(d 1 2 yd 1 2 )dψ = ϕ(T (y)) ≤ ϕ(y) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 2 in L1 2 M+d 1 1 1 2 yd 1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 for any y ∈ M+. We get T ∗ 1 (1) ≤ 1 by the density of d +(M; ψ), and hence Z T ∗ 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 1 1 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 1 2 T (x)d 1 2 xd (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d for any x ∈ M and our claim follows. 9 Combined with condition (1) and the abstract Riesz-Thorin Theorem, we obtain that 1 2p T (x)d 1 2p xd , for 1 < p < ∞, and x ∈ M. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 1 2p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)p ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)d 1 2p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)p 1 1 p 2p M d (M∗, M) (Theorem 36 in [29], or Kosaki's Moreover, since for 1 < p < ∞, Lp(M; ψ) = C 1 noncommutative interpolation theorem in [16]), in accordance with Theorem 4.2.2(a) in [3], 2p is ·p-norm dense in Lp(M; ψ). Therefore, the map Tp defined above we have that d also extends naturally to a positive contraction of Lp(M; ψ) → Lp(M; ψ) for each p and we still denote it by Tp (1 < p < ∞). From the idea in [29], M = L∞(M; ψ), M∗ = L1(M; ψ) and Lp(M; ψ), 1 < p < ∞, can be regarded as injective subspaces in M + M∗ (factually it is just M∗ for the σ-finite cases, see Definition 7.2 in [16]). In this situation it is easily seen that the maps T, Tp (1 < p < ∞) and T1 defined above coincide on L. For this reason, we may have a linear map on M + M∗, and the restriction of this map on M, Lp(M; ψ) (1 < p < ∞) and M∗ will be T , Tp (1 < p < ∞) and T1, respectively when considering M, Lp(M; ψ) (1 < p < ∞) and M∗ in M + M∗. Since this map on M + M∗ is deduced by T on M, it is viewed as the extension of T on M + M∗, and we still denote it by T . Then we arrive at the stage to show our main result. From above we know that the restriction of T on L1(M; ψ) satisfies the following conditions. • T (a)1 ≤ a1, for all a ∈ L1 • T is positive, i.e., T (a) ≥ 0 if a ∈ L1 +(M; ψ). +(M; ψ). Theorem 2.1. We assume T is a linear transform on a σ-finite von Neumann algebra M satisfying the conditions (1) − (3) above, and T extends to a linear positive contraction on L1(M; ψ). Then for any a ∈ L1 +(M; ψ) and any λ > 0, and any n ∈ N, there exists a projection en ∈ M such that enSr(a)en ≤ λenden for all r ∈ {0, 1, ..., n}, and ϕ(1 − en) ≤ 2 λZ adψ, where Sr(a) = 1 Furthermore, for any a ∈ L1 k=0 T k(a), d = dϕ dψ is the spatial derivative and 1 is the identity of M. +(M; ψ), there exists a projection e ∈ M such that for any r ∈ N, r+1Pr Z eSr(a)edψ ≤ 4λ, and ϕ(1 − e) ≤ 2 λZ adψ. 10 compute this value explicitly as follows. Proof. For any x ∈ M+, R T (a)xdψ is a positive linear functional on L1(M; ψ) since T R T (a)xdψ takes positive values for any a ∈ L1 Indeed, the property M = (L1(M; ψ))∗ implies that each x ∈ M+ gives a is positive. positive linear functional on L1(M; ψ), and T (a) is still positive, hence the functional action +(M; ψ). Or in other words, we may also 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T (a)1, 2 ∈ M+, in accordance with Proposition 8(4) in [11],(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)x 2 T (a) ∈ L1(M; ψ). The Proposition in page 159 of [11] yields that R T (a)xdψ = 2 dψ. For any family {ξα} in D(H, ψ) such that Pα θψ(ξα, ξα) = 1, hence from R x the definition ofR ·dψ in page 163 of [6], we have 2 T (a)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)x Since x 2 T (a)x i.e., x 1 1 1 1 1 1 1 2 T (a)x 1 hx 2 T (a)x 1 2 ξα, ξαi 1 2 dψ =Xα Z x =Xα 1 1 hT (a) 2 x 2 ξα, T (a) 1 2 x 1 2 x ≥ 0. 1 2 ξαi =Xα (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)T (a) 2 1 2 ξα(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +(M; ψ) and x ∈ M+. That is to say,R T (a)xdψ ≥ 0 for any a ∈ L1 p + 1 If the reader is more familiar with the properties of Haagerup's Lp-spaces, we may recall the Proposition 1.20 in [9], i.e., let p, q ∈ [1, ∞], 1 q = 1 and let a ∈ Lq(M), then a ≥ 0 if and only if tr(ab) ≥ 0 for any b ∈ Lp(M)+, and Hilsum's spatial Lp-spaces are get the positivity of the functional at the beginning of this proof. isometrically isomorphic to Haagerup's Lp-spaces, by replacing tr(·) byR ·dψ, we may also Therefore there exists some ex ∈ M+ such that R T (a)xdψ = R aexdψ again by the fact M = (L1(M; ψ))∗. If we denoteex by eT (x) for each x ∈ M+, then and eT extends linearly to be a linear transform on M such that eT (x) ≥ 0 if x ≥ 0. Z T (a)xdψ =Z aeT (x)dψ, Also we note that for any x ∈ M+, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)eT (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = sup(cid:26)Z aeT (x)dψ a1 ≤ 1, a ∈ L1 = sup(cid:26)Z T (a)xdψ a1 ≤ 1, a ∈ L1 +(M; ψ)(cid:27) +(M; ψ)(cid:27) ≤ x , where the last inequality is because T (a)1 ≤ a1 for any a ∈ L1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)eT (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ x can also be obtained from the fact that T is a contraction on L1(M; ψ) and eT is just the adjoint map of T . +(M; ψ). This result 11 Let n ≥ 1 be fixed, we put K = {(x0, x1, . . . , xn) xr ∈ M+ for 0 ≤ r ≤ n and xr ≤ 1}, nXr=0 then K is σ-weakly compact in M × M × · · · × M of (n + 1)-copies. 1 1 1 1 2 xd 2 = (x is given by µ1(x) = d We know from Section 2.3 in [29] that the embedding map µ1 of L+ into L1(M; ψ)+ dψ is the spatial dψ dψ = ϕx(1) < ∞ since ϕx is a positive normal linear functional on M. And for any x ∈ M+, as M = L in σ-finite case, 2 ) for x ∈ L+, where d = dϕ derivative. Hence we may consider the valueR d R d 2 dψ < ∞ for any x ∈ M+ is well-defined. Therefore for any a ∈ L+ 1 (M; ψ), we may define a function g on K by 2 =R dϕx 2 )∗(x 2 xd 2 xd 2 d 2 d 1 1 1 1 1 1 g((x0, x1, ..., xn)) = (r + 1)Z Sr(a)xrdψ − λ nXr=0 1 1 2 dψ. 2 xrd (r + 1)Z d nXr=0 +(M; ψ), so R Sr(b)xrdψ is well- Note that Sr(b)xr ∈ L1(M; ψ) as xr ∈ M and Sr(b) ∈ L1 defined and takes finite values for each 0 ≤ r ≤ n. We recall that K is σ-weakly compact and g is σ-weakly continuous on K, hence the finite maximum value of g is attained for some choice of (x0, x1, . . . , xn) in K. Then we explain the reason that g is σ-weakly continuous. dψ xrdψ = φr(xr) for 2 dψ = dψ xrdψ = ϕ(xr), for xr ∈ M, is σ-weakly continuous on M for ϕ is a normal For each r ∈ N∪{0}, there exists a φr ∈ M∗ such thatR Sr(a)xrdψ =R dφr xr ∈ M, and φr is normal, i.e., φr is σ-weakly continuous on M. Also we haveR d R dxrdψ =R dϕ We let 1−Pn on n). For any x ∈ M+ with x ≤ zn, we havePn g((x0, x1, . . . , xn)) ≥ g((xr + δ(r, r0)x)r=0,1,...,n), where δ(r, r0) =(1 r = r0 r=0 xr = zn (we choose this notation because the positive operator zn depends r=0 xr + x ∈ M+ and r=0 xr + x ≤ 1 andPn state on M. Thus g is σ-weakly continuous on K. hence for any fixed r0 ∈ {0, 1, . . . , n}, 0 r 6= r0. 2 xrd 1 1 As a result, we have nXr=0 nXr=0 ≥ 2 dψ 1 1 2 xrd nXr=0 (r + 1)Z d (r + 1)Z Sr(a)xrdψ − λ (r + 1)Z Sr(a)xrdψ + (r0 + 1)Z Sr0(a)xdψ− (r + 1)Z d nXr=0 2 dψ − λ(r0 + 1)Z d 2 xrd 2 xd 1 1 1 1 2 dψ. − λ 12 Then we get (r0 + 1)Z Sr0(a)xdψ ≤ λ(r0 + 1)Z d 1 2 xd 1 2 dψ for any r0 ∈ {0, 1, . . . , n}. From the above inequality, for any x ∈ M+, x ≤ zn, any r in {0, 1, . . . , n}, we have (2.1) (2.2) Take y = (yr)r=0,1,...,n with and we havePn−1 and for any a ∈ L1 Therefore eT (1) ≤ 1 and we getPn−1 As a result, we obtain that g(x0, x1, . . . , xn) ≥ g(y). That is to say, 0 1 1 2 xd 2 dψ. r = n, n−1Xr=0 +(M; ψ), we have Z Sr(a)xdψ ≤ λZ d yn =(eT (xr+1) 0 ≤ r ≤ n − 1, r=0 eT (xr+1) ≤ 1. Indeed, because eT is linear, n−1Xr=0 eT (xr+1) = eT ( xr+1) ≤ eT (1), Z aeT (1)dψ =Z T (a)dψ = T (a)1 ≤ a1 . r=0 eT (xr+1) ≤ 1, so y is in K. (r + 1)Z d (r + 1)Z Sr(a)xrdψ − λ nXr=0 nXr=0 (r + 1)Z T (Sr(a))xr+1dψ − λ n−1Xr=0 n−1Xr=0 2eT (xr+1)d (r + 1)Z T (Sr(a))xr+1dψ (r + 1)Z Sr(a)xrdψ − n−1Xr=0 nXr=0 (r + 1)Z d (r + 1)Z d n−1Xr=0 nXr=0 2eT (xr+1)d (r + 1)Z Sr(a)xrdψ − (r + 1)Z T (Sr(a))xr+1dψ (r + 1)Z d 2 xrd 2 dψ 2 xrd 2 dψ − λ ≥λ 1 1 ≥ 1 1 1 n−1Xr=0 1 2 dψ. 1 1 2 dψ. Hence from the above inequality, it follows We compute the left hand side of (2.2), and it equals nXr=0 =Z S0(a)x0dψ + =Z S0(a)x0dψ + nXr=1 n−1Xr=0 (r + 1)Z Sr(a)xrdψ − n−1Xr=0 (r + 2)Z Sr+1(a)xr+1dψ − (r + 1)Z T (Sr(a))xr+1dψ (r + 1)Z T (Sr(a))xr+1dψ. n−1Xr=0 13 Since (r + 2)Sr+1(a) = a + T (a) + · · · + T r(a) + T r+1(a) and (r + 1)T Sr(a) = T (a) + T 2(a) + · · · + T r(a) + T r+1(a), the left hand side of (2.2) equals Therefore, inequality (2.2) becomes nXr=0Z axrdψ ≥ λ So we obtain the following inequality, ≥λ from the following inequalities, 1 1 1 1 2eT (xr+1)d 1 1 1 1 2 xrd 2 dψ 2 dψ ≥ 0, 2 xrd 2 dψ − λ 2 xrd 2 dψ − λ rZ d 1 2eT (xr)d 1 nXr=0Z d nXr=1 n−1Xr=0Z axr+1dψ nXr=1Z axrdψ = nXr=0Z axrdψ. (r + 1)Z d n−1Xr=0 Z S0(a)x0dψ + =Z ax0dψ + (r + 1)Z d nXr=0 nXr=0Z axrdψ − λ rZ d nXr=0 Z d 2eT (xr)d = sup{Z d 2eT (xr)d = sup{Z xrT (d = sup{Z d ≤ sup{Z d =Z d nXr=0Z axrdψ ≥ λ 0 sdps be the spectral decomposition of 1 −Pn nXr=0Z d 2 T (x)dψ x ∈ M+, x ≤ 1} 2 xrd 2 xdψ x ∈ M+, x ≤ 1} 2 xdψ x ∈ M+, x ≤ 1} 2 )dψ x ∈ M+, x ≤ 1} 2 xrd 2 dψ. 2 xrd 2 dψ, 1 1 2 xrd 1 1 2 xd 1 2 dψ 1 1 1 1 1 1 1 1 0 ≤ ym ≤ 1 − pm−1 ≤ m(1 − xr). nXr=0 14 1 2 dψ. (2.3) r=0 xr, suppose Hence we have for any a ∈ L1 +(M; ψ). Letting 1 −Pn r=0 xr = R 1 y ∈ M+, y ≤ 1. Writing ym = (1 − pm−1)y(1 − pm−1) for each m ∈ N, we have Hence by (2.1), for 0 ≤ r ≤ n. Z Sr(a)ymdψ ≤ λZ d 1 2 ymd 1 2 dψ hence enSr(a)en ≤ λenden for 0 ≤ r ≤ n because M = (L1(M; ψ))∗. r=0 xr ∈ M+, we have Taking the limit as m → ∞, and putting en = 1 − p0, we get 1 1 2 dψ 2 enyend Z enSr(a)enydψ =Z Sr(a)enyendψ ≤ λZ d = λZ endenydψ, Therefore for r = 0, we get enaen ≤ λenden, and sincePn xr)dψ ≤ λZ den( nXr=0 xr)dψ ≥ λZ d(1 − en)( xr)dψ, nXr=0 nXr=0 0 sdps = 0, we obtain that 1 − en = (1 − en)(Pn nXr=0 xr)dψ. Z aen( Z a(1 − en)( r=0 xr = p0R 1 Z d 1 λZ a(1 − en)dψ ≤ 1 λZ adψ. 1 2 (1 − en)d 1 2 dψ ≤ Since p01 − p0Pn we get which together with (2.3) gives r=0 xr). Hence From the above procedure, we get en for each n ∈ N such that ϕ(1 − en) ≤ 1 enSr(a)en ≤ λenden for any r in {0, 1, . . . , n}. Choose a subnet enk which converges weakly to some h ∈ M with 0 ≤ h ≤ 1. λR adψ and We assume that Sr(a) ∈ L1 ∗ for each r ∈ N. Since M acts standardly on H and each φr is normal and hence a vector state, for any fixed r ∈ N, there exists a vector ξr in H such that φr = ωξr,ξr . As a result, +(M; ψ) corresponds to some φr ∈ M + Z hSr(a)hdψ = φr(h2) = hh2ξr, ξri = hξr2 . The subnet enk converges to h in the weak operator topology, and for each r ∈ N, we have enkξr → hξr in the weak topology of H when k → ∞. Since the Hilbert space norm · is lower semicontinuous relative to this topology, we have hξr2 ≤ lim inf k→∞ enkξr2 . (2.4) 15 Hence, we have Combined with Z hSr(a)hdψ ≤ lim inf k→∞ enk ξr2 . dϕ dψ for k ∈ N such that nk is larger than r, we conclude that enkξr2 = φr(enk) =Z enk Sr(a)enkdψ ≤λZ enk denkdψ = λZ enk Z hSr(a)hdψ ≤ λ lim Taking the spectral decomposition for h =R 1 M, let g =R 1 s−1des. Then for each r ∈ N, k→∞ 1 2 dψ = λϕ(enk), ϕ(enk) = λϕ(h), since ϕ is a normal state on M, i.e., ϕ is weakly continuous on the unit ball of M. 0 sdes, and let e = 1 − e 1 2 ∈ M for e 1 2 is in Z eSr(a)edψ =Z ghSr(a)hgdψ ≤ 4Z hSr(a)hdψ, thanks to e = gh and g ≤ 2. Therefore we obtain that Z eSr(a)edψ ≤ 4Z hSr(a)hdψ ≤ 4λϕ(h) ≤ 4λϕ(1) = 4λ. Moreover, for the reason 1 − e = e 1 2 ≤ 2(1 − h), we get ϕ(1 − e) ≤ 2ϕ(1 − h) = 2 lim k→∞ ϕ(1 − enk ) = 2 lim k→∞Z d 1 2 (1 − enk)d 1 2 dψ ≤ 2 λZ adψ, also because ϕ is a normal state on M. We should point out that we learned much from Professor Kosaki and Professor Xu in the proof of this theorem. We turn to the case that M is semifinite, L1(M; ψ) is equivalent to the tracial L1(M). Moreover, since the construction of the Hilsum Lp-spaces is independent on the choice of the normal semifinite faithful weight ψ in the isometrically isomorphic sense, hence we may choose a special case that ψ(·) = ϕ(J · J), then the spatial derivative dϕ dψ becomes △ϕ, i.e., the modular operator associated with ϕ and in the semifinite case, ϕ can be assumed to be a 16 normal faithful semifinite trace. Considering the modular automorphism group induced by a trace is trivial, we get △ϕ = 1 in this case, then the result in the above theorem is for any a ∈ L1 +(M; ψ) and any λ > 0, and any n ∈ N, there exists a projection en ∈ M such that enSr(a)en ≤ λen for all r ∈ {0, 1, ..., n} and ϕ(1 − en) ≤ 2 in [30], Sr(a) hSr(a) λR adψ, then by Yeadon's method 2 h weakly as k → ∞. In fact, for ξ1 ∈ H, ξ2 ∈ D(Sr(a) 2 enk ξ1, ξ2i = henkξ1, Mr(a) 2 ξ2i → hhξ, Sr(a) 2 enk → Sr(a) 2 ), we have 2 ξ2i, and 1 1 1 1 1 1 1 (cid:12)(cid:12)(cid:12)hhξ1, Sr(a) 2 enk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = enkSr(a)enk hSr(a) 1 1 Since(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Sr(a) Hence it follows 2 ξ2i(cid:12)(cid:12)(cid:12) = lim k→∞(cid:12)(cid:12)(cid:12)hSr(a) 2 ≤ λ 1 1 1 2 enkξ1, ξ2i(cid:12)(cid:12)(cid:12) ≤ λ 2 if nk ≥ r, so that hξ1 ∈ D(Sr(a) 1 2 ) and 1 2 ξ1 ξ2 . 2 hξ1, ξ2i = hhξ1, Sr(a) 1 2 ξ2i = lim k→∞ hSr(a) 1 2 enkξ1, ξ2i. 2 1 2 hξ1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 ≤ lim k→∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Sr(a) 1 2 enk ξ1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) λhenkξ1, ξ1i = λhhξ1, ξ1i, henkSr(a)enkξ1, ξ1i ≤ lim k→∞ = lim k→∞ hhSr(a)hξ1, ξ1i =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Sr(a) , g =R 1 2 i.e., hSr(a)h ≤ λh for each r ∈ N. Taking the spectral decomposition for h =R 1 0 sdes, and s−1des. Then we have eSr(a)e ≤ λghg ≤ 2λe, this is just the result of let e = 1 − e 1 1 2 Yeadon's Theorem 1 in [30]. Then we introduce the conceptions of "type" and "weak type" for the action of T on M + M∗. Such conceptions "type" and "weak type" appeared in the classical real analysis first and have been widely used in classical ergodic theorems. They are modified by Junge- Xu (pp 396 -- 397, [15]), for the framework of noncommutative Lp-spaces associated with a semifinite von Neumann algebra. Here we rewrite them for the transform T we constructed above in the framework of Lp(M; ψ)'s (1 ≤ p ≤ ∞). For each n ∈ N, Sn is a linear map on M + M∗ satisfying the conditions (1) − (3). Thus S = (Sn)n∈N is a map which sends a positive element in Lp(M; ψ) for some fixed p, 1 ≤ p ≤ ∞ to a sequence of positive elements in Lp(M; ψ). Here we identify L1(M; ψ) with M∗ and L∞(M; ψ) with M. We say that S is of type (p, p), (1 ≤ p ≤ ∞) if there is a positive constant c such that +(M; ψ) satisfying ap ≤ c xp and Sn(x) ≤ a, for any +(M; ψ), there is a ∈ Lp for any x ∈ Lp n ∈ N. 17 Then we say that S is of weak type (p, p), (1 ≤ p < ∞), if there is a positive constant c such that for any x ∈ Lp +(M; ψ), and any λ > 0 there is a projection e ∈ M such that λ (cid:19)p ϕ(1 − e) ≤(cid:18) c xp and eSn(x)e ≤ λ1, for any n ∈ N, where 1 is the identity of the von Neumann algebra M. Yeadon's Theorem shows that S = (Sr)r∈N is of weak type (1, 1) when M is semifinite. Indeed, by condition (1), we get that for each r ∈ N, if x ∈ M+ = L∞ It is obvious that S = (Sr)r∈N is of type (∞, ∞) for an arbitrary von Neumann algebra M. + (M; ψ), we have T r(x) ≥ 0 for n ∈ N and T r(x) ≤ x and thus Sr(x) ≤ x. Hence if we put a = x 1 ∈ L∞ + (M; ψ), we have a = x and Sr(x) ≤ a, for all r ∈ N. As we have mentioned in the previous section, such a weak type conception is no longer appropriate for the non-tracial cases. We give a pre-version of pre-weak type here, we hope it is of some meaning. We say that S is of pre-weak type (p, p), (1 ≤ p < ∞) if there is a positive constant c such that for any x ∈ Lp +(M; ψ), and any λ > 0 there is a projection e ∈ M such that ϕ(1 − e) ≤(cid:18) c xp λ (cid:19)p and eSn(x)ep ≤ λ, for any n ∈ N, where 1 is the identity of the von Neumann algebra M. Theorem 2.1 shows that S = (Sr)r∈N is of pre-weak type (1, 1) for a σ-finite von Neumann algebra. If we would like to obtain a satisfactory non-tracial extension of Theorem 1.1, one possible method is to consider real interpolation theory (possible non-tracial real interpolation theory for pre-weak type, though we are not very sure about the existence of such theory and it is still in exploring) of von Neumann algebras, because the real interpolation theory always provides us the type for midpoints from the weak type assumption of endpoints. But as is pointed out by Junge-Xu (pp 385 -- 386, [15]), in contrast with the classical theory, the noncommutative nature of weak type (1, 1) inequalities seems a priori unsuitable for classical interpolation arguments. More accurately speaking, Pisier-Xu gave a counterexample saying that the complex interpolation space Lp(M) may not coincide with the real interpolation space Lp,p(M) for non-tracial von Neumann algebras if we establish non-commutative real interpolation theory verbatim from classical one (p.1472, Example 3.3, [23]). We had tried several ways to modify the definition of real interpolation construction in order to suit well to the von Neumann algebra theory and the complex interpolation of Terp ([29]) at the same time. Unfortunately, we have not obtained any valid method for this matter so far. Let us 18 point out that a key obstacle in this work is the absense of generalized singular numbers for the non-tracial cases. The generalized singular number function ([8]) is a powerful tool when dealing with the τ -measurable operators associated with a semifinite von Neumann algebra, and the noncommutative tracial Lp-spaces are just consisting of such operators. In the non-tracial cases, whether there exists such a counterpart theory which contains the generalized singular number theory by [8] has not been sufficiently understood now. We do not know whether there is any other method available and we will continue to explore this problem in the future. Finally, we give three examples as applications of Theorem 2.1. Example 2.2. Let (Ω, F, µ) be a finite measure space and N be a σ-finite von Neumann algebra equipped with a normal faithful state ϕ1. We consider the von Neumann algebra tensor product (M, ϕ2) = (L∞(Ω), µ)⊗(N, ϕ1), where ϕ2 is a normal faithful state since µ is finite and ϕ1 is a normal faithful state. For 1 ≤ p < ∞, the corresponding noncommutative Lp(M; ψ2) is just Lp(Ω, Lp(N; ψ1)), the usual Lp-space of strongly measurable p-integrable functions on Ω with values in Lp(N; ψ1), where ψ1 (resp. ψ2) is a normal faithful state on the commutant of N (resp. M), and we may choose ψ1 (resp. ψ2) to be associated with ϕ1 (resp. ϕ2) by the Tomita-Takesaki theory. Now let S be a linear map on L∞(Ω) satisfying conditions (1) − (3) (with M = L∞(Ω) there), then T = I ⊗ S is a linear map on M verifying the same conditions (with M = L∞(Ω)⊗N there). From Theorem 2.1, for any a ∈ L1(Ω, L1 +(N; ψ1)) and any λ > 0, and any n ∈ N, there exists a projection en ∈ M such that and enSr(a)en ≤ λenden for all r ∈ {0, 1, ..., n}, ϕ2(1 − en) ≤ 2 λZ adψ2, where Sr(a) = 1 M. Furthermore, for any a ∈ L1(Ω, L1 for any r ∈ N, k=0 I ⊗ Sk(a), d = dϕ2 r+1Pr dψ2 is the spatial derivative and 1 is the identity of +(N; ψ1)), there exists a projection e ∈ M such that Z eSr(a)edψ2 ≤ 4λ, and ϕ2(1 − e) ≤ 2 λZ adψ2. Example 2.3. Let M be a von Neumann algebra with a normal faithful state ϕ, and let N be any von Neumann subalgebra of M. The generalized conditional expectation ε : M → N relative to ϕ defined by Accardi-Cecchini is given as ε(x) = JN PN Jπϕ(x)JPN JN for any 19 x ∈ M (see [1]). (1) − (3). Therefore, Theorem 2.1 implies that for any a ∈ L1 any n ∈ N, there exists a projection en ∈ M such that If we regard ε to be a linear map from M to M, ε satisfies conditions +(M; ψ) and any λ > 0, and enSr(a)en ≤ λenden for all r ∈ {0, 1, ..., n}, and ϕ(1 − en) ≤ 2 λZ adψ, where Sr(a) = 1 Furthermore, for any a ∈ L1 k=0 T k(a), d = dϕ dψ is the spatial derivative and 1 is the identity of M. +(M; ψ), there exists a projection e ∈ M such that for any r ∈ N, r+1Pr Z eSr(a)edψ ≤ 4λ, and ϕ(1 − e) ≤ 2 λZ adψ. If N is globally invariant under the modular automorphism group σϕ t , then ε will be the conditional expectation in the sense of [27], i.e., a projection of norm one. As a projection is idempotent, we have enε(a)en ≤ λenden andR eε(a)edψ ≤ 4λ in this case. Example 2.4. Let {(Mi, ϕi)}i∈I be a family of von Neumann algebras and each ϕi is a normal faithful state. Let (M, ϕ) = ∗i∈I(Mi, ϕi) be the von Neumann algebra reduced free product (see [20]), and hence ϕ is a normal faithful state on M. Put M ◦ i = {x ∈ Miϕi(x) = 0}, then Mi = C1Mi ⊕ M ◦ i , and let Ti : Mi → Mi be defined by TiC1Mi = IC1Mi and TiM ◦ i = exp(−1)IM ◦ i , and {Ti}i∈I defines a positive linear map T on M by free product, and T is uniquely deter- mined by its action on monomials: T (x1x2 · · · xn) = Ti1(x1)Ti2(x2) · · · Tin(xn) = exp(−n)x1x2 · · · xn, for any x1, x2, · · ·, xn with xk ∈ M ◦ ik and i1 6= i2 6= · · · 6= in. The map T is called the free product of the family {Ti}i∈I. Then T satisfies conditions (1) − (3) with respect to M since each Ti satisfies conditions (1) − (3) with respect to Mi. Hence T extends to a positive linear map on L1(M; ψ) and by Theorem 2.1, for any a ∈ L1 +(M; ψ) and any λ > 0, and any n ∈ N, there exists a projection en ∈ M such that enSr(a)en ≤ λenden for all r ∈ {0, 1, ..., n}, and ϕ(1 − en) ≤ 2 λZ adψ, 20 where Sr(a) = 1 Furthermore, for any a ∈ L1 k=0 T k(a), d = dϕ r+1Pr Z eSr(a)edψ ≤ 4λ, for any r ∈ N, and ϕ(1 − e) ≤ dψ is the spatial derivative and 1 is the identity of M. +(M; ψ), there exists a projection e ∈ M such that 2 λZ adψ. 21 Acknowledgements The author would like to take this opportunity to express his highest respect and appreci- ation to his supervisor Professor Y.Kawahigashi, who led him into the realm of von Neumann algebra theory and paid great care and patience for him in the past five years. And the au- thor is also very grateful to Professor H.Kosaki in Kyushu University and Professor Q.Xu in Universit´e de Franche-Comt´e from whom he learned many techniques on noncommutative Lp-spaces through e-mail correspondence, and they kindly pointed out mistakes in the previ- ous version of this paper. And he also thanks for Professor N.Ozawa and Professor Y.Ogata and other ones in their group, since he benefited greatly from seminars and discussions in the group. Finally, his thanks go to Mr. Zhao in his research room, for much help from Mr. Zhao. References [1] L. Accardi, C. Cecchini, Conditional expectations in von Neumann algebras and a the- orem of Takesaki, J. Funct. Anal. 45 (1982) 245-273. MR0647075(84j:46088) [2] H. Araki, T. Masuda, Positive cones and Lp-spaces for von Neumann algebras, Publ. Res. Inst. Math. Sci. 18 (1982) 759-831. MR0677270(84h:46082) [3] J. Bergh, J. Lofstrom, Interpolation Spaces, Springer-Verlag, Berlin, 1976. MR0482275(58:2349) [4] O. Bratteli, D. Robinson, Operator Algebras and Quantum Statistical Mechanics. I, C ∗- and W ∗-algebras, symmetry groups, decomposition of states, in: Texts and Monographs in Physics, Springer-Verlag, New York, 1987. MR0887100(88d:46105) [5] H. A. Carteret, D. R. Terno, K. Zyczkowski, Dynamics beyond completely positive maps: some properties and applications, Phys. Rev. A 77 (2008) 042113.1-042113.8. MR2491057(2010d:82072) [6] A. Connes, On the spatial theory of von Neumann algebras, J. Funct. Anal. 35(1980) 153-164. MR0561983(81g:46083) [7] N. Dunford, J. T. Schwartz, Linear Operators, I, General Theory, in: Applied Mathe- matics, Vol. 7, Interscience Publishers, Inc., New York, 1958. MR0117523(22:8302) 22 [8] T. Fack, H. Kosaki, Generalized s-numbers of τ -measurable operators, Pacific J. Math. 123(1986) 269-300. MR0840845(87h:46122) [9] U. Haagerup. Lp-spaces associated with an arbitrary von Neumann algebra, In: Alg`ebres d'op´erateurs et leurs applications en physique math´ematique, Proc. Col- loq., Marseille, 1977, Volume 274 of Colloq. Internat. CNRS, 175-184. Paris, 1979. MR0560633(81e:46050) [10] U. Haagerup, M. Junge, Q. Xu, A reduction method for noncommutative Lp-spaces and applications, Trans. Amer. Math. Soc. 362 (2010) 2125-2165. MR2574890 [11] M. Hilsum, Les espaces Lp d'une alg`ebre de von Neumann d´efinies par la deriv´ee spatiale, J. Funct. Anal. 40 (1981) 151-169. MR0609439(83c:46053) [12] P. Horodecki, R. Augusiak, M. Demianowicz, General construction of noiseless networks detecting entanglement with the help of linear maps, Phys. Rev. A 74 (2006) 052323.1- 052323.6. MR2288089(2007k:81023) [13] H. Izumi, Constructions of non-commutative Lp-spaces with a complex pa- Internat. J. Math. 8 (1997) 1029-1066. rameter arising from modular actions, MR1484866(99a:46114) [14] M. Junge, Doob's inequality for non-commutative martingales, J. Reine Angew. Math. 549 (2002) 149-190. MR1916654(2003k:46097) [15] M. Junge, Q. Xu, Noncommutative maximal ergodic theorems, J. Amer. Math. Soc. 20 (2007) 385-439. MR22767755(2007k:46109) [16] H. Kosaki, Applications of the complex interpolation method to a von Neumann algebra: noncommutative Lp-spaces, J. Funct. Anal. 56 (1984) 29-78. MR0735704(86a:46085) [17] E. C. Lance, Ergodic theorems for convex sets and operator algebras, Invent. Math. 37 (1976) 201-214. MR0428060(55:1089) [18] T. Mei, Operator-valued Hardy spaces, Mem. Amer. Math. Soc. 188, no.881, American Mathematical Society, 2007. MR2327840(2010d:46085) [19] E. Nelson, Notes on non-commutative integration, J. Funct. Anal. 15 (1974) 103-116. MR0355628(50#8102) 23 [20] A. Nica, R. Speicher, Lectures on the combinatorics of free probability, in: London Mathematical Society Lecture Notes Series, 335, Cambridge University Press, Cam- bridge, 2006. MR2266879(2008k:46198) [21] G. Pisier, Non-commutative vector valued Lp-spaces and completely p-summing maps, Ast´erisque 247 (1998). MR1648908(2000a:46108) [22] G. Pisier, The operator Hilbert space OH and type III von Neumann algebras, Bull. London. Math. Soc. 36 (2004) 455-459. MR2069007(2005c:46082) [23] G. Pisier, Q. Xu, Non-commutative Lp-spaces. In: Handbook of the Geometry of Banach Spaces, Vol. 2, 1459-1517, North-Holland, Amsterdam, 2003. MR1999201(2004i:46095) [24] A. Royer, Families of positivity preserving but not completely positive superoperators, Phys. Lett. A 336 (2005) 295-310. MR2119628(2005j:81022) [25] A. Shaji, E. C. G. Sudarshan, Who's afraid of not completely positive maps? Phys. Lett. A 341 (2005) 48-54. MR2144129(2006g:82029) [26] S. J. Szarek, E. Werner, K. Zyczkowski, Geometry of sets of quantum maps: a generic positive map acting on a high-dimensional system is not completely positive, J. Math. Phys. 49 (2008) 032113.1-032113.21. MR2406781(2009b:81025) [27] M. Takesaki, Conditional expectations in von Neumann algebras, J. Funct. Anal. 9 (1972) 306-321. MR03033078(46#2445) [28] M. Takesaki, Theory of Operator Algebras, II, In: Encyclopaedia of Mathematical Sciences, 125, Operator Algebras and Non-commutative Geometry, 6, Springer-Verlag, Berlin, 2003. MR1943006(2004g:46079) [29] M. Terp, Interpolation spaces between a von Neumann algebra and its predual, J. Operator. Theory 8 (1982) 327-360. MR0677418(85b:46075) [30] F. J. Yeadon, Ergodic theorems for semifinite von Neumann algebras I, J. London. Math. Soc. 16 (1977) 326-332. MR0487482(58:7111) 24
1310.7880	1	1310	2013-10-29T17:06:59	Radial multipliers on arbitrary amalgamated free products of finite von Neumann algebras	[ "math.OA", "math.GR" ]	Let $(M_i)_{i}$ be a (finite or infinite) family of finite von Neumann algebras with a common subalgebra $P$. When $\varphi:\IN\rightarrow\IC$ is a function, we define the radial multiplier $M_\varphi$ on the amalgamated free product $M=M_1\free_P M_2\free_P\ldots$ setting $M_{\varphi}(x)=\varphi(n)x$ for every reduced expression $x$ of length $n$. In this paper we give a sufficient condition on $\varphi$ to ensure that the corresponding radial multiplier $M_\varphi$ is a completely bounded map, and moreover we give an upper bound on its completely bounded norm. Our condition on $\varphi$ does not depend on the choice of von Neumann algebras $(M_i)_i$ and $P$. This result extends earlier results by Haagerup and M\"oller, who proved the same statement for free products without amalgamation, and M\"oller showed that the same statement holds when $P$ has finite index in each of the $M_i$.	math.OA	math	RADIAL MULTIPLIERS ON ARIBITRARY AMALGAMATED FREE PRODUCTS OF FINITE VON NEUMANN ALGEBRAS STEVEN DEPREZ Abstract. Let (Mi)i be a (finite or infinite) family of finite von Neumann algebras with a common subalgebra P . When ϕ : N → C is a function, we define the radial multiplier Mϕ on the amalgamated free product M = M1 ∗P M2 ∗P . . . setting Mϕ(x) = ϕ(n)x for every reduced expression x of length n. In this paper we give a sufficient condition on ϕ to ensure that the corresponding radial multiplier Mϕ is a completely bounded map, and moreover we give an upper bound on its completely bounded norm. Our condition on ϕ does not depend on the choice of von Neumann algebras (Mi)i and P . This result extends earlier results by Haagerup and Moller, who proved the same statement for free products without amalgamation, and Moller showed that the same statement holds when P has finite index in each of the Mi. Introduction Let Γ be a countable group. A function f : Γ → C is called a Herz-Schur multiplier if the map mf : C Γ → C Γ that is defined by mf (ug) = f (g)ug extends to a σ-weakly continuous completely bounded map mf : L Γ → L Γ. The Herz-Schur norm kf kHS of f is the completely- bounded norm of mf . The free groups Fn (2 ≤ n ≤ ∞) come equiped with a natural word-length function ·, associated to the canonical generators. Given a map ϕ : N → C, we can look at the associated radial function fϕ : Fn → C, that is given by fϕ(g) = ϕ(g). Haagerup and Szwarc [HSS] showed that fϕ is a Herz-Schur multiplier if and only if the Hankel matrix that is given by Hϕ = (ϕ(n + m) − ϕ(n + m + 2))n,m is trace class. In that case, there exist constants c± ∈ C such that ϕ(n) = ϕ0(n) + c+ + (−1)nc− where ϕ0(n) converges to 0. The Herz-Schur norm of fϕ is bounded by In fact, this inequality becomes an equality if n = ∞, and Haagerup and Szwarc give an exact expression for the Herz-Schur norm, also when n is finite. kfϕkHS ≤ kHϕk1 + c+ + c− . We say that a function ϕ : N → C is in class C if the Hankel matrix Hϕ is trace class, and we will write kϕkC = kHϕk + c+ + c−. corresponding author, University of Copenhagen, [email protected]. Supported by ERC Advanced Grant no. OAFPG 247321. Supported by the Danish National Research Foundation through the Centre for Symmetry and Deformation. (DNRF92). Department of mathematics, Copenhagen university, Universitetsparken 5, 2500, Copenhagen O. 1 2 STEVEN DEPREZ Wysoczanski [W] studied radial multipliers on free products Γ = ∗i Γi of groups. A free product comes equiped with a natural word-length function g 7→ g, so it makes sense to talk about radial multipliers on free products. Wysoczanski showed that a radial function fϕ is a Herz-Schur multiplier if the Hankel matrix Kϕ = (ϕ(n + m) − ϕ(n + m + 1))n,m is trace class. If this is the case, we say that ϕ is in class C′. If ϕ ∈ C′, then the limit c = limn→∞ ϕ(n) exists and the Hankel matrix eKϕ = (ϕ(n + m + 1) − ϕ(n + m + 2))n,m is also trace class. Wysoczanski showed that the Herz-Schur norm of fϕ is bounded by kfϕkHS ≤ kϕkC ′ = kKϕk1 +(cid:13)(cid:13)(cid:13)eKϕ(cid:13)(cid:13)(cid:13)1 + c . When all the groups Γi have the same (possibly infinite) cardinality, he can explicitly compute the Herz-Schur norm of fϕ. Wysoczanski's result has been extended to an operator algebraic framework by Haagerup and Moller [HM]. Namely, when M = ∗i Mi is a free product of von Neumann algebras, then M is σ-weakly densely spanned by the reduced words, i.e. words of the form x1 . . . xn where xk ∈ Mik \ C 1 and i1 6= i2 6= . . . 6= in. Given a function ϕ : N → C, we consider the radial multiplier mϕ that is given by mϕ(x1 . . . xn) = ϕ(n)x1 . . . xn for every reduced word x1 . . . xn. Haagerup and Moller showed that mϕ is a completely bounded map if the Hankel matrix Kϕ defined above is trace class. They also showed that in that case kmϕkcb ≤ kϕkC ′. Recently, Moller extended this result to include amalgamated free products where the amalgam has finite index in each of the factors Mi [M]. The present paper extends this result to arbitrary amalgamated free products of finite von Neumann algebras. We give an overview of the techniques used. It is instructive to first look at the ideas behind Wysoczanski's result. The Haagerup-Szwarc result can be restated as a result about radial Schur multipliers on trees, as we explain now. In this form, the result was published by Haagerup, Steenstrup and Szwarc [HSS], but is based on earlier unpublished notes by Haagerup and Szwarc. Let X be a countable set, and consider a map k : X × X → C. The Schur product k ⋆ T of k with a bounded operator T ∈ B(ℓ2(X)) is given by hδx, (k ⋆ T )δyi = k(x, y) hδx, T δyi . We say that k is a Schur multiplier is mk : T 7→ k ⋆ T is a well-defined completely bounded map from B(ℓ2(X)) to itself, and its Schur norm is kkkS = kmkkcb. Suppose now that d : X × X → R+ is a metric on X. We say that a Schur multiplier k : X × X → C is a radial Schur multiplier if it is of the form k(x, y) = ϕ(d(x, y)) for some function ϕ : R+ → C. The concepts of Schur multipliers and Herz-Schur multipliers are closely related: a function f : Γ → C on a group Γ is a Herz-Schur multiplier if and only if the function defined by k(g, h) = f (g−1h) is a Schur multiplier, and moreover kf kHS = kkkS. Let T be a regular tree, which we identify with its vertex set. We consider T as a metric space with the usual metric d : T × T → N, given by the length of the shortest path between two vertices. Haagerup, Steenstrup and Szwarc showed that a radial function kϕ : T × T → C, RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 3 defined by kϕ(x, y) = ϕ(d(x, y)) is a radial Schur multiplier if and only if ϕ ∈ C, and moreover its Schur norm is bounded by kkϕkS ≤ kϕkC. Using this restatement, the proof of Wysoczanski's result goes as follows. Observe that Γ = ∗i Γi is the fundamental group of the following graph of groups: {e} Γ2 Γ1 Γn Consider the action of Γ on the Bass-Serre tree T associated to this graph of groups. Observe that the vertices of the tree T are indexed by Γ ⊔ Γ/Γ1 ⊔ . . . ⊔ Γ/Γn. Consider the vertex v0 ∈ T that corresponds to the identity in Γ. For every group element g ∈ Γ, we see that the distance between v0 and g · v0 is exactly twice the word-length of g. Let ϕ : N → C be a function. Consider the function ψ : N → C that is defined by ψ(n) = ϕ(n/2) if n is even, and ψ(n) = 0 if n is odd. The function ψ is in class C if and only if ϕ is in class C′, and moreover kψkC = kϕkC ′. Consider the radial functions k = kψ : T × T → C and f = fϕ : Γ → C as before. We define an isometry u : ℓ2(Γ) → ℓ2(T ) ⊗ ℓ2(Γ) by the formula u(δg) = δg−1v0 ⊗ δ−1 g . Let L(Γ) act on ℓ2(Γ) by left multiplication, i.e. ugδh = δgh. We observe that the multipliers mf and mk are related by mf (x) = u∗(mk ⊗ id)(uxu∗)u for all x ∈ L(Γ). In particular, we see that kmf kcb ≤ kmkkcb ≤ kψkC = kϕkC ′ . At no point in the argument above did we use the fact that Γ was a free product as opposed to an amalgamated free product. We prove our result about amalgamated free products of von Neumann algebras using a similar strategy. For this, we have to find a good replacement for the "fundamental group of a graph of groups". We think that the right notion is that of the "relative Gaussian construction", which we describe in short below, and in more detail in section 2. They come naturally with a notion of a "reduced word", so it makes sense to talk about radial multipliers. We show that, for an important subclass of the "relative Gaussian constructions", a radial multiplier mϕ is completely bounded, whenever ϕ is in class C, and kmϕkcb ≤ kϕkC. This is only true for a subclass of the "relative Gaussian constructions", as it fails for the classical Gaussian construction. The relative Gaussian construction is a strong generalization of Voiculescu's free Gaussian construction [V]. Many generalizations of the free Gaussian construction have been intro- duced before. For example, there are the q-Gaussian constructions [BS1]. The construction that is closest to our "relative Gaussian construction" is given by Shlyakhtenko's "A-valued semicircular systems" [S]. One can describe the "relative Gaussian construction" as being a tracial A-valued semicircular system, but deformed in a way similar to the q-Gaussian con- structions. As with the free Gaussian construction and its generalizations, we first define a 4 STEVEN DEPREZ kind of Fock space and creation and annihilation operators on them. Then we define the "rel- ative Gaussian construction" as the algebra generated by certain creation and annihilation operators. For the construction of the relative Fock space, we need three pieces of data: • A von Neumann algebra M with a faithful normal state τ . • A bimodule H over M . • A self-adjoint M -M bimodular contraction F : H ⊗M H → H ⊗M H that satisfies the braid relation (F ⊗ 1H )(1H ⊗ F )(F ⊗ 1H ) = (1H ⊗ F )(F ⊗ 1H)(1H ⊗ F ), as operators on H ⊗M H ⊗M H. Consider the n-fold Connes tensor product bimodule H (n) of H with itself. By convention, we set H (0) = M . We define a number of M -M bimodular operators on H (n). Let σ ∈ Sn be a permutation on n elements. It is well-known that σ can be decomposed as σ = t1 . . . tn where each ti = (ki, ki + 1) is a transposition of two consecutive elements. Moreover, the shortest such decomposition is unique, up to applying the braid relation (k, k + 1)(k, k − 1)(k, k + 1) = (k, k − 1)(k, k + 1)(k, k − 1). We assume that σ = t1 . . . tn is such a shortest decomposition. Put F(k,k+1) = −1H (k−1) ⊗ F ⊗ 1H (n−k−1) , and observe that Fσ = Ft1 . . . Ftn is well-defined. Define an operator Dn by Dn = Xσ∈Sn Fσ. Bozejko and Speicher showed in [BS2] that such an operator Dn is always a positive operator. We define a new inner product h·, ·iF = h·, Dn·i on H (n). Denote by H (n) the M -M -bimodule that we obtain from H (n) by completion and separation with respect to this new inner product h·, ·iF . The relative Fock space is now F F M (H, F ) =Mn≥0 H (n) F In F M (H, F ), the vector 1 ∈ L2(M ) ⊂ F M (H, F ) plays a special role. It will be called the vacuum vector and is denoted by Ω to avoid confusion. For every ξ ∈ H, we define a left creation operator L(ξ) on F M (H, F ), by the formula L(ξ)η = ξ ⊗ η ∈ H (n+1) ⊂ H (n+1) F for all η ∈ H (n). This creation operator is in general a closable unbounded operator. We still denote its closure by L(ξ). The adjoint L(ξ)∗ is called an annihilation operator. When F = 0, then we get Dn = 1 for all n ∈ N, and the creation operators satisfy the relation L(ξ)∗L(η) = hξ, ηiM . In this case, we obtain the Fock space from Shlyakhtenko's A-valued semicircular elements. If M = C and F (ξ ⊗ η) = qη ⊗ ξ for some −1 < q < 1, we obtain the Fock space of the RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 5 (−q)-Gaussian construction, and the creation and annihilation operators satisfy the relation L(ξ)∗L(η) = hξ, ηi + qL(ξ)L(η)∗. In general, a similar relation holds, but it is slightly more complicated to write down. See section 2, proposition 2.4. In any case, the algebra TM (H, F ) = BM (F M (H, F )) of bounded right-M -linear operators on F M (H, F ) is the w∗-closed linear span of the left action of M and of (the spectral projections of the self-adjoint part of) operators of the form L(ξ1) . . . L(ξn)L(ηm)∗ . . . L(η1)∗ where ξ1, . . . , ξn, η1, . . . , ηm ∈ H. For the relative Gaussian construction, we need one more piece of data: an anti-unitary J : H → H that satisfies J(xξy) = y∗J(ξ)x∗, and that is compatible with F is some way. See section 2 for more details. Then we define Γ′′ M (H, F, J) to be the von Neumann algebra on F M (H, F ) that is generated by the left action of M and all the operators of the form L(ξ) + L(Jξ) with ξ ∈ H. Obviously the free Gaussian construction can be obtained from our construction by taking M = C, F = 0 and where J is given by complex conjugation on ℓ2(S) for a countable set S. More generally, the q-Gaussian construction can be obtained by setting F (ξ ⊗ η) = −qη ⊗ ξ. The tracial cases of Shlyakhtenko's A-valued semicircular systems can be obtained by setting F = 0 in our general construction. For us, a more important class of examples is given by amalgamated free products and HNN extensions. If M1 ∗P M2 is an amalgamated free product, then we set M = M1 ⊕ M2, and H = L2(M1) ⊗P L2(M2) ⊕ L2(M2) ⊗P L2(M1) The anti-unitary J interchanges the two components of the direct sum above. The operator F is the projection onto the closed subspace L2(M1) ⊗P L2(M1) ⊕ L2(M2) ⊗P L2(M2) = L2(M1) ⊗P L2(P ) ⊗P L2(M1) ⊕ L2(M2) ⊗P L2(P ) ⊗P L2(M2) ⊂ L2(M1) ⊗P L2(M2) ⊗P L2(M1) ⊕ L2(M2) ⊗P L2(M1) ⊗P L2(M2) = H (2). Then we obtain that Γ′′ M (H, F, J) = M2(C) ⊗ (M1 ∗P M2). More generally, suppose we are given a graph Γ of von Neumann algebras. A graph consists of a set V of vertices, and a set E of edges, together with source and target maps s, t : E → V , and with an involution · : E → E that reverses each edge, so s(e) = t(e), t(e) = s(e) and e = e. A graph of von Neumann algebras is a graph (V, E) together with a family of von Neumann algebras (Mv)v∈V and a family of von Neumann subalgebras Pe ⊂ Ms(e) (e ∈ E) and ∗-isomorphisms αe : Pe → Pe such that αe = α−1 e . Moreover, we assume that there are normal conditional expectations Ee : Ms(e) → Pe. 6 STEVEN DEPREZ Then we set M =Lv∈V Mv and we define H to be the completion of L2(Ms(e)) ⊗Pe L2(Mt(e)), Me∈E where the inclusion of Pe into Mt(e) is given by the isomorphism αe : Pe → Pe ⊂ Mt(e). The Jordan subspace is spanned (as a real vector space) by elements of the form x ⊗ y + y∗ ⊗ x∗ ∈ Ms(e) ⊗Pe Mt(e) + Mt(e) ⊗Pe Ms(e). The operator F is the projection onto the closed subspace of H (2) that is spanned by the spaces of the form Mt(e) ⊗Pe Pe ⊗Pe Mt(e). We also call Γ′′ M (H, F, J) the fundamental von Neumann algebra of the given graph of von Neumann algebras. With this terminology, we do not get that the group von Neumann algebra of a fundamental group of a gaph of groups is exactly the fundamental von Neumann algebra of the corresponding graph of von Neumann algebras. Instead the fundamental von Neumann algebra is the amplification of the group von Neumann algebra of the fundamental group. In fact, the fundamental von Neumann algebra is the groupoid von Neumann algebra of the fundamental groupoid of the graph of groups. Observe that in both of the previous examples, the contraction F is in fact a projection and moreover 1 ⊗ F commutes with F ⊗ 1 on H (3). We can not extend Haagerup-Moller's result to arbitrary relative Gaussian constructions, but we can extend it to all the cases where the contraction F is a projection such that 1 ⊗ F commutes with F ⊗ 1: Theorem 0.1 (see theorem 3.1). Let (M, τ ) be a tracial von Neumann algebra and let H be a Hilbert M -M bimodule. Assume that F is a projection onto an M -M subbimodule of H ⊗M H such that F ⊗ 1 commutes with 1 ⊗ F on H ⊗M H ⊗M H. Let ϕ : N → C be a function in class C defined above. Then there is a unique completely bounded map Φϕ : BM (F M (H, F )) → BM (F M (H, F )) that satisfies Φϕ(L(ξ1) . . . L(ξn)L(ηm)∗ . . . L(η1)) = ϕ(n + m)L(ξ1) . . . L(ξn)L(ηm)∗ . . . L(η1) Φϕ(x) = ϕ(0)x kΦϕkcb ≤ kϕkC . for all x ∈ M acting on the left The result about amalgamated free products follows by restricting this completely bounded map to a corner of a subspace. 1. Preliminaries and Notation In this paper, the set of natural numbers N includes 0. When we want to refer to the natural numbers excluding 0, we write N1. RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 7 1.1. Permutation groups. We consider the permutation group Sn to be the group of per- mutations of the set {1, . . . , n}. This permutation group Sn (1 ≤ n < ∞) is generated by the transpositions of consecutive elements t1 = (1, 2), . . . , tn−1 = (n − 1, n). A complete set of relations for these generators is given by (1) (2) (3) t2 i = e titj = tjti titi+1ti = ti+1titi+1 for all 1 ≤ i ≤ n − 1 whenever j − i ≥ 2 for all 1 ≤ i ≤ n − 2 The permutation group Sn is a Coxeter group with Coxeter system (Sn, {ti}), and most of the results mentioned below are true in the more general context of Coxeter groups. We write σ for the length of an element σ ∈ Sn with respect to the generating set {t1, . . . , tn−2}. Observe that every element σ ∈ Sn is represented by a word ti1 . . . tis of minimal length, and that this word is unique up to relations (2) and (3). The length σ can also be computed as the number of inversions of σ, i.e. the number of pairs 1 ≤ i < j ≤ n such that σ(i) > σ(j). Let π be a partition of {1, ldots, n} into consecutive sets, i.e. π is of the form (4) π =({1, . . . , k1}, {k1 + 1, . . . , k1 + k2}, . . . ,(1 + ki, . . . , s−1Xi=1 ki)) sXi=1 for some k1, . . . , ks ≥ 1 withPi ki = n. Then we can consider the subgroup Sπ of permuta- tions that preserve the partition π, i.e. Sπ = {σ ∈ Sn σ(X) = X ∀X ∈ π} ∼= Sk1 ⊕ . . . ⊕ Sks . Such a subgroup is called a parabolic subgroup of Sn. Every left coset C of Sπ contains a unique element σC of minimal length. So every element σ ∈ C can be written uniquely as a product σ = σCσ0 where σ0 ∈ Sπ. This decomposition satisfies σ = σC + σ0. The set of all element σC is denoted by Vπ. This set Vπ can also be described directly in terms of the partition π: Vπ = {σ ∈ Sn σX is increasing for every X ∈ π} . When π is given explicitly in the form (4), we also write Vk1,...,ks = Vπ. It is convenient to allow ki = 0 for some i. In that case, the corresponding set in π is empty. For σ1 ∈ Sn and σ2 ∈ Sm, we write σ1 × σ2 for the permutation in Sn+m that is given by (σ1 × σ2)(i) =(σ1(i) if i ≤ n σ2(i − n) + n if i > n The identity element in Sn is denoted by idn, or just id when n is clear from the context. We write σk,l for the permutation in Sk,l that satisfies (5) σk,l(i) =(i + l i − k if i ≤ k if i > k Lemma 1.1. With this notation, we get the following relation between sets of the form Vk1,...,ks: 8 STEVEN DEPREZ • For every n, m ∈ N1 and for all k ≥ −n, we get Vn+k,m = min(n,m)Gl=max(−k,0) (Vn−l,l × Vk+l,m−l)(idn−l ×σk+l,l × idm−l). Proof. The first two assertions follow immediately from the fact that every element of σ ∈ Vn,m,k is the unique element of shortest length in the coset σ(Sn × Sm × Sk). The last assertion is most easily proven with the characterisation of Vn,m in terms of the order on {1, . . . , n + m}. The argument is best explained using the following picture: idn−l ×σk+l,l × idm−l σ1 × σ2 m n+k l k+l n m+k The elements σ ∈ Vn+k,m are precisely the permutations that are increasing on each of {1, . . . , n + k} and {n + k + 1, . . . , n + k + m} separately. Set l to be the number of elements of {n + k + 1, . . . , n + k + m} that are mapped into {1, . . . , n}. These elements from necessarily an initial segment {n + k + 1, . . . , n + k + l}, because σ is increasing on {n + k + 1, . . . , n + k + m}. Because σ is a permutation, it maps the elements {n − l + 1, . . . , n + k} into the set {n + 1, . . . , m + n + k}. It is now clear that σ can be uniquely decomposed as σ = (idn−l ×σk+l,l×idm−l)(σ1 ×σ2) where σ1 ∈ Vn−l,l and σ2 ∈ Vk+l,m−l. Conversely, it is easy to see that every such element is increasing on {1, . . . , n+k} and {n+k+1, . . . , n+k+m}. (cid:3) 1.2. Bimodules. Let M be a von Neumann algebra. A left M -module is a Hilbert space H with a normal representation λ : M → B(H). Given a normal tracial state τ on M , we can perform the GNS construction: on M we consider the inner product given by hx, yi = τ (x∗y). The completion of M with respect to this inner product is the GNS construction of M and is written as L2(M, τ ). We also write kxk2 2 = τ (x∗x). The Hilbert space L2(M, τ ) comes with a natural map M ∋ x 7→ x ∈ L2(M, τ ), and this map has dense range. There is a natural which trace we use, we just write L2(M ). Left M -modules over von Neumann algebras are representation of M on L2(M, τ ) that is given by λ(x)y =cxy. If it is clear from the context not very interesting: every left M -module H is isomorphic toLi L2(M, τ )pi where the pi are projections in M and τ is any faithful normal tracial state on M . RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 9 An M -M bimodule is a Hilbert space H with two normal representations λ : M → B(H) and ρ : M op → B(H) such that λ(x) commutes with ρ(y) for all x ∈ M and y ∈ M op. We will write xξy = λ(x)ρ(yop)ξ for all x, y ∈ M . The GNS construction L2(M, τ ) is an M -M -bimodule where ρ is given by ρ(xop)y =cyx. Let M be a von Neumann algebra and fix a trace τ on M . Let H be an M -M -bimodule and ξ ∈ H a vector. Then we can define an unbounded operator l(ξ) : L2(M ) → H by the formula l(ξ)x = ξx. We say that a vector ξ is left-bounded if l(ξ) is a bounded operator. When ξ, η ∈ H are left-bounded vectors, then we can define an operator l(ξ)∗l(η) : L2(M, τ ) → L2(M, τ ). This operator commutes with the right representation of M on L2(M, τ ), so it is of the form λ(x) = l(ξ)∗l(η) for some x ∈ M . We denote this x ∈ M by hξ, ηiM . The map h·, ·iM is called the (right) M -valued inner product on H. It is easy to see that this inner product satisfies hxξy, ηziM = y∗ hξ, x∗ηiM z for all x, y, z ∈ M , and hξ, ηi∗ M = hη, ξiM . Moreover, the norm of ξ is given by kξk2 = τ (hξ, ξiM ). We also consider the right multiplication operator r(ξ) that is given by r(ξ)x = xξ. A vector ξ is said to be right-bounded if r(ξ) is a bounded operator. For right-bounded operators ξ, η ∈ H, we can also consider the operator r(ξ)∗r(η) on L2(M ). This operator now commutes with the left representation of M , so it is of the form ρ(xop)∗ for some x ∈ M . This element x is denoted by Mhξ, ηi. The map Mh·, ·i is called the left M -valued inner product on H. It satisfies the relation Mhxξz, yηi = x Mhξ, ηz∗i y∗. We say that a vector is bi-bounded if it is both left and right-bounded. We denote the space of all bi-bounded vectors in H by ◦ H. The space of all bi-bounded vectors is dense in H. Let H, K be two M -M bimodules. The Connes tensor product H ⊗M K is the M -M bimodule ◦ H ⊗alg ◦ K, we define an inner product by defined as follows. On the algebraic tensor product the formula hξ ⊗ η, ζ ⊗ θi = hη, hξ, ζiM θi = hξ Mhη, θi , ζi = τ ( Mhη, θi∗ hξ, ζi). ◦ The Connes tensor product H ⊗M K is the result of separation and completion of K with respect to this inner product. Observe that when either ξ is left-bounded or η is right- bounded, the elementary tensor ξ ⊗ η ∈ H ⊗M K is well-defined. These elementary tensors satisfy the relation ξx ⊗ η = ξ ⊗ xη for all x ∈ M . ◦ H ⊗alg Lemma 1.2. Let M be a finite von Neumann algebra and let H, K be M -M bimodules • The set of left-bounded vectors in H with the property that hξ, ξiM is a projection in M , densely spans H. • The set of elementary tensors ξ ⊗ η where hξ, ξiM and Mhη, ηi are the same projection in M , densely spans H ⊗M K. Proof. To prove the first point, we will show that every left-bounded vector ξ ∈ H is the sum of two vectors ξ1, ξ2 with the property that hξi, ξiiM is a projection for i = 1, 2. Set x =phξ, ξiM and denote ξn = ξ( n nx+1 increases to p = χ(0,∞)(x). In particular, it follows that xn converges to p in k·k2 and hence is a k·k2-Cauchy sequence. But kξn − ξmk = kxn − xmk2 for all n, m ∈ N. So also the sequence (ξn)n is a Cauchy nx+1 ). Observe that xn :=phξn, ξniM = nx 10 STEVEN DEPREZ sequence, and hence has a limit η. This limit satisfies hη, ηiM = p and moreover, ξ = ηx. It suffices now to observe that x is the linear combination x = 1 2 kxk (u+ + u−) of the two unitaries u± = x kxk ± iq1 − x2 kxk2 . For the second point, we will show that every elementary tensor product ξ ⊗ η where ξ is left-bounded and η is right-bounded, can be written as a linear combination of two elementary tensors ξi ⊗ ηi (i = 1, 2) with the property that hξi, ξiiM and Mhηi, ηii are the same projection (i = 1, 2). As in the proof of the first point, we find vectors ζ ∈ H and θ ∈ K and x ∈ M such that ξ ⊗ η = ζ ⊗ xθ and such that p = hζ, ζiM and q = Mhθ, θi are projections. We can assume that p is the smallest projection in M that satisfies px = x and similarly that q is minimal with the property that xq = x. Write the polar decomposition of x by x = v x. Then we know that vv∗ = p, v∗v = q and x ∈ M is a positive element. We write x = 1 2 kxk (u+ + u−) where u± = x kxk2 as before. Observe that u± commutes with x and hence with its support q. In particular, we get that v± = vu± satisfies v±v∗ ±v± = q. Now we see that the vectors ξ1 = ξ2 = ζ, η1 = v+θ and η2 = v−θ satisfy our condition: ξ ⊗ η = ξ1 ⊗ η1 + ξ2 ⊗ η2. (cid:3) kxk ± ir1 − x2 ± = p and v∗ 2. The relative Gaussian Construction In this section we define the relative Gaussian construction, giving formal definitions for all the concepts. For a more accessible account of the construction, we refer to the introduction. For the rest of this section, fix the following data. • a finite von Neumann algebra M with a trace τ • an M -M bimodule H • a self-adjoint M -M bimodular contraction F : H ⊗M H → H ⊗M H that satisfies the braid relation (F ⊗ 1H )(1H ⊗ F )(F ⊗ 1H ) = (1H ⊗ F )(F ⊗ 1H)(1H ⊗ F ). We denote the n-fold Connes tensor product by H (n) = H ⊗M H . . . ⊗M H. By convention we write H (0) = L2(M ) and H (1) = H. We will write the identity operator of H (n) by 1n. On each H (n), we define M -M bimodular operators(cid:16)F (n) σ (cid:17)σ∈Sn in the following way. We freely use the notations introduced in subsection 1.1. For a transposition t = (i, i + 1) of consecutive numbers, we set If σ = ti1 . . . tik is a decomposition of minimal length, then the M -M bimodular operator F (n) t = −1i−1 ⊗ F ⊗ 1n−i−1. σ = F (n) F (n) ti1 . . . F (n) tik does not depend on the choice of the (minimal) decomposition. When n, m ∈ N, we also write Fn,m = Fσn,m where σn,m is as defined by (5) in subsection 1.1. RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 11 We set D(n) = Xσ∈Sn F (n) σ . By [BS2, theorem 1.1], this operator D(n) is positive definite. By convention, we set D(0) = 1 and D(1) = 1H . We define a new inner product on H (n) by the formula hξ, ηiF = (cid:10)ξ, D(n)η(cid:11). Apply sep- aration/completion to this new inner product and denote the resulting M -M bimodule by H (n) and observe that this is a bounded M -M bimodular operator that has dense range. We denote the identity operator on H (n) F by idn. F . Consider the natural map I (n) : H (n) → H (n) F Lemma 2.1. For every n, m ∈ N, we get the following results • The identity operator from on H (n+m) extends uniquely to a bounded operator In,m : H (n) F ⊗M H (m) F → H (n+m) F . • The operator Fn,m on H (n+m) extends uniquely to a contraction Fn,m : H (n) F ⊗M H (m) F → H (m) F ⊗M H (n). Proof. To prove the first point, observe that D(n+m) = En,m(cid:0)D(n) ⊗ D(m)(cid:1). In particular, n (cid:19)2(cid:16)D(n) ⊗ D(m)(cid:17)2 (cid:16)D(n+m)(cid:17)2 n,mEn,m(cid:16)D(n) ⊗ D(m)(cid:17) ≤(cid:18)n + m =(cid:16)D(n) ⊗ D(m)(cid:17) E∗ . Since the square root is an operator-monotone function, we also get that and hence that D(n+m) ≤(cid:18)n + m ≤(cid:18)n + m n (cid:19)(D(n) ⊗ D(m)) n (cid:19) kξk (n) F ⊗M H (n+m) F H H kξk . (m) F for all ξ ∈ H (n+m). So indeed, the identity operator on H (n+m) extends uniquely to a bounded operator In,m : H (n) ⊗M H (m) → H (n+m). The norm of this operator is less than(cid:0)n+m n (cid:1). As to the last point, remark that σn,m(σ1 × σ2) = (σ2 × σ1)σn,m for all permutations σ1 ∈ Sn and σ2 ∈ Sm. Moreover, counting the number of inversions, it is easy to see that σn,m(σ1 × σ2) = σn,m + σ1 + σ2 = (σ2 × σ1)σn,m . 12 STEVEN DEPREZ It follows that Fn,m(cid:0)D(n) ⊗ D(m)(cid:1) =(cid:0)D(m) ⊗ D(n)(cid:1) Fn,m. In particular, we see that F ∗ commutes with D(n) ⊗ D(m) and hence with the square root(cid:0)D(n) ⊗ D(m)(cid:1) 1 2 . So we see that n,mFn,m kFn,mξk H (m) F ⊗M H (n) F n,m(cid:16)D(m) ⊗ D(n)(cid:17) Fn,mξE n,mFn,m(cid:16)D(n) ⊗ D(m)(cid:17) ξE =Dξ, F ∗ =Dξ, F ∗ =(cid:28)ξ,(cid:16)D(m) ⊗ D(n)(cid:17) 1 ≤Dξ,(cid:16)D(n) ⊗ D(m)(cid:17) ξE 2 F ∗ = kξk . H (n) F ⊗M H (m) F 2 ξ(cid:29) n,mFn,m(cid:16)D(m) ⊗ D(n)(cid:17) 1 for all ξ ∈ H (n+m) F . So Fn,m extends uniquely to a contraction Fn,m : H (n) F ⊗M H (m) F → H (m) F ⊗M H (n) F . Definition 2.2. The relative Fock space of M , H and F is defined to be (cid:3) We denote the algebraic direct sum by F M (H, F ) =Mi algMi M (H, F ) = F 0 H (n) F . H (n) F . Now we can define creation and annihilation operators. Let T : H (n) right M -modular operator. Then we define a right-M -modular operator L0(T ) : F 0 F 0 M (H, F ) by the relation that F → H (m) F be a bounded M (H, F ) → L0(T )η = Im,k(T ⊗ idk)I ∗ n,kη L0(T )η = 0 whenever η ∈ H (n+k) F whenever η ∈ H (k) F with k < n. Each of the operators Im,k is bounded, but as k → ∞, their norm may tend to ∞. This is the case for the classical Gaussian construction. In such a case, the operators L0(T ) need not be bounded. But the operators L0(T ) are closable because they have a densely defined adjoint: L0(T ∗) ⊂ L0(T )∗. Definition 2.3. The creation operator on F M (H, F ) associated to T is the closure L(T ) of the densely defined operator L0(T ). For every right-bounded vector in ξ ∈ H, we get that the formula l(ξ)x = ξx defines a bounded right-M -modular map l(ξ) : L2(M ) → H. The operator L(l(ξ)) is called the creation operator of ξ. We also denote this operator by L(ξ). The operator L(l(ξ)∗) = L(ξ)∗ is the corresponding annihilation operator. RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 13 Proposition 2.4. The creation operators satisfy the following composition rule: for bounded right-M -modular operators S : H (n1) and T : H (n2) , we get that F → H (m2) F F → H (m1) F (6) L(S)L(T ) = min(n1,m2)Xk=0 In particular, the space L Im1,m2−k(S ⊗ idm2−k)(Ik,n1−k ⊗ idm2−k) (idk ⊗Fm2−k,n1−k) (I ∗ k,m2−k ⊗ idn1−k)(T ⊗ idn1−k)In2,n1−k  . TM (H, F ) = span{L(T ) T : H (n) F → H (m) F for some n, m ∈ N} is a ∗-algebra whose commutant is the right action of M on F M (H, F ). Proof. It is clear from the definition of In,m that, for all n, m, l ∈ N, In+m,l(In,m ⊗ idl) = In,m+l(idn ⊗Im,l), since both are the unique extension of the identity operator to a bounded linear map from H (n) F to H (n+m+l) F ⊗M H (m) F ⊗M H (l) F . Translating lemma 1.1 in terms of the operators In,m and Fn,m yields the following relation I ∗ n1,l−n1Im2,l−m2 = min n1,m2Xk=max(0,n1+m2−l) (Ik,n1−k ⊗ Im2−k,l+k−n1−m2) (idk ⊗Fm2−k,n1−k ⊗ idl+k−n1−m2 ) n1−k,l+k−n1−m2) k,m2−k ⊗ I ∗ (I ∗ It is clear that both the left hand side and the right hand side of (6) give 0 when they are evaluated in an η ∈ H (l) F with l < n2 or l < n2 + n1 − m2. Let l ≥ max(n1, m2) and take a vector η ∈ H (l−m2+n2). Then we compute that L(S)L(T )η = Im1,l−n1(S ⊗ idl−n1)I ∗ n1,l−n1Im2,l−m2(T ⊗ idl−m2)I ∗ n2,l−m2η = min n1,m2Xk=max(0,n1+m2−l) Im1,l−n1(S ⊗ idl−n1)(Ik,n1−k ⊗ Im2−k,l+k−n1−m2) (idk ⊗Fm2−k,n1−k ⊗ idl+k−n1−m2) (I ∗ k,m2−k ⊗ I ∗ n1−k,l+k−n1−m2)(T ⊗ idl−m2)I ∗ n2,l−m2η = = as claimed. min n1,m2Xk=max(0,n1+m2−l) L min(n1,m2)Xk=0 Im1+m2−k,l+k−n1−m2(Im1,m2−k ⊗ idl+k−n1−m2) (S ⊗ idl−n1)(Ik,n1−k ⊗ idl−n1) (idk ⊗Fm2−k,n1−k ⊗ idl+k−n1−m2) (I ∗ k,m2−k ⊗ idl−m2)(T ⊗ idl−m2) (I ∗ n2,n1−k ⊗ idl+k−n1−m2)I ∗ n1+n2−k,l+k−m2−n1η Im1,m2−k(S ⊗ idm2−k)(Ikn1 − k ⊗ idm2−k) (idk ⊗Fm2−k,n1−k) (I ∗ k,m2−k ⊗ idn1−k)(T ⊗ idn1−k)In2,n1−k  η. 14 STEVEN DEPREZ It remains to show that the commutant of TM (H, F ) is just the right action of M . It is clear that right multiplication by M commutes with every operator in TM (H, F ). Suppose that T ∈ B(F M (H, F )) satisfies T L(S) ⊂ L(S)T for all bounded right-M -modular operators S : H (n) F be a left-bounded vector. Then we see that F → H (m) F . Let ξ ∈ H (k) T ξ = T L(l(ξ))Ω = L(l(ξ))T Ω, so T is completely determined by its value T Ω. On the other hand, hξ, T Ωi = hΩ, L(l(ξ)∗)T Ωi = hΩ, T L(l(ξ)∗)Ωi = 0 for every ξ ∈ H (k)F with k > 0. Hence, T Ω ∈ L2(M ). Since T is a bounded operator, it is clear that T Ω = x for some x ∈ M , so T is given by right multiplication by x. (cid:3) We write T ′′ M (H, F ) for the bicommutant of TM (H, F ), i.e. all the bounded linear operators that commute with the right action of M . As a von Neumann algebra, this is not a very M (H, F ) ∼= B(ℓ2(N)) ⊗ M , but later on we will consider more interesting object, since T ′′ interesting subalgebras. Consider the orthogonal projection P : F M (H, F ) → L2(M ), and observe that P T P ∗ ∈ B(L2(M )) commutes with the right action of M , for all T ∈ T ′′ M (H, F ). So we can define a map E : T ′′ M (H, F ) → M by the formula E(T ) = P T P ∗. Lemma 2.5. The map E : T ′′ faithful. M (H, F ) → M is a normal conditional expectation, but not Proof. It is clear that E is a normal unital completely positive map. The map E is also M -M bimodular because P intertwines the left actions of M . To show that E is not faithful, take any nonzero bounded right-M linear map T : H (n) F → L2(M ) with n 6= 0. Then we see that L(T ∗)P L(T ) ∈ T ′′ M (H, F ) is a non-zero positive bounded operator with E(L(T ∗)P L(T )) = 0. (cid:3) The relative Gaussian construction will be a special kind of subalgebra of T ′′ M (H, F ) on which E is faithful. In order to define the relative Gaussian construction, we need one more piece of data, namely an anti-unitary operator J : H → H that satisfies the following relations • J is an involution, i.e. J 2 = id1. • J intertwines the left and right representations of M on H, i.e. J(xξy) = y∗J(ξ)x∗ for all x, y ∈ M and ξ ∈ H. • J is compatible with F in the sense that (7) (l(ζ)∗ ⊗ id1)F (Jη ⊗ ξ) = J(l(ξ)∗) ⊗ 1)F (η ⊗ ζ), for all bi-bounded vectors ξ, ζ ∈ H and all vectors η ∈ H. RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 15 Then we want to define the algebraic relative Gaussian construction ΓM (H, F, J) to be the subalgebra of TM (H, F ) that is generated by the left action of M and by elements of the form W (ξ) = L(ξ)+L(Jξ)∗ where ξ is a bi-bounded vector in H. The relative Gaussian construction Γ′′ M (H, F, J) will then be the von Neumann algebra generated by ΓM (H, F, J). Observe that Jξ is a left-bounded vector if and only if ξ is right-bounded. More precisely, l(Jξ)x = Jr(ξ)x∗. In particular, we get that hJξ, JηiM = Mhξ, ηi. We consider the elements of M to be reduced words of length 0 and elements of the form W (ξ) are interpreted as reduced words of length 1. These reduced words satisfy xΩ = x ∈ L2(M ) ⊂ F M (H, F ) and W (ξ)Ω = ξ. We will define longer "reduced words" W (ξ) for vectors ξ ∈ H (n) F by a generalization of the Wick formula. The operator W (ξ) will be the unique element in ΓM (H, F, J) that satisfies W (ξ)Ω = ξ. We extend J to an anti-unitary operator eJ on F M (H, F ). First, define an anti-unitary operator J (n) on H (n) by the relation that J (n)(ξ1 ⊗ . . . ⊗ ξn) = Jξn ⊗ . . . ⊗ Jξ1. Observe that this does indeed define an anti-unitary because by induction we get DJ (n)(ξ1 ⊗ . . . ⊗ ξn), J (n)(η1 ⊗ . . . ⊗ ηn)E =DJ (n−1)(ξ1 ⊗ . . . ⊗ ξn−1), hJξn, JηniM J (n−1)(η1 ⊗ . . . ⊗ ηn−1)E =DJ (n−1)(ξ1 ⊗ . . . ⊗ ξn−1), J (n−1)(η1 ⊗ . . . ⊗ ηn−1 Mhηn, ξni)E = h(η1 ⊗ . . . ⊗ ηn−1) Mhηn, ξni , ξ1 ⊗ . . . ⊗ ξn−1i = hη1 ⊗ . . . ⊗ ηn, ξ1 ⊗ . . . ⊗ ξni By convention, we set J (1) = J and J (0) x =cx∗. The extended operator is now the direct sum of the operators J (n). eJ : F M (H, F ) → F M (H, F ) Lemma 2.6. We get the following relation between J (n) and F (n) σ for σ ∈ Sn: (8) J (n)F (n) σ J (n) = Fγσγ−1 , where γ ∈ Sn is defined by γ(i) = n − i + 1. Proof. Once we prove that J (2) commutes with F , it is clear that J (n)(idk−1 ⊗F idn−k−1)J (n) = idn−k−1 ⊗F ⊗ idk−1 for all 1 ≤ k ≤ n − 1. So (8) holds for the generators of Sn, and hence for all of Sn. The fact that J (2) commutes with F follows from our compatibility relation between F and J: observe that (8) is equivalent to the fact that, for all bi-bounded vectors ξ1, ξ2, η1, η2, we have that hξ1 ⊗ ξ2, F (Jη1 ⊗ η2)i = hF (η1 ⊗ ξ1), η2 ⊗ Jξ2i . 16 STEVEN DEPREZ Now we see that Dξ1 ⊗ ξ2, J (2)F (η1 ⊗ η2)E = hF (η1 ⊗ η2), Jξ2 ⊗ Jξ1i =Dξ1 ⊗ ξ2, F J (2)(η1 ⊗ η2)E = hη2 ⊗ ξ1, F (Jη1 ⊗ Jξ2)i = hF (η2 ⊗ ξ1), Jη1 ⊗ Jξ2i (cid:3) Definition 2.7. Let n, m ∈ N and let ξ ∈ H (n+m) F → H (m) Sn,m(ξ) : H (n) bounded vectors η ∈ H (n). F F by setting Sn,m(ξ)η = J (m)(l(η)∗ ⊗ idm)I ∗ . We define an unbounded operator n,mJ (n+m)ξ for all left- We say that a vector ξ ∈ H (n) F k = 0, . . . , n. The space of all bounded vectors in H (n) F is denoted by ◦ H (n) F . is bounded if Sk,n−k(ξ) is a bounded operator for all integers When ξ ∈ H = H (1) F , then we see that S0,1(ξ) = l(ξ) while S1,0(ξ) = l(Jξ)∗. In particular, a F is bounded if and only if it is bi-bounded, and a vector ξ ∈ L2(M ) = H (0) vector ξ ∈ H = H (1) is bounded if and only if ξ = x for some x ∈ M . Observe that Sn,m(xξy + η) = xSn,m(ξ)y + Sn,m(η). We can also describe Sn,m(ξ) by the relation that F hη, Sn,m(ξ)ζi =DIm,n(cid:16)η ⊗ J (n)ζ(cid:17) , ξE . In the special case where F = 0, we see that L(Sn,m(ξ1 ⊗ . . . ⊗ ξn ⊗ η1 ⊗ . . . ⊗ ηm)) = L(ξ1) . . . L(ξn)L(Jη1)∗ . . . L(Jηm)∗, for all sets of bi-bounded vectors ξ1, . . . , ξn, η1, . . . , ηm ∈ H. Definition 2.8. Let ξ ∈ H (n) F be a bounded vector. Then we define the word of ξ to be (9) W (ξ) = nXk=0 L(Sk,n−k(ξ)) Observe that W (ξ)Ω = ξ. For a bi-bounded vector ξ ∈ H, the formula above reads W (ξ) = L(ξ) + L(Jξ)∗. When F = 0, the formula above reduces to the usual Wick formula W (ξ1 ⊗ . . . ⊗ ξn) = nXk=0 L(ξ1) . . . L(ξk)L(Jξk+1)∗ . . . L(Jξn)∗. For this reason, we still call (9) the Wick formula for the relative Gaussian construction. We define bilinear maps ⊠k : F × ◦ H (k+m) ◦ → H (n+m) H (n+k) ξ ⊠k η = In,m(Sk,n(ξ) ⊗ idm)I ∗ F F k,mη, by the formula RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 17 or equivalently, for all ζ ∈ H (n) F hζ ⊗ θ, ξ ⊠k ηi =DI ∗ and θ ∈ H (m) n,kJ (n+k)ξ ⊗ θ, J (n)θ ⊗ I ∗ k,mηE bounded vectors ξ ∈ F . Observe that J (n+m)(ξ ⊠ η) = J (n+k)η ⊠ J (m+k)ξ for all , η ∈ ◦ H (m+k) . F ◦ H (n+k) F In the case where F = 0, these bilinear maps can also be described by (ξ1 ⊗ ξ2) ⊠k (η1 ⊗ η2) = ξ1 ⊗DJ (k)ξ2, η1EM η2, ◦ H (k) and η ∈ ◦ H (m). ◦ H (n), η1, ξ2 ∈ for all ξ1 ∈ Lemma 2.9. Let ξ ∈ H (n) F be a bounded vector. Then J (n)ξ is still a bounded vector and W (ξ)∗ = W (Sξ). Let ξ ∈ H (n) F vectors, for all k = 0, . . . , min(n, m). Moreover, we get the following product formula be bounded vectors. Then the vectors ξ ⊠k η are still bounded and η ∈ H (m) F W (ξ)W (η) = min(n,m)Xk=0 W (ξ ⊠k η). In particular, the space ΓM (H, F, J) = span{W (ξ) ξ ∈ H (n) F is bounded, n ∈ N} is a ∗-algebra. We call this ∗-algebra the algebraic relative Gaussian construction. Proof. In order to prove this theorem, it will be convenient to use the following variants of Sn,m and ⊠k. Let K, L be M -M bimodules. For a vector ξ ∈ K ⊗M L, we define an operator TL,K(ξ) : L → K by the formula hη, TL,K (ξ)ζi =(cid:10)η ⊗ ζ, ξ(cid:11) , for all η ∈ K and for all left-bounded vectors ζ ∈ L. This is not necessarily a bounded operator, but it is a closable, densely define unbounded operator. Indeed, TL,K is closable because its adjoint is given by TL,K(ξ)∗ ⊃ TK,L(ξ), where we identified K ⊗M L with L ⊗M K. We denote the closure of TL,K still by TL,K. We say that ξ is bounded for the tensor product decomposition K ⊗M L if this operator TL,K(ξ) is bounded. Let K, L1, L2 be M -M bimodules. For vectors ξ ∈ L1 ⊗M K and η ∈ K ⊗M L2 that are bounded in these tensor product decompositions, we write ξ ⊡K η = (TK,L1 ⊗ idL2)η, or equivalently, ξ ⊡K η is the unique vector in L1 ⊗M L2 that satisfies hζ1 ⊗ ζ2, ξ ⊡K ηi =(cid:10)ξ ⊗ ζ2, ζ1 ⊗ η(cid:11) . 18 STEVEN DEPREZ Observe that the anti-unitary operator J : H → H can also be interpreted as an M -M bimodule isomorphism J : H → H. This is also true for the operators J (n) on H (n) F . The original Sn,m and ⊠k can easily be expressed in terms of TL,K and ⊡K. Sn,m(ξ) = T H (n) F ,H (m) F ((idm ⊗J (n))I ∗ m,nξ) for ξ ∈ H (n+m) F ξ ⊠k η = In,m(((id ⊗J (k))I ∗ n,kξ) ⊡ (I ∗ k,mη)) for ξ ∈ H (n+k) F and η ∈ H (k+m) F . Moreover, a vector ξ ∈ H (n) F the tensor product decomposition H (n) is bounded if and only if (idk ⊗J (n−k))E∗ , for all k = 0, . . . , n. F ⊗M H (n−k) F n,n−kξ is bounded for It is now clear that, whenever ξ ∈ H (n)ξ is a bounded vector, then J (n)ξ is still a bounded vector, and Sk,n−k(ξ)∗ = Sn−k,k(J (n)ξ). It follows that W (J (n)ξ) = W (ξ)∗. We prove the product formula in three steps. step 1: For all n, m and for all bi-bounded vectors ξ, θ ∈ H (m) and all vectors η ∈ H (n), we get (10) J (n)(l(ξ)∗ ⊗ idn)Fn,m(η ⊗ ζ) = (l(ζ)∗ ⊗ idn)Fn,m(J (n)η ⊗ ξ). Remark that (10) can be rewritten to the relation that DFn,m(η ⊗ ζ), ξ ⊗ J (n)θE =Dζ ⊗ θ, Fn,m(J (n)η ⊗ ξ)E , for all bi-bounded vectors ξ, ζ ∈ H (m) F . Observe that, for fixed bi-bounded vectors ξ, ζ, the left and right hand sides of the equation are continuous in η, θ, so we only have to check this for dense sets of η, θ ∈ H (n) F . and all vectors η, θ ∈ H (n) F First we observe that the relation (10) is symmetric in n, m: if Fn,m satisfies (10), then we see that, for all bi-bounded vectors ξ, ζ ∈ H (n) F and η, θ ∈ H (m) F DFm,n(η ⊗ ζ), ξ ⊗ J (m)θE =Dη ⊗ ζ, Fn,m(ξ ⊗ J (m)θ)E =DFn,m(J (n)ξ ⊗ η), J (m)θ ⊗ J (n)ζE =Dζ ⊗ θ, Fm,n(J (m)η ⊗ ξ)E , where we used the fact that F ∗ n,m = Fm,n = J (n+m)Fn,mJ (n+m). So also Fm,n satisfies (10). We will prove (10) by induction on m and n. So it suffices to show that Fn+k,m satisfies (10), whenever Fn,m and Fk,m satisfy (10). Note that Fn+k,m = (idm ⊗In,k)(Fn,m ⊗ idk)(idn ⊗Fk,m)(I ∗ n,k ⊗ idm). Let ξ, ζ ∈ H (m) be bi-bounded vectors and let η1, θ2 ∈ H (n), η2, θ1 ∈ H (k) be vectors such that hη1, η1iM = p = Mhη2, η2i and such that hθ1, θ1iM = q = Mhθ2, θ2i for projections p, q ∈ M . F RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 19 Let (χi)i be an orthonormal basis for pH (m) F q. Then we compute that D(Fn,m ⊗ idk)(idn ⊗Fk,m)(η1 ⊗ η2 ⊗ ζ), ξ ⊗ J (n)θ2 ⊗ J (k)θ1E D(Fn,m ⊗ idk)(η1 ⊗ χi ⊗ J (k)θ1), ξ ⊗ J (n)θ2 ⊗ J (k)θ1E =Xi D(idn ⊗Fk,m)(η1 ⊗ η2 ⊗ ζ), η1 ⊗ χi ⊗ J (k)θ1E =Xi DFn,m(η1 ⊗ χi), ξ ⊗ J (n)θ2EDFk,m(η2 ⊗ ζ), χi ⊗ J (k)θ1E =Xi Dζ ⊗ θ1, Fk,m(J (k)η2 ⊗ χi)EDχi ⊗ θ2, Fn,m(J (n)η1 ⊗ ξ)E =Xi DJ (k)η2 ⊗ χi ⊗ θ2, (idk ⊗Fn,m)(J (k)η2 ⊗ J (n)η1 ⊗ ξ)E =Dζ ⊗ θ1 ⊗ θ2, (Fk,m ⊗ idn)(idk ⊗Fn,m)(J (k)η2 ⊗ J (n)η1 ⊗ ξ)E Dζ ⊗ θ1 ⊗ θ2, (Fk,m ⊗ idn)(J (k)η2 ⊗ χi ⊗ θ2)E This implies that Fn+k,m satisfies (10). step 2: Let n, m ∈ N and let L1, L2, K be M -M bimodules. Then for all vectors ξ ∈ L1 ⊗M H (n) F ⊗M L2 that are bounded in all the tensor product decompositions above, we get that F ⊗M K and η ∈ K ⊗M H (m) (11) T L2⊗M H (n) F ,L1⊗M H (m) F ((idL1 ⊗ id H (m) F ⊗J (n) ⊗ idL2 )(idL1 ⊗Fn,m ⊗ idL2 )(ξ ⊡K η)) =(cid:16)T K⊗M H (n) F ,L1(cid:16)(idL1 ⊗J (n) ⊗ idK )ξ(cid:17) ⊗ idm(cid:17) (idK ⊗Fm,n)(cid:16)T L2,K⊗M H (η) ⊗ idn(cid:17) . (m) F The left-hand side in (11) is a closed operator while the right-hand side is bounded, so we only have to check (11) on a dense subset of the domain of the left-hand side. Let ζ1 ∈ H (n) F , ζ2 ∈ L2, θ1 ∈ L1 and θ2 ∈ H (m) be bi-bounded vectors. Then we have to show that Dθ1 ⊗ θ2 ⊗ J (n)ζ2 ⊗ ζ1, (idL1 ⊗Fn,m ⊗ idL2)(ξ ⊡K η)E =DT F (cid:16)(idL1 ⊗J (n) ⊗ idK)ξ(cid:17) θ1 ⊗ θ2, (idK ⊗Fm,n)(cid:16)T L1,K⊗M H (n) L2,K⊗M H (η)ζ1 ⊗ ζ2(cid:17)E . (m) F For fixed bi-bounded vectors ζ1, ζ2, θ1, θ2, both the left hand side and the right hand side of the above expression is continuous in ξ, provided η is bounded for both tensor product decompositions of K ⊗M H (m) F ⊗M L2. The same is true about continuity in the variable ξ. Hence we can assume that ξ = ξ1 ⊗ ξ2 ⊗ ξ2 and η = η1 ⊗ η2 ⊗ η3 where all of ξ1, ξ2, ξ3, η1, η2, η3 are bi-bounded vectors. 20 STEVEN DEPREZ Now the left hand side reduces to Dθ1 ⊗ θ2 ⊗ J (n)ζ2 ⊗ ζ1, (idL1 ⊗Fn,m ⊗ idL2)(ξ ⊡K η)E =Dθ1 ⊗ θ2 ⊗ J (n)ζ2 ⊗ ζ1, (idL1 ⊗Fn,m ⊗ idL2)(ξ1 ⊗ ξ2 ⊗(cid:10)ξ3, η1(cid:11)M η2 ⊗ η3)E =Dhξ1, θ1iM θ2 ⊗ J (n)(ζ2) M(cid:10)ζ1, η3(cid:11) , Fn,m(ξ2 ⊗(cid:10)ξ3, η1(cid:11)M η2)E =Dhξ1, θ1iM θ2 ⊗ J (n)(hη3, ζ1iM ζ2), Fn,m(ξ2 ⊗(cid:10)ξ3, η1(cid:11)M η2)E The right hand side reduces to (n) F L1,K⊗M H ((idL1 ⊗J (n) ⊗ idK )ξ)θ1 ⊗ θ2, (idK ⊗Fm,n)(T DT =Dξ3 ⊗ J (n)ξ2 ⊗ hξ1, θ1iM θ2, (idK ⊗Fm,n)(η1 ⊗ η2 ⊗ hη3, ζ1iM ζ2)E =DJ (n)ξ2 ⊗ hξ1, θ1iM θ2, Fm,n((cid:10)ξ3, η1(cid:11)M η2 ⊗ hη3, ζ1iM ζ2)E =DFn,m(J (n)ξ2 ⊗ hξ1, θ1iM θ2),(cid:10)ξ3, η1(cid:11)M η2 ⊗ hη3, ζ1iM ζ2E L2,K⊗M H (η)ζ1 ⊗ ζ2)E (m) F These expressions are equal by step 1. step 3: An elementary but rather tedious computation using step 2 and proposition 2.4 shows that ξ ⊠k η is indeed a bounded vector for all k, whenever ξ ∈ H (n) are bounded vectors. Moreover we get that F , η ∈ H (m) F W (ξ)W (η) = min(n,m)Xk=0 W (ξ ⊠k η). Definition 2.10. The relative Gaussian construction Γ′′ bra generated by the spectral projections of the operators in ΓM (H, F, J). M (H, F, J) is the von Neumann alge- (cid:3) It follows from the result above that for Γ′′ M (H, F, J). We define a state ϕ on Γ′′ ◦ H (n) F is dense in H (n) F . In particular, Ω is a cyclic vector M (H, F, J) by the relation that ϕ(x) = hΩ, xΩi. Theorem 2.11. The state ϕ defined above is a faithful normal trace on Γ′′ satisfies ϕ = τ ◦ E. faithful. In particular, the conditional expectation E : Γ′′ M (H, F, J) that M (H, F, J) → M is Proof. We know already that Ω is a cyclic vector for Γ′′ M (H, F, J). Moreover, we see that eJW (ξ)eJΩ = eJξ = W (eJξ)Ω. Hence it suffices to show that W (ξ) commutes with eJW (η)eJ . RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 21 For each n ∈ N, we know that the space W (ξ) are bounded when restricted to H (n) is dense in H (n) ◦ H (n) F F . So it suffices to show that F , and moreover, the operators for all bounded vectors ξ ∈ H (n) F . We compute that F and ζ ∈ H (k) F , η ∈ H (m) W (ξ)eJW (η)eJ W (ζ)Ω = eJ W (η)eJW (ξ)W (ζ)Ω, eJW (η)eJ W (ζ)Ω = eJ W (η ⊠i J (k)ζ)Ω = W (ζ ⊠i J (m)η)Ω min(m,k)Xi=0 min(m,k)Xi=1 = W (ζ)W (J (m)η)Ω So we find that W (ξ)eJW (η)eJ W (ζ)Ω = W (ξ)W (ζ)W (J (m)η)Ω = = W (ξ ⊠i ζ)W (η)Ω min(n,k)Xi=0 min(n,k)Xi=0 = eJW (η)eJ W (ξ)W (ζ)Ω. eJ W (η)eJW (ξ ⊠i ζ)Ω It follows that W (ξ) commutes with eJ W (η)eJ, and hence that ϕ is a faithful normal trace. (cid:3) 3. Proof of the Main Result This section is devoted to the proof of the following theorem. The proof resembles very closely the proof of the main result of [HM], but our proof is a little less technical because the radial structure of relative Gaussian constructions is easier to handle. Theorem 3.1 (see theorem 0.1). Let M be a von Neumann algebra and let H be a Hilbert M -M bimodule. Assume that F is a projection onto an M -M subbimodule of H ⊗M H such that F ⊗ 1 commutes with 1 ⊗ F on H ⊗M ⊗M H. Let ψ : N → C be a function in class C defined above. Then there is a unique ultraweakly continuous, completely bounded map Φψ : T ′′ M (H, F ) that satisfies M (H, F ) → BM T ′′ Φψ(L(T )) = ψ(n + m)L(T ) for all bounded right-M -linear operators T : H (n) norm is less than kΦψkcb ≤ kψkC . F → H (m) F . Moreover, the completely bounded 22 STEVEN DEPREZ We prove the theorem by a series of lemmas. For this section, fix a tracial von Neumann algebra (M, τ ), an M -M bimodule H and an M -M bimodular projection F : H ⊗H → H ⊗H that satisfies the relation (F ⊗ id)(id ⊗F ) = (id ⊗F )(F ⊗ id). The first lemma provides a way to estimate the completely bounded norm of an operator of the form Φ(T ) =Pi uiT vi. It was proven by Christensen and Sinclair in [CS]. the formula Φ(T ) = Pi uiT v∗ Lemma 3.2 (see [CS, Corollary 6.2]). Let H, K be Hilbert spaces and (ui), (vi) sequences i vi are bounded operators, then of bounded operators from H to K. i defines a completely bounded map Φ : B(H) → B(K). The completely bounded norm of Φ is bounded by kΦk2 If Pi uiu∗ i and Pi v∗ cb ≤ kPi uiu∗ i k kPi viv∗ i k. Whenever x ∈ ℓ∞(N), we define a radial multiplication operator Mx on F M (H, F ) by the relation that Mxη = x(n)η whenever η ∈ H (n) F . On ℓ∞(N), we consider the one-directional shift S, which is defined by S(x)n = xn−1 when n > 0 and S(x)0 = 0. We denote the shift in the other direction by S∗. Remark 3.3. It is clear that the radial multiplication operators are M -M bimodular and they satisfy the following relations with the creation operators. MxL(T )My = Mx((S ∗)nSmy)L(T ) = L(T )My(Sn(S ∗)mx) for all bounded right-M modular operators T : H (n) F → H (m) F . We consider sequences z, rk ∈ ℓ∞(N) that are defined by zn = (−1)n and rk(n) = 1 whenever n ≥ k and rk(n) = 0 otherwise. Observe that u = Mz is a unitary and that the qk = Mrk are projections. As a left M -module, we can write H in the form M H ∼=Li M L2(M, ϕ)pi. We write ξi for the vector corresponding to 1pi in the i-th component of the direct sum above. Observe that Mhξi, ξji = δi,jpi. Lemma 3.4. Define a completely positive map ρ : B(F M,ϕ(H, F )) → B(F M,ϕ(H, F )) by the formula ρ(T ) =Pi R(ξi)T R(ξi)∗. This operator satisfies ρl(L(T )) = L(Im,l(T ⊗ idl)I ∗ n,l) = L(T )qn+l, for all bounded right-M modular operators T : H (n) subunital completely positive map, because 1 = L(id0) and q1 = L(id1) is a projection. F → H (m) and l ≥ 0. In particular, ρ is a F F → H (m) Proof. We first prove lemma 3.4 with l = 1. Let T : H (n) F modular operator. It is clear that ρ(L(T ))η = 0 = L(Im,1(T ⊗ id)I ∗ be a bounded right-M - n,1)η whenever η ∈ H k for RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 23 some k < n + 1. Let η ∈ H n+k+1 for some k ≥ 0. We compute that Im+k,1(idm+k ⊗r(ξi))Im,k(T ⊗ id1)I ∗ n,k(idn+k ⊗r(ξi)∗)I ∗ n+k,2η ρ(L(T ))η =Xi Xi Im,k,1(T ⊗ idk ⊗r(ξi)r(ξi)∗)I ∗ n,k,1η = Im,k,1(T ⊗ idk ⊗ id1)I ∗ k,1))I ∗ = Im,k+1(T ⊗ (Ik,1I ∗ and L(T ⊗ id)η = Im,1,k(T ⊗ id1 ⊗ idk)I ∗ 1,k))I ∗ m,k+1. and L(T )qn+1η = Im,k+1(T ⊗ idk+1)I ∗ = Im,k+1(T ⊗ (I1,kI ∗ n,1,k n,k,1 n,k+1 n,k+1 When F is a projection such that F ⊗ id1 commutes with id1 ⊗F , then we get that D(n) : H (n) → H (n) is simply the projection onto the orthogonal complement of the closed linear span of the subspaces of the form H (i−1) ⊗M F (H (2)) ⊗M H (n−i−1) for i = 1, . . . , n − 1. So H (n) is a closed subspace of H (n). The operator In,m is simply the projection from H (n) n,m = idn+m, so it follows that L(T ⊗ id)η = L(T )qn+1η. This is true for all η ∈ H k, for all k ∈ N, so ρ(L(T )) = L(T ⊗ id) = L(T )qn+1. onto its closed subspace H (n+m) . Hence we see that In,mI ∗ F ⊗M H (m) F F F Now let l > 1. By induction we get that ρl(L(T )) = L(Im,1,...,1(T ⊗ id1 ⊗ . . . ⊗ id1)In,1,...,1) = L(Im,l(T ⊗ I1,...,1I ∗ 1,...,1)I ∗ n,l. By the same argument as above, we see that 1,...,1I ∗ 1,...,1 = idl, so indeed ρl(L(T )) = L(Im,l(T ⊗ idl)I ∗ n,l) = L(T )qn+l. (cid:3) Lemma 3.5. Let x, y ∈ ℓ2(N), and define an M -M -bimodular map Φx,y : T ′′ T ′′ M (H, F ) by the formula M (H, F ) → Φx,y(T ) =Xn≥0 (S ∗)ny +Xn≥1 M(S ∗)nxT M ∗ MSnxρn(T )M ∗ Sny. Then this operator is completely bounded with completely bounded norm kΦx,ykcb ≤ kxk2 kyk2 that satisfies (12) Φx,y(L(T )) = h(S∗)mx, (S∗)nyi L(T ). lemma 3.2, we have that Pi uiu∗ Proof. Observe that Φx,y is given in the form of lemma 3.2, and that in the notation of i = Φy,y(1). So, once we show (12), it follows from lemma 3.2 that Φx,y is completely bounded and its norm is bounded by kΦx,ykcb ≤ kxk2 kyk2. i = Φx,x(1) and Pi viv∗ 24 STEVEN DEPREZ Let T : H (n) F → H (m) F be a bounded right-M -linear operator. Then we see that Φx,y(L(T )) =Xk≥0 (S ∗)k y +Xk≥1 M(S ∗)k xL(T )M ∗ MSkxρk(L(T ))M ∗ Sk y = L(T )Xk≥0 = L(T )Mf , MSn(S ∗)k+mxM ∗ (S ∗)k y +Xk≥1 MSn+k(S ∗)m+kSkxM ∗ Skyqn+k where the function f is given by f (l) = 0 x(l + m + k)y(l + n + k) + if l < n x(l − k + m)y(l + n − k) if l ≥ 0 lXk=1 f (n + l) =Xk≥0 =Xk≥l x(k + m)y(k + n) + x(k + m)y(k + n) l−1Xk=0 = h(S∗)ny, (S∗)mxi . Since L(T ) = L(T )qn, we see that indeed Φx,y(L(T )) = L(T ) h(S∗)ny, (S∗)mxi . (cid:3) proof of theorem 3.1. It is well-known that every trace class operator T ∈ B(ℓ2 N) can be written as a sum of rank one operators T =Pn xn ⊗ yn with xn, yn ∈ ℓ2(N) and where the sum converges in trace norm. Moreover, the trace norm of T isPn kxnk2 kynk2. In particular, for the Hankel matrix Hψ, we find sequences of vectors xn, yn ∈ ℓ2(N) such that ψ(k + l) − ψ(k + l + 2) = Hψ(k, l) =Xn xn(k)yn(l) for all k, l ∈ N, and such that kHψk1 =Pn kxnk2 kynk2. Moreover, we see that (ψ(k + 2m) − ψ(k + 2m + 2)). So we also see that ψ(k) = c+ + (−1)kc− +Xm ψ(k + l) = c+ + (−1)k+lc− +Xm,n xn(k + m)yn(l + m) = c+ + (−1)k+lc− +Xn D(S∗)lyn, (S∗)kxE RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 25 Now, we set Φψ = c+ + c− Adu +Pk Φxk,yk , where u = Mz with z(k) = (−1)k as before. Let be a bounded right-M -linear operator. Then it follows from lemma 3.5 that F T : H (n) F → H (m) Φψ(L(T )) = c+ + c−L(T )MSm(S ∗)nzMz +Xk = c+ + c−(−1)n+m +Xn = ψ(n + m)L(T ) Φxk,yk (L(T )) h(S∗)my, (S∗)nxi! L(T ) So Φψ is indeed the map we were searching for. Moreover Φψ is completely bounded and its completely bounded norm is bounded by kΦψkcb ≤ c+ + c− +Xk kxkk2 kykk2 = kψkC . (cid:3) 4. Amalgamated free product von Neumann algebras In this section we deduce the following theorem from theorem 3.1. Theorem 4.1. Let (Mi)i be a (finite or countably infinite) family of tracial von Neumann algebras with a common subalgebra P . Denote by M the amalgamated free product over P of the von Neumann algebras Mi. Let ψ : N → C be a function in class C′ defined in the introduction. Then there is a unique ultraweakly continuous completely bounded map Ψψ : M → M that satisfies Ψψ(x1 . . . xn) = ψ(n)x1 . . . xn for every reduced word x1 . . . xn ∈ M. Moreover, the completely bounded norm of Ψψ is bounded above by kΨψkcb ≤ kψkC ′. Proof. Consider the graph of von Neumann algebras that is depicted below M (1) M (2) P P P P M (i) 26 STEVEN DEPREZ As in the introduction, we can consider the fundamental von Neumann algebra fM of this graph. In other words, we set fM = Γ′′ (H, F, J) where eP Mi eP = P ⊕Mi H =Mi J(x) =cx∗ ∈ P L2(Mi)Mi Mi L2(Mi)P ⊕ P L2(Mi)Mi L =Mi ⊂Mi P P P ⊕Mi P ⊕Mi,j P M (i) = H (2). for all x ∈ Mi L2(Mi)P and vice-versa. and where F is the projection onto the eP -eP Hilbert subbimodule M (j)M (i) ⊗P M (i) M (i) M (i)M (i) ⊗P M (j) M (j) Denote by p the central projection in eP ⊂ fM that corresponds to the term P in the direct sum eP = P ⊕Li M (i). Then it follows that (13) p F eP (H, F )p ∼= P L2(P )P ⊕Mi P L2(Mi ⊖P )P ⊕Mi6=j P L2(Mi ⊖P )⊗P L2(Mj ⊖P )P ⊕. . . , a priori just as P -P bimodules. Observe that each term L2(Mi) that appears in the direct sum above is a direct summand of H (2) F . The P -P bimodule on the right is precisely L2(M ). We denote by U : p F eP ,ϕ(H, F )p → L2(M ) the unitary that implements the identification in (13). It is now easy to see that F rather than of H (1) for every y ∈ Mi ⊖ P . upW (y)pu∗ = y ∈ M, In particular, we get that α = Adu∗ : M → pfM p is an isomorphism that maps a reduced word and ψ(2n + 1) = 0 for all n ∈ N. Then we know that ψ ∈ C and its norm is(cid:13)(cid:13)(cid:13) ψ(cid:13)(cid:13)(cid:13)C x1 . . . xn of length n to the reduced word pW (x1⊗. . .⊗xn)p that has length 2n. Let ψ : N → C be a function in class C′. We consider the function ψ : N → C that is defined by ψ(2n) = ψ(n) = kψkC ′. M (H, F )) that By theorem 3.1, we find an ultraweakly continuous map Φ ψ : T ′′ satisfies M (H, F ) → T ′′ Φ ψ(L(T )) = ψ(n + m)L(T ) for all bounded right-M modular operators T : H (n) F → H (m) F . Moreover, the completely bounded norm of Φ ψ is bounded by(cid:13)(cid:13)(cid:13)Φ ψ(cid:13)(cid:13)(cid:13)cb ≤ kψkC. RADIAL MULTIPLIERS ON VON NEUMANN ALGEBRAS 27 In particular, we get that Φ ψ(W (ξ)) = ψ(n)W (ξ) for all bounded vectors ξ ∈ H (n) F . Moreover, Φ ψ is M -M -bimodular, and in particular, Φ ψ(pfM p) ⊂ pfM p. We define Φψ : M → M by the formula Φψ(x) = α−1(Φ ψ(α(x))). We conclude that Φψ(x1 . . . xn) = ψ(2n)x1 . . . xn = ψ(n)x1 . . . xn for all reduced words x1 . . . xn in the amalgamated free product decomposition M = M1 ∗P M2 ∗P . . .. Moreover, the completely bounded norm of Φψ is bounded by kΦψkcb ≤(cid:13)(cid:13)(cid:13)Φ ψ(cid:13)(cid:13)(cid:13)cb ≤(cid:13)(cid:13)(cid:13) ψ(cid:13)(cid:13)(cid:13)C = kψkC ′ . (cid:3) So Φψ satisfies the conditions of theorem 3.1 References [BS1] M. Bozejko and R. Speicher. An example of a generalized Brownian motion. Comm. Math. Phys., 137(3):519 -- 531, 1991. [BS2] M. Bozejko and R. Speicher. Completely positive maps on Coxeter groups, deformed commutation relations, and operator spaces. Math. Ann., 300(1):97 -- 120, 1994. [CS] E. Christensen and A. M. Sinclair. A survey of completely bounded operators. Bull. London Math. Soc., 21(5):417 -- 448, 1989. [HM] U. Haagerup and S. Moller. Radial multipliers on reduced free products of operator algebras. J. Funct. Anal., 263(8):2507 -- 2528, 2012. [HSS] U. Haagerup, T. Steenstrup, and R. Szwarc. Schur multipliers and spherical functions on homogeneous [M] [S] [V] [W] trees. Internat. J. Math., 21(10):1337 -- 1382, 2010. S. Moller. Radial multipliers on amalgamated free products of II1 factors. preprint. ArXiV:1306.5540. D. Shlyakhtenko. A-valued semicircular systems. J. Funct. Anal., 166(1):1 -- 47, 1999. D. Voiculescu. Symmetries of some reduced free product C ∗-algebras. In Operator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983), volume 1132 of Lecture Notes in Math., pages 556 -- 588. Springer, Berlin, 1985. J. Wysocza´nski. A characterization of radial Herz-Schur multipliers on free products of discrete groups. J. Funct. Anal., 129(2):268 -- 292, 1995.
1306.3411	3	1306	2015-03-26T13:01:19	Bilinear Ideals in Operator Spaces	[ "math.OA", "math.FA" ]	We introduce a concept of bilinear ideal of jointly completely bounded mappings between operator spaces. In particular, we study the bilinear ideals $\mathcal{N}$ of completely nuclear, $\mathcal{I }$ of completely integral, $\mathcal{E}$ of completely extendible bilinear mappings, $\mathcal{MB}$ multiplicatively bounded and its symmetrization $\mathcal{SMB}$. We prove some basic properties of them, one of which is the fact that $\mathcal{I}$ is naturally identified with the ideal of (linear) completely integral mappings on the injective operator space tensor product.	math.OA	math	BILINEAR IDEALS IN OPERATOR SPACES VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Abstract. We introduce a concept of bilinear ideal of jointly completely bounded mappings between operator spaces. In particular, we study the bilinear ideals N of completely nuclear, I of completely integral, E of completely extendible bilinear mappings, MB multiplicatively bounded and its symmetrization SMB. We prove some basic properties of them, one of which is the fact that I is naturally identified with the ideal of (linear) completely integral mappings on the injective operator space tensor product. 1. Introduction and Preliminaries Let V, W and X be operator spaces. If we consider the underlying vector space structure, the relations (1) Bil(V × W, X) ν ≃ L(V ⊗ W, X) ρ ≃ L(V, L(W, X)) hold through the two natural linear isomorphisms ν, ρ. In order for ν and ρ to induce natural morphisms in the operator space category, it is necessary to have appropriately defined an oper- ator space tensor norm on V ⊗ W and specific classes of linear and bilinear mappings. This is the case, for instance, of the so called projective operator space tensor norm k · k∧, the completely bounded maps and the jointly completely bounded bilinear mappings, where ν and ρ induce the following completely bounded isometric isomorphisms: J CB(V × W, X) ≃ CB(Vb⊗W, X) ≃ CB(V, CB(W, X)). There are many possible ways to provide V ⊗ W with an operator space tensor norm and, of course, to define classes of mappings. Several authors, inspired by the success that the study of the relations between tensor products and mappings has had in the Banach space setting, have systematically study some analogous relations for operator spaces. This is the case, for instance, of the completely nuclear and completely integral linear mappings (see [7, Section III]). In this paper we follow this approach as well, but with the attention focused on the relations involving ν, the isomorphism in (1) which concerns bilinear mappings. In Section 2 we introduce the notion of an ideal of completely bounded bilinear mappings and study its general properties. In Section 3 we define the ideals of completely nuclear and completely integral bilinear mappings. The main result proved here is that the ideal of completely integral bilinear mappings is naturally identified with the ideal of completely integral linear mappings on the injective operator space tensor product, that is I(V × W, X) ∼= LI(V ∨ ⊗ W, X) (see Theorem 3.8). This implies that, 2010 Mathematics Subject Classification. 47L25,47L22, 46M05. Key words and phrases. Operator spaces, Bilinear mappings, Bilinear ideals. The first author was partially supported by CONICET PIP 0624 and PAI-UDESA 2011. The second author was partially supported by CONACYT 182296. 1 2 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA contrary to the result for Banach spaces, the relation I(V × W ) ∼= LI(V, W ∗) does not always hold. Indeed, it holds if and only if W is locally reflexive. η (cid:18)V ⊗ W(cid:19)∗ The ideal E of bilinear completely extendible mappings is introduced in Section 4. We prove in Proposition 4.4 that E gives rise, through duality, to an operator space tensor product η such that ∼= E(V ×W ). In Section 5 we consider the ideal SMB of symmetrized multiplicatively bounded mappings, which is the symmetrization of the ideal MB of multiplicatively bounded mappings. The following theorem summarizes the inclusion relations among all these bilinear ideals: Theorem 1.1. Let V, W and X be operator spaces. Then, we have the following complete contractive inclusions: (a) N (V × W, X) ⊂ I(V × W, X) ⊂ MB(V × W, X) ⊂ SMB(V × W, X) ⊂ J CB(V × W, X). (b) I(V × W, X) ⊂ E(V × W, X) ⊂ J CB(V × W, X). (c) MB(V × W, L(H)) ⊂ SMB(V × W, L(H)) ⊂ E(V × W, L(H))) ⊂ J CB(V × W, L(H)). In Section 6 we prove the inclusions and provide examples to distinguish the ideals. We now recall some basic concepts about operator spaces, mainly with respect to bilinear operators and tensor products. For a more complete presentation of these topics, see [2, 7, 14]. All vector spaces considered are over the complex numbers. For a linear space V , we let Mn×m(V ) denote the set of all the n × m matrices of elements in V . In the case n = m, the notation is simplified to set Mn×n(V ) = Mn(V ). If V is the scalar field we just write Mn×m and Mn, respectively. For α ∈ Mn×m, its norm kαk will be considered as an operator from ℓm 2 to ℓn 2 . Given v = (vi,j) ∈ Mn(V ) and w = (wk,l) ∈ Mm(V ), v ⊕ w ∈ Mn+m(V ) stands for the matrix v ⊕ w = (vi,j) 0 0 (wk,l) ! . A matrix norm k · k on a linear space V is an assignment of a norm k · kn on Mn(V ), for each n ∈ N. A linear space V is an operator space if it is endowed with a matrix norm satisfying: M1 kv ⊕ wkn+m = max{kvkn, kwkm}, for all v ∈ Mn(V ) and w ∈ Mm(V ). M2 kαvβkm ≤ kαk · kvkn · kβk, for all v ∈ Mn(V ), α ∈ Mm×n and β ∈ Mn×m. We usually omit the subindex n in the matrix norms and simply denote k · k instead of k · kn. The inclusion Mn×m(V ) ֒→ Mmax{n,m}(V ) naturally endows the rectangular matrices with a norm. Throughout the article, V , W , X, Y , Z, U1, U2 will denote operator spaces where the underlying normed space is complete (i.e. it is a Banach space). Every linear mapping ϕ : V → W induces, for each n ∈ N, a linear mapping ϕn : Mn(V ) → Mn(W ) given by ϕn(v) = (ϕ(vi,j)) , for all v = (vi,j) ∈ Mn(V ). It holds that kϕk = kϕ1k ≤ kϕ2k ≤ kϕ3k ≤ .... The completely bounded norm of ϕ is defined by kϕkcb = sup n∈N kϕnk. BILINEAR IDEALS IN OPERATOR SPACES 3 We say that ϕ is completely bounded if kϕkcb is finite, that ϕ is completely contractive if kϕkcb ≤ 1 and that ϕ is a complete isometry if each ϕn : Mn(V ) → Mn(W ) is an isometry. It is easy to see that k·kcb defines a norm on the space CB(V, W ) of all completely bounded linear mappings from V to W . The natural identification Mn (CB(V, W )) ∼= CB (V, Mn(W )) provides CB(V, W ) with the structure of an operator space. Also, since V ∗ = CB(V, C), the dual of an operator space is again an operator space. In contrast to the linear case, a bilinear mapping φ : V × W → X naturally induces not one, but two different bilinear mappings in the matrix levels. Some authors (see, for instance [7, 18]) use the name "complete boundedness" for the first notion and "multiplicative boundedness" or "matrix complete boundedness" for the second one, while others [2, 3, 20] use the name "jointly complete boundedness" for the first concept and "complete boundedness" for the second one. In order to avoid confusion, we will not use the name "complete boundedness" for bilinear mappings. So, given a bilinear mapping φ : V × W → X, consider the associated bilinear mapping φn : Mn(V ) × Mn(W ) → Mn2(X) defined, for each n ∈ N, as follows: φn(v, w) = (φ(vi,j, wk,l)) , for all v = (vi,j) ∈ Mn(V ), w = (wk,l) ∈ Mn(W ). When their norms are uniformly bounded, that is, when kφkjcb ≡ sup n∈N kφnk < ∞, we say that φ is jointly completely bounded. It is plain to see that k · kjcb is a norm on the space J CB(V × W, X) of all jointly completely bounded bilinear mappings from V × W to X. As in the linear setting, the identification Mn (J CB(V × W, X)) ∼= J CB (V × W, Mn(X)) . provides J CB(V × W, X) with an operator space structure. The second way to naturally associate φ with a bilinear mapping φ(n) : Mn(V ) × Mn(W ) → Mn(X), for each n ∈ N, involves the matrix product and it is given by φ(n)(v, w) = nXk=1 φ(vi,k, wk,l)! , for all v = (vi,j) ∈ Mn(V ), w = (wk,l) ∈ Mn(W ). We say that φ is multiplicatively bounded if kφkmb = sup n∈N kφ(n)k < ∞. Again, it is easily seen that k·kmb is a norm on the space MB(V ×W, X) of all multiplicatively bounded bilinear mappings from V × W to X. The identification Mn (MB(V × W, X)) ∼= MB (V × W, Mn(X)) endows MB(V × W, X) with matrix norms that give the structure of an operator space. We finish this section recalling three basic examples from the theory of tensor products of operator spaces (the general notion is in Definition 2.3): the operator space projective tensor norm, the operator space injective tensor norm and the operator space Haagerup tensor norm. 4 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Consider two operator spaces V and W . The definition of the first norm uses the fact that each element u ∈ Mn(V ⊗ W ) can be written as: (2) u = α(v ⊗ w)β with v ∈ Mp(V ), w ∈ Mq(W ), α ∈ Mn×p·q, β ∈ Mp·q×n, for certain p, q ∈ N, where v ⊗ w is the p · q × p · q-matrix given by (3) v ⊗ w =  v1,1 ⊗ w1,1 ... v1,1 ⊗ wq,1 · · · · · · vp,1 ⊗ w1,1 ... vp,1 ⊗ wq,1 · · · ... · · · · · · · · · · · · ... · · · v1,1 ⊗ w1,q ... v1,1 ⊗ wq,q · · · · · · vp,1 ⊗ w1,q ... vp,1 ⊗ wq,q · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · · v1,p ⊗ w1,1 ... v1,p ⊗ wq,1 · · · · · · vp,p ⊗ w1,1 ... vp,p ⊗ wq,1 · · · ... · · · · · · · · · · · · ... · · · v1,p ⊗ w1,q ... v1,p ⊗ wq,q · · · · · · vp,p ⊗ w1,q ... vp,p ⊗ wq,q  The Haagerup tensor norm is defined as: kukh = inf {kvk · kwk : u = v ⊙ w, v ∈ Mn×r(V ), w ∈ Mr×n(W ), r ∈ N} , while the Haagerup tensor product V h ⊗ W is the completion of (V ⊗ W, k · kh). For any operator spaces V and W , k · k∨ and k · k∧ are, respectively, the smallest and the largest operator space cross norms on V ⊗ W . In particular, for each u ∈ Mn(V ⊗ W ) it holds that kuk∨ ≤ kukh ≤ kuk∧. The operator space projective tensor norm of u ∈ Mn(V ⊗ W ) is defined as kuk∧ = inf{kαk · kvk · kwk · kβk : all representations of u as in (2)}. The operator space injective tensor norm of u ∈ Mn(V ⊗ W ) is defined as kuk∨ = sup {k(f ⊗ g)n(u)k : f ∈ Mp(V ∗), g ∈ Mq(W ∗), kf k ≤ 1, kgk ≤ 1} . The operator space projective tensor product Vb⊗W and the operator space in- ∨ ⊗ W are the completion of (V ⊗ W, k · k∧) and the completion jective tensor product V of (V ⊗ W, k · k∨), respectively. There is a natural completely isometric identification: So, in particular: J CB(V × W, X) ∼= CB(Vb⊗W, X) ∼= CB(V, CB(W, X)). J CB(V × W ) ∼= (Vb⊗W )∗ ∼= CB(V, W ∗). ∨ ⊗ W )∗ with a subset of bilinear mappings is done later, in Proposition Every u ∈ Mn(V ⊗ W ) can be written as u = v ⊙ w, for certain matrices v ∈ Mn×r(V ) and The identification of (V 3.11. w ∈ Mr×n(W ), where v ⊙ w = rXk=1 vi,k ⊗ wk,j! . BILINEAR IDEALS IN OPERATOR SPACES 5 The Haagerup tensor product is naturally associated with multiplicatively bounded bilinear operators through the following identifcations: MB(V × W, X) ∼= CB(V h ⊗ W, X) and MB(V × W ) ∼= (V h ⊗ W )∗. Remark 1.2. We will use repeatedly along the text the following extension property for completely bounded linear mappings (see [7, Theorem 4.1.5]): if V is a subspace of an operator space W and H is a Hilbert space, then every completely bounded linear map ϕ : V → L(H) has a completely bounded extension ϕ : W → L(H) with kϕkcb = kϕkcb. Equivalently, this can be stated as in [14, Theorem 1.6]: if V , W are operator spaces, H, K are Hilbert spaces such that V is a subspace of L(H) and W is a subspace of L(K), then every completely bounded linear map ϕ : V → W has a completely bounded extension ϕ : L(H) → L(K) with kϕkcb = kϕkcb. 2. Bilinear ideals The linear structure and the closedness by compositions are the basic properties required of a subset of maps, in order to have a suitable relation between mappings spaces and tensor products. These will be, precisely, the defining properties of a bilinear ideal (see Definition 2.2). To deal with compositions, we need first to prove the following estimate: Lemma 2.1. Let φ ∈ Mn (J CB(V × W, X)), r1 ∈ CB(U1, V ), r2 ∈ CB(U2, W ), s ∈ CB(X, Y ). Then sn ◦ φ ◦ (r1, r2) is jointly completely bounded and ksn ◦ φ ◦ (r1, r2)kjcb ≤ kskcb · kφkjcb · kr1kcb · kr2kcb. Proof. Let ψ = sn ◦ φ ◦ (r1, r2). It is easy to see that ψm = sn·m2 ◦ φm ◦ ((r1)m, (r2)m). Thus, for every m, kψmk ≤ ksn·m2k · kφmk · k(r1)mk · k(r2)mk ≤ kskcb · kφkjcb · kr1kcb · kr2kcb, and the conclusion follows. (cid:3) In accordance with the definition of an operator space ideal of linear mappings (see [6] and [7]), we introduce: Definition 2.2. An operator space bilinear ideal A is an assignment, to each group of three operator spaces V , W and X, of a linear subspace A(V × W, X) of J CB(V × W, X) containing all finite type continuous bilinear maps, together with an operator space matrix norm k · kA such that: (a) For all φ ∈ Mn(A(V × W, X)), kφkjcb ≤ kφkA. (b) For all φ ∈ Mn(A(V × W, X)), r1 ∈ CB(U1, V ), r2 ∈ CB(U2, V ), s ∈ CB(X, Y ), the matrix sn ◦ φ ◦ (r1, r2) belongs to Mn(A(U1 × U2, Y )) and ksn ◦ φ ◦ (r1, r2)kA ≤ kskcb · kφkA · kr1kcb · kr2kcb. We now introduce the notion of tensor norm for operator spaces. 6 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Definition 2.3. We say that α is an operator space tensor norm if α is an operator space matrix norm on each tensor product of operator spaces V ⊗ W that satisfies the following two conditions: (a) α is a cross matrix norm, that is, α(v ⊗ w) = kvk · kwk, for all v ∈ Mp(V ), w ∈ Mq(W ), p, q ∈ N. (b) α fullfils the "completely metric mapping property": for every r1 ∈ CB(U1, V ), r2 ∈ CB(U2, W ), the operator r1 ⊗ r2 : (U1 ⊗ U2, α) → (V ⊗ W, α) is completely bounded and kr1 ⊗ r2kcb ≤ kr1kcb · kr2kcb. α ⊗ W the completion of (V ⊗ W, α). We denote by V This notion is, in principle, less restrictive than the one introduced in [3, Definition 5.9], which the authors called "uniform operator space tensor norm". Whenever the linear isomorphism determined by (3) (the so called algebraic shuffle isomorphism) Mp(V ) ⊗ Mq(W ) → Mpq(V ⊗ W ) extends to a complete contraction Mp(V )⊗α Mq(W ) → Mpq(V ⊗α W ), both notions coincide [20]. That is the case of the three tensor norms defined above (projective, injective and Haagerup). The proof that these main examples satisfy the definition, as well as the fact that the projective tensor norm k · k∧ is the largest operator space tensor norm, can be found in [7]. Every operator space tensor norm determines, through ν in (1), an operator space bilinear ideal according to the following identification: Given V , W , X operator spaces, let Aα(V × W, X) ∼= CB(V α ⊗ W, X). Proposition 2.4. Let α be an operator space tensor norm. Then Aα is an operator space bilinear ideal. α CB(V Proof. From the relation CB(V subspace of J CB(V × W, X). Also, it is clear that all finite type continuous bilinear mappings belong to Aα(V × W, X). ⊗ W, X) ⊂ CB(Vb⊗W, X), it follows that Aα(V × W, X) is a ⊗ W, X)(cid:17) ∼= α α α ⊗ W ) → Mm(Mn(X)) has norm On the other hand, φ also belongs to Mn(J CB(V × W, X)) and it has an associated matrix of keφmk = supneφm(u) : u ∈ Mm(V ⊗ W ), α(u) ≤ 1o . (a) Let φ ∈ Mn(Aα(V × W, X)) then its linear associated eφ belongs to Mn(cid:16)CB(V ⊗ W, Mn(X)). This says that kφkAα = keφkcb = supm keφmk. The mapping eφm : Mm(V linear mappings φ ∈ Mn(cid:0)CB(Vb⊗W, X)(cid:1) ∼= CB(Vb⊗W, Mn(X)). This implies that and the mapping φm : Mm(Vb⊗W ) → Mm(Mn(X)) has norm For each u ∈ Mm(V ⊗ W ), eφm(u) = φm(u) and α(u) ≤ kuk∧. Then, for every m, kφmk ≤ keφmk, and thus kφkjcb ≤ kφkAα. kφmk = sup(cid:8)φm(u) : u ∈ Mm(V ⊗ W ), kuk∧ ≤ 1(cid:9) . kφkjcb = kφkcb = sup m kφmk, BILINEAR IDEALS IN OPERATOR SPACES 7 (b) For φ ∈ Mn(Aα(V ×W, X)), leteφ ∈ Mn(cid:16)CB(V mappings. For any r1 ∈ CB(U1, V ), r2 ∈ CB(U2, W ) and s ∈ CB(X, Y ), the following equality holds. α ⊗ W, X)(cid:17) be its associated matrix of linear A direct computation gives the required inequality. ksn ◦ φ ◦ (r1, r2)kAα = ksn ◦eφ ◦ (r1 ⊗ r2)kcb (cid:3) Example 2.5. Since MB(V × W, X) ∼= CB(V is an operator space bilinear ideal. h ⊗ W, X), from Proposition 2.4 we obtain that MB With similar arguments to those used to prove Proposition 2.4, we obtain: Proposition 2.6. Let α be an operator space tensor norm and B be an operator space ideal of linear mappings. Given the operator spaces V , W and X, let AB α (V × W, X) be the operator space determined by the identification (4) α (V × W, X) ∼= B(V AB α ⊗ W, X). Then, AB α is an operator space bilinear ideal. 3. Completely nuclear and completely integral bilinear mappings In [7, Sections 12.2 and 12.3] the definitions of completely nuclear and completely integral linear mappings are presented. We now introduce and study the analogous bilinear concepts. We will see that they define operator space bilinear ideals. Theorem 3.8 provides a concrete identification of the integral bilinear ideal as in (4). On the contrary, from Proposition 3.12, it will follow that the nuclear bilinear ideal can not be described in such a way. In order to properly define the notion of nuclearity in the context of bilinear mappings on operator spaces, we need to state first some natural mappings. Let Θ : (V ∗ ∨ ⊗ W ∗) ∨ ⊗ X ֒→ J CB(V × W, X) ∨ be the natural complete isometry obtained as a composition of the natural complete isometries V ∗ and (V ∗ ∨ ⊗ X (see [7, Proposition 8.1.2 and Proposition 8.1.5]). Let ⊗ X ֒→ CB(Vb⊗W, X) ∼= J CB(V × W, X) ⊗ W ∗ ֒→ (Vb⊗W )∗, ∨ ⊗ W ∗) ∨ ∨ (Vb⊗W )∗ ⊗ X ֒→ (Vb⊗W )∗ Φ : (V ∗b⊗W ∗)b⊗X → (V ∗ ∨ ⊗ W ∗) ∨ ⊗ X be the canonical complete contraction and let With such a Ψ: Ψ = Θ ◦ Φ : (V ∗b⊗W ∗)b⊗X → J CB(V × W, X). Definition 3.1. A bilinear mapping φ ∈ J CB(V × W, X) is completely nuclear if it belongs to the image of Ψ. The operator space structure in the set of completely nuclear bilinear mappings N (V × W, X), is given by the identification of the image of Ψ with the quotient of its domain by its kernel. That is, N (V × W, X) ∼= (V ∗b⊗W ∗)b⊗X/ ker Ψ. 8 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Proposition 3.2. N is an operator space bilinear ideal. Proof. By definition N (V × W, X) is a linear subspace of J CB(V × W, X) and the contention of finite type elements is plain. The injective mapping N (V × W, X) → J CB(V × W, X) induced on the quotient by the complete contraction Ψ, has norm less or equal than Ψ, and so, it is again a complete contraction. Hence, kφkjcb ≤ kφkN and (a) is proved. (b) Let Ψ denote the quotient map induced by Ψ. Given φ ∈ Mn (N (V × W, X)), r1 ∈ CB(U1, V ), r2 ∈ CB(U2, W ) and s ∈ CB(X, Y ), consider the following diagram: ((r∗ 1 ⊗r∗ Mn(cid:0)(V ∗b⊗W ∗)b⊗X(cid:1) 2 )b⊗Y(cid:1) Mn(cid:0)(U ∗ 1b⊗U ∗ 2 )⊗s)n Ψn Ψn Mn (N (V × W, X)) / Mn (N (U1 × U2, Y )) , where the right vertical arrow is the mapping φ 7→ sn ◦ φ ◦ (r1, r2). It is immediate to check that the mappings are well defined and that the diagram commutes. In particular, sn ◦ φ ◦ (r1, r2) belongs to Mn (N (U1 × U2, Y )). If u ∈ Mn(cid:0)(V ∗b⊗W ∗)b⊗X(cid:1) is such that Ψn(u) = φ it holds sn ◦ φ ◦ (r1, r2) = sn ◦ Ψn(u) ◦ (r1, r2) = Ψn (((r∗ 2) ⊗ s)n (u)) . 1 ⊗ r∗ The estimate we are looking for follows from the fact that the inequality ksn ◦ φ ◦ (r1, r2)kN ≤ k ((r∗ 1 ⊗ r∗ 2) ⊗ s)n (u)kMn((U ∗ 1 b⊗U ∗ 2 )b⊗Y ) (cid:3) holds for every u such that Ψn(u) = φ. Definition 3.3. We say that a bilinear mapping φ ∈ J CB(V × W, X) is completely integral if kφkI = sup {kφF1×F2kN : F1 ⊂ V, F2 ⊂ W of finite dimension} < ∞. Let I(V × W, X) be the space of all completely integral bilinear mappings from V × W to X. We consider in I(V × W, X) the matrix norm given by kφkI = sup {kφF1×F2kN : F1 ⊂ V, F2 ⊂ W of finite dimension} , for every φ ∈ Mn (I(V × W, X)). It is easy to see that this norm endowed I(V × W, X) with the structure of an operator space. Proposition 3.4. Let V, W , X be operator spaces and let φ ∈ Mn (N (V × W, X)). Then The first inequality also holds for φ ∈ Mn (I(V × W, X)). kφkjcb ≤ kφkI ≤ kφkN . Proof. For φ ∈ Mn (I(V × W, X)), consider finite dimensional spaces F1 ⊂ V and F2 ⊂ W . Since kφF1×F2kjcb ≤ kφF1×F2kN and kφkjcb = sup {kφF1×F2kjcb : F1 ⊂ V, F2 ⊂ W of finite dimension} we obtain that kφkjcb ≤ kφkI . / / / BILINEAR IDEALS IN OPERATOR SPACES 9 Now, if φ ∈ Mn (N (V × W, X)) and we denote by j1 : F1 ֒→ V and j2 : F2 ֒→ W the canonical (completely contractive) embeddings, it is clear that kφF1×F2kN = kφ ◦ (j1, j2)kN ≤ kφkN · kj1kcb · kj2kcb = kφkN . (cid:3) Proposition 3.5. I is an operator space bilinear ideal. Proof. By definition I(V × W, X) is a linear subspace of J CB(V × W, X). Finite type continuous bilinear maps are obviously contained in I(V × W, X). Condition (a) was already proved above. (b) Let φ ∈ Mn (I(V × W, X)), r1 ∈ CB(U1, V ), r2 ∈ CB(U2, V ) and s ∈ CB(X, Y ). For finite dimensional spaces F1 ⊂ U1 and F2 ⊂ U2 let j1 : F1 ֒→ U1 and j2 : F2 ֒→ U2 be the canonical (completely contractive) embeddings. We have ksn ◦ φ ◦ (r1, r2)F1×F2kN = ksn ◦ φ ◦ (r1j1, r2j2)kN ≤ kskcb · kφkI · kr1kcb · kr2kcb. (cid:3) A pointwise limit of completely nuclear bilinear contractions is not necessarily completely nuclear, but it is always integral. This result is in the following two lemmas and will be used several times. The statements given here are simpler than their linear analogues given in [7, Lemma 12.2.7 and Lemma 12.3.1]. Lemma 3.6. Let (φλ) and φ in Mn (N (F1 × F2, Mm)), where F1 and F2 are finite dimensional operator spaces. Suppose that there exists a constant C such that kφλkMn(N (F1×F2,Mm)) ≤ C for all λ and that φλ(x, y) → φ(x, y) for every (x, y) ∈ F1 × F2. Then, kφkMn(N (F1×F2,Mm)) ≤ C. Proof. Take {x1, . . . , xk} and {y1, . . . , yl} vector bases of F1 and F2, respectively, and denote by {x∗ l } the corresponding dual bases. Since 1, . . . , x∗ k} and {y∗ 1, . . . , y∗ kφλ − φkMn(N (F1×F2,Mm)) ≤ Xi,j ≤ Xi,j Hence, the result follows. kφλ(xi, yj) − φ(xi, yj)kMn·m · kx∗ i ⊗ y∗ j kN (F1×F2) kφλ(xi, yj) − φ(xi, yj)kMn·m · kx∗ i k · ky∗ j k → 0. (cid:3) Lemma 3.7. Suppose that φ ∈ Mn (J CB(V × W, Mm)) and that there exists a net (φλ) ⊂ Mn (N (V × W, Mm)) with kφλkMn(N (V ×W,Mm)) ≤ C, for all λ and φλ(v, w) → φ(v, w), for all v ∈ V, w ∈ W. Then, φ belongs to Mn (I(V × W, Mm)) and kφkMn(I(V ×W,Mm)) ≤ C. Proof. For a given pair of finite dimensional subspaces F1 ⊂ V and F2 ⊂ W , the net (φλF1×F2) and the map φF1×F2 satisfy the hypothesis of the previous lemma. Thus, kφF1×F2kMn(N (V ×W,Mm)) ≤ C. This implies that φ is completely integral and kφkMn(I(V ×W,Mm)) ≤ C. (cid:3) φλ =Xi,j we have φλ(xi, yj) x∗ i ⊗ y∗ j and φ(xi, yj) x∗ i ⊗ y∗ j φ =Xi,j 10 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA For the classes of completely nuclear and completely integral mappings, it is necessary to recall the linear definitions in order to make precise the relationship between bilinear mappings on operator spaces and linear mappings on operator space tensor products. A complete exposition of this topic is provided in [7, Chapter 12]. A linear mapping ϕ : V → W is said to be completely nuclear, ϕ ∈ LN (V, W ), if it belongs to the image of the canonical completely contractive mapping The operator space structure of LN (V, W ) is given by the identification ⊗ W ֒→ CB(V, W ). LΨ : V ∗b⊗W → V ∗ ∨ LN (V, W ) ∼= V ∗b⊗W/ ker LΨ. A linear mapping ϕ : V → W is said to be completely integral, ϕ ∈ LI(V, W ), if the completely nuclear norms of all its restrictions to finite dimensional subspaces of V are bounded. The operator space matrix norm on LI(V, W ) is given by kϕkMn(LI (V,W )) = sup {kϕF kLN : F ⊂ V of finite dimension} , for each ϕ ∈ Mn (LI(V, W )). So, the relation we were seeking states the following: Theorem 3.8. For every three operator spaces V, W and X, there is a complete isometry I(V × W, X) ∼= LI(V ∨ ⊗ W, X). An analogous relation in the Banach space setting holds, and it is crucial in the study of the bilinear integral mappings (see [19]). The proof for operator spaces is, however, quite more involved. We prove first the particular case of Theorem 3.8 when X is the finite dimensional operator space of n × n-matrices Mn. The operator space dual/pre-dual of Mn is the space Tn of n × n- matrices where the norm is given by kαkTn = trace(α). Remark 3.9. A version of "Goldstine's theorem" holds in operator spaces: If u ∈ Mn(V ∗∗) with kuk ≤ 1, then there exists a net (uλ) ∈ Mn(V ) such that kuλk ≤ 1, for all λ and ϕn(uλ) → u(ϕ), for all ϕ ∈ V ∗ (see [7, Proposition 4.2.5]). Proposition 3.10. There is a complete isometry I(V × W, Mn) ∼= LI(V ∨ ⊗ W, Mn). Proof. Since Mn = T ∗ that there is a completely isometric identity n is a finite-dimensional operator space, from [7, Corollary 12.3.4] we get Thus, the result will be proved once we see that there is a complete isometry LI(V ∨ ⊗ W, Mn) ∼=(cid:18)(V I(V × W, Mn) ∼=(cid:18)(V ∨ ⊗ W ) ∨ ⊗ W ) . ∨ ⊗ Tn(cid:19)∗ ⊗ Tn(cid:19)∗ ∨ . To that end, consider the following applications: BILINEAR IDEALS IN OPERATOR SPACES 11 , which is the canonical completely isometric isomorphism given by the identification • S : J CB(V × W, Mn) → (cid:0)(Vb⊗W )b⊗Tn(cid:1)∗ • bΨ : (V ∗b⊗W ∗)b⊗Mn → N (V × W, Mn), the quotient map. • Ω : (V ∗b⊗W ∗)b⊗Mn →(cid:18)(V J CB(V × W, Mn) ∼= CB(Vb⊗W, Mn) ∼=(cid:0)(Vb⊗W )b⊗Tn(cid:1)∗ ⊗ Tn(cid:19)∗ ∨ ⊗ W ) ∨ ∨ ⊗ W ∨ ⊗ Tn)∗ n → (V . 7→ (v ⊗ w ⊗ φ 7→ v∗(v)w∗(w)φ∗(φ)), V ∗ × W ∗ × T ∗ (v∗, w∗, φ∗) , the linearization of the trilinear mapping which is completely contractive. ∨ ⊗ Tn(cid:19)∗ ∨ ⊗ W ) • Φ∗ : (cid:18)(V (Vb⊗W )b⊗Tn → (V ֒→ (cid:0)(Vb⊗W )b⊗Tn(cid:1)∗ range, Φ∗ results an injective complete contraction. , which is the transpose mapping of Φ : ∨ ⊗ W ) ∨ ⊗ Tn. Since Φ is a complete contraction and it has dense Replicating the argument of the linear case we use the previous mappings to construct a com- mutative diagram: yS N (V × W, Mn) ⊆ I(V × W, Mn) ⊆ J CB(V × W, Mn) bΨx (V ∗b⊗W ∗)b⊗Mn ∨ ∨ ⊗ W ) Ω−→ (cid:18)(V ⊗ Tn(cid:19)∗ Snuc : N (V × W, Mn) →(cid:18)(V Φ∗ −→ (cid:0)(Vb⊗W )b⊗Tn(cid:1)∗ ⊗ Tn(cid:19)∗ The injectivity of both SN and Φ∗ yields that ker(Ω) = ker(bΨ). This allows us to define: in such a way that Snuc◦bΨ = Ω and Φ∗◦Snuc = SN . The mapping Snuc is a complete contraction. Let us suppose now that φ ∈ I(V × W, Mn) with kφkI(V ×W,Mn) ≤ 1. We want to see that S(φ) is continuous with respect to the injective tensor norm of (V ⊗ W ) ⊗ Tn. Given u ∈ (V ⊗ W ) ⊗ Tn with kuk∨ ≤ 1, there exist finite-dimensional spaces Vu ⊂ V and Wu ⊂ W such that u ∈ (Vu ⊗ Wu) ⊗ Tn. Let us call jVu : Vu ֒→ V and jWu : Wu ֒→ W the canonical inclusions, then ∨ ⊗ W ) ∨ hS(φ), ui = hSnuc(φ ◦ (jVu , jWu )), ui. Therefore, hS(φ), ui ≤ kSnuc(φ ◦ (jVu , jWu))k(cid:18)(Vu ∨ ⊗Wu) ∨ ⊗Tn(cid:19)∗ · kuk ∨ ⊗Wu) ∨ ⊗Tn (Vu ≤ kφ ◦ (jVu, jWu )kN (Vu×Wu,Mn) · kuk∨ ≤ kφkI(V ×W,Mn) ≤ 1. Thus, S determines a contractive mapping Sint : I(V × W, Mn) →(cid:18)(V ∨ ⊗ W ) ∨ ⊗ Tn(cid:19)∗ . Through a similar argument it can be seen that Sint is also a complete contraction. 12 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Let us show now that Sint is a complete isometry. For that, get φ ∈ Mm (I(V × W, Mn)) such that k(Sint)m(φ)k Mm(cid:18)(cid:18)(V ∨ ⊗W ) ∨ ⊗Tn(cid:19)∗(cid:19) ≤ 1. We have to prove that kφkMm(I(V ×W,Mn)) ≤ 1. Since (Sint)m(φ) ∈ Mm(cid:18)(cid:18)(V ∨ ⊗ Tn ֒→ J CB(V ∗ × W ∗, Tn) is a complete isometry, by Remark 1.2, (Sint)m(φ) extends to ^(Sint)m(φ) ∈ CB(J CB(V ∗ × W ∗, Tn), Mm) preserving the norm. Now, we have completely isometric identifi- cations ⊗ Tn(cid:19)∗(cid:19) ∼= CB((V ∨ ⊗ Tn, Mm) and (V ∨ ⊗ W ) ∨ ⊗ W ) ∨ ⊗ W ) ∨ , Mm(cid:17) CB(J CB(V ∗ × W ∗, Tn), Mm) ∼= CB(CB(V ∗b⊗W ∗, Tn), Mm) ∼= CB(cid:16)(cid:0)(V ∗b⊗W ∗)b⊗Mn(cid:1)∗ and we thus know that k ^(Sint)m(φ)kMm(((V ∗b⊗W ∗)b⊗Mn)∗∗) ≤ 1. Hence, by Remark 3.9, there exists a net (uλ) in Mm(cid:0)(V ∗b⊗W ∗)b⊗Mn(cid:1) with kuλk ≤ 1 such that, for all ϕ ∈(cid:0)(V ∗b⊗W ∗)b⊗Mn(cid:1)∗ ∼= Mm(cid:16)(cid:0)(V ∗b⊗W ∗)b⊗Mn(cid:1)∗∗(cid:17) , ϕm(uλ) → ^(Sint)m(φ)(ϕ). , In particular, for any v ∈ V , w ∈ W and α ∈ Tn, ((v ⊗ w) ⊗ α)m(uλ) → ^(Sint)m(φ)((v ⊗ w) ⊗ α) = (Sint)m(φ)((v ⊗ w) ⊗ α). Looking into the coordinates of this matrix limit, with the notation uλ = (uk,l λ )k,l and φ = (φk,l)k,l, we obtain λ )(v, w), αi = ((v ⊗ w) ⊗ α)(uk,l λ ) → Sint(φk,l)((v ⊗ w) ⊗ α) = hφk,l(v, w), αi, for every (v, w) ∈ V × W , α ∈ Tn and k, l ∈ {1, . . . , m}. Thus, for each pair (v, w) ∈ V × W , the converges weakly to φk,l(v, w). Being Mn a finite dimensional space, this convergence turns out to be strong and now we can also forget the coordinates and look at the hbΨ(uk,l net(cid:16)bΨ(uk,l λ )(v, w)(cid:17)k,l whole picture again. So we have bΨm(uλ)(v, w) → φ(v, w), for all (v, w) ∈ V × W . Since bΨ is a complete contraction, we know kbΨm(uλ)kMm(N (V ×W,Mn)) ≤ 1 and with an ap- It only remains to prove that Sint is surjective. Let f ∈(cid:18)(V ⊗ Tn(cid:19)∗ . The surjectivity of S tells us that there exists φ ∈ J CB(V × W, Mn) such that Φ∗(f ) = S(φ). Moreover, for finite dimensional spaces F1 ∈ V and F2 ∈ W with canonical inclusions j1 : F1 ֒→ V and j2 : F2 ֒→ W it holds pealing to Lemma 3.7 we derive that kφkMm(I(V ×W,Mn)) ≤ 1. ∨ ⊗ W ) ∨ Φ∗(f ◦ (j1, j2)) = S(φ ◦ (j1, j2)). Since φ ◦ (j1, j2) belongs to N (F1 × F2, Mn) ∼= I(F1 × F2, Mn) it is clear that Sint(φ ◦ (j1, j2)) = f ◦ (j1, j2). Hence, kφ ◦ (j1, j2)kN (F1×F2,Mn) = kφ ◦ (j1, j2)kI(F1×F2,Mn) = kSint(φ ◦ (j1, j2))k = kf ◦ (j1, j2)k ≤ kf k. Thus, φ ∈ I(V × W, Mn) with kφkI(V ×W,Mn) ≤ kf k. (cid:3) BILINEAR IDEALS IN OPERATOR SPACES 13 Now we can prove the general result I(V × W, X) ∼= LI (V ∨ ⊗ W, X) : Proof of Theorem 3.8. Let φ ∈ I(V × W, X) and consider the associated linear application Lφ : V ⊗ W → X. We begin by proving that Lφ is completely bounded from (V ⊗ W, ∨) to X. This will allows us to extend Lφ to V ∨ ⊗ W . For that, we need to find a common bound for the norms of the mappings (Lφ)n : Mn(V ⊗ W, ∨) → Mn(X). Let u ∈ Mn(V ⊗ W ). By [7, Lemma 2.3.4], there exists ξ ∈ CB(X, Mn) with kξkcb ≤ 1 satisfying k(Lφ)n(u)kMn(X) = kξn ((Lφ)n(u)) kMn(Mn) = k(ξ ◦ Lφ)n(u)kMn(Mn) = k(Lξ◦φ)n(u)kMn(Mn). Since ξ ◦ φ : V × W → Mn is completely integral, we know from Proposition 3.10 that Lξ◦φ belongs to LI(V ∨ ⊗ W, Mn). Thus, Lξ◦φ ∈ CB(V ∨ ⊗ W, Mn) and therefore, k(Lξ◦φ)n(u)kMn(Mn) ≤ kLξ◦φkcb · kuk Mn(V ∨ ⊗W ) ≤ kξkcb · kφkI(V ×W,X) · kuk Mn(V . ∨ ⊗W ) This yields that Lφ ∈ CB(V Lφ belongs to Mn(cid:18)LI(V ∨ ∨ ⊗W, X). Let us prove now that, indeed, given φ ∈ Mn (I(V × W, X)), ⊗ W, X)(cid:19). To that end we need to compute the nuclear norms of its ∨ F1 ∨ ⊗ F2)∗ ∼= F ∗ ∨ ⊗ F2, X). Thus, restrictions to finite dimensional spaces. Let F ⊂ V exist finite dimensional subspaces F1 ∈ V and F2 ∈ W such that F ⊂ F1 isometry (F1 LN (F1 ∨ ⊗ W be a finite dimensional subspace. There ∨ ⊗ F2. The complete 2 (see, for instance, [7, (15.4.1)]) yields that N (F1 × F2, X) ∼= 1b⊗F ∗ kLφF kMn(LN (F,X)) ≤(cid:13)(cid:13)(cid:13)Lφ ⊗F2(cid:13)(cid:13)(cid:13)Mn(cid:18)LN (F1 ⊗F2,X)(cid:19) = kφF1×F2kMn(N (F1×F2,X)) ≤ kφkMn(I(V ×W,X)) Hence, it follows that Lφ ∈ Mn(cid:18)LI(V ⊗ W, X)(cid:19) and kLφk To prove the opposite contention, consider L ∈ Mn(cid:18)LI (V ⊗F2(cid:13)(cid:13)(cid:13)Mn(cid:18)LN (F1 kφF1×F2kMn(N (F1×F2,X)) =(cid:13)(cid:13)(cid:13)Lφ L is Lφ, for some φ ∈ Mn (J CB(V × W, X)) . The same argument as above shows that for any finite dimensional subspaces F1 ∈ V and F2 ∈ W , Mn(cid:18)LI (V ⊗W,X)(cid:19) ≤ kφkMn(I(V ×W,X)). ⊗ W, X)(cid:19). It is plain to see that Consequently, φ ∈ Mn (I(V × W, X)) and kφkMn(I(V ×W,X)) ≤ kLφk ⊗F2,X)(cid:19) ≤ kLφk Mn(cid:18)LI (V ∨ ⊗W,X)(cid:19). ∨ F1 (cid:3) ∨ ∨ ∨ ∨ ∨ Mn(cid:18)LI (V ∨ ⊗W,X)(cid:19). The scalar valued case. Let V and W be operator spaces and let ν be the linear isomorphism in (1). As a corollary of Theorem 3.8 we have that ν induces the following complete isometry: Proposition 3.11. I(V × W ) ∼= (V ∨ ⊗ W )∗. In contrast, in the case of the nuclear bilinear ideal we have: Proposition 3.12. The following are equivalent: 14 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA (i) There exists an operator space tensor norm α such that N (V × W ) ∼= (V (ii) N (V × W ) = I(V × W ). α ⊗ W )∗. In this case, α coincides with the injective operator space tensor norm. Proof. (i) follows from (ii) by Proposition 3.11. To prove the other implication, recall that k · k∨ ≤ k · kα for any operator space tensor norm α. Thus, if (i) holds for some α, then I(V × W ) ∼= (V α ⊗ W )∗ ∼= N (V × W ). ∨ ⊗ W )∗ ⊂ (V (cid:3) It is worth noticing that there are examples of completely integral scalar valued bilinear mappings which are not completely nuclear (see Example 6.1). Thus, the completely nuclear bilinear ideal is not of the type described in Proposition 2.6. Something more can be said about a tensorial representation of N (V × W ). First, recall the following definition Definition 3.13. An operator space V is said to have the operator space approximation ∨ ⊗ V and for every ε > 0 there exists a finite rank property (OAP) if for every u ∈ K(H) mapping T on V such that ku − (I ⊗ T )(u)k < ε. By [7, Theorem 11.2.5], V has OAP if and only if the canonical inclusion Vb⊗W ֒→ V ∨ ⊗ W is one-to-one, for every operator space W (or just for V ∗). Recall that the standard translation of this result to the Banach space setting was also valid. As a direct consequence we can state the following: Proposition 3.14. If V ∗ or W ∗ has OAP then there is a complete isometry: N (V × W ) ∼= V ∗b⊗W ∗. As an example we can consider a reflexive operator space V such that its dual V ∗, looked as a Banach space has the (Banach) approximation property but as an operator space V ∗ has not OAP (see [1, 11] for examples of such spaces). In this case, the space of (Banach) nuclear bilinear forms on V × V ∗ has a canonical representation as a projective tensor product while the space of completely nuclear bilinear forms has not: N B(V × V ∗) ∼= V ∗ ⊗π V ∗∗ and N (V × V ∗) 6∼= V ∗b⊗V ∗∗. Remark 3.15. The argument in Proposition 3.14 can be easily extended to the vector valued case. Hence, we have whether two of the three spaces V ∗, W ∗ and X have OAP. N (V × W, X) ∼= (V ∗b⊗W ∗)b⊗X, Looking at the equivalence J CB(V ×W ) ≃ CB(V, W ∗) and taking into account the situation in the Banach space setting, we question about the existence of an operator space identification for completely nuclear bilinear/linear mappings and for completely integral bilinear/linear mappings. For the nuclear case, a careful look to the definitions of the spaces of completely nuclear bilinear and linear mappings, easily gives the following. Proposition 3.16. N (V × W ) ∼= LN (V, W ∗). BILINEAR IDEALS IN OPERATOR SPACES 15 The situation for completely integral mappings is quite different: since LI(V, W ∗) is not ∨ ⊗ W )∗ [7, Section 12.3] then neither the spaces I(V × W ) and always completely isometric to (V LI(V, W ∗) are always completely isometric. In the Banach space setting, the space of integral bilinear forms from two Banach spaces is isometrically isomorphic to the space of integral linear mappings from one of the spaces to the dual of the other (see, for instance, [17, Proposition 3.22]). The hidden reason behind this different behavior is the Principle of Local Reflexivity, which is valid for every Banach space while its operator space version does not always hold (see [7, Section 14.3] or [14, Definition 18.1] for a precise definition). Indeed, [7, Theorem 14.3.1] along with Proposition 3.11 give us the statement below. Proposition 3.17. Let W be an operator space.Then the following are equivalent: (i) W is locally reflexive. (ii) For every operator space V , there is a complete isometry I(V × W ) ∼= LI(V, W ∗). 4. Completely extendible bilinear mappings Within the scope of Banach spaces, the non-validity of a Hahn-Banach theorem for multilinear mappings and homogeneous polynomials motivates the study of the 'extendible' elements (those that can be extended to any superspace). We propose and study here a version of this concept for bilinear mappings between operator spaces. Our approach was strongly inspired by the results and arguments of [4] (see also [9]). Definition 4.1. A mapping φ ∈ J CB(V × W, Z) is completely extendible if for any operator spaces X and Y such that V ⊂ X, W ⊂ Y there exists a jointly completely bounded extension φ : X × Y → Z of φ. By the Representation Theorem for operator spaces (see, for instance [7, Theorem 2.3.5]), any operator space can be seen, through a complete isometry, as a subspace of certain L(H). Given V and W , let us denote the complete isometries that realize these spaces by ΩV : V → L(HV ) and ΩW : W → L(HW ). Following the idea of [4, Theorem 3.2], we obtain: Proposition 4.2. A jointly completely bounded mapping φ : V × W → Z is extendible if and only if it can be extended to L(HV ) × L(HW ). In this case, if φ0 is such an extension, then for every X ⊃ V and Y ⊃ W there exists an extension φ : X × Y → Z with kφkjcb ≤ kφ0kjcb. Proof. Let φ0 : L(HV ) × L(HW ) → Z be an extension of φ. By Remark 1.2, ΩV and ΩW have complete contractive extensions ΩV : X → L(HV ) and ΩW : Y → L(HW ). Then, φ : X ×Y → Z given by φ(x, y) = φ0(ΩV (x), ΩW (y)), for all x ∈ X, y ∈ Y, extends φ and kφkjcb ≤ kφ0kjcb · kΩV kcb · kΩW kcb = kφ0kjcb. (cid:3) 16 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA Let E(V × W, Z) = {φ ∈ J CB(V × W, Z) : φ is extendible} . It is clear that E(V × W, Z) is a subspace of J CB(V × W, Z). Moreover, it is an operator space if we consider the following norm: for each φ ∈ Mn (E(V × W, Z)), let kφkE be the infimum of the numbers C > 0 such that for all X ⊃ V and Y ⊃ W there exists φ ∈ Mn (J CB(X × Y, Z)) which extends φ, kφkjcb ≤ C. The previous proposition tells us that we can define equivalently kφkE = inf {kφ0kjcb : φ0 extension of φ to Mn (J CB(L(HV ) × L(HW ), Z))} . Proposition 4.3. E is an operator space bilinear ideal. Proof. Since continuous functionals are completely extendible, it is clear that all finite type continuous bilinear mappings belong to this subspace. (a) For any φ ∈ Mn (E(V × W, Z)) we know that kφkjcb ≤ kφ0kjcb for every extension φ0 ∈ Mn (J CB(L(HV ) × L(HW ), Z)). Thus, kφkjcb ≤ kφkE . (b) Consider φ ∈ Mn (E(V × W, Z)), r1 ∈ CB(U1, V ), r2 ∈ CB(U2, W ) and s ∈ CB(Z, Y ). Since φ is a matrix of completely extendible maps, given ε > 0, there exists an extension φ0 ∈ Mn (J CB(L(HV ) × L(HW ), Z)) such that kφ0kjcb ≤ kφkE + ε. According to Remark 1.2, let R1 : L(HU1) → L(HV ) and R2 : L(HU2) → L(HW ) be completely bounded extensions of r1 and r2, respectively, with kr1kcb = kR1kcb and kr2kcb = kR2kcb. Then, sn ◦ φ0 ◦ (R1, R2) is an extension of sn ◦ φ ◦ (r1, r2) to L(HU1) × L(HU2) and ksn ◦ φ0 ◦ (R1, R2)kjcb ≤ kskcb · kφ0kjcb · kR1kcb · kR2kcb ≤ kskcb · (kφkE + ε) · kr1kcb · kr2kcb. Therefore, sn ◦ φ ◦ (r1, r2) ∈ Mn (E(U1 × U2, Z)) and ksn ◦ φ ◦ (r1, r2)kE ≤ kskcb · kφkE · kr1kcb · kr2kcb. (cid:3) Motivated by what is done in the Banach space setting (see [4, Corollary 3.9] or [9, Proposition 3]), we now define an operator space tensor norm η such that for any V, W , the dual operator space η ⊗ W )∗ coincides with the scalar-valued completely extendible bilinear mappings E(V × W ). (V To that end, consider the tensor product of the canonical operator space inclusions where the range is endowed with the operator space projective tensor norm: Let η be the operator space tensor norm in V ⊗ W induced by this application. Thus, for any u ∈ Mn (V ⊗ W ), ΩV ⊗ ΩW : V ⊗ W → L(HV )b⊗L(HW ). η(u) = k(ΩV ⊗ ΩW )n(u)k∧. It is plain to see that η is an operator space matrix norm that does not depend on the representations of ΩV and ΩW but just on the operator space structure of V and W . Also, since ΩV and ΩW are complete isometries it easily follows that η is a cross matrix norm. Moreover, it can be proved evidently that η is an operator space tensor norm according to Definition 2.3. Let V η ⊗ W denote the completion of (V ⊗ W, η). Proposition 4.4. There is a complete isometry ∼= E(V × W ). (cid:18)V η ⊗ W(cid:19)∗ BILINEAR IDEALS IN OPERATOR SPACES 17 η Since V η ⊗ W(cid:19)∗ and denote by φ the associated bilinear form, φ : V × W → C. Proof. Let ϕ ∈ (cid:18)V of L(HV )b⊗L(HW ). By Remark 1.2, ϕ can be extended to ϕ0 : L(HV )b⊗L(HW ) → C with ⊗ W ֒→ L(HV )b⊗L(HW ) is a complete isometry, we can see V kϕ0kcb = kϕkcb. It is easy to see that the bilinear map φ0 : L(HV ) × L(HW ) → C associated to ϕ0 is an extension of φ. Also, η ⊗ W as a subspace Then, φ is completely extendible and kφkE ≤ kϕk. kφ0kjcb = kϕ0kcb = kϕkcb. Reciprocally, let φ ∈ E(V × W ) and denote its linear associated by ϕ : V ⊗ W → C. Let φ0 : L(HV ) × L(HW ) → C be an extension of φ and consider its linear associated ϕ0 ∈ . Thus, for each u ∈ V ⊗ W , ϕ(u) = ϕ0(ΩV ⊗ ΩW )(u) ≤ kϕ0kcb · k(ΩV ⊗ ΩW )(u)k∧ = kϕ0kcb · kukη. This implies that ϕ is η-continuous and so it can be extended continuously to V η ⊗ W . Hence, (cid:0)L(HV )b⊗L(HW )(cid:1)∗ ϕ ∈(cid:18)V ⊗ W(cid:19)∗ The isometry between(cid:18)V η that the isometry is complete. with kϕk ≤ kφkE . η ⊗ W(cid:19)∗ and E(V × W ) is now proved and a similar argument shows (cid:3) 5. The symmetrized multiplicatively bounded bilinear ideal Given a bilinear mapping φ : V × W → Z, its transposed φt : W × V → Z is defined by the relation φt(w, v) = φ(v, w). We will say that an operator space bilinear ideal A is symmetric when satisfies that if φ ∈ A(V × W, Z) then φt ∈ A(W × V, Z) with kφkA = kφtkA. The bilinear ideals J CB, N , I and E are clearly symmetric, while MB is not (see Example 6.2). Definition 5.1. A bounded bilinear mapping φ : V ×W → Z is symmetrized multiplicatively bounded, φ ∈ SMB(V × W, Z) if it can be decomposed as φ = φ1 + φ2 with φ1 ∈ MB(V × W, Z) and φt 2 ∈ MB(W × V, Z). The space SMB(V ×W, Z) is equiped with an operator space structure through the identifica- tion with the sum MB(V ×W, Z)+ tMB(W ×V, Z) in the sense of operator spaces interpolation theory (see [12, Chapter 2]). In this way, the norm of a matrix φ ∈ Mn(SMB(V × W ; Z)) is given by kφksmb = inf(cid:8)k(φ1, φ2)kMn(MB(V ×W,Z)⊕1 tMB(W ×V,Z)) : φ = φ1 + φ2(cid:9) . Proposition 5.2. SMB is a symmetric operator space bilinear ideal. Proof. By means of [12, Proposition 2.1] it is easy to see that whenever A1 and A2 are operator space bilinear ideals then the same holds for A1 + A2. Hence, this is valid for SMB = MB + tMB. (cid:3) 18 W VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA We denote by (V h ⊗ V . Appealing again to interpolation theory, we can see (V h ⊗ V ) the set of elements u in V h ⊗ W ) ∩ (W h ⊗ W ) ∩ (W h ⊗ W such that ut belongs to h ⊗ V ) as an operator space with the structure inherited by the canonical inclusion in (V h ⊗ W ) ⊕∞ (W h ⊗ V ). The completely isometric identity (X ∩ Y )∗ ∼= X∗ + Y ∗ [12, page 23] applied to our case says: SMB(V × W ) ∼=(cid:16)(V h ⊗ W ) ∩ (W h ⊗ V )(cid:17)∗ , completely isometrically. In the vector-valued case, there is also some interplay between the space of symmetrized multiplicatively bounded bilinear mappings and the intersection of both Haagerup tensor products: Proposition 5.3. Let V , W and Z be operator spaces. Then: h ⊗ V ), Z) is a complete contraction. h ⊗ W ) ∩ (W (a) The inclusion SMB(V × W, Z) ֒→ CB((V (b) If Z = L(H) there is a complete isomorphism SMB(V × W, L(H)) ∼= CB((V h ⊗ W ) ∩ (W h ⊗ V ), L(H)). Proof. (a) Composing the restriction with the usual identification we naturally have the following complete contractions: MB(V ×W, Z) ֒→ CB((V h ⊗W )∩(W h ⊗V ), Z) and tMB(W ×V, Z) ֒→ CB((V h ⊗W )∩(W h ⊗V ), Z). Thus, the classical interpolation property (see [12, Proposition 2.1]) gives that the mapping SMB(V × W, Z) = MB(V × W, Z) + tMB(W × V, Z) ֒→ CB((V h ⊗ W ) ∩ (W h ⊗ V ), Z) is also a complete contraction. h ⊗ W ) ∩ (W h h ⊗ W ) ∩ (W h (b) In the case Z = L(H), let us see that the injective mapping of (a) is actually a surjective to prove that the bilinear associate φ belongs to Mn (SMB(V × W, L(H))) with kφk ≤ 2kLφk. ⊗ V ), L(H))(cid:17). We have complete isomorphism. For that, consider Lφ ∈ Mn(cid:16)CB((V Since Lφ ∈ Mn(cid:16)CB((V ⊗ V ), L(H n)(cid:17) and an extension Leφ ∈ CB(cid:16)(V ⊗ V ), L(H n)(cid:17) with the same completely bounded norm. Then, we should have that the bilinear associated to Leφ is written as φ1 + φ2 with kφ1kMB(V ×W,L(H n)) ≤ kLeφk and kφt 2kMB(W ×V,L(H n)) ≤ kLeφk. Hence, ⊗ V ), L(H))(cid:17) ∼= CB(cid:16)(V h ⊗ V ) is completely isometrically contained in (V h ⊗ W ) ⊕∞ (W h ⊗ W ) ⊕∞ (W h ⊗ V ) there is h ⊗ W ) ∩ (W h h ⊗ W ) ∩ (W (V h kφ1kMB(V ×W,L(H n)) + kφt 2kMB(W ×V,L(H n)) ≤ 2kLφk. Now, the usual identification MB(V × W, L(H n)) ∼= Mn (MB(V × W, L(H))) yields: kφkMn(SMB(V ×W,L(H))) ≤ kφ1kMn(MB(V ×W,L(H))) + kφt 2kMn(MB(W ×V,L(H))) ≤ 2kLφk Mn(cid:16)CB((V h ⊗W )∩(W h ⊗V ),L(H))(cid:17). (cid:3) BILINEAR IDEALS IN OPERATOR SPACES 19 The case of scalar valued mappings is of special interest and was extensively studied in the literature in relation with the so called Non-commutative Grothendieck's Theorem. In the next section there is a briefly exposition of this. We thank the referee for suggesting us to study the symmetrized multiplicatively bounded mappings and for his/her very valuable comments. 6. Proof of Theorem 1.1 and Examples Now we study the relationships between the bilinear ideals: we prove the inclusion relations that always hold, and provide examples that distinguish them when they are different. Proof of Theorem 1.1(a). It is clear, by definition, that every completely nuclear bilinear map- ping is completely integral. Also, the fact that k · k∨ is smaller than k · kh implies that LI (V ∨ ⊗ W, X) ⊂ CB(V ∨ ⊗ W, X) ⊂ CB(V h ⊗ W, X). h Moreover, since I(V × W, X) ∼= LI (V ⊗ W, X), we obtain that I(V × W, X) ⊂ MB(V × W, X). From the very definition of SMB, the relation MB(V × W, X) ⊂ SMB(V × W, X) always holds. ∨ ⊗ W, X) and MB(V × W, X) ∼= CB(V All these inclusions are strict as we can see in the following examples. Recall that in the Banach space setting, a classical example of an integral non-nuclear bilinear mapping is φ : ℓ1 × ℓ1 → C given by φ(x, y) =Pn xnyn. For operator spaces, a similar example works. Example 6.1. A completely integral bilinear form which is not completely nuclear. Let us consider the operator space τ (ℓ2) of trace class operators from ℓ2 to ℓ2. Naturally, each element x ∈ τ (ℓ2) is identified with an infinite matrix (xs,t). We define a bilinear map φ : τ (ℓ2) × τ (ℓ2) → C by φ(x, y) =Xs xs,s · ys,s The bilinear map φ is jointly completely bounded but not completely nuclear. Indeed, by Propo- sition 3.16, if φ is completely nuclear so is Lφ : τ (ℓ2) → L(ℓ2) given by Lφ(x) = .  x1,1 0 0 0 ... 0 x2,2 0 · · · · · · 0 x3,3 0 · · · · · · 0 x4,4 · · · · · · · · · · · · · · · · · · · · · · · ·  Lφ could not be completely nuclear because it is not compact [7, Proposition 12.2.1]. Now we want to see that φ is completely integral. Invoking Lemma 3.7, we want to estimate the completely nuclear norms of the mappings φm : τ (ℓ2) × τ (ℓ2) → C given by φm(x, y) = xs,s · ys,s. mXs=1 20 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA For each s ∈ N, let us denote by εss the element in L(ℓ2) represented by the matrix with a number 1 in position (s, s) and numbers 0 in all the other places. Recall that N (τ (ℓ2) × τ (ℓ2)) ∼= defined in Section 3. Since L(ℓ2)b⊗L(ℓ2)/ ker Ψ, where Ψ : L(ℓ2)b⊗L(ℓ2) → J CB(τ (ℓ2) × τ (ℓ2)) is the canonical mapping we have kφmkN ≤ kPm mXs=1 φm = Ψ mXs=1 2m Xδ∈{−1,1}m mXs=1 s=1 εss ⊗ εssk∧. In order to compute this norm, consider the following δsεss! ⊗ mXs=1 εss ⊗ εss! , δsεss! . usual way of expressing it: εss ⊗ εss = (5) 1 It is easy to prove that for vectors v1, . . . , vp in any operator space V we have the following representation: vj ⊗ vj = α · ((v1 ⊕ · · · ⊕ vp) ⊗ (v1 ⊕ · · · ⊕ vp)) · β, pXj=1 εss ⊗ εss(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∧ where α ∈ M1×p2, β ∈ Mp2×1 and both α and β have '1' in p of the places and '0' in the others. Applying this representation to the expression (5), we obtain ≤ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mXs=1 others. Since kαk = kβk = 2m/2 and kPm where α ∈ M1×22m , β ∈ M22m×1 and both α and β have '1' in 2m of the places and '0' in the s=1 εss ⊗ εssk∧ ≤ 1. Hence, kφmkN ≤ 1 (in fact, it is equal to 1) and by Lemma 3.7, φ is completely integral with kφkI = 1. 1 2m kαk · max · kβk, δsεss(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) δ∈{−1,1}m(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mXs=1 s=1 δsεsskL(ℓ2) = maxs δs = 1, we derive kPm L(ℓ2) 2 Example 6.2. A multiplicatively bounded bilinear mapping which is not completely integral / A symmetrized multiplicatively bounded bilinear mapping which is not multiplicatively bounded. Let H be a Hilbert space and denote by Hc the column space associated to H. An exam- ple of non commutativity of Haagerup tensor product is given through the canonical complete isometries (see, for instance [7, Propositions 9.3.1, 9.3.2 and 9.3.4]): h ⊗ (Hc)∗ ∼= Hc ∨ ⊗ (Hc)∗ ∼= K(H) Hc and (Hc)∗ h A close look to these mappings allows us to state that the application ⊗ Hc ∼= (Hc)∗b⊗Hc ∼= τ (H). Hc ⊗ (Hc)∗ → (Hc)∗ h v ⊗ w 7→ w ⊗ v ⊗ Hc could not be extended as a completely bounded mapping defined on Hc h ⊗ (Hc)∗. Consider φ : (Hc)∗ × Hc → (Hc)∗ h ⊗ Hc and φt : Hc × (Hc)∗ → (Hc)∗ h ⊗ Hc (w, v) 7→ w ⊗ v (v, w) 7→ w ⊗ v. It turns out that φ is multiplicatively bounded while φt is not. Hence, φ could not be completely integral (because the ideal of completely integral bilinear mappings is symmetric). Therefore, φ BILINEAR IDEALS IN OPERATOR SPACES 21 is multiplicatively bounded but not completely integral and φt is symmetrized multiplicatively bounded but not multiplicatively bounded. We also see in [8, Example 3.6], or in Example 6.5 below, that the bilinear ideals SMB and J CB do not coincide. Proof of Theorem 1.1(b). The ideal of completely extendible bilinear mappings cannot be placed as a link in the chain of inclusions in Theorem 1.1 (a): It contains the ideal of completely integral bilinear operators (see arguments below), but it has not a relation with the ideal of multiplicatively bounded bilinear mappings holding for every operator space. Examples 6.4 and 6.6 prove this. We will see, though, that in the particularly relevant cases when the range is C or L(H) there are relations between them. In the Banach space setting, Grothendieck-integral bilinear mappings are always extendible [5, Proposition 7]. Let us see that an analogous contention holds in the operator space framework. Pisier (personal communication) made us realize that completely integral linear mappings being completely 2-summing are hence completely extendible [13, Proposition 6.1]. This linear result allows us to derive the bilinear one. Indeed, from Theorem 3.8, we know I(V × W, X) ∼= LI(V ∨ ⊗ W, X). Now, the previous linear ∨ ⊗ W, X) ⊂ LE (V inclusion gives us LI (V k · k∨ is smaller than η, and LE is an ideal, we have LE (V η conclusion follows once we see that given any ϕ ∈ LE (V ⊗ W, X), its associated bilinear mapping φ : V × W → X belongs to E(V × W, X). η ⊗ W, X). Now, the ∨ ⊗ W, X). Also, since the operator space tensor norm ∨ ⊗ W, X) ⊂ LE (V The extendibility of ϕ along with the inclusion V ⊗ W ֒→ L(HV )b⊗L(HW ) produce that, for any ε > 0 there exists a completely bounded linear mapping ϕ0 : L(HV )b⊗L(HW ) → X that extends φ with kϕkLE ≤ kϕ0kcb ≤ kϕkLE + ε. η It is clear now that the bilinear map associated to ϕ0, φ0 : L(HV ) × L(HW ) → X, is an extension of φ that satisfies Hence, φ is completely extendible with kφkE ≤ kϕkLE . kφkE ≤ kφ0kjcb = kϕ0kcb ≤ kϕkLE + ε. Therefore, (b) in Theorem 1.1 is proved: I(V × W, X) ⊂ E(V × W, X) ⊂ J CB(V × W, X). Examples 6.4 and 6.6 below, will show that both inclusions could be strict. It is known [20, page 45] that multiplicatively bounded bilinear mappings with range L(H) are completely extendible. This can also be seen as a consequence of Arvenson-Wittstock extension theorem for completely bounded mappings (Remark 1.2) along with the fact that the Haagerup tensor norm preserves complete isometries. Moreover, the inclusion MB(V × W, L(H)) ⊂ E(V × W, L(H)) is a complete contraction. Since E is a symmetric ideal, appealing once more to [12, Proposition 2.1] we derive the complete contractive inclusion SMB(V × W, L(H)) ⊂ E(V × W, L(H)), which proves (c) in Theorem 1.1. 22 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA We do not know whether this last inclusion is strict. Actually, for scalar-valued bilinear map- pings we do know that the equality isomorphically holds. This is a consequence of Grothendieck's Theorem for C∗-algebras. In [15] one may find a broad exposition on the topic. For the moment let us recall just some relevant results in a terminology according to our presentation. Pisier and Shlyakhtenko [16] obtain the result for exact operator spaces (and also for C ∗-algebras satisfying some conditions). In [16, Theorem 0.4] they prove: Theorem (Pisier-Shlyakhtenko). If V and W are exact operator spaces, then the following iso- morphism holds: SMB(V × W ) = J CB(V × W ). Haagerup and Musat [8] prove the theorem for general C ∗-algebras. Combining [8, Theorem 1.1] with [8, Lemma 3.1] (which relies on Pisier and Shlyakhtenko's result) produces: Theorem (Haagerup-Musat). If A and B are C ∗-algebras, then the following isomorphism holds: SMB(A × B) = J CB(A × B). As a consequence, for any operator spaces V and W the following (Banach space) isomorphism holds: SMB(V × W ) = E(V × W ). Indeed, let φ ∈ E(V × W ). For V → L(HV ) and W → L(HW ) complete isometries and ε > 0, let ψ : L(HV ) × L(HW ) → C be a jointly completely bounded extension of φ with kψkjcb ≤ kφkE + ε. By Haagerup-Musat's Theorem (for A = L(HV ) and B = L(HW )), ψ can be decomposed as ψ = ψ1 + ψ2, with ψ1 ∈ MB(L(HV ) × L(HW )), ψt 2 ∈ MB(L(HW ) × L(HV )) and kψ1kmb + kψt 2kmb ≤ Kkψkjcb. Restricting the domains of ψ1 and ψ2 to V × W , we complete the proof. A predual version of the last expression reads as η ⊗ W = (V V h ⊗ W ) ∩ (W h ⊗ V ) isomorphically. It is worth noticing that Oikhberg and Pisier in [10] proved that the sum of these Haagerup h ⊗ W ) + (W h ⊗ V ) is completely isometric to the "maximal" tensor product tensor products (V V µ ⊗ W which was introduced and studied in that article. Let us now show that the other two inclusions of Theorem 1.1 (c) are strict. We have already distinguished the space of multiplicatively bounded bilinear forms from its symmetrized relative. These spaces may be different even when the range is L(H). To construct an example, first we need an easy observation: Remark 6.3. Let φ : V × W → X be a jointly completely bounded bilinear mapping and j : X → Y be a complete isometry. Then, φ is multiplicatively bounded if and only if j ◦ φ is multiplicatively bounded. Indeed, for any v ∈ Mn(V ) and w ∈ Mn(W ), since (j ◦ φ)(n)(v, w) = jn(cid:0)φ(n)(v, w)(cid:1) we have (cid:13)(cid:13)(j ◦ φ)(n)(v, w)(cid:13)(cid:13) =(cid:13)(cid:13)jn(cid:0)φ(n)(v, w)(cid:1)(cid:13)(cid:13) =(cid:13)(cid:13)φ(n)(v, w)(cid:13)(cid:13) . Thus, kj ◦ φkmb = kφkmb. BILINEAR IDEALS IN OPERATOR SPACES 23 Example 6.4. A symmetrized multiplicatively bounded bilinear mapping with range L(H), which is not multiplicatively bounded / A completely extendible bilinear mapping which is not completely integral. We recover the mappings φ and φt of Example 6.2. Denoting by V = (Hc)∗ h ⊗ Hc, we consider ΩV : V → L(HV ) the usual completely isometric inclusion. Now, let ψ = ΩV ◦ φ : (Hc)∗ × Hc → L(HV ). The previous remark and the fact that φt is not multiplicatively bounded, imply that ψt = ΩV ◦ φt : Hc × (Hc)∗ → L(HV ) neither is multiplicatively bounded. On the other hand, φ ∈ MB((Hc)∗ × Hc, (Hc)∗ h ⊗ Hc) and so ψ ∈ MB((Hc)∗ × Hc, L(HV )). Hence, ψt ∈ SMB((Hc)∗ × Hc, L(HV )). Example 6.5. A jointly completely bounded bilinear mapping (with range C) which is not ex- tendible (and hence not symmetrized multiplicatively bounded). Consider a non-complemented copy of ℓ2 in L(H), and let V be the operator space determined by ℓ2 with the matrix structure inherited from L(H). Let φ : V × V ∗ → ((ai)i, (bi)i) 7→ P∞ C i=1 aibi. φ is jointly completely bounded but there is not a jointly completely bounded extension of φ defined on L(H) × V ∗, since this extension would give rise to a bounded projection on L(H) onto that copy of ℓ2. Now we prove that the inclusion of the space of multiplicatively bounded bilinear mappings (and hence symmetrized multiplicatively bounded) into the space of completely extendible bilin- ear mappings is not longer true when the range space is an arbitrary operator space. For that, it is convenient to introduce the concept of completely extendible linear mapping. We say that a mapping ϕ ∈ CB(V, Z) is completely extendible if for any operator space X such that V ⊂ X, there exists a completely bounded extension ϕ : X → Z of ϕ. The set of completely extendible linear mappings from V to Z is denoted by LE (V, Z). Following the same steps as in the proofs of Proposition 4.2 it is obtained that ϕ ∈ CB(V, Z) is completely extendible if and only if it can be extended to L(HV ) and that LE (V, Z) is an operator space with the norm given by kϕkLE = inf{kϕ0kcb : ϕ0 extension of ϕ to Mn (CB(L(HV ), Z))}, for every ϕ ∈ Mn (LE (V, Z)). As in Proposition 4.3 it is also obtained that LE is a (linear) mapping ideal. Example 6.6. A multiplicatively bounded bilinear mapping which is not extendible. Let V be the operator space of Example 6.5. The canonical mapping V h ⊗ C → V is a complete isometry. Hence, its associated bilinear map φ : V ×C → V is multiplicatively bounded. However, since id : V → V is not extendible, φ neither is so. Acknowledgements. The first author wishes to thank the Centro de Investigación en Matemáti- cas (Guanajuato) for its kind hospitality during the months of January and February 2012, when this work was initiated. 24 VERÓNICA DIMANT AND MAITE FERNÁNDEZ-UNZUETA References [1] Arias, Alvaro. Operator Hilbert spaces without the operator approximation property. Proc. Amer. Math. Soc. 130 (2002), no. 9, 2669 -- 2677. [2] Blecher, David P.; Le Merdy, Christian. Operator algebras and their modules: an operator space approach. Oxford University Press, USA, 2005. [3] Blecher, David P.; Paulsen, Vern I. Tensor products of operator spaces. J. Funct. Anal. 99 (1991), no. 2, 262 -- 292. [4] Carando, Daniel. Extendible polynomials on Banach spaces. J. Math. Anal. Appl. 233 (1999), no. 1, 359 -- 372. [5] Carando, Daniel; Lassalle, Silvia. Extension of vector-valued integral polynomials. J. Math. Anal. Appl. 307 (2005), no. 1, 77 -- 85. [6] Effros, Edward G.; Junge, Marius; Ruan, Zhong-Jin. Integral mappings and the principle of local reflexivity for noncommutative L1-spaces Ann. of Math. 151 (2000), 59-92. [7] Effros, Edward G.; Ruan, Zhong-Jin. Operator spaces. London Mathematical Society Monographs. New Series, 23. The Clarendon Press, Oxford University Press, New York, 2000. [8] Haagerup, Uffe; Musat, Magdalena. The Effros-Ruan conjecture for bilinear forms on C∗-algebras. Invent. Math. 174 (2008), no. 1, 139-163. [9] Kirwan, Pádraig; Ryan, Raymond. Extendiblity of homogeneous polynomials on Banach spaces. Proc. Amer. Math. Soc. 126 (1998), 1023 -- 1029. [10] Oikhberg, Timur; Pisier, Gilles. The "maximal" tensor product of operator spaces. Proc. Edinburgh Math. Soc. (2) 42 (1999), no. 2, 267 -- 284. [11] Oikhberg, Timur; Ricard, Éric. Operator spaces with few completely bounded maps. Math. Ann. 328 (2004), no. 1-2, 229 -- 259. [12] Pisier, Gilles. The operator Hilbert space OH, complex interpolation and tensor norms. Mem. Amer. Math. Soc. 122 (1996), no. 585. [13] Pisier, Gilles. Non-commutative vector valued Lp-spaces and completely p-summing maps. Astérisque 247 (1998). [14] Pisier, Gilles. Introduction to operator space theory. London Mathematical Society Lecture Note Series, 294. Cambridge University Press, Cambridge, 2003. [15] Pisier, Gilles. Grothendieck's theorem, past and present. Bull. Amer. Math. Soc. (N.S.) 49 (2012), no. 2, 237-323. [16] Pisier, Gilles; Shlyakhtenko, Dimitri. Grothendieck's theorem for operator spaces. Invent. Math. 150 (2002), no. 1, 185-217. [17] Ryan, Raymond. Introduction to tensor products of Banach spaces. Springer Monographs in Mathematics. Springer-Verlag London, Ltd., London, 2002. [18] Schreiber, Bertram M. Operator spaces: Basic theory and applications. Advanced Courses of Mathematical Analysis III, World Scientific Publishing, 2008. [19] Villanueva, Ignacio. Integral mappings between Banach spaces. J. Math. Anal. Appl. 279 (2003), no. 1, 56 -- 70. [20] Wittstock, Gerd, et al. "What are operator spaces." An online dictionary. URL: http://www. math. uni-sb. de/ ag-wittstock/projekt2001. html (2001). Departamento de Matemática, Universidad de San Andrés, Vito Dumas 284, (B1644BID) Vic- toria, Buenos Aires, Argentina and CONICET. E-mail address: [email protected] Centro de Investigación en Matemáticas (Cimat), A.P. 402 Guanajuato, Gto., México E-mail address: [email protected]
1304.6337	1	1304	2013-04-23T16:19:40	A Resolvent Approach to the Real Quantum Plane	[ "math.OA" ]	Let $q\neq \pm 1$ be a complex number of modulus one. This paper deals with the operator relation $AB=qBA$ for self-adjoint operators $A$ and $B$ on a Hilbert space. Two classes of well-behaved representations of this relation are studied in detail and characterized by resolvent equations.	math.OA	math	A Resolvent Approach to the Real Quantum Plane VASYL OSTROVSKYI, KONRAD SCHM UDGEN Abstract. Let q 6= ±1 be a complex number of modulus one. This paper deals with the operator relation AB = qBA for self-adjoint operators A and B on a Hilbert space. Two classes of well-behaved representations of this relation are studied in detail and characterized by resolvent equations. AMS Subject Classification (2000). 47D40, 81R50, 47B25 Key words: real quantum plane, q-commutation relations 1. Introduction The algebraic relation ab = qba is a basic ingredient of the theory of quantum groups. Let us assume for a moment that this relation holds for a complex number q and some elements a and b of a unital ∗-algebra with involution x → x+. There are three important cases in which this relation is invariant under the involution. The first one is when a is unitary (that is, a+a = aa+ = 1) and b is hermitian (that is, b+ = b), while in the second case we have a = b+. In both cases q is real. From an operator-theoretic point of view these two cases are closely related (for instance, by taking the polar decomposition of a in the second case). In the third case a and b are hermitian and q is of modulus one. All three cases occur in the definitions of real forms of quantum groups and quantum algebras, see e.g. [7, Subsections 6.1.7, 9.2.4, 9.2.5]. The present paper deals with operator representations of the relation ab = qba in this third case. The corresponding ∗-algebra generated by a and b is the coordinate algebra of the real quantum plane [13] and of the quantum ax + b-group [14]. The general operator relation ab = qba has been studied in many papers such as [9], [2], [11], [12], [10], [15], [6], [3]. Throughout this paper q is a fixed complex number of modulus one such that q2 6= 1 and A and B are self-adjoint operators on a Hilbert space H. We write (1) q = e−iθ0, where 0 < θ0 < π. Our aim is to study the operator relation (2) AB = qBA. It turns out that this simple operator relation leads to unexpected technical dif- ficulties and interesting operator-theoretic phenomena. If A and B are bounded and AB = 0, then (2) is obviously satisfied. Let us call representations of (2) with AB = 0 trivial. Since q2 6= 1, these are the only representations of (2) given by bounded operators (see [2] or [9]). Operator representations of algebraic relations have been extensively studied in [9], but the methods developed therein lead only to trivial representations of (2). Further, as noted in [11, p. 1031], in contrast to Lie algebra relations the method of analytic vectors fails for the relation (2). 1 2 VASYL OSTROVSKYI, KONRAD SCHM UDGEN Representations of (2) by unbounded self-adjoint operators A and B have been investigated in [11] and [12]. Some classes of well-behaved representations of (2) have been introduced and classified in [12]. The present paper is devoted to an approach to the operator relation (2) that is based on the resolvents of the self- adjoint operators A and B. For two classes C0 and C1 (see Definition 2) of well- behaved representations this approach is developed in detail. This paper is organized as follows. In Section 2 we give a number of reformu- lations of the operator relation (2) in terms of the resolvent Rλ(A) and B, the resolvent Rµ(B) and A, and the resolvents Rλ(A) and Rµ(B), and we study the largest linear subspace Dq(A, B) on which relation (2) holds. In Section 4 the two classes C0 and C1 of well-behaved representations of relation (2) are defined and investigated in detail. All irreducible pairs of these classes are built of self-adjoint operators eαQ and eβP on the Hilbert space L2(R), where Q = x, P = i d dx and α, β ∈ R. We prove that the weak resolvent forms qRλq(A)B ⊆ BRλ(A) and Rµ(B)A ⊆ qARµq(B) of relation (2) hold for all pairs {A, B} of these classes and for all complex numbers λ resp. µ outside certain critical sectors. Section 5 contains the main results of this paper. These are various theorems which char- acterize (under additional technical assumptions) well-behaved representations, es- pecially pairs {A, B} of the classes C0 and C1, by weak resolvent relations such as qRλq(A)B ⊆ BRλ(A). Let A := eαQ, B := eβP , and q := e−iαβ, where α, β ∈ R. As shown in Section 4, the resolvent relations (12) and (13) are satisfied on L2(R) if αβ < π and λ, µ are not in the critical sector S(q)+. From Propositions 1 and 2 it follows that for arbitary numbers α, β, λ, µ the relations (12) and (13) holds for vectors of the closures of subspaces (B − µI)(A− λI)D0 and (A− λqI)(B − µqI)D0, respectively, where D0 = Lin {e−εx2+γx; ε > 0, γ ∈ C}. In Section 6 the orthogonal complements of these two subspaces are explicitely described and the resolvent actions on these complements are computed. Some technical preliminaries are contained in Section 3. Amongs these are prop- erties of the operator eβP and a formula for fractional powers of sectorial operators. Let us collect some basic notations on operators. Let T be a densely defined closed operator on a Hilbert space. We denote its domain by D(T ), its resolvent set by ρ(T ) and its resolvent (T − λI)−1 by Rλ(T ). Let UT be the phase operator occuring in the polar decomposition T = UTT of the operator T . The symbol L2(R) stands for the L2-space with respect to the Lebesgue measure on R. 2. General considerations on the relation (2) The following two propositions contain some simple reformulations of equation (2) in terms of the resolvents of the self-adjoint operators A and B. Proposition 1. Suppose that λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). then (i) If D is a linear subspace of D(AB) ∩ D(BA) and (2) holds for all f ∈ D, (3) BRλ(A)g = qRλq(A)Bg for all g ∈ E := (A − λI)D and E is a linear subspace of D(B). satisfied for all g ∈ E, then (2) holds for all f ∈ D := Rλ(A)E. (ii) If E is a linear subspace of D(B) such that Rλ(A)g ∈ D(B) and (3) is A Resolvent Approach to the Real Quantum Plane 3 (iii) If E is a linear subspace of D(B) and (3) holds for all g ∈ E, then (4) Rλ(A)Rµ(B)h = qRµq(B)Rλq(A)h + µλq(q − 1)Rµq(B)Rλq(A)Rλ(A)Rµ(B)h (iv) If F is a linear subspace of H such that (4) holds for all h ∈ F , then (3) is for all h ∈ F := (B − µI)E. fulfilled for all g ∈ E := Rµ(B)F . Proof. (i): Clearly, (2) implies that (A − λqI)Bg = qB(A − λI)g for f ∈ D. Hence, for all vectors of the form g = (A − λI)f , where f ∈ D, we have Rλ(A)g ∈ D(B) and Rλq(A)(A − λqI)BRλ(A)g = qRλq(A)B(A − λI)Rλ(A)g, so that BRλ(A)g = qRλq(A)Bg. (iii): Let g ∈ E. From equation (3) we obtain (B − µqI)Rλ(A)g = (BRλ(A) − µqRλ(A))g = (qRλq(A)B − µqRλ(A))g = (qRλq(A)(B − µI) + µqRλq(A) − µqRλ(A))g = (qRλq(A)(B − µI) + µq(λq − λ)Rλq(A)Rλ(A))g. (5) Setting h = (B − µI)g, we have g = Rµ(B)h. Inserting this into (5) and applying Rµq(B) to both sides yields (4) for h ∈ (B − µI)E. (ii) and (iv) follow by reversing the preceding arguments of proofs of (i) and (iii), respectively. (cid:3) Using the equalities Rλ(qA) = qRλq(A) and Rµ(qB) = qRµq(B) one can rewrite (3) in the form and (4) as BRλ(A)g = Rλ(qA)Bg qRλ(A)Rµ(B)h = Rµ(qB)Rλ(qA)h + µRµ(qB)(qRλ(qA) − Rλ(A))Rµ(B)h. In a similar manner the following proposition is derived. Proposition 2. Suppose that λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). then (i) If equation (2) is satisfied for all f of a linear subspace D ⊆ D(AB)∩D(BA), Rµ(B)Ag = qARµq(B)g (6) for all g ∈ E := (B − µqI)D and E is a subspace of D(A). for all g ∈ E, then (2) is true for all f ∈ D := Rµq(B)E. (ii) If E is a linear subspace of D(A) such that Rµq(B)g ∈ D(A) and (6) holds (iii) If E is a linear subspace of D(A) and (6) is satisfied for all g ∈ E, then (7) Rλ(A)Rµ(B)h = qRµq(B)Rλq(A)h + µλq(q − 1)Rλ(A)Rµ(B)Rµq(B)Rλq(A)h for all h ∈ F := (A − λqI)E. satisfied for all g ∈ E := Rλq(A)F . (iv) If equation (7) holds for all h of a linear subspace F ⊆ H, then (6) is 4 VASYL OSTROVSKYI, KONRAD SCHM UDGEN Comparing Propositions 1 and 2, especially formulas (4) and (7), we obtain Corollary 3. Let λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). If equation (2) holds on a linear subspace D ⊆ D(AB)∩D(BA), then the operators Rλ(A)Rµ(B) and Rµq(B)Rλq(A) commute on the linear space (A − λqI)(B − qµI)D ∩ (B − µI)(A − λI)D. Without further assumptions the linear subspace (A − λI)D of D(B) is neither a core for B nor the subspace (B − µI)E is dense in H. Note that (2) for all f ∈ D implies that (4) holds for all vectors h ∈ (B − µI)(A − λI)D and (7) is valid for h ∈ (A − λqI)(B − µqI)D. Definition 1. Dq(A, B) := {f ∈ D(BA) ∩ D(AB) : ABf = qBAf}. Obviously, Dq(A, B) is the largest linear subspace of H on which relation (2) holds. Of course, for arbitary self-adjoint operators A and B it may happen Dq(A, B) = {0}. From Proposition 1 we immediately obtain the following de- scriptions of the space Dq(A, B): Dq(A, B) = Rλ(A){g ∈ D(B) : Rλ(A)g ∈ D(B) and BRλ(A)g = qRλq(A)Bg} = Rλ(A)Rµ(B){h ∈ H : Rλ(A)Rµ(B)h = qRµq(B)Rλq(A)h + µλq(q − 1)Rµq(B)Rλq(A)Rλ(A)Rµ(B)h}. Similarly, Proposition 2 leads to the following descriptions of Dq(A, B): Dq(A, B) = Rµq(B){g ∈ D(A) : Rµq(B)g ∈ D(A) and Rµ(B)Ag = qARµq(B)g} = Rλ(A)Rµ(B)(cid:8)h ∈ H : Rλ(A)Rµ(B)h = qRµq(B)Rλq(A)h + µλq(q − 1)(A)Rλ(A)Rµ(B)Rµq(B)Rλqh(cid:9). In particular, we have Dq(A, B) ⊂ Rλ(A)D(B) ∩ Rµq(B)D(A). The operator relation (2) is obviously equivalent to the the relation (8) Hence Dq(A, B) = Dq(B, A). BAf = qABf. If equation (3) holds for all vectors g of the whole domain D(B), that is, if (9) qRλq(A)B ⊆ BRλ(A), we shall say that relation (9) is the weak A-resolvent form of equation (2) for λ, λq ∈ ρ(A). If equation (6) holds for all vectors g of the domain D(B), that is, if (10) Rµ(B)A ⊆ qARµq(B), we say that relation (10) is the weak B-resolvent form of equation (2) for µ, µq ∈ ρ(B). Setting ν = µq relation (10) can be rewritten as (11) qRνq(B)A ⊆ ARν (B). The form (11) of the weak B-resolvent relation of (2) corresponds to the weak A-resolvent form of equation (8) which is obtained by interchanging A and B and replacing q by q . Further, if equation (4) is satisfied for all h ∈ H, that is, if (12) Rλ(A)Rµ(B) = qRµq(B)Rλq(A) + µλq(q − 1)Rµq(B)Rλq(A)Rλ(A)Rµ(B), A Resolvent Approach to the Real Quantum Plane 5 The resolvent relations (12) and (13) can be rewritten as then equation (4) is called the (A, B)-resolvent form of equation (2) for λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). Likewise, if equation (7) holds for all h ∈ H, that is, if (13) Rλ(A)Rµ(B) = qRµq(B)Rλq(A) + µλq(q − 1)Rλ(A)Rµ(B)Rµq(B)Rλq(A), then equation (7) is called the (B, A)-resolvent form of equation (2) for λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). (cid:16)Rµq(B)Rλq(A) − λ2µ2q(q − 1)2 I, (cid:16)Rλ(A)Rµ(B) + λ2µ2q(q − 1)2 I, respectively. They hold for all vectors from the subspaces (B−µI)(A−λI)Dq (A, B) and (A− λqI)(B − µqI)Dq(A, B), respectively. In Section 6 we derive for a class of representations of (2) the form of resolvent relations on the complements of these subspaces. I(cid:17)(cid:16)Rλ(A)Rµ(B) + I(cid:17)(cid:16)Rµq(B)Rλq(A) − I(cid:17) = − I(cid:17) = − µλq(q − 1) µλq(q − 1) 1 1 1 µλ(q − 1) 1 µλ(q − 1) 1 1 Proposition 4. The weak A-resolvent form (9) is equivalent to the (A, B)-resolvent form (12) of equation (2). The weak B-resolvent form (10) and the (B, A)-resolvent form (13) of (2) are equivalent. Proof. First suppose that (9) holds. This means that (3) is satisfied for all vectors g ∈ D(B). Therefore, by Proposition 1(ii), equation (3) holds for h ∈ (B−µI)D(B). Since µ ∈ ρ(B), (B − µI)D(B) is equal to H which yields (12). Conversely, assume that (12) is fulfilled. Let g ∈ D(B). We set h = (B − µqI)g in (4). Since the ranges of resolvents of B are contained in the domain of B, the vector in (4) is in D(B), so we can apply the operator B − µI to both sides of (4). Then we obtain (3) which proves (9). The equivalence of (13) and (10) follows by a similar reasoning. (cid:3) The next proposition collects a number of basic facts concerning the weak resol- vent identities. Proposition 5. Suppose that λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). (i) qRλq(A)B ⊆ BRλ(A) if and only if Dq(A, B) = Rλ(A)D(B). (ii) Rµ(B)A ⊆ qARµq(B) if and only if Dq(A, B) = Rµq(B)D(A). (iii) qRλq(A)B ⊆ BRλ(A) if and only if qRλ(A)B ⊆ BRλq(A). (iv) Rµ(B)A ⊆ qARµq(B) if and only if Rµq(B)A ⊆ qARµ(B). (v) If qRλq(A)B ⊆ BRλ(A), then Dq(A, B) is a core for A. (vi) If Rµ(B)A ⊆ qARµq(B) then Dq(A, B) is a core for B. Proof. We carry out the proofs of (i), (iii), and (v). The proofs of (ii), (iv), and (vi) follows by a similar reasoning. (i): Throughout this proof let us set DB := Rλ(A)D(B). First suppose that qRλ(A)B ⊆ BRλ(A). Obviously, DB ⊂ D(A). The inclusion DB ⊂ D(B) follows from qRλ(A)Bf = BRλ(A)f , f ∈ D(B). Further, we have BDB ⊂ D(A) since qRλ(A)Bf = BRλ(A)f and Rλ(A)Bf ⊆ D(A), f ∈ D(B). Also, ADB ⊂ D(B), since A− λI maps DB onto D(B). Therefore, DB ⊂ Dq(A, B). Since g = (A − λI)f ∈ D(B) for any f ∈ Dq(A, B), we see that f = Rλ(A)g, so that Dq(A, B) ⊆ DB. Thus, DB = Dq(A, B). Conversely, assume that DB = Dq(A, B). Then (A − λI)Dq(A, B) = D(B) and by Proposition 1(i), we have qRλ(A)Bf = BRλ(A)f for all f ∈ D(B). 6 VASYL OSTROVSKYI, KONRAD SCHM UDGEN (iii): Suppose that qRλq(A)B ⊆ BRλ(A). Since Rλq(A) is bounded, we have (Rλq(A)B)∗ = B∗(Rλq(A))∗ = BRλq(A) and hence q BRλq(A) = (qRλq(A)B)∗ ⊇ (BRλ(A))∗ ⊇ Rλ(A)B, so that qRλ(A)B ⊆ BRλq(A). The converse direction follows by applying the same implication once again. (v): Since Dq(A, B) = Rλ(A)D(B) by (i), (A − λI)Dq(A, B) = D(B) is dense in (cid:3) H. Hence Dq(A, B) is a core for A. An immediate consequence of Proposition 5 is the following corollary. Corollary 6. Let λ, λq ∈ ρ(A) and µ, µq ∈ ρ(B). Assume that qRλ(A)B ⊆ BRλ(A) and Rµ(B)A ⊆ qARµq(B). Then (14) Dq(A, B) = Rλ(A)D(B) = Rµq(B)D(A) and Dq(A, B) is a core for A and B. Corollary 7. Suppose that µ, µq, µq2 ∈ ρ(B). Rµq(B)A ⊆ qARµq2 (B), then Dq2 (A2, B) is a core for B. Proof. From the assumptions we derive Rµ(B)A2 ⊆ qARµq(B)A ⊆ q2A2Rµq2 (B), that is, the weak B-resolvent form for the relation A2B = q2BA2 is satisfied. Therefore, Dq2 (A2, B) is a core for B by Proposition 5(vi). (cid:3) If Rµ(B)A ⊆ qARµq(B) and The next proposition shows how the resolvent relations (12) and (13) follow from the essential self-adjointness of a certain symmetric operator. Let us fix a, b ∈ R and choose the branch of the square root such that (15) q1/2 = q1/2 . We define an operator T with domain D(T ) := Dq(A, B) by (16) T f = ¯q1/2(A − aq1/2)(B − bq1/2)f + Lemma 8. The operator T is symmetric. ¯q1/2 − q1/2 2 abf, f ∈ D(T ). Proof. Clearly, T f = (cid:0)¯q1/2AB − bA − aB + q1/2+¯q1/2 derive 2 ab(cid:1)f . Using this formula we ab(cid:1)f, gi ab(cid:1)gi ab(cid:1)gi = hT f, gi q1/2 + ¯q1/2 2 q1/2 + ¯q1/2 q1/2 + ¯q1/2 2 2 hT f, gi = h(cid:0)¯q1/2AB − bA − aB + = hf,(cid:0)q1/2BA − bA − aB + = hf,(cid:0)¯q1/2AB − bA − aB + for f, g ∈ D(T ), that is, T is symmetric. Proposition 9. Assume that ab 6= 0 and q2 6= 1. If the operator T is essentially self-adjoint, then both resolvent relations (12) and (13) hold on H for λ = a¯q1/2, µ = b¯q1/2 and the operator Rbq1/2 (B)Raq1/2 (A) is normal. (cid:3) A Resolvent Approach to the Real Quantum Plane 7 Proof. Setting τ = ¯q1/2 − q1/2 2 ab, the operator T can be rewritten as T f = ¯q1/2(A − aq1/2)(B − bq1/2)f + τ f = q1/2(B − b¯q1/2)(A − a¯q1/2)f − τ f for f ∈ D(T ). Therefore, since T is essentially self-adjoint and τ is purely imagi- nary and nonzero (by the assumptions ab 6= 0 and q2 6= 1), the set F0 := (T − τ I)D(T ) = (A − aq1/2)(B − bq1/2)D(T ) is dense in H. By Proposition 2,(i) and (iii), equation (7) is satisfied for λ = a¯q1/2, µ = b¯q1/2 and all vectors h ∈ F0. Since F0 is dense and all resolvent operators are bounded, equation (7) holds for all h ∈ H. That is, we have Ra¯q1/2 (A)Rb¯q1/2 (B) = qRbq1/2 (B)Raq1/2 (A) (17) + ab(q − 1)Ra¯q1/2 (A)Rb¯q1/2 (B)Rbq1/2 (B)Raq1/2 (A). Thus, the (B, A)-resolvent relation (13) is satisfied. Similarly, we conclude that F1 := (T + τ I)D(T ) = (B − b¯q1/2)(A − a¯q1/2)D(T ) is dense in H and equation (4) holds for λ = a¯q1/2, µ = b¯q1/2 and h ∈ F1 by Proposition 1,(i) and (iii), and hence for all vectors h ∈ H. That is, the (A, B)- resolvent relation (12) is valid and we have Ra¯q1/2 (A)Rb¯q1/2 (B) = qRbq1/2 (B)Raq1/2 (A) (18) + ab(q − 1)Rbq1/2 (B)Raq1/2 (A)Ra¯q1/2 (A)Rb¯q1/2 (B). Comparing (17) and (18) we conclude that Ra¯q1/2 (A)Rb¯q1/2 (B)Rbq1/2 (B)Raq1/2 (A) = Rbq1/2 (B)Raq1/2 (A)Ra¯q1/2 (A)Rb¯q1/2 (B) which means that the operator Rbq1/2 (B)Raq1/2 (A) is normal. (cid:3) 3. Operator-theoretic preliminaries We denote by P = i d dx the momentum operator and by Q = x the position operator acting on the Hilbert space L2(R) with respect to the Lebesgue measure on R. Fix β > 0. Lemma 10. (i) Suppose that f (z) is a holomorphic function on the strip Iβ := {z ∈ C : 0 < Imz < β} such that (19) −∞ f (x + iy)2 dx < ∞. 0<y<β Z +∞ sup Set fy(x) := f (x + iy). Then the limits f0 := limy↓0 fy(x) and fβ := limy↑β fy(x) exist in L2(R) and we have f0 ∈ D(eβP ) and eβP f0 = fβ. such that f0 := limy↓0 fy(x) in L2(R) and eβP f0 = fβ. (ii) For each function f0 ∈ D(eβP ) there exists a unique function f as in (i) Proof. [12, Lemma 1.1]. (cid:3) 8 VASYL OSTROVSKYI, KONRAD SCHM UDGEN If f is a function as in Lemma 10(i), we write simply f (x) for f0(x) and f (x+ iβ) for fβ(x). Then the operator eβP acts by (20) (eβP )(x) = f (x + iβ), f ∈ D(eβP ). For a nonzero complex number q we denote by S(q)+ the closed sector in the plane with opening angle less than π between the positive x-axis and the half-line through the origin and q and set S(q) := S(q)+ ∪ (−S(q)+). q ¯q S(q)+ S(q) q ¯q Figure 1. The sectors S(q)+ and S(q) We fix two reals α, β such that β > 0 and 0 < αβ < π. Put q := e−iαβ. Now we define positive selfadjoint operators A and B on the Hilbert space L2(R) by A := eαQ and B := eβP . (21) Corollary 11. If f ∈ D(BA) ∩ D(B), then f ∈ D(AB) and ABf = qBAf . Proof. Using the description of the domain D(B) = D(eβP ) in Lemma 10 and formula (20) we derive (qBAf )(x) = qB(eαxf (x)) = e−αiβeα(x+iβ)f (x + iβ) = eαxf (x + iβ) = (ABf )(x). (cid:3) Clearly, the linear space D0 = Lin {e−εx2+γx; ε > 0, γ ∈ C} is contained in D(A) ∩ D(B) and it is invariant under A and B and also under the Fourier transform and its inverse. By Corollary 11, we have D0 ⊆ Dq(A, B). As noted in [12], D0 is a core for both selfadjoint operators A and B. Proposition 12. Suppose that λ ∈ C\S(q)+. Then λ and λq are in ρ(A) and qRλq(A)B ⊆ BRλ(A). Proof. If z runs through the strip {z : 0 ≤ Im z ≤ β}, then the number eαz fills the sector S(q)+. Hence the infimum of the function eαz − λ on the strip Iβ is equal to the distance of λ from S(q)+. In particular, this infimum is positive, since λ /∈ S(q)+. Let f ∈ D(B) and let f (z) be the corresponding holomorphic function from Lemma 10. Since eαz − λ has a positive infimum on the strip Iβ, the function g(z) = (eαz − λ)−1f (z) is holomorphic on Iβ and it satisfies condition (19) as well, because f does. Therefore, from Lemma 10 we conclude that g ∈ D(B) and (BRλ(A)f )(x) = (Bg)(x) = g(x + iβ) = (eα(x+iβ) − λ)−1f (x + iβ) = q(eαx − λq)−1f (x + iβ) = (qRλq(A)Bf )(x). (cid:3) A Resolvent Approach to the Real Quantum Plane 9 Another technical ingredient used below is Balakrishnan's theory of fractional powers of nonnegative operators on Banach spaces [1], see e.g. [5]. Suppose that T is a closed linear operator on a Banach space such that (−∞, 0) ⊆ ρ(T ) and sup {λ(T + λI)−1 : λ > 0} < ∞. (22) Then, for any γ ∈ C, 0 < Re γ < 1, the Balakrishnan operator J γ (see [5], p. 57) is defined by (23) J γf = sin(ε + it)π π Z ∞ 0 λγ−1(T + λI)−1T f dλ, f ∈ D(J γ) := D(T ). Here the integral is meant as an improper Riemann integral of a continuous function on (0, +∞) with values in the underlying Banach space. The operator J γ (or its closure) is considered as a power of the operator T with exponent γ. For our investigations the following special case is sufficient. Proposition 13. Suppose that A is a positive self-adjoint operator on a Hilbert space H such that ker A = {0} and let ϑ ∈ R, ϑ < π. Let T denote the normal operator eiϑA in H. Then, for any 0 < ε < 1, t ∈ R and f ∈ D(T ) = D(A) we have (24) T ε+itf = eiϑεe−ϑtAε+itf = sin(ε + it)π π Z ∞ 0 λε+it−1(T + λI)−1T f dλ, where the operators T ε+it and Aε+it are defined by the spectral functional calculus. Proof. Using that ϑ < π and A ≥ 0 it is easily verified that the operator T satisfies the conditions stated in (22). Hence formula (23) for the Balakrishnan operator J ε+it holds. For the normal operator T the closure of the operator J ε+it is just the power T ε+it defined by the functional calculus (see Example 3.3.2 in [5]), where the principal branch of the complex power has to be taken. Further, since ϑ < π, we have T ε+it = eiϑεe−ϑtAε+it. Hence formula (24) follows from (23). (cid:3) Lemma 14. If A is a positive self-adjoint operator with trivial kernel, then lim ε→+0 Aεf = f for f ∈ D(A). Proof. By the spectral calculus we have kAεf − fk2 = Z ∞ 0 λε − 12 dhE(λ)f, fi. Passing to the limit ε → +0 and using Lebesgue's dominated convergence theorem (by the assumption f ∈ D(A)) we obtain the assertion. (cid:3) 4. Two Classes of Well-behaved Representations of Relation (2) In this section we describe some well-behaved representations of relation (2). For this we also restate some results from [12]. Recall that q = e−iθ0 and 0 < θ0 < π by (1). Set θ1 := θ0 − π if θ0 > 0, θ1 := θ0 + π if θ0 < 0. Then we also have −q = e−iθ1 and 0 < θ1 < π. If A = 0 or if B = 0, then Dq(A, B) = D(B) resp. Dq(A, B) = D(A) and it is obvious that the pair {A, B} satisfies the relation (2) and the resolvent relations (3) and (4). We call pairs of the form {0, B} and {A, 0} trivial representations of relation (2). 10 VASYL OSTROVSKYI, KONRAD SCHM UDGEN Interesting representations of relation (2) are the classes C0 and C1 defined as follows. Definition 2. Suppose that ker A = ker B = {0}. We say that the pair {A, B} is a representation of the class C0 if (25) and that the pair {A, B} is a representation of the class C1 if AitB ⊆ eθ0tBAit , t ∈ R, and UAB ⊆ BUA. AitB ⊆ eθ1tBAit , t ∈ R, and UAUB = −UBUA. (26) Definition 3. The trivial pairs {A2, 0}, {0, B2} and pairs {A0, B0} and {A1, B1} of the classes C0 and C1, respectively, and orthogonal direct sum of such pairs are called well-behaved representations of relation (2). Remarks. 1. Note that the class C0 defined above is precisely the class C0 in [12], while the class C1 according to Definition 2 corresponds to C1 if θ0 < 0 and to C−1 if θ0 > 0 in [12]. 2. Suppose that {A, B} is a well-behaved representation of relation (2). If A ≥ 0 and ker A = ker B = {0}, then UA = I and {A, B} is a pair of the class C0. Further, if A ≥ 0, then the well-behaved representation {A, B} cannot have an orthogonal summand of the class C1. 3. As it is usual for relations having unbounded operator representations there are many "bad" unbounded representations of relation (2). In [11] pairs of self- adjoint operators A and B have been constructed for which Dq(A, B) is a core for A and B, but the pair {A, B} is not a well-behaved representation of relation (2) and it is not in one of classes Cn, n ∈ Z, defined in [12]. Let us describe all pairs of the classes C0 and C1 up to unitary equivalence. We fix real numbers α, α1, β, β1, where β > 0, β1 > 0, such that αβ = θ0 α1β1 = θ1, where q = e−iθ0 and −q = e−iθ1. Let u, v be two commuting self-adjoint unitaries on K. We define self-adjont (27) and Let K be a Hilbert space. operators A and B on the Hilbert space H = K ⊗ L2(R) a by (28) and self-adjoint operators A1 and B1 on the Hilbert space H1 = (K ⊕ K) ⊗ L2(R) by the operator matrices A0 = u ⊗ eαQ, B0 = v ⊗ eβP (29) A1 = (cid:18)eα1Q 0 −eα1Q(cid:19) , B1 = (cid:18) 0 eβ1P eβ1P 0 (cid:19) . Proposition 15. The pairs {A0, B0} and {A1, B1} belong to the classes C0 and C1, respectively. Each pair of the class C0 resp. C1 is unitarily equivalent to a pair {A0, B0} resp. {A1, B1} of the form (28) resp. (29). Corollary 16. Up to unitary equivalence there are precisely five nontrivial ir- reducible well-behaved representations of relation (2). These are the fours pairs {A = ε1eαQ, B = ε2eβP} on L2(R), where ε1, ε2 ∈ {+1,−1}, and the pair {A, B} on C2 ⊗ L2(R) given by (29) with K = C. Any well-behaved representation {A, B} of relation (2) satisfying ker A = ker B = {0} is a direct orthogonal sum of these representations. A Resolvent Approach to the Real Quantum Plane 11 Corollary 17. Let {A, B} be a well-behaved representation of relation (2) for which ker A = ker B = {0}. Then there is a linear subspace D ⊆ D(A) ∩ D(B) such that (i) AD = D, BD = D, and AitD = D, BitD = D for t ∈ R, (ii) D is a core for A and B, (iii) ABf = BAf for f ∈ D. Corollary 18. A pair {A, B} is a well-behaved representation (resp. of the class C0 or C1) of relation (2) if and only if {B, A} is a well-behaved representation (resp. of the class C0 or C1) of relation (8). Proposition 15 and Corollaries 16 -- 18 are contained in [12, Section 2]. The next proposition is essentially used in the proofs of various theorems in Section 5. Proposition 19. Let k = 0, 1. Suppose that ker A = ker B = {0} and Dq(A, B) is a core for B. If for t ∈ R, AitB ⊆ eθktBAit (30) then {A, B} is a pair of the class Ck. Proof. Putting qk := (−1)kq we have qk = eiθk. By (30), Proposition 2.3 in [12] applies to the pair {A, B} and the relation AB = qkBA. Hence there exists a linear subspace D of Dqk (A, B) such that D = AD is a core for B. Then ABg = qkBAg for g ∈ D by the definition of Dqk (A, B). Since A is self-adjoint and ker A = {0}, UA is self-adjoint unitary and A = AUA. Let f ∈ Dq(A, B) and g ∈ D. Using the preceding facts we derive hUAf, BAgi = hUAf, qkABgi = hf, qkUAABgi = hf, qkABgi = hqkAf, Bgi = h(−1)kqBAf, gi = h(−1)kABf, gi = h(−1)kUABf,Agi. Since D = AD is a core for B, from the preceding equality we conclude that UAf ∈ D(B) and BUAf = (−1)kUABf for f ∈ Dq(A, B). By assumption Dq(A, B) is a core for B, so the latter implies that UAB ⊆ (−1)kBUA. Since UA is a self-adjoint unitary, we get UABUA ⊆ (−1)kB, that is, the self- adjoint operator (−1)kB is an extension of the self-adjoint operator UABUA on H. This is only possible if UABUA = (−1)kB. From the latter it follows that UABUA = B and hence BUA = UAUABUA = BUA. Therefore, UAUBB = UAB ⊆ (−1)kBUA = (−1)kUBBUA = (−1)kUBUAB. Since ker B = {0}, the range of B is dense in H, so we get UAUB = (−1)kUAUB. Thus, {A, B} is in Ck. (cid:3) (31) Now let us return to the weak resolvent equations qRλq(A)B ⊆ BRλ(A), qRµq(B)A ⊆ ARµ(B). (32) For the trivial representations {0, B} resp. {A, 0} they are obviously fulfilled for all λ 6= 0 resp. µ 6= 0. The classes C0 and C1 are treated in the next theorem. Theorem 20. (i) If {A, B} is a pair of the class C0 for the relation (2), then (31) and (32) are satisfied for all λ ∈ C\S(q) and µ ∈ C\S(q). If in addition A ≥ 0 resp. B ≥ 0, then (31) resp. and (32) holds for λ ∈ C\S(q)+ resp. µ ∈ C\S(q)+. 12 VASYL OSTROVSKYI, KONRAD SCHM UDGEN (ii) If {A, B} is a pair of the class C1 for the relation (2), then (31) and (32) are fulfilled for λ ∈ C\S(−q) and µ ∈ C\S(−q). Proof. By Corollary 18 it suffices to prove all assertions for the first relation (31). Clearly, (31) is preserved under orthogonal direct sums. Therefore, by Corollary 16, it is sufficient to prove (31) for the corresponding irreducible representations listed in Corollary 16. (i): Let {A0 = eαQ, B0 = eβP} and suppose that λ /∈ S(q)+. Then, by Propo- sition 12, relation (31) is valid. Then, obviously, (31) holds also for the pair {A0,−B0}. Since Rz(−A0) = −R−z(A0), it follows that (31) is satisfied for the pairs {−A0,±B0} provided that λ /∈ −S(q)+. (ii): We have to show that (31) holds for the pair {A1, B1} given by (29). Put A := eα1Q, B := eα1P and q := −q. Then, since (33) Rz(A) = (cid:18)Rz(A) 0 0 Rz(−A)(cid:19) , B = (cid:18) 0 B B 0(cid:19) , the weak resolvent relation qRλq(A)B ⊆ BRλ(A) reduces to the relations qRλq(A)B ⊆ BRλ(−A), qRλq(−A)B ⊆ BRλ(A). Since Rz(−A) = −R−z(A) and q := −q, the latter equalities are equivalent to qR(−λ)q(A)B ⊆ BR(−λ)(A), qRλq(A)B ⊆ BRλ(A). But these relations follow from Proposition 12, now applied to q := −q. (cid:3) 5. Characterizations of Classes C0 and C1 by Weak Resolvent Identities Recall that A and B always denote self-adjoint operators acting on a Hilbert space H and that the linear subspace Dq(A, B) was defined in Definition 1. Let {A, B} be a pair of class C0. If λ > 0, then −λ is not in the sector S(q)+, so the weak resolvent identity qR−λq(A)B ⊆ BR−λ(A) holds by Theorem 20(i). Similarly, if µ ∈ Ri, µ 6= 0, and θ < π 2 , then µ /∈ S(q)+ and hence qRµq(A)B ⊆ BRµ(A). The following two theorems state some converses of these assertions. Theorem 21. Let A is a positive operator such that ker A = {0}. Suppose that 0 < θ0 < π and the domain Dq(A, B) is a core for B. Assume that qR−λq(A)B ⊆ BR−λ(A) for λ > 0. Then the pair {A, B} is an orthogonal direct sum of a trivial representation {A2, 0} and a pair {A0, B0} of the class C0. Proof. Let f ∈ Dq(A, B). Cleary, the positive self-adjoint operator A and the nor- mal operator qA (because of q = e−iθ0 with θ0 < π) satisfy the assumptions of Proposition 13. By the definition of Dq(A, B), the vectors f and Bf are in D(A), so formula (24) applies to the operator T = A and the vector Bf and also to the operator T = qA and the vector f . The assumptions qR−λq(A)B ⊆ BR−λ(A) and ABf = qBAf imply that (qA + λI)−1(qA)Bf = B(A + λI)−1Af. Next we apply Proposition 13 to the operator T = qA. Since q = eiθ0 with θ0 < π, the assumptions of Proposition 13 are fulfilled. Interchanging the closed operator B A Resolvent Approach to the Real Quantum Plane 13 and the integral in formula (24) (by considering the integral as a limit of H-valued Riemann sums) we therefore obtain (34) (qA)ε+itBf = BAε+itf for f ∈ Dq(A, B), t ∈ R. By the first equality in (24) and the relation Aε+itf = AεAitf we have (qA)ε+itBf = (eiθ0A)ε+itBf = eiεθ0e−θ0t AitAεBf, so by (34) we obtain eiεθ0e−θ0t AitAεBf = BAεAitf. (35) Recall that f ∈ D(A) and AitBf ∈ D(A) by the assumption f ∈ Dq(A, B). Passing to the limit ε → +0 in (35) by using Lemma 14 and the fact that the operator B is closed it follows that AitBf = eθ0tB Aitf for all f ∈ Dq(A, B). Since Dq(A, B) is a core for B by assumption, we conclude that AitB ⊆ eθtBAit for all t ∈ R. The latter implies that Ait leaves the closed subspace H2 := ker B invariant. Hence H2 is reducing for A and B, so we have B = 0⊕ B0 and A = A2 ⊕ A0 on H = H2 ⊕H⊥ such that ker B0 = {0} and (A0)itB0 ⊆ eθ0tB0(A0)it for t ∈ R. Since A ≥ 0 and ker A = {0}, the pair {A0, B0} belongs to C0. Theorem 22. Suppose that 0 < θ0 < π Dq2 (A2, B) is a core for B and 2 and ker A = {0}. Assume that 2 (cid:3) qRµiq(A)B ⊆ BRµi(A) for µ ∈ R, µ 6= 0. Then {A, B} is an orthogonal sum of a trivial representation {A2, 0} and a pair {A0, B0} such that {A0, B0} belongs to the class C0. If in addition Dq(A, B) is a core for B, then {A0, B0} is a pair of the class C0. Proof. From the relation qRµiq(A)B ⊆ BRµi(A) it follows that each resolvent Rµi(A) and its adjoint R−µi(A) leaves the closed linear subspace H2 := ker B invariant. This implies that H2 is a reducing subspace for Rµi(A) and therefore for the operator A. Obviously, H2 reduces B. Hence the pair {A, B} is an orthogonal sum of a trivial representation {A2, 0} and a pair {A0, B0} such that ker B0 = {0}. For notational simplicity let us assume already that ker B = {0}. Our aim is to prove that {A, B} is in C0. First we recall a simple operator-theoretic fact: If T is a closed operator such that ν,−ν ∈ ρ(T ), then ν2 ∈ ρ(T 2) and (36) Rν2 (T 2) = Rν(T )R−ν(T ) = 1 2ν (Rν (T ) − R−ν(T )). Suppose now that λ > 0. Putting µ = √λ, we have (iµ)2 = −λ. Let f ∈ D(B). Using the identity (36) twice, for ν = µiq and for ν = µi, and the assumptions qR±µiq(A)B ⊆ BR±µi(A), we obtain q2R−λq2 (A2)Bf = (37) = q2 2µiq 1 2µi (Rµiq(A) − R−µiq(A))Bf B(Rµi(A) − R−µi(A))f = BR−λ(A2)f. Thus, since q2 = e−2iθ0, 2θ0 < π and Dq2 (A2, B) is a core for B by assumption, the pair {A2, B} satisfies all assumptions of Theorem 21 for the relation A2B = q2BA2. Therefore, by this theorem we have (A2)itB ⊆ e2θ0tB(A2)it for t ∈ R, so that AisB ⊆ eθ0sBAis for s ∈ R. Hence the pair {A, B} belongs to C0 (see e.g. 14 VASYL OSTROVSKYI, KONRAD SCHM UDGEN Remark 2 in Section 4). Further, if in addition Dq(A, B) is a core for B, it follows from Proposition 19 that {A, B} is in the class C0. (cid:3) The next theorem contains a characterization of the class C1. Recall that for the class C1 the weak resolvent relation (31) holds for λ ∈ C\S(−q) by Theorem 20(ii). Theorem 23. Suppose that 0 < θ0 < π and ker A = ker B = {0}. Assume that both domains Dq2 (A2, B) and Dq(A, B) are cores for the operator B and that there exists a number p ∈ C\S(−q) such that qRµpq(A)B ⊆ BRµp(A) for all µ ∈ R, µ 6= 0. Then the pair {A, B} belongs to the class C1. Proof. Without loss of generality we can choose the number p ∈ C\S(−q) of modu- lus one and contained in the open sector with angle less than π between the positive x-axis and the half-line through the origin and q. We modify some arguments that have been used already in the proofs of Theorems 21 and 22. −¯q −q S(−q) q p ¯q S(−q) q ¯q S(−q) −¯q S(−q) p −q Figure 2. The S(−q) sector and admissible ranges for p as 0 < arg q < π 2 < arg q < π 2 and as π Using the identity (36) the assumption qRµpq(A)B ⊆ BRµp(A) implies that q2Rλp2q2 (A2)B ⊆ BRλp2 (A2), λ > 0. (38) Let f ∈ Dq2 (A2, B). From (38) and the relation A2Bf = q2BA2f we derive (39) (−p2q2A2 + λI)−1(−p2q2A2)Bf = B(−p2A2 + λI)−1(−p2A2)f, λ > 0. Our aim is to apply Lemma 13, especially formula (24) therein, to the operators −p2q2A2 and −p2A2. To fullfill the assumptions of Lemma 13 it is crucial to write the number −p2q2 and −p2 (of modulus one) in the form eiϑ with ϑ ∈ (−π, π). Let us write p as p = eiψ with 0 < ψ < π and let s(ψ) denote the sign of ψ. Note that q = e−iθ and θ < π. Also we recall that by definition we have θ1 = θ − π if θ > 0 and θ1 = θ + π if θ < 0. We shall prove that (40) (41) −p2 = ei(−2ψ+s(ψ)π) −p2q2 = ei(2θ1−2ψ+s(ψ)π) and − 2ψ + s(ψ)π ∈ (−π, π), and 2θ1 − 2ψ + s(ψ)π ∈ (−π, π). First suppose that ψ > 0. Then π > θ > ψ > 0 by the choice of the number p. Hence −2ψ + π ∈ (−π, π) and −p2 = −(e−iψ)2 = ei(−2ψ+π). Further, 2θ1 − 2ψ + π = 2(θ − π) − 2ψ + π = 2(θ − ψ) − π ∈ (−π, π) and −p2q2 = −(e−iψ)2(eiθ)2 = ei(2θ−2ψ−π) = ei(2θ1−2ψ+π). A Resolvent Approach to the Real Quantum Plane 15 Next we treat the case ψ < 0. Then 0 > ψ > θ > −π by the definition of p. Therefore, −2ψ − π ∈ (−π, π) and −p2 = −(e−iψ)2 = ei(−2ψ−π). Moreover, 2θ1 − 2ψ − π = 2(θ + π) − 2ψ − π = 2(θ − ψ) + π ∈ (−π, π) and −p2q2 = −(e−iψ)2(eiθ)2 = ei(2θ−2ψ+π) = ei(2θ1−2ψ−π). This proves (40) and (41) in both cases. By (40) and (41) it follows from the first equality of formula (24) that (42) (−p2A2)ε+itg = ei(−2ψ+s(ψ)π)ε e−(−2ψ+s(ψ)π)t (A2)ε+itg, (−p2q2A2)ε+itg = ei(2θ1−2ψ+s(ψ)π)ε e−(2θ1−2ψ+s(ψ)π)t (A2)ε+itg (43) for any g ∈ D(A2). Since f, Bf ∈ D(A2), the second equality of formula (24) yields (44) (−p2q2A2)ε+itBf = B(−p2A2)ε+itf. e−2θ1t(A2)itBf = B(A2)itf, t ∈ R, Inserting (42) with g = f and (43) with g = Bf into (44) and passing to the limit ε → +0 we obtain and hence AisBf = eθ1sBAisf , s ∈ R, for all f ∈ Dq2 (A2, B). Since Dq2 (A2, B) is a core for B by assumption, the latter implies that AisB ⊆ eθ1sBAis for s ∈ R. Therefore, since we also assumed that Dq(A, B) is a core for B, the assumptions of Proposition 19 are fulfilled with k = 1, so the pair {A, B} belongs to the class C1. (cid:3) Related characterizations of the class C1 can be also given by requiring the weak resolvent relation for A rather than A. The following theorem is a sample of such a result. Theorem 24. Suppose that 0 < θ0 < π and ker A = ker B = {0}. Suppose that the domains Dq(A, B) and D−q(A, B) are cores for B and −qRλq(A)B ⊆ BR−λ(A) for λ > 0. Then the pair {A, B} is in C1. Proof. The assumptions of Theorem 24 imply that the pair {A, B} satisfies the assumptions of Theorem 21 for the relation (2) with q = e−iθ0 replaced by −q = e−iθ1. Note that 0 < θ1 < π, since we assumed that 0 < θ0 < π. Therefore, by Theorem 21, we have AitB ⊆ eθ1tBAit for all t ∈ R. Since Dq(A, B) is a core for B, it follows from Proposition 19 that the pair {A, B} is in C1. (cid:3) The crucial assumption in the preceding Theorems 21, 22, and 23 is that the weak A-resolvent identity (9) holds on some line that intersects the critical sector only at the origin. In addition there have been technical assumptions such as ker A = {0} and the requirement that Dq(A, B) resp. Dq2 (A2, B) is a core for the operator B. These technical assumptions can be avoided if we assume in addition that the weak B-resolvent identity (10) holds for some points of the resolvent set ρ(B). Theorem 25. Let A is a positive operator. Suppose that 0 < θ0 < π and there exist a number ν ∈ ρ(B) such that νq, ν q ∈ ρ(B), (45) q Rνq(B)A ⊆ ARν(B) and q Rν q(B)A ⊆ ARν (B). Assume that (46) qR−λq(A)B ⊆ BR−λ(A) for all λ > 0. 16 VASYL OSTROVSKYI, KONRAD SCHM UDGEN Then the pair {A, B} is an orthogonal direct sum of trivial representations {A2, 0} and {0, B2} and a pair {A0, B0} of the class C0. Proof. From (45) it follows that the operator Rν (B) and its adjoint Rν (B) leave the closed subspace G2 := ker A invariant. Therefore, G2 is reducing for Rν (B) and hence for B. Since G2 ≡ ker A is obviously reducing for A, the pair {A, B} decomposes as an orthogonal sum of pairs {0, B2} and { A, B} such that ker A = {0}. Further, by Proposition 5(vi), (45) implies that Dq(A, B) is core for B. Hence Dq( A, B) is core for B. Clearly, (46) leads to the relation R−λq( A) B ⊆ BR−λ( A) for λ > 0. Thus, the pair { A, B} satisfies the assumptions of Theorem 21 which gives the assertion. Theorem 26. Suppose that 0 < θ0 < π that νq, ν q, ν q2 ∈ ρ(B) and (47) 2 and there exists a number ν ∈ ρ(B) such for λ = νq , ν, ν. (cid:3) Assume that q Rλq(B)A ⊆ ARλ(B) qRµiq(A)B ⊆ BRµi(A) for all µ ∈ R, µ 6= 0. Then {A, B} is an orthogonal sum of trivial representations and a pair {A0, B0} belonging to the class C0. Proof. Arguing in a similar manner as in the proof of Theorem 25 the assertion is reduced to Theorem 22. We sketch only the necessary modifications: From the relations (47), applied for λ = ν, ν, it follows that ker A is reducing for the pair {A, B}. By Proposition 5(vi) and Corollary 7, the relations (47), applied with λ = ν, νq, imply that the domains Dq(A, B) and Dq2 (A2, B) are cores for B. (cid:3) Theorem 27. Suppose that π 2 < θ0 < π and there exist numbers ν ∈ ρ(B) and p ∈ C\S(−q) such that νq, ν q, ν q2 ∈ ρ(B), (48) for λ = νq , ν, ν, for all µ ∈ R, µ 6= 0. (50) Then {A, B} is an orthogonal sum of trivial representations and a pair {A1, B1} of the class C1. Proof. As in the proof of Theorem 26 it follows from (48) that ker A is reducing for the pair {A, B} and that Dq(A, B) and Dq2(A2, B) are cores for B. Likewise the relations qRp q(A)B ⊆ BRp(A) and qRpq(A)B ⊆ BRp(A) (by (49) and (50)) imply that ker B is reducing for the pair {A, B}. Using these facts the assertion is derived from Theorem 23. (cid:3) Remarks. By Theorem 20, the weak B-resolvent relation qRµq(B)A ⊆ ARµ(B) is satisfied for a pair {A, B} of the class C0 if µ ∈ C\S(q) and for a pair {A, B} in C1 if µ ∈ C\S(−q). From this result it follows that for the corresponding pairs in Theorems 25 -- 27 the assumptions (45), (47), and (48) can be fulfilled. That is, if {A, B} is a pair of the class C0 we can choose ν ∈ C\S(q) such that ν ∈ C\S(q) (then (45) holds) and if in addition 0 < θ0 < π 2 there exists ν ∈ C\S(q) such that νq, ν ∈ C\S(q) (which implies (47)). If the pair {A, B} is in C1 and π 2 < θ0 < π we can find ν ∈ C\S(−q) such that νq , ν ∈ C\S(−q) (these conditions imply (48)) (49) q Rλq(B)A ⊆ ARλ(B) qRp q(A)B ⊆ BRp(A), qRµpq(A)B ⊆ BRµp(A) A Resolvent Approach to the Real Quantum Plane 17 and there exists p ∈ C\S(−q) such that p ∈ C\S(−q) (then (46) and (50) are satisfied by Theorem 20(ii)). 6. Deficiency subspaces and their dimensions Let A and B be positive self-adjoint operators with trivial kernels acting on a Hilbert space H. Then, by Theorems 20 and 21 and by Proposition 5, the pair {A, B} belongs to the class C0 of representations of equation (2) if and only if the resolvent relations (12) and (13) are satisfied for all λ, µ ∈ C such that λ, µ /∈ S(q)+. Moreover, up to unitary equivalence, the only irreducible such pair is {eαQ, eβP}, where αβ < π and q = e−iαβ; see Section 3. In this section we study the resolvent relations for the pair {eαQ, eβP} in the general case, that is, we do not assume that αβ < π or λ, µ /∈ S(q)+. First let us fix some assumptions and notations that will be kept throughout this section. Suppose that α, β ∈ R and q := e−iαβ 6= ±1. We consider the pair {A := eαQ, B := eβP} of self-adjoint operators on the Hilbert space H := L2(R). Recall that A and B act by (51) (Af )(x) = eαxf (x) and (Bf )(x) = f (x + iβ) for all functions f of the dense domain D0 = Lin {e−εx2+γx; ε > 0, γ ∈ C}. Since (2) is satisfied for f ∈ D0 (by (51)), we know from Section 2 that (4) holds for all h ∈ (B− µI)(A− λI)D0 and that (7) holds for all h ∈ (A− λqI)(B− µqI)D0. Therefore, in order to find the resolvent equations in the present case it suffices to describe the form of resolvent equations on the orthogonal complements of spaces (B − µI)(A− λI)D0 and (A− λqI)(B − µqI)D0, respectively. This will be achieved by the formulas at the end of this section. Definition 4. HA(λ, µ) = {ψ ∈ H : ψ ⊥ (B − µI)(A − λI)D0}, HB(λ, µ) = {η ∈ H : η ⊥ (A − λqI)(B − µqI)D0}. Assume that λ, µ ∈ C r (R+ ∪ ¯qR+), where R+ = [0, +∞), and write λ = er+is, µ = eu+iv, r > 0, u > 0, s < π, v < π. Further, we let so that q = e−iθ0, and set αβ = θ0 + 2πm, m ∈ Z, θ0 ∈ (−π, π), ε1 = sign(v), ε2 = sign(v − θ0), ε3 = sign(s), ε4 = sign(s − θ0). Theorem 28. (i) The vector space HA(λ, µ) is spanned by the functions ψj(x) = ¯µ−ix/βeπ(1+ε1)x/β e2πx/β − ¯λ2π/αβ e−4iπ2j/αβ , where j ∈ Z and 0 < (s − 2πj)/αβ < 1, and its dimension is ε3 − ε4 (cid:17). dimHA(λ, µ) = sign(αβ)(cid:16)m + 2 18 VASYL OSTROVSKYI, KONRAD SCHM UDGEN (ii) The space HB(λ, µ) is the linear span of functions ηk(x) = (¯µ¯q)−ix/βe2πkx/β . eαx − ¯λ¯q where k ∈ Z and 0 < (θ0 + 2πk − v)/αβ < 1, and it has the dimension dimHB(λ, µ) = sign(αβ)(cid:16)m + ε1 − ε2 2 (cid:17). Proof. (i): First we study the space HB(λ, µ). Let η ∈ HB(λ, µ). Applying (51) to φ(x) = e−εx2+itx ∈ D0 we compute ((A − λq)(B − µq)φ)(x) = (eαx − λq)(e−ε(x+iβ)2+it(x+iβ) − µqe−εx2+itx) = (eαx − λq)(e−2iβεx+εβ2−tβ − µq)e−εx2+itx. Therefore, since η ⊥ (A − λq)(B − µq)φ(x), we obtain η(x) (eαx − λq)(e−2iβεx+εβ2−tβ − µq)e−εx2+itx dx 0 = ZR = eεβ2−tβ ZR − µqZR ei(t−2βε)x η(x) (eαx − λq) e−εx2 eitx η(x) (eαx − λq) e−εx2 dx dx which implies that (52) eεβ2−tβgε(t − 2βε) = µq gε(t), gε(t) = ZR eitx η(x) (eαx − λq) e−εx2 dx. Since µ = eu+iv, v < π, it is easily seen that the function Gε(t) = e− t2 4ε − u+i(v−θ0 ) 2εβ t satisfies (52). Therefore, any solution of (52) has the form Gε(t)Hε(t), where Hε is a periodic function on R with period −2εβ, that is, Hε(t − 2εβ) = Hε(t) for t ∈ R. The crucial step of this proof is contained in the following lemma. Lemma 29. Hε is a trigonometric polynomial, that is, there are an integer l ≥ 0 and numbers ck ∈ C, k = −l, . . . , l such that Hε(t) = Pl k=−l dke iπkt εβ . Proof. We have (53) Hε(t) = (Gε(t))−1gε(t) = e t2 4ε + u+i(v−θ0 ) 2εβ tZR eitx η(x) (eαx − λq) e−εx2 dx. Because the above arguments are valid for complex numbers t as well, Hε becomes a periodic function on a whole complex plane C. is also in L2(R) for any τ > 0. Therefore, the integral in (53) is an entire function on the complex plane C. Since η ∈ L2(R), the function eτ xη(x) (eαx − λq) e−εx2 We show that Hε is of exponential type. For any s ∈ R we have Hε(t + is) = e 4ε + u+i(v−θ0) (t+is)2 2εβ t2 −s2 4ε + ut−(v−θ0 )s 2εβ t2 −s2 4ε + ut−(v−θ0 )s 2εβ = e ≤ e ZR (t+is)(cid:12)(cid:12)(cid:12) ei(t+is)x η(x) (eαx − λq) e−εx2 ZR (cid:12)(cid:12)(cid:12) eitx η(x) (eαx − λq) e−εx2−sx dx(cid:12)(cid:12)(cid:12) kηk(cid:16)ZR eαx − λq2 e−2εx2−2sx dx(cid:17)1/2 . dx(cid:12)(cid:12)(cid:12) A Resolvent Approach to the Real Quantum Plane 19 Since ZR eαx − λq2 e−2εx2−2sx dx = r π 2ε e for 0 ≤ t ≤ 2εβ we have s2 2ε(cid:0)λ2 + e α(α−2s) 2ε − (λq + ¯λ¯q)e α(α−4s) 8ε (cid:1), Hε(t + is) ≤ M e− (v−θ0 )s 2εβ kηk(cid:16)r π 2ε(cid:0)λ2 + e α(α−2s) 2ε − (λq + ¯λ¯q)e α(α−4s) 8ε (cid:1)(cid:17)1/2 , t2 4ε + ut 2εβ on the interval [0, 2εβ]. Since Hε(t + is) where M is the supremum of e is periodic in t, the latter estimate holds for all t ∈ R. Hence Hε(t + is) has exponential growth. Thus, Hε is an entire periodic function of exponential type. By a result from complex analysis (see, e.g., [8, p. 334]), such a function Hε has to be a trigonometric polynomial Hε(t) = l Xk=−l cke iπkt βε . (cid:3) By Lemma 29 the equality gε(t) = Gε(t)Hε(t) takes the form ZR eitx η(x) (eαx − λq) e−εx2 dx = e− t2 4ε − u+i(v−θ0 ) 2εβ l t Xk=−l cke iπkt βε . Applying the Fourier transform, we obtain η(x) (eαx − λq) e−εx2 e−itxgε(t) dt 1 = 2π ZR = e−εx2 eixu/βe−(v−θ0)x/β l Xk=−l dke2πkx/β, where Obviously, the factor e−εx2 dk = cke(u+i(v−θ0−2kπ))2/(4εβ2). √ε √π cancels, so we get eixu/βe−(v−θ0)x/β (54) η(x) = eαx − λq Let us introduce the functions l Xk=−l dke 2πkx β = (µq)ix/β eαx − λq l Xk=−l dke2πkx/β. (¯µ¯q)−ix/βe2πkx/β e−iux/βe(θ0−v+2πk)x/β ηk(x) := = eαx − ¯λ¯q . eαx − ¯λ¯q ¯dkηk(x). k=−l Then we have η(x) = Pl Next we want to decide which functions ηk belong to L2(R). Let us begin with the case where α > 0. Then the function ηk(x) behaves like e(θ0−v+2πk)x/β as x → −∞ and like e(θ0−v+2πk−αβ)x/β as x → +∞, so that ηk ∈ L2(R) if and only 0 < (θ0 − v + 2πk)/β < α. In the case where α < 0 a similar reasoning shows that ηk ∈ L2(R) if and only α < (θ0 − v + 2πk)/β < 0. Thus, in both cases we have ηk ∈ L2(R) if and only if if and only if 0 < (θ0 − v + 2πk)/αβ < 1, that is, v − θ0 < 2πk < v + 2πm for αβ > 0, for αβ < 0, v + 2πm < 2πk < v − θ0 20 VASYL OSTROVSKYI, KONRAD SCHM UDGEN or equivalently, 1 + ε2 2 ≤ k ≤ m − 1 − ε1 2 for αβ > 0, m + 1 + ε1 2 ≤ k ≤ ε2 − 1 2 for αβ < 0. The functions ηk are obviously linearly independent. From the asymptotic be- ¯dkηk(x) is in L2(R) if and only each ηk haviour of ηk it follows that η(x) = Pl with nonvanishing coefficient ¯dk is in L2(R). Therefore, we obtain k=−l dimHB = sign(αβ)(cid:16)m + ε1 − ε2 2 (cid:17). (ii): Now we turn to the space HA(λ, µ). Let ψ ∈ L2(R). Recall that ψ ∈ HA(λ, µ) if and only if ψ ⊥ (B − µI)(A − λI)D0 = (eβP − µI)(eαQ − λI)D0 For the Fourier transform F we have eαQ = F e−αPF −1 and eβP = F eβQF −1. Hence the latter is equivalent to ψ ⊥ F (eβQ − µI)(e−αP − λI)F ∗D0. Since D0 is invariant under the Fourier transform, ψ ∈ HA(λ, µ) if and only if the inverse Fourier transform F ∗ψ of ψ satisfies F ∗ψ ⊥ (eβQ − µI)(e−αP − λI)D0. From the proof of (i) we already know that ψ ∈ HA(λ, µ) if and only if F ∗ψ is a linear combination of functions where k ∈ Z and 0 < (s − 2πk)/αβ < 1. The condition on k can be rewritten as φk(x) = (¯λ)ix/αe−2πkx/α , eβx − ¯µ 1 + ε4 2 − m ≤ k ≤ 2 ≤ k ≤ ε4 − 1 1 + ε3 ε3 − 1 2 2 − m for αβ < 0. for αβ > 0, Therefore, we have dimHA = sign(αβ)(cid:16)m + ε3 − ε4 2 (cid:17). To calculate the Fourier transform of φk, we shall apply the following formula (see, e.g., [4, 3.311.9]) (55) ZR e−δx e−x + γ dx = πγδ−1 sin πδ , arg γ < π, 0 < Re δ < 1. After some computations using (55) we obtain F φk(x) = Ckψk(x), where i√2πei(u−iv−iπ(1−ε1))(r−is+i2πk)/αβ Ck = ψk(x) = ¯µβ ¯µ−it/βeπ(1+ε1)t/β (e2πt/β − ¯λ2π/αβe4iπ2k/αβ ) . , (cid:3) The following corollary restates the result on the dimensions of Theorem 28 in an important special case. A Resolvent Approach to the Real Quantum Plane 21 Corollary 30. Suppose that {A, B} is an irreducible nontrivial representation of the class C0 for relation (2). Then dimHA(λ, µ) = 0 dimHB(λ, µ) = 0 for λ /∈ S(q)+, for µ /∈ S(q)+, dimHA(λ, µ) = 1 dim HB(λ, µ) = 1 for λ ∈ S(q)+, for µ ∈ S(q)+. Proof. By Corollary 16, {A, B} is unitarily equivalent to a pair {δ1eαP , δ2eβQ}, where δ1, δ2 ∈ {1,−1} and αβ = θ0, θ0 < π. Hence the assertion follows at once from Theorem 28. (cid:3) We now continue the considerations towards the modified resolvent relations. Proposition 31. (i) R¯µ(B)R¯λ(A) − (ii) R¯λ¯q(A)R¯µ ¯q(B) + 1 ¯q¯λ¯µ(1−¯q) I maps HA(λ, µ) into HB(λ, µ). 1 ¯λ¯µ(1−¯q) I maps HB(λ, µ) into HA(λ, µ). Proof. (i): Indeed, for φ ∈ D0 we have (B − µI)(A − λI)φ = ¯q(A − λqI)(B − µqI)φ + λµ(1 − q)φ. Therefore, for η ∈ HB(λ, µ) we derive 1 (cid:10)(cid:0)R¯µ(B)R¯λ(A) − ¯λ¯µ(1 − ¯q) = hR¯µ(B)R¯λ(A)η, (B − µI)(A − λI)φi − I(cid:1)η, (B − µI)(A − λI)φ(cid:11) ¯λ¯µ(1 − ¯q) hη, (B − µI)(A − λI)φi 1 q = − ¯λ¯µ(1 − ¯q) hη, (A − λqI)(B − µqI)φi = 0, ¯λ¯µ(1−¯q) I(cid:1)η ∈ HA(λ, µ). 1 that is, (cid:0)R¯µ(B)R¯λ(A) − (ii) is proved in a similar manner. From Proposition 31 it follows that the operator (56) (cid:16)R¯µ(B)R¯λ(A) − 1 ¯λ¯µ(1 − ¯q) I(cid:17)(cid:16)R¯λ¯q(A)R¯µ¯q(B) + 1 ¯q¯λ¯µ(1 − ¯q) maps the subspace HA(λ, µ) into itself and that the operator I(cid:17)(cid:16)R¯µ(B)R¯λ(A) − (cid:16)R¯λ¯q(A)R¯µ¯q(B) + (57) 1 ¯q¯λ¯µ(1 − ¯q) 1 ¯λ¯µ(1 − ¯q) (cid:3) I(cid:17) I(cid:17) maps HB(λ, µ) into itself. In the special case when HA(λ, µ) = HB(λ, µ) = {0} we know from Section 2 that the corresponding operators in (56) and (57) are both equal to −q¯λ−2 ¯µ−2(1 − ¯q)−2 I on the whole Hilbert space. 22 VASYL OSTROVSKYI, KONRAD SCHM UDGEN In the general case some lengthy but straightforward computations using (55) and the relations eαQ = F e−αPF −1, eβP = F eβQF −1, lead to the formulas (cid:16)R¯λ¯q(A)R¯µ¯q(B) + I(cid:17)ψj(x) 1 ¯λ¯µ¯q(1 − ¯q) ¯λ2π(m−(1−ε1)/2)/αβe−4iπ2(m−(1−ε1)/2)j/αβ = m−(1−ε1)/2 Xl=(1+ε2)/2 ¯λ−2πl/αβ e4iπ2lj/αβ ηl(x), ¯λ¯q(1 − ¯q)¯µ 1 (cid:16)R¯µ(B)R¯λ(A) − = −2π¯λ−2π(m−k+(ε1−1)/2)/αβ ¯λ¯µ(1 − ¯q) I(cid:17)ηk(x) αβ¯λ¯µ(1 − ¯q) Therefore, we finally obtain m−(ε4+1)/2 Xj=(1−ε3)/2 e4iπ2j(m−k+(ε1−1)/2)/αβψj(x). 1 1 = = (cid:16)R¯µ(B)R¯λ(A) − 1 ¯λ¯µ(1 − ¯q) −2π αβ¯λ2 ¯µ2 ¯q(1 − ¯q)2 I(cid:17)(cid:16)R¯λ¯q(A)R¯µ¯q(B) + ¯λ¯µ¯q(1 − ¯q) I(cid:17)ψj(x) m−1−(ε2−ε1)/2 m−(ε4+1)/2 Xk=0 Xl=(1−ε3)/2 e−4iπ2kj/αβ e4iπ2lk/αβ ψl(x), (cid:16)R¯λ¯q(A)R¯µ¯q(B) + 1 ¯λ¯µ¯q(1 − ¯q) −2π m−(ε4+1)/2 I(cid:17)(cid:16)R¯µ(B)R¯λ(A) − Xl=(1+ε2)/2 Xj=(1−ε3)/2 m−(1−ε1)/2 αβ¯λ2 ¯µ2 ¯q(1 − ¯q)2 I(cid:17)ηk ¯λ¯µ(1 − ¯q) e−4iπ2j(k−l)/αβ ¯λ2π(k−l)/αβ ηl(x). The preceding two equations are the versions of the resolvent equations for basis elements of the subspaces HA(λ, µ) and HB(λ, µ), respectively. Acknowledgments The results of this paper were obtained during research visits of the first author at Leipzig University supported by the DFG grant SCHM 1009/5-1. Excellent working conditions and warm hospitality are acknowledged. The authors express their gratitude to Prof. Lyudmila Turowska and Prof. Andrew Bakan for helpful discussions of topics related to this research. References [1] Balakrisnan, A.V., Fractional powers of closed operators and the semigroup generated by them, Pacific J. Math. 10(1960),419 -- 437. [2] Brooke, J. A., Busch, P. and Pearson, D. B., Commutativity up to a factor of bounded operators in complex Hilbert space, R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci. 458 (2002), 109 -- 118. [3] Cho, M., Harte, R. and Ota, Sch., Commutativity to within scalars on Banach space, Funct. Anal. Approx. 3 (2011), 69 -- 77. [4] Gradshteyn, I.S. and Ryzhik, I.M. Table of Integrals, Series and Products. 7-th ed. Acad. Press, 2007. [5] Martinez Carracedo, C. and Sanz Alix, M., The Theory of Fractional Powers of Operators, North-Holland, Amsterdam, 2001. [6] Mortad, M.H., Commutativity up to a factor: more results and the unbounded case. Z. Anal. Anwend. 29(2010), 303-307. [7] Klimyk, A. and Schmudgen, K., Quantum Qroups and Their Representations, Springer, 1997. A Resolvent Approach to the Real Quantum Plane 23 [8] A.V.Markushevich, The theory of analytic functions. Vol. II. (Russian) [9] Ostrovskyi, V. and Samoilenko, Yu., Introduction to the Theory of Representations of Finitely Presented ∗-Algebras, Gordon and Breach, London, 1999. [10] Ota, S. and Szafraniec, F.H., Notes on q-normal operators, Studia Math. 165(2004), 295 -- 301. [11] Schmudgen, K.: Operator representations of R2 q, Publ. RIMS Kyoto Univ. 29(1992), 1030 -- 1061. [12] Schmudgen, K.: Integrable representations of R2 q, Xq,γ , and SLq(2, R), Commun. Math. Phys. 159(1994), 27 -- 237. [13] Schmudgen, K.: The quantum quarter plane and the real quantum plane, Inter. J. Math. 13(2002), 279 -- 321. [14] Woronowicz, S. L. and Zakrzewski, S., Quantum 'ax + b' group. Rev. Math. Phys. 14 (2002), 797 -- 828. [15] Yang, J. and Du, H.-P., A note on commutativity up to a factor of bounded operators, Proc. Amer. Math. soc. 132 (2004), 1713 -- 1720. Institute of Mathematics, Ukrainian Academy of Sciences, Tereshchenkivska 3, Kyiv 01601, Ukraine E-mail address: [email protected] Mathematisches Institut, Universitat Leipzig, Augustusplatz 9/10, 04109 Leipzig, Germany E-mail address: [email protected]
1811.06923	1	1811	2018-11-16T17:09:12	Constructing KMS states from infinite-dimensional spectral triples	[ "math.OA", "math.DS", "math.KT", "math.SP" ]	We construct KMS-states from $\mathrm{Li}_1$-summable semifinite spectral triples and show that in several important examples the construction coincides with well-known direct constructions of KMS-states for naturally defined flows. Under further summability assumptions the constructed KMS-state can be computed in terms of Dixmier traces. For closed manifolds, we recover the ordinary Lebesgue integral. For Cuntz-Pimsner algebras with their gauge flow, the construction produces KMS-states from traces on the coefficient algebra and recovers the Laca-Neshveyev correspondence. For a discrete group acting on its Stone-\v{C}ech boundary, we recover the Patterson-Sullivan measures on the Stone-\v{C}ech boundary for a flow defined from the Radon-Nikodym cocycle.	math.OA	math	Constructing KMS states from infinite-dimensional spectral triples Magnus Goffeng∗, Adam Rennie‡, Alexandr Usachev∗ ∗ Department of Mathematical Sciences, Chalmers University of Technology and University of Gothenburg, Gothenburg, Sweden ‡School of Mathematics and Applied Statistics, University of Wollongong, Northfields Ave Wollongong, Australia November 19, 2018 Abstract We construct KMS-states from Li1-summable semifinite spectral triples and show that in several important examples the construction coincides with well-known direct construc- tions of KMS-states for naturally defined flows. Under further summability assumptions the constructed KMS-state can be computed in terms of Dixmier traces. For closed ma- nifolds, we recover the ordinary Lebesgue integral. For Cuntz-Pimsner algebras with their gauge flow, the construction produces KMS-states from traces on the coefficient algebra and recovers the Laca-Neshveyev correspondence. For a discrete group acting on its Stone- Cech boundary, we recover the Patterson-Sullivan measures on the Stone- Cech boundary for a flow defined from the Radon-Nikodym cocycle. Contents 1 Introduction 1.1 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Connection to some earlier work . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Structure of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Preliminaries 2.1.1 2.2 Examples 2.1 Semifinite spectral triples and summability . . . . . . . . . . . . . . . . . . . . Semifinite spectral triples from unbounded Kasparov modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Dirac operators on closed manifolds . . . . . . . . . . . . . . . . . . . . 2.2.2 Graph C ∗-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Cuntz-Pimsner algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.4 Group C ∗-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 3 5 7 7 8 8 8 12 13 14 17 19 24 ∗email: [email protected], [email protected], [email protected] 1 3 KMS states constructed from Li1-summable spectral triples 3.1 The positive part of the spectrum and heat traces . . . . . . . . . . . . . . . . 3.2 To-plitz or not To-plitz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Modular spectral triples and modular index theory . . . . . . . . . . . . . . . . 4 The KMS-state φω and Dixmier traces 5 The KMS-state φω in examples 5.1 Dirac operators on closed manifolds . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Graph C ∗-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Group C ∗-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 KMS-states on Cuntz-Pimsner algebras with their gauge action 6.1 KMS-states on Cuntz-Pimsner algebras from traces on the coefficient algebra . 6.2 The Toeplitz construction vs the Laca-Neshveyev correspondence . . . . . . . . 6.3 Obstructions to bi-Hilbertian bimodule structures . . . . . . . . . . . . . . . . . 26 27 31 37 38 43 43 47 48 50 50 53 57 1 Introduction The construction of the JLO cocycle [26, 36, 37] from θ-summable spectral triples [14] has from the start been closely linked with the idea of KMS states. A θ-summable spectral triple (A, H, D) on a C ∗-algebra A gives rise to a state φ(a) := Tr(ae−D2 ) on A and under suitable conditions this is a KMS-state on the saturation of A by the R-action defined from the wave operators eitD2 . By [37] the JLO-cocycle can be defined starting from this KMS-state. On the other hand, [14] shows that a finitely summable spectral triple (A, H, D) on a C ∗-algebra A defines a tracial state on A. Similar constructions were studied in [63]. The idea since then has been to understand the measure theory associated to θ-summable spectral triples in terms of 'twisted traces', and more specifically KMS states. Indeed this idea was present early in the development, [37]. Two viewpoints make it interesting to study states associated with spectral triples having specified summability degrees: the associated states obstructs summability degrees, and the states provide a notion of measure theory. In this paper we present a construction of KMS states from Li1-summable spectral triples. By definition, a spectral triple (A, H, D) is Li1-summable if and only if e−tD is trace class for t large enough -- a slight strengthening of being θ-summable. It is an important observation that large classes of examples of θ-summable spectral triples are also Li1-summable. For the spectral triple defined from a Dirac operator on a closed manifold, our construction recovers the Lebesgue integral. For Cuntz-Pimsner algebras we also relate our construction to previous work of Laca and Neshveyev [45], and the authors [31, 56]. We also examine spectral triples arising from certain Hilbert space valued cocycles on discrete groups. In the examples we consider, the KMS-states are associated to flows that are well-suited to the geometries. This is usually not the case for the KMS-state φ(a) = Tr(ae−D2 ) associated with a θ-summable spectral triple. It is our hope that our construction provides a more natural approach to the KMS-states appearing in the JLO-cocycle and that in the future it will have a bearing on the index theory of Li1-summable spectral triples. 2 1.1 Main results We now state our main results. All our results make sense for general semifinite spectral triples, and so we fix a semifinite trace T for this discussion. First, we state the main technical construction of KMS-states from Li1-summable spectral triples. After that, we state the implications of this construction to more specific examples. We use the notation PD for the non-negative spectral projection of D, i.e. PD := χ[0,∞)(D). If for some βD ≥ 0, T(PDe−tD) is finite for t > βD and diverges as t ց βD, we say that D has positive T-essential spectrum. We define the C ∗-algebra AD as the saturation of A under the action of the wave group eitD, that is AD := C ∗ (∪t∈Rσt(A)) , where σt(a) := eitDae−itD. At this stage, we formulate our results in terms of AD. In Subsection 3.2 we refine the con- struction to a smaller C ∗-algebra. In examples, the construction often applies to A directly. Recall from [4, Definition 5.3.1] that a state φ on an R-C ∗-algebra σ : R y A is said to be KMS at inverse temperature β if φ(ab) = φ(σ−iβ(b)a) for a, b from an R-invariant norm dense ∗-subalgebra of A. If φ is a state on an R-von Neumann algebra σ : R y A we say that it is KMS if the same condition holds on an R-invariant σ-weakly dense ∗-subalgebra of A. The following theorem is the main result of the paper. Theorem 1. Let (A, H, D, N, T) be a unital Li1-summable semifinite spectral triple such that D has positive T-essential spectrum (see Definition 3.1 on page 27) and is β-analytic (see Definition 3.18 on page 33). Define βD := inf{t > 0 : T(PDe−tD) < ∞}. For any extended limit ω ∈ L∞(βD,∞)∗ as t → βD (see Definition 3.14 on page 32), we define the state φω on AD as φω(a) := lim t→ω T(PDae−tD) T(PDe−tD) . Then φω is a KMS-state at inverse temperature βD for the R-action defined from σt. particular, if βD = 0 then φω is a tracial state on A. If βD = 0, and there is a decreasing function ψ : [0,∞) → (0,∞) with regular variation of index −1, satisfying the conditions (4.2) and (4.3), and for some d > 0 we have that µT(t, PDD) ∼ ψ(t)−1/d as t → ∞, then for any exponentiation invariant extended limit ω as t → ∞, In φω(a) = Tω,ψ(PDa(1 + D2)−d/2), where ω is an extended limit as t → 0 defined in Theorem 4.9 (see page 41), and Tω,ψ is the Dixmier trace defined from T and ω on the weak ideal Lψ(N) := {T ∈ KN : µT(t, T ) = O(ψ(t))}. The first part of this result can be found as Corollary 3.21 (see page 34) in the body of the text and the second part as Corollary 4.10 (see page 42). Remark 1.1. If βD = 0, any unital Li1-summable semifinite spectral triple with T(PD) = ∞ has positive T-essential spectrum and is β-analytic. Therefore, Theorem 1 shows that any unital Li1-summable semifinite spectral triple (A, H, D, N, T) with T(PDe−tD) < ∞ for t > 0 and T(PD) = ∞ gives rise to a tracial state on A. This extends a result of Voiculescu [63, Proposition 4.6]. For details on this case, see Theorem 3.22 (see page 34). 3 The following three results compute the KMS-state in specific examples. Theorem 2. Let M be a closed Riemannian manifold, A := C ∞(M ), D be a Dirac operator on a Clifford bundle S → M and H := L2(M, S). Then the KMS-state φω constructed in Theorem 1 is independent of ω and is a tracial state on C(M ) that takes the form φω(a) =ZM− a dV, where dV denotes the volume measure defined from the Riemannian metric on M and Z− the normalized integral. This result appears as Theorem 5.1 (see page 43) in the body of the text. Theorem 3. Let A be a unital C ∗-algebra, E be a strictly W-regular fgp bi-Hilbertian bimodule (see Definitions 2.21 and 2.25 on pages 19 and 21, respectively) and (OE, ΞA, D) the associated unbounded (OE, A)-cycle as in [31]. If τ is a positive trace on A, then the semifinite spectral triple (OE, ΞA ⊗A L2(A, τ ), D ⊗ 1A, (End∗ Moreover, if τ is critical for E (see Definition 6.2 on page 51), the assumptions in Theorem 1 are satisfied and the state φω is KMS for the gauge action on OE. If τ satisfies the Laca- Neshveyev condition for α ≥ 0 (see Definition 6.7 on page 53), then φω is independent of ω and takes the form φω = φLN,τ where φLN,τ is the KMS-state defined from τ via the Laca-Neshveyev correspondence. A(ΞA) ⊗ 1)′′, Trτ ) is Li1-summable. This result is found in Section 6 (starting on page 50). We also discuss extensions of these results to more general A−A-correspondences in Subsection 6.3 (starting on page 57) dispensing the assumption of strict W-regularity. Theorem 4. Let Γ be a discrete group and c : Γ → H0 a Hilbert space valued proper 1- cocycle defining a length function of at most exponential growth. The semifinite spectral triple (A, H, D, N, T) constructed from c in Subsection 2.2.4 is an Li1-summable semifinite spectral triple on Cb(Γ) ⋊ Γ. Moreover, if c is critical (see Definition 3.10 on page 30) the assumptions of Theorem 1 are satisfied and the associated KMS-state φω on C(∂SCΓ) ⋊ Γ is given by where µω is a quasi-invariant Patterson-Sullivan measure on the Stone-Cech boundary ∂SCΓ. The state φω extends to a KMS-state on the von Neumann algebra L∞(∂SC Γ, µω)⋊Γ where it is KMS with inverse temperature 1 for the R-action defined from the Radon-Nikodym cocycle φωXg∈Γ agλg =Z∂Γ ae dµω, σtXg∈Γ agλg :=Xg∈Γ(cid:18) dg∗µ dµ (cid:19)it agλg. This result appears as Theorem 5.10 (see page 49) below. Our method extends to proper quasi-cocycles, and as such would allow for the construction of KMS-states from semifinite spectral triples with possible K-homological content on a-TT-menable groups. Remark 1.2. We will prove that the spectral triple of a length function (which is K-homologically trivial) gives rise to the same KMS state as that appearing in Theorem 4. 4 1.2 Connection to some earlier work Here we show how our approach relates to some results obtained by Connes in [15, Section IV.8.α, Theorem 4]. Connes proves that θ-summable Fredholm modules can be lifted to θ- summable spectral triples. We show that Connes' result can be extended to Lis-summability for 0 < s ≤ 1, and discuss obstructions to summability properties of K-homology classes. For terminology and notations concerning summability and operator ideals, the reader is referred forward to Subsection 2.1. Recall [17, 6] that a semifinite Fredholm module is a collection (A, H, F, N, T) where A acts on the Hilbert space H by operators from N and F ∈ N is an operator with a(F − F ∗), a(F 2 − 1), [F, a] ∈ KT for all a ∈ A. We say that (A, H, F, N, T) is unital if A acts unitally. A unital semifinite Fredholm is said to be Lis-summable if [F, a] ∈ Lis(T) for all a ∈ A and F 2 − 1, F − F ∗ ∈ Li2s(T). If the same conditions holds with Lis(T) replaced by Lp(T), and Li2s(T) by Lp/2(T), we say that (A, H, F, N, T) is p-summable. If (A, H, F, N, T) is a semifinite Fredholm module we say that a semifinite spectral triple (A, H, D, N, T) is a lift if F −sign(D) ∈ KT. Theorem 5. Let s ∈ (0, 1] and (A, H, F, N, T) be a unital semifinite Lis-summable Fredholm module with F 2 = 1 and F = F ∗. Assume that A is countably generated. Then there is a self-adjoint operator D affiliated with N making (A, H, D, N, T) into a unital semifinite Lis- summable spectral triple with Moreover, (A, H, D, N, T) satisfies that a Dom(D1/s) ⊆ Dom(D1/s) and [D1/s, a] has a bounded extension for all a ∈ A. F = FD := DD−1. This theorem is found in [15, Section IV.8.α, Theorem 4] in the special case s = 1/2 and N = B(H). We will not give the full details of the proof in the general case, but merely indicate how Connes' proof extends. The starting point of Connes' proof is a reduction to the case that A contains F and is generated by a countable group of unitaries Γ generated by a countable set of unitaries (uµ)µ∈N. This argument extends to a general von Neumann algebra N. Connes introduces the operator G :=Xµ∈N [F, uµ]∗[F, uµ] 2µk[F, uµ]∗[F, uµ]kLi1 . Since [F, uµ] ∈ Li1/2 for all µ, the series converges in Li1. The proof proceeds by using an average procedure Θ over the group Γ applied to G and Connes proves that D := F Θ(G)−1/2 fulfils the statement of the theorem. For general s ∈ (0, 1], the proof goes mutatis mutandis using the operator Gs :=Xµ∈N ([F, uµ]∗[F, uµ])2s 2µk([F, uµ]∗[F, uµ])2skLi1 ∈ Li1(T), and setting D := F Θ(G)−s. In the special case s = 1, we obtain that (A, H, F, N, T) lifts to a unital semifinite Lis-summable for all a ∈ A the commutator spectral triple (A, H, D, N, T) which is Lipschitz regular, i.e. [D, a] is bounded. 5 The lifting theorem for Lis-summable spectral triples (Theorem 5) stands in sharp contrast to the finitely summable setup, or even the Li(0),s-summable setup. The two upcoming theorems show that a statement as in Theorem 5 could not extend to the ideal Li(0),1. Theorem 6. Let A be a unital C ∗-algebra with no tracial states and (A, H, D, N, T) be a unital semifinite spectral triple on A defining a non-trivial class in KK1(A, KT). Then PD(i± D)−1 /∈ Li(0),1(T). Proof. Consider a unital C ∗-algebra A and an Li(0),1-summable unital semifinite spectral triple (A, H, D, N, T) on A defining a non-trivial class in KK1(A, KT). In particular, T(PD) = ∞; otherwise PD ∈ KT which contradicts the non-triviality of the KK1(A, KT)-class defined by (A, H, D, N, T). By Remark 1.1 (see page 3) all assumptions of Theorem 1 reduces to T(PD) = ∞ in the Li(0),1-summable case. Therefore, the existence of Li(0),1-summable unital semifinite spectral triples on A being non-trivial in KK implies that A admits a tracial state. This argument shows that if A admits no tracial states, it admits no Li(0),1-summable unital semifinite spectral triple. In fact, a careful inspection of the results used show that as soon as there is a unital semifinite spectral triple on A with PD(i ± D)−1 ∈ Li(0),1(T), there is an associated tracial state on A. The theorem follows. There are several C ∗-algebras carrying no traces, for instance any purely infinite C ∗-algebra. Using Theorems 1 and 6, we will give an example of a finitely summable Fredholm module that can not lift to an Li(0),s-summable spectral triple. In particular, lifting of finite summability and Li(0),s-summability fails in general. Theorem 7. There is a C ∗-algebra A with K 1(A) 6= 0, such that for a dense ∗-subalgebra A ⊆ A we can represent any x ∈ K 1(A) by a Fredholm module (A, Hx, Fx) with F 2 x = 1, F ∗ x = Fx and for any a ∈ A, [Fx, a] is of finite rank. Moreover, any lift (A, Hx, Dx) of (A, Hx, Fx) will satisfy that (1 + D2 x)−1/2 /∈ Li(0),1(H). Proof. Consider the Cuntz algebra A = ON and A the ∗-algebra generated by isometries S1, S2, . . . , SN ∈ ON with orthogonal ranges. It is a well known fact that K 1(ON ) ∼= Z/(N − 1)Z 6= 0. By [29], we can represent the generator of K 1(ON ) ∼= Z/(N − 1)Z by the Fredholm module (A, L2(ON , φ), 2P − 1) where φ is the KMS-state on ON and P is the orthogonal projection onto the closed linear span of Sµ, where µ ranges over all finite words on the alpabet {1, . . . , N}. By the results of [29, Section 2.2], [2P − 1, a] = 2[P, a] is finite rank for all a ∈ A. The first statement of the theorem follows. There are no tracial states on ON since 1 N − 1 NXj=1 S∗ j Sj − NXj=1 SjS∗ j = 1 N − 1 [S∗ j , Sj]. NXj=1 1ON = 1 N − 1 (N − 1)1ON = We can now deduce the second statement of the theorem from Theorem 6. Remark 1.3. It is not of importance that K 1(ON ) is torsion for the argument in Theorem 7 to work. In [29], non-torsion examples satisfying the conclusions of Theorem 7 can be found. The reader should also note that the proof of Theorem 7 obstructs all lifts (A, Hx, Dx) of (A, Hx, Fx) with PDx (1 + D2 x)−1/2 ∈ Li(0),1(H). 6 In the nonunital case, the techniques of [5] will likely be required. The substantial technical considerations in the nonunital case goes beyond this paper, and is left to future work. 1.3 Structure of the paper Section 2 recalls the basics of (unital) semifinite spectral triples and their summability. We also recall our main examples from the literature in this section for later use. Section 3 presents our construction of KMS states from Li1-summable spectral triples. We close the section by discussing connections to modular spectral triples. We consider the case βD = 0 in Section 4 and compute the tracial states constructed in Section 3 by means of Dixmier traces. In Section 5 we apply the techniques of Section 3 to the examples. The final Section 6 examines the construction of KMS states for Cuntz-Pimsner algebras. In this case we apply our ideas to derive obstructions to the existence of fgp bi-Hilbertian bimodule structures compatible with the underlying correspondence of the Cuntz-Pimsner algebra. 1.4 Notations A(X) N T KN End∗ KA(X) A A′ A AD PD := χ[0,∞)(D) FD := 2PD − 1 µT(·, T ) nT(·, T ) Lis(T), Li(0),s(T) Lψ(T), L(0),ψ(T) Tω,ψ(T ) T(A) L∞(a,∞), a ≥ 0 C0(a,∞), a ≥ 0 limt→ω f (t) ℓ∞(N) limk→ω xk Lg f ∼ g semifinite von Neumann algebra positive, faithful, normal, semifinite trace on N ideal of T-compact operators C ∗-algebra of adjointable endomorphisms of an A-Hilbert C ∗-module X C ∗-algebra of compact endomorphisms of an A-Hilbert C ∗-module X ∗-algebra the commutant of an algebra A C ∗-closure of an algebra A saturation of A under the action of the wave group eitD non-negative spectral projection of an operator D singular values function of an operator T affiliated with N distribution function of an operator T affiliated with N ideals of compact operators in Definition 2.6 on page 9 ideals of compact operators in Definition 2.6 on page 9 Dixmier trace on Lψ(T) set of positive traces on a unital C ∗-algebra A space of essentially bounded functions on (a,∞) equipped with the essential supremum norm subspace of L∞(a,∞) of all continuous functions vanishing at infinity value of an extended limit ω on a function f space of bounded sequences equipped with the supremum norm value of an extended limit ω on a sequence x transfer operator defined by formula (2.13) on page 23 for two functions or sequences f and g if f = g + o(f ) and g = f + o(g) 7 1.5 Acknowledgements A. R. thanks the Gothenburg Centre for Advanced Studies in Science and Technology for funding and the University of Gothenburg and Chalmers University of Technology for their hospitality in 2017 when this work was begun. M. G. and A. U. were supported by the Swedish Research Council Grant 2015-00137 and Marie Sklodowska Curie Actions, Cofund, Project INCA 600398. The authors acknowledge the support of the Erwin Schrodinger Institute where part of this work was conducted. The authors are grateful to Alan Carey, Heath Emerson and Bram Mesland for inspiring discussions. We also thank Branimir ´Ca´ci´c for sharing his construction of semifinite spectral triples from proper group cocycles, and Edward McDonald for references on previous work in that direction. 2 Preliminaries Before entering into the body of the paper, we recall some basic definitions that we will require and provide some examples that motivated this work. These examples will be studied further in the later sections of the paper. The results will be formulated for semifinite spectral triples. We do however remark that there are several examples of 'vanilla' spectral triples that will be used throughout the paper. 2.1 Semifinite spectral triples and summability To set the stage for the paper, we summarize the basic definitions and properties of semifinite spectral triples. The reader familiar with semifinite spectral triples and symmetrically normed operator ideals can skip this subsection. We let N denote a semifinite von Neumann algebra and we fix a positive, faithful, normal, semifinite trace T on N. The T-compact operators are denoted by KN. The C ∗-algebra KN can be defined as the norm closed ideal generated by the projections E ∈ N with T(E) < ∞. Equivalently, one can define KN := {T ∈ N : µT(t, T ) = o(1) as t → ∞} where the singular value function µT(t, T ) is defined as µT(t, T ) := inf(cid:8)kT (1 − E)kN : where E ∈ N is a projection with T(E) ≤ t(cid:9). Definition 2.1. A semifinite spectral triple (A, H, D, N, T) consists of (2.1) • A ∗-algebra A represented on a Hilbert space H as operators in N ⊆ B(H), that is, we have a specified ∗-homomorphism π : A → N. • A densely defined self-adjoint operator D : dom D ⊂ H → H which is affiliated with N such that for all a ∈ A we have a · domD ⊂ domD and 1. [D, π(a)] := Dπ(a) − π(a)D initially defined on dom(D) is bounded in operator norm. 2. π(a)(1 + D2)−1/2 ∈ KN. 8 Remark 2.2. Sometimes we write a spectral triple as a collection of three objects (A, H, D). In this case, it is implicitly assumed that N = B(L2(M, S)) and T is the standard trace. Remark 2.3. If in addition to the data (A, H, D, N, T) we have specified an operator γ ∈ B(H) with γ = γ∗, γ2 = 1, Dγ + γD = 0 on Dom(D), and for all a ∈ A we have γπ(a) = π(a)γ, we call the semifinite spectral triple even, or sometimes graded. If γ has not been specified, we say that the semifinite spectral triple is odd, or ungraded. This distinction plays an important role in the topological properties of the spectral triple, but since this paper deals with measure theory it will not play a role in this paper. Remark 2.4. We will nearly always dispense with the representation π, treating A as a subal- gebra of N ⊆ B(H). Remark 2.5. In the sequel we assume that the algebra A is unital and that 1 ∈ A acts as In particular, the operator (1 + D2)−1/2 is a T-compact the identity of the Hilbert space. operator. To emphasize this assumption, we refer to the data (A, H, D, N, T) as a unital semifinite spectral triple. Examples of semifinite spectral triples often satisfy a finer summability structure, i.e. a re- finement of the condition (1 + D2)−1/2 ∈ KN. We formulate such conditions in terms of symmetrically quasi-normed operator ideals. We will use the Schatten ideals, the Li-ideals and more generally weak ideals. Definition 2.6. Let N denote a semifinite von Neumann algebra and T a positive, faithful, normal, semifinite trace on N. For parameters p, d ∈ [1,∞) and s > 0 we define the following operator ideals. • Lp(T) := {T ∈ KN : µT(·, T ) ∈ Lp(0,∞)}. • L(d,∞)(T) := {T ∈ KN : µT(·, T ) = O(t−1/d) as t → ∞}. • Lis(T) := {T ∈ KN : µT(t, T ) = O((log(t))−s) as t → ∞}. • Li(0),s(T) := {T ∈ KN : µT(t, T ) = o((log(t))−s) as t → ∞}. • If ψ : [0,∞) → (0,∞) is a decreasing function satisfying that supt>0 the associated weak ideal and its separable subspace Lψ(T) := {T ∈ KN : µT(t, T ) = O(ψ(t))}, L(0),ψ(T) := {T ∈ KN : µT(t, T ) = o(ψ(t))}, ψ(t) ψ(2t) < ∞, we define The condition supt>0 quasi-Banach space in the quasi-norm ψ(2t) < ∞ guarantees that Lψ(T) is a vector space, and in fact even a ψ(t) kTkLψ := sup t>0 µT(t, T ) ψ(t) . Note that Lis(T) = Lψ(T) and Li(0),s(T) = L(0),ψ(T) for ψ(t) := (log(2 + t))−s. It is immediate from the definition that Lp(T) ⊆ Li(0),s(T) for any p and s. More generally, if ψ1, ψ2 : [0,∞) → ψj (2t) < ∞, then Lψ1(T) ⊆ Lψ2(T) (0,∞) are two decreasing functions satisfying that supt>0 as soon as ψ1 = O(ψ2). ψj (t) 9 Remark 2.7. Our definition of symmetrically normed operator ideals in the semifinite setting differs slightly from the standard definition unless N is atomic. In the usual definition, the sym- metrically normed operator ideals are defined from operators affiliated with N that potentially are unbounded. Since we only use bounded operators from these ideals, we have incorporated this fact in our definition. Definition 2.8. Let (A, H, D, N, T) be a unital semifinite spectral triple. • (A, H, D, N, T) is said to be p-summable if (1 + D2)−1/2 ∈ Lp(T). • (A, H, D, N, T) is said to be (d,∞)-summable if (1 + D2)−1/2 ∈ L(d,∞)(T). • (A, H, D, N, T) is said to be Lis-summable if (1 + D2)−1/2 ∈ Lis(T). • (A, H, D, N, T) is said to be Li(0),s-summable if (1 + D2)−1/2 ∈ Li(0),s(T). • (A, H, D, N, T) is said to be ψ-summable if (1 + D2)−1/2 ∈ Lψ(T). The standard terminology in the literature for the special case s = 1/2 is to refer to Li1/2- summability as weak θ-summability and to Li(0),1/2-summability as θ-summability. Since L(d,∞)(T) ⊆ Lp(T) for all p > d, (d,∞)-summability refines p-summability. The notion of (d,∞)-summability is a noncommutative generalization of being d-dimensional as the spectral triple defined from a Dirac operator on a closed d-dimensional manifold (as in Subsection 2.2.1) is (d,∞)-summable. We shall see an abundance of Li1-summable, truly noncommutative, examples where p-summability and (d,∞)-summability fails for all p and d. The notion of ψ-summability generalizes both (d,∞)-summability and Li1-summability, and appears naturally in examples of (semi-) group actions on manifolds (see Subsection 5.1 and [19, 32]). We will make use of this notion in Section 4 where certain conditions on ψ allows one to compute the tracial state defined from a ψ-summable unital semifinite spectral triples in terms of Dixmier traces on Lψ(T). Remark 2.9. It is readily verified that Lis-summability is equivalent to T(e−tD1/s ) < ∞, for t > t0 for some critical value t0, and that Li(0),s-summability is equivalent to T(e−tD1/s ) < ∞, for t > 0. In particular, (A, H, D, N, T) is θ-summable if and only if T(e−tD2 ) < ∞, for all t > 0. Historically, θ-summability has been studied more in depth than Li1-summability. This can in part be explained from the two facts that the JLO-cocycle only requires θ-summability and classically, the heat operator e−tD2 is geometrically more interesting than e−tD to study on a manifold. The two operators e−tD2 and e−tD can be compared by explicit integral formulas, see [28, Chapter 4]. We will exploit the observation that large classes of examples of θ-summable spectral triples are also Li1-summable. 10 In the bulk of the paper, we are interested in computing asymptotics of heat traces of the form T(Be−tD1/s ) for B ∈ N as t approaches a critical value. When (A, H, D, N, T) is Li(0),s- summable, the critical value of t is 0, and in several classical examples (e.g. on closed manifolds) the heat trace T(Be−tD1/s ) admits an asymptotic expansion. The following result is useful for relating heat trace asymptotics to zeta function asymptotics in the case of the nice behaviour appearing when t0 = 0. Theorem 2.10. Let s ∈ (0, 1]. Let (A, H, D, N, T) be an Li(0),s-summable semifinite spectral triple and B ∈ N. The following are equivalent. 1. There are constants pheat > 0, ǫ > 0 and cheat T(Be−tD1/s ) = cheat B t−spheat B ∈ C such that + O(t−spheat+ǫ), as t → 0. 2. The ζ-function ζ(z; B,D1/s) := T(BD−z/s) is well-defined for large Re(z) and there B ∈ C and a function f = f (z) holomorphic in the region are constants pζ > 0, ǫ′ > 0, cζ Re(z) > spζ − ǫ′ such that ζ(z; B,D1/s) = 1 Γ(spζ) cζ B z − spζ + f (z). In this case, pheat = pζ and cheat it holds for all s ∈ (0, 1]. If B = 1 and either of the conditions above hold, then (A, H, D, N, T) is pζ-summable. B. Moreover, if the conditions above hold for one s ∈ (0, 1], B = cζ The proof of Theorem 2.10 follows by noting that Γ(z)ζ(z; B,D1/s) is the Mellin transform of T(Be−tD1/s ) and using [33, Proposition 5.1]. Recall that we use the notation PD := χ[0,∞)(D) for the non-negative spectral projection of D. Let us state a fundamental lemma on the commutators of A with the function of D defined by FD := 2PD − 1. Note that FD differs from the phase DD−1 by the T-finite kernel projection of D. Lemma 2.11. Let (A, H, D, N, T) be a semifinite spectral triple. Then for any a ∈ A [FD, a] ∈ KN. Moreover if (A, H, D, N, T) is unital, then if (A, H, D, N, T) is p-summable, then [FD, a] ∈ Lp(T) for all a ∈ A, and if (A, H, D, N, T) is Lis-summable, then [FD, a] ∈ Lis(T) for all a ∈ A. More generally, if (A, H, D, N, T) is ψ-summable then [FD, a] ∈ Lψ(T) for all a ∈ A. Proof. The proof of the operator inequality −k[D, a]kD−1 ≤ [FD, a] ≤ k[D, a]kD−1 for invertible D and a = −a∗ is found in the proof of [59, Proposition 1]. The assertion follows from the definition of p-, Lis- and ψ-summability, resp. In the non-invertible case we replace (A, H, D, N, T) by (cid:18)(cid:18)A 0 0(cid:19) , H ⊕ H, Dµ =(cid:18)D µ µ −D(cid:19) , M2(N), T ⊗ TrM2(cid:19) , 0 (2.2) 11 for µ ∈ [0, 1]. When µ > 0 we are back in the invertible case. In [6, Proposition 2.25] it is shown that for any a ∈ A we have the norm limit [FDµ, a] µ→0 (cid:18)[FD, a] 0 0(cid:19) = lim 0 and so [FD, a] is compact. Indeed, the proof of [6, Proposition 2.25] shows that if (1 + D2)−1/2 is in a symmetrically quasi-normed ideal J of T-compact operators then [FD, a] ∈ J for all a ∈ A. Again the assertion follows from the definitions of summability. 2.1.1 Semifinite spectral triples from unbounded Kasparov modules For several kinds of C ∗-algebras one can capture the noncommutative geometry through an unbounded Kasparov module. This is a bivariant generalization of spectral triples. Localizing an unbounded Kasparov module in a positive trace gives rise to a semifinite spectral triple as in Theorem 2.12 below. Several of the examples in this paper arises in this way. We briefly recall this construction, which has been informally used for some years. Let A and B be unital C ∗-algebras. A unital unbounded (B, A)-Kasparov module is a collection (B, X, D) where • B ⊆ B is a dense ∗-subalgebra, • X is an A-Hilbert C ∗-module carrying a left action of B as adjointable operators, • D is an A-linear, densely defined, self-adjoint, regular operator on X with A-compact resolvent (i ± D)−1 ∈ KA(X) and • for a ∈ B the operator [D, a] is defined on dom(D) and is bounded in the norm on X. If τ is a positive trace on A, we write L2(X, τ ) := X ⊗A L2(A, τ ) where L2(A, τ ) is the GNS-representation associated with τ . For ξ, η ∈ X, we write Θξ,η for the rank one operator Θξ,η(ν) = ξ(ην)A. The von Neumann algebra Nτ (X) := (End∗ A(X)⊗1A)′′ ⊆ B(L2(X, τ )) coincides with the weak closure of the set of operator spanned by {Θξ,η ⊗ 1A : ξ, η ∈ X} and carries a positive, normal, semifinite, faithful trace Trτ characterized by Trτ (Θξ,η ⊗ 1A) := τ ((ηξ)A), see [45, Section 3]. The following theorem also appears in [49]. Theorem 2.12. Let (B, XA, D) be a unital unbounded (B, A)-Kasparov module and τ : A → C a faithful norm densely defined norm lower semicontinuous tracial weight. Then the data (B, L2(X, τ ), D⊗ 1, Nτ (X), Trτ ) defines a semifinite spectral triple. The von Neumann algebra is Nτ (X) = (End∗ A(X)⊗ 1)′′ and Trτ : N → C is the (positive faithful semifinite normal) trace dual to the normal extension of τ to A′′ ⊆ B(L2(A, τ )). Proof. The operator D ⊗ 1 is self-adjoint by [46, Proposition 9.10]. The commutant of Nτ (X) in X ⊗A L2(A, τ ) is the algebra A′′ (acting by right multiplication). Every unitary in A′′ thus preserves the domain of D⊗1 and so D⊗1 is affiliated to Nτ (X). Plainly commutators of D⊗1 with B remain bounded. Since we start with an unbounded Kasparov module, the operator 12 (1 + D2)−1/2 ∈ KA(X). So we can approximate (1 + D2)−1/2 in norm by finite rank operators Pj Θxj,yj and we can take the xj, yj ∈ X. Hence (1 + (D ⊗ 1)2)−1/2 = (1 + D2)−1/2 ⊗ 1 is in the norm closure of the finite trace operators in Nτ (X), and so T-compact. A conceptual viewpoint is that (B, L2(X, τ ), D ⊗ 1A, (End∗ A(X) ⊗ 1A)′′, Trτ ) is a semi-finite refinement of the unbounded Kasparov product of (B, X, D) with the Morita morphism A → KA(X) and the ∗-homomorphism KA(X) → K A(X)⊗1A)′′. There is a close relationship between the semifinite index and the Kasparov product, described in [6, 39]. (End∗ We remark at this stage that there is to date no general theory of symmetrically quasi-normed operator ideals in Hilbert C ∗-modules, and, as such, no satisfactory way of describing summa- bility. In concrete applications, it is possible to circumvent this problem by choosing a frame on X, implicitly using the machinery of [52]. In the examples of most relevance to this paper the spectrum of D is discrete, and we can use the following proposition to study summability. Proposition 2.13. Let (B, X, D) be a unital unbounded (B, A)-Kasparov module, where D has discrete spectrum σ(D) ⊆ R, and τ be a positive trace on A. Set Pλ := χ{λ}(D) ∈ KA(X) for λ ∈ σ(D). Then it holds that 1. (B, L2(X, τ ), D ⊗ 1A, Nτ (X), Trτ ) is Lis-summable if and only if Xλ∈σ(D) e−tλ1/s Trτ (Pλ) < ∞, for t large enough. 2. (B, L2(X, τ ), D ⊗ 1A, Nτ (X), Trτ ) is p-summable if and only if Xλ∈σ(D) (1 + λ2)−p/2 Trτ (Pλ) < ∞. The proof follows from the following formula: Trτ (f (D)) = Xλ∈σ(D) f (λ) Trτ (Pλ), which holds for every positive Borel function f . 2.2 Examples To give some further context before entering into the main construction of this paper, let us recall some well known examples that we will further explore later on in the paper. The focus in our presentation is on Li1-summability and heat traces. We remark that the constructions in this subsection are rather lengthy, and the reader familiar with the literature can at a first read restrict themself to glancing through this subsection. 13 2.2.1 Dirac operators on closed manifolds The prototypical example of a spectral triple arises from Dirac operators on a closed Rieman- if S → M nian manifold M . We can work with a rather general type of Dirac operators: is a Clifford module on M and /D is a first order elliptic operator acting on C ∞(M, S) being symmetric in the L2-inner product and /D2 is a Laplacian type operator1, we say that /D is a Dirac operator. In this case, the closure of /D in its graph norm defines a self-adjoint oper- ator on L2(M, S) that we by an abuse of notation also denote by /D. It is well-known that (C ∞(M ), L2(M, S), /D) is a spectral triple on C ∞(M ). We summarize the main properties of its heat traces in the following proposition. Proposition 2.14. Let M be an n-dimensional Riemannian closed manifold, /D a Dirac op- erator on M , and (C ∞(M ), L2(M, S), /D) the associated spectral triple. This spectral triple is (n,∞)-summable and for any classical zero-th order pseudo-differential operator A on S with principal symbol a ∈ C ∞(S∗M, End(S)) and every s ∈ (0, 1] we have TrL2(M,S)(Ae−t /D1/s ) = Γ(sn + 1)cs nt−snZS∗M TrS(a) dV + O(t−sn+ǫ), as t → 0, for some dimensional constant cn > 0 and ǫ > 0. Here TrS(a) ∈ C ∞(S∗M ) denotes the fibrewise trace of a. The dimensional constant cn is determined by the Weyl law for /D describing the ordered sequence (λk( /D))k∈N of eigenvalues as λk( /D) = 1 cnvol(M )1/nrank(S)1/n k1/n + O(k1/n−ǫ0), for some ǫ0 > 0. The Weyl law is proven in many places, for instance [27]. The general heat trace asymptotics follows from Theorem 2.10 and [27]. We return to this example below in Example 3.5 (see page 28) and Subsection 5.1 (see page 43). Later on in the paper, we will make use of a modification of the spectral triple coming from a Dirac operator that is also compatible with non-isometric semigroup actions. Similar con- structions were previously considered in [19, 32]. Definition 2.15. Let ψ : [0,∞) → (0,∞) be a positive measurable function. • We say that ψ is regularly varying of index ρ if for all λ > 0 lim t→∞ ψ(λt) ψ(t) = λρ. (2.3) • We say that ψ is smoothly regularly varying of index ρ if ψ ∈ C ∞ and for any k ∈ N, tkψ(k)(t) ψ(t) lim t→∞ = ρ(ρ − 1)··· (ρ − k + 1). (2.4) 1I.e. the symbol of /D2 coincides with the Riemannian metric as a function on T ∗M . 14 ψ(2t) < ∞ and there is an associated weak ideal Lψ A regularly varying function satisfies supt>0 as in Definition 2.6. By [32, Lemma 7.1], any smoothly regularly varying function ψ satisfies t ψ(t) = O(ψ(t)(1 + t2)−k/2) so if ψ additionally is bounded, ψ belongs to the Hormander that ∂k class S0. ψ(t) Smooth regular variation is a strengthening of having regular variation -- a condition used below in Section 4 in the context of defining and computing Dixmier traces. See more also in [32]. By [3, Theorem 1.8.2], any regularly varying function asymptotically behaves like a smoothly regularly varying function. Smooth regular variation allows for defining associated classes of pseudo-differential operators and computing Dixmier traces of geometric operators by means of a Connes trace theorem, see [32, Section 7 and 9]. For a decreasing smoothly varying function ψ : [0,∞) → (0,∞) with limt→0 ψ(t) = 0, we define the self-adjoint operator /Dψ := F /Dψ( /Dn)−1. It follows from [32, Proposition 10.1] that /Dψ ∈ L0 ψ−1(M, S) (see [32] for the meaning of this symbol) and its ψ-principal symbol is cS(ξ)ξ−1ψ(ξ)−1, where cS : T ∗M → End(S) denotes Clifford multiplication. The next result follows from [32, Proposition 10.3]. Proposition 2.16. Let M be an n-dimensional Riemannian closed manifold, /D a Dirac op- erator on M , and ψ as above with ψ(t)−1 = O(t1/n), as t → ∞. Then (C ∞(M ), L2(M, S), /Dψ) is a ψ-summable spectral triple whose associated K-homology class coincides with that of (C ∞(M ), L2(M, S), /D). If ψ(t)−1 = o(t1/n), then for any a ∈ C ∞(M ), the operator [ /Dψ, a] is compact with µ(t, [ /Dψ, a]) = O(t−1/nψ(t)−1), as t → ∞. We return to the problem of computing heat traces involving /Dψ below in the Section 4 (see page 38) and Subsection 5.1 (see page 43). The spectral triple (C ∞(M ), L2(M, S), /Dψ) does in some cases extend to a crossed product by a semigroup action. For simplicity, we consider an action of N by a local diffeomorphism g : M → M . Following [32], we say that g acts conformally if there is a function cg ∈ C ∞(M, R>0) such that g∗gM = cggM where gM denotes the Riemannian metric on M . We say that g lifts to the Clifford bundle S → M if there is a unitary Clifford linear morphism ug : g∗S → S. For simplicity, we assume M to be connected and define N := #g−1({x}) for some x ∈ M . Since g is a local diffeomorphism and M is connected, N is independent of the choice of x. If g : M → M is a surjective local diffeomorphism, acting conformally and lifting to S, we can define the isometry Vg : L2(M, S) → L2(M, S), Vgξ := cn/4 g N −1/2ug(ξ ◦ g). (2.5) The isometry Vg satisfies the following for a ∈ C(M ): VgaV ∗ g = (a ◦ g)VgV ∗ g , and V ∗ g aVg = Lg(a), 15 where Lg(a)(x) := Pg(y)=x a(y). See more in [19, Proposition 8.3]. Define A as the ∗- algebra generated by C ∞(M ) and Vg. By [19, Proposition 8.6], the C ∗-closure of A is the image in a representation on L2(M, S) of the Cuntz-Pimsner algebra OEg defined from Eg := C(M )Vg (see more in Example 2.30 (see page 23) and Example 6.6 (see page 52)). The space Eg = C(M )Vg ⊆ B(L2(M, S)) is considered as a bimodule over C(M ) and is an fgp bi-Hilbertian C(M )-bimodule because Eg can be identified with C(M ) with the bimodule structure (af b)(x) = a(x)f (x)b(g(x)) for a, b, f ∈ C(M ), for details, see Example 2.30 below on page 23 or [19, Section 8]. Definition 2.17. Let ψ : [0,∞) → (0,∞) be a decreasing function and g : M → M a local diffeomorphism of a Riemannian manifold. • We say that ψ and g are compatible if there is a constant C ≥ 0 such that for all x ∈ M and ξ ∈ T ∗ x M , the differential Dg satisfies (cid:12)(cid:12)ψ((Dg)T x ξ) − ψ(ξ)(cid:12)(cid:12) ≤ Cψ(ξ)2. • We say that g acts isometrically if Dg acts isometrically on each fibre. The following results poses restrictions on a function ψ compatible with a local diffeomorphism which acts non-isometrically. Proposition 2.18. Let ψ : [0,∞) → (0,∞) be a decreasing function with limt→∞ ψ(t) = 0 and g : M → M a compatible local diffeomorphism of a Riemannian manifold. If ψ has regular variation, then either g acts isometrically or ψ has regular variation of index 0. Proof. Assume that g acts non-isometrically, we shall prove that in this case ψ has regular ψ(rt) ψ(t) = 1 for some variation of index 0. Since ψ is decreasing, it suffices to show that limt→∞ r 6= 1 by [32, Proposition 2.15]. If g acts non-isometrically, there is a point x ∈ M and a unit vector ξ0 ∈ T ∗ x M such that (Dg)T x ξ0g(x) 6= 1. Set r := (Dg)T x ξ0g(x) 6= 1. Since ψ is compatible with g, there is a constant C ≥ 0 such that and by setting ξ = tξ0, we arrive at the inequality (cid:12)(cid:12)ψ((Dg)T x ξ) − ψ(ξ)(cid:12)(cid:12) ≤ Cψ(ξ)2 ψ(rt) − ψ(t) ≤ Cψ(t)2. After dividing by ψ(t), taking the limit t → ∞ and using that limt→∞ ψ(t) = 0 we arrive at the desired equality limt→∞ ψ(rt) ψ(t) = 1. A prototypical example of a function with regular variation of index 0 is ψ(t) := log(1 + t). This function is compatible with any conformal local diffeomorphism and has been considered in the context of constructing spectral triples in [19, 29, 30, 32, 50]. 16 Proposition 2.19. Let M be an n-dimensional Riemannian closed manifold, /D a Dirac op- erator on S → M , g : M → M a surjective local diffeomorphism acting conformally and lifting to S and ψ : [0,∞) → (0,∞) a decreasing function compatible with g, having smooth regular variation and satisfying limt→∞ ψ(t) = 0 and ψ(t)−1 = O(t1/n) as t → ∞. Let A denote the ∗-algebra generated by C ∞(M ) and the isometry Vg from Equation (2.5) (see page 15). Then (A, L2(M, S), /Dψ) is a unital ψ-summable spectral triple. This result follows in the same manner as in [19, Section 8] and [32, Theorem 10.6]. Clearly, if ψ(t)−1 = O(log(t)) as t → ∞, then (A, L2(M, S), /Dψ) is Li1-summable. The reader should note that the K-homology class [(A, L2(M, S), /Dψ)] ∈ K ∗(OEg ), where OEg is the Cuntz-Pimsner algebra of the module Eg = C(M )Vg, discussed later. The class [(A, L2(M, S), /Dψ)] restricts to [(C ∞(M ), L2(M, S), /D)] ∈ K ∗(C(M )). Thus the class [(C ∞(M ), L2(M, S), /D)] ∈ K ∗(C(M )) obstructs [(A, L2(M, S), /Dψ)] ∈ K ∗(OEg ) being a Kasparov product with the Cuntz-Pimsner boundary extension in KK 1(OEg , C(M )) -- we will return to study this boundary extension and its associated semifinite spectral triples below in Subsection 2.2.3. 2.2.2 Graph C ∗-algebras A class of examples carrying interesting noncommutative geometries with a well-studied set of KMS-states is that of graph C ∗-algebras. With a finite directed graph G = (V, E), with edge set E and vertex set V , one associates a C ∗-algebra C ∗(G) [54]. For simplicity we suppose that we have no sources nor sinks. The C ∗-algebra C ∗(G) is generated by partial isometries (Se)e∈E and projections (pv)v∈V satisfying the relations S∗ e Se = pr(e), and pv = Xs(e)=v SeS∗ e , where r(e) denotes the range of the vertex e and s(e) its source. The C ∗-algebra C ∗(G) can be described as a Cuntz-Pimsner algebra in several ways, a class of C ∗-algebras carrying noncommutative geometries that we will study in the next subsection. In this subsection, we focus on a construction of noncommutative geometries along an orbit in the infinite path space of G -- a construction based in the model of C ∗(G) as a groupoid C ∗-algebra. The associated noncommutative geometries come from [29]. As explained in [31] the groupoid model can be seen as a Cuntz-Pimsner model of C ∗(G) using the one-sided infinite path space ΩG := {x = e1e2 ··· ∈ E N : s(ej) = r(ej+1) ∀j}. The path space ΩG is a compact Hausdorff space in the subspace topology ΩG ⊆ E N. It carries a shift mapping σG : ΩG → ΩG, σG(e1e2e3 ··· ) := e2e3 ··· . The shift mapping is a surjective local homeomorphism. We can define an ´etale groupoid GG over ΩG by GG :=(cid:8)(x, n, y) ∈ ΩG × Z × ΩG : ∃k ≥ max(0,−n) such that σn+k G (x) = σk The range mapping is defined by r(x, n, y) = x, the source mapping as s(x, n, y) := y and the product by G(y)(cid:9). (x, n, y)(y, m, z) := (x, n + m, z). 17 The topology of GG is uniquely determined by declaring the groupoid to be ´etale and the mappings (x, n, y) 7→ n and κG(x, n, y) := min{k ≥ max(0,−n) : σn+k G (x) = σk G(y)}, to be continuous. There is an isomorphism πG : C ∗(G) → C ∗(GG) determined by defining πG(Se) to be the characteristic function of the set {(x, 1, σG(x)) : x ∈ Ce} where Ce := {e1e2 ··· ∈ ΩG : e1 = e}. See [18]. Define the function ψ0 : {(n, k) ∈ Z × N : n + k ≥ 0} → Z by k = 0 k > 0. ψ0(n, k) :=(n, −n − k, For a point y ∈ ΩG, we define the discrete set Vy := d−1({y}) = {(x, n) : (x, n, y) ∈ GG}. The set Vy is the union of all forward orbits of all backward orbits of y under σG where we keep track of the lag n. Since Vy is a fibre of the domain mapping, point evaluation in y induces a representation πy : C ∗(G) → B(ℓ2(Vy)). If G is primitive C ∗(G) is simple and πy is faithful. At the level of the generators, πy(Se)δ(x,n) =(δ(ex,n+1), 0, r(x) = s(e), r(x) 6= s(e). Proposition 2.20. Define the operator Dy densely on ℓ2(Vy) as the self-adjoint operator with Dyδ(x,n) := ψ0(n, κ(x, n, y))δ(x,n). The triple (Cc(GG, ℓ2(Vy), Dy) is an Li1-summable spectral triple. Moreover, under the isomor- phism K 1(C ∗(G)) ∼= ZE/(1− Aedge)ZE, of odd K-homology with the cokernel of the edge adja- cency matrix, the class [(Cc(GG, ℓ2(Vy), Dy)] is mapped to the element δe mod (1 − Aedge)ZE where δe ∈ ZE denotes the basis element corresponding to e ∈ E. Proof. It is proven in [29, Theorem 5.2.3] that (Cc(GG, ℓ2(Vy), Dy) is a spectral triple whose K-homology class corresponds to the element δe mod (1 − Aedge)ZE under K 1(C ∗(G)) ∼= ZE/(1− Aedge)ZE. It remains to prove that (Cc(GG, ℓ2(Vy), Dy) is Li1-summable. We compute that Tr(e−tDy) = X(x,n)∈Vy However, if (x, n) is such that κG(x, n, y) = k the path x is determined by y except for its first n + k steps so #{(x, n) : κG(x, n, y) = k} ≤ En+k ≤ elog(E)(n+k). We conclude that e−t(n+κG(x,n,y)) =Xn∈Z #{(x, n) : κG(x, n, y) = k}e−t(n+k). ∞Xk=max(0,−n) Tr(e−tDy) < ∞ if t > log(E). We return to this example below in Example 3.6 (see page 28) and Subsection 5.2 (see page 47). 18 2.2.3 Cuntz-Pimsner algebras In this subsection we consider the construction of semi-finite spectral triples on Cuntz-Pimsner algebras -- a broad class of examples which include both Cuntz-Krieger algebras and crossed products by Z. Quite general techniques for constructing spectral triples for these algebras were developed in a series of papers (in rough chronological order) [29, 56, 31, 30, 57]. We consider the set up of [56] and [31], which provide a means of lifting data from the (unital) coefficient algebra of a bi-Hilbertian bimodule to its Cuntz-Pimsner algebra. We start with a unital, separable C ∗-algebra A, and a finitely generated projective (fgp) bi- Hilbertian bimodule E over A, i.e. a module fulfilling the conditions of the following definition. Definition 2.21. An fgp bi-Hilbertian bimodule E over A is an A-bimodule equipped with the following structures: • E has both left and right A-valued inner products which induce equivalent norms on E. • The left and right actions are both injective and adjointable. • E is finitely generated and projective as both a left and right module. To separate the left and the right structures, we write EA when we want to emphasize the right module structure and (··)A for the right inner product. Similarly, AE denotes the left module defined from E and A(··) the left inner product. The algebraic Fock space Falg space FE is defined as the right A-Hilbert C ∗-module completion of Falg algebra TE ⊆ End∗ for µ ∈ Falg E is the algebraic direct sum of the A-modules E⊗Ak. The Fock E . The Cuntz-Toeplitz A(FE) is the C ∗-algebra generated by the creation operators Tµξ := µ ⊗ ξ E . The Cuntz-Pimsner algebra OE is defined from the short exact sequence 0 → KA(FE) → TE → OE → 0. We call this short exact sequence the defining extension of OE. A set (ej)N e = Pj ej(eje)A, and similarly for a left frame (fk)N j=1 ⊂ E of vectors is a frame for the right module EA if for all e ∈ E we have k=1. Since E is finitely generated and projective, there exists left and right frames and we can for simplicity assume that they have the same cardinality. For e and f in the right Hilbert module EA, we denote the associated rank-one operator by Θe,f := e(f·). Then the frame condition can be expressed as Θej,ej = IdE NXj=1 and similarly for fσ. The frame (ej)N a multi-index and eρ = eρ1 ⊗ ··· ⊗ eρk . We define the right Watatani index of E⊗k as the element of A given by j=1 induces a frame for E⊗k A , namely (eρ)ρ=k where ρ is eβk = Xρ=k A(eρeρ) = Xρ′=k−1 A(eρ′ eβeρ′). (2.6) 19 The right Watatani index is positive, central and since the left action is injective, also invertible. Therefore βk is a well defined self-adjoint central element in A. The key assumptions we make concern the asymptotic behaviour of the right Watatani indices. In [56, Section 3.2] we define an A-bilinear functional Φ∞ : OE → A. This functional gives us an A-valued inner product on OE. The construction of Φ∞ begins by defining Φk : End∗ A(E⊗k) → A, A(T eρeρ). (2.7) Φk(T ) = Xρ=k Here we use the notation End∗ A(E⊗k) for the C ∗-algebra of A-linear adjointable operators on E⊗k. It follows from [41, Lemma 2.16] that Φk does not depend on the choice of frame. We note that eβk = Φk(IdE⊗k). Since Φk is independent of the choice of frame, so is eβk . We extend the functional Φk to a mapping End∗ A(FE) → A by compressing along the orthogonally complemented submodule E⊗k ⊆ FE. To obtain a good "limiting functional" Φ∞(T ) := limk→∞ Φk(T )e−βk on the Cuntz-Toeplitz algebra, we impose the following condition on the Watatani indices. Definition 2.22. Let E be an fgp bi-Hilbertian bimodule over the unital C ∗-algebra A. We say that E is W-regular if for every k ∈ N and ν ∈ E⊗k there exists a ν ∈ E⊗k satisfying ke−βnνeβn−k − νkE⊗k → 0 as n → ∞. In [56] the reader can find several examples of Cuntz-Pimsner algebras for which a stronger version of W-regularity as defined in Definition 2.22 holds. There are no known examples of modules that are not W-regular. When E is W-regular, [56, Proposition 3.5] guarantees that limk→∞ Φk(T )e−βk is well defined for T from the ∗-algebra generated by the set of creation operators {Tν : ν ∈ Falg E } and is continuous in the C ∗-norm. In Section 6.3 we shall see that there are ways around W-regularity, and even the existence of an A-valued left inner product, when constructing semifinite spectral triples giving rise to KMS-states. We thus obtain a unital positive A-bilinear functional Φ∞ : TE → A. The functional Φ∞ annihilates the compact endomorphisms, and descends to a well-defined functional on the Cuntz-Pimsner algebra OE. By an abuse of notation, we also denote this functional by Φ∞ : OE → A. Since Φk and eβk do not depend on the choice of frame, neither does Φ∞. We define the inner product S1, S2 ∈ OE. When computing these inner products, the following fact is useful. (S1S2)A := Φ∞(S∗ 1 S2), Lemma 2.23. Let E be a W-regular fgp bi-Hilbertian bimodule. For homogeneous elements µ, ν ∈ Falg E we have Φ∞(SµS∗ ν ) = lim k→∞ A(µe−βk νeβk−ν) = A(µν). (2.8) In particular, if S ∈ OE is homogeneous of degree n 6= 0, then Φ∞(S) = 0. Completing OE modulo the vectors of zero length (with respect to Φ∞) yields a right A- Hilbert C ∗-module that we denote by ΞA. The module ΞA carries a left action of OE given by extending the multiplication action of OE on itself. By considering the linear span of the image of the generators Sν, ν ∈ Falg E , inside the module ΞA, we obtain an isometrically embedded and complemented copy of the Fock space. We let Q be the projection on this copy of the Fock space. 20 Theorem 2.24 (Proposition 3.14 of [56]). Let E be a W-regular fgp bi-Hilbertian bimodule over a unital C ∗-algebra A. The tuple (OE, ΞA, 2Q−1) is an odd Kasparov module representing the class of the defining extension 0 → KA(FE) → TE → OE → 0. (2.9) To construct an unbounded representative of (OE, ΞA, 2Q − 1), we will add an additional assumption regarding the fine structure of the operation ν 7→ ν in the definition of W-regularity (see Definition 2.22). Assuming W-regularity, we can define the operator qk : E⊗k → E⊗k by qkν := ν = lim n→∞ e−βnνeβn−k . Definition 2.25. Let E be an fpg bi-Hilbertian bimodule over the unital C ∗-algebra A. We say that E is strictly W-regular if it is W-regular and for any k, we can write qk = ckPk = Pkck where Pk ∈ End∗ A(E⊗k) is a (necessarily A-bilinear) projection and ck is given by left- multiplication by an element in A. Remark 2.26. As with W-regularity, the reader can in [56] find several examples of Cuntz- Pimsner algebras for which strict W-regularity holds. There are no known examples of modules that are not strictly W-regular. Remark 2.27. If there is a decomposition qk = ckPk as in the definition of strict W-regularity, [31, Lemma 3.8] shows that it is unique and of a very specific form. Indeed, each ck is central, invertible and ck = Φk(Pk)−1. Strict W-regularity is readily verified in practice using [31, Lemma 3.8]. For instance, if β1 is central for the module action on E, ck = e−βk = e−kβ1 is central for the module action on E and Pk = 1E⊗k . When E is strictly W-regular, an unbounded self-adjoint regular operator Dψ on ΞA is con- structed in [31] making (OE, ΞA, Dψ) into an unbounded Kasparov module representing the KK-class of the defining extension (2.9). The operator Dψ is of the form Dψ =Xn∈Z Xr≥max{0,n} ψ(n, r)Pn,r where ψ : Z × N → [0,∞) is a function with certain Lipschitz properties (see [31, Remark 3.20]), and the Pn,r are projections on finitely generated projective subspaces, [31]. While the particular choice of function ψ does not matter much, we will take the function ψ(n, r) =(cid:26) n n = r . −(2r − n) otherwise r=max(0,n)Pn,r is the projection onto the A-linear span of {SµS∗ The projections Pn,r form a sequence of mutually orthogonal projections satisfying that the direct sum ⊕r0 ν : µ − ν = In particular, Pn,n is the projection onto E⊗n for n ∈ N. More n, max(0, n) ≤ r ≤ r0}. precisely, the projections are defined by Pn,r :=(Qn,r − Qn,r−1, (2.10) Qn,r, r > max{0, n} r = max{0, n} 21 where the projections Qn,r are defined in terms of the right frame (ej)N (fj)N j=1 as j=1 and the left frame Qn,r := Xρ−σ=n, ρ=r ΘW eρ,c −1/2 σ PF fσ ,W eρ,c −1/2 σ PF fσ (2.11) where PF = ⊕Pk is the projection on the Fock module coming from Definition 2.25. Here we have written Wξ,η ∈ ΞA for the element defined from SξS∗ η ∈ OE where ξ ∈ E⊗r and η ∈ E⊗k. For details, see [31, Lemma 3.10 and Proposition 3.11]. To obtain a semifinite spectral triple, we localize (OE, ΞA, Dψ) in a positive trace on A. Fol- lowing Proposition 2.12, we consider the semifinite spectral triple (OE , L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ). Here L2(ΞA, τ ) := ΞA ⊗A L2(A, τ ) and Trτ is the dual trace on Nτ (ΞA) := (End∗ which satisfies Trτ (Θe,f ) = τ ((fe)A), [45, 62]. Lemma 2.28. Assume that the fgp bi-Hilbertian bimodule E is strictly W-regular and that τ is a positive trace on A. Then the semifinite spectral triple A(ΞA) ⊗ 1)′′ is Li1-summable. (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) Remark 2.29. The assumptions on the existence of limiting behaviour for the Watatani indices are really just for convenience here. These assumptions relate to existence and behaviour of norm limits, but we have passed to the 'measurable setting' and so really only need weak limits. We will explore this point of view in Subsection 6.3. Proof. We need to prove that the following expression is finite for t large enough: Trτ (e−tDψ ) =Xn∈Z Xr≥max{0,n} e−tψ(n,r) Trτ (Pn,r). By definition (see (2.10)), Pn,r = Qn,r − Qn,r−1 when r > max(0, n) and Pn,r = Qn,r when r = max(0, n) and Qn,r is defined as in (2.11). Using the computations of [31, Lemma 2.8], we see that eρ,c−1/2 σ PF fσW eρ,c−1/2 σ PF fσ (2.12) τ(cid:18)(W τ ◦ Φ∞(cid:18)S Trτ (Qn,r) = Xρ−σ=n, ρ=r = Xρ−σ=n, ρ=r = Xρ−σ=n, ρ=r c−1/2 σ PF fσ S∗ eρSeρS∗ τ (A(PF fσPF fσ(eρeρ)A)) . )A(cid:19) σ PF fσ(cid:19) c−1/2 Using the fact that the elements of the frame have norm bounded by 1, we see that Trτ (Qn,r) ≤ Xρ−σ=n, ρ=r ≤ Xρ−σ=n, ρ=r τ (A(PF fσPF fσ(eρeρ)A)) k(A(PF fσPF fσ(eρeρ)A)kA ≤ N 2r−n ≤ elog(N )ψ(n,r), 22 where N is the number of elements in the left frame and the right frame. We can now estimate (cid:12)(cid:12)(cid:12)Trτ (e−tDψ )(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xn∈Z Xr>max{0,n} +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xn∈Z Xr=max{0,n} ≤2Xn∈Z Xr≥max{0,n} ≤2Xn∈Z Xr≥max{0,n} e−(t−log(N ))ψ(n,r) < ∞, e−tψ(n,r) Trτ (Qn,r − Qn,r−1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) e−tψ(n,r) Trτ (Qn,r)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) e−tψ(n,r) Trτ (Qn,r) if t > log(N ). We return to Cuntz-Pimsner algebras below in Example 3.7 (see page 29) and Section 6 (see page 50). Let us discuss a special case of Cuntz-Pimsner algebras arising on a commutative coefficient algebra. Example 2.30. Let Y be a compact Hausdorff space and g : Y → Y a surjective local homeomorphism. We consider the module Eg = C(Y ) with the bimodule action (af b)(x) = a(x)f (x)b(g(x)), a, b ∈ C(Y ), f ∈ Eg. This is an fgp bi-Hilbertian bimodule in the inner products C(Y )(f1, f2) = f1f2 and (f1f2)C(Y ) := Lg(f1f2), where Lg : C(Y ) → C(Y ) is the transfer operator Lg(f )(x) := Xy∈g−1(x) f (y). (2.13) The Cuntz-Pimsner algebra OEg can be realized as a groupoid C ∗-algebra as in [18] (see also [19, Theorem 3.2]) over the solenoid X = {x = y1y2 ··· ∈ Y N : g(yk+1) = yk ∀k}. The case that X equipped with the shift mapping is a Smale space was studied in [19]. j=1 where ej = √χj for a partition of unity (χj)N The module Eg has a right frame (ej)N subordinate to an open covering (Uj)N j=1 of Y such that gUj is injective for all j. Using this partition of unity, one sees that βk = 0 for all k. It follows that Eg is a strictly W-regular module. The case that g : M → M was a surjective local diffeomorphism acting conformally was con- sidered in Subsection 2.2.1 (see page 14). However, the spectral triple considered Proposition 2.19 on OEg differs greatly from the semifinite spectral triples considered in Lemma 2.28 -- the latter are in the image of the boundary mapping in KK1(OEg , C(M )) defined from Equation (2.9) while the former is not if [ /D] 6= 0 ∈ K ∗(C(M )). j=1 23 Another class of examples already considered arises from a finite graph G as in Subsection 2.2.2 (see page 17) where the shift mapping σG : ΩG → ΩG is a surjective local homeomorphism and C ∗(G) ∼= OEσG . The spectral triples in Proposition 2.20 (see page 18) arises from the construction of Lemma 2.28 by taking the trace τ : C(ΩG) → C to be defined from point evaluation in y. 2.2.4 Group C ∗-algebras We now turn our attention to examples coming from the reduced group C ∗-algebra of a discrete group. A well known construction associates a spectral triple with a length function on the group, we consider this example and a semifinite modification thereof which is possible for a-T-menable groups, i.e. groups with the Haagerup property. The methods extend to a-TT- menable groups, a class of groups containing all hyperbolic groups, see more in [51, Chapter 7.2]. Let Γ denote a countable discrete group. Recall that a length function ℓ : Γ → R≥0 is a function satisfying ℓ(e) = 0 for e ∈ Γ the identity, and ℓ(γγ′) ≤ ℓ(γ) + ℓ(γ′) for all group elements γ, γ′ ∈ Γ. We say that ℓ is a proper length function if ℓ is a proper function, i.e. the set {γ ∈ Γ : ℓ(γ) ≤ R} is finite for any R ≥ 0. If there exists a constant β ≥ 0 such that #{γ ∈ Γ : ℓ(γ) ≤ R} = O(eβR) as R → ∞ we say that (Γ, ℓ) has at most exponential growth. Define the operator Dℓ densely on ℓ2(Γ) as the self-adjoint operator with Dℓδγ := ℓ(γ)δγ . The space of compactly supported functions cc(Γ) is a core for Dℓ. We define the ∗-algebra cb(Γ) ⋊alg Γ as the ∗-algebra generated by multiplication operators (by bounded functions on Γ) cb(Γ) ⊆ B(ℓ2(Γ)) and all left translation operators. Proposition 2.31. Let ℓ be a proper length function on Γ. The triple (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) is a spectral triple defining the trivial class in the K-homology of the C ∗-algebra cb(Γ) ⋊r Γ. Moreover, if (Γ, ℓ) has at most exponential growth the spectral triple (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) is Li1-summable. Proof. Since ℓ is proper, it is clear that Dℓ has compact resolvent and if (Γ, ℓ) has at most exponential growth, then there is C > 0 such that Tr(e−tDℓ) =Xγ∈Γ e−tℓ(γ) = ∞Xn=0 #{γ ∈ Γ : ℓ(γ) = n}e−tn ≤ C ∞Xn=0 e−(t−β)n = C 1 − eβ−t < ∞. To show that (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) is a spectral triple, it remains to show that cb(Γ) ⋊alg Γ preserves the domain of Dℓ and has bounded commutators with Dℓ. Domain preservation is clear. For an element aλγ ∈ cb(Γ) ⋊alg Γ and a function f ∈ cc(Γ) we compute that [Dℓ, aλγ]f (g) = a(g)(ℓ(g) − ℓ(γ−1g))f (γ−1g). It follows that k[Dℓ, aλγ]kB(ℓ2(Γ) ≤ kakcb(Γ) sup g∈Γ ℓ(g) − ℓ(γ−1g) ≤ kakcb(Γ)ℓ(γ). 24 The K-homology class of (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) is trivial. We shall now consider a topolog- ically more interesting semifinite spectral triple that can be constructed on groups with the Haagerup property. We are grateful to Branimir ´Ca´ci´c for sharing this construction with us. Similar ideas appeared in [38, Appendix B]. Let Γ be a discrete group with the Haagerup property. Then there is a proper isometric action of Γ on a real Hilbert space HΓ. By the Mazur-Ulam theorem there exists an orthogonal representation on the Hilbert space HΓ and a proper cocycle cΓ for πΓ, meaning that cΓ : Γ → H is a proper function satisfying the cocycle identity πΓ : Γ → O(HΓ) The cocycle identity allows us to define a length function on Γ by cΓ(γ1γ2) = cΓ(γ1) − πΓ(γ1)cΓ(γ2). (2.14) ℓ(γ) := kcΓ(γ)kHΓ. Since cΓ is proper, so is ℓ. Remark 2.32. The existence of a proper isometric action of a group Γ on a Hilbert space is equivalent to Γ having the Haagerup property, also known as a-T-menability. Our construction extends to the case when there exists an orthogonal representation πΓ : Γ → O(HΓ) and a proper quasi-cocycle cΓ : Γ → HΓ. That is, when Q(cΓ) := sup γ1,γ2 kcΓ(γ1γ2) − cΓ(γ1) + πΓ(γ1)cΓ(γ2)kHΓ < ∞. In this case, ℓ(γ) := kcΓ(γ)kHΓ could fail to be a length function but still satisfies ℓ(γγ′) ≤ ℓ(γ) + ℓ(γ′) + Q(cΓ) which suffices for our purposes. The existence of a proper quasi-cocycle on a Hilbert space is equivalent to Γ being a-TT-menable. Hyperbolic groups are a-TT-menable. For notational simplicity, we restrict our attention to cocycles. Definition 2.33. Let Cℓ(HΓ) denote the the complex Clifford algebra of HΓ and assume that cS : Cℓ(H) → B(SH) is a representation of Cℓ(H). We say that a unitary representation πS : Γ → U (SH) is a lift of πΓ to SH if for all v ∈ H and g ∈ Γ we have πS(g)cS (v)πS (g−1) = cS(πΓ(g)v). (2.15) When H is finite dimensional this is just the well-known Clifford algebra, but when H is infinite dimensional we refer to [8, 65] for a description of this algebra. For a representation SH of Cℓ(H) we consider the new Hilbert space ℓ2(Γ, SH). Assuming that πS lifts πΓ to SH the Hilbert space ℓ2(Γ, SH) carries a representation of Γ defined by π : Γ → U (ℓ2(Γ, SH)), (cid:0)π(g)f(cid:1)(γ) = πS(g)f (g−1γ). On the Hilbert space ℓ2(Γ, SH) define a self-adjoint operator Dc by declaring (Dcf )(γ) := cS(cΓγ))f (γ), f ∈ cc(Γ, SH). Since cS(v)2 = kvk2 H for all v ∈ H, the domain for Dc can be deduced from (D2 c f )(γ) = ℓ(γ)2f (γ). 25 (2.16) (2.17) The compatibility requirement Equation (2.15) and cocycle property Equation (2.14) imply that for g, γ ∈ Γ we have πS(g)cS (cΓ(γ)) = cS (π(g)cΓ(γ))πS (g) = cS(cΓ(g))πS (g) + cS(cΓ(gγ))πS (g). Then the commutator of Dc and a group element is ([Dc, π(g)]f )(γ) = cS(cΓ(γ))πS(g)f (g−1γ) − πS(g)cS (cΓ(g−1γ))f (g−1γ) = cS(cΓ(g))πS (g)f (g−1γ) = (cS(cΓ(g))π(g))(f )(γ). Hence the commutators between Dc and group elements are bounded. It is moreover clear that these commutators lie in N0 ⋊ Γ where N0 = cS(Cℓ(H))′′. We define N as the von Neumann algebraic tensor product B(ℓ2(Γ)) ¯⊗N0. Finally, (1 + D2 c )−1 ∈ K(ℓ2(Γ)) ⊗ 1 ⊂ K(ℓ2(Γ)) ⊗ Kτ where Kτ is the compacts in N0 for a choice of normalized positive trace τ . Let Trτ be the c )−1 ∈ KTrτ . Finally, if ℓ trace on N defined from the trace τ on N0. We conclude that (1 + D2 has at most exponential growth then Trτ (e−tDc) =Pγ∈Γ e−tℓ(γ) < ∞ for t large enough. As such, (i ± Dc)−1 ∈ Li1 if ℓ has at most exponential growth. We conclude the following result. Proposition 2.34. Assume that cS : Cℓ(H) → B(SH) is a representation of Cℓ(H) and that the unitary representation πS : Γ → U (SH) lifts πΓ to SH. Let cb(Γ) be the (continuous) bounded functions on Γ, and define a representation of cb(Γ) ⋊ Γ on ℓ2(Γ, SH) by πS(aλg)f (γ) = a(γ)[π(g)f ](γ), where π is as in (2.16). Then the triple (cb(Γ) ⋊alg Γ, ℓ2(Γ, SH), Dc, N, Trτ ) is a semifinite spectral triple which is Li1-summable if ℓ has at most exponential growth. We return to the example of this subsubsection in Example 3.9 (see page 30) and Subsection 5.3 (see page 48). Example 2.35. The following example of a proper group cocycle shows the construction's geometric advantage compared to only using a length function. Consider the trivial action of the discrete group Γ = Zn on HΓ = Rn. The inclusion Zn ֒→ Rn is additive and proper, and therefore a proper group cocycle for the trivial action. The semifinite spectral triple associated with a finite dimensional Clifford representation cS : Rn → EndC(S) can when restricted to C ∗(Zn) ∼= C(Tn) be identified with the semifinite spectral triple (C ∞(Tn), L2(Tn, S), /DTn ⊗ 1S, B(L2(Tn)) ⊗ Cℓn, Trτ ) using Fourier theory on the dual torus Tn = cZn. We observe that to extract K-homological content from this construction we need to specify a grading if n is even. In particular, it is unclear how to interpret the construction above in K-homology when HΓ is infinite-dimensional. 3 KMS states constructed from Li1-summable spectral triples This section contains the fundamental technical construction of the paper. Starting from a semifinite Li1-summable spectral triple, we use the associated algebra of Toeplitz operators to construct an action from the operator D and a KMS-state from the operator D. 26 3.1 The positive part of the spectrum and heat traces Throughout this section we suppose that (A, H, D, N, T) is a unital semifinite spectral triple. The spectral triple can be semifinite, in which case we let T denote the given positive faithful normal semifinite trace. In general T is not unique, and coincides with a non-zero multiple of the operator trace in the "usual" non-semifinite case N = B(H). We write KN for the compacts for T. Again, KN coincides with the usual compacts in the case of the type I factor N = B(H). We write N+ for the von Neumann algebra PDNPD. By an abuse of notation, we write T also for the induced faithful normal semifinite trace on N+. The T-compacts on N+ will be denoted by K+ N. Definition 3.1. The operator D is said to have positive T-essential spectrum if for some β ∈ [0,∞), we have T(PDe−tD) < ∞ for t > β and T(PDe−tD) ր ∞ as t ց β. Proposition 3.2. Let D be a self-adjoint densely defined operator affiliated with N satisfying that (1 + D2)−1/2 ∈ Li1(T). We define the number βD := inf{t > 0 : T(PDe−tD) < ∞}. (3.1) Then βD ∈ [0,∞) and D has positive T-essential spectrum if and only if lim tցβD T(PDe−tD) = ∞. In particular, if βD = 0 then D has positive T-essential spectrum if and only if T(PD) = ∞. Proof. By definition, if (1 + D2)−1/2 ∈ Li1 then T(e−tD) < ∞ for t large enough. Therefore, T(PDe−tD) = T(PDe−tD) < ∞ for t large enough and βD := inf{t > 0 : T(PDe−tD) < ∞} will be a number in [0,∞). By definition, T(PDe−tD) < ∞ for t > βD and if limtցβD T(PDe−tD) = ∞ then D has positive T-essential spectrum with β = βD. Conversely, if D has positive T- essential spectrum there is a β ∈ [0,∞) with T(PDe−tD) < ∞ for t > β and T(PDe−tD) ր ∞ as t ց β, and in this case it is clear that β = βD. Remark 3.3. We will often impose the assumption of positive T-essential spectrum. If for some β ∈ [0,∞), T((1 − PD)etD) < ∞ for t > β and T((1 − PD)etD) ր ∞ as t ց β, we can equally well use −D in our construction. Proposition 3.4. Let D be a self-adjoint densely defined operator affiliated with N satisfying that (1 + D2)−1/2 ∈ Li1(T). Then D has positive T-essential spectrum if and only if both of the following conditions fail: 1. PD has finite T-trace. 2. There exists a p > 0 such that PDe−D ∈ Lp(T) \ ∩q>pLq(T). The reader should note that conditions 1. and 2. are mutually exclusive. Proof. If D has positive T-essential spectrum, then clearly 1. fails. Also 2. fails if D has positive T-essential spectrum because condition 2. is equivalent to βD = p and limtցp T(PDe−tD) being finite. 27 fails, then either βD = 0 and limtց0 T(PDe−tD) must be Conversely, if Condition 1. and 2. infinite not to violate T(PD) being infinite or βD > 0 and the set {p > 0 : PDe−D ∈ Lp(T)} is open (due to condition 2. failing) showing that T(PDe−tD) ր ∞ as t ց βD. Example 3.5. Dirac operators on closed manifolds, as considered in Subsection 2.2.1 (see page 14), have positive essential spectrum. In this case, we can compute βD = 0 and the leading term in the heat trace asymptotics using Proposition 2.14. If (C ∞(M ), L2(M, S), /D) is the spectral triple associated with a Dirac operator, Proposition 2.14 implies that TrL2(M,S)(PDe−t /D) = n!cnt−nZS∗M TrS(p /D)dV + O(t−n+ǫ), as t → 0, 2 (cS(ξ) + 1) for x ∈ M and ξ ∈ S∗ where p /D is the principal symbol of the zeroth order pseudo-differential operator P /D. A direct computation shows that p /D(x, ξ) = 1 xM . Here cS : T ∗M → End(S) denotes Clifford multiplication. More generally, Proposition 2.14 allows us to conclude that for a ∈ C ∞(M ), TrL2(M,S)(PDae−t /D) = n!cnt−nZS∗M = n!cnt−nZMZS∗ = cnt−nZM TrS(p /D)a dV + O(t−n+ǫ) = xM (ξ)dV (x) + O(t−n+ǫ) = a(x)TrS(p /D(x, ξ)) dVS∗ a dV + O(t−n+ǫ), as t → 0, xM for a new constant cn > 0, depending only on the dimension of M and the rank of S. In the last equality we used that p /D − 1/2 is an antisymmetric function under the involution (x, ξ) 7→ (x,−ξ) of S∗M and therefore ZS∗ xM TrS(p /D(x, ξ)) dVS∗ TrS(p /D(x,−ξ)) dVS∗ xM = xM =ZS∗ xM = rank(S) 2 ZS∗ xM dVS∗ xM = rank(S)πdim(M )/2 Γ(dim(M )/2) . Example 3.6. The spectral triples for graph C ∗-algebras from Proposition 2.20 (see page 18) also have positive essential spectrum. We compute βD and the heat trace asymptotics assuming that G is primitive. In this example, PDy is the projection onto the subspace ℓ2(Vy ∩ κ−1 G (0)). We use the notation y := Vy ∩ κ−1 V+ G (0) = {(x, n) ∈ Vy : n ≥ 0, σn G(x) = y}. (3.2) . Note that G(x) = y then x is uniquely determined by y and a finite path σ = σ1σ2 ··· σn with x = σy. G ({y}). The space PDy ℓ2(Vy) is therefore spanned by the orthonormal basis (δ(x,n))(x,n)∈V+ if σn Note that paths of the form x = σy, with s(σn) = r(y), exhaust all possible x ∈ σ−n Using that PDy Dyδ(x,n) = nδ(x,n) for (x, n) ∈ V+ y , we compute that y Tr(PDy e−tDy) = ∞Xn=0 Xx∈σ−n G ({y}) e−tn = ∞Xn=0 #{σ ∈ En : s(σn) = r(y)}e−tn. 28 Let A denote the edge adjacency matrix of G and rσ(A) its spectral radius. If G is primitive, (i.e. all entries of Ak are positive for some integer k > 0) we let w ∈ CE its ℓ2-normalized Perron-Frobenius vector. It follows from [56, Lemma 3.7] that there is an α0 ∈ [0, 1) such that #{σ ∈ En : s(σn) = r(y)} = kwkℓ1 wr(y)rσ(A)n + O((α0rσ(A))n), as n → ∞. We can conclude that there is a function f holomorphic in Re(t) > log rσ(A)+log α0 such that Tr(PDy e−tDy) = kwkℓ1 wr(y) 1 − rσ(A)e−t + f (t). Therefore, Tr(PDy e−tDy) − log α0 whenever G is primitive. kwkℓ1 wr(y) t−log(rσ(A)) has a holomorphic extension to Re(t) > log rσ(A) + More generally, if G is primitive, the method above shows that for two finite paths µ and ν we can compute that Tr(PDy SµS∗ ν e−tDy) =δµ,ν kS∗ G ({y}) ℓ2e−tn µδ(x,n)k2 ∞Xn=0 Xx∈σ−n ∞Xn=µ #(cid:8)σ ∈ En−µ : r(σ1) = s(µ), s(σn−µ) = r(y)(cid:9)e−tn µ−1Xn=0 Xx∈σ−n kS∗ µδ(x,n)k2 ℓ2e−tn G ({y}) = + δµ,ν =δµ,νws(µ)wr(y) e−tµ 1 − rσ(A)e−t + δµ,ν fµ,y(t), for a function fµ,y holomorphic in Re(t) > log rσ(A) + log α0. We conclude that Tr(PDy SµS∗ ν e−tDy) − δµ,νwd(µ)wr(y) rσ(A)−µ t − log(rσ(A)) , has a holomorphic extension to Re(t) > log rσ(A) + log α0 whenever G is primitive. As such, βDy = log(rσ(A)) and Dy has positive essential spectrum. Example 3.7. Let OE be a Cuntz-Pimsner algebra defined from a strictly W-regular (recall Definition 2.25) finitely generated and projective bi-Hilbertian bimodule EA and a positive trace τ on the coefficient algebra A. The semifinite spectral triple considered in Lemma 2.28 (see page 22) also has positive Trτ -essential spectrum assuming a criticality condition on τ that we formulate below (see Definition 6.2 on page 51). The heat trace asymptotics are slightly more involved, and we compute these explicitly in Subsection 6.2 under a condition on τ previously studied by Laca-Neshveyev [45] in the context of KMS-states. However, for a general τ we can proceed as in the proof of Lemma 2.28 to deduce the following. Proposition 3.8. For any strictly W -regular fgp bi-Hilbertian bimodule E over the unital C ∗-algebra and a positive trace τ on A, Trτ (PDψ e−tDψ ) = e−tnτ∗(E⊗An), ∞Xn=0 29 where τ∗ : K0(A) → R denotes the map induced by τ on K-theory. In particular, Trτ (PDψ e−tDψ ) does not depend on the choice of inner products on E but only on τ and the bimodule structure on E. Proof. We compute that Trτ (PDe−tDψ ) = ∞Xn=0 e−tψ(n,n) Trτ (Pn,n) = e−tn Trτ (Qn,n) = ∞Xn=0 ∞Xn=0 Xρ=n e−tnτ ((eρeρ)A). On the other hand, τ∗(E⊗An) = (τ ⊗ TrMN(n))(pE⊗An) where pE⊗An ∈ MN (n)(A) is a projection representing E⊗An. Using the choice of frame (ej)N j=1, we can take N (n) := N n and pE⊗An := ((eµeν )A)µ=ν=n. In this choice of representing projection, τ∗(E⊗An) = (τ ⊗ TrMN(n))(pE⊗An) = TrMN(n)((τ ((eµeν )A)µ=ν=n)) = Xρ=n τ ((eρeρ)A). This computation shows that it is in general difficult to compute βD. In this case βD depends on the asymptotic properties of the sequence (τ∗(E⊗An))n∈N as n → ∞. For a simple tensor σ ∈ E⊗m write σ = σσ, where the initial segment σ will be of a length understood from context (σ = µ in the next computation). With this notation, we can ν ∈ OE, where µ ∈ E⊗k, ν ∈ E⊗l are simple tensors. compute our functional on a typical SµS∗ We find Trτ (PDSµS∗ ν e−tDψ) = δµ,ν = δµ,ν = δµ,ν ∞Xn=0 ∞Xn=µ Xσ=n ∞Xn=µ Xσ=n e−tn Trτ (SµS∗ ν Qn,n) (3.3) e−tnτ ((S∗ µeσS∗ ν eσ)E⊗(n−µ)) e−tnτ ( ( (µeσ)Eµeσ (νeσ)Eµeσ)E⊗(n−µ)). Example 3.9. Consider a length function ℓ on a countable group Γ as in Subsection 2.2.4. Definition 3.10. We define the critical value of (Γ, ℓ) as β(Γ, ℓ) := inf{t ≥ 0 :Xγ∈Γ e−tℓ(γ) < ∞}. IfPγ∈Γ e−tℓ(γ) ր ∞ as t ց β(Γ, ℓ), we say that ℓ is critical. It follows directly from Definition 3.1 that the operator Dℓ appearing in Proposition 2.31 (see page 24) has positive essential spectrum as long as ℓ is critical. The heat trace of an element aλg ∈ cb(Γ) ⋊alg Γ is given by Tr(PDℓ aλge−tDℓ) = Tr(aλge−tDℓ) = δe,gXγ∈Γ a(γ)e−tℓ(γ). 30 Similar computations can be carried out for the semifinite spectral triple constructed in Propo- sition 2.34 using a Hilbert space valued cocycle cΓ (see page 26). Note that Therefore, the heat trace of an element aλg ∈ cb(Γ) ⋊alg Γ is given by + 1(cid:19) . PDc f (g) = 1 kcΓ(g)kHΓ 2(cid:18) cS(cΓ(g)) 2Xγ∈Γ(cid:28)δγ, τ(cid:18) cS(cΓ(g)) δe,gXγ∈Γ kcΓ(g)kHΓ 1 a(γ)e−tℓ(γ) = 2 Trτ (PDc aλge−tDc) = = 1 1 2 + 1(cid:19) a(g−1γ)δgγ(cid:29) e−tℓ(γ) = Tr(PDℓ aλge−tDℓ). Here we use that τ (cS(v)) = 0 for any v ∈ HΓ which holds due to the fact that we can pick a w ∈ HΓ orthogonal to v and compute that τ (cS (v)) = τ (cS (w)cS (v)cS (w)) = −τ (cS (w)2cS(v)) = −τ (cS(v)). If the length function ℓ(γ) := kcΓ(γ)kHΓ associated with cΓ is critical, we say that cΓ is critical. We conclude that the semifinite spectral triple from Proposition 2.34 has positive essential spectrum if cΓ is critical. 3.2 To-plitz or not To-plitz We proceed under the same working conditions as in the previous section to construct states from spectral triples. Recall that N+ = PDNPD and that K+ N = PDKNPD. Definition 3.11. Let (A, H, D, N, T) be a semifinite spectral triple. We define the Toeplitz algebra of (A, H, D, N, T) as The saturated Toeplitz algebra of (A, H, D, N, T) is defined by TA := PDAPD + K+ N ⊆ N+. TA,D := C ∗ [s∈R eisDTAe−isD! = C ∗ [s∈R eisDPDAPDe−isD + K+ N! ⊆ N+. Proposition 3.12. The Toeplitz algebra TA of a unital semifinite spectral triple (A, H, D, N, T) is a C ∗-algebra. Proof. It follows from Lemma 2.11 that [PD, a] ∈ KN for all a ∈ A. Therefore, the mapping βD : A → N+/K+ N, is a ∗-homomorphism and βD(A) ⊆ N+/K+ the preimage of βD(A) under the quotient mapping N+ → N+/K+ algebra. a 7→ PDaPD mod K+ N, N is a closed C ∗-subalgebra. By definition, TA is N and is therefore a C ∗- The reader should note that Proposition 3.12 also holds in the non-unital setting because it only relies on Lemma 2.11 which holds non-unitally. 31 Proposition 3.13. The saturated Toeplitz algebra of (A, H, D, N, T) carries an R-action σ+ : R → Aut(TA,D) defined by σ+ s (T ) := PDeisDT e−isDPD = PDeisDT e−isDPD, s ∈ R, T ∈ TA,D. This proposition is a consequence of that TA,D is constructed as the saturation of TA under the action σ+ extended to N+. We note that if we identify TA,D with a subalgebra of N, we can also write σ+ s (T ) := eisDT e−isD = eisDT e−isD. Define a one-parameter family of states (φt,0)t>β on TA,D by φt,0(T ) := T(PDT e−tD) T(PDe−tD) . We shall compose the family (φt,0)t>β with an "extended limit" as t → β: Definition 3.14. An extended limit as t → β is a state ω ∈ L∞(β,∞)∗ such that ω(f ) = 0 whenever limt→β f (t) = 0. For an extended limit ω and f ∈ L∞(β,∞), we write Let ω be any extended limit and define lim t→ω f := ω(f ). φω,0 : TA,D → C, φω,0(T ) := ω ◦ φt,0(T ) = lim t→ω φt,0(T ). Lemma 3.15. Let (A, H, D, N, T) be a semifinite Li1-summable spectral triple with positive T-essential spectrum (see Definition 3.1 on page 27). For any extended limit ω, the functional φω,0 is a state on TA,D. Moreover φω,0(T ) = 0 for all T ∈ K+ N. Proof. It is immediate that φω,0 is a state. For the statement that φω,0(T ) = 0 for all T ∈ K+ N, we observe that since φω,0 is a state, it is also norm-continuous. It therefore suffices to prove that φω,0(T ) = 0 for all projections T ∈ N+ with T(T ) < ∞. For such T , we can estimate that φt,0(T ) = T(PDT e−tD) T(PDe−tD) ≤ T(T ) T(PDe−tD) . Since, T(PDe−tD) ր ∞ as t ց β, it follows that limt→β φt,0(T ) = 0. We conclude that for all projections T ∈ N+, with T(T ) < ∞, and any extended limit ω as t → β, ω ◦ φt,0(T ) = 0. Due to Lemma 3.15, we can make the following definition. Definition 3.16. Define the C ∗-algebra AD := TA,D/K+ N and the state φω on AD as φω(T mod K+ N) := φω,0(T ). The state φω,0 also restricts to a state on TA, and Lemma 3.15 implies that φω,0TA descends to a state on A via the ∗-epimorphism βD : A → TA/K+ N (see the proof of Proposition 3.12 on page 31). To analyse the situation of the state on AD versus that on A, we consider the following ideal I := {a ∈ A : PDaPD ∈ K+ N} 32 so that PDIPD = PDAPD ∩ K+ N. Since TA ⊆ TA,D, we obtain a commuting diagram 0 0 / K+ N = / K+ N TA A/I / 0 / TA,D / AD / 0, with exact rows. The mapping A/I → AD is indeed injective by the four lemma. We identify A/I with a subalgebra of AD. The induced mapping γ : A → AD is compatible with the states φω and φω,0 in the sense that φω,0(PDaPD) = φω(γ(a)) for a ∈ A. The R-action σ+ sition follows from the construction of TA,D as the saturation of TA under σ+. s (T ) := eisDT e−isD on TA,D induces an R-action on A/I. The following propo- Proposition 3.17. Let β ∈ R. The algebra AD carries an R-action σ : R → Aut(AD) defined by declaring the quotient mapping TA,D → AD to be equivariant. The C ∗-algebra AD is the saturation of A/I under the action σ, i.e. AD is generated in N+/K+ N by ∪s∈Rσs(A/I). The aim of our construction is to obtain a KMS state on A, or, failing that, on A/I. As a first step we introduce conditions ensuring that we at least get a KMS state on TA,D, and so on AD. To this end we make the following assumption Definition 3.18. Let (A, H, D, N, T) be a semifinite spectral triple. We say that a subset S ⊂ A is analytically generating at β if the set PDSPD + K+ generates the Toeplitz algebra TA as a C ∗-algebra and satisfies that eβDPDSPDe−βD ⊂ N+. We say that the semifinite spectral triple (A, H, D, N, T) is β-analytic if it admits an analyti- cally generating set at β. We note that this condition is empty if β = 0. The condition of being β-analytic is just requiring that we have enough analytic elements in TA,D to verify the KMS condition. Indeed we have the following. Proposition 3.19. Let (A, H, D, N, T) be a semifinite spectral triple and β ∈ R. Consider the following statements: i) The semifinite spectral triple (A, H, D, N, T) is β-analytic. ii) There is a dense σ+-invariant subspace T 0 A,D ⊆ TA,D of elements satisfying that for any t (T ) has a bounded holomorphic extension A,D the function fT : R → N, fT (t) := σ+ T ∈ T 0 to the strip {z ∈ C : Im(z) ∈ (−β, 0)}. iii) There is a dense σ-invariant subspace A0 D ⊆ AD of elements satisfying that for any a ∈ AD the function fa : R → N, fa(t) := σt(a) has a bounded holomorphic extension to the strip {z ∈ C : Im(z) ∈ (−β, 0)}. It holds that i) implies ii) which implies iii). If I := {a ∈ A : PDaPD ∈ K+ i). N} = 0, iii) implies 33 / / / / / _ _ / / / / / Proof. It is clear that i) implies ii) since for an analytically generating set S at β we can take A,D to be the σ+-invariant ∗-algebra generated by PDSPD ∪ FD, where FD ⊆ K+ T 0 N is the dense two sided ideal in N generated by the spectral projections of D over compact intervals in R. The implication ii)⇒iii) is seen from taking A0 If I = 0, the set S = A0 a bounded operator in N. Thus eβDPDSPDe−βD ⊂ N+ and PDSPD + K+ TA. We conclude that S is analytically generating at β and that iii) implies i) if I = 0. D ∩ A is dense in A, and for every s ∈ S we find that eβDse−βD is N plainly generates A,D ∩ K+ N). D := T 0 A,D/(T 0 Proposition 3.20. Let (A, H, D, N, T) be a βD-analytic Li1-summable spectral triple with pos- itive T-essential spectrum. For any extended limit ω as t → βD, the state φω,0 is a βD-KMS state on TA,D for the one-parameter group σ+. Proof. It follows from Proposition 3.19 that the dense subalgebra T 0 A,D ⊆ TA,D consists of βD-analytic elements of TA,D. The twisted trace property relative to the one-parameter group σ+ holds on T 0 A,D by direct computation: for all T1, T2 ∈ T 0 A,D φω(T1T2) = lim t→ω T(T1T2e−tD) T(PDe−tD) = lim t→ω T(etDT2e−tDT1e−tD) T(PDe−tD) = φω(σ−iβ(T2)T1). By definition, φω,0 is a βD-KMS-state for σ+. Corollary 3.21. Let (A, H, D, N, T) be a βD-analytic Li1-summable spectral triple with positive T-essential spectrum. For any extended limit ω as t → βD, the state φω is a βD-KMS state on AD for the one-parameter group σ, defined in Proposition 3.17. Proof. This follows from Proposition 3.19 and Proposition 3.20 because φω,0 vanishes on the compacts and the fact that φω is induced from φω,0. In practice, for a βD-analytic Li1-summable spectral triple with positive T-essential spectrum, we will want to check that in fact TA,D = TA. In this case AD = A/I and φω induces a KMS-state on A/I. In the often occuring special case PDAPD ∩ K+ N = 0, we obtain a KMS state on the algebra A. In practice, these things are all checkable and we will do so in several examples in the subsequent sections. The special case β = 0 has been addressed by Voiculescu, [63, Proposition 4.6] under the assumption that D is positive. Exponential β-compatibility is superfluous when βD = 0. By Proposition 3.2, when βD = 0, positive T-essential spectrum is equivalent to T(PD) = ∞. Theorem 3.22. Let (A, H, D, N, T) be a unital Li1-summable semifinite spectral triple with βD = 0 and T(PD) = ∞. Then A has a tracial state. Indeed, for any extended limit ω as t → 0, is a tracial state. φω(a) := lim t→ω T(PDae−tD) T(PDe−tD) Theorem 3.22 applies to unital Li(0),1-summable semifinite spectral triples with T(PD) = ∞ since Li(0),1-summability implies βD = 0. 34 Remark 3.23. If the heat trace has an asymptotic expansion as in Theorem 2.10, then Theorem 3.22 can be further simplified. Assume that there is a p > 0 and that for any a ∈ A, there is a φ0(a) ∈ C such that φ0(1) 6= 0 and T(PDae−tD) = φ0(a)t−p + O(t−p+ǫ), as t → 0, for some ǫ > 0 (which can depend on a). Since φ0(1) 6= 0, it follows that φ0 is continuous in the C ∗-norm on A and φω(a) = φ0(a) φ0(1) for all a ∈ A. In fact, φ0(a) = Γ(p)Resz=pζ(z; PDa,D). The construction of the state φω in Corollary 3.21 involves the operator PD. We shall now provide a result allowing us to remove PD from the definition of φω in the presence of certain symmetries on the spectral triple. The result provides a checkable set of conditions to compute φω by means of asymptotics of e−tD. Lemma 3.24. Let (A, H, D, N, T) be a unital Li1-summable semifinite spectral triple and β ≥ 0 a number such that T(e−tD) < ∞ for t > β and T(e−tD) ր ∞ as t ց β. Assume that there exists self-adjoint operators γ1, . . . , γN ∈ N ∩ A′ such that 1. PN 2. γj Dom(D) ⊆ Dom(D) and [D, γj]+ := Dγj + γj D has a bounded extension to H; 3. For j = 1, . . . , N and some ǫ > 0, the function t 7→ e−tDγjetD extends to a norm j is invertible; j=1 γ2 continuous function from the interval [β, β + ǫ) to N with e−tDγjetD = γj. lim t→β Then βD = β and D has positive T-essential spectrum. Moreover, for any extended limit ω as t → βD, φω(a) = lim t→ω T(ae−tD) T(e−tD) . Example 3.25. Before proceeding with the proof of the lemma, we give some examples of how the operators γ1, . . . , γN ∈ N ∩ A′ can arise. The most trivial instance is when (A, H, D, N, T) is even, in which case the grading γ will satisfy the conditions of Lemma 3.24. Here is a more geometric example. Let (C ∞(M ), L2(M, S), /D) denote the spectral triple de- fined from a Dirac operator on a closed Riemannian manifold M as in Proposition 2.14 (see page 14). If we take a collection X1, . . . , XN ∈ C ∞(M, T M ) of vector fields spanning the tan- gent bundle T M in all points, the collection of Clifford multiplication operators γj := cS(Xj), j = 1, . . . , N is readily verified to satisfy the conditions of Lemma 3.24. This construction ex- tends to semi-finite spectral triples defined from the fibrewise Dirac operator of a Riemannian spinc submersion π : M → B (see more in [40]) and a measure on B by taking X1, . . . , XN to be vertical vector fields spanning the vertical tangent bundle ker dπ in all points of M . Proof. Define C := k(PN j )−1k and recall that FD := DD−1 modulo a finite trace projec- j=1 γ2 35 tion. For any self-adjoint a ∈ A, we estimate T(FDae−tD) = C 2 γj(FDa + aFD)e−tDγj = γ2 1 2 + − + = C 2 C 2 C 2 C 2 C 2 = − T T(cid:16)(FDa + aFD)e−tD(cid:17) ≤ NXj=1 γj(FDa + aFD)γje−tD + T NXj=1 γj(FDa + aFD)(e−tDγjetD − γj)e−tD = T NXj=1 j (FDa + aFD)e−tD + T NXj=1 γj([FD, γj]+a + a[FD, γj]+)γje−tD T NXj=1 γj(FDa + aFD)(e−tDγjetD − γj)e−tD ≤ T NXj=1 T γj(FDa + aFD)(e−tDγjetD − γj)e−tD NXj=1 γj([FD, γj]+a + a[FD, γj]+)γje−tD = o(T(e−tD)), τ NXj=1 γj(FDa + aFD)(e−tDγjetD − γj)e−tD = o(T(e−tD)), τ NXj=1 ≤ − T(cid:16)FDae−tD(cid:17) − T NXj=1 + C 2 C 2 γj([FD, γj]+a + a[FD, γj]+)γje−tD Since [D, γj ]+ is bounded, [FD, γj]+ is compact and an approximation argument by finite T- rank operators shows that as t ց β. By norm continuity of t 7→ e−tDγjetD we can also deduce that as t ց β. In conclusion, for a self-adjoint a, T(FDae−tD) = −T(FDae−tD) + o(T(e−tD)). We can conclude that τ (FDae−tD) = o(T(e−tD)) as t ց β. Since 2PD = FD + 1, we have for any a ∈ A that T(ae−tD) = 2T(PDae−tD) + o(T(e−tD)), as t ց β. 36 In particular β = βD and T(PDe−tD) ր ∞ as t ց β. We compute that T(ae−tD) T(e−tD) = T(PDae−tD) T(PDe−tD) + o(1), as t ց β. In particular, for any extended limit ω as t → β, T(ae−tD) T(e−tD) lim t→ω = lim t→ω T(PDae−tD) T(PDe−tD) . This concludes the proof of the lemma. 3.3 Modular spectral triples and modular index theory Modular spectral triples and their (equivariant) index theory were considered in [10, 12, 7], with the definition laid out most clearly in [58, Definition 2.1]. These were defined in order to study the (equivariant) index theory of KMS weights associated to periodic flows, so one might wonder how modular spectral triples fit into our scheme. Given a KMS state ψ : B → C with inverse temperature β for a one-parameter group σ : R → Aut(B) on a unital C ∗-algebra and a faithful expectation onto the fixed point algebra Φ : B → Bσ, we can emulate the constructions that inspired the definition of modular spectral triples. First we construct the right C ∗-module X over Bσ by completing B in the norm coming from the inner product (xy)Bσ = Φ(x∗y). Then we can use [45] to construct Trψ : KBσ (X) → C the trace dual to ψBσ . The action σ induces a one parameter unitary group on L2(X, ψ), and we let D be the generator of this one parameter group. By [7], when the action σ is periodic, the data (B, L2(X, ψ), D, KBσ (X)′′, φD), where φD(a) := Trψ(e−βD/2ae−βD/2), gives us a modular spectral triple. The operator D is affiliated to the semifinite algebra (K(X)σφD )′′ Proposition 3.26. Let (A, H, D, N, T) be a semifinite spectral triple such that that for all t > β T(PDe−tD) < ∞ and limtցβ T(PDe−tD) = ∞. Assume that PDAPD ∩ KN = {0} and the spectral triple is β- analytic. Then we obtain the KMS state φω : AD → C, and provided that D has discrete spectrum we obtain an expectation Φ : AD → Aσ Provided that the group σ is periodic we then obtain a finitely summable modular spectral triple D onto the fixed point subalgebra. (AD, L2(AD, φω), D, (KAσ D (AD))′′, φD), where the operator D generates the one-parameter group induced by σ on L2(AD, φω) and φD := Trφω (e−βD/2 · e−βD/2). The spectral dimension is 1. 37 Proof. The existence of the KMS state φω comes from Corollary 3.21. In general the action σ is real, but assuming that the operator D has discrete spectrum, the action will factor through a (compact) torus. To see this, one takes a rational basis of the eigenvalues (possibly an infinite basis), and takes a product over the circles corresponding to these individual actions. Consequently, by averaging over this torus, there is an expectation Φ : N → Nσ onto the fixed point algebra for t 7→ (T 7→ eitDT e−itD). Of course D is affiliated to the fixed-point algebra. Finally if the action σ is periodic then [7] proves that we have a modular spectral triple, and that φD((1 + D2)−s/2) < ∞ for s > 1. For circle actions there is a local index formula in twisted cyclic theory, but for real actions factoring through a torus there is not. One serious issue that comes up in this more general setting is the compactness of the resolvent of D, and determining summability. We leave this issue to another place. 4 The KMS-state φω and Dixmier traces Dixmier traces. For a decreasing function ψ : [0,∞) → (0,∞) we denote Ψ(t) :=R t In the present section we discuss a relation between the trace φω from Theorem 3.22 and 0 ψ(s)ds. Let Lψ(T) be the principal ideal defined as in Definition 2.6 (see page 9). Let ET be the spectral projection of an operator T affiliated with N and let nT(s, T ) := T(ET (s,∞)) be its distribution function. The following result extends [48, Lemma 12.6.3]. Lemma 4.1. Let ψ be a regularly varying function of index −1. Let T ∈ Lψ(T) be strictly positive and µT(t, T ) ∼ ψ(t), t → ∞. For every q > 0 we have T(e−T −q/t) ∼ Γ(1 + 1 q )ψ−1(t− 1 q ), t → ∞. Proof. The assumption µT(t, T ) ∼ ψ(t) implies µT(t, T q) ∼ [ψ(t)]q, t → ∞. Since the distribu- 1 tion function is an inverse of the singular values function, it follows that nT(s, T q) ∼ ψ−1(s q ), s → 0+. Next we have T(ET −q (t)) ∼ nT(1/t, T q), t → ∞. Thus, q ), t → ∞. T(ET −q (t)) ∼ ψ−1(t− 1 Since ψ is varying regularly with index −1, [3, Theorem 1.5.12] implies that ψ−1 varies regularly with index −1, too. Thus, T(ET −q ) varies regularly with index 1 q . Writing the heat trace as a Laplace transform T(e−T −q/t) =Z ∞ 0 e−z/t dT(ET −q (z)) and using the Karamata theorem [44, Chapter IV, Theorem 8.1] we obtain T(e−T −q/t) ∼ Γ(1 + 1 q )ψ−1(t− 1 q ), t → ∞. 38 Let Pa : L∞(0,∞) → L∞(0,∞) be the exponentiation operator defined by (Paf )(t) := f (ta), t > 0. Definition 4.2. [11] An extended limit ω as t → ∞ on L∞(0,∞) is said to be exponentiation invariant if lim t→ω (Paf )(t) = lim t→ω f (t) for every f ∈ L∞(0,∞) and every a > 0. Definition 4.3. Let ψ : [0,∞) → (0,∞) be a regularly varying function of index −1. For any extended limit ω as t → ∞ on L∞(0,∞) a linear extension of the weight Tω,ψ(T ) := lim t→ω µT(s, T ) ds, 0 ≤ T ∈ Lψ(T) 0 1 Ψ(t)Z t is said to be a Dixmier trace on Lψ(T). Remark 4.4. Usually Dixmier traces are defined on Lorentz ideals corresponding to the function Ψ (which are strictly larger than Lψ(T)) by exactly the same formula as in Defninition 4.3 (see e.g. [20, 15, 48]). Then, Dixmier traces on Lψ(T) are restrictions of those on Lorentz ideal to Lψ(T). Since we do not deal with Lorentz ideals here, it is convenient to define Dixmier traces directly on Lψ(T). Also, it should be pointed out that on Lorentz ideals to define Dixmier traces one needs an additional assumption on ω: either dilation invariance [20, 48] or exponentiation invariance [25, 61]. As it was shown in [60, Theorem 17] these requirements are redundant on Lψ(T). The proof of the following theorem is the same as that of [48, Theorem 8.5.1] and thus omitted. Note however, that in [48] the result was proved for Lorentz ideals and required dilation invariance of the extended limit ω. For the case of Lψ(T) one can refer to [60, Lemma 15] to remove this assumption. Theorem 4.5. Let f ∈ C 2[0,∞) be a bounded function such that f (0) = f ′(0) = 0. Let T ∈ Lψ(T) be positive and let B ∈ N. For every extended limit ω as t → ∞ on L∞(0,∞) we have t→ω(cid:18) 1 Ψ(t)Z t lim 1 T(f (sT )B) ds s2(cid:19) =Z ∞ 0 f (s) ds s2 · lim t→ω(cid:18) 1 Ψ(t)Z t 1 T(e−(sT B)−1 ) ds s2(cid:19) . Below we will need the relation between generalised heat kernels and Dixmier traces on Lψ(T), which was proved in [25] under the additional assumption that AΨ(α) := lim t→∞ Ψ(tα) Ψ(t) exists for every α > 0. (4.1) Recall the notation Ψ(t) =R t 0 ψ(s)ds. The corresponding (natural) assumption on ψ is that α · lim t→∞ ψ(tα)tα−1 ψ(t) exists for every α > 0. (4.2) The number appearing in Equation (4.2) equals AΨ(α). Note that condition (4.2) implies the condition (4.1) and regular variation of ψ with index −1 [25, Proposition 1.7]. 39 Let H : L∞(0,∞) → L∞(0,∞) be the Cesaro mean defined as follows: (Hf )(t) := f (s) ds, t > 0. 1 tZ t 0 Let MΨ : L∞(0,∞) → L∞(0,∞) be the Cesaro mean twisted by Ψ, that is (MΨf )(t) : = [(H(f ◦ Ψ−1)) ◦ Ψ](t) = Lemma 4.6. If ψ satisfies condition (4.2), then 1 Ψ(t)Z t 0 f (s)ψ(s) ds, t > 0. MΨ ◦ Pa − Pa ◦ MΨ : L∞(0,∞) → C0(0,∞) for every a > 0. Proof. For f ∈ L∞(0,∞) we have Ψ(t)Z t (MΨ ◦ Paf )(t) = 1 0 f (sa)ψ(s) ds = 1 Ψ(t)Z ta 0 f (s)ψ(s1/a) s1/a−1 a ds. Using conditions (4.1) and (4.2) we obtain (MΨ ◦ Paf )(t) ∈ AΨ(a)AΨ( 1 a ) 1 Ψ(ta)Z ta 0 f (s)ψ(s) ds + C0. Direct calculations show that AΨ(a)AΨ( 1 a ) = 1. Thus, (MΨ ◦ Pa − Pa ◦ MΨ)f ∈ C0. Lemma 4.7. If ψ : [0,∞) → (0,∞) is a decreasing function with regular variation of index −1. Then for any d > 0, as t → ∞, ψ(t) = o((log(t))−d). Proof. Since ψ is decreasing and limt→∞ Cǫ > 0 such that for t ∈ (2k, 2k+1], ψ(2t) ψ(t) = 2−1, we can for any ǫ ∈ (0, 1) find a constant ψ(t) ≤ Cǫ(2 − ǫ)−k. For details, see [32, Proposition 2.21]. Therefore, ψ(t) = o(t−1/2) and the lemma follows. For the next result we need to assume the (stronger) condition, that the function ψ satisfies t2ψ(t) ψ−1(1/t) lim t→∞ = c, (4.3) for some constant c > 0. Remark 4.8. For every k ∈ Z, the functions ψ(t) = logk t tion (4.3). The functions ψ = Ψ′ with Ψ(t) = elogβ t do not satisfy (4.3) for any β > 0. and ψ(t) = logk(log t) t·log t t satisfy condi- 40 In the following result we use a singular values function of an unbounded operator affiliated with N. For such operators the formula (2.1) cannot be used. The singular values function of an operator T affiliated with N is defined [24] as follows: µT(t, T ) := inf{s ≥ 0 : nT(s, T ) ≤ t}. Theorem 4.9. Let d > 0 and ψ : [0,∞) → (0,∞) be a decreasing function satisfying condi- tions (4.2) and (4.3). Assume that (A, H, D, N, T) is a unital semifinite spectral triple such that 1. T is infinite; 2. D is positive; 3. µT(t, D) ∼ ψ(t)−1/d. Then (A, H, D, N, T) is Li(0),1-summable with positive T-essential spectrum. Moreover, for every a ∈ AD and every exponentiation invariant extended limit ω as t → ∞ we have φω(a) = Tω,ψ(a(1 + D2)−d/2), where ω := ω◦(JMΨJ) and J : L∞(0,∞) → L∞(0,∞) is defined as the pullback along t 7→ t−1. Proof. Condition (4.2) on ψ implies that ψ has regular variation of index −1 so ψ(t) = o((log(t))−d) by Lemma 4.7. Therefore, µT(t, D) = o(log(t)) and the semifinite spectral triple (A, H, D, N, T) is Li(0),1-summable. Since T is infinite, T(e−tD) ր ∞ as t ց 0. The operator D therefore has positive T-essential spectrum. Thus, φω(a) = lim t→ω T(ae−tD) T(e−tD) is a trace on A by Theorem 3.22. The assumption on D implies that µT(t, (1+ D2)−d/2) ∼ ψ(t), t → ∞. Applying Lemma 4.1 with T = (1 + D2)−d/2 and q = 1/d yields Using the properties of extended limits, the definitions of Pa and J and Lemma 4.6, we obtain Γ(1 + d) lim t→ω T(e−D/t) ∼ Γ(1 + d)ψ−1(t−d), t → ∞. (J ◦ MΨ) T(ae−D/t) ψ−1(t−d)! (J ◦ MΨ ◦ Pd) T(ae−(tD−d)−1/d (J ◦ Pd ◦ MΨ) T(ae−(tD−d)−1/d ψ−1(1/t) ψ−1(1/t) lim t→ω lim t→ω Γ(1 + d) 1 1 1 Γ(1 + d) φω(a) = = = ) ) ! ! . (4.4) Using the definition of MΨ and assumption (4.3) we obtain MΨ T(ae−(tD−d)−1/d ψ−1(1/t) ) ! = T(ae−(sD−d)−1/d ) ψ−1(1/s) ψ(s) ds T(ae−(sD−d)−1/d ) ds s2 + C0(0,∞). (4.5) 0 1 Ψ(t)Z t Ψ(t)Z t 1 0 41 ∈ Since ω is exponentiation invariant extended limit, combining (4.4) and (4.5) we obtain φω(a) = 1 Γ(1 + d) · 1/t→ω(cid:18) 1 Ψ(t)Z t lim 1 T(ae−(sD−d)−1/d ) ds s2(cid:19) . Now we apply Lemma 4.5 twice with T = D−d, f (x) = e−x−1/d Since and then with f (x) = e−x−1 . we obtain φω(a) = lim 1/t→ω 0 Z ∞ J(cid:18) 1 Ψ(t)Z t 1 e−x−q dx x 1 q = Γ(1 + ), T(ae−(sD−d)−1 ) ds s2(cid:19) = Tω,ψ(cid:16)a(1 + D2)−d/2(cid:17) , where the last equality follows from [25, Theorem 4.7]. We can now provide sufficient conditions on a general Li(0),1-summable unital semifinite spec- tral triples ensuring a relation between the trace φω of Theorem 3.22 and Dixmier traces. Corollary 4.10. Let (A, H, D, N, T) be a unital semifinite spectral triple with T(PD) = ∞. Assume that there is a number d > 0 and a decreasing function ψ : [0,∞) → (0,∞) with regular variation of index −1 satisfying conditions (4.2) and (4.3) and µ(t, PDD) ∼ ψ(t)−1/d. Then, βD = 0 and for any exponentiation invariant extended limit ω as t → ∞ and a ∈ AD, φω(a) = Tω,ψ(PDa(1 + D2)−d/2), where ω is as in Theorem 4.9. The corollary follows immediately from Theorem 4.9 applied to the unital semifinite spectral triple (T, PDH, PDD, N+, T) where T is the ∗-algebra generated by PDAPD. Example 4.11. Let us revisit the spectral triple constructed in Proposition 2.16 (see page 15). If ψ : [0,∞) → (0,∞) is a smoothly varying function with limt→0 ψ(t) = 0 and ψ(t)−1 = O(t1/n) as t → ∞, and /D a Dirac operator on a Riemannian closed n-dimensional manifold M , a ψ-summable spectral triple (C ∞(M ), L2(M, S), /Dψ) was constructed in Proposition 2.16, where /Dψ := F /Dψ( /Dn)−1. The Weyl law for /D and Theorem 2.10 applied to B = P /D shows that the order of the spectral asymptotics of /D coincides with the order of the spectral asymptotics of P /D /D so µ(t, P /D /Dψ) ∼ cψ(t)−1 for some constant c > 0. If ψ has regular variation of index −d for a d > 0, Corollary 4.10 shows that the tracial state on C(M ) defined from the spectral triple (C ∞(M ), L2(M, S), /Dψ) takes the form φω(a) = c′ Trω,ψ(P /Daψ( /Dn)1/d), for some proportionality constant c′. Applying Connes' trace theorem as in [32, Theorem 9.1], it follows that φω(a) =ZM− adV, 42 where dV denotes the Riemannian volume measure on M andZM− the normalized integral. The computation above requires ψ to have strictly negative index of regular variation. We note that by Proposition 2.18 (see page 16), the computation above can only extend to the spectral triple from Proposition 2.19 (see page 17) on a crossed product by a local diffeomorphism if the local diffeomorphism acts isometrically. 5 The KMS-state φω in examples We are now in a state where we can compute the KMS-states associated with the spectral triples considered in Subsection 2.2 (see page 13). The computations for the KMS-states associated with the unbounded Kasparov cycle on Cuntz-Pimsner algebras considered in Subsection 2.2.3 (see page 19) are more involved and dedicated a separate section, Section 6 (see page 50). 5.1 Dirac operators on closed manifolds For a closed manifold M with a Dirac operator /D acting on a Clifford bundle S → M , we consider the spectral triple (C ∞(M ), L2(M, S), /D) as in Proposition 2.14 (see page 14). The following theorem can be deduced immediately from Example 3.5 (see page 28) or from Corollary 4.10 and Connes' trace theorem for pseudo-differential operators. Theorem 5.1. Let (C ∞(M ), L2(M, S), /D) be the spectral triple associated with a Dirac op- erator on a closed manifold, ω an extended limit as t → 0 and φω the associated tracial state from Theorem 3.22 (see page 34). The trace φω is the normalized volume integral on M , i.e. for a ∈ C(M ), φω(a) =ZM− a dV. For completeness, let us describe the Toeplitz algebras TC(M ), TC(M ), /D and the flow σ on C(M ) /D in this example. We remark that since φω is a trace in this case, the flow induced from our construction is irrelevant for the study of φω but it could nevertheless serve to clarify the constructions of Subsection 3.2. The relevant algebras are all contained in the C ∗-algebra C∗(M, S) defined as the C ∗-closure of the ∗-algebra Ψ0 Ψ0 cl(M, S) of zeroth order classical pseudo- differential operators acting on L2(M, S). It is well known that Ψ0 C∗(M, S) fits into a short exact sequence 0 → K(L2(M, S)) → Ψ0 C∗(M, S) symb −−−→ C(S∗M, End(S)) → 0, where symb denotes the continuous extension of the principal symbol mapping Ψ0 C ∞(S∗M, End(S)). Since P /D is a projection in Ψ0 Ψ0 cl(M, S) → cl(M, S), we can consider the C ∗-algebra C∗,+(M, S) := P /DΨ0 C∗(M, S)P /D and we obtain a short exact sequence 0 → K(P /DL2(M, S)) → Ψ0 C∗,+(M, S) symb −−−→ C(S∗M, End(p /DS)) → 0, where p /D ∈ C ∞(S∗M, End(S)) denotes the principal symbol of P /D. The algebras TC(M ) and 43 TC(M ), /D are characterized by the following commuting diagram with exact rows 0 0 0 / K(P /DL2(M, S)) TC(M ) = / K(P /DL2(M, S)) TC(M ), /D = symb symb C(M ) C(M ) /D / 0 / 0 / K(P /DL2(M, S)) / Ψ0 C∗,+(M, S) symb / / C(S∗M, End(p /DS)) / 0, The composition of the mappings in the right most column coincides with the pull back ho- momorphism C(M ) → C(S∗M ) composed with the inclusion C(S∗M ) ⊆ C(S∗M, End(p /DS)). To describe the flow σ on C(M ) /D, we describe it on C ∞(S∗M, End(p /DS)). Surjectivity of the principal symbol mapping implies that any a ∈ C ∞(S∗M, End(p /DS)) is the symbol of an operator A ∈ P /DΨ0 cl(M, S)P /D. By Egorov's theorem [22] (see also [35, Section IV] and [21]), eis /DAe−is /D is again an element of P /DΨ0 cl(M, S)P /D and the expression σs(a) := symb(eis /DAe−is /D) gives a well defined flow on C ∞(S∗M, End(p /DS)). Again by Egorov's theorem, using that σs(a) = symb(eis /DAe−is /D) for A ∈ P /DΨ0 cl(M, S)P /D, we have that σs(a) = g∗ s (a) where gs : S∗M → S∗M is the Hamiltonian flow associated with the symbol ξ of /D. On the cosphere bundle, this Hamiltonian flow coincides with the geodesic flow. We conclude that the flow σ is induced by geodesic flow and C(M ) /D ⊆ C(S∗M ) is a closed subalgebra invariant under geodesic flow. This discussion should be compared with that in [16]. As in Subsection 2.2.1, we consider a local diffeomorphism g : M → M . Assuming that g acts conformally and lifts to S → M , it is readily verified that g is compatible with the decreasing . We use the notation /Dlog := /Dψ for this particular choice of ψ. function ψ(t) := Note that log(1+t2/n) 1 /Dlog = F /D log(1 + /D2) and e−t /Dlog = (1 + /D2)−t. By Proposition 2.19, we arrive at a spectral triple (A, L2(M, S), /Dlog) where A is the ∗-algebra generated by C ∞(M ) and an isometry Vg. Let us compute KMS-state constructed from (A, L2(M, S), /Dlog). Before diverting into this computation, we recall that the C ∗-closure of A coincides with the image of a representation of OEg . As such, we can write elements of A as linear span of elements of the form SµS∗ ν where µ = a1Vg ··· akVg and ν = b1Vg ··· blVg, where a1, . . . , ak, b1, . . . , bl ∈ C ∞(M ). Proposition 5.2. Set S := C ∞(M ) ∪ C ∞(M )Vg ⊆ A. For any β ∈ R, the set S is an analytically generating set at β for (A, L2(M, S), /Dlog). Proof. For notational convenience, write D = /Dlog. The set S generates A, so PDSPD + K generates TA. For any β ∈ R, and a ∈ S e−βDPDaPDeβD = (1 + /D2)−βP /DaP /D(1 + /D2)β. The proposition follows from that Dom((1 + /D2)β) = H 2β(M, S) as Banach spaces and any a ∈ S extends to a continuous operator on the Sobolev spaces H 2β(M, S) for all β. 44 / / / / / _ _ / / / / / / _ _ / / / / Theorem 5.3. Let M be a connected n-dimensional Riemannian manifold, /D a Dirac operator on S → M , and g : M → M a local diffeomorphism acting conformally and lifting to S. Then the spectral triple (A, L2(M, S), /Dlog) is Li1-summable, has positive essential spectrum with βD = n/2 and is n/2-analytic. Assume for all m ∈ N+, that the set of fixed points {x ∈ M : gm(x) = x}, has measure zero with respect to the volume measure. Then the state φω on A constructed from Corollary 3.21 (see page 34) is independent of ω and takes the form φω(SµS∗ ν ) = δµ,νZM− Lg(cn/2 g kLg(cn/2 b∗ g k−1Lg(··· Lg(cn/2 b∗ g b∗ 1a1)a2)··· ak−1)ak) dV, (5.1) For µ = a1Vg ··· akVg and ν = b1Vg ··· blVg. Here dV denotes the Riemannian volume form. The state φω viewed as a state on OEg via its representation on L2(M, S) is KMS for the action defined by γt(aVg) := citn/2 aVg with inverse temperature 1. g We note that the state Theorem 5.3 is a KMS-state on a Cuntz-Pimsner algebra, but not for its gauge action. If cg(x) < 1 for some x ∈ M , the generator of γ is not positive on Eg and the Laca-Neshveyev correspondence [45] does not apply in the context of OEg with the action γ. In the case k = l = 0, the formula (5.1) should be interpreted as φω(a) = ZM− a dV for a ∈ C ∞(M ). This special case follows from Connes' trace theorem. Proof. For t > n/2, standard techniques of pseudo-differential operators show that the integral kernel Kt of the trace class operator P /D(1 + /D2)−t belongs to C(M × M, END(S))∩ C ∞(M × M \ ∆M , END(S)). Here ∆M ⊆ M × M denotes the diagonal and END(S) denotes the big endomorphism bundle defined by END(S)(x,y) := Hom(Sx, Sy) for (x, y) ∈ M × M . By [19, Proposition 8.3], V ∗ g takes the form where LS,g is the L2-extension of the operator V ∗ g = N 1/2LS,gc−n/4 g , LS,g : C(M, S) → C(M, S), LS,gξ(x) := Xg(y)=x ug(y)−1ξ(y). Take a1, . . . , ak, b1, . . . , bl ∈ C ∞(M ) and write µ = a1Vg ··· akVg and ν = b1Vg ··· blVg. We introduce the notation aj := ajcn/4 . We can compute for t > n/2 that and bj = bjc−n/4 g g TrL2(M,S)(P /DSµS∗ ν e−t /Dlog) = TrL2(M,S)(a1Vg ··· akVgV ∗ g b∗ l V ∗ g b∗ g ··· V ∗ 1P /D(1 + /D 2 )−t) = N −(k−l)/2 TrL2(M,S)(a1cn/4 g ugg∗LS,gc−n/4 g b∗ l ··· c−n/4 g LS,gb∗ 1P /D(1 + /D 2 )−t) aj(xj ) k+lYj=k+1 bj−k(xj )∗ Kt(xk+1, x1)dV (x), = N −(k−l)/2ZM g ugg∗ ··· akcn/4 kYj=1 X  (x1,...,xk+l)∈M(x,k,l) 45 where M (x, k, l) ⊆ M k+l is defined as the k + l-tuples (x1, . . . , xk+l) such that for j = 1, . . . , k, xj = gj−1(x), xk = g(xk+l) and for j = k + 1, . . . k + l − 1, g(xj ) = xj+1. Note that M (x, k, l) is finite for all x, because g is a local homeomorphism, and that x1 = x for (x1, . . . , xk+l) ∈ M (x, k, l). Define M 0(x, k, l) ⊆ M (x, k, l) as the k + l-tuples (x1, . . . , xk+l) where xk+1 = x1. The set M 0(x, k, l) can be characterized as the k + l-tuples (x1, . . . , xk+l) with x = x1 = xk+1 and xk = gl(xk+l) such that for j = 1, . . . , k, xj = gj−1(x), and for j = k + 1, . . . k + l − 1, xj+1 = g(xj). In particular, if M 0(x, k, l) is non-empty then gk(x) = gl(x). In other words, (x1, . . . , xk+l) ∈ M 0(x, k, l) if and only if gk(x) = gl(x), and xj = gj−1(x) for j = 1, . . . , k and xk+j = gj−1(x) for j = 1, . . . , l. Therefore, M 0(x, k, l) contains at most one element. In particular, since M 0(x, k, l) is non-empty then gk(x) = gl(x) and the set of fixed points {x ∈ M : gm(x) = x} has measure zero for all m ∈ N+ by assumptions, then M 0(x, k, l) = ∅ if k 6= l a.e. in x. As t approaches n/2, the integral kernel Kt localizes (up to lower order term) to the diagonal and the leading order terms come from the sum over M 0(x, k, l). The Weyl law for /D2 and an explicit pseudo-differential computation of the principal symbol of P /D(1 + /D2)−t implies that there is a constant c and an ǫ > 0 only depending on /D such that TrL2(M,S)(P /DSµS∗ ν e−t /Dlog) = =cδk,l(t − n/2)−1ZM X (x1,...,xk+l)∈M 0(x,k,l)  kYj=1 aj(xj ) k+lYj=k+1 bj−k(xj)∗ dV (x) + fµ,ν(t), where fµ,ν is holomorphic on a neighborhood of the intervall [n/2 − ǫ, n/2]. Recall the notation aj := ajcn/4 we write and bj := bjc−n/4 . For k = l and (x1, . . . , xk+l) ∈ M 0(x, k, k) g g aj(gj−1(x))b∗ j (gj−1(x)) = 2kYj=k+1 aj(xj)(cid:17) (cid:16) kYj=1 kYj=1 = kYj=1 bj−k(xj)∗ = j (gj−1(x)) =(cid:16)[a1b∗ aj(gj−1(x))b∗ 1][g∗(a2b∗ 2)][(g2)∗(a2b∗ 2)]··· [(gk)∗(akb∗ k)](cid:17) (x). , we can partially integrate RM ag∗(b)dV = By the same argument that V ∗ g = N −1/2LS,gc−n/4 g RM Lg(cn/2 g a)bdV for a, b ∈ C(M ). By partially integrating k − 1 times we deduce that for some function fµ,ν holomorphic on a neighborhood of the intervall [n/2 − ǫ, n/2]. TrL2(M,S)(P /DSµS∗ ν e−t /Dlog) = =cδk,l(t − n/2)−1ZM = cδk,l(t − n/2)−1ZM [a1b∗ 1][g∗(a2b∗ 2)][(g2)∗(a2b∗ Lg(cn/2 g b∗ kLg(cn/2 g k)]dV + fµ,ν(t) = 2)]··· [(gk)∗(akb∗ b∗ 1a1)a2)··· ak−1)ak)dV + fµ,ν(t) b∗ k−1Lg(··· Lg(cn/2 g We conclude that formula (5.1) holds. It remains to show that φω defines a KMS-state on OEg for the action defined by γt(aVg) := citn/2 aVg. Let τ denote the tracial state on C(M ) defined from integrating against the volume g 46 form and L ∈ End∗ products shows that for µ = a1Vg ⊗ ··· ⊗ akVg, ν = b1Vg ⊗ ··· ⊗ bkVg ∈ E computation C(M )(Eg) the generator of γt, i.e. L = n 2 log(cg). Some yoga with inner , we have the ⊗C(M )k g φω(SµS∗ ν ) =ZM Lg(cn/2 g kLg(cn/2 b∗ g k−1Lg(··· Lg(cn/2 b∗ g b∗ 1a1)a2)··· ak−1)ak)dV = = τ(cid:0)b1Vg ⊗ ··· ⊗ bkVge−L(a1Vg) ⊗ ··· ⊗ e−L(akVg)(cid:1)E = φω(S∗ ν γi(Sµ)). ⊗C(M )k g We conclude that φω(ab) = φω(bγi(a)) for a, b ∈ A and φω is a KMS-state on OEg in the action γ. 5.2 Graph C ∗-algebras For a finite graph G we consider the spectral triple on C ∗(G) constructed in Proposition 2.20 from the choice of a point y ∈ ΩG in the infinite path space. We will assume that G is primitive, in which case C ∗(G) is simple. For an element (x, n) ∈ Vy and a finite path µ, we compute that eisDy PDy SµPDy e−isDy δ(x,n) = eisµPDy SµPDy δ(x,n). ν ) = eis(µ−ν)SµS∗ It follows that σt(SµS∗ ν and σ coincides with the gauge action on the graph C ∗-algebra C ∗(G). We can conclude that C ∗(G) = C ∗(G)Dy is closed under the flow σ. The reader can recall from Example 3.6 (see page 28) that PDy ℓ2(Vy) = ℓ2(V+ y is defined y )) is the C ∗-algebra generated by the creation in Equation (3.2). Moreover, TC∗(G) ⊆ B(ℓ2(V+ operators Teδ(x,n) :=(δ(ex,n), y ) where V+ 0, r(e) = s(x), r(e) 6= s(x). Proposition 5.4. Let β ∈ R. The set S = {Se : e ∈ E} ⊆ Cc(GG) is an analytically generating set at β for (Cc(GG), ℓ2(Vy), Dy). Proof. Since PDy SePDy = Te, it is clear that S satisfies that PDy SPDy + K(ℓ2(V+ y ) generates the Toeplitz algebra TC∗(G) as a C ∗-algebra. Moreover, eβDy PDSePDe−βDy = eβPDSePD is bounded and the proposition follows. We conclude the following theorem from Example 3.6. Theorem 5.5. Let G be a finite primitive graph with edge adjacency matrix A and y ∈ ΩG. For any extended limit ω as t → log rσ(A), the KMS-state φω on C ∗(G) associated with the spectral triple (Cc(GG), ℓ2(Vy), Dy) (see Proposition 2.20 on page 18) as in Corollary 3.21 (see page 34) is given by φω(SµS∗ ν ) = δµ,ν rσ(A)−µ, ws(µ) kwkℓ1 where ν and µ are finite paths and w ∈ CE is the ℓ2-normalized Perron-Frobenius vector. The state φω is KMS for the gauge action and its inverse temperature is log(rσ(A)). The KMS-state φω on C ∗(G) in Theorem 5.5 is the unique KMS-state by [23]. Numerous authors present constructions of this state and for more general graphs, eg [1, 2, 13, 42]. 47 5.3 Group C ∗-algebras For the reduced group C ∗-algebra of a countable discrete group we considered two types of (semifinite) spectral triples in Subsection 2.2.4 (see page 24). We now compute the associated KMS-states. We fix a length function ℓ on the discrete countable group Γ. For technical simplicity, we assume that Γ is an exact group ensuring that Γ acts amenably on its Stone-Cech boundary ∂SCΓ (see [53]). We assume that (Γ, ℓ) is of at most exponential growth and that ℓ is critical (see Definition 3.10 on page 30). Let (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) denote the associated Li1-summable spectral triple as in Proposition 2.31. Since Dℓ ≥ 0, we have that Tcb(Γ)⋊rΓ = cb(Γ) ⋊r Γ + K(ℓ2(Γ)) = cb(Γ) ⋊r Γ. The last equality follows from that c0(Γ) ⋊r Γ = K(ℓ2(Γ)). We conclude that we have a short exact sequence 0 → K(ℓ2(Γ)) → Tcb(Γ)⋊r Γ → C(∂SCΓ) ⋊r Γ → 0. The flow σ+ on Tcb(Γ)⋊rΓ = cb(Γ) ⋊r Γ is given on an element aλg ∈ cb(Γ) ⋊ Γ by s (aλg) = eis(ℓ(·)−ℓ(g−1·)aλg, σ+ (5.2) and we conclude that Tcb(Γ)⋊rΓ is invariant under σ+. Therefore Tcb(Γ)⋊rΓ = Tcb(Γ)⋊rΓ,Dℓ. Proposition 5.6. Let β ∈ R. The ∗-algebra cb(Γ) ⋊alg Γ is an analytically generating set at β for (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ). Proof. Note that PDℓ = 1 because Dℓ is positive. For aλg ∈ cb(Γ) ⋊alg Γ, we compute that eβDℓ aλge−βDℓ = e−β(ℓ(·)−ℓ(g−1·))aλg. Since ke−β(ℓ(·)−ℓ(g−1·)akcb(Γ) ≤ eβℓ(g−1)kakcb(Γ), it holds that eβDℓ aλge−βDℓ ∈ cb(Γ) ⋊alg Γ and the proposition follows. Proposition 5.7. Let β ∈ R. The ∗-algebra cb(Γ) ⋊alg Γ is an analytically generating set at β for (cb(Γ) ⋊alg Γ, ℓ2(Γ, SH), Dc, N, Trτ ). Proof. For aλg ∈ cb(Γ) ⋊alg Γ, we compute for f ∈ ℓ2(Γ, SH) that eβDc πS(aλg)e−βDcf (γ) = e−β(ℓ(γ)−ℓ(g−1γ))a(γ)[π(g)f ](γ). The estimate ke−β(ℓ(·)−ℓ(g−1·)akcb(Γ) ≤ eβℓ(g−1)kakcb(Γ) shows that eβDc πS(aλg)e−βDc ∈ π(cid:0)cb(Γ) ⋊alg Γ(cid:1). Therefore, eβDc PDc(cb(Γ) ⋊alg Γ)PDc e−βDc = PDc eβDc(cb(Γ) ⋊alg Γ)e−βDcPDc ⊆ PDc(cb(Γ) ⋊alg Γ)PDc ⊆ N+, and the proposition follows. 48 Definition 5.8. If ω is an extended limit as t → β(Γ, ℓ), and ℓ is critical, we define the Patterson-Sullivan measure µω on the Stone- Cech boundary ∂SCΓ as Z∂SC Γ a dµω := lim t→ωPγ∈Γ a(γ)e−tℓ(γ) Pγ∈Γ e−tℓ(γ) , for a function a ∈ C(∂SCΓ) and where a ∈ cb(Γ) is any function with a = a mod c0(Γ). Remark 5.9. It is possible to define the Patterson-Sullivan measure µω as an extended weak- limit of the family of probability measures on Γ µt = Pγ∈Γ δγe−tℓ(γ) Pγ∈Γ e−tℓ(γ) . In the literature, Patterson-Sullivan measures are usually defined as weak accumulation points of (µt)t>β(Γ,ℓ) but we allow for a slightly more general construction with extended limits. A priori, µω is a probability measure on the Stone- Cech compactification of Γ. Since the support of µω is contained in the closed subspace ∂SCΓ we consider µω as a measure on ∂SCΓ. Theorem 5.10. Let Γ be a discrete group and φω the KMS-state on C(∂SCΓ) ⋊r Γ constructed as in Corollary 3.21 (see page 34) using an extended limit ω as t → β(Γ, ℓ) and either of the following two semifinite spectral triples: • The spectral triple (cb(Γ) ⋊alg Γ, ℓ2(Γ), Dℓ) associated with a critical length function of at most exponential growth as in Proposition 2.31 (see page 24). • The semifinite spectral triple (cb(Γ) ⋊alg Γ, ℓ2(Γ, SH), Dc, N, Trτ ) associated with a critical proper Hilbert space valued cocycle of at most exponential growth as in Proposition 2.34 (see page 26). Then φω is given in terms of the Patterson-Sullivan measure µω by φω(aλg) = δe,gZ∂SC Γ a dµω. The state φω is KMS at inverse temperature β(Γ, ℓ) for the flow on C(∂SCΓ) ⋊r Γ induced by the action σ+ on cb(Γ) ⋊r Γ given in Equation (5.2). Moreover, φω extends to a KMS-state at inverse temperature 1 on the von Neumann algebra L∞(∂SCΓ, µω)⋊Γ with its Radon-Nikodym flow σRN s (aλg) =(cid:18) dg∗µω dµω (cid:19)is aλg. Proof. By the computations of Example 3.9 (see page 30), the spectral triple associated with a length function as in Proposition 2.31 have the same heat traces as the semifinite spectral triple associated with a proper Hilbert space valued cocycle as in Proposition 2.34. In both cases, Example 3.9 shows that for aλg ∈ cb(Γ) ⋊alg Γ we have φt,0(aλg) = δe,gPγ∈Γ a(γ)e−tℓ(γ) Pγ∈Γ e−tℓ(γ) . 49 It follows that φω(aλg) = δe,gR∂SC Γ a dµω in both cases. To relate φω to the Radon-Nikodym flow, we first show that µω is strictly positive, i.e. that µω(U ) > 0 for any open set U ⊂ ∂SCΓ. For any open set U ⊂ ∂SCΓ, the translates (γU )γ∈Γ cover ∂SC Γ. If µω(U ) = 0, then by quasi-invariance µω(γU ) = 0 which contradicts µω being a probability measure. The fact that µω is strictly positive ensures that the Radon-Nikodym derivatives dg∗µω dµω defined and strictly positive. The mapping g 7→ dg∗µω is a cocycle, i.e. for h, g ∈ Γ, are well dµω d(gh)∗µω dµω = dg∗µω dµω dµω (cid:21) . (g−1)∗(cid:20) dh∗µω The proof is completed by computing that for aλg, bλh ∈ C(∂SCΓ) ⋊alg Γ, we have the identity φω(aλgbλh) = δh,g−1φω(ah∗(b)) = δh,g−1Z∂SC Γ ah∗(b)dµω = g∗(a)bd(h∗µω) = dµω = bg∗(a) dh∗µω dµω h∗ (g∗(a)b) dµω = δh,g−1Z∂SC Γ = δh,g−1Z∂SC Γ = δh,g−1Z∂SC Γ b(h−1)∗ a(cid:18) dg∗µω dµω (cid:19)−1! dµω = = δh,g−1Z∂SC Γ aλg! = φω(cid:0)bλhσRN = φω bλh(cid:18) dg∗µω dµω (cid:19)−1 s=i (aλg)(cid:1) (h−1)∗(cid:18) dg∗µω dµω (cid:19) dh∗µω = 1. dµω In the third last identity we used the cocycle identity implying that if hg = e, then Remark 5.11. The reader should note that φωC∗ r (Γ) coincides with the ℓ2-trace. 6 KMS-states on Cuntz-Pimsner algebras with their gauge action In this section we consider the constructions of Corollary 3.21 in a broad class of examples which include both Cuntz-Krieger algebras and crossed products by Z. Here we use the techniques from Section 3 in conjunction with those from Subsection 2.2.3 to analyse the KMS states on Cuntz-Pimsner algebras, and compare them to the Laca-Neshveyev correspondence establishing a bijection between KMS-states on Cuntz-Pimsner algebras and tracial states on its coefficient algebra. 6.1 KMS-states on Cuntz-Pimsner algebras from traces on the coefficient algebra Firstly, we shall show that a critical positive trace on A (see Definition 6.2 below) gives rise to a KMS-state on the Cuntz-Pimsner algebra OE assuming that E is strictly W-regular. Recall Lemma 2.28 (see page 22) giving the construction of the Li1-summable semifinite spectral triple (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ), where Nτ (ΞA) := (End∗ A(ΞA) ⊗A 1)′′. 50 Lemma 6.1. Let E be a strictly W-regular fgp bi-Hilbertian bimodule over A, β ∈ R and τ a positive trace on A. The set S = {Se : e ∈ E} ⊂ OE is an analytically generating set at β for (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ). Proof. Since TOE,D is precisely the Toeplitz algebra TE, it is immediate that the set of operators {P SeP : e ∈ E} generates TOE ,D. The analyticity condition on the Fock space is likewise obvious from the computation eβDψ PDψ SePDψ e−βDψ = eβPDψ SePDψ . Definition 6.2. Let E be an A-bimodule which is fgp from the right and τ a positive trace on A. We define the critical value of (E, τ ) as β(E, τ ) := inf{t ≥ 0 : ∞Xn=0 τ∗(E⊗An)e−tn < ∞}. We say that τ is critical for E if lim tցβ(E,τ ) ∞Xn=0 τ∗(E⊗An)e−tn = ∞. Remark 6.3. Note that the critical value of a trace and the notion of it being critical only depends on [E] ∈ KK0(A, A), in fact only on the sequence (ch0(E⊗An))n∈N ⊆ HC0(A) in cyclic homology. Just as in the proof of Lemma 2.28 we obtain the estimate 0 ≤ β(E, τ ) ≤ log(N ), where N is the number of elements in the left frame and the right frame. It follows from Definition 3.1 and Proposition 3.8 that the Li1-summable semifinite spectral triple (OE , L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) of Lemma 2.28 has positive Trτ -essential spectrum if and only if τ is critical. By construction, the projection PDψ is the projection onto the Fock module FE and therefore Nτ (ΞA)+ = (End∗ A(FE)⊗A1)′′ is a subalgebra of B(L2(FE, τ )). We let N denote the number op- erator on L2(FE, τ ) -- the self-adjoint operator defined from NE⊗n⊗AL2(A,τ ) = nIdE⊗n⊗AL2(A,τ ). The next proposition follows from the definition of the Toeplitz algebra of a semifinite spectral triple. Proposition 6.4. The Toeplitz algebra TOE of the semifinite spectral triple (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) is given by TOE = TE ⊗A 1A + K(End∗ A(FE )⊗A1)′′ , where TE ⊆ End∗ A(FE) is the Cuntz-Toeplitz algebra of E. Moreover, the action σ+ preserves TOE and is generated by the number operator in the sense that for µ ∈ Ek, ν ∈ E⊗l and K ∈ K A(FE )⊗A1)′′, we have (End∗ σ+ s (TµT ∗ ν + K) = eis(µ−ν)TµT ∗ ν + eisN Ke−isN . An immediate consequence of Proposition 6.4 is that OE = TOE /K action σ on OE coincides with the gauge action σs(SµS∗ theorem is readily deduced from the computations of Example 3.7 (see page 29). ν ) = eis(µ−ν)SµS∗ A(FE ))′′ and that the ν . The following (End∗ 51 Theorem 6.5. Let E be a strictly W-regular fgp bi-Hilbertian bimodule over A. For any positive trace τ on A which is critical for E and any extended limit ω as t → β(E, τ ), the KMS- state φω on OE associated with the semifinite spectral triple (OE , L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) as in Theorem 3.21 (see page 34) is given by φτ,ω(SµS∗ ν ) = δµ,νe−β(E,τ )µ lim n=0 TrE⊗n τ t→ωP∞ P∞ ((νµ)Eµ) e−tn , n=0 TrE⊗n τ (1)e−tn where ν ∈ E⊗k and µ ∈ E⊗l. The state φτ,ω is KMS for the gauge action on OE and its inverse temperature is β(E, τ ). Proof. By Lemma 6.1 the semifinite spectral triple (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) is β- analytic for any β ∈ R. By Definition 3.1 and Proposition 3.8 it has positive Trτ -essential spectrum. Thus the state φτ,ω as in Theorem 3.21 is KMS for the gauge action on OE with inverse temperature β(E, τ ). Using the computation in Equation (3.3), we see that φτ,ω(SµS∗ ν ) = δµ,ν lim t→ωP∞ n=µPσ=n e−tnτ(cid:0)( (µeσ)Eµ eσ (νeσ)Eµ eσ)E⊗(n−µ)(cid:1) P∞ n=0Pρ=n τ ((eρeρ)A)e−tn , for ν ∈ E⊗k and µ ∈ E⊗l. However, this expression can be vastly simplified using that φτ,ω is KMS. Using φτ,ω(SµS∗ ν ) = δµ,νe−β(E,τ )µφτ,ω(S∗ ν Sµ) = δµ,νe−β(E,τ )µφτ,ω((νµ)E⊗µ ), we obtain φτ,ω(SµS∗ = δµ,νe−β(E,τ )µ lim ν ) = δµ,νe−β(E,τ )µφω((νµ)E⊗µ) n=0Pσ=n e−tnτ(cid:0)( eσ (νµ)Eµ eσ)E⊗n(cid:1) t→ωP∞ P∞ n=0Pρ=n τ ((eρeρ)A)e−tn t→ωP∞ ((νµ)Eµ) e−tn P∞ = δµ,νe−β(E,τ )µ lim n=0 TrE⊗n n=0 TrE⊗n (1)e−tn . τ τ The reader should note that in the formula computing φτ,ω, it is only the right inner product on E that appears. Example 6.6. Let us consider the construction from Theorem 6.5 in a specific example. As in Example 2.30 (see page 23), we consider a compact Hausdorff space Y , a surjective local homeomorphism g : Y → Y and the associated bimodule Eg. Let us compute φτ,ω starting from a positive trace τ on C(Y ), i.e. a positive measure λ on Y . By the argument of Theorem 6.5, the KMS-condition on φτ,ω guarantees that it suffices to describe φτ,ω(a) for a ∈ C(Y ). By the Riesz representation theorem, for a probability measure λω. We compute that for a ∈ C(Y ), TrE⊗n τ (a) = Xσ=n Ln g (a) dλ =ZY a d[(Ln g )∗λ]. a dλω, φτ,ω(a) =ZY τ ((eσaeσ)C(Y )) =ZY 52 λω = lim t→ω P∞ P∞ n=0 e−tn(Ln g )∗λ . n=0 e−tn[(Ln g )∗λ](Y ) From Theorem 6.5, we conclude that λω is given by an extended weak* limit of measures The KMS-condition on φτ,ω translates into (Lg)∗λω = eβ(E,τ )λω which is readily verified for the measure λω. We remark that the construction above is reminiscent of the method in [64] to construct equilibrium measures. In the case that g is mixing, i.e. for all open subsets U, V ⊆ Y there is an N ≥ 0 such that gn(U ) ∩ V 6= ∅ for all n ≥ N , then there exists a unique KMS-state on OEg (see [19, Theorem 6.1]). In particular, for mixing g, the KMS-state φτ,ω on OEg does not depend on the choice of trace τ . 6.2 The Toeplitz construction vs the Laca-Neshveyev correspondence In the previous subsection we saw that there is a mapping from the set of positive critical traces on a C ∗-algebra A to the set of KMS-states on the Cuntz-Pimsner algebra OE when E is strictly W-regular. As Example 6.6 shows, this mapping is not injective in general, but it is surjective in some cases (e.g. when g is mixing). We now compare our construction to the bijective correspondence between a certain set of tracial states on A with KMS-states on the Cuntz-Pimsner algebra OE first discovered by Laca-Neshveyev [45]. Definition 6.7. The positive trace τ : A → C satisfies the Laca-Neshveyev condition for α ≥ 0 if where La denotes the left action of a on E. TrE τ (La) = eατ (a), For notational simplifity we often write TrE τ (La). Given a positive trace τ : A → C satisfying the Laca-Neshveyev condition, it was proven by Laca-Neshveyev [45, Theorem 2.1 and 2.5] that the expression τ (a) instead of TrE φLN,τ (SµS∗ ν ) := δµ,νe−αµτ ((νµ)A), defines an α-KMS state on OE. Moreover, Laca-Neshveyev proved that the construction τ 7→ φLN,τ is a bijection between tracial states on A satisfying the Laca-Neshveyev condition for α ≥ 0 and α-KMS states on OE. The work of Laca-Neshveyev [45] gives more context to the construction in Theorem 6.5 (see page 52). For a unital C ∗-algebra A, we let T(A) denote the set of positive traces on A. If E is an A − A-correspondence which is finitely generated and projective as a right module, we can also define CTE,α(A) as the set of positive critical traces τ with β(E, τ ) = α. We also define LNE,α(A) as the set of positive traces satisfying the Laca-Neshveyev condition. Following [45, Discussion proceeding Definition 2.3], we define FE,α : T(A) → T(A), FE,ατ (a) = TrE τ (a)e−α. (6.1) Proposition 6.8. For any positive trace τ on a unital C ∗-algebra and an A-A-correspondence E which is fgp from the right, it holds that E,ατ (a) = e−αn TrE⊗n F n τ (a), n ∈ N+. 53 computation shows that Proof. By definition, FE,ατ (a) = e−αPN E,ατ (a) = e−αn Xσ=n F n τ (eσaeσ)A = e−αn TrE⊗n τ (a). j=1 τ (ejaej )A for a right frame (ej)N j=1 of E. A direct Proposition 6.9. Let E be an A-A-correspondence which is fgp from the right. Then τ ∈ T(A) is a fixed point of FE,α if and only if τ satisfies the Laca-Neshveyev condition from Definition 6.7 (see page 53). Proposition 6.9 is a direct consequence of Definition 6.7 and the formula (6.1). Proposition 6.10. Let E be an A-A-correspondence which is fgp from the right. If τ ∈ T(A) is a tracial state satisfying the Laca-Neshveyev condition for α ≥ 0, then τ∗(E⊗An) = eαnF n E,ατ (1) = eαn. Proof. The proposition follows from the computation τ∗(E⊗An) = TrE⊗n τ (1) = eαnF n E,α(τ )(1), and Proposition 6.9. Proposition 6.11. Let E be an A-A-correspondence which is fgp from the right and α ≥ 0. Then the following holds: 1. A positive trace τ ∈ T(A) is critical for α if and only if the positive trace St E,ατ := ∞Xn=0 e−tnF n E,ατ, is finite for t > 0 and satisfies St inclusion of sets E,ατ (1) → ∞ as t → 0. In particular, we have an 2. For any extended limit ω as t → 0, the mapping LNE,α(A) ⊆ CTE,α(A). Sω E,α : CTE,α(A) → LNE,α(A), Sω E,ατ := lim t→ω E,ατ St St E,ατ (1) , is well defined. Moreover, (Sω states in LNE,α(A). E,α)2τ = Sω E,ατ E,ατ (1) and Sω Sω E,α surjects onto the set of tracial Proof. Statement 1 is an immediate consequence of Proposition 6.10. The first part of state- ment 2 follows from the computation St E,ατ (1) − FE,αSt E,ατ (1) = τ (1) = o(St E,ατ (1)), as t → 0 for τ ∈ CTE,α(A). The second part of statement 2 follows from the fact that Proposition 6.9 implies that for τ ∈ LNE,α(A), St E,ατ := (1 − e−t)−1τ. 54 Remark 6.12. For Example 6.6, the mapping FE,α takes the form FE,ατ = e−αL∗ gτ . In par- ticular, the computations of Example 6.6 is a special case of the constructions in Proposition 6.11. We shall see that this holds in general below in Proposition 6.14. Using our previous results, Proposition 3.8 (see page 29) and Proposition 6.10 (see page 54), we can deduce a computation of heat traces. Proposition 6.13. Let E be a strictly W-regular fpg bi-Hilbertian bimodule over a unital C ∗- algebra A. If τ is a tracial state on A satisfying the Laca-Neshveyev condition for α ≥ 0, then Trτ (PDe−tDψ) = 1 1 − eα−t . In particular, for any tracial state τ satisfying the Laca-Neshveyev condition for α ≥ 0 the semifinite spectral triple (OE, L2(ΞA, τ ), Dψ, Nτ (ΞA), Trτ ) has positive Trτ -essential spectrum with βD = β(E, τ ) = α. Proof. We compute that Trτ (PDe−tDψ ) = τ∗(E⊗An)e−tn = ∞Xn=0 ∞Xn=0 e−(t−α)n = 1 1 − eα−t . In the first step we used Proposition 3.8 and in the second step we used Proposition 6.10. We can now reformulate Theorem 6.5 in terms of the map FE,α and the constructions of Proposition 6.11. Proposition 6.14. Assume that E is a strictly W-regular fgp bi-Hilbertian bimodule over A. Let α ≥ 0, ω be an extended limit as t → α and τ ∈ CTE,α(A) a critical trace. The KMS-state φτ,ω defined from Theorem 6.5 takes the following form: φτ,ω(SµS∗ ν ) = δµ,νe−αµSωα E,ατ ((νµ)A), µ, ν ∈ Falg E , where ωα is the extended limit at 0 obtained from translating ω by α. Proof. The KMS-condition on OE reduces the proof to showing that φτ,ω(a) = Sωα E,ατ (a) for a ∈ A, just as in the proof of Theorem 6.5. Using Proposition 6.8 we can compute for a ∈ A that φτ,ω(a) = lim n=0 TrE⊗n n=0 TrE⊗n τ τ (a)e−tn (1)e−tn t→ω0P∞ P∞ = lim t→ωP∞ P∞ n=0 F n n=0 F n E,ατ (a)e−(t−α)n E,ατ (1)e−(t−α)n = Sωα E,ατ (a). Let KM Sα(OE ) denote the set of α-KMS states on OE for the gauge action and LNSE,α(A) for the set of tracial states satisfying the Laca-Neshveyev condition. Theorem 6.15. Let α ≥ 0 and let ω be an extended limit as t → α. Assume that E is a strictly W-regular fgp bi-Hilbertian bimodule over A. The mapping LNSE,α(A) → KM Sα(OE), τ 7→ φτ,ω, defined from Theorem 6.5, and revisited in Proposition 6.14, is a well defined bijection of sets. More precisely, φτ,ω : OE → C is a KM Sα state for the gauge action, which is independent of ω and coincides with φLN,τ . 55 Proof. By Theorem 6.5 and Proposition 6.13, the mapping τ 7→ φτ,ω is a well defined mapping from the set of positive traces on A satisfying the Laca-Neshveyev condition for α and α-KMS- states on OE for the gauge action. By the Laca-Neshveyev correspondence, the KMS-state φτ,ω is uniquely determined by the trace φτ,ωA. We can therefore deduce that φτ,ω = φLN,τ and the Theorem upon proving the identity φτ,ωA = τ . This statement follows immediately from Proposition 6.14 and the second part of Proposition 6.11. There are some quasi-invariance assumptions on traces that allows us to compare the KMS- states constructed in Theorem 6.5 and the Laca-Neshveyev correspondence to a simpler con- struction involving Φ∞. While the construction of Φ∞ depends on the left inner product on E, the quasi-invariance condition we impose also depends on the left inner product. The following quasi-invariance condition is a refinement of the notion of E-invariant functionals from [56]. Definition 6.16. Let α ≥ 0 and E a finitely generated projective bi-Hilbertian bimodule. We say that a positive trace τ on A is α-quasi-invariant with respect to E and the extended limit ω0 ∈ ℓ∞(N)∗ if for all n ∈ N and µ, ν ∈ E⊗n we have e−αµτ ((νµ)A) = lim k→ω0 τ (Φk(TµT ∗ ν )e−βk ) = lim k→ω0 τ (A(µνeβk−ν)e−βk ). (6.2) If τ is α-quasi-invariant with respect to E and some extended limit, we simply say that τ is α-quasi-invariant with respect to E. Note, that Φk are defined on page 20 (formula (2.7) and the paragraph after). Observe that if E is W-regular then the limit in the definition of quasi-invariance exists, and so is independent of the extended limit ω0. Remark 6.17. Just like in [56, Lemma 4.2], if E is full as a right module, then any positive functional τ : A → C which is quasi-invariant in the sense of Definition 6.16 is a positive trace. To see this, observe that for all µ, ν ∈ Falg E and a ∈ A, the centrality of the Watatani indices eβk ∈ A (see formula (2.6) for the definition) gives e−αµτ ((νµ)Aa) = e−αµτ ((νµa)A) = lim τ (A(µaνeβk−ν )e−βk ) τ (A(µνa∗eβk−ν)e−βk ) = e−αµτ ((νa∗µ)A) k→ω0 = lim k→ω0 = e−αµτ (a(νµ)A). Example 6.18. We consider the module Eg over C(Y ) defined from a surjective local home- omorphism g : Y → Y as in Example 6.6 (see page 52). In this case, βn = 0 for all n and quasi-invariance of a positive trace τ given by a positive measure λ on Y is equivalent to the condition (Lg)∗λ = eαλ. Another computation shows that this condition is equivalent to the Laca-Neshveyev condition. In particular, for the module Eg, quasi-invariance is equivalent to satisfying the Laca-Neshveyev condition. Example 6.19. Let us consider the Cuntz algebra ON defined as the Cuntz-Pimsner algebra of the C-bimodule CN . In this case, (νµ)A = A(µν) for all µ, ν ∈ (CN )⊗n and βn = n log(N ). Therefore, quasi-invariance of a trace τ on C is equivalent to eατ = N τ. 56 That is, any non-zero trace on C is log(N )-quasi invariant. Our immediate aim is to connect the quasi-invariance of Definition 6.16 with the condition imposed by Laca-Neshveyev, [45]. Lemma 6.20. Let E be an fgp bi-Hilbertian bimodule. Suppose that τ satisfies the α-quasi- invariance condition of Definition 6.16 with respect to E. Then τ satisfies the Laca-Neshveyev condition for α. Proof. This is a computation using Definition 6.16 of quasi-invariance and a frame (ej) for EA. So τ ((ejaej )A) = eα lim k→ω0 Taej T ∗ ej )e−βk ) TrE τ (La) = Trτ(cid:18)Xj aΘej,ej(cid:19) =Xj τ (Xj = eα lim k→ω0 τ (Φk(Xj aA(ejej eβk−1)e−βk ) = eα lim k→ω0 τ (aeβk e−βk ) = eατ (a). If τ : A → C satisfies the α-quasi-invariance condition of Definition 6.16, then we can rewrite the Laca-Neshveyev KMS state φLN,τ : OE → C as φLN,τ (TµT ∗ ν ) = δµ,νe−αµτ ((µν)A) = lim k→ω0 τ (Φk(TµT ∗ ν )e−βk ). This computation proves the following. Note that the next result does not require any W- regularity from the module E. Proposition 6.21. Let E be an fgp bi-Hilbertian bimodule and consider an α-quasi-invariant positive trace τ : A → C with respect to E. The state on OE defined by SµS∗ ν 7→ lim k→ω0 τ (Φk(TµT ∗ ν )e−βk ), (6.3) is α-KMS for the gauge action on OE and coincides with φLN,τ . ν ) := limk→∞ Φk(TµT ∗ If E is W-regular, the definition Φ∞(SµS∗ Equation (6.3) coincides with τ ◦ Φ∞. We conclude the following. Corollary 6.22. Let τ : A → C be an α-quasi-invariant positive trace with respect to an fgp bi-Hilbertian bimodule E. Assume that E is a W-regular. Then, the state τ ◦ Φ∞ on OE is α-KMS for the gauge action on OE and coincides with φLN,τ . ν )e−βk shows that the state in By Theorem 6.15 and Corollary 6.22, we have that φω,τ = τ ◦ Φ∞ for any α-quasi-invariant positive trace τ and extended limit ω at α assuming that E is strictly regular. 6.3 Obstructions to bi-Hilbertian bimodule structures In the two previous subsections, we assumed our A-bimodule E to be an fgp bi-Hilbertian bimodule, and imposed the additional assumption of strict W-regularity (see Definitions 2.22 and 2.25). The assumptions allowed us to construct a semifinite spectral triple from a Kasparov module relying on both the left and the right inner product on E. Instead, we can just use [45] 57 to proceed from a KMS-state directly to a semifinite spectral triple whose associated KMS- state as in Corollary 3.21 coincides with the original KMS-state. In order to compare the indirect approach for strictly W-regular modules to the direct approach from the KMS-state we will need to extend our module to von Neumann algebra coefficients, and along the way we derive obstructions to having the structure of a strictly W-regular bi-Hilbertian bimodule structure on an A − A-correspondence. We suppose that we have a finitely generated projective right A-module EA, with A unital, and carrying a unital left action of A. Let τ : A → C be a faithful positive trace satisfying the Laca-Neshveyev condition for α ≥ 0 (see Definition 6.7 on page 53), and define the associated KMS-state on OE by φLN,τ (SµS∗ ν ) := δµ,ντ ((νµ)A)e−αµ, µ, ν ∈ Falg E . By construction, φLN,τA = τ . The Cuntz-Pimsner algebra OE acts on the GNS-space L2(OE, φLN,τ ) by left multiplication and, since φLN,τ restricts to a trace on A, A acts by both left and right multiplication on L2(OE, φLN,τ ). For notational simplicity, we identify τ with its normal extension to A′′. Note that A′′ is independent of whether we take the bicommutant in L2(OE, φLN,τ ) or L2(A, τ ) and by faithfulness of τ we can identify A with its image under the GNS-representation and obtain an inclusion A ⊆ A′′. Proposition 6.23. Let E be an fgp right A-Hilbert C ∗-module with a left unital action of A, τ a faithful positive trace on A satisfying the Laca-Neshveyev condition and let P0 : L2(OE, φLN,τ ) → L2(A, τ ) ⊂ L2(OE , φLN,τ ) denote the orthogonal projection. It then holds that A′′ = P0O′′ EP0. Proof. It is clear that A′′ ⊆ P0O′′ write T = P0T0P0 where T0 is the WOT-limit of a net Tλ =Pj Sµλ,j S∗ EP0. To prove the converse inclusion, take T ∈ P0O′′ P0TλP0 = Xj:µλ,j=νλ,j=0 EP0 and νλ,j ∈ OE. We have that Sµλ,j S∗ νλ,j , so P0TλP0 ∈ A and T ∈ A′′. We can define a conditional expectation Φ∞ : O′′ E → A′′, Φ∞(S) := P0SP0, which is well defined by Proposition 6.23. Using the expectation Φ∞ we can define a right module ΞA′′ by completing OE in the norm defined by the inner product (S1S2)A′′ := Φ∞(S∗ 1 S2), S1, S2 ∈ OE. It is clear that L2(OE, φLN,τ ) = L2(ΞA′′, τ ). The construction of ΞA′′ does not require EA to be biHilbertian, just an A-A-correspondence. The following result follows from the relations defining the Cuntz-Pimnser algebra and the fact that Φ∞ is a conditional expectation. 58 Proposition 6.24. For µ ∈ E⊗k and ν ∈ E⊗l, Φ∞(S∗ µSν) = δµ,ν(µν)A. In particular, the map µ 7→ Sµ extends to an A′′-linear isometric embedding FE ⊗A A′′ → ΞA′′ of the Fock module. We now turn to describing the WOT-closure of E inside O′′ A-sub-bimodule of OE via µ 7→ Sµ. Lemma 6.25. Let A be a unital C ∗-algebra. Let EA be a finitely generated projective right A-module with a unital left action, and suppose that τ : A → C satisfies the Laca-Neshveyev condition on E for α ≥ 0. Then E. We will identify E with an WOT is an A′′-bimodule. Moreover, the following hold: E′′ := E ⊂ O′′ E 1. As right A′′-modules, E′′ ∼= E ⊗A A′′ and the isomorphism is an isomorphism of A′′- Hilbert C ∗-modules when equipping E′′ with the right inner product (µν)A′′ := Φ∞(S∗ µSν), µ, ν ∈ E′′. 2. The right A′′-Hilbert C ∗-module E′′ is finitely generated and projective. 3. If E is finitely generated and projective as a left A-module, then E′′ is finitely generated and projective as a left A′′-module. 4. If the implication holds, the expression P0ee∗P0 = 0 ⇒ e = 0 ∀e ∈ E′′, (6.4) gives a left inner product on E′′ making it into a bi-Hilbertian bimodule. The right Watatani index of E′′ is 1 and qk = Id(E′′)⊗k for all k (and so is invertible). A(ef ) := P0SeS∗ f P0 5. If E′′ is a finitely generated projective module from the left and the implication (6.4) holds, then E′′ is a strictly W -regular fgp bi-Hilbertian bimodule over A′′. Proof. The initial statement of the lemma is clear since E is an A-sub-bimodule of OE, so its WOT closure is a bimodule over A′′. To prove statement 1. we note that E⊗AA′′ ∼= EA′′ ⊆ O′′ E. Moreover, using that E is finitely generated and projective, it follows that E′′ = EA′′ and therefore E′′ ∼= E ⊗A A′′ follows. Statement 1. now follows from Proposition 6.24. Statement 2. follows from statement 1. because E is finitely generated and projective over A. Statement 3. is proven in a similar way as Statement 1., indeed if E is finitely generated and projective as a left A-module then E′′ ∼= A′′ ⊗A E as a left A-module. Statement 4. is less trivial. Assuming that the implication (6.4) holds, it is straight-forward to verify that the left and right actions are compatible, i.e. that the A-action from the left/right is adjointable for the right/left inner product. For E′′ to be a bi-Hilbertian bimodule it remains 59 to show that the norm arising from the left inner product is equivalent to the norm arising from the right inner product. For any e ∈ E, we compute that kA(ee)kA = kP0SeS∗ e P0kL2(A,τ ) = kS∗ e P0SekL2(E∗,φτ ) = kS∗ e SekL2(E∗,φτ ) = (6.5) are equivalent. where the norm of S∗ e Se is attained by S∗ = k(ee)AkL2(E∗,φτ ) = k(ee)AkA e P0Se on L2(E∗, φτ ) by the assumption that A(··) is positive definite. Equation (6.5) shows that the two normspk(··)AkA andpkA(··)kA on E To finalize the proof of statement 4., we compute the right Watatani index of E, which exists be- cause E′′ is finitely generated projective by statement 2. We compute on L2(A, τ ) ⊂ L2(OE, φτ ) that ha,Xj P0π(1)(Θej,ej )P0ai = φτ (a∗P0π(1)(IdE)P0a) = φτ (a∗1Aa) = τ (a∗a). A(ejej)ai = ha,Xj By the faithfulness of τ and Cuntz-Pimsner covariance we can now deduce that the right Watatani index is equal to 1A. This immediately implies that qk = Id(E′′)⊗k for all k. Finally, statement 5. follows from that under the stated assumptions, E′′ is an fgp bi-Hilbertian bimodule (using statements 1., 2. and 4.) and by statement 4. qk satisfies the condition in Definition 2.25, so E′′ is strictly W-regular. Remark 6.26. There are examples of A-A-correspondences E that are fgp from the right but not fgp from the left such that E′′ is fgp from the left. These examples come from (certain) self-similar dynamical systems, see [43]. Here is a simple example. Let A = C([0, 1]), γ1 : [0, 1] → [0, 1] γ1(x) = x/2, γ2 : [0, 1] → [0, 1] γ2(x) = 1/2 + x/2, and E = C({(γ1(x), x) : x ∈ [0, 1]} ∪ {(γ2(x), x) : x ∈ [0, 1]}). The correspondence structure is defined for a, b ∈ A and e ∈ E by (a · e · b)(γj(x), x) = a(γj(x))e(γj (x), x)b(x) e1(γj(x), x)e(γj (x), x). and (e1e2)A(x) = Xj=1,2 The graphs of γ1 and γ2 in [0, 1]× [0, 1] are disjoint and their respective characteristic function χ1 and χ2 are elements of E. One checks directly that {χ1, χ2} is a right frame for EA, and since (χ1χ2) = 0, there is an isomorphism of right Hilbert C ∗-modules EA ∼= C[0, 1] ⊕ C[0, 1]. We conclude that EA is fgp from the right. Using the frame {χ1, χ2}, we can identify AE ∼= C[0, 1/2]⊕C[1/2, 1] as a left C[0, 1]-module. As a left module, AE is therefore finitely generated but clearly not projective as the rank of Ex := E/C0([0, 1]\{x})E is discontinuous at x = 1/2. We shall now see that it is even impossible for a left inner product compatible with the right inner product to exist. Let 0 ≤ φ ∈ A be 1 on [0, 1/2]. Then if we have a compatible left A-valued inner product A(χ1χ2)(x) = A(φ · χ1χ2)(x) = φ(x)A(χ1χ2)(x) = A(χ1χ2)(x)φ(x) = A(χ1φ · χ2)(x). 60 Taking the infimum over such φ we see that the support of A(χ1χ2) is contained in {1/2}. Then for arbitrary a, b ∈ A A(aχ1 + bχ2aχ1 + bχ2) = aA(χ1χ1)a∗ + bA(χ2χ2)b∗ + a(1/2)b∗(1/2)A(χ1χ2) + b(1/2)a∗(1/2)A(χ2χ1). From here one can show that any inner product taking values in the continuous functions takes values in the functions vanishing at 1/2. Then one shows that the associated norm can not be equivalent to the right inner product. The situation is better for E′′. We consider the trace τ (a) :=R 1 0 a(x)dx on C[0, 1]. A short computation shows that TrE τ = 2τ so τ satisfies the Laca-Neshveyev condition for α = log(2) and extends to a KMS-state on OE at inverse temperature log(2). It is readily verified that C[0, 1]′′ = L∞[0, 1], E′′ ∼= L∞[0, 1] ⊕ L∞[0, 1] as a right module and E′′ = L∞[0, 1/2] ⊕ L∞[1/2, 1] as a left module. In particular, E′′ is an fgp bi-Hilbertian bimodule over L∞[0, 1]. The same discussion applies to any 'graph separated' iterated function system satisfying the open set condition. See [43] for more details. Theorem 6.27. Let EA be a right A-Hilbert C ∗-module with a unital left action and assume that E is finitely generated both as a left and a right A-module and E′′ is finitely generated and projective both as a left and a right A′′-module. Then EA has a left inner product such that E is an fgp bi-Hilbertian bimodule, has finite right Watatani index and is W -regular with qk invertible for all k if and only if Φ∞ is faithful and Φ∞(OE ) ⊆ A. We remark that if E is W-regular and qk is invertible for all k, then E is strictly W-regular by [31, Lemma 3.8]. Proof. First suppose that Φ∞ is faithful and Φ∞(OE) ⊆ A. Lemma 6.25 shows that E′′ has a left inner product making E (not just E′′) bi-Hilbertian with right Watatani index 1A and (invertible) qk = Id(E)⊗k . Since E, both as a left and a right module, is finitely generated and admits a Hilbert C ∗-module structure it is also projective. Conversely, if E is an fgp bi-Hilbertian bimodule with finite right Watatani index and is W- regular with invertible qk, then [56] proves that the map Φ∞(SµS∗ ν ) = A(µqν(ν)) is a (faithful) conditional expectation. That it agrees with Φ∞ is a computation. Remark 6.28. The issue with qk being non-invertible is as follows. Since qk = q1⊗ q1⊗···⊗ q1, the non-invertibility occurs with q1. Supposing OE to be strictly W-regular, we can define the right module ΞA as the completion of OE for the norm coming from Φ∞. Then Φ∞(SeS∗ f ) = A(eq1(f )), and so if q1 is not invertible, we do not get a left inner product in this way. See [56, Example 3.10] for an example where we have strict W-regularity with q1 not being invertible. Heuristically, one should view the passage to ΞA as erasing the information about the left inner product on E, corresponding to the kernel of q1. On the other hand, if qk is invertible for all k then we can replace our left inner products A(··)E⊗k and obtain an equivalent inner product structure with right Watatani index 1. by A(·qk(·))E⊗k We can now relate our constructions above back to KMS-states. 61 Proposition 6.29. Let A be a unital C ∗-algebra, E a finitely generated projective right A- Hilbert C ∗-module with a unital adjointable left A-action, α ≥ 0 and τ a positive trace on A. We assume that this data satisfies the following conditions: • τ satisfies the Laca-Neshveyev condition. • The A′′-bimodule E′′ is finitely generated and projective from the left. • The implication (6.4) (see page 59) holds. We define a semifinite spectral triple (OE, L2(OE, φLN,τ ), Dψ, N, T) using that L2(OE , φLN,τ ) = L2(OE′′, φLN,τ ) = L2(ΞA′′, τ ) and pulling back the semi-finite spectral triple defined from the fgp bi-Hilbertian A′′-bimodule E′′ as in Lemma 2.28 along the inclusion OE → OE′′. This semi-finite spectral triple is Li1-summable, α-analytic, has positive essential T-spectrum with βDψ = α and its associated KMS-state for the gauge action (as in Corollary 3.21 on page 34) coincides with φLN,τ . References [1] A. An Huef, M. Laca, I. Raeburn, A. Sims, KMS states on the C ∗-algebras of finite graphs, J. Math. Anal. and App., 405 (2), (2013), 388-399. [2] A. an Huef, I. Raeburn, Equilibrium States on Graph Algebras, In: Carlsen T.M., Larsen N.S., Neshveyev S., Skau C. (eds) Operator Algebras and Applications. Abel Symposia, vol 12. Springer, Cham, (2016). [3] N. H. Bingham, C. M. Goldie, J. L. Teugels, Regular variation, volume 27 of Ency- clopedia of Mathematics and its Applications. Cambridge University Press, Cambridge, 1989. [4] O. Bratteli, D. W. Robinson, Operator algebras and quantum statistical mechanics. 2. Equilibrium states. Models in quantum statistical mechanics. Second edition. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1997. [5] A. Carey, V. Gayral, A. Rennie, F. Sukochev, Integration on locally compact noncom- mutative spaces, J. Funct. Anal., 263 (no. 2) (2012), 383 -- 414. [6] A. Carey, V. Gayral, A. Rennie, F. Sukochev, Index theory for locally compact noncom- mutative geometries, Mem. Amer. Math. Soc. 231 no. 1085, (2014), vi+130 pp. [7] A. Carey, S. Neshveyev, R. Nest, A. Rennie, Twisted cyclic theory, equivariant KK- theory and KMS states, J. reine angew. Math., 650 (2011), 161 -- 191. [8] A. Carey, D. O'Brien, Automorphisms of the infinite dimensional Clifford algebra and the Atiyah-Singer mod 2 index theorem, Topology, 22 no. 4 (1983), 437 -- 448. [9] A. Carey, J. Phillips, Spectral flow in Fredholm modules, eta invariants and the JLO cocycle, K-Theory 31 no. 2 (2004), 135 -- 194. [10] A. Carey, J. Phillips, A. Rennie, Twisted cyclic theory and an index theory for the gauge invariant KMS state on Cuntz algebras, J. K-Theory, 6 no. 2 (2010), 339 -- 380. 62 [11] A. L. Carey, J. Phillips, F. A. Sukochev, Spectral flow and Dixmier traces, Adv. Math., 173 (2003), 68 -- 113. [12] A. Carey, A. Rennie, K. Tong, Spectral flow invariants and twisted cyclic theory for the Haar state on SUq(2), J. Geom. Phys., 59 (2009), 1431 -- 1452. [13] J. Christensen, K. Thomsen, Finite digraphs and KMS states, J. Math. Anal. Appl. 433 (2016), no. 2, 1626 -- 1646. [14] A. Connes, Compact metric spaces, Fredholm modules, and hyperfiniteness, Ergodic Theory Dynam. Systems 9 (1989), no. 2, pp. 207 -- 220. [15] A. Connes, Noncommutative geometry, Academic Press, Inc., San Diego, CA, 1994. [16] A. Connes, Geometry from a spectral point of view, Lett. Math. Phys. 34 (1995), 203 -- 238. [17] A. Connes, J. Cuntz, Quasi homomorphismes, cohomologie cyclique et positivit´e, Com- mun. Math. Phys. 114 (1988), 515 -- 526. [18] V. Deaconu, Generalized solenoids and C ∗-algebras, Pacific J. Math. 190 (1999), 247 -- 260. [19] R. J. Deeley, M. Goffeng, B. Mesland, and M. F. Whittaker, Wieler solenoids, Cuntz- Pimsner algebras and K-theory, Ergodic Theory Dynam. Systems 38 (2018), no. 8, 2942 -- 2988. [20] J. Dixmier, Existence de traces non normales, C. R. Acad. Sci. Paris S´er. A-B 262 (1966), A1107 -- A1108. [21] J. J. Duistermaat, I. M. Singer, Order-preserving isomorphisms between algebras of pseudo-differential operators, Comm. Pure Appl. Math. 29 (1976), no. 1, 39 -- 47. [22] J. V. Egorov, The canonical transformations of pseudodifferential operators, Uspehi Mat. Nauk 24 1969 no. 5 (149) 235 -- 236. [23] M. Enomoto, M. Fujii, Y. Watatani, KMS states for gauge action on OA, Math. Japon., 29 (4). (1984), 607 -- 619. [24] T. Fack, H. Kosaki, Generalised s-numbers of τ -measurable operators, Pacific J. Math., 123 (1986), 269 -- 300. [25] V. Gayral, F. Sukochev, Dixmier traces and extrapolation description of noncommuta- tive Lorentz spaces. J. Funct. Anal., 266 (10):6256 -- 6317, 2014. [26] E. Getzler, A. Szenes, On the Chern character of a theta-summable Fredholm module, J. Funct. Anal. 84 (1989), no. 2, 343 -- 357. [27] P. Gilkey, Invariance theory, the heat equation, and the Atiyah-Singer index theorem, 2nd edition, CRC Press, 1995. [28] H. Gimperlein, G. Grubb, Heat kernel estimates for pseudodifferential operators, frac- tional Laplacians and Dirichlet-to-Neumann operators, J. Evol. Equ. 14 (2014), no. 1, 49 -- 83. 63 [29] M. Goffeng, B. Mesland, Spectral triples and finite summability on Cuntz-Krieger alge- bras, Doc. Math. 20 (2015), 89 -- 170. [30] M. Goffeng, B. Mesland, Spectral triples on ON , Conference proceedings, MATRIX- program "Refining C ∗-algebraic invariants for dynamics using KK-theory" in Creswick, Australia, 2016. [31] M. Goffeng, B. Mesland, A. Rennie, Shift-tail equivalence and an unbounded represen- tative of the Cuntz-Pimsner extension, Ergodic Theory Dynam. Systems 38 (2018), no. 4, 1389 -- 1421. [32] M. Goffeng, A. Usachev, Dixmier traces and residues on weak operator ideals. arXiv:1710.08260 [33] G. Grubb, R. T. Seeley, Zeta and eta functions for Atiyah-Patodi-Singer operators, J. Geom. Anal. 6 no. 1 (1996), 31 -- 77. [34] N. Higson, J. Roe, Analytic K-Homology, Oxford Mathematical Monographs. Oxford Science Publications. Oxford University Press, Oxford, 2000. xviii+405 pp. [35] L. Hormander, Fourier integral operators. I. Acta Math. 127 (1971), no. 1-2, 79 -- 183. [36] A. Jaffe, A. Lesniewski, and K. Osterwalder, Quantum K-theory. I. The Chern charac- ter, Comm. Math. Phys. 118 no. 1 (1988), 1 -- 14. [37] A. Jaffe, A. Lesniewski, and K. Osterwalder, On super-KMS functionals and entire cyclic cohomology, K-Theory 2 no. 6 (1989), 675 -- 682. [38] M. Junge, T. Mei, J. Parcet, Noncommutative Riesz transforms?dimension free bounds and Fourier multipliers, J. Eur. Math. Soc. (JEMS) 20 (2018), no. 3, 529 -- 595. [39] J. Kaad, R. Nest, A. Rennie, KK-theory and spectral flow in von Neumann algebras, J. K-Theory, 10 (no. 2) (2012), 241 -- 277. [40] J. Kaad, W. D. van Suijlekom, Riemannian submersions and factorization of Dirac operators, arXiv:1610.02873. [41] T. Kajiwara, C. Pinzari, Y. Watatani, Jones index theory for Hilbert C ∗-bimodules and its equivalence with conjugation theory, J. Funct. Anal. 215 (2004), 1 -- 49. [42] T. Kajiwara, Y. Watatani, KMS states on finite-graph C ∗-algebras, Kyushu J. Math. 67 (1) (2013), 83 -- 104. [43] T. Kajiwara, Y. Watatani, C ∗-algebras associated with self-similar sets, J. Operator Theory, 56 (2) (2006), 225 -- 247. [44] J. Korevaar, Tauberian theory, volume 329 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 2004. A century of developments. [45] M. Laca, S. Neshveyev, KMS states of quasi-free dynamics on Pimsner algebras, J. Funct. Anal. 211 (2004), 457 -- 482. 64 [46] E. C. Lance, Hilbert C ∗-modules. A toolkit for operator algebraists. London Mathemat- ical Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995. x+130 pp. [47] S. Lord, A. Rennie, J. V´arilly, Riemannian manifolds in noncommutative geometry, J. Geom. Phys. 62 (2012), 1611 -- 1638. [48] S. Lord, F. Sukochev, D. Zanin, Singular Traces: Theory and Applications, volume 46 of Studies in Mathematics. De Gruyter, 2012. [49] L. MacDonald, A. Rennie, The Godbillon-Vey invariant in equivariant KK-theory, arXiv:1811.04603. [50] B. Mesland, M. H. Sengun, Hecke operators in KK-theory and the K-homology of Bianchi groups, to appear in Journal of Noncommutative Geometry, arXiv:1610.06808. [51] M. Mimura, Rigidity Theorems For Universal And Symplectic Universal Lattices, PhD-thesis University of Tokyo, 2011, https://www.math.tohoku.ac.jp/∼mimura- mas/mimura phdthesis.pdf [52] V. Nistor, A bivariant Chern-Connes character, Ann. of Math. (2) 138 no. 3 (1993), 555 -- 590. [53] N. Ozawa, Amenable actions and exactness for discrete groups, C. R. Acad. Sci. Paris Sr. I Math. 330 (2000), no. 8, 691 -- 695. [54] I. Raeburn, Graph Algebras, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 2005, vi+113. [55] I. Raeburn, D. P. Williams, Morita Equivalence and Continuous-Trace C ∗-algebras, Mathematical Surveys and Monographs 60, American Mathematical Society, Provi- dence, RI, 1998. [56] A. Rennie, D. Robertson, A. Sims, The extension class and KMS states for Cuntz -- Pimsner algebras of some bi-Hilbertian bimodules, Journal of Topology and Analysis, 09 no. 02, (2017) 297 -- 327. [57] A. Rennie, D. Robertson, A. Sims, Poincar´e duality for Cuntz-Pimsner algebras of bimodules, arXiv:1804.08114. [58] A. Rennie, R. Senior, The resolvent cocycle in twisted cyclic cohomology and a local index formula for the Podle´s sphere, J. Noncommut. Geom., 8 (no. 1) (2014), 1 -- 43. [59] E. Schrohe, M. Walze, J.-M. Warzecha, Construction de triplets spectraux ´a partir de modules de Fredholm, C. R. Acad. Sci. Paris S´er. I Math. 326 no. 10 (1998), 1195 -- 1199. [60] A. Sedaev, F. Sukochev, Dixmier measurability in Marcinkiewicz spaces and applica- tions, J. Funct. Anal. 265 no. 12 (2013), 3053 -- 3066. [61] F. Sukochev, A. Usachev, D. Zanin, Dixmier traces generated by exponentiation invari- ant generalised limits, J. Noncommut. Geom. 8 no. 2 (2014), 321 -- 336. 65 [62] M. Takesaki, Theory of operator algebras. I. Reprint of the first (1979) edition. En- cyclopaedia of Mathematical Sciences, 124. Operator Algebras and Non-commutative Geometry, 5. Springer-Verlag, Berlin, 2002. xx+415 pp. [63] D. Voiculescu, On the existence of quasicentral approximate units relative to normed ideals. Part I, J. Funct. Anal. 91 (1990), 1 -- 36. [64] P. Walters, Convergence of the Ruelle operator for a function satisfying Bowen's condi- tion, Trans. Amer. Math. Soc. 353 no. 1 (2001), 327 -- 347. [65] G. P. Wene, The Clifford algebra of an infinite-dimensional space, J. Math. Phys. 30 (1989), 249 -- 251. 66
1604.08774	3	1604	2017-04-03T16:35:54	Just-infinite C*-algebras	[ "math.OA", "math.GR" ]	By analogy with the well-established notions of just-infinite groups and just-infinite (abstract) algebras, we initiate a systematic study of just-infinite C-algebras, i.e., infinite dimensional C-algebras for which all proper quotients are finite dimensional. We give a classification of such C-algebras in terms of their primitive ideal space that leads to a trichotomy. We show that just-infinite, residually finite dimensional C-algebras do exist by giving an explicit example of (the Bratteli diagram of) an AF-algebra with these properties. Further, we discuss when C-algebras and -algebras associated with a discrete group are just-infinite. If $G$ is the Burnside-type group of intermediate growth discovered by the first named author, which is known to be just-infinite, then its group algebra $C[G]$ and its group C-algebra $C^(G)$ are not just-infinite. Furthermore, we show that the algebra $B = \pi(C[G])$ under the Koopman representation $\pi$ of $G$ associated with its canonical action on a binary rooted tree is just-infinite. It remains an open problem whether the residually finite dimensional C-algebra $C^_\pi(G)$ is just-infinite.	math.OA	math	Just-infinite C -algebras Rostislav Grigorchuk∗, Magdalena Musat: and Mikael Rørdam† Dedicated to Efim Zelmanov on the occasion of his 60th birthday Abstract By analogy with the well-established notions of just-infinite groups and just-infinite (abstract) algebras, we initiate a systematic study of just-infinite C -algebras, i.e., infi- nite dimensional C -algebras for which all proper quotients are finite dimensional. We give a classification of such C -algebras in terms of their primitive ideal space, that leads to a trichotomy. We show that just-infinite, residually finite dimensional C -algebras do exist by giving an explicit example of (the Bratteli diagram of) an AF-algebra with these properties. Further, we discuss when C -algebras and -algebras associated with a discrete group are just-infinite. If G is the Burnside-type group of intermediate growth discovered by the first-named author, which is known to be just-infinite, then its group algebra CrGs and its group C -algebra C pGq are not just-infinite. Furthermore, we show that the algebra B " πpCrGsq under the Koopman representation π of G associated with its canonical action on a binary rooted tree is just-infinite. It remains an open problem whether the residually finite dimensional C -algebra C π pGq is just-infinite. 1 Introduction A group is said to be just-infinite if it is infinite and all its proper quotients are finite. Just-infinite groups arise, e.g., as branch groups (including the Burnside-type group of intermediate growth discovered by the first named author, see [19]). A trichotomy describes the possible classes of just-infinite groups, see [22, Theorem 3]. Each finitely generated infinite group has a just-infinite quotient. Therefore, if we are interested in finitely generated infinite groups satisfying a certain exotic property preserved by homomorphic images, if such a group exists, then one is also to be found in the class of just-infinite groups. The purpose of this paper is to investigate just-infinite dimensional C -algebras, defined to be infinite dimensional C -algebras for which all proper quotients by closed two-sided ideals are finite dimensional. (In the future, we shall omit "dimensional" and refer to these C -algebras as just-infinite. The well-established notion of infiniteness of a unital C -al- gebra, that is, its unit is Murray-von Neumann equivalent to a proper subprojection, is unrelated to our notion of just-infiniteness.) Analogous to just-infiniteness in other cate- gories, any infinite dimensional simple C -algebra is just-infinite for trivial reasons. It is also easy to see that if a C -algebra A contains a simple essential closed two-sided ideal I such that A{I is finite dimensional, then A is just-infinite. (A closed two-sided ideal in a C -algebra is essential if it has non-zero intersection with every other non-zero closed ∗The first named author was supported by NSF grant DMS-1207699 and NSA grant H98230-15-1-0328 †The second and third named authors were supported by the Danish National Research Foundation (DNRF) through the Centre for Symmetry and Deformation at University of Copenhagen, and The Danish Council for Independent Research, Natural Sciences. 1 two-sided ideal.) Hence, e.g., any essential extension of the compact operators on an infinite dimensional separable Hilbert space by a finite dimensional C -algebra is just-infinite. We give in Theorem 3.10 a classification of just-infinite C -algebras into three types, depending on their primitive ideal space. In more detail, if A is a separable just-infinite C -algebra, then its primitive ideal space is homeomorphic to one of the T0-spaces Yn, 0 ď n ď 8, defined in Example 3.7. The case n " 0 corresponds to A being simple, while the case 1 ď n ă 8 occurs when A is an essential extension of a simple C -algebra by a finite dimensional C -algebra with n simple summands. If the primitive ideal space of a separable just-infinite C -algebra A is infinite, then it is homeomorphic to the T0-space Y8, and in this case A is residually finite dimensional (i.e., there is a separating family of finite dimensional representations of A). The C -algebra A has an even stronger property, described in Section 2, that we call strictly residually finite dimensional. We refer the reader to the survey paper [6] for a more comprehensive treatment of residually finite groups and residually finite dimensional group C -algebras. To our knowledge, residually finite dimensional, for short RFD, just-infinite C -alge- bras have not been previously considered in the literature. A priori it is not even clear that they exist. This issue is settled in Section 4, where we construct a RFD just-infinite unital AF-algebra, by giving an explicit description of its Bratteli diagram. Residually finite dimensional C -algebras have been studied extensively, see for example [18], [15], [14], [6], [13]. They are always quasidiagonal (see, e.g., [12] or [13, Chapter 7]). Interesting classes of C -algebras, such as the full group C -algebras of the free groups and subhomogenous C -algebras, are RFD. Among RFD C -algebras, the just-infinite ones are distinguished by having the smallest possible ideal lattice. In Section 5, we show that unital, separable, RFD just-infinite C -algebras need not be AF-algebras, nor nuclear, or even exact. Using a construction of Dadarlat from [14], we show that the just-infinite, residually finite dimensional AF-algebra constructed in Section 4 contains a RFD just-infinite, non-nuclear sub-C -algebra. Moreover, this AF-algebra is contained in a non-exact C -algebra, which, likewise, is RFD and just-infinite. Just-infiniteness for C -algebras is less prevalent than the corresponding property in the category of groups. Not every infinite dimensional C -algebra has a just-infinite quotient, since, for example, no abelian C -algebra is just-infinite; cf. Example 3.3. There seems to be no natural condition ensuring that a C -algebra has a just-infinite quotient. We discuss in Section 6 when a group C -algebra is just-infinite, depending on properties of the group. We prove that the full group C -algebra C pGq of a discrete group G is just- infinite if and only if its group algebra CrGs has a unique (faithful) C -norm and it is -just-infinite, i.e., is just-infinite as a -algebra. The former property holds trivially for any locally finite group. We do not know of any non-locally finite group for which CrGs has unique C -norm. We show that there are locally finite just-infinite groups G, for which the group C -algebra C pGq and the group algebra CrGs are just-infinite. If the reduced group C -algebra C λ pGq must be simple, or G must be λpGq coincides with the full group C -algebra C pGq. It seems amenable, in which case C plausible that the group C -algebra associated with unitary representations other than the universal or the left-regular one might be just-infinite for a larger class of groups. λ pGq is just-infinite, then either C If the group algebra CrGs of a group G is -just-infinite, then G must be just-infinite, but the converse does not hold. Indeed, we show in Section 7 (Theorem 7.10) that CrGs is not -just-infinite whenever G is a branch group, while there are many branch groups which are just-infinite, e.g., the group G of intermediate growth mentioned above. We show that the image B " πpCrGsq of CrGs under the Koopman representation π of G, associated with the canonical action of G on a binary rooted tree, is just-infinite. We leave open the 2 question whether or not the C -completion C π pGq of B is just-infinite. In the affirmative case, this would provide an example of a RFD just-infinite C -algebra arising from a group. 2 Preliminaries As we shall later describe just-infinite C -algebras in terms of their primitive ideal space, and as the interesting cases of just-infinite C -algebras are those that are residually finite dimensional, we review in this section the relevant background. 2.1 The primitive ideal space of a C -algebra A C -algebra A is said to be primitive if it admits a faithful irreducible representation on some Hilbert space. It is said to be prime if, whenever I and J are closed two-sided ideals in A such that I X J " 0, then either I " 0, or J " 0. It is easy to see that every primitive C -algebra is prime, and it is a non-trivial result that the converse holds for all separable C -algebras; cf. [31, Proposition 4.3.6]. However, there are (complicated) examples of non-separable C -algebras that are prime, but not primitive, see [37]. A closed two-sided ideal I in a C -algebra A is said to be primitive if I ‰ A and I is the kernel of an irreducible representation of A on some Hilbert space. The primitive ideal space, PrimpAq, is the set of all primitive ideals in A. A closed two-sided ideal I of A is primitive if and only if the quotient A{I is a primitive C -algebra. In particular, 0 P PrimpAq if and only if A is primitive. The primitive ideal space is a T0-space when equipped with the hull-kernel topology, which is given as follows: the closure F of a subset F Ď PrimpAq consists of all ideals I P PrimpAq which contain ŞJPF J. If A is primitive, so that 0 P PrimpAq, then t0u " PrimpAq. In the commutative case, the primitive ideal space is the usual spectrum: PrimpC0pXqq is homeomorphic to X, whenever X is a locally compact Hausdorff space. The following fact will be used several times in the sequel: Remark 2.1. Each finite dimensional C -algebra A is (isomorphic to) a direct sum of full matrix algebras, A – Mk1pCq ' Mk2pCq ' ' MknpCq, for some positive integers n, k1, k2, . . . , kn. As each matrix algebra is simple, PrimpAq can be naturally identified with the set t1, 2, . . . , nu, equipped with the discrete topology. The primitive ideal space is Hausdorff in this case. A closed subset F of a T0-space X is said to be prime if, whenever F 1 and F 2 are closed subsets of X such that F Ď F 1 Y F 2, then F is contained in one of F 1 and F 2. The closure of any singleton is clearly prime. A spectral space is a T0-space for which the converse holds: each closed prime subset is the closure of a singleton. The results listed in the proposition below can be found in [30, Sect. 5.4], or [31, Sect. 4.3]: Proposition 2.2. Let A be a C -algebra. (i) If A is unital, then PrimpAq is a compact1 T0-space. (ii) Let I P PrimpAq. Then tIu is closed in PrimpAq if and only if I is a maximal proper ideal in A, i.e., if and only if the quotient A{I is simple. (iii) If A is separable, then PrimpAq is a second countable spectral space. 1A (possibly non-Hausdorff) topolotical space is said to be compact if it has the Heine-Borel property: each open cover can be refined to a finite open cover. Sometimes this property is referred to as quasi-compactness. 3 By Remark 2.1, the only finite dimensional C -algebras which are primitive are those which are isomorphic to full matrix algebras. Hence, the following holds: Proposition 2.3. Let A be a separable C -algebra, and let I P PrimpAq be such that A{I is finite dimensional. Then A{I – MkpCq, for some k P N, and tIu is closed in PrimpAq. A T0-space X is said to be totally disconnected if there is a basis for its topology consisting of compact-open sets. If the projections in a C -algebra A separate its ideals, then PrimpAq is totally disconnected. In this situation, we have the following result, which will be discussed in more detail in Section 4: Theorem 2.4 (Bratteli–Elliott, [11]). Let X be a second countable, totally disconnected spectral space. Then X is homeomorphic to PrimpAq, for some separable AF-algebra A. If X is compact, then A can be taken to be unital. Recall that an AF-algebra is a C -algebra which is the completion of an increasing union of finite dimensional sub-C -algebras. We end this section by recalling that there is a one-to-one correspondence between open subsets U of PrimpAq and closed two-sided ideals IpUq of A, given by IpUq " čJPPrimpAqzU J, with the convention that IpHq " 0 and IpPrimpAqq " A. Moreover, PrimpA{IpUqq " PrimpAqzU , (2.1) (2.2) for each open subset U of PrimpAq. Consequently, each closed subset of PrimpAq is the primitive ideal space of a quotient of A (see [31, Theorem 4.1.3]). Note furthermore that if J P PrimpAq, then J " IpUq, where U is the complement of the closure of tJu. 2.2 Residually finite dimensional C -algebras This section is devoted to discussing residually finite dimensional C -algebras and their primitive ideal spaces. We also introduce the class of so-called strictly residually finite dimensional C -algebras, and describe them in terms of their primitive ideal space. A C -algebra A is said to be residually finite dimensional (RFD), if it admits a separating family of finite dimensional representations. The finite dimensional representations can be taken to be irreducible and pairwise (unitarily) inequivalent. (We note that two irreducible finite dimensional representations are equivalent if and only if they are weakly equivalent, i.e., they have the same kernel.) Assume that tπiuiPI is a family of irreducible and pairwise inequivalent finite dimensional representations of a C -algebra A. Let ki be the dimension of the representation πi, and identify the image of πi with MkipCq. We then get a -homomorphism ΨI " àiPI πi : A Ñ źiPI MkipCq. Note that ΨI is injective if and only if ŞiPI Kerpπiq " t0u, which again happens if and only if tKerpπiq : i P Iu is a dense subset of PrimpAq. Therefore, the following lemma holds; cf. Proposition 2.3: Lemma 2.5. A C -algebra A is RFD if and only if PrimpAq contains a dense subset P such that A{I is a full matrix algebra, for each I P P. 4 Observe that tIu is closed in PrimpAq, for each I P P, by Proposition 2.2 (ii). separable, then one can choose the set P in the lemma above to be countable. If A is Since the ideals Kerpπiq are maximal and pairwise distinct (by the assumed inequiva- lence of the finite dimensional representations πi, which implies that they are also weakly inequivalent), it follows from the Chinese Remainder Theorem that the map ΨF " àiPF πi : A Ñ źiPF MkipCq is surjective, for each finite subset F of I. Definition 2.6. A unital C -algebra A is said to be strictly residually finite dimensional (strictly RFD ) if there exists an infinite family tπi : A Ñ MkipCquiPI of irreducible, pairwise inequivalent, finite dimensional representations of A such that the map ΨT " àiPT πi : A Ñ źiPT MkipCq (2.3) is injective, for each infinite subset T of I. The following characterizes strictly RFD C -algebras in terms of their primitive ideal space: Proposition 2.7. A unital separable C -algebra A is strictly RFD if and only if there exists an infinite subset P of PrimpAq such that each of its infinite subsets is dense in PrimpAq, and such that A{I is finite dimensional, for each I P P. Note that if such a subset P of PrimpAq exists, then each infinite subset of P has the same properties, and hence one can take P to be countably infinite. Proof. Suppose first that A is a strictly RFD unital separable C -algebra witnessed by an infinite family tπi : A Ñ MkipCquiPI of irreducible, pairwise inequivalent, finite dimensional representations. Set P " tKerpπiq : i P Iu Ď PrimpAq. If I " Kerpπiq P P, then A{I – πipAq is finite dimensional. Let T be an infinite subset of I, then 0 " KerpΨT q " ŞiPT Kerpπiq. Therefore tKerpπiq : i P T u is dense in PrimpAq. Suppose conversely that A is a unital separable C -algebra for which there exists an infinite subset P " tIiuiPI of PrimpAq satisfying the hypotheses. For each i P I, find an irreducible representation πi : A Ñ BpHiq with Kerpπiq " Ii. Then πipAq – A{Ii is finite dimensional, so Hi is finite dimensional and πipAq " BpHiq – MkipCq, where ki " dimpHiq. Let T be an infinite subset of I. The associated map ΨT then satisfies KerpΨT q " čiPT Kerpπiq " čiPT Ii " 0, by the assumption that tIiuiPT is dense in PrimpAq. This shows that A is strictly RFD. 3 Just-infinite C -algebras: A classification result By analogy with the notion of just-infiniteness in the category of groups and of abstract algebras, see [29], we define a C -algebra to be just-infinite as follows: Definition 3.1. A C -algebra A is said to be just-infinite if it is infinite dimensional, and for each non-zero closed two-sided ideal I in A, the quotient A{I is finite dimensional. 5 Lemma 3.2. Every just-infinite C -algebra is prime. Proof. Let A be a just-infinite C -algebra, and let I and J be two non-zero closed two-sided ideals in A. Consider the natural homomorphism π : A Ñ A{I ' A{J. By the assumption that A is just-infinite, the image is finite dimensional. It follows that π cannot be injective, so I X J " Kerpπq ‰ 0. Example 3.3. The group Z is just-infinite, and it is the only abelian just-infinite group. It is also known, see [22, Proposition 3(a)], that every finitely generated infinite group has a just-infinite quotient. The corresponding statements for C -algebras are false: No commutative C -algebra is just-infinite, since no commuative C -algebra other than C is prime. This also shows that no commutative C -algebra has a just-infinite quotient. It is well-known that every unital C -algebra has a maximal proper closed two-sided ideal, and hence a quotient which is simple. If, moreover, such a simple quotient is infinite dimensional, then it is just-infinite. There seems to be no satisfactory description of unital C -algebras having an infinite dimensional simple quotient. Lemma 3.4. Each non-zero closed two-sided ideal in a just-infinite C -algebra is essential and infinite dimensional. Proof. It is easy to see that a C -algebra is prime if and only if each non-zero closed two- sided ideal is essential, so the first statement of the lemma follows from Lemma 3.2. If a closed two-sided ideal I in a C -algebra A has a unit e, then e is a central projection in A and I " Ae. Thus Ae and Ap1 ´ eq are closed two-sided ideals in A with zero intersection. So if I is essential, then Ap1 ´ eq " 0 and I " A. Now, as each finite dimensional C -algebra has a unit, we see that no non-zero closed two-sided ideal in a just-infinite C -algebra can be finite dimensional. The class of just-infinite C -algebras does not have good permanence properties. In fact, almost all natural operations on C -algebras (such as passing to sub-C -algebras, extensions, passing to ideals and quotients, taking inductive limits, Morita equivalence, forming crossed products by suitable groups) fail to be consistent with the class of just-infinite C -algebras. However, the following permanence-type properties of just-infinite C -algebras do hold: Proposition 3.5. (i) If B is an infinite dimensional hereditary2 sub-C -algebra of a just-infinite C -algebra In particular, each non-zero closed two-sided ideal in a A, then B is just-infinite. just-infinite C -algebra is again just-infinite. (ii) Let 0 Ñ I Ñ A Ñ Q Ñ 0 be a short exact sequence of C -algebras, where I is an essential ideal in A. Then A is just-infinite if and only if Q is finite dimensional and I is just-infinite. Proof. (i). Let J be a non-zero closed two-sided ideal in B, and let I be the (necessarily non-zero) closed two-sided ideal in A generated by J. Then J " B X I, so B{J is isomorphic to a (hereditary) sub-C -algebra of A{I. The latter is finite dimensional, so B{J must also be finite dimensional. 2A sub-C -algebra B of a C -algebra A is hereditary if whenever b P B and a P A are such that 0 ď a ď b, then a P B. 6 The second part of (i) follows from the fact that each closed two-sided ideal in a C - algebra is a hereditary sub-C -algebra, together with Lemma 3.4, which ensures that each non-zero closed two-sided ideal in A must be infinite dimensional. (ii). Suppose that I is just-infinite and Q is finite dimensional. Let J be a non-zero ideal in A. Then we have a short exact sequence 0 Ñ I{pI X Jq Ñ A{J Ñ Q{πpJq Ñ 0, where π : A Ñ Q is the quotient mapping. Now, I X J is non-zero (since I is an essential ideal), so I{pI X Jq is finite dimensional. This implies that A{J is finite dimensional, being an extension of two finite dimensional C -algebras. Conversely, if A is just-infinite, then Q, which is isomorphic to the quotient A{I, is finite dimensional (because I is non-zero). The ideal I cannot be finite dimensional (since otherwise A would be finite dimensional), so it follows from (i) that I is just-infinite. The observation made above that the class of just-infinite C -algebras is not closed under Morita equivalence, can be justified as follows. If A is a just-infinite C -algebra and if K denotes the C -algebra of compact operators on a separable Hilbert space, then A b K is just-infinite if and only if A is simple (since all proper non-zero quotients of Ab K are stable, and therefore infinite dimensional). Remark 3.6 (Hereditary just-infiniteness). If G is a residually finite group and all its normal subgroups of finite index, including G itself, are just-infinite, then G is said to be hereditarily just-infinite. Just-infinite branch groups are residually finite, but not hereditarily just- infinite; cf. [22, Section 6], so a finite index normal subgroup of a just-infinite group need not be just-infinite. We shall say more about hereditarily just-infinite groups and just- infinite branch groups in Examples 6.12 and 6.13 and in Theorem 7.10. It follows from Proposition 3.5 above that just-infinite C -algebras automatically have a property analogous to being hereditarily just-infinite for groups: Any non-zero closed two- sided ideal in a just-infinite C -algebra is itself just-infinite. Note also that the following three conditions for a closed two-sided ideal I in a just-infinite C -algebra A are equiva- lent (cf. Lemma 3.4 and the definition of being just-infinite): I is non-zero, I is infinite dimensional, and I has finite co-dimension in A. We proceed to describe the primitive ideal space of a just-infinite C -algebra. They turn out to be homeomorphic to one of the T0-spaces in the following class: Example 3.7. For each n P t0, 1, 2, . . . , 8u, consider the T0-space Yn defined to be the disjoint union Yn " t0u Y Y 1 n has countably infinitely many elements, if n " 8. Equip Yn with the topology for which the closed subsets of Yn are precisely the following sets: H, Yn, and all finite subsets of Y 1 n. n is a set with n elements, if n is finite, and Y 1 n, where Y 1 We shall usually take Y 1 n to be t1, 2, . . . , nu, if 1 ď n ă 8, and Y 1 8 to be N. The spaces Yn have the following axiomatic properties: Lemma 3.8. A (non-empty) second countable T0-space X is homeomorphic to Yn, for some n P t0, 1, 2, . . . , 8u, if and only if it the following conditions hold, for some point x0 P X: (A) tx0u is dense in X, (B) txu is closed, for all x P Xztx0u, (C) each infinite subset of X is dense in X. Moreover, if X is any T0-space satisfying conditions (A), (B) and (C) above, then 7 (i) the closed subsets of X are the following sets: H, X, and all finite subsets of Xztx0u, (ii) X is second countable if and only if X is countable, (iii) each subset of X is compact (in particular, X is totally disconnected), (iv) X is a spectral space.3 Proof. The spaces Yn satisfy conditions (A), (B), and (C) with x0 " 0. We show below that (A), (B) and (C) imply (i)–(iv). Any second countable T0-space X satisfying (i) and (ii) is homeomorphic to Yn, where n is the cardinality of X 1 " Xztx0u. Indeed, X is countable by (ii), and any bijection f : X Ñ Yn, with f px0q " 0, is a homeomorphism by (i). Let now X be a T0-space satisfying (A), (B) and (C). We show that (i), (ii), (iii) and (iv) hold. Set X 1 " Xztx0u. It follows from (B) that each finite subset of X 1 is closed, and so are H and X. Conversely, if F is a closed subset of X and if F ‰ X, then F must be a finite subset of X 1 by (A) and (C). Hence (i) holds. Suppose now that X is second countable and X ą 1. Let tUnu8 for some n, whence x P Fn. Thus X 1 is contained in the countable set Ť8 n"1 be a basis for the topology on X consisting of non-empty open sets. For n ě 1, set Fn " XzUn, and observe that Fn is finite (or empty) by (i). Let x P X 1. Then Xztxu is open by (i), so Un Ď Xztxu n"1 Fn, so (ii) holds. Let K be an arbitrary subset of X and let tWiuiPI be an open cover of K. Take any i0 P I such that Wi0 is non-empty. Then the set F " XzWi0 is finite. Hence F X K is covered by finitely many open sets from the collection tWiuiPI which, together with Wi0, gives a finite open cover of K. This proves (iii). Finally, let F ‰ H be a closed subset of X which is prime. If F " X, then F is the closure of tx0u. If F ‰ X, then F is a finite subset of X 1, by (i). Write F " ŤxPF txu, and note that each singleton txu, x P F , is closed. Hence F can have at most one element, so it is in particular the closure of a singleton. This proves that X is a spectral space. Lemma 3.9. Let A be a separable C -algebra. The following hold: (i) PrimpAq is homeomorphic to Yn, for some n P t0, 1, 2, . . . , 8u, if and only if the following three conditions hold: (a) A is primitive, (b) A{I is simple, for each non-zero primitive ideal I in A, (c) if PrimpAq is infinite, then ŞIPP I " 0, for each infinite subset P of PrimpAq. (ii) If PrimpAq is infinite and A satisfies (b) and (c), then it automatically satisfies (a). If A{I is finite dimensional, for each non-zero I P PrimpAq, then condition (b) holds. (iii) If A is just-infinite, then PrimpAq " Yn, for some n P t0, 1, 2, . . . , 8u. Proof. (i). It follows from Proposition 2.2 (iii) that PrimpAq is second countable. It therefore suffices to show that the conditions (a), (b) and (c) are equivalent to items (A), (B) and (C) of Lemma 3.8 (with X " PrimpAq and x0 " 0). By definition, A is primitive if and only if 0 P PrimpAq, so (a) is equivalent to (A). The equivalence of (b) and (B) follows from Proposition 2.2 (ii), while the equivalence of (c) and (C) follows from the fact that a subset P of PrimpAq is dense if and only if ŞIPP I " 0. (ii). Suppose that PrimpAq is infinite and that (b) and (c) are satisfied. We assert that (a) holds, as well. By [31, Proposition 4.3.6] it suffices to check that PrimpAq is prime, i.e., 3See definition above Proposition 2.2 8 whenever PrimpAq " F Y G, where F and G are closed subsets of PrimpAq, then one of F and G is equal to PrimpAq. However, one of F and G must be infinite, and hence dense in PrimpAq by (c) (which is equivalent to (C)), and therefore one of F and G must be equal to PrimpAq. The remaining assertion follows from Proposition 2.3. (iii). Suppose that A is just-infinite. To see that PrimpAq – Yn, for some n, it suffices by (i) and (ii) to check that (b) and (c) hold. Moreover, we conclude from (ii) that (b) holds because A{I must be finite dimensional, for each non-zero ideal I. Suppose that P is an infinite subset of PrimpAq, and set J " ŞIPP I. For each finite subset F of Pzt0u, let JF " ŞIPF I Ě J. Then A{JF is isomorphic to ÀIPF A{I, which has dimension at least F, so A{J also has dimension at least F. As this holds for all finite subsets F of P, we conclude that A{J must be infinite dimensional. Hence J " 0, since A is just-infinite. This proves that P is dense in PrimpAq, so (c) holds. Just-infinite C -algebras are classified as follows (to be compared with [22, Theorem 3]): Theorem 3.10. Let A be a separable C -algebra. Then A is just-infinite if and only if PrimpAq is homeomorphic to Yn, for some n P t0, 1, 2, . . . , 8u, and each non-faithful irre- ducible representation of A is finite dimensional. (If n " 0, we must also require that A is infinite dimensional; this is automatic when n ě 1.) Moreover: (α) PrimpAq " Y0 if and only if A is simple. Every infinite dimensional simple C -algebra is just-infinite. (β) PrimpAq " Yn, for some integer n ě 1, and A is just-infinite, if and only if A con- tains a simple non-zero essential infinite dimensional ideal I0 such that A{I0 is finite dimensional. In this case, n is equal to the number of simple summands of A{I0. (γ) The following conditions are equivalent: (i) A is just-infinite and PrimpAq " Y8, (ii) A is just-infinite and RFD, (iii) PrimpAq is an infinite set, all of its infinite subsets are dense, and A{I is finite dimensional, for each non-zero I P PrimpAq, (iv) PrimpAq is an infinite set, the direct sum representation ÀiPT πi is faithful for each infinite family tπiuiPT of pairwise inequivalent irreducible representations of A, and each non-faithful irreducible representation of A is finite dimensional. We shall occasionally refer to a just-infinite C -algebra as being of type (α), (β) and (γ), respectively, if it satisfies the corresponding condition in the theorem above. In view of the theorem, we shall also, more frequently, refer to a just-infinite C -algebra of type (γ) as a RFD just-infinite C -algebra. Proof. If A is just-infinite and separable, then PrimpAq " Yn, for some n P t0, 1, 2 . . . , 8u, by Lemma 3.9, and each non-faithful irreducible representation of A is finite dimensional (by the definition of being just-infinite). Suppose conversely that PrimpAq " Yn, for some n P t0, 1, 2, . . . , 8u, and that each non- faithful irreducible representation of A is finite dimensional. We show that A then must be just-infinite (if it is also infinite dimensional). This is clear if n " 0, since A is simple in this case. This also shows that (α) holds. 9 Suppose that 1 ď n ď 8. Since Yn is non-Hausdorff, when n ą 0, and the primitive ideal space of any finite dimensional C -algebra is Hausdorff; cf. Remark 2.1, A must be infinite dimensional. Write PrimpAq " t0u Y tIjun j"1. (3.1) Any non-zero proper ideal J of A is the intersection of the primitive ideals in A that contain it. By Lemma 3.9 (c), any intersection of infinitely many distinct primitive ideals of A is zero. Hence J " ŞjPF Ij, for some, necessarily finite, subset F of N (or of t1, 2, . . . , nu, if n ă 8). Now, A{J is isomorphic to ÀjPF A{Ij, and each quotient A{Ij is finite dimensional by assumption, whence A{J is finite dimensional. This shows that A is just-infinite. We proceed to verify the claims in (β) and (γ). (β). The "if" part follows from Proposition 3.5 (ii). Moreover, PrimpAq consists of 0 (cf. Lemma 3.2) and the kernels of the maps onto the n simple summands of A{I0, so PrimpAq has cardinality n ` 1. Also, PrimpAq is homeomorphic to Yk, for some k, by Lemma 3.9 (iii), and by cardinality considerations, we conclude that k " n. some n P N. Retain the notation set forth in (3.1), and let I0 " Şn Let us prove the "only if" part. Suppose that A is just-infinite and PrimpAq " Yn, for j"1 Ij. In the notation from (2.1), we have I0 " Ipt0uq (observe that t0u is an open subset of PrimpAq, when n ă 8). We deduce that I0 is non-zero and simple. Each non-zero ideal in a primitive C - algebra is essential, so I0 is an essential ideal in A, by Lemma 3.2. Since A is just-infinite, A{I0 is finite dimensional. Finally, by (2.2), PrimpA{I0q " PrimpAqzt0u " tI1, I2, . . . , Inu, and since A{I0 is finite dimensional, n is the number of direct summands of A{I0; cf. Remark 2.1. (γ). (i) ñ (iii). If A is just-infinite, then A{I is finite dimensional, for each non-zero ideal I in A; and if PrimpAq " Y8, then each infinite subset of PrimpAq is dense (by Lemma 3.9 (i)(c)), and PrimpAq is an infinite set. (iii) ñ (ii). The assumptions in (iii) imply that A is RFD; cf. Lemma 2.5. If π is a non-faithful irreducible representation of A, then Kerpπq " I is a non-zero primitive ideal in A, so πpAq – A{I is finite dimensional. To conclude that A is just-infinite we show that PrimpAq is homeomorphic to Y8. For this it suffices to verify conditions (b) and (c) of Lemma 3.9 (i). Item (b) holds because A{I is finite dimensional, for each non-zero I P PrimpAq; cf. Lemma 3.9 (ii). Item (c) is equivalent to condition (C) in Lemma 3.8, which holds by assumption. (ii) ñ (i). If A is RFD, then A cannot be just-infinite of type (α) or (β), so PrimpAq must be homeomorphic to Y8. (iii) ñ (iv). We already saw that (iii) implies that A is just-infinite, and hence that each non-faithful irreducible representation is finite dimensional. Let tπiuiPT be an infinite family of pairwise inequivalent irreducible representations of A. Since tKerpπiq : i P T u is an infinite set, and hence by assumption a dense subset of PrimpAq, it follows that the kernel of ÀiPT πi, which is equal to ŞiPT Kerpπiq, must be zero. (iv) ñ (iii). Let P be an infinite subset of PrimpAq, and choose pairwise inequivalent irreducible representations tπiuiPT of A such that P " tKerpπiq : i P T u. The assumptions in (iv) now yield 0 " Ker´àiPT πi¯ " čiPT Kerpπiq " čIPP I, which implies that P is dense in PrimpAq. 10 If I is a non-zero primitive ideal in A, then I " Kerpπq, for some (non-faithful) irreducible representation of A, so A{I – πpAq is finite dimensional. The following result follows immediately from Proposition 2.7 and Theorem 3.10: Corollary 3.11. Each separable RFD just-infinite C -algebra is strictly RFD. We note that not all strictly RFD C -algebras are just-infinite; cf. Section 4.3 below. Corollary 3.12. The primitive ideal space of a separable just-infinite C -algebra is count- able. Moreover, any RFD just-infinite separable C -algebra has countably infinitely many equivalence classes of finite dimensional irreducible representations. Proof. The first claim follows from Lemma 3.9 (iii). The second claim follows from Theo- rem 3.10 (γ), by the fact that there is a one-to-one correspondence between weak equivalence classes of irreducible representations and the primitive ideal space of a separable C -algebra (given by mapping an irreducible representation to its kernel), and by the fact, observed earlier, that two finite dimensional irreducible representations are unitarily equivalent if they are weakly equivalent. Remark 3.13. It is shown in Theorem 3.10 that a separable C -algebra A is just-infinite if and only if the following two conditions hold: PrimpAq " Yn, for some n P t0, 1, 2, . . . , 8u and each non-faithful irreducible representation is finite dimensional. These two conditions are independent, i.e., none of them alone implies that A is just-infinite, as shown below. If X is a Hausdorff space and k is a positive integer, then all irreducible representations of MkpCpXqq have dimension k, and PrimpMkpCpXqqq " X. If X is not a point, then X is not homeomorphic to Yn, for any n, because Yn is non-Hausdorff for all n ą 0. Therefore MkpCpXqq is not just-infinite. For each n P t0, 1, 2, . . . , 8u, there is a unital AF-algebra whose primitive ideal space is Yn, by Theorem 2.4 and Lemma 3.8. The AF-algebras obtained in this way may or may not have the property that each non-faithful irreducible representation is finite dimensional. Tensoring such an AF-algebra by a UHF-algebra, we obtain a unital separable C -algebra whose primitive ideal space is Yn, and which has no finite dimensional irreducible represen- tations. Therefore, it is not just-infinite. We show in the next example and in Section 4 below that each space Yn can be realized as the primitive ideal space of a just-infinite AF-algebra. Example 3.14 (Existence of just-infinite C -algebras). Any simple infinite dimensional C -algebra is just-infinite of type (α) (and there are many examples of such, both in the unital and the non-unital case). To exhibit examples of just-infinite C -algebras of type (β), let n P N, and let F be a finite dimensional C -algebra with n simple summands, e.g., F " C ' ' C with n summands. Let H be an infinite dimensional separable Hilbert space, and let π : BpHq Ñ BpHq{K be the quotient mapping onto the Calkin algebra, where as before K denotes the compact operators on H. Let τ : F Ñ BpHq{K be a unital injective -homomorphism. Set A " π´1`τ pF q Ď BpHq. (3.2) Then K is a simple essential ideal in A and A{K is isomorphic to F . Hence A is just-infinite of type (β), and PrimpAq " Yn; cf. Theorem 3.10 (β). Since A is an extension of two AF-algebras, it is itself an AF-algebra. 11 Each just-infinite C -algebra A arising as in (3.2) above is of type I : for each irreducible representation of A on a Hilbert space H, the image of A contains the compact operators on H. Conversely, a separable C -algebra A of type I is just-infinite if and only if it is isomorphic to the compact operators K on a separable Hilbert space, or it is of the form described in (3.2) above for some finite dimensional C -algebra F . Indeed, if A is separable, just-infinite and of type I, then A is prime by Lemma 3.2, hence primitive (because it is separable), so it admits a faithful irreducible representation ρ on some (separable) Hilbert space. Being of type I, ρpAq contains the compact operators K. If ρpAq ‰ K, then the quotient B :" ρpAq{K is finite dimensional, because A is just-infinite, so A – ρpAq " π´1pBq is as in (3.2). It requires more work to establish the existence of RFD just-infinite C -algebras, i.e., those of type (γ). This will be done in Section 4. Remark 3.15 (Characteristic sequences of just-infinite C -algebras). Let A be a unital separable just-infinite C -algebra. If A is non-simple, then PrimpAq " Yn, for some n P t1, 2, . . . , 8u. Let tIjun j"1 be the non-zero primitive ideals of A. Then A{Ij – Mkj pCq, for some kj P N; cf. Proposition 2.3. The resulting n-tuple, or sequence, tkj un j"1 (as an unordered set) is an invariant of A, which we shall call the characteristic sequence of A. that such a sequence tπjun irreducible representations of A. Equivalently, tπjun For each j, choose an irreducible representation πj : A Ñ Mkj pCq with kernel Ij. We say j"1 is an exhausting sequence of pairwise inequivalent non-faithful j"1 is an exhausting sequence of pairwise inequivalent non-faithful irreducible representations of A if PrimpAqzt0u " Kerpπjq : j " 1, 2, . . . , n(, and Kerpπjq ‰ Kerpπiq when i ‰ j. If n P N and if I0 is the (unique) simple essential ideal in A, then (as in the proof of Theorem 3.10) we have the following isomorphisms A{I0 – n àj"1 A{Ij – n àj"1 Mkj pCq. (3.3) It follows from Example 3.14 (and Remark 2.1) that for all positive integers k1, k2, . . . , kn, there exists a just-infinite C -algebra A, which is necessarily an AF-algebra, such that (3.3) holds with I0 " K. This argument shows in particular that each finite characteristic sequence tkjun j"1, where n P N, is realized by a just-infinite AF-algebra (of type (β)). We end this section by showing that the characteristic sequence tkju8 j"1 of a RFD just- infinite C -algebra must tend to infinity. The proof of this fact involves results about subhomogeneous C -algebras. Recall that a C -algebra is said to be subhomogeneous if it is isomorphic to a sub-C -algebra of MkpCpXqq, for some compact Hausdorff space X, and some k P N. The next proposition is well-known, but we include a brief proof for the sake of completeness. Proposition 3.16. For a C -algebra A, the following conditions are equivalent: (i) A is subhomogeneous, (ii) the bidual A of A is isomorphic to Àn j"1 Mkj pCpΩjqq, for some positive integers n, k1, k2, . . . , kn, and some (extremally disconnected) compact Hausdorff spaces Ω1, Ω2, . . . , Ωn, (iii) there exists a positive integer k such that each irreducible representation of A has dimension at most k, 12 (iv) there exist a positive integer k and a separating family tπiuiPT of irreducible represen- tations of A such that each πi has dimension at most k. Proof. The implication (ii) ñ (i) holds because A is a sub-C -algebra of A. If (iii) holds, then A, which is a von Neumann algebra, cannot have central summands of type In, for n ą k, or of type II or III. Therefore (ii) holds. The implication (i) ñ (iv) follows easily from the definition of subhomogeneity. Suppose now that (iv) holds, and that there exists an irreducible representation of A of dimension strictly greater than k (possibly infinite dimensional). By (a version of) Glimm's lemma, see, e.g., [33, Proposition 3.10], there is a non-zero -homomorphism ρ : C0pp0, 1sq b Mk`1 Ñ A. However, there is no non-zero -ho- momorphism C0pp0, 1sq b Mk`1 Ñ BpHq when dimpHq ď k, so it follows that πi ρ " 0, for each i P T . As the family tπiuiPT is separating, we conclude that ρ " 0, a contradiction. Lemma 3.17. No separable subhomogeneous C -algebra is just-infinite. Proof. Let A be a separable just-infinite C -algebra. Then A is prime; cf. Lemma 3.2, hence primitive, and so A admits a faithful irreducible representation. Such a representa- tion cannot be finite dimensional, because A is infinite dimensional. Hence A cannot be subhomogeneous; cf. Proposition 3.16. Proposition 3.18. Let A be a separable RFD just-infinite C -algebra with characteristic sequence tkju8 j"1. Then limjÑ8 kj " 8. Proof. Let I1, I2, . . . be the non-zero primitive ideals of A, and for each j, let πj be an irreducible representation of A whose kernel is Ij, such that the dimension of πj is kj. We must show that for each k, Tk :" tj P N : kj ď ku is finite. Suppose that the set T " Tk is infinite. Then the -homomorphism ΨT " ÀjPT πj is injective, which implies that tπjujPT is a separating family of irreducible representations of A, each of which having dimension less than or equal to k. Then Proposition 3.16 implies that A is subhomogeneous, but this is impossible by Lemma 3.17. 4 Examples of RFD just-infinite AF-algebras We construct an example of a RFD just-infinite AF-algebra. By Theorem 3.10, its primitive ideal space must be Y8. The existence of a unital AF-algebra whose primitive ideal space is homeomorphic to Y8 follows from Theorem 2.4 (Bratteli–Elliott). To conclude that such an AF-algebra is just-infinite, we must also ensure that its non-faithful irreducible representations are finite dimensional; cf. Theorem 3.10. This is accomplished by taking a closer look at the construction by Bratteli and Elliott, done in Proposition 4.1 below. 4.1 Construction of a RFD just-infinite AF-algebra Recall that a Bratteli diagram is a graph pV, Eq, where V " Ť8 n"1 En (disjoint unions), all Vn and all En are finite sets, and where each edge e P En connects a vertex v P Vn to a vertex in w P Vn`1. In this case, we write speq " v and rpeq " w, thus giving rise to the source and the range maps s, r : E Ñ V . It was shown by Bratteli, [10], that there is a bijective correspondence between Bratteli diagrams (modulo a natural equivalence class of these) and AF-algebras (modulo Morita equivalence). n"1 Vn and E " Ť8 An ideal in a Bratteli diagram pV, Eq is a subset U Ď V with the following properties: • for all e in E, if speq belongs to U , then so does rpeq, 13 • for all v in V , if trpeq e P s´1pvqu is contained in U , then v belongs to U . The ideal lattice of an AF-algebra associated with a given Bratteli diagram is isomorphic to the ideal lattice of the Bratteli diagram, see [17] or [16]. The following proposition is contained in [11]: Proposition 4.1 (Bratteli–Elliott). Let X be a second countable, compact, totally discon- nected T0-space. Let G1, G2, . . . be finite families of compact-open subsets of X such that: (i) X " ŤGPGn G, for each n ě 1, (ii) for each n ě 1, Gn`1 is a refinement of Gn, i.e., each set in Gn`1 is contained in a set in Gn, and each set in Gn is the union of sets from Gn`1, 8 (iii) ďn"1 Gn is a basis for the topology on X. Consider the Bratteli diagram for which the vertices at level n are the sets in Gn, and where there is one edge from G P Gn to G1 P Gn`1 if G1 Ď G, and none otherwise. Then there is a one-to-one correspondence between open subsets of X and ideals of the Bratteli diagram, given as follows: the ideal in the Bratteli diagram associated with an open subset U of X consists of all vertices G P Ť8 n"1 Gn for which G Ď U . If, in addition, X is a spectral space, and if A is an AF-algebra associated with the Bratteli diagram constructed above, then PrimpAq is homeomorphic to X. In the following, we construct a sequence G1, G2, G3, . . . of finite families of compact-open subsets of X satisfying the conditions of Proposition 4.1 in the case where X " Y8. Recall that Y8 " t0u Y N, that the open subsets of Y8 are H, Y8, and all co-finite subsets of N, and that every subset of Y8 is compact. For all n ě 1, set Fn,k " t1, 2, . . . , nuztku, Gn,k " Y8zFn,k, for 1 ď k ď n, and let Gn " tGn,1, Gn,2, . . . , Gn,nu. Observe that each Gn,k is open and (automatically) compact. Moreover, the sets tGnu8 satisfy conditions (i), (ii) and (iii) in Proposition 4.1. Furthermore, for 1 ď k ď n, n"1 Gn`1,k Ď Gn,k, Gn`1,n`1 Ď Gn,k. No other inclusions between sets in Gn`1 and sets in Gn hold. Therefore, the Bratteli diagram associated with this sequence of compact-open subsets of Y8 as in Proposition 4.1 is: ‚ ‚ ‚ ‚ ‚ ... ‚ ❆❆❆❆❆❆❆ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ‚ ‚ ‚ ‚ ‚ ... ‚ ‚ ... ‚ ... C C ' C C ' C ' M2pCq C ' C ' M2pCq ' M4pCq (4.1) C ' C ' M2pCq ' M4pCq ' M8pCq ‚ ... 14 The sequence of finite dimensional C -algebras on the right-hand side, equipped with unital connecting mappings given by the Bratteli diagram, defines a unital AF-algebra A, associ- ated with the Bratteli diagram. The one-to-one correspondence between (non-empty) open subsets G Ď Y8 " t0u Y N and ideals U pGq of the Bratteli diagram above is given as follows: U pGq " tGn,k Gn,k Ď Gu " tGn,k k P G, n ě max Y8zGu. E.g., U pY8zt1, 3uq " tGn,k n ě 3, k ‰ 1, 3u and U pY8ztjuq " tGn,k n ě j, k ‰ ju, j ě 1. The quotient of the AF-algebra A by the ideal in A corresponding to U pGq is given by the Bratteli diagram that arises by removing U pGq from the original diagram. The two pictures below show the ideal U pGq (in blue) and the Bratteli diagram of the quotient (in red) in the cases where G " Y8zt2u, respectively, G " Y8zt1, 3u: ‚ ‚ ‚ ‚ ‚ ... ‚ ❆❆❆❆❆❆❆ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ‚ ‚ ‚ ‚ ‚ ‚ ... ‚ ... ‚ ... ‚ ‚ ‚ ‚ ‚ ... ‚ ... ‚ ❆❆❆❆❆❆❆ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ PPPPPPPPPPPPPP ❆❆❆❆❆❆❆ ‚ ‚ ‚ ‚ ‚ ‚ ... ‚ ... ‚ ... ‚ ... The quotient of A by the ideal in A corresponding to U pGq, with G " Y8zt2u, is the AF- algebra associated to the red part of the Bratteli diagram, which is C. The quotient of A in the case where G " Y8zt1, 3u is similarly seen to be C ' M2pCq. By construction, and by Proposition 4.1, we have PrimpAq – Y8. In more detail, we have PrimpAq " t0u Y tI1, I2, I3, . . . u, where Ij is the primitive ideal in A corresponding to the ideal U pY8ztjuq of the Bratteli diagram. Arguing as in the two examples above, we see that A{Ij – MkpjqpCq, where kp1q " kp2q " 1 and kpjq " 2j´1, for j ě 2. Hence A{I is finite dimensional, for each non-zero primitive ideal I of A. It now follows from Theorem 3.10 that A is just-infinite and RFD, as desired. The characteristic sequence of A is precisely the sequence tkpjqu8 j"1 defined above. One can modify the Bratteli diagram in various ways to construct new RFD just-infinite AF-algebras with other characteristic sequences. For example, one can delete the first n ´ 1 rows and let row n correspond to an arbitrary finite dimensional C -algebra with n summands. (The remaining finite dimensional C -algebras are then determined by the one chosen and by the Bratteli diagram.) One is also allowed to change the multiplicity of the edges connecting the vertex at position pn, kq, 1 ď k ď n, to the vertex at position pn ` 1, n ` 1q. In these examples, the characteristic sequences all grow exponentially. By Proposition 3.18, we know that they must tend to infinity. This leaves open the following: Question 4.2. What are the possible characteristic sequences tkj u8 C -algebras? j"1 of RFD just-infinite 15 4.2 The dimension group We compute the dimension group pK0pAq, K0pAq`, r1sq of the just-infinite AF-algebra A constructed above (associated with the Bratteli diagram (4.1)). Recall that the dimension group, pH, H `, vq, associated with the Bratteli diagram (4.1) is the inductive limit of the ordered abelian groups αnpx1, x2, . . . , xnq " px1, x2, . . . , xn, x1 ` x2 ` ` xnq, px1, . . . , xnq P Zn, Z α1 / Z2 α2 / Z3 α3 / , where v P H ` is the image of 1 in the first copy of Z. It follows from standard theory of AF-algebras that pK0pAq, K0pAq`, r1sq is isomorphic to pH, H `, vq. We proceed to identify the latter more explicitly. Let śjPN Z denote the (uncountable) group of all sequences x " txju8 j"1 of integers, equipped with the usual order: x ě 0 if and only if xj ě 0, for all j ě 1. Let G be the countable subgroup of śjPN j"1 for which the identity xj`1 " x1 `x2 ` `xj holds eventually, and equip G with the order inherited from śjPN Z. Set u " p1, 1, 2, 4, 8, q. We show below that pH, H `, vq – pG, G`, uq. In conclusion, Z consisting of those sequences txju8 pK0pAq, K0pAq`, r1sq – pH, H `, vq – pG, G`, uq. For this, define ρn : Zn Ñ G by ρnpx1, x2, . . . , xnq " px1, x2, . . . , xn, xn`1, xn`2, . . . q, where xj`1 " x1 ` x2 ` ` xj, for all j ě n. Then ρn`1 αn " ρn, for all n, and each ρn is positive. It follows that the ρn's extend to a positive group homomorphism ρ : H Ñ G. Each ρn is injective, so ρ is injective. To complete the proof that ρ is an order isomorphism, we show that ρpH `q " G`. Take j"1 P G`, and let n ě 1 be such that xj`1 " x1 ` x2 ` ` xj, for all j ě n. Then x " txju8 x " ρnpx1, x2, . . . , xnq " x1ρnpepnq " x1ρpf pnq 1 2 q ` ` xnρnpepnq n q q ` ` xnρpf pnq n q, 1 q ` x2ρnpepnq q ` x2ρpf pnq 2 1 , epnq 2 , . . . , epnq where epnq are the corresponding images in H ` Ď H. This shows that x P ρpH `q. Finally, ρpvq " ρ1p1q " u, as wanted. n is the standard basis for pZnq` Ď Zn, and f pnq , . . . , f pnq , f pnq n 1 2 Unital AF-algebras are completely classified by their ordered K0-group, together with the position of the class of the unit. It is therefore an interesting question to classify, or charac- terize, those dimension groups which are the K0-group of a RFD just-infinite AF-algebra. In the light of the computation above, one may first wish to consider those dimension Z. In addition, one should assume that G Z, for each finite subset Z. The dimension group groups G which are (ordered) subgroups of ś8 is a subdirect product of ś8 F of N, where ϕF is the canonical projection of ś8 Z, in the sense that ϕF pGq " śjPF Z onto śjPF considered above has this property. j"1 j"1 j"1 16 / / / 4.3 A strictly RFD C -algebra which is not just-infinite It was shown in Corollary 3.11 that all RFD just-infinite C -algebras are strictly RFD. We show here that the converse does not hold, by constructing an example of a unital AF-algebra which is strictly RFD and not just-infinite. Let us first describe the example at the level of its primitive ideal space. Let X be the disjoint union of two copies of Y8, i.e., X " X1 > X2, where X1 " X2 " Y8. Equip X with the following topology: A non-empty subset U of X is open if and only if U X X1 is non-empty and open, and U X X2 is open. That this indeed defines a topology on X follows from the fact that the intersection of any two non-empty open subsets of X1 is non-empty, or, equivalently, that the set X1 is prime. Observe that X2 is an infinite closed subset of X. Hence X2 is a non-dense infinite subset of X. This shows that X cannot be the primitive ideal space of a just-infinite C -algebra; cf. Lemma 3.9. The set X1, on the other hand, is an open and dense subset of X, and each infinite subset of X1 is dense in X1, and therefore also dense in X. The space X is the primitive ideal space of the unital AF-algebra B whose Bratteli diagram is given as follows (ignoring at first the coloring of the vertices and edges): ‚ ‚ ‚ ‚ ‚ ... ‚ ❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜❜ ❆❆❆❆❆❆❆ ⑥⑥⑥⑥⑥⑥⑥ ❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝❝ PPPPPPPPPPPPPP ♥♥♥♥♥♥♥♥♥♥♥♥♥♥ ❆❆❆❆❆❆❆ ⑥⑥⑥⑥⑥⑥⑥ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ ❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡ PPPPPPPPPPPPPP ✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐ ♥♥♥♥♥♥♥♥♥♥♥♥♥♥ ❆❆❆❆❆❆❆ ⑥⑥⑥⑥⑥⑥⑥ ❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲❲ ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ ❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣ PPPPPPPPPPPPPP ✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐ ❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥ ♥♥♥♥♥♥♥♥♥♥♥♥♥♥ ❆❆❆❆❆❆❆ ⑥⑥⑥⑥⑥⑥⑥ ‚ ‚ ‚ ‚ ‚ ‚ ‚ ... ‚ ‚ ... ‚ ‚ ... ‚ ‚ ... ‚ ‚ ... ‚ ‚ ... ‚ ... ‚ ... ‚ ‚ ‚ ‚ ‚ ... The left-hand half of this Bratteli diagram is an essential ideal in the Bratteli diagram4 and therefore corresponds to an essential ideal I of the AF-algebra B. The right-hand half is the Bratteli diagram of the quotient B{I. Hence B{I is equal to the RFD just-infinite AF-algebra A described in Section 4.1, and I is Morita equivalent to A. Hence B cannot be just-infinite. For each k ě 1, let Uk be largest ideal of the Bratteli diagram which does not contain any vertex from the kth column of the left-hand half of the Bratteli diagram. Furthermore, let Ik be the ideal of B corresponding to the ideal Uk. To illustrate this definition, in the diagram above, the ideal U3 is marked in blue and the Bratteli diagram of the quotient B{I3 is marked in red The quotient B{I3 is seen to be isomorphic to M4pCq. In general, for each k ě 1, we see that B{Ik is a full matrix algebra, and (hence) that this, observe that Uk contains no vertices from the top k ´ 1 rows of the left-hand half of the each Ik is a primitive ideal. Moreover, ŞkPT Ik " 0, for each infinite subset T of N. (To see Bratteli diagram, or from the top k ´ 2 rows of the right-hand half. Hence ŞkPT Uk " H, for each infinite subset T of N.) This shows that B is a strictly RFD AF-algebra which is not just-infinite. 4An ideal U in a Bratteli diagram is said to be essential, if U X V ‰ H, for all non-empty ideals V . 17 5 Subalgebras and superalgebras In this section, which is addressed to specialists in C -algebras, we investigate when subalge- bras and superalgebras of just-infinite C -algebras are again just-infinite, and we also show that not all RFD just-infinite C -algebras are nuclear, or even exact. The third named author thanks Jose Carrion for his suggestion to use Theorem 5.3 below of Dadarlat to conclude that there are non-nuclear, and even non-exact, RFD just-infinite C -algebras. Recall that a C -algebra A has real rank zero if each self-adjoint element in A is the norm limit of self-adjoint elements in A with finite spectra. A commutative C -algebra CpXq has real rank zero if and only if X is totally disconnected (or, equivalently, dimpXq " 0). Real rank zero is therefore viewed as a non-commutative analog of being zero-dimensional. A C -algebra has real rank zero if it has "sufficiently many projections". Each closed two- sided ideal of a C -algebra of real rank zero again has real rank zero and, as a consequence, is generated by its projections. We denote by IdealpAq the lattice of closed two-sided ideals in A. If B is a sub-C -alge- bra of A, then there is a natural map Φ : IdealpAq Ñ IdealpBq, given by ΦpIq " I X B. The map Φ is, in general, neither injective nor surjective, but it is both in the special situation of the lemma below. We use the symbol p „A q to denote that p and q are Murray-von Neumann equivalent projections, relatively to the C -algebra A. Lemma 5.1. Let B Ď A be unital C -algebras of real rank zero, and suppose that there is a -homomorphism κ : A Ñ B such that κppq „A p, for all projections p P A, and κpqq „B q, for all projections q P B. Then the map Φ : IdealpAq Ñ IdealpBq is a lattice isomorphism. Proof. We first show that Φ is injective. Let I ‰ I 1 P IdealpAq be given. Since A has real rank zero, and ideals in A are generated by their projections, there exists a projection p P I such that p R I 1 (or vice versa). Set q " κppq „ p. Then q P I X B " ΦpIq, but q R I 1 X B " ΦpI 1q. Hence ΦpIq ‰ ΦpI 1q. Let now J P IdealpBq be given, and let I " AJ A be the closed two-sided ideal in A generated by J. Then, clearly, J Ď I X B " ΦpIq. To see that ΦpIq Ď J, it suffices to show that each projection q in ΦpIq belongs to J. Being a projection in I, q belongs to the j"1 aj xjbj for some aj, bj P A and xj P J. The conditions on κ, together with the fact that B is a C -algebra of real rank zero, imply that κ maps J into itself, so algebraic two-sided ideal in A generated by J, so q " řn q „B κpqq " n ÿj"1 κpajqκpxjqκpbj q P J. This shows that q belongs to J, as desired. Lemma 5.2. Let A be a unital separable RFD just-infinite C -algebra of real rank zero, and let tπnu8 n"1 be an exhausting5 sequence of pairwise inequivalent non-faithful irreducible representations of A. (i) Suppose that B is a unital sub-C -algebra of A such that the map Φ : IdealpAq Ñ IdealpBq is an isomorphism, and such that each projection in A is equivalent to a n"1 is an projection in B. exhausting sequence of pairwise inequivalent non-faithful irreducible representations of B, and that πnpBq " πnpAq, for all n. In particular, A and B have the same characteristic sequence. It follows that B is just-infinite and RFD, that tπnBu8 5See Remark 3.15. 18 (ii) Suppose that C is a unital C -algebra of real rank zero which contains A and is asymp- totically homotopy equivalent6 to A. Suppose also that Φ : IdealpCq Ñ IdealpAq is an isomorphism.7 It follows that C is just-infinite and RFD with an exhausting se- quence tνnu8 n"1 of pairwise inequivalent non-faithful irreducible representations for which KerpνnAq " Kerpπnq and νnpCq – πnpAq, for all n. In particular, A and C have the same characteristic sequence. Proof. (i). The lattice isomorphism Φ : IdealpAq Ñ IdealpBq restricts to a homeomorphism PrimpAq Ñ PrimpBq, and so PrimpBq is homeomorphic to PrimpAq, which again is homeo- morphic to Y8. Moreover, PrimpBqzt0u " tKerpπnq X B n P Nu " tKerpπnBq n P Nu. Let I be a non-zero primitive ideal of B. Then I " KerpπnBq, for some n, and B{I is isomorphic to πnpBq, which is a subalgebra of the finite dimensional C -algebra πnpAq, so B{I is finite dimensional. It now follows from Theorem 3.10 that B is just-infinite. Let us also show that πnpBq " πnpAq, for all n. Since πnpBq Ď πnpAq and both C -al- gebras are full matrix algebras, it suffices to show that πnpBq contains a minimal projection in πnpAq. Let e P πnpAq be such a projection and lift it to a projection p P A (which is possible because A is assumed to have real rank zero). Find a projection q P B which is equivalent to p. Then πnpqq is equivalent to e, which implies that πnpqq itself is a minimal projection in πnpAq. (ii). As in (i), the given lattice isomorphism Φ : IdealpCq Ñ IdealpAq restricts to a homeomorphism PrimpCq Ñ PrimpAq, so PrimpCq is homeomorphic to Y8. Given n ě 1, let Jn P PrimpCq be such that ΦpJnq " Kerpπnq. Since Φ is an isomorphism, each non-zero primitive ideal in C is of this form. Identify πnpAq with MkpCq, for some positive integer k. Find an irreducible representation νn : C Ñ BpHq on some Hilbert space H, with Kerpνnq " Jn. Then Kerpπnq " ΦpJnq " Jn X A " KerpνnAq. Let ι : MkpCq Ñ BpHq be the inclusion mapping making the following diagram commutative: A πn C νn MkpCq ι / BpHq We show that dimpHq " k, which by Theorem 3.10, will imply that C is just-infinite. It will also imply that ι is an isomorphism, and that πnpAq – νnpAq " νnpCq " BpHq. It is clear that dimpHq ě k. Suppose that dimpHq ą k. Then we can find pairwise orthogonal non-zero projections f1, f2, . . . , fk`1 in νnpCq. (Indeed, νnpCq acts irreducibly on H, so if dimpHq is finite, then νnpCq " BpHq. If dimpHq is infinite, then νnpCq is infinite dimensional and of real rank zero. In either case, one can find the desired pro- jections.) Since C has real rank zero, we can lift the projections f1, f2, . . . , fk`1 to mutu- ally orthogonal projections p1, p2, . . . , pk`1 in C. Applying the asymptotic homomorphism C Ñ A to the projections p1, p2, . . . , pk`1, and using that Ck`1 is semiprojective (see [8]), we obtain mutually orthogonal projections q1, q2, . . . , qk`1 in A. Since the asymptotic ho- momorphism C Ñ A composed with the inclusion mapping A Ñ C is homotopic to the 6This means that there exists an asymptotic morphism C Ñ A, so that the asympotic morphism C Ñ C (obtained by composing it with the inclusion mapping A Ñ C) is homotopic to the identity on C in the category of asymptotic morphism. See also [14]. 7In fact, the assumptions on A and C imply that Φ is an isomorphism. This can be shown along the same lines as the proof of Lemma 5.1. 19 / / / identity mapping on C, we further get that qj is equivalent to pj, for each j. In particular, pι πnqpqjq " νnpqjq „ νnppjq " fj, for each j, so πnpqjq is non-zero. But MkpCq does not contain k ` 1 mutually orthogonal non-zero projections. This proves that dimpHq " k. We shall combine Lemma 5.2 with the following results due to Dadarlat: Theorem 5.3 (Dadarlat, [14, Theorem 11 and Proposition 9]). Let A be a unital AF-algebra not of type I. Then: (i) A contains a unital non-nuclear sub-C -algebra B of real rank zero and stable rank one, for which there exists a unital -monomorphism κ : A Ñ B such that ι κ is homotopic to idA and κ ι is asymptotically homotopic to idB, where ι is the inclusion mapping B Ñ A. Moreover, Φ : IdealpAq Ñ IdealpBq is an isomorphism of lattices. (ii) A is contained in a unital separable non-exact C -algebra C of real rank zero and stable rank one, which is asymptotically homotopy equivalent to A, and for which Φ : IdealpCq Ñ IdealpAq is an isomorphism of lattices. The statements (i) and (ii) that Φ is an isomorphism between the ideal lattices of A and B, respectively, of C and A, are included in the quoted results of Dadarlat, and it also follows from Lemma 5.1 in the situation considered in (i). To apply Theorem 5.3, we need the following: Lemma 5.4. A separable just-infinite C -algebra is of type I if and only if either it is isomorphic to K, the compact operators on a separable infinite dimensional Hilbert space, or it is an essential extension of K by a finite dimensional C -algebra. In the former case, A is just-infinite of type (α), and in the latter case A is just-infinite of type (β); cf. Theorem 3.10. In particular, no just-infinite C -algebra of type (γ), i.e., RFD, is of type I. Proof. Let A be a separable just-infinite C -algebra of type I. By Lemma 3.2, A is prime, and hence primitive, so we can find a faithful irreducible representation π of A on a separable, necessarily infinite dimensional, Hilbert space H. Since A is a C -algebra of type I, the algebra K of compact operators on H is contained in the image of π. Hence I " π´1pKq is a non-zero closed two-sided ideal in A, which is isomorphic to K. As A is just-infinite, either I " A, or A{I is finite dimensional. Corollary 5.5. Let A be a unital separable RFD just-infinite AF-algebra, and let tπnu8 n"1 be an exhausting sequence of pairwise inequivalent non-faithful irreducible representations of A. It follows that A contains a unital non-nuclear RFD just-infinite sub-C -algebra B of real rank zero such that tπnBu8 n"1 is an exhausting sequence of pairwise inequivalent non-faithful irreducible representations of B, and πnpBq " πnpAq, for all n. Proof. By Lemma 5.4, we can now apply Theorem 5.3 (i) to find a sub-C -algebra B of A with the properties listed therein. Each projection p P A is equivalent to a projection in B. Indeed, set q " κppq P B. Then q " pι κqppq is homotopic (and hence equivalent) to p. The desired conclusion now follows from Lemma 5.2 (i). Corollary 5.6. Let A be a unital separable RFD just-infinite AF-algebra. Then A is con- tained in a separable non-exact unital RFD just-infinite C -algebra C of real rank zero, equipped with an exhausting sequence of pairwise inequivalent non-faithful irreducible repre- sentations tνnu8 n"1, such that their restrictions to A form an exhausting sequence of pairwise inequivalent non-faithful irreducible representations for A, and νnpAq " νnpCq, for all n. 20 Proof. This follows immediately from Lemma 5.4, Theorem 5.3 (ii) and Lemma 5.2 (ii). The two above corollaries, in combination with the existence of a RFD just-infinite AF- algebra (see Section 4), now yield the following: Corollary 5.7. There exist non-nuclear exact RFD just-infinite C -algebras, and there also exist non-exact RFD just-infinite C -algebras. It is shown in [32, Theorem 4.3] that each unital C -algebra A of real rank zero contains a unital AF-algebra B such that each projection in A is equivalent to a projection in B, and such that Φ : IdealpAq Ñ IdealpBq is an isomorphism. Together with Lemma 5.2 (i), this proves the following: Proposition 5.8. Let A be a unital separable RFD just-infinite C -algebra of real rank zero, and let tπnu8 n"1 be an exhausting sequence of pairwise inequivalent non-faithful irreducible representations of A. It follows that A contains a unital RFD just-infinite AF-sub-C -alge- bra B such that tπnBu8 n"1 is an exhausting sequence of pairwise inequivalent non-faithful irreducible representations of B, and πnpBq " πnpAq, for all n. By combining Corollary 5.5 with Proposition 5.8, one obtains the following fact: Suppose that A is a unital separable RFD just-infinite C -algebra of real rank zero, and tπnu8 n"1 is an exhausting sequence of pairwise inequivalent non-faithful irreducible representations. Then there is a strictly decreasing sequence A Ą A1 Ą A2 Ą A3 Ą of unital sub-C -algebras Ak of A such that each Ak is a RFD just-infinite C -algebra, and πnpAkq " πnpAq, for all k and n. (In fact, every other C -algebra in the sequence tAku can be taken to be an AF-algebra and the remaining ones to be non-nuclear.) In particular, a unital separable RFD just-infinite C -algebra of real rank zero can never be minimal in the sense that it contains no proper RFD just-infinite sub-C -algebras. 6 Just-infiniteness of group C -algebras We discuss in this section when C -algebras associated with groups are just-infinite. The group algebra CrGs of a group G is in a natural way a -algebra in such a way that each group element g P G becomes a unitary in CrGs, and it can be completed to become a C -algebra, usually in many ways. The universal C -algebra of G, denoted by C pGq, is the completion of CrGs with respect to the maximal C -norm on CrGs. Each unitary rep- resentation π of the group G on a Hilbert space gives rise to unital -representations (again denoted by π) of the -algebras CrGs and C pGq on the same Hilbert space. Respectively, each unital -representation π of C pGq restricts to a -representation of CrGs, and if this restriction is faithful, then it creates a C -norm } }π on this algebra. Each C -norm on CrGs arises in this way, where by a C -norm on CrGs we mean a (faithful) norm such that the completion of CrGs with respect to this norm is a C -algebra. Given a unitary representation π of G, we let C π pGq denote the completion of πpCrGsq. This is equal to the completion of CrGs with respect to the norm } }π, if π is faithful on CrGs. The reduced group C -algebra of G, denoted by C λpGq, arises in this way from the left-regular representation λ of G on ℓ2pGq. It is well-known that the maximal and the reduced C -norms on CrGs are equal, i.e., C pGq " C λ pGq, if and only if G is amenable (see [13, Theorem 2.6.8]). It is also well-known (see, e.g., [13, Exercise 6.3.3]) that if the reduced group C -algebra C λ pGq has a finite-dimensional representation, then G must be amenable. Hence the following holds: 21 Proposition 6.1. Let G be a group and suppose that C C λ pGq is simple, or G is amenable. λ pGq is just-infinite. Then either Whereas CrGs always has one maximal C -norm, there may or may not be a minimal C - λ pGq is a simple norm on CrGs, depending on the group G. C -algebra, then the norm } }λ on CrGs is minimal. If G is C -simple, i.e., if C Proposition 6.2. Let G be a group, and let π be a representation of G which gives a faithful representation of CrGs. If C π pGq is just-infinite, then } }π is a minimal C -norm on CrGs. Proof. Any C -norm on CrGs which is smaller than } }π arises from a unitary representation ν of G on a Hilbert space, which factors through C π pGq. Since ν is injective on CrGs, the image νpC π pGqq cannot be finite dimensional, so ν is injective, and hence isometric, on π pGq. (Recall that each injective -homomorphism between C -algebras automatically is C isometric.) The norm arising from ν is therefore equal to the norm arising from π. If G is infinite and if C then C π pGq is simple, for some unitary representation π of the group G, π pGq is just-infinite and } }π is a minimal norm on CrGs; cf. Proposition 6.2. The group algebra CrGs is said to be -just-infinite if each -representation of CrGs either is injective, or has finite dimensional image. Note that -just-infinite is a formally weaker condition than "just-infinite", as CrGs can have non-self-adjoint two-sided ideals. Proposition 6.3. Let G be an infinite group. Then C pGq is just-infinite if and only if CrGs is -just-infinite and CrGs has a unique C -norm. Proof. Suppose first that C pGq is just-infinite. Let π be a unital -representation of CrGs, and extend it to a -representation of C pGq. Then π is either injective on C pGq, or πpC pGqq is finite dimensional. If π is injective on C pGq, then it is also injective on CrGs, while if πpC pGqq is finite dimensional, then so is πpCrGsq. Hence CrGs is -just-infinite. Each C -norm on CrGs arises as } }π, for some -representation π of C pGq which is faithful on CrGs. Thus πpC pGqq is infinite dimensional, so π must be injective on C pGq. This entails that } }π is the maximal norm on CrGs, and thus the only C -norm on CrGs. Suppose now that CrGs has a unique C -norm, and that CrGs is -just-infinite. Let π be a non-faithful unital -representation of C pGq. If the restriction of π to CrGs were faithful, then it would induce a C -norm on CrGs, which by uniqueness would be equal to the maximal C -norm on CrGs. This contradicts that π is non-faithful on C pGq. Hence π is not faithful on CrGs, whence πpCrGsq is finite dimensional. In this case, πpC pGqq is equal to πpCrGsq. This proves that C pGq is just-infinite. Corollary 6.4. Let G be a group for which C pGq is just-infinite. Then G is amenable, and hence C pGq " C λpGq is nuclear. Proof. It follows from Proposition 6.3 that the reduced and the maximal norm on CrGs coincide, so G is amenable. Corollary 6.5. For each group G, if C pGq is just-infinite, then CrGs is -just-infinite, which in turn implies that G is just-infinite. Proof. The first implication follows from Proposition 6.3. To see that the second implication holds, suppose that CrGs is -just-infinite, and let N be a non-trivial normal subgroup of G. The quotient map G Ñ G{N lifts to a necessarily non-injective -homomorphism CrGs Ñ CrG{N s. Hence CrG{N s must be finite dimensional, whence G{N is finite. 22 None of the reverse implications above hold; cf. Examples 6.6 and 7.3. Example 6.6. The group algebra CrZs is -just-infinite, and the group Z is just-infinite; but C pZq is not just-infinite, and CrGs has no minimal C -norm. Proof. Each unital -representation π of CrZs on a Hilbert space H admits a natural fac- torization CrZs Ñ CpKq Ñ BpHq, where K Ď T is the spectrum of the unitary operator u " πp1q, and where CpKq Ñ BpHq is injective. It is easy to see that π is faithful on CrZs if and only if K is an infinite set. If π is not faithful, then K is finite, which entails that πpCrZsq is finite dimensional. This shows that CrZs is -just-infinite. As there is no minimal closed infinite subset of T, there is no minimal C -norm on CrZs, and we conclude from Proposition 6.3 that C pZq is not just-infinite. This conclusion also follows from Example 3.3. Proposition 6.7. If G is a locally finite group, then CrGs has a unique C -norm. Proof. Each element x P CrGs is a linear combination of finitely many elements from G, and each finitely generated subgroup of G is finite, by assumption. Hence there is a finite subgroup H of G such that x P CrHs Ď CrGs. Now, CrHs is a (finite dimensional) C -al- gebra, so it has a unique C -norm. Thus any two C -norms on CrGs must agree on x. As x was arbitrarily chosen, we conclude that CrGs has a unique C -norm. Question 6.8. Let G be a group and suppose that CrGs has a unique C -norm. Does it follow that G is locally finite? The augmentation ideal of the full group C -algebra C pGq of a group G is the kernel of the trivial representation C pGq Ñ C. If G is infinite and if the augmentation ideal is simple, or, more generally, just-infinite, then C pGq is just-infinite by Proposition 3.5 (ii), since the augmentation ideal always is essential when G is infinite. There are locally finite groups whose augmentation ideal is simple, such as Hall's univer- sal groups, see [9] and [27]. It follows from Theorem 3.10 that C pGq is just-infinite of type (β), for any such group G. It is easy to see that if an amenable group G has simple augmen- tation ideal, then it must be simple; however, simple groups (even locally finite ones) need not have simple augmentation ideal: the infinite alternating group A8 is a counterexample. Lemma 6.9. Let G be a residually finite group for which C pGq is just-infinite. Then C pGq is RFD (and hence of type (γ); cf. Theorem 3.10). Proof. Let tNiuiPI be a decreasing net of finite index normal subgroups of G with ŞiPI Ni " teu, and consider the -homomorphism Φ : C pGq Ñ źiPI C pG{Niq. It suffices to show that Φ is injective; and by the assumption that C pGq is just-infinite, it further suffices to show that the image of Φ is infinite dimensional. The latter follows from the fact that G is infinite (as C pGq is just-infinite) and (hence) that supiPI G : Ni " 8. Question 6.10. Does there exist an infinite, residually finite group G such that C pGq is just-infinite? If such a group G exists, then C pGq will be a RFD just-infinite C -algebra by Lemma 6.9. If the answer to Question 6.8 is affirmative, then G must be locally finite. This leads to the following: 23 Question 6.11. Does there exist an infinite, residually finite, locally finite (necessarily just-infinite) group G such that CrGs is -just-infinite? If such a group G exists, then C pGq will be a RFD just-infinite C -algebra by Lemma 6.9, Proposition 6.7 and Proposition 6.3. After the first version of this paper was made public, Question 6.11 has been answered in the affirmative in [7]. Just-infinite groups are divided into three disjoint subclasses (the trichotomy for just- infinite groups), see [22, Section 6]: The non-residually finite ones (which contain a finite index normal subgroup N which is the product of finitely many copies of a simple group), branch groups (see more about those in Theorem 7.10 below), and the hereditarily just- infinite groups, i.e., the residually finite groups for which all finite index normal subgroups are just-infinite. It is shown in Theorem 7.10 below that if G is a just-infinite branch group, then CrGs is not -just-infinite, whence C pGq is not just-infinite. Hence, if there exists a residually finite group G for which C pGq is just-infinite (and hence also RFD), then G must be hereditarily just-infinite. Consider the following three (classes of) examples of hereditarily just-infinite groups: the integers Z, the infinite dihedral group D8, and PSLnpZq, for n ě 3. As shown below, if G is any of these groups, then C pGq is not just-infinite. Moreover, there is no unitary representation π of G such that C π pGq is RFD and just-infinite. If G " Z, then this claim follows immediately from Example 3.3. In the two examples below we discuss the situation for the two other (classes of) groups. Example 6.12 (PSLnpZq, n ě 3). The groups PSLnpZq, n ě 3, are renowned for being the first examples of infinite groups with Kazdan's property (T), as first shown by Kazdan. For a different and nice proof by Shalom, see [34]. They are residually finite, as witnessed by the finite quotient groups PSLnpZ{N Zq, N P N; and they are hereditarily just-infinite by Margulis' normal subgroup theorem. Bekka–Cowling–de la Harpe proved in [4] that PSLnpZq is C -simple, for all n ě 2. In particular, PSLnpZq is an ICC group (all its non-trivial conjugacy classes are infinite). We conclude from these facts that the C -algebra C λ pPSLnpZqq is just-infinite (being simple) for all n ě 2, while the full group C -algebra C pPSLnpZqq is not just-infinite, because PSLnpZq is non-amenable, for n ě 2. Bekka proved in [3] that the set of extremal characters on PSLnpZq, for n ě 3, is a countably infinite set consisting of the trivial character δe and a sequence tδku8 k"1 of characters, each of which factors through a finite quotient, PSLnpZ{N Zq, of PSLnpZq for a suitable integer N (depending on k). Recall that each (extremal) character on a group corresponds to an extremal trace on its full group C -algebra. The trivial character δe on PSLnpZq corresponds to the canonical trace τ0 on C pPSLnpZqq; while for k ě 1, the character δk corresponds to a trace, denoted by τk, whose GNS-representation πτk is finite dimensional. Bekka also shows that τk Ñ τ0 in the weak topology. Furthermore, observe that C pPSLnpZqq has a just-infinite quotient, namely the simple λ pPSLnpZqq. However, as shown below, there is no unitary representation π C -algebra C of PSLnpZq such that C π pPSLnpZqq is RFD and just-infinite. Indeed, assume that π is such a unitary representation of PSLnpZq. As in Remark 3.15, let tπju8 j"1 be an exhausting sequence of pairwise inequivalent non-faithful irreducible rep- resentations of C π pPSLnpZqq. Then ρj " πj π, j ě 1, is a sequence of pairwise inequivalent non-faithful irreducible representations of C pPSLnpZqq. Hence ρj is equivalent to πτkpjq, for some kpjq ě 1, by the above mentioned result of Bekka. Suppose now that x P C pPSLnpZqq belongs to the kernel of π. Then, for all j ě 1, πτkpjqpxq " 0, so τkpjqpxxq " 0. It follows that τ0pxxq " limjÑ8 τkpjqpxxq " 0, so λpxq " 0. This shows that λ is weakly contained 24 in π. We conclude that the left-regular representation λ factors through π, so the simple C -algebra C π pPSLnpZqq. Each simple quotient of a RFD just-infinite C -algebra is finite dimensional, so C π pPSLnpZqq cannot both be RFD and just-infinite. λpPSLnpZqq is a quotient of C Example 6.13 (The infinite dihedral group D8). The infinite dihedral group D8 is an example of a hereditarily residually finite just-infinite group, see [22], and it is isomorphic to the free product Z2 Z2, which is an amenable group (of linear growth). The group C - algebra C pZ2 Z2q is known to be a sub-C -algebra of M2pCpr0, 1sqq (being the universal unital C -algebra generated by two projections), and is hence subhomogeneous (cf. Proposi- tion 3.16). Clearly, any quotient of a subhomogeneous C -algebra is again subhomogeneous, so we conclude from Lemma 3.17 that C pZ2 Z2q is not just-infinite, and neither is any of its quotients. 7 Algebras associated with groups of intermediate growth In this section we present some results concerning algebras associated with the 3-generated infinite torsion group constructed in [19], which we here will denote by G. This group is a simple example of a group of Burnside-type, and it is investigated more deeply in [20] and many other papers (see the surveys [23], [24], and the references therein). Among its unusual properties, most notably G is of intermediate growth (between polynomial and exponential), and, as a consequence, it is amenable, but not elementary amenable, thus answering questions by Milnor and Day, respectively; cf. [20]. Furthermore, G is a just- infinite group of branch type (and hence residually finite), and moreover, it is a self-similar group (i.e., a group generated by states of a Mealy-type automaton). There are indications that various C -algebras associated with G (including C pGq and some of its quotients, discussed below) may be new types of C -algebras with properties unseen yet in the theory of operator algebras. Our main conjecture in this direction is the following: Conjecture 7.1. The (self-similar RFD) C -algebra C representation π of G is just-infinite. π pGq generated by the Koopman The Koopman representation π of G will be described below, along with the notion of self- π pGq is a just-infinite C -algebra of similarity. type (γ) as described in Theorem 3.10; cf. Lemma 6.9, and it is the first example of such a C -algebra associated with a group. If the conjecture above is correct, then C Recall that the group G is generated by four elements a, b, c, d satisfying the relations 1 " a2 " b2 " c2 " d2 " bcd " σk`padq4 " σk`padacacq4, for k " 0, 1, 2, . . . , where the permutation σ is given by the substitutions: (7.1) a Ñ aca, b Ñ d, c Ñ b, d Ñ c. This presentation was found by I. Lysenok in [28], and it is a minimal presentation (in the sense that no relator in (7.1) can be deleted without changing the group, see [21]). In fact, G is generated by 3 elements, as d " bc. The set t1, b, c, du is the Klein group Z{2Z ' Z{2Z. For our purposes it will be most important to know that G has a faithful self-similar action by automorphisms on an infinite binary rooted tree T , as shown, in part, here: 25 ‚ rrrrrrrrrrrr ✿✿✿✿✿✿✿ ✑✑✑✑✑✑ ✲✲✲✲✲✲ ‚ ‚ ‚ v0 ‚ ✴✴✴✴✴✴ ‚ ‚ ✑✑✑✑✑✑ ‚ ▲▲▲▲▲▲▲▲▲▲▲▲ ☎☎☎☎☎☎☎ ✲✲✲✲✲✲ ✑✑✑✑✑✑ ‚ ‚ ‚ ‚ v1 ❃❃❃❃❃❃❃ ✎✎✎✎✎✎ ‚ ‚ ✲✲✲✲✲✲ ‚ (7.2) The generators a, b, c, d act on T as follows: The root of the tree (marked in red) is a common fixed point. The generator a just permutes the two vertices v0 and v1 at the first level and acts trivially inside the subtrees T0 and T1 with roots v0 and v1, respectively. The generators b, c, d fix the vertices v0 and v1 (and hence leave the subtrees T0 and T1 invariant), and they are defined recursively by: bT0 " a, bT1 " c, cT0 " a, cT1 " d, dT0 " 1, dT1 " b, (7.3) when identifying the subtrees T0 and T1 with T in the natural way, and where 1 stands for the identity automorphism. For more details on this definition, and other definitions of G, we refer to [19, 20, 23, 24]. The relations (7.3) imply that G is a self-similar group in the sense that it has a natural embedding ψ : G Ñ G ≀ pZ{2Zq – pG Gq ¸ pZ{2Zq, (7.4) where Z{2Z " te, εu acts on G G by permuting the two copies of G (e is the identity element and ε is a transposition). The embedding ψ is given as follows: ψpaq " p1, 1qε " ε, ψpbq " pa, cqe " pa, cq, ψpcq " pa, dqe " pa, dq, ψpdq " p1, bqe " p1, bq. To further illlustrate this action of G on the tree T it is convenient to label the vertices of the nth level, Vn, of T by the set t0, 1un and equip each Vn with the lexicographic ordering: ♠♠♠♠♠♠♠♠♠♠♠♠♠♠♠♠♠ ❈❈❈❈❈❈❈❈ ☛☛☛☛☛☛ ✸✸✸✸✸✸ 01 0 ④④④④④④④④ ✸✸✸✸✸✸ 001 00 ☛☛☛☛☛☛ 000 H ◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗ ④④④④④④④④ ✸✸✸✸✸✸ ☛☛☛☛☛☛ 10 (7.5) 1 ❈❈❈❈❈❈❈❈ ☛☛☛☛☛☛ 110 11 ✸✸✸✸✸✸ 111 010 011 100 101 The action of the group G on T yields an action of G by homeomorphisms on the boundary BT of T , which consists of geodesic rays joining the root H with infinity. The boundary BT can in a natural way be identified with the Cantor set t0, 1uN of infinite binary sequences equipped with the Tychonoff topology. n"1 t 1 Let µ " Ś8 2 u be the uniform Bernoulli measure on BT . It is invariant with respect to the action of the entire group AutpT q of automorphisms on T , and hence with respect to the action of G on T . The topological dynamical system pG, BT q can be converted into 2 , 1 26 a metric dynamical system pG, BT, µq which is ergodic (while pG, BT q is minimal), because the action of G on each level Vn is transitive, see [25, Proposition 6.5]. Let π be the (unitary) Koopman representation of G on the Hilbert space L2pBT, µq given by `πpgqfpxq " f pg´1xq, where f P L2pBT, µq, g P G, and x P BT . We denote the image of the group algebra CrGs under the representation π by B, and we let as usual C denote the completion of CrGs with respect to the norm induced by π. π pGq The following theorem carries some evidence in support of Conjecture 7.1. Theorem 7.2. Let G " xa, b, c, dy be the infinite torsion group of intermediate growth from above, let π be the Koopman representation of G, and let B " πpCrGsq. Then: (i) B is self-similar, infinite dimensional and RFD. (ii) C π pGq is self-similar, infinite dimensional, RFD, and it posseses a faithful trace. (iii) The natural surjection π : CrGs Ñ B is not injective, whence CrGs is not -just-infinite. (iv) B is just-infinite. The notions of self-similarity of the algebras B and C rem 7.2 above is proved at the end of this section. π pGq will be explained below. Theo- The type of just-infinite algebras (also called "thin algebras") considered above were studied for the first time by Sidki in [35]. The group used by Sidki was the Gupta–Sidki 3-group H, and the algebra was defined over a field F3 in a rather involved way as a certain inductive limit. If one considers the "Koopman" representation of H over the field F3, then the image of the group algebra F3rHs will be isomorphic to Sidki's thin algebra. The C -algebra generated by the Koopman representation of the group G (considered in this section) was considered in [2], and so was the algebra B, even though it was not explicitly defined there. Vieira, [36], used Sidki's approach to define a "thin algebra" of the group G over the field F2, and proved that it is just-infinite. Thin algebras under the name "Tree enveloping algebras" were considered by Bartholdi in [1]. He defines algebras, similar to the algebra B in Theorem 7.2, however, over arbitrary fields. He considers a vector space with a basis consisting of all points of the boundary of the rooted tree, and then defines an algebra as the image of the group algebra in the algebra of endomorphisms of this huge vector space. One can show that if the field is complex numbers and the group is the group G, then Bartholdi's algebra is isomorphic to the algebra B we are considering here. In [1, Theorem 3.9], a sufficient condition is given for the tree enveloping algebra to be just-infinite. This condition is satisfied in the case of the group G. Example 7.3. As mentioned above, the group G is just-infinite. We can therefore deduce from Theorem 7.2(iii) that just-infiniteness of a group G does not imply that its complex group algebra CrGs is -just-infinite. Self-similarity of graphs, Hilbert spaces, representations and algebras Let X " tx1, x2, . . . , xdu be an alphabet on d ě 2 letters, let X " Ů8 n"0 X n be the set of words over X, and let T " TX be the d-arnery rooted tree whose vertices are in bijection with the elements of X (so that the nth level Vn of T corresponds to X n). The action of an arbitrary group G on T by automorphisms induces an action G ñ X . This action is said to be self-similar if for all g P G and all x P X, there are h P G and y P X such that 27 gpxwq " yhpwq, for all w P X . If this holds, then for every v P X , there exists a unique h P G satisfying, for all w P X , gpvwq " gpvqhpwq. (7.6) The element h is called the section (or restriction) of g in v, and is denoted by h " gv. For example, for the group G " xa, b, c, dy under examination, we have; cf. (7.3), that av0 " av1 " 1, bv0 " a, bv1 " c, cv0 " a, cv1 " d, dv0 " 1, dv1 " b. Let H be a separable infinite dimensional Hilbert space, and fix an integer d ě 2. A unitary operator u : H Ñ H d " H ' H ' ' H is called a d-similarity of H. Each d-similarity arises from d isometries s1, . . . , sd on H, satisfying řd ÿj"1 upξ1, . . . , ξdq " upξq " ps 1 ξ, . . . , s d ξq, d sjξj, j"1 sjs j " 1, as follows: for ξ, ξ1, . . . , ξd P H. Observe that s1, . . . , sd define a representation of the Cuntz algebra Od, and that every representation of Od is obtained in this way. For each x " pxi1 , xi2 , . . . , xinq P X consider the isometry on H given by Sx " si1 si2 . . . sin, and observe that SxSy " Sxy. A unitary representation ρ of a group G on a Hilbert space H is said to be self-similar with respect to the d-similarity ψ considered in (7.4) above, if ρpgqSx " Syρphq, (7.7) for all g, h P G and all x, y P X satisfying gpxwq " yhpwq, for all w P X . In other words, ρpgqSx " Sgpxqρpgxq, for all g P G and x P X . The image Bρ " ρpCrGsq, where ρ is a self-similar representation, is called a self-similar (abstract) algebra. The C -algebra C ρ pGq associated with a self-similar representation ρ is called a self-similar C -algebra. One of the features of the self-similar algebra Bρ (or of the C -algebra C ρ pGq) is the existence of the unital embedding ψρ : Bρ Ñ MdpBρq, b ÞÑ s 1 bs1 ... s dbs1 s 1bsd ... s dbsd , ‹‚ (7.8) It follows from (7.7) that s for b P Bρ. j Bρsi Ď Bρ, for all i, j. The embedding ψρ is typically not surjective. Nonetheless, it has many interesting and non-trivial features, see, for example, Lemma 7.6 below. In the case of our main example G " xa, b, c, dy and of the Koopman representation π of G on H " L2pBT, µq, we have an explicit self-similarity H Ñ H ' H arising from the two isometries s0, s1 on H defined by psif qpxq " f pixq, (7.9) for i " 1, 2, where f P L2pBT, µq and x P BT , and where ix P BT " t0, 1uN is the word obtained by putting the letter i in front of the word x. The resulting embedding ψπ : B Ñ M2pBq is given as follows on the generators: ψπp¯aq " 0 1 1 0 , ψπp¯bq " ¯a 0 0 ¯c , ψπp¯cq " ¯a 0 ¯d , ψπp ¯dq " 1 0 0 ¯b , 0 (7.10) (as can be deduced from (7.7) and (7.8)), where we have introduced the notation ¯g " πpgq, for g P G. (The Koopman representation is faithful on G, so the map g ÞÑ ¯g is injective, but the Koopman representation is not faithful on CrGs; cf. Theorem 7.2, so it is pertinent to distinguish between g and πpgq.) 28 More on the Koopman representation What we are going to present here is known in the more general situation of groups acting on rooted trees, [2, 5, 23]. Consider the binary rooted tree T (as described in (7.2) and (7.3)), and the Koopman representation π of the group G " xa, b, c, dy on L2pBT, µq. For each n ě 1, let vn,1, vn,2, . . . , vn,2n be the order preserving enumeration of the set i"1 En,i, Vn " t0, 1un (equipped with the lexicographic ordering); cf. (7.5), and write BT " Ů2n where En,i is the set of infinite words in BT " t0, 1uN that start with vn,i. Set Hn " spantχEn,i i " 1, 2, . . . , 2nu Ď H " L2pBT, µq, which is a subspace of dimension 2n. Since En,i " En`1,2i´1 Y En`1,2i, we see that Hn Ď Hn`1. Moreover, as the cylinder sets En,i, n ě 1, 1 ď i ď 2n, form a basis for the topology on BT , it follows that Ť8 n"1 Hn is dense in H. The subspaces Hn are π-invariant. Let πn be the restriction of π to Hn, for n ě 1. Observe that πn is unitarily equivalent to the representation of G on ℓ2pVnq arising from its action on the nth level Vn of the tree T . More specifically, identify Hn with ℓ2pVnq via the isomorphism that identifies χEn,i with δvn,i. Write Hn`1 " Hn ' H K n denote the restriction of π to H K n has dimension 2n. It is shown in the appendix of [5] that the representation πK n of G is irreducible, for each n ě 1. Thus we have decompositions n . Note that H K n , and let πK H " C ' 8 àn"0 H K n , π " 1 ' 8 àn"0 πK n , (7.11) of the Hilbert space H and of the representation π into irreducible representations, where we identify H0 with C, and π0 with the trivial representation 1. The proof of Theorem 7.2 Proof of Theorem 7.2 (i): Recall from (7.9) that we have isometries s0, s1 on the Hilbert space H " L2pBT, µq satisfying the Cuntz relation s0s 1 " 1. The range of the isometry si is L2pBTi, µiq, where T0 and T1 are the subtrees of T with roots v0 and v1, respectively; cf. (7.2), and where µ0 and µ1 are the normalized restrictions of µ to the subsets BT0 and BT1, respectively, (making them probability measures). The Koopman representation π is self- similar with respect to the 2-similarity of H given by the isometries s0, s1, so B " πpCrGsq is self-similar. 0 `s1s By (7.11) and irreducibility of the representations πK n , we see that B is a subalgebra of M :" C ' 8 źn"0 M2npCq, (7.12) with the property that the projection of B onto each summand in (7.12) is surjective. Hence B is infinite dimensional and RFD. This completes the proof of (i). Proof of Theorem 7.2 (ii): It follows from (7.11) that the inclusion of B into M is isometric, when B is equipped with the norm arising from the Koopman representation π. Thus C π pGq, which is the completion of B with respect to this norm, embeds into M . Hence C π pGq is RFD. Moreover, it is infinite dimensional because it contains the infinite dimensional algebra B, and it is self-similar because the Koopman representation π is self-similar. Finally, M has a faithful trace, for example the one given by αjτjpxjq, τ pxq " 8 ÿj"´1 29 where x " px1, x0, x1, . . . q P M , τn is the normalized trace on M2npCq, for each n ě 1 (and τ´1 and τ0 are the normalized traces on C), and tαju8 j"´1 is any sequence of strictly positive π pGq has a faithful trace, being a sub-C -algebra of M . numbers summing up to 1. Hence C Proof of Theorem 7.2 (iii): The first claim of (iii) is proved in the lemma below, and the second claim follows from the first claim and the fact, proved in (i), that B is infinite dimensional. The result below can be found in [26]. We include its proof for completeness of the exposition. Lemma 7.4. p1 ´ dqap1 ´ dq is a non-zero element in the kernel of π : CrGs Ñ B. Proof. We observe first that z :" a ´ da ´ ad ` dad is non-zero in CrGs. Indeed, if z " 0, then a ` dad " da ` ad, which can happen only if either a " da and dad " ad, or a " ad and dad " da. Both are impossible, because d ‰ e. (It is also easy to see, for example using the action of G of the tree T , that the four elements a, da, ad, dad are pairwise distinct.) By (7.10) (and retaining the notation ¯g " πpgq, for g P G), we have ψπpπpzqq " ψπ`p1 ´ ¯dq¯ap1 ´ ¯dq " 0 0 0 1 ´ ¯b " 0, 0 0 1 ´ ¯b0 1 1 00 where ψπ is the embedding of B into M2pBq arising from self-similarity. As ψπ is injective, this implies that πpzq " 0. The proof of part (iv) of Theorem 7.2 is somewhat lengthy and is divided into several lemmas. The proof mimics the proof of the fact that G is a just-infinite group, as well as the idea from the proof of [22, Theorem 4] showing that a proper quotient of an arbitrary branch group is virtually abelian. In our situation, the following statement from [22, Proposition 8] is useful: Proposition 7.5. The normal subgroup K " xpabq2yG has finite index 16 in G, and it is of self-replicating type, written K K ă K, i.e., K K Ď ψpKq, where ψ is given by (7.4). Let ψ : G Ñ pG Gq ¸ Z{2Z be as defined in (7.4). For each m ě 1, the stabilizer subgroup StGpmq of G, with respect to the action of G on the tree T , is the set of elements g P G In particular, if g P StGp1q, then that fix all vertices at level m, i.e., all vertices in Vm. ψpgq P G G. The group K is a subgroup of StGp1q. It is also shown in [19] that G itself is self-replicating (or recurrent), in the sense that StGp1q ďS G G, where ďS is subdirect product. This means that the group homomorphisms StGp1q ψ / G G πj / G, where πj, j " 0, 1, are the coordinate homomorphisms, are surjective. Let ∆ be the ideal in B generated by the set t¯k ´ 1 k P Ku Ď B, where K is as in Proposition 7.5 above. Then B{∆ has dimension at most 16. To see this, let tt1, t2, . . . , t16u be representatives of the cosets of K in G. For each g P G, there exist i in t1, 2, . . . , 16u and k in K such that g " tik " ti ` tipk ´ 1q, so ¯g P ¯ti ` ∆. This shows that B{∆ is the linear span of the elements ¯t1 ` ∆, ¯t2 ` ∆, . . . , ¯t16 ` ∆. Let ψπ : B Ñ M2pBq be as defined in (7.8), and let ψn π : B Ñ M2npBq " B b M2n pCq denote the "nth iterate of ψπ", in the sense that ψn π " pψπ b idM2n´1 pCqq pψπ b idM2pCqq ψπ. The homomorphisms ψn π are not surjective, but the following holds: 30 / / Lemma 7.6. For each n ě 1, M2np∆q Ď ψn π p∆q. Proof. The lemma follows easily by induction on n, once the base step n " 1 has been verified. So let us show that M2p∆q Ď ψπp∆q. It follows from Proposition 7.5 that for each k P K we can find k1 P K such that ψpk1q " pk, 1q. Hence ψπp ¯k1q " ¯k 0 0 1 , ψπp ¯k1 ´ 1q " ¯k ´ 1 0 0 . 0 Let x, x1 P ∆ be such that ψπpx1q " x 0 0 0 . (7.13) (7.14) Since G is self-replicating; cf. the comments below Proposition 7.5, we can for each f P G find g P StGp1q and h P G, such that ψpgq " pf, hq. Then ψπp¯gx1q " ¯f x 0 0 , 0 ψπpx1¯gq " x ¯f 0 0 0 . Together with (7.13), this shows that ∆ 0 0 Ď ψπp∆q. 0 (7.15) If x, x1 P ∆ are such that (7.14) holds, then ψπpx1aq " 0 0 x 0 , ψπpax1q " 0 x 0 0 , ψπpax1aq " 0 0 0 x . Together with (7.15), this completes the proof. Lemma 7.7. dimpB{∆2q ď G : rK, Ks ă 8. Proof. Let ∆1 be the ideal in B generated by the set t¯k ´ 1 k P rK, Ksu. Exactly as in the argument above, showing that the dimension of B{∆ is at most G : K " 16, we see that the dimension of B{∆1 is at most G : rK, Ks. Now, K is finitely generated, and so is the quotient K{rK, Ks, which, moreover, is an abelian torsion group. Hence K{rK, Ks is finite, so G : rK, Ks " G : KK : rK, Ks is finite. For all k1, k2 P K, which shows that ∆1 Ď ∆2. This proves the lemma. 2 `pk1 ´ 1qpk2 ´ 1q ´ pk2 ´ 1qpk1 ´ 1q P ∆2, rk1, k2s ´ 1 " k´1 1 k´1 One more property of G, that we are going to exploit, is the so-called contracting property, already used in [19]. Let g denote the length of g P G with respect to the canonical generating set ta, b, c, du. With ψ : G Ñ pG Gq ¸ Z{2Z as defined in (7.4), and g P G, we have ψpgq " pg0, g1qη, where g1, g2 P G and η P te, ǫu. By [19], see also [23, Lemma 3.1], gi ď g ` 1 2 , (7.16) for i " 0, 1. In particular, gi ă g if g ě 2. The set of elements g P G for which g ď 1 is equal to N " t1, a, b, c, du, which is called the nucleus of G. We can repeat this process and obtain for each g P G and v P t0, 1un a section gv " gv P G (defined underneath (7.6)), such that ψpgvq " pg0v, g1vqηv, where ηv P te, ǫu and giv ď pgv ` 1q{2, for i " 0, 1. It follows that, for each g P G, there exists n ě 1 such that gv P N , for all v in t0, 1un. By the construction of the self-similarity map ψπ : B Ñ M2pBq, this leads to the following: 31 Lemma 7.8. For each x in B, there exists n ě 1 such that the 2n 2n matrix ψn M2npBq has entries in the linear span of the element in the nucleus ¯N " t1, ¯a, ¯b, ¯c, ¯du. π pxq P Next we will prove: Lemma 7.9. Let J be a non-zero ideal in B. There is m ě 1 so that M2m p∆2q Ď ψm π pJq. Proof. Let x be a non-zero element in J. Suppose that there exists m ě 1 such that one of π pxq is a non-zero scalar λ. Denote by epmq the 2m 2m entries, say the ps, tqth entry, of ψm , i, j " 1, 2, . . . , 2m, the standard matrix units of M2m pCq. Then, upon identifying M2m pBq with B b M2m, we have ij pp b epmq is qψm π pxqpq b epmq tj q " λpq b epmq ij , (7.17) for all p, q P B and all i, j " 1, 2, . . . , 2m. It follows from (7.17) and from Lemma 7.6 that pq b epmq π pJq, for all p, q P ∆. We conclude that z b epmq π pJq, for all z P ∆2 and all i, j " 1, 2, . . . , 2m, and hence that M2mp∆2q Ď ψm ij belongs to ψm belongs to ψm ij π pJq. To complete the proof, we show below that one of the entries of ψm π pxq is a non-zero scalar, for some m ě 1. Let n ě 1 be as in Lemma 7.8 (associated with our given x P B). Write ψn π pxq " s,t"1 with xs,t P B. By the choice of n, we deduce that xs,t belongs to the span of π pxq is non-zero, so we can find s, t pxs,tq2n ¯N " t1, ¯a, ¯b, ¯c, ¯du, for all s, t. Since ψn such that xs,t is non-zero. Write π is injective, ψn xs,t " ρ 1 ` ξ¯a ` β¯b ` γ¯c ` δ ¯d, for suitable ρ, ξ, β, γ, δ P C. Observe that, by (7.10), ψπpxs,tq " pβ ` γq¯a ` δ ` ρ ξ ξ β¯c ` γ ¯d ` δ¯b ` ρ . (7.18) The proof is now divided into three cases: 1). Assume that ξ ‰ 0. In this case both off diagonal entries of ψπpxs,tq are non-zero pxq, at least one scalars, and since ψπpxs,tq is a sub-matrix of the 2n`1 2n`1 matrix ψn`1 of the entries of ψn`1 pxq is a non-zero scalar. π π 2). Assume that either β ` γ ‰ 0, or δ ` ρ ‰ 0. Use (7.10) to compute the 2 2 matrix ψπppβ ` γq¯a ` δ ` ρq " δ ` ρ β ` γ δ ` ρ . β ` γ By assumption, one of the scalar entries in this matrix is non-zero. Further, it is a sub- πpxs,tq and hence a sub-matrix of the 2n`2 2n`2 matrix matrix of the 4 4 matrix ψ2 ψn`2 pxq. Thus at least one of the matrix entries of ψn`2 pxq is a non-zero scalar. π π 3). Assume that ξ " β ` γ " δ ` ρ " 0. Then ψπpβ¯c ` γ ¯d ` δ¯b ` ρq " pβ ` γq¯a ` δ ` ρ 0 0 β¯c ` γ ¯d ` δ¯b ` ρ " 0 0 β¯c ´ β ¯d ` δ¯b ´ δ , 0 and ψπpβ¯c ´ β ¯d ` δ¯b ´ δq " pβ ` δq¯a ` β ` δ 0 0 β ¯d ´ β¯b ` δ¯c ´ δ . 32 π If β ` δ ‰ 0, then, as in step 2), ψπppβ ` δq¯a ` β ` δq is a non-zero scalar 2 2 matrix, which is a sub-matrix of the 16 16 matrix ψ4 πpxs,tq, whence at least one of the entries of ψn`4 pxq is a non-zero scalar. If β ` δ " 0, then β ‰ 0 (because xs,t ‰ 0), so β¯c ´ β ¯d ` δ¯b ´ δ " βp¯c ´ ¯d ´ ¯b ` 1q, and ψπp¯c ´ ¯d ´ ¯b ` 1q " 0 ¯b ´ ¯c ´ ¯d ` 1 . ¯d ´ ¯b ´ ¯c ` 1 , ψπp ¯d ´ ¯b ´ ¯c ` 1q " 2 ´ 2¯a 0 0 0 0 Arguing as in step 2), we see that ψπp2 ´ 2¯aq is a non-zero scalar 2 2 matrix, which is a sub-matrix of the 32 32 matrix ψ5 pxq is a non-zero scalar. πpxs,tq, so at least one of the entries of ψn`5 π We are now ready to complete the proof of Theorem 7.2. Proof of Theorem 7.2 (iv): Let J be a non-zero ideal in B. Use Lemma 7.9 to find n ě 1 such that M2np∆2q Ď ψn π is injective, it follows that π pJq. Since ψn dimpB{Jq " dimpψn π pBq{ψn π pJqq ď dimpM2n pB{∆2qq " 22n dimpB{∆2q ă 8, by Lemma 7.7. This completes the proof. We end our paper by showing that if G is a residually finite group for which CrGs is -just- infinite, then G is hereditarily just-infinite (see also the discussion at the end of Section 6). Indeed, if CrGs is -just-infinite, then G is just-infinite, by Corollary 6.5. By the trichotomy for just-infinite groups, [22, Section 6], G must be either a branch group or hereditarily just-infinite, and the theorem below rules out the former possibility. We remind the reader about some facts concerning branch groups (see also [22]). Con- sider a spherically homogeneous rooted tree T " T ¯m, where ¯m " tmnu8 n"0, is the branching index of the tree (each mn ě 2 is an integer). For each vertex v in the kth level of the tree T , let Tv be the sub-tree of T consisting of all vertices "below" v, so that Tv is a rooted tree with root v and branching index tm1 nu8 n"0, where m1 n " mn`k. Suppose that G is a group that acts on such a spherically homogeneous rooted tree T . Then G fixes the root of the tree and hence leaves each level of the tree invariant. The rigid stabilizer of a vertex v P T , denoted by ristGpvq, is the subgroup of G consisting of all g P G which act trivially outside Tv (and fix v). The rigid stabilizer, ristGpnq, at level n P N is the subgroup of G generated by the rigid stabilizers ristGpvq of all vertices v at level n. It is easy to see that ristGpnq is, in fact, the direct product of the groups ristGpvq, where v is a vertex at level n. A group G is said to be a branch group if it admits a faithful action on such a spherically homogeneous rooted tree T " T ¯m, such that the index G : ristGpnq is finite, for all n P N, and such that T acts transitively on each level of the tree. Theorem 7.10. If G is a branch group, then CrGs is not -just-infinite, whence C pGq is not just-infinite. Proof. Fix an action of G on a spherically homogeneous rooted tree T " T ¯m satisfying the above mentioned conditions. Let π be the Koopman representation of G into the unitary n"0 µn, and µn is the uniform prob- ability measure on the set t1, 2, . . . , mnu. Denote also by π the associated -representation CrGs Ñ BpHq. group of the Hilbert space H " L2pBT, µq, where µ " Ś8 We show that π : CrGs Ñ BpHq is not injective, and that πpCrGsq is infinite dimensional. This will imply that CrGs is not -just-infinite, and hence (by Corollary 6.5) that C pGq 33 is not just-infinite. Since G acts level transitively on T , we conclude that G is infinite and that πpCrGsq is infinite dimensional. is isomorphic to Śm Let m " mp0q and let v1, v2, . . . , vm be the vertices at the first level of the tree T (below the root of the tree). The condition that G : ristGp1q is finite implies that ristGp1q, which j"1 ristGpvjq, is infinite. Moreover, by level transitivity of the action of G on T , the rigid stabilizers ristGpvjq are pairwise conjugate, so they are, in particular, non-trivial. We can therefore choose gj P ristGpvjq, for j " 1, 2, such that gj ‰ 1. Observe that p1 ´ g1qp1 ´ g2q " 1 ´ g1 ´ g2 ` g1g2 is non-zero in CrGs, because g1 ‰ 1 and g2 ‰ 1. For i " 1, 2, . . . , m, let Xi be the subset of BT consisting of words that start with vi, i.e., Xi " BTvi , so that BT is the disjoint union of the sets X1, X2, . . . Xm. Set Hi " L2pXi, µq. i"1 Hi. Let Pi be the projection from H onto Hi. Since gj acts trivially on the sub-trees Tvi, for i ‰ j, we conclude that Pi commutes with πpgjq for i " 1, 2, . . . , m and j " 1, 2, and Piπpgjq " Pi, when j ‰ i. Hence πp1 ´ gjqPi " 0, for i ‰ j. It follows that Then H " Àm πpp1´g1qp1´g2qq " πpp1´g1qp1´g2qq m ÿi"1 Pi " πpp1´g1qp1´g2qqP2 " πp1´g1qP2πp1´g2q " 0, so π : CrGs Ñ BpHq is not injective, as wanted. The theorem above (and its proof) contains item (iii) of Theorem 7.2, since G is a branch group. As in the conclusion of Theorem 7.2, it can happen, at least for some just-infinite branch groups G (for example, when G " G), that πpCrGsq is just-infinite. It may also π pGq is a RFD just-infinite C -alge- happen, for some just-infinite branch groups G, that C bra, where π as above is the Koopman representation of G arising from its action on a tree. We conjecture that C π pGq is a RFD just-infinite C -algebra. References [1] L. Bartholdi, Branch rings, thinned rings, tree enveloping rings, Israel J. Math. 154 (2006), 93–139. [2] L. Bartholdi and R. I. Grigorchuk, On the spectrum of Hecke type operators related to some fractal groups, Proc. Steklov Inst. Math. 231 (2000), no. 4, 1–41, English translation. [3] M. Bekka, Operator-algebraic superridigity for SLnpZq, n ě 3, Invent. Math. 169 (2007), no. 2, 401–425. [4] M. Bekka, M. Cowling, and P. de la Harpe, Simplicity of the reduced C -algebra of PSLpn, Zq, Internat. Math. Res. Notices (1994), no. 7, 285ff., approx. 7 pp. (electronic). [5] M. B. Bekka and P. de la Harpe, Irreducibility of unitary group representations and reproducing kernels Hilbert spaces, Expo. Math. 21 (2003), no. 2, 115–149, Appendix by the authors in collaboration with Rostislav Grigorchuk. [6] M. B. Bekka and N. Louvet, Some properties of C -algebras associated to discrete linear groups, C -algebras (Munster, 1999), Springer, Berlin, 2000, pp. 1–22. [7] V. Belyaev, R. Grigorchuk and P. Shumyatsky, On just-infiniteness of locally finite groups and their C -algebras, Bull. Math. Sciences, to appear, arXiv:1606.07648. [8] B. Blackadar, Shape theory for C -algebras, Math. Scand. 56 (1985), no. 2, 249–275. [9] K. Bonvallet, B. Hartley, D. S. Passman, and M. K. Smith, Group rings with simple augmenta- tion ideals, Proc. Amer. Math. Soc. 56 (1976), 79–82. [10] O. Bratteli, Inductive limits of finite dimensional C -algebras, Trans. Amer. Math. Soc. 171 (1972), 195–234. 34 [11] O. Bratteli and G. A. Elliott, Structure spaces of approximately finite-dimensional C -algebras. II, J. Funct. Anal. 30 (1978), no. 1, 74–82. [12] N. P. Brown, On quasidiagonal C -algebras, Operator algebras and applications, Math. Soc. Japan, Tokyo, Adv. Stud. Pure Math. 38 (2004), 19–64. [13] N. P. Brown and N. Ozawa, C -algebras and finite-dimensional approximations, Graduate Stud- ies in Mathematics, vol. 88, American Mathematical Society, Providence, RI, 2008. [14] M. Dadarlat, Nonnuclear subalgebras of AF-algebras, American J. Math. 122 (2000), no. 3, 581–597. [15] M. Dadarlat, Residually finite dimensional C -algebras and subquotients of the CAR algebra., Math. Res. Lett. 8 (2001), no. 4, 545–555. [16] K. R. Davidson, C -algebras by Example, Fields Institute Monographs, Amer. Math. Soc., Providence, R.I., 1996. [17] E. G. Effros, Dimensions and C -algebras, CBMS Regional Conference Series in Mathematics, vol. 46, Amer. Math. Soc., Washington, D.C., 1981. [18] R. Exel and T. Loring, Finite-dimensional representations of free product C -algebras, Internat. J. Math. 3 (1992), no. 4, 469–476. [19] R. I. Grigorchuk, On Burnside's problem on periodic groups, Funktsional. Anal. i Prilozhen. 14 (1980), no. 1, 53–54. [20] [21] [22] [23] [24] , Degrees of growth of finitely generated groups and the theory of invariant means, Izv. Akad. Nauk SSSR Ser. Mat. 48 (1984), no. 5, 939–985. , On the system of defining relations and the Schur multiplier of periodic groups generated by finite automata, Groups St. Andrews 1997 in Bath, I, London Math. Soc. Lecture Note Ser., vol. 260, Cambridge Univ. Press, Cambridge, 1999, pp. 290–317. , Just infinite branch groups, New horizons in pro-p groups, Progr. Math., vol. 184, Birkhauser Boston, Boston, MA, 2000, pp. 121–179. , Solved and unsolved problems around one group, Infinite groups: geometric, combina- torial and dynamical aspects, Progr. Math., vol. 248, Birkhauser, Basel, 2005, pp. 117–218. , Some problems of the dynamics of group actions on rooted trees, Tr. Mat. Inst. Steklova 273 (2011), no. Sovremennye Problemy Matematiki, 72–191. [25] R. I. Grigorchuk, V. V. Nekrashevich, and V. I. Sushchanskiı, Automata, dynamical systems, and groups, Tr. Mat. Inst. Steklova 231 (2000), no. Din. Sist., Avtom. i Beskon. Gruppy, 134–214. [26] R. I. Grigorchuk and V. V. Nekrashevych, Self-similar groups, operator algebras and Schur complement, J. Mod. Dyn. 1 (2007), no. 3, 323–370. [27] F. Leinen and O. Puglisi, Cofined subgroups in periodic simple finitary linear groups, Israel J. Math. 128 (2002), 285–324. [28] I. G. Lysenok, A set of defining relations for the Grigorchuk group, Mat. Zametki 38 (1985), no. 4, 503–516, 634. [29] D. McCarthy, Infinite groups whose proper quotien groups are finite, I, Comm. Pure and Applied Math. XXI (1968), 545–562. [30] G. J. Murphy, C -algebras and operator theory, Academic Press, London, 1990. [31] G. K. Pedersen, C -algebras and their automorphism groups, Academic Press, London, 1979. [32] F. Perera and M. Rørdam, AF-embeddings into C -algebras of real rank zero, J. Funct. Anal. 217 (2004), no. 1, 142–170. [33] L. Robert and M. Rørdam, Divisibility properties for C -algebras, Proc. Lond. Math. Soc. (3) 106 (2013), no. 6, 1330–1370. 35 [34] Y. Shalom, Bounded generation and Kazhdan's property (T), Inst. Hautes ´Etudes Sci. Publ. Math. (1999), no. 90, 145–168 (2001). [35] S. Sidki, A primitive ring associated to a Burnside 3-group, J. London Math. Soc. (2) 55 (1997), no. 1, 55–64. [36] A. C. Vieira, Modular algebras of Burnside p-groups, Mat. Contemp. 21 (2001), 287–304, 16th School of Algebra, Part II (Portuguese) (Bras´ılia, 2000). [37] N. Weaver, A prime C -algebra that is not primitive, J. Funct. Anal. 203 (2003), no. 2, 356–361. Rostislav Grigorchuk Mathematics Department Texas A& M University College Station, TX 77843-3368 USA [email protected]. Magdalena Musat Department of Mathematical Sciences University of Copenhagen Universitetsparken 5, DK-2100, Copenhagen Ø Denmark [email protected] Mikael Rørdam Department of Mathematical Sciences University of Copenhagen Universitetsparken 5, DK-2100, Copenhagen Ø Denmark [email protected] 36
1708.03471	2	1708	2017-08-29T13:01:33	A bicategorical interpretation for relative Cuntz--Pimsner algebras	[ "math.OA" ]	We interpret the construction of relative Cuntz-Pimsner algebras of correspondences in terms of the correspondence bicategory, as a reflector into a certain sub-bicategory. This generalises a previous characterisation of absolute Cuntz-Pimsner algebras of proper correspondences as colimits in the correspondence bicategory.	math.OA	math	A BICATEGORICAL INTERPRETATION FOR RELATIVE CUNTZ -- PIMSNER ALGEBRAS RALF MEYER AND CAMILA F. SEHNEM Abstract. We interpret the construction of relative Cuntz -- Pimsner algebras of correspondences in terms of the correspondence bicategory, as a reflector into a certain sub-bicategory. This generalises a previous characterisation of absolute Cuntz -- Pimsner algebras of proper correspondences as colimits in the correspondence bicategory. 1. Introduction Many important C∗-algebras may be described as (relative) Cuntz -- Pimsner algebras, see [12, 16, 17]. These are defined by triples (A,E, J), where A is a C∗-algebra, E is a C∗-correspondence from A to itself, that is, a Hilbert A-module with a nondegenerate left action of A by adjointable operators, ϕ: A → B(E), and J / A is an ideal that acts on E by compact operators, that is, ϕ(J) ⊆ K(E). The Cuntz -- Pimsner covariance condition is only required on J. We view the correspondence E as a generalised endomorphism of A. If E comes from an automorphism α of A, then the relative Cuntz -- Pimsner algebra for J = A is naturally isomorphic to the crossed product A (cid:111)α Z. So we may view Cuntz -- Pimsner algebras as analogues of crossed products for automorphisms. This is made precise in [2] by viewing both crossed products and Cuntz -- Pimsner algebras as colimits of diagrams in the bicategory of C∗-correspondences. The interpretation of Cuntz -- Pimsner algebras in [2] is limited, however, to proper correspondences, that is, ϕ(A) ⊆ K(E), and the "absolute" case J = A. This article is concerned with another bicategorical interpretation of the Cuntz -- Pimsner algebra construction, which needs no properness and extends to the relative case. Our results use the equivalence between C∗-algebras with a T-action and Fell bundles over Z, see [1]. The spectral decomposition of a T-action β on a C∗-algebra B gives a Fell bundle (Bn)n∈N over the group Z whose section C∗-algebra C∗((Bn)n∈N) is canonically isomorphic to B; namely, Bn := {b ∈ B : βz(b) = zn · b} for n ∈ Z with the multiplication, involution and norm from B. Conversely, the section C∗-algebra of any Fell bundle over Z carries a canonical gauge action of T. The Fell bundle underlying a Cuntz -- Pimsner algebra is semi-saturated, that is, Bn · Bm = Bn+m if n, m ≥ 0 (or if n, m ≤ 0). Here and below, X · Y means the closed linear span of {x · y : x ∈ X, y ∈ Y }. By the results of [1], a semi-saturated Fell bundle is determined by its fibres B0 and B1: B0 is a C∗-algebra, B1 is a Hilbert B0-bimodule, and the crossed product for the Hilbert B0-bimodule B1 is isomorphic to the section C∗-algebra of the Fell bundle generated by B0 and B1. Thus we split the construction of Cuntz -- Pimsner algebras with their canonical T-action into two steps. The first builds the Hilbert bimodule O1 J,E, the second takes the crossed product for this Hilbert bimodule. When we include the J,E over O0 Key words and phrases. relative Cuntz -- Pimsner algebra; correspondence bicategory. The second author was supported by CNPq (Brazil). 1 2 RALF MEYER AND CAMILA F. SEHNEM J,E ,O1 gauge action, then the second step is reversible using the spectral decomposition. This article interprets the first step in the construction as a reflector to a sub- bicategory. A Hilbert bimodule is a C∗-correspondence with an additional left inner product, which is unique if it exists. Thus Hilbert bimodules form a full sub-bicategory in the correspondence bicategory. We describe a bicategory whose objects are the triples (A,E, J) needed to define a relative Cuntz -- Pimsner algebra. Those triples where E is a Hilbert bimodule and J is Katsura's ideal for E form a full sub-bicategory. We show that the construction of (O0 J,E) is a reflector onto this sub-bicategory. Roughly speaking, a reflector approximates a given object by an object in the sub-bicategory in the optimal way. More precisely, it is a left (bi)adjoint to the inclusion of the sub-bicategory. We gradually work up to such bicategorical considerations. Section 2 deals with known properties of relative Cuntz -- Pimsner algebras. We also discuss their Fell bundle structure coming from the gauge action, and we show that the Cuntz -- Pimsner algebra OJ,E is the crossed product of its gauge-fixed point algebra O0 J,E by the Hilbert O0 J,E. Section 2 culminates in a result about the functoriality of relative Cuntz -- Pimsner algebras, which goes back to an idea of Schweizer [18]. We correct his idea and extend it to the relative case by defining proper covariant correspondences between triples (A,E, J) so that they induce correspondences between the associated relative Cuntz -- Pimsner algebras. J,E-bimodule O1 T. The objects of C This construction is upgraded in Section 3 to a homomorphism of bicategories N pr to the T-equivariant correspondence (or "functor") from a certain bicategory C pr are the triples (A,E, J) needed to define a rela- N bicategory C tive Cuntz -- Pimsner algebra, the arrows are the proper covariant correspondences introduced in Section 2, and the 2-arrows are isomorphisms of covariant correspon- dences. Whereas Schweizer reduces to ordinary categories by identifying isomorphic correspondences, bicategories are crucial for our purposes, as in [2]. N Then we define a sub-bicategory C pr,∗ by restricting to Hilbert bimodules instead of correspondences. We prove a crucial statement about covariant correspondences, namely, that proper covariant correspondences (A,E, J) → (B,G, IG) are "equivalent" J,E , IE) → (B,G, IG) for all (B,G, IG) to proper covariant correspondences (O0 pr,∗, that is, for a Hilbert B-bimodule G and Katsura's ideal IG. N in C Section 4 introduces the bicategorical language to understand this fact: it says J,E ,O1 that a certain arrow υ(A,E,J) : (A,E, J) → (O0 J,E ,O1 J,E , IE) is a universal arrow from (A,E, J) to C N pr,∗. The existence of universal arrows implies an adjunction (see [8]). So general bicategory theory upgrades the "equivalence" pr,∗ ⊆ C N N observed above to our main statement, namely, that the sub-bicategory C pr pr → C N N is reflective and that the reflector homomorphism C pr,∗ acts on objects by mapping (A,E, J) to (O0 J,E). We describe this reflector in detail pr,∗ → C T N and show that its composite with the crossed product homomorphism C T described in is the relative Cuntz -- Pimsner algebra homomorphism C Section 3. The definitions of bicategories, homomorphisms, transformations, and modifications are recalled in the appendix, together with some examples related to the correspondence bicategory. pr,∗ ⊆ C N J,E ,O1 J,E , IO1 2. Preliminaries In this section, we recall basic results on Cuntz -- Pimsner algebras, and their gauge action and Fell bundle structure. We correct and generalise an idea by Schweizer on the functoriality of Cuntz -- Pimsner algebras for covariant correspondences. A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 3 2.1. Correspondences. Let F1, F2 be Hilbert B-modules. Let B(F1,F2) be the space of adjointable operators from F1 to F2. Let ξihη ∈ B(F1,F2) for ξ ∈ F2 and η ∈ F1 be the generalised rank-1 operator defined by ξihη(ζ) := ξhη ζiB. Let K(F1,F2) be the closed linear span of ξihη for ξ ∈ F1 and η ∈ F2. Elements of K(F1,F2) are called compact operators. We abbreviate B(F) := B(F,F) and K(F) := K(F,F) if F = F1 = F2. Lemma 2.1. Let E1 ⊆ F1 and E2 ⊆ F2 be Hilbert B-submodules. There is a unique map K(E1,E2) → K(F1,F2) that maps ξihη ∈ K(E1,E2) to ξihη ∈ K(F1,F2) for all ξ ∈ E2, η ∈ E1. This map is injective. Definition 2.2. Let A and B be C∗-algebras. A correspondence from A to B is a Hilbert B-module F with a nondegenerate left action of A through a ∗-homomorphism ϕ: A → B(F). A correspondence is proper if ϕ(A) ⊆ K(F). It is faithful if ϕ is injective. We write F : A (cid:32) B to say that F is a correspondence from A to B. Definition 2.3. A Hilbert A, B-bimodule is a (right) Hilbert B-module F with a left Hilbert A-module structure hh··iiA such that hhξ ηiiAζ = ξhη ζiB for all ξ, η, ζ ∈ F. If F is a Hilbert A, B-bimodule, then A acts by adjointable operators on F and B acts by adjointable operators for the left Hilbert A-module structure, that is, hhξb ηiiA = hhξ ηb∗iiA for all ξ, η ∈ F and all b ∈ B. In particular, E is an A, B- bimodule. The next lemma characterises which correspondences may be enriched to Hilbert bimodules: Lemma 2.4 (see [7, Example 1.6]). A correspondence F : A (cid:32) B carries a Hilbert A, B-bimodule structure if and only if there is an ideal I / A such that left action on F restricts to a ∗-isomorphism I ∼= K(F). In this case, the ideal I and the left inner product are unique, and I = hhF FiiA. Definition 2.5. Let F1,F2 : A (cid:32) B be C∗-correspondences. A correspondence isomorphism F1 ⇒ F2 is a unitary A, B-bimodule isomorphism from F1 to F2. We write "⇒" because these isomorphisms are the 2-arrows in bicategories that we are going to construct. Let F be a Hilbert B-module and let ϕ: A → B(F) be a ∗-homomorphism. For ξ ∈ E, we define an operator Tξ : F → E ⊗ϕ F, ξ (ζ ⊗ η) = ϕ(hξ ζi)η on elementary tensors, see [17]. Hence It is adjointable with T ∗ ζ = ξihζ ⊗ 1, TξT ∗ ζ Tξ = ϕ(hζ ξi), T ∗ η 7→ ξ ⊗ η. where ξihζ ⊗ 1 is the image of ξihζ under the canonical map B(E) → B(E ⊗ϕ F), T 7→ T⊗1. Hence the operator Tξ for ξ ∈ E is compact if and only if ϕ(hξ ξi) = T ∗ ξ Tξ is compact. Lemma 2.6 ([17, Corollary 3.7]). Let J := ϕ−1(K(F)) / A and let T ∈ K(E). The operator T ⊗ 1 on E ⊗A F is compact if and only if T ∈ K(E · J) (see Lemma 2.1 for the inclusion K(E · J) ⊆ K(E)). In particular, if ϕ(A) ⊆ K(F), then T ⊗ 1 ∈ K(E ⊗ϕ F) for all T ∈ K(E). over A. By convention, E⊗0 := A. Let E+ :=L∞ 2.2. C∗-algebras of correspondences. Let E : A (cid:32) A be a correspondence over A. Let ϕ: A → B(E) be the left action. Let E⊗n be the n-fold tensor product of E n=0 E⊗n be the Fock space of E, see [17]. Define ξ : E⊗n → E⊗n+1, tn η 7→ ξ ⊗ η, RALF MEYER AND CAMILA F. SEHNEM 4 for n ≥ 0 and ξ ∈ E; this is the operator Tξ above for F = E⊗n. The operators tn combine to an operator tξ ∈ B(E+), that is, tξE⊗n = tn ξ . Let ϕ∞ : A → B(E+) be the obvious representation by block diagonal operators and let t∞ : E → B(E+) be the linear map ξ 7→ tξ. Definition 2.7. The Toeplitz C∗-algebra TE of E is the C∗-subalgebra of B(E+) generated by ϕ∞(A) + t∞(E). ξ Let J be an ideal of A with ϕ(J) ⊆ K(E). Let P0 be the projection in B(E+) that is the identity on A ⊆ E+ and that vanishes on E⊗n for n ≥ 1. Then J0 := ϕ∞(J)P0 is contained in TE. The ideal in TE generated by J0 is equal to K(E+J) ⊆ K(E+). Definition 2.8 ([16, Definition 2.18]). The relative Cuntz -- Pimsner algebra OJ,E of E with respect to J is TE /K(E+J). IE := ϕ−1 E (K(E)) ∩ (ker ϕE)⊥, The following three cases are particularly important. First, if J = {0}, then OJ,E is the Toeplitz C∗-algebra TE. Secondly, if J = ϕ−1(K(E)) and ϕ is injective, then OJ,E is the algebra OE defined by Pimsner [17]. Third, if J is Katsura's ideal (2.9) then OIE ,E is Katsura's Cuntz -- Pimsner algebra as defined in [12]. Proposition 2.10. Katsura's ideal IE in (2.9) is the largest ideal J in A with ϕ(J) ⊆ K(E) for which the canonical map A → OJ,E is injective. Proof. That πIE is injective is [12, Proposition 4.9]. The ideal IE is maximal with this property because any ideal J / A with ϕ(J) ⊆ K(E) and J 6⊆ (ker ϕ)⊥ must contain a ∈ J with ϕ(a) = 0. Then ϕ∞(a) ∈ ϕ∞(J) · P0 becomes 0 in OJ,E. (cid:3) Definition 2.11. Let E : A (cid:32) A be a correspondence and B a C∗-algebra. A representation of E in B is a pair (π, t), where π : A → B is a ∗-homomorphism, t: E → B is a linear map, and (1) π(a)t(ξ) = t(ϕ(a)ξ) for all a ∈ A and ξ ∈ E; (2) t(ξ)∗t(η) = ϕ(hξ ηiA) for all ξ, η ∈ E. These conditions imply t(ξ)π(a) = t(ξa) for all ξ ∈ E and a ∈ A. In particular, (ϕ∞, t∞) is a representation of E in the Toeplitz C∗-algebra TE. This representation is universal in the following sense: Proposition 2.12. Any representation (π, t) of E in a C∗-algebra B is of the form (π ◦ ϕ∞, π ◦ t∞) for a unique ∗-homomorphism π : TE → B. Conversely, (π ◦ ϕ∞, π ◦ t∞) is a representation of E for any ∗-homomorphism π : TE → B. Lemma 2.13. For any representation (π, t) of E, there is a unique ∗-homomorphism π1 : K(E) → B with π1(ξihη) = tξt∗ Proposition 2.14 ([16, Theorem 2.19]). The representation π of TE associated to a representation (π, t) of E factors through the quotient OJ,E of TE if and only if (2.15) In this case, we call the representation covariant on J. η for all ξ, η ∈ E. π(a) = π1(ϕ(a)) for all a ∈ J. Let (πJ , tJ) be the canonical representation of E in OJ,E. Proposition 2.14 says that (πJ , tJ) is the universal representation of E that is covariant on J. Proposition 2.16. A representation (π, t) in B is covariant on J if and only if π(J) ⊆ t(E) · B. A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 5 Proof. Let a ∈ J. Then π1(ϕ(a)) is contained in the closed linear span of t(E)t(E)∗ and hence in t(E)·B. So π(a) ∈ t(E)·B is necessary for π(a) = π1(ϕ(a)). Conversely, assume π(a) ∈ t(E) · B for all a ∈ J. We have π(a) · t(ξ) = t(ϕ(a)ξ) = π1(ϕ(a))t(ξ) for all ξ ∈ E (see [12, Lemma 2.4]). Hence (π(a) − π1(ϕ(a))) · t(E) · B = 0. Since π(a∗), π1(ϕ(a∗)) ∈ t(E) · B, we get (π(a) − π1(ϕ(a))) · (π(a) − π1(ϕ(a)))∗ = 0. This (cid:3) is equivalent to π(a) = π1(ϕ(a)). 2.3. Gauge action and Fell bundle structure. Let E : A (cid:32) A be a correspon- dence and let J / A be an ideal with ϕ(J) ⊆ K(E). If (π, t) is a representation of E that is covariant on J, then so is (π, z·t) for z ∈ T. This operation on representations comes from an automorphism of the relative Cuntz -- Pimsner algebra OJ,E by its universal property. These automorphisms define a continuous action γ of T on OJ,E, called the gauge action. Let On J,E := {b ∈ OJ,E : γz(b) = znb for all z ∈ T} and (On J,E · Om J,E ⊆ On+m J,E J,E)∗ = O−n for n ∈ Z be the nth spectral subspace. These spectral subspaces form a Fell bundle over Z, that is, On J,E for all n, m ∈ Z. In particular, for J = {0} we get a gauge action on TE and corresponding spectral subspaces T nE ⊆ TE. Explicitly, the gauge action on TE comes from the obvious N-grading on E+: if x ∈ TE, then x ∈ T nE if and only if x(E⊗k) ⊆ E⊗n+k for all k ∈ N; this means xE⊗k = 0 if k + n < 0. And OnE,J is the image of T nE in OE,J. Lemma 2.17. Let n ∈ Z. The subspace On J,E in OJ,E is the closed linear span of tJ(ξ1)tJ(ξ2)··· tJ(ξk) · t∗ J(η1) for ξi, ηj ∈ E, k − l = n. If n ∈ N, then J(ηl)··· t∗ J(η2)t∗ J,E ∼= E⊗n ⊗A O0 On J,E J,E. The Fell bundle (Ok T J,E · Ol as a correspondence A (cid:32) O0 is, Ok Proof. Let b ∈ On of monomials tJ(ξ1)tJ(ξ2)··· tJ(ξk) · t∗ J,E = Ok+l J,E if k, l ≥ 0. J,E and let > 0. Then b is -close to a finite linear combination b Z J(η1) with k, l ∈ N. Define x ∈ OJ,E . J(ηl)··· t∗ z−nγz(x) dz, pn(x) := J(η2)t∗ J,E)k∈Z is semi-saturated, that This is a contractive projection from OJ,E onto On we have kb − pn(b)k ≤ as well. J(ηl)··· t∗ tJ(ξ1)tJ(ξ2)··· tJ(ξk)·t∗ Hence On The monomials generating Ok+l J,E is the closed linear span of such monomials with k − l = n. J,E · Ol J(η2)t∗ J,E for k, l ≥ 0 are obviously in Ok J,E. Since pn(b) = b and kpnk ≤ 1, Inspection shows that pn maps a monomial J(η1) to itself if k−l = n and kills it otherwise. the first statement immediately implies the last one. There is an isometric A,O0 bimodule map J,E. Hence J,E- E⊗n ⊗A O0 J,E → On J,E , ξ1 ⊗ ··· ⊗ ξn ⊗ y 7→ tJ(ξ1)··· tJ(ξn) · y. The first statement implies that its image is dense, so it is unitary. The Fell bundle (On J,E)n∈Z need not be saturated, that is, On J,E · O−n J,E. from O0 Theorem 2.18. The relative Cuntz -- Pimsner algebra is T-equivariantly isomorphic J,E by the Hilbert O0 to the crossed product of O0 J,E and to the full or reduced section C∗-algebra of the Fell bundle (On Proof. The Fell bundle (On of Abadie -- Eilers -- Exel [1] imply our claims. J,E)n∈Z is semi-saturated by Lemma 2.17. Now the results (cid:3) J,E-bimodule O1 J,E)n∈Z. (cid:3) J,E may differ 6 RALF MEYER AND CAMILA F. SEHNEM J,E-bimodule O1 Theorem 2.18 splits the construction of relative Cuntz -- Pimsner algebras into two steps. The first builds the Hilbert O0 J,E, the second takes the crossed product for this Hilbert bimodule. A Hilbert bimodule G on a C∗-algebra B is the same as a Morita -- Rieffel equivalence between two ideals in B or, briefly, a partial Morita -- Rieffel equivalence on B (this point of view is explained in [4]). The crossed product O0 J,E generalises the partial crossed product for a partial automorphism. Many results about crossed products for automorphisms extend to Hilbert bimodule crossed products. In particular, the standard criteria for simplicity and detection and separation of ideals are extended in [14]. Proposition 2.19. The following conditions are equivalent: J,E (cid:111) O1 (1) the map πJ : A → O0 J,E is an isomorphism; (2) the map ϕ: J → K(E) is an isomorphism; (3) the correspondence E comes from a Hilbert bimodule and J = IE. IG ,G as a Hilbert B-bimodule. IG ,G (cid:111) O1 IG ,G ∼= G ⊗A O0 IG ,G, and the isomorphism O1 Proof. If J = IE is Katsura's ideal, then everything follows from [12, Proposi- tion 5.18]. So it remains to observe that (1) and (2) fail if J 6= IE. Lemma 2.4 shows that E comes from a Hilbert bimodule if and only if there is an ideal I in A so that ϕI : I → K(E) is an isomorphism. In this case, I is the largest ideal on which ϕ restricts to an injective map into K(E). So I = IE. Thus (2) ⇐⇒ (3). If J 6⊆ IE, then A → OJ,E is not injective by Proposition 2.10. So (1) implies J ⊆ IE. If J ⊆ IE and (1) holds, then the map A → OIE ,E is still surjective because OIE ,E is a quotient of OJ,E, and it is also injective by Proposition 2.10. Hence OIE ,E = OJ,E. This implies K(E+IE) = K(E+J) and hence IE = J because of the direct summand A in E+. (cid:3) Proposition 2.20. Let G be a Hilbert B-bimodule and let IG be Katsura's ideal for G. Then OIG ,G ∼= B (cid:111) G T-equivariantly. Proof. Theorem 2.18 identifies OIG ,G ∼= O0 IG ,G. Proposition 2.19 gives B ∼= O0 IG ,G from Lemma 2.17 implies that G ∼= O1 (cid:3) 2.4. Functoriality of relative Cuntz -- Pimsner algebras. Schweizer [18] has defined "covariant homomorphisms" and "covariant correspondences" between self- correspondences and has asserted that they induce ∗-homomorphisms and corre- spondences between the associated Toeplitz and absolute Cuntz -- Pimsner algebras. For the proof of functoriality for covariant correspondences he refers to a preprint that never got published. In fact, there are some technical pitfalls. We correct his statement here, and also add a condition to treat relative Cuntz -- Pimsner algebras. Throughout this subsection, let E : A (cid:32) A and G : B (cid:32) B be correspondences and let JA ⊆ ϕ−1(K(E)) and JB ⊆ ϕ−1(K(G)) be ideals. Definition 2.21. A covariant correspondence from (A,E, JA) to (B,G, JB) is a pair (F, V ), where F is a correspondence A (cid:32) B with JA · F ⊆ F · JB and V is a correspondence isomorphism E ⊗A F ⇒ F ⊗B G. A covariant correspondence is proper if F is proper. Proposition 2.22. A proper covariant correspondence (F, V ) from (A,E, JA) to (B,G, JB) induces a proper T-equivariant correspondence OF ,V : OJA,E (cid:32) OJB ,G. Schweizer [18] claims this also for non-proper correspondences, and he allows V to be a non-adjointable isometry. In fact, a pair (F, V ) where V is only a non- adjointable isometry induces a correspondence between the Toeplitz C∗-algebras. It is unclear, however, when this correspondence descends to one between the absolute or relative Cuntz -- Pimsner algebras. And we need F to be proper. Alternatively, we may require E instead of F to be proper. This situation is treated in [2]. A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 7 Proof. We use the canonical ∗-homomorphism πJB : B → OJB ,G to view OJB ,G as a proper correspondence B (cid:32) OJB ,G. Thus FO := F ⊗B OJB ,G becomes a proper correspondence A (cid:32) OJB ,G, that is, a Hilbert OJB ,G-module with a representation π : A → K(FO). The T-action on OJB ,G induces a T-action on FO because πJB(B) ⊆ O0 JB ,G. We are going to define a map t: E → K(FO) such that (π, t) is a representation of (A,E) on FO that is covariant on JA. Then Proposition 2.14 yields a representation π : OJA,E → K(FO). This is the desired correspondence OJA,E (cid:32) OJB ,G. There is an isometry µG : G ⊗B OJB ,G ⇒ OJB ,G, ζ ⊗ y 7→ t∞(ζ) · y, of correspon- dences B (cid:32) OJB ,G. Usually, it is not unitary. We define an isometry 1⊗µG===⇒ F ⊗B OJB ,G = FO. V ! : E ⊗A FO = E ⊗A F ⊗B OJB ,G V ⊗1===⇒ F ⊗B G ⊗B OJB ,G It yields a map t from E to the space of bounded operators on FO by t(ξ)(η) := V !(ξ⊗ η). To show that t(ξ) is adjointable, we need that FO is a proper correspondence A (cid:32) OJB ,G: then Tξ ∈ K(FO,E ⊗A FO), and composition with V ! maps this into K(FO) by Lemma 2.1. So even t(ξ) ∈ K(FO) for all ξ ∈ E. We claim that the pair (π, t) is a representation. We have π(a)t(ξ) = t(ϕ(a)ξ) because V ! is a left A-module map. And t(ξ1)∗t(ξ2) = π(hξ1 ξ2i) holds because ht(ξ1)η1 t(ξ2)η2i = hV !(ξ1 ⊗ η1) V !(ξ2 ⊗ η2)i = hξ1 ⊗ η1 ξ2 ⊗ η2i = hη1 π(hξ1 ξ2i)η2i. If JA = 0, then we are done at this point, and we have not yet used that V is unitary. So the Toeplitz C∗-algebra of a correspondence remains functorial for proper covariant correspondences where V is not unitary. It remains to prove that π is covariant on JA. By Proposition 2.16, this is equivalent to π(JA)(FO) ⊆ t(E)(FO). And JB · OJB ,G ⊆ tJB(G) · OJB ,G holds because the canonical representation of (B,G) on OJB ,G is covariant on JB. Since JA · F ⊆ F · JB by assumption, JA · FO ⊆ F ⊗ JB · OJB ,G ⊆ F ⊗ tJB(G) · OJB ,G = (1 ⊗ µG)(F ⊗B G ⊗B OJB ,G). Since V is unitary, we may rewrite this further as V !(E ⊗A F ⊗B OJB ,G) = t(E)·FO. This finishes the proof that (π, t) is covariant on JA. The operators t(ξ) for ξ ∈ E (cid:3) are homogeneous of degree 1 for the T-action. Thus π is T-equivariant. Example 2.23. Let A = B and J = JA = JB 6= {0} and let E ⊆ G be an A-invariant Hilbert submodule. Then the identity correspondence F = A with the inclusion map E ⊗A F ∼= E ,→ G ∼= G ⊗B F is a covariant correspondence in the notation of Schweizer. There is indeed a canonical ∗-homomorphism TE → TG. But it need not descend to the relative Cuntz -- Pimsner algebras because ϕG(a) ∈ K(G) for a ∈ J need not be the extension of ϕE(a) ∈ K(E) given by Lemma 2.1. So the Cuntz -- Pimsner covariance conditions for OE,J and OG,J may be incompatible. We ask V to be unitary to avoid this problem. Example 2.24. Turn O0 J,E with the obvious left action of A. The proper correspondence O0 J,E with the isomorphism from Lemma 2.17 is a proper covariant correspondence from E : A (cid:32) A J,E : O0 with the ideal J to O1 J,E (cid:32) O0 J,E. It remains to show that J ·O0 J,E ·IO1 J,E ⊆ O0 J,E = IO1 J,E is a Hilbert bimodule, Katsura's ideal is equal to the range ideal of the left inner product, that is, the closed linear span of xy∗ for all x, y ∈ O1 J,E. This contains K(E) for x, y ∈ E, which in turn contains J by the Cuntz -- Pimsner covariance condition on J, see Proposition 2.14. So J · O0 J,E) is J,E, into a proper C∗-correspondence A (cid:32) O0 J,E : A (cid:32) O0 J,E with Katsura's ideal IO1 J,E. Since O1 J,E , IO1 J,E ⊆ IO1 J,E. The relative Cuntz -- Pimsner algebra of (O0 J,E ,O1 J,E by Proposition 2.19. The correspondence O0 J,E (cid:32) O0 J,E associated to the J,E. RALF MEYER AND CAMILA F. SEHNEM 8 again O0 covariant correspondence above is just the identity correspondence on O0 Remark 2.25. If JA = 0 or JB = ϕ−1(K(G)), then the condition JA · F ⊆ F · JB for covariant transformations (A,E, JA) → (B,G, JB) always holds and so may be left out. This is clear if JA = 0. Let JB = ϕ−1(K(G)). Since F is proper, JA acts on E ⊗A F ∼= F ⊗B G by compact operators by Lemma 2.6. Again by Lemma 2.6, this implies JA ⊆ K(F · JB). Thus JA · F ⊆ F · JB. Example 2.26. Covariant correspondences are related to the T-pairs used by Kat- sura [13] to describe the ideal structure of relative Cuntz -- Pimsner algebras. For this, we specialise to covariant correspondences out of (A,E, J) where the underlying correspondence comes from a quotient map A → A/I. That is, F = A/I : A (cid:32) A/I for an ideal I / A. When is this part of a covariant correspondence from (A,E, J) to (A/I,E0, J0) for some E0, J0? There are natural isomorphisms E ⊗A F ∼= E/EI and F ⊗A/I E0 ∼= E0 as corre- spondences A (cid:32) A/I. So the only possible choice for E0 is E0 := E/EI with a left A/I-action which gives the canonical A-action when composed with the quotient map A → A/I. Such a correspondence E/EI : A/I (cid:32) A/I exists if and only if E is positively invariant, that is, IE ⊆ EI. Assume this to be the case. An ideal J0 / A/I is equivalent to an ideal I0 / A that contains I. For a covariant correspondence, we require JF ⊆ FJ0, which means that J ⊆ I0. And in order for (A/I,E0, J0) to be an object of C pr, the ideal J0 or, equivalently, I0, should act by N compact operators on E0 := E/EI. It is unique up to an automorphism of E/EI, that is, a unitary operator on E/EI that also commutes with the left action of A or A/I, but this shall not concern us. So we get a covariant correspondence in this case. This induces a correspondence from OJ,E to OJ0,E0 by Proposition 2.22. Actually, our covariant correspondence is a covariant homomorphism, and so the correspondence from Proposition 2.22 comes from a T-equivariant ∗-homomorphism, which turns out to be surjective. So a pair of ideals (I, I0) as above induces a T-equivariant quotient or, equivalently, a T-invariant ideal in OJ,E. If I0/I contains elements that act by 0 on K(E/EI), then the map A/I → OJ0,E0 is not injective by Proposition 2.10. Then we may enlarge I without changing the relative Cuntz -- Pimsner algebra. When we add the condition that no non-zero element of I0/I acts by a compact operator on E/E · I, then we get exactly the T-pairs with J ⊆ I0 of [13]. The T-pairs (I, I0) with J ⊆ I0 correspond bijectively to gauge-invariant ideals of OJ,E by [13, Proposition 11.9]. Sometimes different pairs (I, I0) produce the same quotient of OJ,E. Then there is an isomorphism E ⊗A F ∼= F ⊗A E0. 3. Bicategories of correspondences and Hilbert bimodules We are going to enrich the relative Cuntz -- Pimsner algebra construction to a homomorphism (functor) from a suitable bicategory of covariant correspondences to the T-equivariant correspondence bicategory. Most of the work is already done in Proposition 2.22, which describes how this homomorphism acts on arrows. It remains to define the appropriate bicategories and write down the remaining data of a homomorphism. The correspondence bicategory of C∗-algebras and related bicategories have been discussed in [2,4 -- 6]. We recall basic bicategorical definitions in the appendix for the convenience of the reader. Here we go through these notions much more quickly. Let C be the correspondence bicategory. It has C∗-algebras as objects, A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 9 C∗-correspondences as arrows, and correspondence isomorphisms as 2-arrows. The composition is the tensor product ⊗B of C∗-correspondences. Given any bicategory D, there is a bicategory CD with homomorphisms D → C as objects, transformations between these homomorphisms as arrows, and mod- ifications between these transformations as 2-arrows (see the appendix for these notions). There is also a continuous version of this for a locally compact, topological bicategory D. In particular, we shall use the T-equivariant correspondence bicat- T. Its objects are C∗-algebras with a continuous T-action. Its arrows are egory C T-equivariant C∗-correspondences, and 2-arrows are T-equivariant isomorphisms of C∗-correspondences. When D is the monoid (N, +), we may simplify the bicategory CD, see [2, Section 5]. An object in it is equivalent to a C∗-algebra A with a self-correspondence E : A (cid:32) A. An arrow is equivalent to a covariant correspondence (without the condition JAF ⊆ FJB), and a 2-arrow is equivalent to an isomorphism between two N where N covariant correspondences. The bicategory C pr that we need is a variant of C we add the ideal J and allow only proper covariant correspondences as arrows. N pr has the following data (see Definition A.1): • Objects are triples (A,E, J), where A is a C∗-algebra, E : A (cid:32) A is a C∗-correspondence, and J ⊆ ϕ−1(K(E)) is an ideal. • Arrows (A,E, J) → (A1,E1, J1) are proper covariant correspondences (F, u) from (A,E, J) to (A1,E1, J1), that is, F is a proper correspondence A (cid:32) A1 with JF ⊆ FJ1 and u is a correspondence isomorphism E ⊗AF ⇒ F ⊗A1 E1. • 2-Arrows (F0, u0) ⇒ (F1, u1) are isomorphisms of covariant correspondences, that is, correspondence isomorphisms w: F0 ⇒ F1 for which the following diagram commutes: Definition 3.1. The bicategory C E ⊗A F0 1E⊗w E ⊗A F1 u0 + u1 + F0 ⊗A1 E1 w⊗1E1 3 F1 ⊗A1 E1 • The vertical product of 2-arrows w0 : (F0, u0) ⇒ (F1, u1), w1 : (F1, u1) ⇒ (F2, u2) pr((A,E, J), (A1,E1, J1)). N is the usual product w1 · w0 : F0 → F2. This is indeed a 2-arrow from (F0, u0) to (F2, u2). And the vertical product is associative and unital. Thus the arrows (A,E, J) → (A1,E1, J1) and the 2-arrows between them form a category C • Let (F, u): (A,E, J) → (A1,E1, J1) and (F1, u1): (A1,E1, J1) → (A2,E2, J2) be arrows. Their product is (F1, u1)◦(F, u) := (F⊗A1F1, u•u1), where u•u1 is the composite correspondence isomorphism E ⊗A F ⊗A1 F1 • The horizontal product for a diagram of arrows and 2-arrows u⊗1F1−−−−→ F ⊗A1 E1 ⊗A1 F1 1F⊗u1 −−−−→ F ⊗A1 F1 ⊗A2 E2. (A,E, J) (F ,u) w (eF ,eu) 3(A1,E1, J1) (F1,u1) w1 (eF1,eu1) 3(A2,E2, J2) 3 + + 3 + + 3 • The unit arrow on the object (A,E, J) is the proper covariant correspondence (A, ιE), where A is the identity correspondence, that is, A with the obvious A-bimodule structure and the inner product hx yi := x∗y, and ιE is the canonical isomorphism E ⊗A A ∼= E ∼= A ⊗A E 10 RALF MEYER AND CAMILA F. SEHNEM is the 2-arrow (A,E, J) (F⊗A1F1,u•u1) w⊗w1 (eF⊗A1eF1,eu•eu1) 2 (A2,E2, J2). This horizontal product and the product of arrows combine to composition bifunctors pr((A,E, J), (A1,E1, J1)) × C N C pr((A1,E1, J1), (A2,E2, J2)) N → C pr((A,E, J), (A2,E2, J2)). N built from the right and left actions of A on E. • The associators and unitors are the same as in the correspondence bicategory. Thus they inherit the coherence conditions needed for a bicategory. pr → C T that maps each object (A,E, J) N Theorem 3.2. There is a homomorphism C to its relative Cuntz -- Pimsner algebra and is the construction of Proposition 2.22 on arrows. Proof. The construction in Proposition 2.22 is "natural" and thus functorial for isomorphisms of covariant correspondences, and it maps the identity covariant correspondence to the identity T-equivariant correspondence on the relative Cuntz -- Pimsner algebras. Let (F, u): (A,E, J) → (A1,E1, J1) and (F1, u1): (A1,E1, J1) → (A2,E2, J2) be covariant correspondences and let OF ,u and OF1,u1 be the associated correspondences of relative Cuntz -- Pimsner algebras. By definition, OF ,u ⊗OJ1 ,E1 OF1,u1 and OF⊗A1F1,u•u1 are equal to (F ⊗A1 OJ1,F1)⊗OJ1 ,F1 (F1 ⊗A2 OJ2,F2) and (F ⊗A1 F1) ⊗A2 OJ2,F2 as T-equivariant correspondences A (cid:32) OJ2,F2. Associators and unit transformations give a canonical T-equivariant isomorphism between these correspondences. This isomorphism also intertwines the representations of E. Hence it is an isomorphism of correspondences OJ,F (cid:32) OJ2,F2. These canonical isomorphisms satisfy the coherence conditions for a homomorphism of bicategories (cid:3) in Definition A.5. J,E (cid:111) O1 The relative Cuntz -- Pimsner algebra OJ,E is the crossed product O0 J,E by Theorem 2.18. So OJ,E with the gauge T-action and the Hilbert O0 J,E-bimodule O1 J,E contain the same amount of information. We now study the construction that sends (A,E, J) to the Hilbert O0 J,E. The appropriate bicategory of Hilbert bimodules is a sub-bicategory of C pr,∗ ⊆ C N N Definition 3.3. Let C pr be the full sub-bicategory whose objects are triples (B,G, IG), where G is a Hilbert B-bimodule and IG is Katsura's ideal for G, which is also equal to the range ideal hhG Gii of the left inner product on G. The arrows N N and 2-arrows among objects of C pr, including the condition pr,∗ are the same as in C IEF ⊆ FIG for covariant correspondences. J,E-bimodule O1 N pr: pr → C N T When we restrict the relative Cuntz -- Pimsner algebra construction C N pr,∗, we get the (partial) crossed product construction for Hilbert bimodules by to C , , 2 A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 11 T. J,E ,O1 Proposition 2.20. Thus Theorem 3.2 also completes the crossed product for Hilbert pr,∗ → C N bimodules to a functor C The map that sends (A,E, J) to (O0 pr → C N N J,E) is part of a functor C pr,∗ pr,∗ → C T, gives the N which, when composed with the crossed product functor C relative Cuntz -- Pimsner algebra functor of Theorem 3.2. We do not prove this now because it follows from our main result. The key step is the following universal property of (O0 Proposition 3.4. Let (A,E, J) and (B,G, IG) be objects of C N pr and C tively. Let N pr,∗, respec- J,E ,O1 J,E , IO1 J,E , IO1 J,E): υ(A,E,J) : (A,E, J) → (O0 J,E ,O1 J,E , IO1 J,E) be the covariant correspondence from Example 2.24. Composition with υ(A,E,J) induces a groupoid equivalence (cid:0)(A,E, J), (B,G, IG)(cid:1) ' C pr,∗(cid:0)(O0 N N pr C J,E), (B,G, IG)(cid:1). J,E ,O1 J,E , IO1 Recall that C pr ((A,E, J), (A1,E1, J1)) for objects (A,E, J) and (A1,E1, J1) of C N N pr denotes the groupoid with arrows (A,E, J) → (A1,E1, J1) as objects and 2-arrows among them as arrows. Proof. We begin with an auxiliary construction. Proposition 2.20 identifies OIG ,G ∼= IG ,G ∼= G, O−1 IG ,G ∼= G∗ B(cid:111)G as Z-graded C∗-algebras. In particular, O0 (B,G, IG). It induces a proper, T-equivariant correspondence OF ,V =L as Hilbert B-bimodules. Let (F, u) be a proper covariant correspondence (A,E, J) → n∈Z OnF ,V from OJ,E to OIG ,G by Proposition 2.22. By construction, OnF ,V = F ⊗B On IG ,G. IG ,G ∼= F ⊗B G. The Thus O0F ,V = F ⊗B O0 left action on OF ,V is a nondegenerate, T-equivariant ∗-homomorphism OJ,E → K(OF ,V ). So O0 J,E acts on OF ,V by grading-preserving operators. Restricting to the J,E → K(O0F ,V ) ∼= K(F). degree-0 part, we get a nondegenerate ∗-homomorphism O0 Let F # be F viewed as a correspondence O0 IG ,G ∼= F ⊗B B ∼= F and O1F ,V = F ⊗B O1 IG ,G ∼= B and O1 J,E (cid:32) B in this way. We now construct an isomorphism of correspondences J,E F # ⇒ F # ⊗B G. J,E ⊗O0 u# : O1 We need two descriptions of u#. The first shows that it is unitary, the second that it intertwines the left actions of O0 J,E. The first formula for u# uses Lemma 2.17, which gives unitary Hilbert B-module maps J,E F # ∼= E ⊗A O0 J,E F # ∼= E ⊗A F. O1 J,E ⊗O0 J,E ⊗O0 J,E ⊗O0 ∼= F # ⊗B G. J,E O0F ,V → O1F ,V J,E F # = O1 J,E-linear because the isomorphism F # ⊗B On Composing with u: E ⊗A F ⇒ F ⊗B G gives the desired unitary u#. The second formula for u# restricts the left action of OJ,E on OF ,V to a multiplication map O1 J,E ⊗O0 (3.5) IG ,G ∼= OnF ,V is by This is manifestly O0 right multiplication and so intertwines the left actions of O0 J,E. The map in (3.5) maps tJ(ξ) ⊗ η 7→ u(ξ ⊗ η) for all ξ ∈ E, η ∈ F. This determines it by Lemma 2.17. So both constructions give the same map u#. J,E · F # ⊆ F# · IG holds, so that the pair (F #, u#) is a proper J,E) to (B,G, IG). The ideal IO1 J,E is J,E. Using the Fell bundle structure, covariant correspondence from (O0 J,E , IO1 equal to the range of the left inner product on O1 we may rewrite this as O1 J,E · O−1 J,E · O−1 J,E. Thus J,E · O0F ,V ⊆ O1 F ,V = E · O−1 F ,V . We claim that IO1 F ,V = E · O0 J,E · O0F ,V = O1 J,E · O−1 J,E · O−1 J,E ,O1 IO1 12 The product E · O−1 Proposition 2.22. So E · O−1 RALF MEYER AND CAMILA F. SEHNEM F ,V uses the representation of E on OF ,V built in the proof of E ⊗A F ⊗B G∗ ∼= F ⊗B G ⊗B G∗ = F · IG. F ,V is the image of the map J,E , IO1 J,E ,O1 J,E ,O1 J,E) to (B,G, IG). J,E·O0F ,V ⊆ F·IG as claimed. We have turned a proper covariant correspondence So IO1 (F, u) from (A,E, J) to (B,G, IG) into a proper covariant correspondence (F #, u#) from (O0 J,E ∼= O0 Conversely, take a proper covariant correspondence (F, u) from (O0 J,E) to (B,G, IG). Composing it with υ(A,E,J) gives a proper covariant correspondence from (A,E, J) to (B,G, IG). We now simplify this product of covariant correspon- J,E in υ(A,E,J) is O0 dences. The underlying correspondence A → O0 J,E, and the isomorphism E ⊗A O0 J,E = O1 J,E is the one from Lemma 2.17. We identify the tensor product O0 J,E F with F by the canonical map. Thus the product of (F, u) with υ(A,E,J) is canonically isomorphic to a covariant cor- respondence (F [, u[) with underlying correspondence F [ = F : A (cid:32) B with the left A-action through πJ : A → O0 J,E. The isomorphism u[ : E ⊗A F [ ⇒ F [ ⊗B G J,E F ⇒ F ⊗B G with the is the composite of the given isomorphism u: O1 isomorphism E ⊗A O0 J,E from Lemma 2.17. Now let (F, u) be a proper covariant correspondence from (A,E, J) to (B,G, IG). J,E O1 J,E ⊗O0 J,E ∼= O1 J,E ⊗Ou J,E ⊗O0 J,E , IO1 J,E , IO1 J,E ,O1 (F #[, u#[) = (F, u). J,E is the original action of A. The isomorphism E ⊗A O0 We claim that (3.6) By construction, the underlying Hilbert B-module of F #[ is F. We even have F #[ = F as correspondences A (cid:32) B, that is, the left O0 J,E-action on F # composed J,E ∼= O1 with πJ : A → O0 J,E is used both to get u# from u and to get u#[ from u#. Unravelling this shows that u#[ = u. Now we claim that the map that sends a proper covariant correspondence J,E) → (B,G, IG) to (F [, u[) is injective. This claim and (3.6) (F, u): (O0 imply (F [#, u[#) = (F, u), that is, our two operations are inverse to each other. To prove injectivity, we use Proposition 2.22 to build a correspondence OF ,u : OJ,E (cid:32) OIG ,G from (F, u). This correspondence determines (F, u): we can get back F as its degree-0 part because OIG ,G = B (cid:111) G, and because u and the left O0 J,E-module structure on F are both contained in the left OJ,E-module structure on OF ,u. An OJ,E-module structure on OIG ,G is already determined by a representation of (A,E). Since On IG ,G, this representation is determined by its IG ,G ∼= B. And (F [, u[) determines the representation of (A,E) on B. restriction to O0 Thus (F [, u[) determines (F, u). The constructions of (F #, u#) and (F [, u[) are clearly natural for isomorphisms (cid:0)(A,E, J), (B,G, IG)(cid:1) ∼= C of covariant correspondences. So they form an isomorphism of categories J,E), (B,G)(cid:1). pr,∗(cid:0)(O0 IG ,G = O0 IG ,G · On J,E ,O1 J,E , IO1 N C pr N One piece in this isomorphism is naturally equivalent to the functor that composes (cid:3) with υ(A,E,J). Hence this functor is an equivalence of categories, as asserted. 4. The reflector from correspondences to Hilbert bimodules We now strengthen Proposition 3.4 using some general results on adjunctions of homomorphisms between bicategories. We first recall the related and better known results about ordinary categories and functors. C(c, x) → B(b, R(x)), f 7→ R(f) ◦ υ, A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 13 Let C and B be categories. Let R: C → B be a functor and b ∈ obB. An object c ∈ obC with an arrow υ : b → R(c) is called a universal arrow from b to R if, for each x ∈ obC and each f ∈ B(b, R(x)), there is a unique g ∈ C(c, x) with R(g) ◦ υ = f. Equivalently, the maps (4.1) are bijective for all x ∈ obC. The functor R has a left adjoint L: B → C if and only if such universal arrows exist for all x ∈ obC. The left adjoint functor L: B → C is uniquely determined up to natural isomorphism. It maps b 7→ c on objects, and the isomorphisms (4.1) become natural in both b and x when we replace c by L(b). An adjunction between L and R may also be expressed through its unit and counit, that is, natural transformations L◦ R ⇒ idC and idB ⇒ R◦ L such that the induced transformations L ⇒ L◦R◦L ⇒ L and R ⇒ R◦L◦R ⇒ R are unit transformations. A subcategory C ⊆ B is called reflective if the inclusion functor R: C → B has a left adjoint L: B → C. The functor L is called reflector. The case we care about is a bicategorical version of a full subcategory. If C ⊆ B is a full subcategory, then we may choose L ◦ R to be the identity functor on C and the counit L ◦ R ⇒ idC to be the unit natural transformation. Fiore [8] carries the story of adjoint functors over to homomorphisms between 2-categories (which he calls "pseudo functors"), that is, bicategories where the associators and unitors are identity 2-arrows. The bicategories we need are not 2-categories. But any bicategory is equivalent to a 2-category by MacLane's Coher- ence Theorem. Hence Fiore's definitions and results apply in bicategories as well. We shorten notation by speaking of "universal" arrows and "adjunctions" instead of "biuniversal" arrows and "biadjunctions." A 2-category is also a category with some extra structure. So leaving out the prefix "bi" may cause confusion in that setting. But it will always be clear whether we mean the categorical or bicategorical notions. Definition 4.2 ([8, Definition 9.4]). Let B and C be bicategories, R: C → B a homomorphism, and b ∈ obB. Let c ∈ obC and let g : b → R(c) be an arrow in B. The pair (c, g) is a universal arrow from b to R if, for every x ∈ obC, the following functor is an equivalence of categories: g∗ : C(c, x) → B(b, R(x)), f 7→ R(f) · g, w 7→ R(w) • 1g. Universal arrows are called left biliftings by Street [19]. We can now reformulate Proposition 3.4: Proposition 4.3. Let (A,E, J) ∈ ob C N pr. The covariant correspondence υ(A,E,J) J,E ,O1 J,E , IE) is a universal arrow from (A,E, J) to the inclusion from (A,E, J) to (O0 pr,∗ → C (cid:3) N N pr. homomorphism C There are two alternative definitions of adjunctions, based on equivalences between morphism categories or on units and counits. These are spelled out, respectively, by Fiore in [8, Definition 9.8] and by Gurski in [11, Definition 2.1]. We shall use Fiore's definition. Definition 4.4 ([8, Definition 9.8]). Let B and C be bicategories. An adjunction between them consists of • two homomorphisms L: B → C, R: C → B; • equivalences of categories ϕb,c : C(L(b), c) ' B(b, R(c)) for all b ∈ obB, c ∈ obC; 14 RALF MEYER AND CAMILA F. SEHNEM • natural equivalences of functors C(L(b1), c1) f∗ C(L(b2), c1) g∗ / C(L(b2), c2) ϕb1,c1 B(b1, R(c1)) / B(b2, R(c1)) / B(b2, R(c2)) g∗ ϕb2,c2 f∗ for all arrows f : b2 → b1, g : c1 → c2 in B and C. These are subject to a coherence condition. In brief, the functors ϕb,c and the natural equivalences form a transformation between the homomorphisms Bop × C ⇒ Cat, (b, c) 7→ C(L(b), c), B(b, R(c)). Here Cat is the bicategory of categories, see Example A.2. Theorem 4.5 ([8, Theorem 9.17]). Let B and C be bicategories and let R: C → B be a homomorphism. It is part of an adjunction if and only if there are universal arrows from c to R for each object c ∈ obC. More precisely, let cb ∈ obC and υb : b → R(cb) for b ∈ obC be universal arrows from b to R. Then there is an adjoint homomorphism L: B → C that maps b 7→ cb on objects. In particular, this assignment is part of a homomorphism of bicategories. Theorem 4.6 ([8, Theorem 9.20]). Two left adjoints L, L0 : B ⇒ C of R: C → B are equivalent, that is, there are transformations L ⇒ L0 and L0 ⇒ L that are inverse to each other up to invertible modifications. Using these general theorems, we may strengthen Proposition 3.4 (in the form of pr,∗ ⊆ C N N pr is reflective, that is, the inclusion pr,∗ → C pr → C N N N N pr has a left adjoint (reflector) L: C pr,∗. On Proposition 4.3) to an adjunction theorem: Corollary 4.7. The sub-bicategory C homomorphism R: C objects, this adjoint homomorphism maps (A,E, J) 7→ (O0 J,E ,O1 J,E , IO1 J,E). The homomorphism L is determined uniquely up to equivalence by Theorem 4.6. So we have characterised the construction of relative Cuntz -- Pimsner algebras pr,∗ ⊆ C N N in bicategorical terms, as the reflector for the full sub-bicategory C pr. By Corollary 4.7, the relative Cuntz -- Pimsner algebra construction is part of a homomorphism L: C Corollary 4.8. The relative Cuntz -- Pimsner algebras OJ,E and OJ1,E1 are Morita equivalent if there is a Morita equivalence F between E and E1 as in [15, Definition 2.1] with J · F ⊆ F · J1. pr → C N N pr,∗. For instance, this implies the following: The proof of Theorem 4.5 also describes the adjoint functor. We now describe pr → C N N the reflector L: C pr,∗ explicitly, thereby explaining part of the proof of Theorem 4.5. Much of the work in this proof is needed to check that various diagrams of 2-arrows commute. We do not repeat these computations here. The homomorphism L maps (A,E, J) 7→ (O0 J,E) on objects. Let (A,E, J) and (A1,E1, J1) be objects of C pr and let (F, u): (A,E, J) → (A1,E1, J1) be N proper covariant correspondences. We use the notation of the proof of Proposition 3.4 ∼= O1 and write ¯ιE1 for the canonical isomorphism E1 ⊗A1 O0 O0 J1,E1 from Lemma 2.17, which is the covariance part of υ(A1,E1,J1). Let J1,E1 ⊗O0 J,E ,O1 J,E , IO1 J1,E1 J1,E1 L(F, u): (O0 J,E ,O1 J,E , IO1 J,E) → (O0 J1,E1 ,O1 L(F, u) :=(cid:0)(F ⊗A1 O0 J1,E1, IO1 J1,E1 ), J1,E1)#, (u • ¯ιE1)#(cid:1) . / / / o w / / (F0 ⊗A1 O0 J1,E1) ⊗O0 given by the left action of O0 of Proposition 2.22. (F1 ⊗A2 O0 J1,E1 J1,E1 on (F1 ⊗A2 O0 J2,E2) ∼= (F0 ⊗A1 F1) ⊗A2 O0 J2,E2 A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 15 In other words, we first compose (F, u) with υ(A1,E1,J1) to get a covariant corres- pondence (F ⊗A1 O0 ) and then apply the equivalence in Proposition 3.4. J1,E1, u • ¯ιE1) from (A,E, J) to (O0 The construction on covariant correspondences above is clearly "natural", that is, functorial for isomorphisms. Explicitly, L maps an isomorphism of covariant correspondences w: (F, u) ⇒ (F0, u0) to J1,E1,O1 J1,E1 , IO1 J1 ,E1 L(w) := (w ⊗ 1O0 J1 ,E1 )# : L(F, u) ⇒ L(F0, u0). # J,E ,O1 J,E ,O1 (A,E,J) : (O0 J,E , IE) → (O0 To make L a homomorphism, we also need compatibility data for units and composition of arrows. The construction of L above maps the identity covariant correspondence on (A,E, J) to υ J,E , IE). This is canonically isomorphic to the identity covariant correspondence on (O0 J,E , IE) because the equivalence in Proposition 3.4 is by composition with υ(A,E,J). This is the unit part in our homomorphism L. Let (F, u): (A,E, J) → (A1,E1, J1) and (F1, u1): (A1,E1, J1) → (A2,E2, J2) be proper covariant correspondences. Then the homomorphism L contains isomor- phisms of covariant correspondences (4.9) which are natural for isomorphisms of covariant correspondences and satisfy some coherence conditions when we compose three covariant correspondences or compose with identity covariant correspondences. We take λ to be the isomorphism λ(cid:0)(F, u), (F1, u1)(cid:1): L(F, u) ◦ L(F1, u1) ⇒ L(cid:0)(F, u) ◦ (F1, u1)(cid:1), J,E ,O1 J2,E2) that is constructed in the proof The proof of Theorem 4.5 builds λ using only the universality of the arrows υ(A,E,J). By the equivalence of categories in Proposition 3.4, whiskering (horizontal composi- tion) with υ(A,E,J) maps isomorphisms as in (4.9) bijectively to isomorphisms (4.10) L(F1, u1) ∼= (F, u) ◦ (F1, u1) ◦ υ(A2,E2,J2) and υ(A,E,J) ◦ L(cid:0)(F, u) ◦ (F1, u1)(cid:1) ∼= The construction of L implies υ(A,E,J) ◦ L(F, u) ◦ L(F1, u1) ∼= (F, u) ◦ υ(A1,E1,J1) ◦ (cid:0)(F, u) ◦ (F1, u1)(cid:1) ◦ υ(A2,E2,J2), where we disregard associators. Hence there is υ(A,E,J) ◦ L(F, u) ◦ L(F1, u1) ⇒ υ(A,E,J) ◦ L(cid:0)(F, u) ◦ (F1, u1)(cid:1). a canonical isomorphism of covariant correspondences as in (4.10). This Ansatz produces the same isomorphisms λ as above. We have now described the data of the homomorphism L. Fiore's arguments in [8] show that it is indeed a homomorphism. Proposition 4.11. The composite of L and the crossed product homomorphism pr,∗ → C N T of Theorem 3.2. C Proof. Our homomorphisms agree on objects by Proposition 2.19. The proof of Proposition 3.4 constructed the covariant correspondence (F #, u#) by taking the degree-0 part in the correspondence constructed in the proof of Proposition 2.22. Thus we may build a natural isomorphism between the functors in question out of the nondegenerate left action of O0 (cid:3) pr → C N T is naturally isomorphic to the homomorphism C J1,E1 on OJ1,E1. So the reflector L lifts the Cuntz -- Pimsner algebra homomorphism C T to a N homomorphism with values in C pr,∗. Such a lifting should exist because a Hilbert bimodule and its crossed product with the T-action determine each other. An adjunction also contains "natural" equivalences of categories ϕb,c : C(L(b), c) ' B(b, R(c)), where naturality is further data, see Definition 4.4. In the case at hand, pr → C N 16 RALF MEYER AND CAMILA F. SEHNEM these equivalences are exactly the equivalences of categories (cid:0)(A,E, J), (B,G, IG)(cid:1) ' C pr,∗(cid:0)(O0 N υ∗ (A,E,J) : C N pr J,E), (B,G, IG)(cid:1). J,E ,O1 J,E , IO1 in Proposition 3.4. Their naturality boils down to the canonical isomorphisms of correspondences υ(A,E,J) ◦ L(F, u) ∼= (F, u) ◦ υ(A,E,J), which we have already used above to describe the multiplicativity data λ in the homomorphism L. Finally, we relate our adjunction to the colimit description of Cuntz -- Pimsner algebras in [2]. Let C and D be categories. Let CD be the category of functors D → C, which are also called diagrams of shape D in C. Identify C with the subcategory of "constant" diagrams in CD. This subcategory is reflective if and only if all D-shaped diagrams in C have a colimit, and the reflector maps a diagram to its colimit. This remains true for the bicategorical colimits in [2]: by definition, the colimit of a diagram is a universal arrow to a constant diagram. In our context, a constant N diagram in C pr is an object of the form (B, B, B) that is, the Hilbert B-bimodule is N the identity bimodule and J = B as always for objects of C pr,∗. Since the condition J · F ⊆ F · B always holds, the ideal J plays no role, compare Remark 2.25. A proper covariant correspondence (A,E, J) → (B, B, B) is equivalent to a proper correspondence F : A (cid:32) B with an isomorphism E ⊗A F ⇒ F because F ⊗B B ∼= F. As shown in [2], such a pair is equivalent to a representation (ϕ, t) of the correspondence E on F that is nondegenerate in the sense that t(E) · F = F. The properness of F means that ϕ(A) ⊆ K(F), which implies t(E) ⊆ K(F). It is shown in [2] that all diagrams of proper correspondences of any shape have a colimit. This is probably false for diagrams of non-proper correspondences, such as the correspondence '2(N): C (cid:32) C that defines the Cuntz algebra O∞. The way around this problem that we found here is to enlarge the sub-bicategory of constant diagrams, allowing diagrams of Hilbert bimodules. In addition, we added an ideal J to have enough data to build relative Cuntz -- Pimsner algebras. Since the sub-bicategory C ⊆ C N pr of constant diagrams is contained in C N pr,∗, we pr,∗ as follows. Let (A,E, J) be N may relate universal arrows to objects in C and C J,E ,O1 pr. Then υ(A,E,J) : (A,E, J) → (O0 N J,E) is a universal arrow an object of C J,E , IO1 N to an object of C pr,∗ by Proposition 4.3. The universality of υ(A,E,J) implies that a universal arrow from (A,E, J) to a constant diagram factors through υ(A,E,J), and J,E ,O1 that an arrow from (O0 J,E) to a constant diagram is universal if and only if its composite with υ(A,E,J) is universal. In other words, the diagram (A,E, J) has a colimit if and only if (O0 J,E) has one, and then the two colimits are the same. We are dealing with the same colimits as in [2] because the ideal J in (A,E, J) plays no role for arrows to constant diagrams. Appendix A. Bicategories J,E , IO1 J,E ,O1 J,E , IO1 We recall some basic definitions from bicategory theory, following [3,10]. We also give a few examples with Sections 3 and 4 in mind. Definition A.1. A bicategory B consists of the following data: • a set of objects obB; • a category B(x, y) for each pair of objects (x, y); objects of B(x, y) are called arrows (or morphisms) from x to y, and arrows in B(x, y) are called 2-arrows (or 2-morphisms); the category structure on B(x, y) gives us a unit 2-arrow 1f on each arrow f : x → y, and a vertical composition of 2-arrows: w0 : f0 ⇒ f1 and w1 : f1 ⇒ f2 compose to a 2-arrow w1 · w0 : f0 ⇒ f2; • composition functors ◦: B(y, z) × B(x, y) → B(x, z) A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS 17 for each triple of objects (x, y, z); this contains a horizontal composition of 2-arrows as displayed below: f0 x w0 ; y f1 g0 w1 g1 g0·f0 7→ x ; z w1•w0 5 z. g1·f1 • a unit arrow 1x ∈ B(x, x) for each x; • natural invertible 2-arrows (unitors) rf : f · 1x ⇒ f and lf : 1y · f ⇒ f for • natural isomorphisms all f ∈ B(x, y); B(x, y) × B(y, z) × B(z, w) (1,◦) B(x, y) × B(y, w) (◦,1) a ◦ B(x, z) × B(z, w) ◦ B(x, w); that is, natural invertible 2-arrows, called associators, a(f1, f2, f3): (f3 · f2) · f1 ⇒' f3 · (f2 · f1), where f1 : x → y, f2 : y → z and f3 : z → w. This data must make the following diagrams commute: ((f4 · f3) · f2) · f1 (f4 · f3) · (f2 · f1) 3 f4 · (f3 · (f2 · f1)) (f4 · (f3 · f2)) · f1 3 f4 · ((f3 · f2) · f1), (f2 · 1y) · f1 f2 · (1y · f1) f2 · f1, where f1, f2, f3, and f4 are composable arrows, and the 2-arrows are associators and unitors and horizontal products of them with unit 2-arrows. We write "·" or nothing for vertical products and "•" for horizontal products. Example A.2. Categories form a bicategory Cat with functors as arrows and natural transformations as 2-arrows. Here the composition of morphisms is strictly associative and unital, that is, Cat is even a 2-category. Example A.3. A category C may be regarded as a bicategory in which the categories C(x, y) have only identity arrows. Example A.4. The correspondence bicategory C is defined in [6] as the bicategory with C∗-algebras as objects, correspondences as arrows, and correspondence isomorphisms as 2-arrows. The unit arrow 1A on a C∗-algebra A is A viewed as a Hilbert A-bimodule in the canonical way. The A, B-bimodule structure on F provides the unitors A ⊗A F ⇒ F and F ⊗B B ⇒ F for a correspondence F : A (cid:32) B. The associators (E ⊗A F) ⊗B G ⇒ E ⊗A (F ⊗B G) are the obvious isomorphisms. Definition A.5. Let B,C be bicategories. A homomorphism F : B → C consists of • a map F : obB → obC between the object sets; • functors Fx,y : B(x, y) → C(F 0(x), F 0(y)) for all x, y ∈ obB; # # ; # # ; ) ) 5 / / / / / 7 + 3 + + K S + 3 % - 18 RALF MEYER AND CAMILA F. SEHNEM • natural transformations B(y, z) × B(x, y) (Fy,z,Fx,y) ◦ ϕxyz B(x, z) Fx,z C(F(y), F(w)) × C(F(x), F(y)) ◦ C(F(x), F(z)) for all triples x, y, z of objects of B; explicitly, these are natural 2-arrows ϕ(f1, f2): Fy,z(f2) · Fx,y(f1) ⇒ Fx,z(f2 · f1); • 2-arrows ϕx : 1F (x) ⇒ Fx,x(1x) for all objects x of B. This data must make the following diagrams commute: (A.6) (Fz,w(f3) · Fy,z(f2)) · Fx,y(f1) ϕ(f2,f3)•1Fx,y(f1) a0 Fz,w(f3) · (Fy,z(f2) · Fx,y(f1)) 1Fz,w(f3)•ϕ(f1,f2) Fy,w(f3 · f2) · Fx,y(f1) ϕ(f1,f3·f2) Fx,w((f3 · f2) · f1) Fx,w(a) Fz,w(f3) · Fx,z(f2 · f1) ϕ(f2·f1,f3) 3 Fx,w(f3 · (f2 · f1)); Fx,y(f1) · Fx,x(1x) ϕ(1x,f1) 3 Fx,y(f1 · 1x) (A.7) 1Fx,y(f1)•ϕx Fx,y(f1) · 1F (x) r0 Fx,y(f1) Fy,y(1y) · Fx,y(f1) ϕ(f1,1y) (A.8) ϕy•1Fx,y(f1) 1F (y) · Fx,y(f1) l0 Fx,y(f1) Fx,y(rf1 ) Fx,y(f1); 3 Fx,y(1y · f1) Fx,y(lf1 ) Fx,y(f1). Example A.9. A semigroup P may be viewed a category with one object and P as its set of arrows. It may be viewed as a bicategory as well as in Example A.3. A homo- morphism from P to C is equivalent to an essential product system (A, (Ep)p∈P op, µ) over P op as defined by Fowler [9]. The condition (A.6) says that the multiplication '→ Eqp are associative. The conditions (A.7) and (A.8) mean maps µp,q : Ep ⊗A Eq that µ1,p(a ⊗ ξ) = ϕp(a)ξ and µp,1(ξ ⊗ a) = ξa for a ∈ A, ξ ∈ Ep. A morphism f : x → y in a bicategory B induces functors f∗ : B(c, x) → B(c, y), f∗ : B(y, c) → B(x, c) for c ∈ obB by composing arrows with f and composing 2-arrows horizontally with 1f on one side (this is also called whiskering with f). Definition A.10. Let F, G: B ⇒ C be homomorphisms. A transformation α: F ⇒ G consists of • morphisms αx : F(x) → G(x) for all x ∈ obB; • natural transformations B(x, y) Fx,y / C(F(x), F(y)) Gx,y C(G(x), G(y)) αx,y ∗ αx αy∗ / C(F(x), G(y)), / / / / / 7 + 3 + + + 3 K S + + 3 K S / p x / A BICATEGORICAL INTERPRETATION FOR CUNTZ -- PIMSNER ALGEBRAS that is, 2-arrows αx,y(f): αyFx,y(f) ⇒ Gx,y(f)αx for all x, y ∈ obB. This data must make the following diagrams commute: 19 αz(Fy,z(g)Fx,y(f)) 1•ϕF (f,g) + 3 αzFx,z(gf) αx,z(gf) 3 Gx,z(gf)αx (αzFy,z(g))Fx,y(f) αy,z(g)•1 (Gy,z(g)αy)Fx,y(f) k ϕG(f,g)•1 (Gy,z(g)Gx,y(f))αx Gy,z(g)(αyFx,y(f))1•αx,y(f)+ 3 Gy,z(g)(Gx,y(f)αx). αxFx,x(1x) 1αx•ϕF x αx1F (x) r αx,x(1x) Gx,x(1x)αx x •1αx ϕG 3 αx l−1 1G(x)αx; Example A.11. Let G be a group. A transformation between homomorphisms G → C consists of a correspondence F : A (cid:32) B and isomorphisms αs : Es ⊗A F ' F ⊗B Gs so that the following diagrams commute for all s, t ∈ G: (Es ⊗A Et) ⊗A F s,t⊗1 w1 3 Est ⊗A F αst 3 F ⊗B Gst 1⊗w2 s,t F ⊗B (Gs ⊗B Gt) Es ⊗A (Et ⊗A F) 1⊗αt Es ⊗A (F ⊗B Gt) k (Es ⊗A F) ⊗B Gt αs⊗1 3 (F ⊗B Gs) ⊗B Gt. This is called a correspondence of Fell bundles (see [6, Proposition 3.23]). Definition A.12. Let α, β : F ⇒ G be transformations between homomorphisms. A modification ∆: α (cid:86) β is a family of 2-arrows ∆x : αx ⇒ βx such that for every 2-arrow w: f1 ⇒ f2 for arrows f1, f2 : x → y, the following diagram commutes: αyFx,y(f1) ∆y•Fx,y(w) 3 βyFx,y(f2) αx,y(f1) βx,y(f2) Gx,y(f1)αx Gx,y(w)•∆x + 3 Gx,y(f2)βx References [1] Beatriz Abadie, Søren Eilers, and Ruy Exel, Morita equivalence for crossed products by Hilbert C∗-bimodules, Trans. Amer. Math. Soc. 350 (1998), no. 8, 3043 -- 3054, doi: 10.1090/S0002- 9947-98-02133-3. MR 1467459 [2] Suliman Albandik and Ralf Meyer, Colimits in the correspondence bicategory, Münster J. Math. 9 (2016), 51 -- 76, doi: 10.17879/45209432019. MR 3549542 [3] Jean Bénabou, Introduction to bicategories, Reports of the Midwest Category Seminar, Springer, Berlin, 1967, pp. 1 -- 77, doi: 10.1007/BFb0074299. MR 0220789 Math. 47 (2017), no. 1, 53 -- 159, doi: 10.1216/RMJ-2017-47-1-53. MR 3619758 [4] Alcides Buss and Ralf Meyer, Inverse semigroup actions on groupoids, Rocky Mountain J. [5] Alcides Buss, Ralf Meyer, and Chenchang Zhu, Non-Hausdorff symmetries of C∗-algebras, , A higher category approach to twisted actions on C∗-algebras, Proc. Edinb. Math. Math. Ann. 352 (2012), no. 1, 73 -- 97, doi: 10.1007/s00208-010-0630-3. MR 2885576 [6] Soc. (2) 56 (2013), no. 2, 387 -- 426, doi: 10.1017/S0013091512000259. MR 3056650 K S + K S s + 3 K S k s + k s K S + + K S s + 3 + K S + 20 RALF MEYER AND CAMILA F. SEHNEM [7] Siegfried Echterhoff, Steven P. Kaliszewski, John Quigg, and Iain Raeburn, A categorical approach to imprimitivity theorems for C∗-dynamical systems, Mem. Amer. Math. Soc. 180 (2006), no. 850, viii+169, doi: 10.1090/memo/0850. MR 2203930 [8] Thomas M. Fiore, Pseudo limits, biadjoints, and pseudo algebras: categorical founda- tions of conformal field theory, Mem. Amer. Math. Soc. 182 (2006), no. 860, x+171, doi: 10.1090/memo/0860. arXiv: math/0408298. MR 2229946 [9] Neal J. Fowler, Discrete product systems of Hilbert bimodules, Pacific J. Math. 204 (2002), no. 2, 335 -- 375, doi: 10.2140/pjm.2002.204.335. MR 1907896 [10] John W. Gray, Formal category theory: adjointness for 2-categories, Lecture Notes in Math- ematics, vol. 391, Springer-Verlag, Berlin-New York, 1974. doi: 10.1007/BFb0061280 MR 0371990 available at http://www.tac.mta.ca/tac/volumes/26/14/26-14abs.html. MR 2972968 [11] Nick Gurski, Biequivalences in tricategories, Theory Appl. Categ. 26 (2012), No. 14, 349 -- 384, [12] Takeshi Katsura, On C∗-algebras associated with C∗-correspondences, J. Funct. Anal. 217 , Ideal structure of C∗-algebras associated with C∗-correspondences, Pacific J. Math. (2004), no. 2, 366 -- 401, doi: 10.1016/j.jfa.2004.03.010. MR 2102572 [13] 230 (2007), no. 1, 107 -- 145, doi: 10.2140/pjm.2007.230.107. MR 2413377 [16] [14] Bartosz Kosma Kwaśniewski and Ralf Meyer, Aperiodicity, topological freeness and pure out- erness: from group actions to Fell bundles, Studia Math. (2016), accepted. arXiv: 1611.06954. [15] Paul S. Muhly and Baruch Solel, On the Morita equivalence of tensor algebras, Proc. London , Tensor algebras over C∗-correspondences: Math. Soc. (3) 81 (2000), no. 1, 113 -- 168, doi: 10.1112/S0024611500012405. MR 1757049 representations, dilations, and C∗-envelopes, J. Funct. Anal. 158 (1998), no. 2, 389 -- 457, doi: 10.1006/jfan.1998.3294. MR 1648483 [17] Mihai V. Pimsner, A class of C∗-algebras generalizing both Cuntz -- Krieger algebras and crossed products by Z, Free probability theory (Waterloo, ON, 1995), Fields Inst. Commun., vol. 12, Amer. Math. Soc., Providence, RI, 1997, pp. 189 -- 212. MR 1426840 [18] Jürgen Schweizer, Crossed products by C∗-correspondences and Cuntz -- Pimsner algebras, C∗-Algebras (Münster, 1999), Springer, Berlin, 2000, pp. 203 -- 226, doi: 10.1007/978-3-642- 57288-3. MR 1798598 [19] Ross Street, Fibrations in bicategories, Cahiers Topologie Géom. Différentielle 21 (1980), no. 2, 111 -- 160. MR 574662 E-mail address: [email protected] E-mail address: [email protected] Mathematisches Institut, Georg-August-Universität Göttingen, Bunsenstrasse 3 -- 5, 37073, Göttingen, Germany
1107.3965	1	1107	2011-07-20T13:08:24	An Ergodic Dilation of Completely Positive Maps	[ "math.OA" ]	We shall prove the following Stinespring-type theorem: there exists a triple $(\pi,\mathcal{H},\mathbf{V})$ associated with an unital completely positive map $\Phi:\mathfrak{A}\rightarrow \mathfrak{A}$ on C* algebra $\mathfrak{A}$ with unit, where $\mathcal{H}$ is a Hilbert space, $\pi:\mathfrak{A\rightarrow B}(\mathcal{H})$ is a faithful representation and $\mathbf{V}$ is a linear isometry on $\mathcal{H}$ such that $\pi(\Phi(a)=\mathbf{V}^*\pi(a)\mathbf{V}$ for all $a$ belong to $\mathfrak{A}$. The Nagy dilation theorem, applied to isometry $\mathbf{V}$, allows to construct a dilation of ucp-map, $\Phi$, in the sense of Arveson, that satisfies ergodic properties of a $\Phi $-invariante state $\phi$ on $\mathfrak{A}$, if $\Phi$ admit a $\phi $-adjoint.	math.OA	math	An Ergodic Dilation of Completely Positive Maps Carlo Pandiscia Abstract We shall prove the following Stinespring-type theorem: there exists a triple (π, H, V) associated with an unital completely positive map Φ : A → A on C-algebra A with unit, where H is a Hilbert space, π : A → B(H) is a faithful representation and V is a linear isometry on H such that π(Φ(a) = π(a)V for all a belong to A. The Nagy dilation theorem, applied to isometry V, allows to construct V∗ a dilation of ucp-map, Φ, in the sense of Arveson, that satisfies ergodic properties of a Φ-invariante state ϕ on A, if Φ admit a ϕ-adjoint. 1 Introduction A discrete quantum process is a pair (M, Φ) consisting of a von Neumann algebra M and a normal unital completely positive map Φ on M. In this work we shall prove that any quantum process is possible dilate to quantum process where the dynamic Φ is a -endomorphism of a larger von Neumann algebra. In dynamical systems, the process of dilation has taken different meanings. Here we adopt the following definition (See Ref. Muhly-Solel [6]): Suppose M acts on Hilbert space H, a dilation of a quantum process (M, Φ) is a quadruple (R, Θ, K, z) where (R, Θ) is a quantum process with R acts on Hilbert space K and Θ is a homomorphism (i.e. -endomorphism on von Neumann algebra R) with z : H → K isometric embedding such that: • zMz∗ ⊂ R and z∗Rz ⊂ M; • Φn(a) = z∗Θn(zaz∗)z for all a ∈ M and n ∈ N; • z∗Θn(X)z = Φn(z∗Xz) for all X ∈ R and n ∈ N. Many authors in the past have been applied to problems very similar to the one we described above. We remember the work of Arveson [2] on the Eo-semigroups, of Baht-Parthasarathy on the dilations of nonconservative dynamical semigroups [3] and finally, the most recent work of Mhulay-Solel [6]. We shall prove the existence of dilation using the Nagy theorem for linear contraction (See Fojas-Nagy Ref.[7]) and of a particular covariat representation obtained through the Stinespring's theorem for com- pletely positive maps (See Stinespring Ref.[10]). We recall that a covariant representation of discrete quantum process (M, Φ) is a triple (π, H, V) where π : M → B(H) is a normal faithful representation on the Hilbert space H and V is an isometry on H such that for a ∈ M and a ∈ N, π(Φn(a)) = Vn∗π(a)Vn. Since the covariant representation is faithful and normal, we identify the von neuman algebra M with π(M) and in sec. 3 we construct a dilation of the quantum process (π(M), Ψ) where Ψ is the following completely positive map Ψ(π(x)) = π(Φ(x)) for all n ∈ M. In fact, if the triple (bV,bH, z) is the minimal unitary dilation of isometry V, we can construct a von Neumann algebrascM ⊂ B(bH) with following properties: bV∗cMbV ⊂cM and z∗cMz = M. Of fundamental importance to quantum process theory, is the ϕ-adjointness properties. The dynamic Φ admit a ϕ-adjoint (See Kummerer Ref.[4]) relative to the normal Φ-invariant state ϕ on M, if there is a normal unital completely positive map Φ♮ : M → M such that for a, b ∈ M, ϕ(Φ(a)b) = ϕ(aΦ♮(b)). 1 The relationship between reversible process, modular operator and ϕ-adjointness has been studied by Accardi-Cecchini in [1] and Majewski in [5]. In sec. 4 we shall prove that our dilation satisfies ergodic properties of a Φ-invariante state ϕ on M if the dynamic Φ admit a ϕ-adjoint. More precisely, let (R, Θ) be our dilation of quantum process (M, Φ), we shall prove that if lim n→∞ 1 n + 1 nXk=0 ϕ(aΦk(b)) − ϕ(a)ϕ(b) = 0, for all a, b ∈ M, we have lim n→∞ 1 n + 1 nXk=0 ϕ(z∗XΘk(Y )z) − ϕ(z∗Xz)ϕ(z∗Y z) = 0, for all X, Y ∈ R. For generality, we will work with concrete unital C-algebras A and unital completely positive map Φ (briefly ucp-map). The results obtained are easily extended to the quantum process (M, Φ). Before introducing the proof about existence of dilation of discrete quantum process, it is necessary to recall the fundamental Nagy dilation theorem, subject of the next section. 2 Nagy dilation theorem If V is a linear isometry on Hilbert space H, there is a triple (bV,bH, Z) where bH is a Hilbert space, Z : H→bH is a lineary isometry, while bV is an unitary operator on bH such that for n ∈ N, (1) with the following minimal properties: bVnZ = ZVn, bH = _k∈ZbVkZH. bVn =(cid:12)(cid:12)(cid:12)(cid:12) Vn C(n) 0 Wn (cid:12)(cid:12)(cid:12)(cid:12) , 2 For our purposes it is useful to recall here the structure of the unitary minimal dilation of a contraction (See Fojas-Nagy Ref.[7]). Let K be a Hilbert space, by l2(K) we denote the Hilbert space {ξ : N → K : Pn≥0 We now get the orthogonal projection F = I − VV∗ and the following Hilbert space bH = H ⊕ l2(FH) and define the following unitary operator on the Hilbert space bH: 0 W (cid:12)(cid:12)(cid:12)(cid:12) , where for each j ∈ N we have set with Πj : l2(FH) → H the canonical projections: bV =(cid:12)(cid:12)(cid:12)(cid:12) kξ(n)k2 < ∞}. V FΠ0 while W : l2(FH) → l2(FH) is the linear operator Πj(ξ0, ξ1...ξn...) = ξj , W(ξ0, ξ1...ξn...) = (ξ1, ξ2...), for all (ξ0, ξ1...ξn...) ∈ l2(FH). If Z : H → bH is the isometry defined by ZΨ = Ψ ⊕ 0 for all Ψ ∈ H, it's simple to prove that the relationships 1 and 2 are given. We observe that for each n ∈ N we have (2) (3) where C(n) : l2(FH) → H are the following operators: C(n) = Furthermore, for each n, m ∈ N we obtain: Vn−j FΠj−1, n ≥ 1. nPj=1 ΠnWm = Πn+m and ΠnWm∗ since Wm∗(ξ0, ξ1...ξn..) = (0, 0....0, while for each k and p natural number, we obtain: 0 ΠpC(k)∗ =(cid:26) FV(k−p−1)∗ } FV(k−1)∗ k−time z Ψ......FV C(k)∗Ψ = ( k > p elsewhere ∗ Ψ, FΨ, 0, .0..). { since for all Ψ ∈ H. 0 n < m =(cid:26) Πn−m n ≥ m z}{ξ0 , ξ1...), m+1 , (4) (5) (6) 3 Invariant algebra Let be A ⊂ B(H) a C-algebras with unit and V an isometry on Hilbert space H such that V∗AV ⊂ A. sition: If (bV,bH, Z) denotes the minimal unitary dilation of the isometry V we shall prove the following propo- Proposition 1. There exists a C-algebra with unit bA ⊂ B(bH) such that: 1 - ZAZ∗ ⊂bA and Z∗bAZ ⊂ A, 2 - bV∗bAbV ⊂bA, 3 - Z∗bV∗XbVZ = V∗Z∗XZV, 4 - Z∗bV∗(ZAZ∗)bV = V∗AV, first of all we want to consider some special operators on Hilbert space H. for all X ∈bA, for all A ∈ A. 3.1 The gamma operators associated to pair (A, V ) The sequences of elements of type α = (n1, n2....nr, A1, A2...Ar), with nj ∈ N and Aj ∈ A for all j = 1, 2...r, are called strings of A of length r and weight ni. For each α string of A, we associate the following operators of B(H): nPi=1 α) = A1Vn1 · · · ArV and (α = Vn∗ r Ar · · · Vn∗ 1 A1, ni and l(α) = r, while n) denote the set operators α) with · α = n and usually furthermore . α = nPi=1 n)A =nα)A : A ∈ A and α-string of A with · α = no . The symbols (n and A (n have the same obvious meaning of above. 3 Proposition 2. Let α and β are strings of A for each R ∈ A we have: and with a simple calculation Proof. For each m, n ∈ N and R ∈ A we have: · α ≥ · α < · β · β , · · if α − · β − β(cid:12)(cid:12)(cid:12)(cid:12) α(cid:19) A if β(cid:19) . (α R β) ∈ A(cid:18) · (cid:12)(cid:12)(cid:12)(cid:12) α) R β) ∈(cid:12)(cid:12)(cid:12)(cid:12) RVn∈(cid:26) V(m−n)∗ Vm∗ · α + · A m ≥ n AV(n−m) m < n (7) (8) (9) · · · α < Let α = (m1, m2....mr, A1, A2...Ar) and β = (n1, n2....ns, B1, B2...BS) strings of A, we obtain: eα < r Ar · · · Vm∗ (α R β) = Vm∗ · β it follows that while if In fact if m1 ≥ n1 we obtain: 1 A1RB1Vn1 · · · BsVns = (eα I(cid:12)(cid:12)(cid:12)eβ(cid:17) where eα and eβ are strings of A with l (eα) + l(cid:16)eβ(cid:17) = l (α) + l (β) − 1. Moreover if R1B2Vn2 · · · BsVns = (eα I(cid:12)(cid:12)(cid:12)eβ(cid:17) , 2 A2R1V(n1−m1)B2 · · · BsVns = (eα I(cid:12)(cid:12)(cid:12)eβ(cid:17) , 1 A1RB1Vn1 , eα = (m1 − n1, m2....mr, R1, A2...Ar) and eβ = (n2....ns, B2...BS). 1 A1RB1Vm1 , eα = (m2....mr, A2...Ar) and eβ = (n1 − m1, n2....ns, R1, B2...BS). where R1 = Vm∗ Then by induction on number ν = l(α) + l(β) we have the relationship 7. where R1 = Vn∗ If m1 < n1 we can write: r Ar · · · A2V(m1−n1)∗ (α R β) = Vm∗ (α R β) = Vm∗ r Ar · · · Vm∗ eβ. · α ≥ · α ≥ 1, we define the linear operators: For each α string of A with · β we have · eα ≥ · eβ Γ(α) = (αFΠ · α−1 , that will be the gamma associated operators to the pair (A, V). Proposition 3. For each α and β strings of A with satisfy the following relationship: · α, · β ≥ 1, the gamma operators associated to (A, V) Γ(α) · Γ(β)∗ ∈ A. Proof. We obtain: in fact since we have (α V ∈(cid:16) · ∗ , 0 Π α−1 · β−1 · β · β (α F β) · α = · α 6= Γ(α) · Γ(β)∗ = (αFΠ · (α F β) = (α (I − VV∗) α) = (α I α) − (α VV∗ α) ∈ A, F β) = α − 1(cid:12)(cid:12)(cid:12) while V∗ α) ∈(cid:12)(cid:12)(cid:12) α − 1(cid:17) and by relationship 7 follows that: α − 1(cid:12)(cid:12)(cid:12) I(cid:12)(cid:12)(cid:12) (cid:16) · α − 1(cid:17) ⊂ A. · · 4 We have an operator system Σ of B(l2(FH)) this is: Σ =(cid:8)T ∈ B(l2(FH)) : Γ1TΓ∗ We observe that I ∈ Σ and Γ∗ 2 ∈ A f or all gamma operators Γi associated to (A, V(cid:9) . 1AΓ2 ∈ Σ for all gamma operators Γi. Moreover Σ is a norm closed, (10) while it is a weakly closed if A is a W-algebra. 3.2 The napla operators For each α, β strings of A, A ∈ A and k ∈ N we define the napla operators of B(l2(FH)): ∆k(A, α, β) = Π ∗ · α+k Fα)A(βFΠ · β+k . For each h, k ≥ 0 we obtain the following results: ∆k(A, α, β)∗ = ∆k(A∗, β, α), and ∆k(A, α, β)·∆h(B, γ, δ) = In fact we have: 0 ∆k (R, α, ϑ) k + ∆h (R, ϑ, δ) k + . β = h + . β = h + k + . γ, h − k ≥ 0, with . γ, k − h > 0, with . . β 6= h + γ, . ϑ = . ϑ = . δ + h − k and R ∈ A . δ + k − h and R ∈ A (11) ∆k (A, α, β) · ∆h (B, γ, δ) = Π ∗ · α+k F α) A (β FΠ · β+k Π · γ+h ∗ F γ) B (δ FΠ · δ+h and if k + . β 6= h + . γ follows that Π · ∗ Π get h ≥ k, and we obtain . β = . γ + h − k ≥ = 0, while if k + · γ+h . γ . Moreover by relationship 7 . β = h + β+k . γ, without losing generality we can β − (β F γ) ∈ A(cid:16) . A (β F γ) B (δ ∈ A(cid:18) · . γ(cid:12)(cid:12)(cid:12) . β − δ + · δ + . β − . γ and a R ∈ A such that: . γ(cid:12)(cid:12)(cid:12)(cid:12) , then there exists ϑ string of A with · ϑ = Since · ϑ = · δ + h − k we have: ∆k (A, α, β) · ∆h (B, γ, δ) = Π ∗ · α+k A (β F γ) B (δ = R (ϑ . F α) R (ϑ FΠ · δ+h ∗ = Π · α+k F α) R (ϑ FΠ · ϑ+k = ∆k (R, α, ϑ) . Proposition 4. The linear space Xo generated by napla operators, is a -subalgebra of B(cid:0)l2 (FH)(cid:1) included in the operator systems Σ defined in 10. Proof. From relationship 11 the linear space Xo is a -algebra. Moreover for each gamma operators Γ (α) and Γ (β) we obtain: Γ (α) ∆k (A, γ, δ) Γ (β)∗ = (α FΠ · α−1 since by the relationships 7 and 8 we have ∗ Π · γ+k F γ) A (δ FΠ · δ+k Π · β−1 F β) ∈ A, (α FΠ · α−1 Π ∗ · γ+k F γ) A (δ FΠ · δ+k Π · β−1 F β) ∈(cid:26) (k + 1 A k + 1) 0 . α − 1 = . γ + k, . β − 1 = . δ + k elsewhere 5 In fact if . α = . γ + k + 1 we can write: (α FΠ · α−1 since ∗ Π · γ+k F γ) = (α F γ) = (α I γ) − (α VV∗ γ) ∈ A (k + 1 while if . β = . δ + k + 1 we obtain (α I γ) ∈ A (k + 1 and (α VV∗ γ) ∈ A (k + 1 (δ FΠ · δ+k Π · β−1 F β) ∈ (k + 1 A. Corollary 1. The -algebra Xo and the operator systems Σ are W-invariant: W∗XoW ⊂ Xo and W∗ΣW ⊂ Σ. Proof. Let be T belong to Σ, for each gamma operators Γ (α) and Γ (β) we have: Γ (α) (W∗TW) Γ (β)∗ = (α FΠ · α−1 W∗TWΠ · β−1 F β) = = (α FΠ · α−2 TΠ · β−2 F β) ∈ AV∗Γ1 (αo) TΓ2 (βo) VA ⊂ V∗AV ⊂A. where αo and βo are strings of A with In fact let α = (m1, m2....mr, A1, A2...Ar) by definition of gamma operator, there is i ≤ r with mi ≥ 1 such that . αo = . α − 1 and . βo = . β − 1. (α FΠ · α−2 =A1 · · · ·AiV∗(αo FΠ · α−2 =A1 · · · ·AiV∗Γ (αo) , where αo = (0, ..0, mi − 1, mi+1..mr, A1, A2...Ar) with · αo = · α − 1. Let X be the closure in norm of the -algebra Xo. Since Σ is a norm closed set, we have X ⊂ Σ while if A is a von Neumann algebra of B (H) then Σ is weakly closed and X′′ o ⊂ Σ. Proposition 5. The set S =(cid:26)(cid:12)(cid:12)(cid:12)(cid:12) A Γ1 Γ∗ 2 T (cid:12)(cid:12)(cid:12)(cid:12) : A ∈ A, T ∈ X and Γi are gamma op.of (A, V)(cid:27) , (12) is an operator system of B(cid:16)bH(cid:17) such that: Furthermore where C ∗ (S) is the C-algebra generated by the set S. bV∗SbV ⊂ S. bV∗C ∗ (S)bV ⊂ C ∗ (S) , Proof. We obtain: bV∗SbV =(cid:12)(cid:12)(cid:12)(cid:12) V∗AV C (1)∗ AV + W∗Γ∗ 2V C (1)∗ AC (1) + W∗Γ∗ V∗AC (1) + V∗Γ1W 2C (1) + C (1)∗ Γ1W + W∗TW (cid:12)(cid:12)(cid:12)(cid:12) , where the operators V∗Γ (α) W and V∗AC (1) are gamma operators associated to pair (A, V), while C (1)∗ AC (1), C (1)∗ Γ (α) W, and W∗T W are operators belonging to X. In fact we have the following relationships: V∗AC (1) = V∗AFΠ0 = Γ (ϑ) with ϑ = (1, A) . 6 while if α = (m1, m2..mr, A1, A2...Ar) we obtain: V∗Γ (α) W = V∗ (α FΠ · α−1 W =Γ (ϑ) , with ϑ = (m1 + 1, m2..mr, A1, A2...Ar) since Π · Furthermore W = Π · α . α−1 C (1)∗ AC (1) = Π ∗ 0FAFΠ0 = ∆0 (A, α, β) with α = β = (0, I) while C (1)∗ Γ (α) W = Π ∗ 0F (α FΠ · α−1 ∗ W = Π 0F γ) (α FΠ · α+0 = ∆0 (I, γ, α) with γ = (0, I) . We observe that the -algebra A∗ (S) generated by the operator system S is given by Now we can easily prove proposition 1. A∗ (S) =(cid:12)(cid:12)(cid:12)(cid:12) A XΓ∗A AΓX X (cid:12)(cid:12)(cid:12)(cid:12) . (13) Proof. We get C ∗ (S), the C-algebra generated by S defined in 12, by the definition ZAZ∗ ⊂ S then Moreover for X ∈ C ∗(S) we have: Z∗C ∗ (S) Z ⊂ A. Z∗bV∗XbVZ = VZ∗XZV, family F with inclusion is partially ordered set, then for Zorn lemma's exists a minimal element that we since bVZ = ZV. Let be F the family of C-subalgebras bB with unit of C ∗(S) such that ZAZ∗ ⊂ bB and bV∗bBbV ⊂ bB. The shall denote with bA. 4 Stinespring's theorem and dilations We examine a concrete C-algebra A of B(H) with unit and an ucp-map Φ : A → A. By the Stinespring theorem for the ucp-map Φ, we can deduce a triple (VΦ, σΦ, LΦ) constituted by a Hilbert space LΦ, a representation σΦ : A → B(LΦ) and a linear contraction VΦ : H → LΦ such thata for ∈ A, Φ(a) = V∗ ΦσΦ(a)VΦ. (14) We recall that on the algebraic tensor A ⊗ H we can define a semi-inner product by ha1 ⊗ Ψ1, a2 ⊗ Ψ2iΦ = hΨ1, Φ (a∗ 1a2) Ψ2iH , for all a1, a2 ∈ A and Ψ1, Ψ2 ∈ H furthermore the Hilbert space LΦ is the completion of the quotient space A⊗ΦH of A ⊗ H by the linear subspace {X ∈ A ⊗ H : hX, XiΦ = 0} with inner product induced by h·, ·iΦ. We shall denote the image at a ⊗ Ψ ∈ A ⊗ H in A⊗ΦH by a⊗ΦΨ, so that we have ha1⊗ΦΨ2, a2⊗ΦΨ2iLΦ = hΨ1, Φ (a∗ 1a2) Ψ2iH , for all a1, a2 ∈ A and Ψ1, Ψ2 ∈ H. Moreover σΦ (a) (x⊗ΦΨ) = ax ⊗Φ Ψ, for each x⊗ΦΨ ∈ LΦ and VΦΨ = 1⊗ΦΨ for each Ψ ∈ H. Since Φ is unital map, the linear operator VΦ is an isometry with adjoint V∗ Φ defined by V∗ Φa⊗ΦΨ = Φ(a)Ψ, 7 for all a ∈ A and Ψ ∈ H. We recall that the multiplicative domain of the ucp-map Φ : A → A is the C-subalgebra of A such defined: DΦ = {a ∈ A : Φ(a∗)Φ(a) = Φ(a∗a) and Φ(a)Φ(a∗) = Φ(aa∗)}, we have the following implications (See Paulsen Ref.[9]): a ∈ DΦ if and only if Φ(a)Φ(x) = Φ(ax) and Φ(x)Φ(a) = Φ(xa) for all x ∈ A. Proposition 6. The ucp-map Φ is a multiplicative if and only if VΦ is an unitary. Moreover if x ∈ D (Φ) we have: σΦ (x) VΦV∗ Φ = VΦV∗ ΦσΦ (x) . Proof. For each Ψ ∈ H we obtain the following implications: a⊗ΦΨ = 1⊗ΦΦ (a) Ψ ⇐⇒ Φ (a∗a) = Φ (a∗) Φ (a) , since ka⊗ΦΨ − 1⊗ΦΦ (a) Ψk = hΨ, Φ (a∗a) Ψi − hΨ, Φ (a∗) Φ (a) Ψi . Furthermore, for each a ∈ A and Ψ ∈ H we have VΦV∗ Φa⊗ΦΨ = 1⊗ΦΦ (a) Ψ. Now we prove the following Stinespring-type theorem (See Zsido Ref.[11]): Proposition 7. Let A be a concrete C-subalgebra with unit of B (H) and Φ : A → A an ucp-map, then there exists a faithful representation (π∞, H∞) of A and an isometry V∞ on Hilbert Space H∞ such that for a ∈ A, V∗ ∞π∞ (a) V∞ = π∞ (Φ (a)) , (15) where and (Vn, σn+1, Hn+1) is the Stinespring dilation of Φn for every n ≥ 0, σ0 = id, Φn = σn ◦ Φ H∞ = ∞Mj=0 and Hj , Hj = A⊗Φj−1 Hj−1, for j ≥ 1 and H0 = H; (16) V∞(Ψ0, Ψ1, Ψ2, ...) = (0, V0Ψ0, V1Ψ1, ...) for each (Ψ0, Ψ1, Ψ2, ...) ∈ H∞. Furthermore the map Φ is a homomorphism if and only if V∞V∗ ′ ∞ ∈ π∞ (A) . Proof. By the Stinespring theorem there is triple (V0, σ1, H1) such that for each a ∈ A we have Φ(a) = V∗ 0σ1(a)V0. The application a ∈ A → σ1(Φ(a)) ∈ B(H1) is a composition of cp-maps therefore it is also a cp map. Set Φ1(a) = σ1(Φ(a)). By appling the Stinespring's theorem to Φ1, we have a new triple (V1, σ2, H2) such that Φ1(a) = V∗ 1σ2(a)V1. By induction for n ≥ 1 we define Φn(a) = σn(Φ(a)) and we have a triple (Vn, σn+1, Hn+1) such that Vn : Hn → Hn+1 and Φn(a) = V∗ We get the Hilbert space H∞ defined in 16 and the injective representation of the C-algebra A on H∞ : nσn+1(a)Vn. π∞(a) =Mn≥0 σn(a) with σ0(a) = a, for each a ∈ A. Let V∞ : H∞ → H∞ be the isometry defined by V∞(Ψ0, Ψ1....Ψn...) = (0, V0Ψ0, V1Ψ1....VnΨn...), (17) (18) 8 for all Ψi ∈ Hi with i ∈ N. The adjoint operator of V∞ is V∗ ∞(Ψ0, Ψ1, ....Ψn...) = (V∗ 0Ψ1, V∗ 1Ψ2....V∗ n−1Ψn...) (19) for all Ψi ∈ Hi with i ∈ N, therefore V∗ ∞π∞ (a) V∞Ln≥0 Ψn = Ln≥0 = Ln≥0 V∗ nσn+1 (a) VnΨn = Ln≥0 σn (Φ (a)) Ψn = π∞ (Φ (a))Ln≥0 Ψn. Φn (a) Ψn = We notice that En = VnV∗ n be the orthogonal projection of B (Hn−1), we have: E (Ψ0, Ψ1...Ψn..) = (0, E0Ψ1, E1Ψ2, ...EnΨn+1...) . Finally for the proof of the last statement we only need to note that x belong to multiplicative domains D (Φ) if and only if we have: π∞ (x) V∞V∗ ∞ = V∞V∗ ∞π∞ (x) . . Remark 1. Let (M, Φ) be a quantum process, the representation π∞(a) : M → B(H∞) defined in proposition 7 is normal, since the Stinespring representation σΦ : A → B(LΦ) is a normal map. Then (π∞, H∞, V∞) is a covariant representation of quantum process. 4.1 Dilations of ucp-Maps If (H∞, π∞, V∞) is the Stinespring representation of proposition 7, we have that V∗ ∞π∞ (A) V∞ ⊂ for all X ∈bA. π∞ (A) and by proposition 1 there exists a C-algebra with unit of B(cid:16)bH(cid:17) such that: 1 - Zπ∞ (A) Z∗ ⊂bA, 2 - Z∗bAZ = π∞ (A) , 3 - Z∗bV∗XbVZ = Vπ∞ (Z∗XZ) V, Furthermore, we have a homomorphism bΦ :bA →bA thus defined bΦ(X) = bV∗XbV for all X ∈bA, such that for A ∈ A, X ∈bA and n ∈ N we have: Φn(A) = Z∗bΦn(ZAZ∗)Z Z∗bΦn(X)Z = Φn(Z∗XZ). and Φ : A → A. The quadruple (bΦ,bA, H, Z) with the above properties, is said to be a multiplicative dilation of ucp-map Remark 2. It is clear that these results are easily extended to the von Neumann algebras M with Φ normal ucp-map. In this way we obtain a dilation of discrete quantum process (M, Φ). 5 Ergodic properties Let A be a concrete C-algebra of B (H) with unit, Φ : A → A an ucp-map and ϕ a state on A such that ϕ ◦ Φ = ϕ. We recall (See N.S.Z. Ref.[8]) that the state ϕ is a ergodic state, relative to the ucp-map Φ, if lim n→∞ 1 n + 1 nXk=0 [ϕ(aΦk(b)) − ϕ(a)ϕ(b)] = 0, 9 for all a, b ∈ A, while is weakly mixing if lim n→∞ 1 n + 1 nXk=0 ϕ(aΦk(b)) − ϕ(a)ϕ(b) = 0, for all a, b ∈ A. We observe that by the Stinepring-type theorem 7 we can assume, without losing generality, that A is a concrete C-algebra of B (H), and that there is a linear isometry V on H such that: Φ (A) = V∗AV for all A ∈ A. Then(cid:16)bV,bH, Z(cid:17) is the minimal unitary dilation of (V, H) and the C-algebra bA defined in proposition 1 is included in B(bH). We want to prove the following ergodic theorem, for dilation ucp-map (bΦ,bA, H, Z) previously defined: Proposition 8. If the ucp-map Φ admits a ϕ-adjoint and ϕ is a ergodic state, we obtain: lim N→∞ 1 N + 1 NXk=0 while if ϕ is weakly mixing: [ϕ(Z∗XbVk∗ Y bVkZ) − ϕ(Z∗XZϕ(Z∗Y Z)] = 0, lim N→∞ 1 N + 1 NXk=0 ϕ(Z∗XbVk∗ Y bVkZ) − ϕ(Z∗XZ)ϕ(Z∗Y Z) = 0, obtain: for all X, Y ∈bA. If we write every element X of B(cid:16)bH(cid:17) in matrix form X =(cid:12)(cid:12)(cid:12)(cid:12) ϕ(cid:16)Z∗XbVk∗ Y bVkZ(cid:17) = ϕ(cid:0)X1,1VkY1,1Vk(cid:1) + ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) + ϕ(cid:16)X1,2Wk∗ Y2,1Vk(cid:17) Lemma 1. Let X ∈ A∗(S), the -algebra generated by operator system S defined in 12 and Y ∈bA, X1,1 X1,2 X2,1 X2,2 (cid:12)(cid:12)(cid:12)(cid:12) with bH = H ⊕ l2 (FH) we and the proof of previous proposition is an easy consequence of the following lemma: a] if ϕ is an ergodic state we have: (20) (21) b] if ϕ is weakly mixing we have: 1 lim N→∞ N + 1 lim N→∞ 1 N + 1 ϕ(cid:16)X1,2C (k)∗ Y1,1Vk + X1,2Wk∗ NPk=0 NPk=0(cid:12)(cid:12)(cid:12)ϕ(cid:16)X1,2C (k)∗ Y1,1Vk + X1,2Wk∗ =( A (γ F α) B (β FΠ · Y2,1Vk(cid:17) = 0, Y2,1Vk(cid:17)(cid:12)(cid:12)(cid:12) = 0. β+m β+m 0 Now we observe taht there is a natural number ko such that for each k > ko we obtain: X1,2Wk∗ Y2,1Vk = 0 10 Proof. Since X ∈ A∗ (S) we can assume that X1,2 = AΓ (γ) ∆m (B, α, β) with A, B ∈ A and γ string of A. Then: X1,2 = A (γ FΠ · γ−1 Π ∗ · α+m F α) B (β FΠ · . γ − 1 = . α + m elsewhere (22) In fact we have that for all (ξ0, ξ1...ξn..) ∈ l2 (FH) then Πβ+mWk∗ It follows that: k−time Wk∗ (ξ0, ξ1...ξn...) = ϕ(cid:16)X1,2C (k)∗ Y1,1Vk + X1,2Wk∗ 0, ...0 , ξ0, ξ1... , z}{ Y2,1Vk(cid:17) = lim = 0 for all k > . β + m. 1 1 N + 1 lim N→∞ NPk=0 ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) , Then we compute only the term ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) and by relationship 22 we can write that: NPk=0 N + 1 N→∞ X1,2C (k)∗ Y1,1Vk = A (γ F α) B (β FΠ · C (k)∗ Y1,1Vk β+m moreover by relationship 6 for k > . β + m we have: C (k)∗ = FV(k−β−m−1)∗ , Π · β+m it follows that X1,2C (k)∗ Y1,1Vk = A (γ F α) B (β FV(k−β−m−1)∗ Y1,1Vk = A (γ F α) B (β FΦ(k−β−1) (Y1,1) Vβ+m+1. Since · γ = · α + m + 1, by relationship 7 we obtain: it follows that there exists a ϑ string of A with · ϑ = · β + m + 1 and an operator R ∈ A, such that A (γ F α) B (β ∈ A(cid:16) . β + m + 1(cid:12)(cid:12)(cid:12) , A (γ F α) B (β = R (ϑ . Then X1,2C (k)∗ Y1,1Vk = R (ϑ FΦ(k−β−1) (Y1,1) Vβ+m+1. If we set ϑ = (n1, n2, ...nr, A1,A2, ....Ar) . we have n1 + n2 + ... + nr = . β + m + 1 and R (ϑ FΦ(cid:18)k− · β−1(cid:19) (Y1,1) V · β+m+1 = RVn∗ r ArVn∗ r−1 Ar−1 · · · A2Vn∗ 1 A1FΦ(cid:18)k− · β−1(cid:19) (Y1,1) V · β+m+1 = where We have: = RΦnr (ArΦnr−1 (Ar−1 · · · Φn2 (A2Rk))) , Rk = Φnr (Ar) Φ(k−β−1) (Y1,1) − Φnr−1(cid:16)Φ (Ar) Φ(k−β) (Y1,1)(cid:17) . ♮ ♮ ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) = ϕ (RΦnr (ArΦnr−1 (Ar−1 · · · Φn2 (A2Rk)))) = (R) ArΦnr−1 (Ar−1 (· · ·Φn2 (A2Rk) )(cid:17) = = ϕ(cid:16)Φnr (cid:16)Φnr = ϕ(cid:16)Φnr−1 = ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 ♮ (cid:16)Φn3 (cid:16)Φnr (R) Ar(cid:17) · · · A3(cid:17) A2Rk = (R) Ar(cid:17) Ar−1 (Ar−2 · · · A3Φn2 (A2Rk)(cid:17) = (R) Ar(cid:17) · · · A3(cid:17) A2Rk(cid:17) and replacing Rk, we obtain: Φn2 ♮ · · · Φnr−1 (cid:16)Φnr · · · Φnr−1 ♮ ♮ ♮ ♮ ♮ ♮ 11 ♮ ♮ ♮ ♮ · β(cid:19) (Y1,1)!!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Then: 1 ♮ ♮ ♮ ♮ ♮ ♮ ♮ ♮ ♮ ♮ ♮ 1 N + 1 N + 1 1 N + 1 ♮ ♮ = − = Φn2 −Φn2 ♮ ♮ ♮ lim N→∞ · · · Φnr−1 · · · Φnr−1 ♮ ♮ · · · Φnr−1 ♮ ♮ while lim N→∞ · · · Φnr−1 ♮ N + 1 1 It follows that : ♮ · · · Φnr−1 · · · Φnr−1 ♮ · · · Φnr−1 · · · Φnr−1 If the state ϕ is ergodic we have: (cid:16)Φnr ♮ ♮ · · · Φnr−1 (R) Ar(cid:17) · · · A3(cid:17) A2Φn1 (A1) Φ(k−β−1) (Y1,1) − ♮ (cid:16)Φn3 (cid:16)Φnr (R) Ar(cid:17) · · · A3(cid:17) A2Φn1−1(cid:0)Φ (A1) Φ(k−β) (Y1,1)(cid:1) . ♮ (cid:16)Φn3 (cid:16)Φnr ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) = = ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr (R) Ar(cid:17) · · · A3(cid:17) A2Φn1 (A1) Φ(k−β−1) (Y1,1)(cid:17) − (R) Ar(cid:17) · · · A3(cid:17) A2Φn1−1(cid:0)Φ (A1) Φ(k−β) (Y1,1)(cid:1)(cid:17). −ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) = NPk=0 (R) Ar(cid:17) · · · A3(cid:17) A2Φn1 (A1) Φ(k−β−1) (Y1,1)(cid:17) − ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr NPk=0 (R) Ar(cid:17) · · · A3(cid:17) A2Φn1−1(cid:0)Φ (A1) Φ(k−β) (Y1,1)(cid:1)(cid:17). ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr NPk=0 (R) Ar(cid:17) · · · A3(cid:17) A2Φn1−1(cid:0)Φ (A1) Φ(k−β) (Y1,1)(cid:1)(cid:17) = ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr NPk=0 (R) Ar(cid:17) · · · A3(cid:17) A2Φn1 (A1)(cid:17) ϕ (Y1,1) = = ϕ(cid:16)Φn2 ♮ (cid:16)Φn3 (R) Ar(cid:17) · · · A3(cid:17) A2(cid:17) A1(cid:17) ϕ (Y1,1) = ϕ(cid:16)Φn1 ♮ (cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr (cid:16)Φnr ♮ (cid:16)Φn3 (cid:16)Φn2 ϕ(cid:16)Φn1−1 NPk=0 (R) Ar(cid:17) · · · A3(cid:17) A2(cid:17) Φ (A1)(cid:17) ϕ (Y1,1) = (cid:16)Φnr ♮ (cid:16)Φn3 (cid:16)Φn2 = ϕ(cid:16)Φn1−1 (R) Ar(cid:17) · · · A3(cid:17) A2(cid:17)(cid:17) A1(cid:17) ϕ (Y1,1) , = ϕ(cid:16)Φ♮(cid:16)Φn1−1 (cid:16)Φn2 ♮ (cid:16)Φn3 (cid:16)Φnr ϕ(cid:0)X1,2C (k)∗ Y1,1Vk(cid:1) = 0. NPk=0 kY1,1Vk(cid:1)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) β−1(cid:19) (Y1,1)! − ϕ BΦn1−1 Φ (A1) Φ(cid:18)k− ϕ BΦn1 (A1) Φ(cid:18)k− (R) Ar(cid:17) · · · A3(cid:17) A2. ♮ (cid:16)Φn3 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) β−1(cid:19) (Y1,1)! − ϕ BΦn1−1 Φ (A1) Φ(cid:18)k− ϕ BΦn1 (A1) Φ(cid:18)k− β−1(cid:19) (Y1,1)! − ϕ (BΦn1 (A1)) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ BΦn1 (A1) Φ(cid:18)k− β(cid:19) (Y1,1)!! − ϕ (BΦn1 (A1)) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ BΦn1−1 Φ (A1) Φ(cid:18)k− (R) Ar(cid:17) · · · A3(cid:17) A2(cid:17) Φ (A1) Φ(k−β) (Y1,1)(cid:17) = Adding and subtracting the element ϕ (BΦn1 (A1)) ϕ (Y1,1) we can write: β(cid:19) (Y1,1)!!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) In weakly mixing case, using the previous results, we obtain: (cid:12)(cid:12)ϕ(cid:0)X1,2C∗ ♮ ♮ · · · Φnr−1 (cid:16)Φnr ♮ where B = Φn2 then we obtain · · · Φnr−1 ♮ · · · Φnr−1 ♮ · · · Φnr−1 ♮ ♮ ♮ lim N→∞ N + 1 1 ♮ ♮ · · + ♮ ♮ N + 1 ♮ 1 · · ≤ . · 12 Moreover · (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ BΦn1−1 Φ (A1) Φ(cid:18)k− =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ Φn1−1 (B) Φ (A1) Φ(cid:18)k− ♮ · 1 lim N→∞ N + 1 and lim N→∞ 1 N + 1 and by the weakly mixing properties we obtain: ♮ , = β(cid:19) (Y1,1)!! − ϕ (BΦn1 (A1)) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) β(cid:19) (Y1,1)! − ϕ(cid:16)Φn1−1 NXk=0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ BΦn1 (A1) Φ(cid:18)k− NXk=0(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ Φn1−1 (B) Φ (A1)(cid:17) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) β−1(cid:19) (Y1,1)! − ϕ (BΦn1 (A1)) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (B) Φ (A1)(cid:17) ϕ (Y1,1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) β(cid:19) (Y1,1)! − ϕ(cid:16)Φn1−1 (B) Φ (A1) Φ(cid:18)k− = 0, · · ♮ ♮ = 0. Finally, the proof of proposition 8 is a simple result of the previous lemma. References [1] L. Accardi and C. Cecchini: Conditional expectations in von Neumann algebras and a theorem of Takesaki,J. Funct. An.45 (1982) 245-273. [2] W. Arveson: Non commutative dynamics and Eo-semigroups. Monograph in mathematics. Springer- Verlag (2003). [3] B.V. Bath and K.R. Parthasarathy: Markov dilations of nonconservative dynamical semigroups and quantum boundary theory. Annales de lI. H. P., section B, tome 31, No 4 (1995) 601-651 [4] B. Kummerer: Markov dilations on W-algebras. - J. Funct. Anal. 63 (1985), 139-177 . [5] W.A. Majewski: On the relationship between the reversibility of dynamics and balance conditions - Annales de l'I. H. P. section A, tome 39, no.1 (1983), 45-54. [6] P.S. Muhly and B. Solel: Quantum Markov Processes (correspondeces and dilations). - Int. J. Math Vol.13, No. 8 (2002), 863-906. [7] B.Sz. Nagy and C. Foia¸s: Harmonic analysis of operators on Hilbert space - Regional Conference Series in Mathematics, n.19 (1971). [8] c. Niculescu, A. Stroh and L.Zsid´o: Non commutative extensions of classical and multiple recurrence theorems - J. Operator Theory 50 (2002), 3-52. [9] V.I. Paulsen: Completely bounded maps and dilations - Pitman Research Notes in Mathematics 146, Longman Scientific & Technical, 1986. [10] F. Stinesring: Positive functions on C algebras - Proc. Amer. Math. Soc. 6 (1955) 211-216. [11] L. Zsido: Personal communication - 2008. 13

Subsets and Splits

No saved queries yet

Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.